Advances in Intelligent Systems and Computing 1096

Hasmat Malik
Atif Iqbal
Amit Kumar Yadav   Editors

Soft Computing in
Condition Monitoring
and Diagnostics
of Electrical and
Mechanical Systems
Novel Methods for Condition Monitoring
and Diagnostics
Advances in Intelligent Systems and Computing

Volume 1096

Series Editor
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences,
Warsaw, Poland

Advisory Editors
Nikhil R. Pal, Indian Statistical Institute, Kolkata, India
Rafael Bello Perez, Faculty of Mathematics, Physics and Computing,
Universidad Central de Las Villas, Santa Clara, Cuba
Emilio S. Corchado, University of Salamanca, Salamanca, Spain
Hani Hagras, School of Computer Science and Electronic Engineering,
University of Essex, Colchester, UK
László T. Kóczy, Department of Automation, Széchenyi István University,
Gyor, Hungary
Vladik Kreinovich, Department of Computer Science, University of Texas
at El Paso, El Paso, TX, USA
Chin-Teng Lin, Department of Electrical Engineering, National Chiao
Tung University, Hsinchu, Taiwan
Jie Lu, Faculty of Engineering and Information Technology,
University of Technology Sydney, Sydney, NSW, Australia
Patricia Melin, Graduate Program of Computer Science, Tijuana Institute
of Technology, Tijuana, Mexico
Nadia Nedjah, Department of Electronics Engineering, University of Rio de Janeiro,
Rio de Janeiro, Brazil
Ngoc Thanh Nguyen , Faculty of Computer Science and Management,
Wrocław University of Technology, Wrocław, Poland
Jun Wang, Department of Mechanical and Automation Engineering,
The Chinese University of Hong Kong, Shatin, Hong Kong
The series “Advances in Intelligent Systems and Computing” contains publications
on theory, applications, and design methods of Intelligent Systems and Intelligent
Computing. Virtually all disciplines such as engineering, natural sciences, computer
and information science, ICT, economics, business, e-commerce, environment,
healthcare, life science are covered. The list of topics spans all the areas of modern
intelligent systems and computing such as: computational intelligence, soft comput-
ing including neural networks, fuzzy systems, evolutionary computing and the fusion
of these paradigms, social intelligence, ambient intelligence, computational neuro-
science, artificial life, virtual worlds and society, cognitive science and systems,
Perception and Vision, DNA and immune based systems, self-organizing and
adaptive systems, e-Learning and teaching, human-centered and human-centric
computing, recommender systems, intelligent control, robotics and mechatronics
including human-machine teaming, knowledge-based paradigms, learning para-
digms, machine ethics, intelligent data analysis, knowledge management, intelligent
agents, intelligent decision making and support, intelligent network security, trust
management, interactive entertainment, Web intelligence and multimedia.
The publications within “Advances in Intelligent Systems and Computing” are
primarily proceedings of important conferences, symposia and congresses. They
cover significant recent developments in the field, both of a foundational and
applicable character. An important characteristic feature of the series is the short
publication time and world-wide distribution. This permits a rapid and broad
dissemination of research results.
** Indexing: The books of this series are submitted to ISI Proceedings,
EI-Compendex, DBLP, SCOPUS, Google Scholar and Springerlink **

More information about this series at http://www.springer.com/series/11156

Hasmat Malik Atif Iqbal Amit Kumar Yadav
• •

Editors

Soft Computing in Condition

Monitoring and Diagnostics
of Electrical and Mechanical
Systems
Novel Methods for Condition Monitoring
and Diagnostics

123
Editors
Hasmat Malik Atif Iqbal
Department of Instrumentation Department of Electrical Engineering
and Control Engineering (ICE) Qatar University
NSUT Doha, Qatar
Delhi, India
BEARS
University Town, NUS Campus
Singapore, Singapore

Amit Kumar Yadav

Department of Electrical
and Electronics Engineering (EEE)
National Institute of Technology Sikkim
Sikkim, India

ISSN 2194-5357 ISSN 2194-5365 (electronic)

Advances in Intelligent Systems and Computing
ISBN 978-981-15-1531-6 ISBN 978-981-15-1532-3 (eBook)
https://doi.org/10.1007/978-981-15-1532-3
© Springer Nature Singapore Pte Ltd. 2020
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Preface

First of all, we are thankful to the contributors of this edited book. We are indebted
to 44 author experts in the fields of soft computing, condition monitoring, and fault
diagnosis of electrical and non-electrical equipment and substance who have
authored and co-authored their original research findings. We have combined the
state-of-the-art and advanced techniques of condition monitoring and fault diag-
nosis that will be highly beneficial to the researchers, students, and practicing
engineers. There are 22 chapters included in this volume which have undergone a
strict quality check by peer review process in several rounds. We thankfully
acknowledge the help, cooperation, and support in each stage of production of the
book by all authors. This is a unique single volume that has bring all the relevant
topics ranging from condition monitoring and fault diagnosis of electrical machines
and drives, mechanical equipment, air quality, distributed energy generations, and
smart buildings. The reliability of process and equipment is highly dependent on the
condition monitoring system used, and hence, basic understanding of methods
adopted is crucial.
Electrical machines and mechanical equipment are widely employed in all
spheres such as in industries, transportation systems, hospitals, defense systems,
homes, smart buildings, power generation plants, renewable energy systems, and
other goods and services. These equipment are used continuously for long hours
under harsh and variable environmental conditions. The maintenance requirements
vary widely depending upon numerous factors. An intelligent and effective main-
tenance scheduling is required for each equipment. The technological advancement
has brought highly complex equipment, systems, and processes. Condition moni-
toring and fault diagnosis are nowadays an integral part of any complex system, and
they play a pivotal role in the successful operation of the overall system. The failure
of equipment is related to huge economic and life losses, and hence, condition
monitoring, fault diagnosis, and fault isolation need careful considerations.
Tremendous work is carried out, and loads of literature are available in condition
monitoring, fault diagnosis, and use of soft computing in this area. This book has
made an attempt to concisely present the most important relevant literature useful
for practicing engineers. The researchers can make effective use of the available

v
vi Preface

literature, and this can help them develop and refine the process and come up with
more efficient techniques.
Condition monitoring is one of the tools used for the maintenance activities in
industries. In condition monitoring, the health of the equipment or process is
assessed by collecting important and critical data from sensors and instrumentation.
The collected data are translated into meaningful information by using different
algorithms as elaborated in the book. Soft computing employed in condition
monitoring and fault diagnosis are the latest advanced techniques that are being
adopted in the present day. Artificial intelligence-based techniques such as artificial
neural network, fuzzy logic, and combination of artificial neural network and fuzzy
logic called ANFIS are presented in this book. The processed data give an indi-
cation of the preventive maintenance of the equipment too. Smarter the data pro-
cessing algorithm, better planning and decision making can be done to reduce or
eliminate the downtime.
Over the last two decades, there has been a steady growth in the use of electric
motors as drivers in various industries, power plants, transportation sectors such as
more electric aircraft, ship propulsion, hybrid and full electric vehicles as well as oil
and gas processing plants. In particular, compressor trains are now commonly
driven by electric motors, especially for LNG production. Replacing the traditional
practice of using gas turbines as drivers, electric motors offer many advantages,
such as higher availability and reliability, reduced maintenance and shut-down
costs, more operational flexibility, less dependence on ambient conditions, and
perhaps most significantly reduced gas emissions and air pollution. With today’s
focus on global warming, there is significant interest in reducing emissions glob-
ally, while at the same time increasing energy production to meet increasing
demands. Speed and torque control in various types of electric motors require a
greater understanding of the design and the characteristics of these motors. For
example, vibration is one common characteristic of electric motors. This issue
arises from the inherent electromechanical design and characteristics of these
motors, such as the torsional pulsations resulting from unbalanced circuits in
three-phase induction motors. The resulting vibration can be used as an indicator of
many electrical problems such as wire insulation wear. Static and dynamic
eccentricities also result in vibrations that could indicate the source of fault in both
the stator and rotor. Other mechanical sources of vibration include bearing damage,
unbalance, misalignment, shaft cracks, and other fatigue-induced damages. The
resulting vibration signal can be used for health monitoring and early detection of
such faults. In addition, the use of electronic variable speed drives is another source
of vibrations, which introduce other harmonics in the driving signal. This vibration
is typically undesirable as it is transmitted to the driven equipment and may cause
failure and extended shut-down periods. These phenomena are adequately covered
in this book.
Chapter “Fault Tolerant Controller Schemes for Single and Multiple Mobile
Robots” of the book covered the fault-tolerant control strategies for actuators in a
single and multiple mobile robots. In recent times, autonomous wheeled mobile
robots (WMRs) have been widely used in different industries such as
Preface vii

manufacturing, health care, and space exploring. Mobile robots can access haz-
ardous and unsafe areas and terrain instead of humans such as nuclear power plants
and under fire buildings and mines. Autonomous mobile robots can explore other
planets in the space and log necessary data of the environmental and geographical
conditions for human benefits and scientific study. In factories, manpower is
replaced by mobile robots to decrease the costs and energy consumption. The
WMRs are presenting the future of life in any aspect where less manpower is
required and robotic technology is used in all applications. Thus, the importance of
autonomous WMRs is obvious for the coming era of technology.
Chapter “Faults and Fault Detection Methods in Electric Drives” of the book
presents a review of faults and fault detection methods in electric motor drives.
Typical faults are presented that arise from the induction motor, which is valued in
the industry for its robust construction and cost-effective production. A summary is
also presented of detectable faults in conjunction with the required physical
information that allows detection of specific faults. The reliability and efficiency of
a drive system depend not only on the control structure and inverter technology, but
also on the condition of the machine and the mechanical system. For this reason, the
industry is interested in reliable, durable, and fault-tolerant drive systems, especially
those with high responsibility, and in single-unit production drives for special
applications. Common mechanical faults are presented along with the summary of
physical forces that arise during specific types of forces. It also focuses on the
industry application and fault detection possibility in sensorless drives; the reader
will be introduced to the important international standards regarding permitted
vibrations, economic aspects of a fault detection application, and fault detection
methods.
Chapter “Introduction to Condition Monitoring of Wide Area Monitoring
System” describes the health assessment of wide area monitoring (WAM) power
system network. A massive penetration of asynchronously connected renewable
energy generation, the generation connected over inverters is significantly changing
the dynamics of modern power systems. Lately, the wide area monitoring
(WAM) system helps the power system operators to continuously analyze all the
features of a large power network in real time. Utilizing phasor measurement units
(PMUs), information can be recorded and monitored. Wide area monitoring pro-
tection and control (WAMPAC) enables the power system to detect distur-
bances and improve knowledge of network behavior under dynamic conditions,
enabling system operators to maximize power flow and network stability. This
chapter gives a basic overview of WAM and describes the components associated
with it and its applications.
In chapter “Introduction to Condition Monitoring of Electrical Systems,” a
pragmatic review has been given regarding developments related to monitoring
techniques of electrical power transmission systems. This system performs a dis-
tinguishing role of ensuring a reliable power supply for utilization by consumers for
various applications. The preponderance of faults can be minimized by monitoring
certain critical parameters and predicting a fault before it actually occurs. In this
chapter, an electric power transmission system is branched into power transformers,
viii Preface

transmission lines, power transmission towers, and protective relays, and primarily,
focus has been laid to highlight perspective view on the techniques which assist in
the development of health prognosis-based (or condition-based) maintenance tools
and practices to safeguard the health of these systems.
Chapters “Condition Monitoring, and Fault Detection and Diagnostics of Wind
Energy Conversion System (WECS)” and “Maximum Power Extraction and
Monitoring from Wind Power Generation System Using Intelligent Controllers”
elaborate the condition monitoring of wind energy system which is one of the most
important renewable energy sources. Failure of WECS results in wind turbine
(WT) shutdown or expenses maintenance that extensively affects the generated
revenue. The WECS failures result in frequent unplanned maintenance to guarantee
the reliability of wind power (WP) generation. So, condition monitoring (CM) and
fault detection and diagnosis are methodologies universally adopted for the pre-
mature fault detection in order to reduce downtime period and increase availability.
Therefore, CM and FDD of WECS’s each part (i.e., rotor, gearbox, drive trains,
generators, and power electronics) are needed. This chapter introduces and provides
an analysis of the current state of CM and FDD for each key element in WECS.
Another important distributed energy generation system is solar PV which is
elaborated in chapters “Condition Monitoring, Fault Detection and Diagnosis
(FDD) of Photovoltaic System and Its Approaches,” “Introduction to Condition
Monitoring of PV System,” “Novel Applications of Soft Computing Techniques for
Comparative Analysis of Maximum Power Point Tracking in Solar Photo-Voltaic
System Under Perturb Conditions,” and “Novel Application of Relief Algorithm in
Cascade ANN Model for Prognosis of Photovoltaic Maximum Power Under Sunny
Outdoor Condition of Sikkim India: A Case Study.” These chapters provide an
introduction about various techniques dedicated to identify and categorize the faults
along with different types of modeling in solar PV system. These chapters also deal
with the faults and their level of severity and their effects on the working and
performance of the solar PV module. Different types of computational techniques
have been explained in detail. Finally, a recommendation has been presented in the
form of the research gap of the study. Researchers suggest that automation of the
monitoring techniques, real-time processing of information, and unmanned aerial
system (UAS) technology can play an important role in the future in condition
monitoring of solar PV system. Chapter “Soft Computing Methods and Its
Applications in Condition Monitoring of DGS—A Review” provides some global
applications of soft computing methods in condition monitoring of distributed
generation system which have been outlined as an introduction. Chapter “Novel
Applications of Soft Computing Techniques for Comparative Analysis of
Maximum Power Point Tracking in Solar Photo-Voltaic System Under Perturb
Conditions” elaborates the application of soft computing techniques in maximum
power point tracking of solar PV modules. Chapter “Novel Application of Relief
Algorithm in Cascade ANN Model for Prognosis of Photovoltaic Maximum Power
Under Sunny Outdoor Condition of Sikkim India: A Case Study” presents a case
study of Sikkim, India.
Preface ix

Chapters “Introduction to Conditioning Monitoring of Mechanical Systems,”

“Theoretical and Experimental Investigation of Spur Gearbox with Cracked Pinion
Tooth Using Dynamic Model for Condition Monitoring,” “Experimental Study of
Sideband Harmonics in Vibration Spectrum of Two-Stage Planetary Gear Box for
Condition Monitoring,” and “Vibration-Based Fault Diagnosis of a Bevel and Spur
Gearbox Using Continuous Wavelet Transform and Adaptive Neuro-Fuzzy
Inference System” discuss the impact of condition monitoring in mechanical sys-
tem. The extensive use of mechanical equipment, in severe conditions, is subject to
failure and progressive deterioration from their initial condition. In mechanical
system, fault occurs due to imbalance of forces, fault in gearing system, bearing
fault, loading conditions on shaft, and improper machine selection. To detect the
fault at an early stage, the system parameters like performance, vibrations, noise,
temperature, pressure, wear at movable, and lubrication monitoring are required.
This chapter is intended to provide the fundamentals of various condition moni-
toring techniques as well as signal processing methods with application in the
mechanical system. The chapter also elaborates various transducers used to provide
the corresponding signals for condition monitoring. This chapter also discusses the
applications of condition monitoring in various mechanical systems like identifi-
cation of bearing defects, vibration monitoring of rotary machine, cutting tool
monitoring, machine tool monitoring, power plant monitoring, vehicles, guiding
rails, 3D printing, identification of gear defects, lubricant condition monitoring,
railway crossing, wind turbines, surface texture condition monitoring, and grinding
wheel condition monitoring. Modeling of gearbox is presented in chapter
“Theoretical and Experimental Investigation of Spur Gearbox with Cracked Pinion
Tooth Using Dynamic Model for Condition Monitoring.” Dynamic modeling of
gearbox is used to describe the effect of presence of gear tooth fault and the role of
change in mesh stiffness on vibration response using digital simulation of a geared
system. In chapter “Theoretical and Experimental Investigation of Spur Gearbox
with Cracked Pinion Tooth Using Dynamic Model for Condition Monitoring,” a ten
degrees-of-freedom dynamic model is formulated to demonstrate the effect of tooth
fault and mesh stiffness variation on dynamic of the geared system. Chapter
“Experimental Study of Sideband Harmonics in Vibration Spectrum of Two-Stage
Planetary Gear Box for Condition Monitoring” investigates and develops an
understanding of the sideband activities around gearmesh frequencies of a
double-stage planetary gearbox. Chapter “Vibration-Based Fault Diagnosis of a
Bevel and Spur Gearbox Using Continuous Wavelet Transform and Adaptive
Neuro-Fuzzy Inference System” is based on a research study aimed at identifying
faults in two different types of gearboxes under non-stationary conditions. The
method proposed for fault diagnosis employs the independent angular re-sampling
technique for processing the raw gearbox vibration signatures along with a hybrid
intelligent classifier, namely ANFIS, for fault diagnosis.
Population growth, urbanization, and air pollutants (AP) in the environment are
increasing significantly, posing severe health hazards. AP data provide information
about the quality of air and health risk in the surrounding which is important for
environmental management. Chapters “Novel Application of Artificial Neural
x Preface

Network Techniques for Prediction of Air Pollutants Using Stochastic Variables for
Health Monitoring: A Review” and “Artificial Neural Networks Based Condition
Monitoring of Air Pollutants for Allahabad Cities in India” reported the application
of soft computing techniques for predicting air pollution. In chapter “Novel
Application of Artificial Neural Network Techniques for Prediction of Air
Pollutants Using Stochastic Variables for Health Monitoring: A Review,” a
review is reported on air pollutant prediction using ANN techniques which are
dependent on types of prediction intervals, i.e., monthly, daily, and hourly. It is
found that the influence of different input variables, training algorithms, and
architecture changes the prediction accuracy of ANN models. A detailed discussion
is given in the chapter “Novel Application of Artificial Neural Network Techniques
for Prediction of Air Pollutants Using Stochastic Variables for Health Monitoring:
A Review.” Chapter “Artificial Neural Networks Based Condition Monitoring of
Air Pollutants for Allahabad Cities in India” discusses a case study of a city in
India.
Chapter “Intelligent System for Monitoring and Recognising the Type of
Domestic Harmonic Loads” discusses the power quality issues due to nonlinear
loads. Due to the fast progress in the power sector, most of the customers are
concerned about the effective utilization of their home appliances. However,
stepping into the next generation, unawareness in customers about the quality of
power being supplied is prevailing massively. Though there is a considerable
development in mass media and telecasting techniques, commercialization has
become the cornerstones of these organizations. Consequently, many customers are
still in darkness regarding power quality. To further contribute to the research
domain, monitoring, analysis, and recognition of the household nonlinear using
S-transform have been taken up in this chapter. Chapter “Fault Analysis of Variable
Frequency Transformer (VFT) for Power Transfer In-Between Synchronous Grids”
discusses the variable frequency transformer.
Chapter “Performance Enhancement and Extension of DGA-Based Transformer
Fault Diagnosis Methods Using Soft-Computing Techniques” discusses incipient
faults that are slowly growing faults found within oil-immersed power transformers.
These anomalies develop due to the deteriorating transformer insulating system.
Various types of electrical, mechanical, and environmental stress, manufacturing
defects, as well as normal aging phenomenon contribute to the degrading health of
transformer insulation. The degradation results in the release of gases which remain
dissolved in oil. The concentration of these dissolved gases has a direct relationship
with incipient transformer faults. This is further elaborated in this chapter. It is
possible to enhance the performance of DGA-based methods by combining more
than one soft-computing-aided DGA methods and by energy weighting of con-
centration of fault gases. A comprehensive study of application of soft computing
techniques in DGA-based transformer condition monitoring is presented in this
chapter.
Chapters “Big-Data Analytics Based Energy Analysis and Monitoring for
Multi-storey Hospital Buildings: Case Study” and “Data-Driven Intelligent Model
for Sale Price Prediction and Monitoring of a Building” are dedicated to smart
Preface xi

building. Big-Data analytics-based energy analysis and monitoring for multi-story

buildings play an important role in proper management and control of the operation
of the hospital. A case study is reported for a hospital building.
We wish to thank our colleagues and friends for their insight and helpful discussion
during the production of this edited book. We would like to highlight the contribution
of Prof. Haitham Abu-Rub, Texas A&M University at Qatar; Prof. Imtiaz Ashraf,
Aligarh Muslim University, India; Prof. Kouzou Abdellah, Djelfa University,
Algeria; Prof. Jaroslaw Guzinski, Gdansk University of Technology, Poland;
Prof. Akhtar Kalam, Victoria University of Technology, Australia; Prof. A.P. Mittal,
NSUT Delhi, India; Prof R. K Jarial, NIT Hamirpur, Himachal Pradesh, India;
Prof. Rajesh Kumar, GGSIPU, India; and Prof. Anand Parey, IIT Indore, India.
We further would like to express our love and affection to our family members,
Shadma (wife of Prof. Atif Iqbal), Abuzar and Abubaker (sons of Prof. Atif Iqbal),
Noorin (daughter of Prof. Atif Iqbal), Dr. Vibha Yadav (wife of Dr. Amit Kumar
Yadav), Avyan (son of Dr. Amit Kumar Yadav), whose passions, care, and love
have made this book a reality. We would like to express our gratitude to
Dr. N. Fatema (wife of Dr. Hasmat Malik) and Zainub and Ayesha (daughters of
Dr. Hasmat Malik) for their intense feeling of deep affection.

Delhi, India/Singapore, Singapore Dr. Hasmat Malik

Doha, Qatar Prof. Dr. Atif Iqbal
Sikkim, India Dr. Amit Kumar Yadav
Contents

A Conceptual Introduction and Review on Condition Monitoring

and Diagnostics of Electrical and Mechanical Systems
Fault Tolerant Controller Schemes for Single and Multiple
Mobile Robots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Parisa Yazdjerdi and Nader Meskin
Faults and Fault Detection Methods in Electric Drives . . . . . . . . . . . . . 57
Patrick D. Strankowski and Jarosław Guziński
Introduction to Condition Monitoring of Wide Area Monitoring
System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Papia Ray and Debani Prasad Mishra
Introduction to Condition Monitoring of Electrical Systems . . . . . . . . . 91
Atul J. Patil, Arush Singh and R. K. Jarial
Condition Monitoring, and Fault Detection and Diagnostics
of Wind Energy Conversion System (WECS) . . . . . . . . . . . . . . . . . . . . . 121
Tarkeshwar Mahto, Hasmat Malik and V. Mukherjee
Condition Monitoring, Fault Detection and Diagnosis (FDD)
of Photovoltaic System and Its Approaches . . . . . . . . . . . . . . . . . . . . . . 155
Omkar Singh, Amit Kumar Yadav and Anjan Kumar Ray
Introduction to Condition Monitoring of PV System . . . . . . . . . . . . . . . 169
Peeyush Kala, Puneet Joshi, Sanjay Agrawal, Lokesh K. Yadav
and Medha Joshi
Soft Computing Methods and Its Applications in Condition
Monitoring of DGS—A Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
Puneet Joshi, Sanjay Agrawal, Lokesh K. Yadav, Medha Joshi, Vikas Patel
and Peeyush Kala

xiii
xiv Contents

Introduction to Conditioning Monitoring of Mechanical Systems . . . . . . 205

Akhilesh Kumar Choudhary and Dilshad Ahmad Khan
Novel Application of Artificial Neural Network Techniques
for Prediction of Air Pollutants Using Stochastic Variables
for Health Monitoring: A Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Vibha Yadav and Satyendra Nath

Application of Soft Computing in Condition Monitoring

and Diagnostics of Electrical and Mechanical Systems
Intelligent System for Monitoring and Recognising the Type
of Domestic Harmonic Loads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
Pullabhatla Srikanth and Chiranjib Koley
Fault Analysis of Variable Frequency Transformer (VFT)
for Power Transfer In-Between Synchronous Grids . . . . . . . . . . . . . . . . 269
Farhad Ilahi Bakhsh, Dheeraj Kumar Khatod and Hasmat Malik
Performance Enhancement and Extension of DGA-Based
Transformer Fault Diagnosis Methods Using Soft-Computing
Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
Shakeb Ahmad Khan, M. Ajmal Khan, Shufali Ashraf Wani
and Md. Danish Equbal
Big-Data Analytics Based Energy Analysis and Monitoring
for Multi-storey Hospital Buildings: Case Study . . . . . . . . . . . . . . . . . . 325
Nuzhat Fatema, Hasmat Malik and Atif Iqbal
Maximum Power Extraction and Monitoring from Wind Power
Generation System Using Intelligent Controllers . . . . . . . . . . . . . . . . . . 345
Aabid Hussain Sheikh and Farhad Ilahi Bakhsh
Novel Applications of Soft Computing Techniques for Comparative
Analysis of Maximum Power Point Tracking in Solar Photo-Voltaic
System Under Perturb Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
Mohammad Junaid Khan, Lini Mathew and Amit Kumar Yadav
Novel Application of Relief Algorithm in Cascade ANN Model
for Prognosis of Photovoltaic Maximum Power Under Sunny Outdoor
Condition of Sikkim India: A Case Study . . . . . . . . . . . . . . . . . . . . . . . 387
Vinay Kumar Reddy Chimmula, Amit Kumar Yadav and Hasmat Malik
Data-Driven Intelligent Model for Sale Price Prediction
and Monitoring of a Building . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
Nuzhat Fatema, Hasmat Malik and Atif Iqbal
Contents xv

Artificial Neural Networks Based Condition Monitoring

of Air Pollutants for Allahabad Cities in India . . . . . . . . . . . . . . . . . . . 423
Vibha Yadav and Satyendra Nath
Theoretical and Experimental Investigation of Spur Gearbox
with Cracked Pinion Tooth Using Dynamic Model for Condition
Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
Yogesh Pandya and Anand Parey
Experimental Study of Sideband Harmonics in Vibration Spectrum
of Two-Stage Planetary Gear Box for Condition Monitoring . . . . . . . . . 465
Yogesh Pandya and Anand Parey
Vibration-Based Fault Diagnosis of a Bevel and Spur Gearbox
Using Continuous Wavelet Transform and Adaptive Neuro-Fuzzy
Inference System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473
Amandeep Singh Ahuja, Dada Saheb Ramteke and Anand Parey
About the Editors

Hasmat Malik (M’16) received Diploma in Electrical

Engineering from Aryabhatt Govt. Polytechnic Delhi,
B.Tech. degree in electrical and electronics engineering
from the GGSIP University, Delhi, M.Tech. degree in
electrical engineering from National Institute of
Technology (NIT), Hamirpur, Himachal Pradesh, and
Ph.D. in power system from Electrical Engineering
Department, Indian Institute of Technology (IIT) Delhi,
India. He is currently a Postdoctoral Scholar at BEARS,
University Town, NUS Campus, Singapore, and an
Assistant Professor (on leave) at Division of
Instrumentation and Control Engineering, Netaji
Subhas University of Technology, Delhi, India. He is
a Life Member of Indian Society for Technical
Education (ISTE), Institution of Electronics and
Telecommunication Engineering (IETE), International
Association of Engineers, Hong Kong (IAENG),
International Society for Research and Development,
London (ISRD), and Member of the Institute of
Electrical and Electronics Engineers (IEEE), USA,
and MIR Labs, Asia. He has published more than 100
research articles, including papers in international
journals, conferences, and book chapters. He is a
Guest Editor of Special Issue of Journal of Intelligent &
Fuzzy Systems, 2018. (SCI, Impact Factor 2019:1.637),
(IOS Press). He received the POSOCO Power System
Award (PPSA-2017) for his Ph.D. work for research
and innovation in the area of power system. He has
received best research papers awards at IEEE
INDICON-2015, and full registration fee at IEEE
SSD-2012 (Germany). His principle areas of research

xvii
xviii About the Editors

interests are artificial intelligence, machine learning and

big data analytics for smart building & automation,
condition monitoring and online fault detection &
diagnosis (FDD).

Atif Iqbal, Fellow IET (UK), Fellow IE (India) and

Senior Member IEEE, DSc (Poland), PhD (UK)-
Associate Editor IEEE ACCESS, Editor-in-Chief,
I’manager journal of Electrical Engineering, Former
Associate Editor IEEE Trans. On Industry Application.
Full Professor at the Dept. of Electrical Engineering,
Qatar University and Former Full Professor at Electrical
Engineering, Aligarh Muslim University (AMU),
Aligarh, India.Recipient of Outstanding Faculty Merit
Award academic year 2014-2015 and Research excel-
lence awards 2015 and 2019 at Qatar University, Doha,
Qatar. He received his B.Sc. (Gold Medal) and M.Sc.
Engineering (Power System & Drives) degrees in 1991
and 1996, respectively, from the Aligarh Muslim
University (AMU), Aligarh, India and PhD in 2006
from Liverpool John Moores University, Liverpool, UK.
He obtained DSc (Habiliation) from Gdansk University
of Technology in Control, Informatics and Electrical
Engineering in 2019. He has been employed as a
Lecturer in the Department of Electrical Engineering,
AMU, Aligarh since 1991 where he served as Full
Professor until Aug. 2016. He is recipient of Maulana
Tufail Ahmad Gold Medal for standing first at B.Sc.
Engg. (Electrical) Exams in 1991 from AMU. He has
received several best research papers awards e.g. at IEEE
ICIT-2013, IET-SEISCON-2013, SIGMA 2018, and
IEEE CENCON 2019. He has published widely in
International Journals and Conferences his research
findings related to Power Electronics, Variable Speed
Drives and Renewable Energy Sources. Dr. Iqbal has
authored/co-authored more than 390 research papers and
two books and three chapters in two other books. He has
supervised several large R&D projects worth more than 8
million USD. He has supervised and co-supervised
several PhD students. His principal area of research
interest is Smart Grid, Complex Energy Transition,
Active Distribution Network, Electric Vehicles drive-
train, Sustainable Development and Energy Security,
Distributed Energy Generation.
About the Editors xix

Dr. Amit Kumar Yadav received B.Tech. in

Electrical and Electronics Engineering in 2009 from
UCER Naini Allahabad U.P., India, M.Tech. in Power
System in 2011 and Ph.D. in artificial neural
network-based prediction of solar radiation for opti-
mum sizing of photovoltaic systems for power gener-
ation in 2016 from Centre for Energy and
Environmental Engineering National Institute of
Technology, Hamirpur, Himachal Pradesh, India.
Currently, he is an Assistant Professor in Electrical
and Electronics Engineering Department, National
Institute of Technology, Sikkim. His research interests
include solar radiation and wind speed prediction for
power generation, hybrid systems, artificial intelli-
gence, optimization techniques and condition monitor-
ing of power apparatus. He supervised more than 10
undergraduate projects and 1 M.Tech. projects. He is
Editorial Board Member in Turkish Journal of
Forecasting. He was awarded with “Research Ratna
Awards 2019” for “Best Researcher In Solar
Photovoltaic Systems For Maximum Power
Generation” by Research Under Literal Access
(RULA) International Awards. He has authored 11
science citation index international journals, 10 Scopus
index international journals, 5 Springer and Elsevier
book chapters and 12 IEEE Conference Publications.
Most of the research papers are of impact factor 10.59.
The h-index of research papers is 12, i-10 index is 14,
and total citation of papers is more than 950.
He is reviewer of IET Science, Measurement &
Technology, Neural Computing and Applications
(Springer), Applied Energy (Elsevier), International
Energy Journal, Electric Power Components and
Systems Journal (Wiley), ISA Transactions (Elsevier),
Sustainable Energy Technology and Assessment
(Elsevier), Journal of Renewable and Sustainable
Energy (American Institute of Physics), Jordanian
Journal of Computers and Information Technology,
IEEE Transaction on Industrial Electronics, International
Journal of Electrical Power and Energy System (IJEP),
Elsevier, Journal of Cleaner Production, Elsevier,
Renewable and Sustainable Energy Review, Elsevier,
Solar Energy Elsevier, Science and Technology for the
Built Environment Taylor and Francis Journal.
A Conceptual Introduction and Review on
Condition Monitoring and Diagnostics of
Electrical and Mechanical Systems
Fault Tolerant Controller Schemes for
Single and Multiple Mobile Robots

Parisa Yazdjerdi and Nader Meskin

Abstract In this chapter, an actuator fault tolerant controller is developed for both
single and multiple differential drive mobile robots. First, a fault tolerant controller
is proposed for loss of effectiveness actuator faults in differential drive mobile robots
while tracking the desired trajectory. The actuator loss of effectiveness fault is injected
on the kinematic equation of the robot as a multiplicative gain in the left and right
wheels angular velocity. Accordingly, the goal is to estimate and tolerate the injected
actuator faults. A fault diagnosis method based on joint state and parameter estimation
scheme is proposed to estimate the actuator loss of effectiveness gains as well as the
states of the system. The estimated values of actuator faults are then used in the
controller to compensate their effects on the mobile robots performance. Next, the
proposed Fault Tolerant Controller (FTC) method is extended for the leader–follower
formation control of mobile robots in the presence of actuator fault in a leader and
followers robots. An Extended Kalman Filter (EKF) is utilized for each robot to
obtain the parameters and states of the system and as the actuator fault is detected
in any of the robots, the corresponding controller compensates the fault. Finally, an
obstacle avoidance feature is added to the designed actuator fault tolerant controller,
for single mobile robot. Toward this goal, a go-to goal controller is designed for the
mobile robot to reach the desired destination. Meanwhile, the robot can detect any
obstacle with a given color specification using Kinect sensor mounted on the robot
and consequently avoid collisions. The efficacy of the proposed FTC framework is
demonstrated for both single and multiple mobile robots by real-time simulation and
implementation on Qbot-2 from Quanser.

P. Yazdjerdi · N. Meskin (B)

Qatar University, Doha, Qatar
e-mail: [email protected]
P. Yazdjerdi
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 3

H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_1
4 P. Yazdjerdi and N. Meskin

1 Introduction

In recent years, autonomous Wheeled Mobile Robots (WMR) have been widely used
in several industries such as manufacturing, agriculture, service industry, health care,
and space exploring. Mobile robots can access hazardous areas instead of human such
as nuclear power plant [1] and under fire area [2]. Autonomous mobile robots can
explore other planets in the space and log environmental and geographical data. In
manufacturing factories, manpower is replaced by robots to cut down the costs and
energy consumption [3]. For instance, WMRs can transfer material from a random
origin to a desired target position [4, 5]. The WMRs are presenting the future of life
in any aspect where less manpower is required and robotic technology is used in all
applications. Thus, the significance of autonomous WMRs is obvious for the coming
era of technology [6].
Wheeled mobile robots are of different types depending on the number of actua-
tors, type of wheels, and the position of wheels. A brief description for the well-known
type of wheeled mobile robots is provided in this paragraph. Unicycle mobile robot
or self-balancing robot with two-dimensional movement has only one wheel with an
actuator [7]. Unicycle robots with Omni wheel provides more flexibility and smooth-
ness in movement [8]. A differential drive mobile robot has two separately driven
wheels on each side of the robot and it may have an Omni wheel in the front side of
the robot [9]. Differential drive mobile robots with Omni wheels provide slippage in
the motion of the robot. Tricycle robot has two fixed wheels at back and one steer-
ing wheel embedded with an actuator in the front or two steering wheels attached
with motors in front and a rear wheel [10]. Four-wheel drive mobile robots can have
different types, for example, all four wheels have a separate actuator, only two front
or rear wheels have a separate actuator, and each front wheel has an actuator while
rear wheels have one actuator on its axis [11]. The dynamic and kinematic equations
of the robot differ based on the position of actuators. Different types of the mobile
robot can be chosen according to the desired application and environment.
Multiple mobile robots have also a wide range of applications in the recent years
[1–3, 12, 13]. Multiple mobile robots can be used instead of manpower under the
fired area to rescue people in the fire. Also, multiple mobile robots can be used to
carry things by keeping the desired formation [14]. In addition, they are useful in
agriculture to ease human life [12, 15]. For mining [16] and military applications
[13], mobile robots are used to rescue human health due to the risky environment
in such applications. However, to accomplish the described tasks successfully, they
should have intelligence and operate safe and reliable. Due to the exponential growth
of systems complexity and automation, the demand for safe and reliable operation is
increasing which results in ameliorating the quality and robustness of the system’s
components such as sensors and actuators. Thus, the main challenge is to identify the
main source of the fault or failure among all components in the system, distinguish
its effect, take appropriate action, and differentiate faults from disturbances such as
wind and friction [17, 18].
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 5

Fault Detection and Isolation (FDI) of sensors and actuators and fault accommoda-
tion are among the most important problems for the development of an autonomous,
safe, and reliable WMRs. The allocated tasks to WMRs are generally complex, and
hence, they are prone to different faults in their components. This fact highlights
the importance of the development of fault detection and isolation as well as fault
tolerant control algorithms in WMRs.
The general objective of this chapter is to design a fault tolerant controller for
differential drive mobile robots considering different missions, namely a trajectory
tracking and obstacle avoidance controller for a single mobile robot, as well as a
formation control scenario for a group of mobile robots.

2 Literature Review on Fault Detection and Isolation

of Mobile Robots

2.1 Single Mobile Robot

In this section, previous works on fault detection and isolation for different types of
mobile robots are presented.

2.1.1 Differential Drive Mobile Robot

For detection and identification of component faults in a mobile robot, a bank of

Kalman filters is proposed in [19]. Left wheel flat, right wheel flat, and periodic-
bump (attaching a small object to one of the wheels) are the three types of process
faults which are considered and modeled in the kinematic equation. It should be
noted that any changes in the radius of the wheels will affect both linear and angular
velocities. The proposed approach is experimentally validated on Pioneer-I robot.
A prediction error based fault detection method is presented in [20] where process
faults such as wheel radius change and general kinematic parameter changes are
considered. Based on the kinematic and dynamic model of the robot, a torque filtering
technique is employed to develop a prediction error based fault detection scheme.
In [21], a neural network fault diagnosis approach along with change detection
algorithm was proposed for actuators fault detection on the mobile robots where total
actuator failure is considered. The proposed approach is validated by the implemen-
tation on a mobile robot Robosoft.
A fault detection approach based on deviation from the expected behavior of
the robot is described in [22]. This approach can successfully detect and identify
different faults such as total failure in sensors and motors. In addition, using the
described algorithm provides the probability of fault occurrence during the operation.
Experimental results prove the efficiency and accuracy of the described algorithm. A
sensor fault detection approach is proposed in [23] based on an extended Kalman filter
6 P. Yazdjerdi and N. Meskin

and using redundant positioning systems. A cumulative sum is also used to monitor
the residual of the extended Kalman filter. The proposed approach is implemented
and demonstrated on a differential drive mobile robot.
A model-based sensor fault detection and isolation of the differential drive
unmanned ground vehicle is proposed and experimentally validated in [24]. Fault
detection for additive sensor fault is done subject to the evaluation of residual gener-
ation module which is built analytically. A fault detection and isolation algorithm is
designed in [25] based on local neural network framework. Two types of faults are
considered as actuator Loss Of Effectiveness (LOE) fault and slippage in the wheels.
The LOE fault is detected by comparing the output of the neural model and the
encoder reading. Also, using an accelerometer, slippage fault is detected by tracking
the robot velocity. Experimental results validate the proposed FDI method.
A Robust Nonlinear Analytical Redundancy (RNLAR) FDI technique is pre-
sented in [26] where both sensor and actuator faults are considered as additive terms
in the kinematic model. The proposed RNLAR method can detect faults in the pres-
ence of model-plant-mismatch and process disturbances. Primary residual vectors
are designed using RNLAR such that they are highly sensitive to the sensor and actu-
ator faults and less sensitive to model-plant-mismatch and process disturbances. To
isolate the fault, the primary residual vectors are transformed into a set of structured
residual vectors. The experimental results on a Pioneer 3-DX are used to verify the
effectiveness of the proposed approach.
A data-driven fault detection method is presented in [27] based on local linear
models and adaptive thresholding. Actuator loss of effectiveness faults, change in the
motor resistance, and the wheel radius change are considered. The proposed method
based on the adaptive threshold is more sensitive to faults compared to the one with
a fixed threshold value. Only simulation results are presented in [27].
A model-based fault diagnosis scheme is proposed for WMR in [28] where three
types of process faults, namely right wheel radius change, left wheel radius change,
and parameter change are considered in the dynamic model of the robot. The model
of each fault is obtained by changing the parameter value of the faulty component in
the dynamic equation. To detect the fault, a nonlinear observer is designed based on
the robot dynamic equation. After fault detection, a warning signal is sent to activate
the fault isolation module. By using three generalized observer scheme in the fault
isolation module, all considered process faults are successfully isolated. The results
are verified by simulation in this paper.
A novel fault detection approach is presented in [29] where a partial loss of
effectiveness fault in the actuator is considered. The proposed method is based on
the comparison between the simulation and the real-time behavior of the robot. The
experiment has been done on the obstacle avoidance behavior of the robot. The fault
is injected by reducing the speed of the actuator using a modified controller.
An implementation and experimental validation of trajectory tracking and model-
based fault detection and isolation algorithm of Automatic Guided Vehicle (AGV)
system based on multiple positioning modules are described in [30]. The kinematic
model of the differential drive AGV is described and two main types of faults are
considered, namely sensor power cut fault and actuator total failure. A fault detection
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 7

algorithm is designed based on multiple positioning modules, where encoders, laser

scanner, and laser navigation system have been used to obtain the position infor-
mation. The fault is detected through comparison between two or more positioning
systems using EKF. Fault isolation is achieved by examining the largest residual.
A model-based fault detection method is used in [31] for different types of actu-
ator faults namely bias fault, drift fault, and power cut. The fault is detected using
analytical redundancy approach in which an observer tracks the deviation between
the nominal dynamic of robot and sensor readings. The fault is detected if the value of
calculated deviation exceeds a predefined threshold. Experimental results on iRobot
Roomba verify the efficiency of the described algorithm.
Table 1 summarizes the developed FDI algorithms for differential drive mobile
robots.

2.1.2 Car-Like Mobile Robot

A hybrid fault detection and identification approach is proposed in [32] where wheels
radius change faults and two sensor faults namely total failure and gain degradation
fault are considered. The model of the system behavior under each fault type is
embedded in a multiple parallel Kalman Filter (KF) estimator where each KF is tuned
to a particular fault and predict the expected values for the sensor reading using its
embedded model. The residual is used as an indicator to check the efficiency of each

Table 1 Summary of FDI for differential drive mobile robot

Ref# Approach Fault type Method Simulation(S)
Implementation(I)
Roumeliotis et al. [19] Model-based Process Bank of KFs I
Dixon et al. [20] Model-based Process Prediction error None
Skoundrianos and Data-driven Actuator Neural network I
Tzafestas [21]
Verma [22] Model-based Sensor and actuator Deviation from I
expected value
Sundvall and Model-based Sensor EKF and redundant I
Jensfelt [23] positioning
Monteriu et al. [24] Model-based Sensor Enhanced residual I
analysis
Sidek and Sarkar [25] Data-driven Actuator and slippage Local neural network I
Halder and Sarkar [26] Model-based Sensor and actuator Analytical redundancy I
Baghernezhad and Data-driven Actuator and process Adaptive threshold S
Khorasani [27]
Hoang and Kang [28] Model-based Process Nonlinear observer S
Millard et al. [29] Data-driven Actuator Comparison between I
simulation and real
Pratama et al. [30] Model-based Sensor and actuator Multiple positioning I
Stavrou et al. [31] Model-based Actuator Analytical redundancy I
8 P. Yazdjerdi and N. Meskin

Table 2 Summary of FDI for car-like mobile robot

Ref# Approach Fault type Method Simulation(S)
Implementation(I)
Goel et al. [32] Model-based Sensor and Multiple-model I
process estimation and
neural network
Hashimoto et Model-based Sensor Variable structure I
al. [33] interacting
multiple-model
estimator
Tsarouhas and Data-driven Sensor, actuator Statistical I
Fourlas [34] and process approach

filter and a neural network approach is used for residual evaluation. Implementation
on a physical robot is presented to validate the proposed approach.
A multi-model-based fault detection approach of internal sensors for a four-wheel
mobile robot is presented in [33] where three sensor faults, namely bias fault, loss of
accuracy fault, and gain degradation fault of the sensors are handled. The detection
and diagnosis of the bias/loss of accuracy fault is based on the variable structure inter-
acting multiple-model estimator where the failure modes are modeled as switching
from one mode to another in a probabilistic manner and the fault decision is done
based on the probability of each mode. The detection and diagnosis of the gain
degradation fault is achieved via a single-model-based Kalman filter. This approach
is validated by experimental results.
The statistical approach is used in [34] to study the reliability of the failure data.
Five different types of faults are considered such as sensor drift fault, process fault (flat
tire), loss of wireless communication, and power cut for both sensor and actuator. The
reliability of the proposed fault detection scheme is verified according to experiments
done on Pioneer 3-AT.
Table 2 summarizes the developed FDI algorithms for car-like drive mobile robots.

2.1.3 Robots with Tracks and Wheels (Treels)

A generalized synthesizing fault detection method for autonomous treels is proposed

in [35]. Two types of faults are analyzed where the first type of fault causes one or
both treels stop moving and the second type occurs if the software crashes, leaving
the one or both motors running at some undefined speed. Basically, the sensory
data and control signals to actuators recorded over a period of time when a robot
is operating as intended and then when various types of actuator faults have been
injected. Neural networks are trained to detect fault from the changes in the flow
of data. Experimental results are used to verify the performance of the FDI system
using S-bot.
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 9

Back-propagation neural network is used in [36] to detect the fault based on the
collected sensory data during the robot operation. Several sensor and actuator faults
are considered such as total failure in treels and loss of accuracy fault on each or
both treels. The described approach is experimentally validated on S-bot.
In [37], a sensor fusion methodology based on the extended Kalman filter is
proposed that not only improves the robot navigation, but provides fault awareness
of Inertial Measurement Unit (IMU) sensor failure, encoder drift, and other failures
(such as a temporary total failure fault). The described method is validated through
implementation.

2.2 Multi-Mobile Robots

An external sensor is used to detect and isolate faults in [38] for a group of robots. Each
robot has two sensors which provide velocity information to other robots (encoder
and gyro). Thus, each robot in the team can monitor states (fault or fault-free) of in-
vehicle sensors by comparing velocities. On the other hand, using laser scan images
captured by each robot’s Laser Range Sensor (LRS) and a leader’s LRS, the position
is estimated. When the robot detects a fault, it compares two estimated position values
and identifies the faulty component (dead reckoning or LRS). Hardware redundancy
approach is used to design the FDI system. The effectiveness of results is validated
experimentally.
A distributed fault diagnosis scheme for faults in actuator is designed in [39]
for group of robots operating with a distributed observer controller approach. Each
robot has a local observer and it receives a vector from neighbor robots to estimate
the overall system state. Based on the estimated state of each mobile robot, a residual
signal is generated in each robot. The residual dynamics are analytically investigated
in both the absence and presence of the fault. Thus, to detect and isolate faults of
each robot, the residual vectors are compared with an adaptive threshold. Healthy
robots can detect and isolate the faulty robots even if they are not directly connected.
Results are experimentally validated using four differential drive mobile robots.

3 Literature Review on Fault Tolerant of Mobile Robots

3.1 Single Mobile Robot

A fault-adaptive control methodology for mobile robots is presented in [40] for

actuator faults. The physical processes of the robot are modeled using bond graphs,
combined qualitative FDI and quantitative estimation schemes are used for fault
detection and isolation. For the fault tolerant control design, two cases are considered
where in the first one, a robust motion controller based on the dynamic and the
10 P. Yazdjerdi and N. Meskin

geometry of the path is designed and in the second case, a switching controller is
developed. It is shown that for a small incipient actuator fault, a robust controller
achieves fault adjustment without a considerable performance loss. However, for
larger faults, the supervisory controller needs to switch among several controllers to
provide a good performance. Simulation results are presented to show the efficiency
of this method.
A fault tolerant controller based on radial basis function neural network is pro-
posed in [41]. This method does not need fault detection and diagnosis methods. A
trajectory tracking control problem for the speed and azimuth of the robot is solved
using this method. It is shown by simulation that in the presence of actuator faults,
the system can track the desired trajectory.
A self-tuning fuzzy PD controller is designed to enhance a control strategy for a
differential drive mobile robot in [42]. The dynamic model of the robot is simulated
in MATLAB/Simulink. Temporary disturbances or partial failure of the sensor is
modeled and self-tuning fuzzy PD controller is used to improve the efficiency of the
controller in the presence of actuator faults. The described method is validated by
simulation.
A method of active fault tolerant control based on fault compensation is presented
in [43] for differential drive mobile robots. The kinematic model of the mobile robot
is investigated and three types of actuator faults are added to the kinematic model
namely power cut, partial failure, and bias fault. The method which is used to solve
the problem of fault detection and isolation is based on an evolutionary algorithm.
The fault tolerant controller is designed based on fault compensation corresponding
to the identified fault. Experimental results verify the effectiveness of the designed
approach.
A fault tolerant controller for the four-wheel mobile robot in the presence of
actuator faults is studied in [44]. The fault is modeled as a multiplicative loss of
effectiveness gain. FDI is achieved by generating residuals using structural analy-
sis method. Then, FTC is designed using recursive least squares approximation to
the residuals to calculate new inputs for the system to compensate for the gains.
Experimental result on Pioneer 3-AT mobile robot has validated the efficiency of
the described approach. A fault tolerant controller for differential drive mobile robot
is designed in [45] for loss of actuator effectiveness. In addition, considering the
dynamic model of the system, backstepping method is used to design the controller.
A tracking controller is developed for nominal cases and as soon as the fault occurs
in the system, an auxiliary controller is activated to achieve the desired goal. The
efficiency of this approach is verified by simulation.
Fault tolerant control in [46] consists of two blocks which are fault estimation
and fault tolerant control. An adaptive observer is designed for a differential drive
mobile robot dynamic to estimate the power cut actuator fault. Then, a linear quadratic
method is used to reconfigure the controller in the fault tolerant block to adapt with the
corresponding fault. Simulation is carried out to verify the approach. Fault tolerant
control based on inverse dynamic control algorithm is designed in [47]. An actuator
fault is modeled and added to the dynamic model of the system. Then, a joint state
and parameter estimation method is used to estimate the actuator effectiveness gain.
Finally, the active failure model is incorporated into the inverse dynamic controller to
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 11

be able to reconfigure the controller in the presence of the actuator fault. Simulation
is done to verify the performance of the proposed method.
To design a fault tolerant controller for a differential drive mobile robot to com-
pensate for the actuator faults, a fault hiding method is used in [48]. First, a sliding
mode controller is used to accomplish the trajectory tracking task for the kinematic
equation of the robot. Then, a fault hiding block is added to the system in which the
effect of the LOE fault will be hidden. Simulation is done to estimate the effectiveness
of the proposed method.

3.2 Multi-Mobile Robots

In [49], a FTC approach is described for multi-mobile robot formation supported

by the hierarchy graph. In this model, all of the robots are in communication with
each other regardless if they are leader or follower. Two types of fault are considered
namely, communication loss or fault in the robot mechanism. Fault detection is done
based on the communication over all robots in the team. For instance, if one of the
robots losses its communication, other robots in the team can notify the disconnection.
Also, if the failure in the robot mechanism cause the faulty robot to stop moving,
other robots are able to identify the faulty robot as its position will be constant for a
long time. Hierarchy graph which is introduced for the formation of robots is used to
reconfigure the formation of the team in the presence of the fault. Experiments on the
team of mobile robots are done to verify the performance of the proposed method.
An adaptive robust controller is described in [50] for multiple mobile robots
in the existence of actuator loss of effectiveness fault which is based on dynamic
surface control method. Simulation results verify that the trajectory tracking of multi-
mobile robots in the existence of actuator loss of effectiveness fault is accomplished
successfully.
A new fault tolerant controller for the group of differential drive mobile robots
is described in [51]. The LOE fault is modeled in the kinematic equation and fault
detection and isolation is achieved using a two-stage EKF. Fault tolerant controller
is designed to take decision according to the fault severity. If the faulty robot is not
able to complete the mission, FTC reassigns the mission to the other healthy robot or
the faulty robot can continue with degraded performance. In addition, the obstacle
avoidance algorithm is designed to avoid any collision with the faulty robot based
on the sensory data of healthy robots. Simulation results are used to validate the
performance of the team of robots in the presence of the fault.
As mentioned above, wide range of methods are used in the literature to tackle the
problem of fault detection, isolation, and identification for different types of robots.
However, in all the model-based methods for differential drive mobile robots, all three
states of the robot are required to estimate the fault information and compensate
for it which is counted as a limitation for the current algorithms. Moreover, the
reconfiguration of the controller in the presence of the fault is another limitation of
12 P. Yazdjerdi and N. Meskin

the introduced approaches in the literature. In addition, using Kalman filter for the
nonlinear kinematic of robots is subjected to less accuracy and flexibility when the
joint state and parameter estimation method is applied in the existing approaches.
The proposed methods in this chapter are designed to overcome the limitations of
the existing algorithms.

4 Qbot-2 and Kinect

Proposed FTC algorithms in this chapter are implemented on Qbot-2 from Quanser
and this section presents a brief description of this robot and its featurs. The spec-
ification of Qbot-2 and Kinect sensor mounted on the robot is provided briefly in
this section. Also, the process of detecting and measuring the size of the obstacle is
presented in detail. According to the data-sheet, Qbot-2 from Quanser is equipped
with the Microsoft Kinect and a Quanser embedded target with a wireless embedded
computer. Qbot-2 consist of a Yujin Robot Kobuki platform, a Microsoft Kinect RGB
camera, and a depth sensor. Qbot-2 from Quanser is a differential drive mobile robot
with two central wheels located on a common axis and two Omni wheels at the front
and back for forward and backward movement of the robot. Encoders are used to
measure the movement of each wheel and orientation angel of the robot is measured
using a gyro sensor. The software platform of Qbot-2 is compatible with MATLAB
and Simulink with Quarc library to provide real-time communication as an external
target by extending the code generation capabilities of Simulink. Qbot-2 is accessi-
ble through three different block sets: Quanser Hardware-In-the-Loop (HIL) block
set to read from sensors and/or write to outputs, Quanser Stream API block set to
perform communications over wired and wireless communication channels, Quanser
Multimedia block set to read RGB and depth image data from the Kinect sensor, and
MathWorks Computer Vision System Toolbox to perform image processing. Qbot-2
physical specification is presented in Table 3.

Table 3 Specification of Qbot-2

Symbol Description Value Unit
D Qbot-2 diameter 0.35 Meter
p Distance between the left and right wheel 0.235 Meter
r Wheels radius 0.035 Meter
d Distance between the robot center and 0 Meter
wheels axle
h Height of the robot 0.27 Meter
m Total mass of the robot 3.29 Kilogram
vmax Maximum linear velocity of robot 0.7 Meter/Seconds
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 13

A Microsoft Kinect sensor is mounted on Qbot-2 from Quanser which provides

the depth image and the distance to the object. This sensor has the ability to detect
objects which are located between 0.5–5 m from the robot. The horizontal field of
view of Kinect sensor is 57◦ and the depth image is presented as a 640 × 480 image.
It should be noted that Kinect sensor needs to identify the object prior to measuring
its coordinate. The identification is done based on the color of the object by using a
Simulink block called “Find Object” in which the user can specify the RGB values
of the desired object to be detected. Accordingly, using the “Find Object” block in
Simulink, the RGB image of the object is extracted from the surrounding areas and
the object is shown as black while other area in the image is ignored. Figures 1 and
2 show the RGB image and the extracted black and white image, respectively. The
distance from the camera to the object in millimeter is presented by the value of
each pixel in the depth image. The (x, y) coordinate of each pixel is calculated by
mapping the depth image to the world coordinate frame which can be captured by
Kinect sensor at the same time of the depth image. For instance, as shown in Fig. 3,
the (x P , y P ) coordinate of the object is calculated using the 400 pixel in the depth
image which is the distance from the camera toward the point “p” on the object.
It should be noted that only one row of the depth image is considered in Fig. 3 for

Fig. 1 RGB image from Kinect “Get image” block

Fig. 2 Extracted black and

white image from “Find
Object” block
14 P. Yazdjerdi and N. Meskin

Fig. 3 Depth image data from Kinect sensor

processing simplicity. According to this figure, the (x P , y P ) coordinate in the mobile

reference frame is calculated as follows:

x p = dP ,
(1)
y p = L sin θ = d tan θ

The angle θ can be calculated based on the distance to the desired point (at 400th
pixel) and horizontal field of view as follows:

57
θ = (400 − 320) × , (2)
640
where 320 refers to the central pixel of sensor data. In this chapter, the size of
the image is reduced to 128 × 96 for simplicity in the processing. The Cartesian
coordinates of an obstacle is measured as follows using depth image and extracted
black and white image of the obstacle. The black and white image of the obstacle
is presented as a matrix with ones and zeros that has same size of the depth image.
Each pixel of the black object is presented as zero and each white pixel corresponds
to the unwanted area is assigned to one. Accordingly, the black pixels are extracted
from the whole image and results in an only black image. Then, the xobst coordinate
of the center of the obstacle is calculated by finding the median pixel in rows and
columns of the black image and the corresponding value of that pixel in the depth
image. Similarly, the distance to the first and last points of the object is measured
by mapping the first and last pixels of the black image (See Fig. 4). The angles α
and β between the horizontal field of view and lines to the first and last points are
calculated as follows, respectively:
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 15

Fig. 4 Image processing

57 π
α = (64 − min(column)) × ×
128 180 (3)
57 π
β = (max(column) − 64) × ×
128 180
where 64 refers to the central pixel of the image, 57 degree is camera’s total field
of view, and min(column) and max(column) present the first and last points of the
object, respectively. The summation of α and β results in the angle between the
first and last points of the object γ . Now, considering that the distance to the first
point of the object is a, the distance to the last point of the object is b, and the angle
between them is γ , the radius robst of the obstacle is calculated as follows using law
of Cosines:
1 2
robst = (a + b2 + 2ab cos γ ). (4)
2
Finally, the yobst coordinate of the center of the obstacle is calculated as follows:

y1 + y2
yobst = , (5)
2
where

y1 = a sin α,
(6)
y2 = b sin β.
16 P. Yazdjerdi and N. Meskin

5 FDI and FTC of Mobile Robots Using Joint State

and Parameter Estimation

In this section, a fault tolerant controller is proposed for loss of effectiveness actua-
tor faults in differential drive mobile robots. A joint state and parameter estimation
approach is utilized to estimate the actuator loss of effectiveness gains as the parame-
ters of the system based on the extended Kalman filter technique [52]. The estimated
actuator faults are then used in the controller to compensate their effects. Moreover, a
fault tolerant leader–follower scheme is presented to control multiple robots moving
on a desired trajectory with the desired formation in the presence of actuator faults
in the followers or the leader. The performance of the proposed method is validated
by the real-time simulation and implementation on Qbot-2. The fault tolerant con-
trol algorithm for single and multiple differential drive mobile robots using the joint
state and parameter estimation approach is widely studied in the literature. In [53],
this method is used to estimate the parameters fault on robots and it is validated
by simulation. Also, in [54], augmented EKF is used for the localization of robots.
This method is used in the biomedical field for parameter and state estimation of the
hemodynamic nonlinear dynamic system representation [55].

5.1 Kinematic Equation

Differential drive mobile robots are used in this chapter due to their simple dynamic
and kinematic model and their low costs. For a differential drive mobile robot shown
in Fig. 5, x(t) and y(t) are the coordinates of the robot along x-axis and y-axis,
respectively. In addition, d is the distance between the robot center and wheels axle
and φ(t) denotes as heading of the robot. The kinematic model can be written as
follows:
⎡ ⎤ ⎡ ⎤
ẋ(t) cos φ(t) −d sin φ(t)  
⎣ ẏ(t) ⎦ = ⎣ sin φ(t) d cos φ(t) ⎦ v(t) . (7)
ω(t)
φ̇(t) 0 1

Assuming the non-rolling condition, the relation between right and left wheel
velocities and the linear and angular velocities of the robot are as follows:

rR ωR (t) − rL ωL (t) rR ωR (t) + rL ωL (t)

ω(t) = , v(t) = , (8)
p 2

where ωR (t) and ωL (t) represent the angular velocities of right and left wheels,
respectively, p is the axle length, rL and rR are the radius of left and right wheels,
respectively. Moreover, the wheels angular velocities which are used to control the
robot, can be obtained as follows:
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 17

v(t) pω(t)
ωL (t) = − ,
rL 2rL
(9)
v(t) pω(t)
ωR (t) = + .
rR 2rR

In this chapter, the LOE fault studied in the wheels/ actuators can be formulated
as follows [6]:

ω L f (t) = (1 − K L )ω L (t),
(10)
ω R f (t) = (1 − K R )ω R (t),

where ω L f (t) and ω R f (t) present the angular velocities of the right and left faulty
wheels, respectively. The primary objective is now to approximate K L and K R using
the joint state and parameter estimation method and to cancel out the consequence
of the approximated fault on the performance of mobile robots. In order to reach this
objective, the augmented discrete-time nonlinear model of the robot is obtained as fol-
lows:
⎡ ⎤ ⎡ ⎤
x(k + 1) x(k) + T v(k) cos φ(k)
⎢ y(k + 1) ⎥ ⎢ y(k) + T v(k) sin φ(k) ⎥
⎢ ⎥ ⎢ ⎥
⎢ φ(k + 1) ⎥ = ⎢ φ(k) + T ω(k) ⎥,
⎢ ⎥ ⎢ ⎥
⎣ K L (k + 1)⎦ ⎣ K L (k) ⎦
K R (k + 1) K R (k)

which can also be expressed as

⎡ ⎤
f 1 (Z (k), u(k))
⎢ f 2 (Z (k), u(k))⎥
⎢ ⎥
Z (k + 1) = f (Z (k), u(k)) = ⎢ ⎥
⎢ f 3 (Z (k), u(k))⎥ , (11)
⎣ K L (k) ⎦
K R (k)

where

Fig. 5 Schematic of a
differential drive robot
18 P. Yazdjerdi and N. Meskin

Z (k) = [x(k), y(k), φ(k), K L (k), K R (k)]T ,

u(k) = [ω L (k), ω R (k)]T ,


Tr    
f 1 (Z (k), u(k)) = x(k) + 1 − K R (k) ω R (k) + 1 − K L (k) ω L (k) cos φ(k),
2 (12)


Tr    
f 2 (Z (k), u(k)) = y(k) + 1 − K R (k) ω R (k) + 1 − K L (k) ω L (k) sin φ(k),
2


Tr    
f 3 (Z (k), u(k)) = φ(k) + 1 − K R (k) ω R (k) − 1 − K L (k) ω L (k) .
p

5.2 Joint State and Parameter Estimation—EKF

In this part, an extended Kalman filter is utilized to approximate the augmented states
Z (k) of the nonlinear system (11) which results in the approximation of the loss of
effectiveness fault values K L and K R . The predication step of the extended Kalman
filter is given as follows [52]:

Ẑ (k + 1|k) = f ( Ẑ (k|k), u(k)),

(13)
P(k + 1|k) = Φ(k)P(k|k)AT (k) + Q(k),

where
∂ f 
Φ(k) =  . (14)
∂ x Ẑ (k+1|k),u(k)

Then, the measurement update step is done as follows:

v(k + 1) = z(k + 1) − C Ẑ (k + 1|k),

S(k + 1) = CP(k + 1|k)C T + R,
K (k + 1) = P(k + 1|k)CS−1 (k + 1), (15)
P(k + 1|k + 1) = (I − K (k + 1))P(k + 1|k),
Ẑ (k + 1|k + 1) = Ẑ (k + 1|k) + K (k + 1)v(k + 1).

The estimated numerical quantity of LOE actuator faults, K̂ L and K̂ R estimated

by the EKF algorithm will be used to compensate effects of actuator fault on the
trajectory tracking operation of robots. Based on the obtained numerical quantities
of loss of effectiveness faults in the right and left wheels, the robot’s control input
ω L (t) and ω R (t) are calculated as follows:
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 19

Fig. 6 General architecture of the proposed FTC

1 pω(t) 
 v(t)
ω L (t) = − ,
1 − K̂ L r 2r
1  v(t) pω(t) 
(16)
ω R (t) = + .
1 − K̂ R r 2r

It is clear from (10) and (16) that as far as K̂ L ≈ K L and K̂ R ≈ K R , the outcome of
loss of effectiveness faults is compensated by the described FTC approach such that
ω R (t) and ω L (t) in (16) do not surpass the highest admissible speed of the wheels
of the robot. Figure 6 shows the general architecture of the proposed FTC algorithm.
It should be noted that the Qbot-2 gyro sensor generates a sensor measurement
in radian which varies as [−π, π ]. It is observed during the implementation that
the jump from −π to π leads to jump in the parameter estimation. Therefore, the
measurement used for the proposed EKF approach is z(k) = [x(k), y(k)]T and it is
demonstrated that fault values K L and K R can be estimated perfectly without the
heading measurement information [6]. The Q and R values are selected as follows
for the EKF:
⎡ ⎤
1 0 0 0 0
⎢0 1 0 0 0⎥  
⎢ ⎥ 1 0
−5 ⎢ ⎥
Q = 5 × 10 × ⎢0 0 1 0 0⎥ , R = 0.001 ×
⎣0 0 0 1 0⎦ 0 1
0 0 0 0 1
20 P. Yazdjerdi and N. Meskin

5.3 Fault Tolerant Control of Single Mobile Robot in

Trajectory Tracking Mission

The real-time simulation and implementation of the proposed FTC algorithm are
presented in this section for a single mobile robot to track a trajectory in the presence
of multiplicative LOE actuator fault. The real-time simulation is done using a “Real-
time Synchronization” block in Simulink that synchronizes the simulation time with
real time.
To verify the efficacy of the proposed method, a trajectory tracking controller is
considered as follows [56]:

 
v(t) = vd (t) cos(φd (t) − φ(t)) + k1 (t) cos φ(xd (t) − x(t)) + sin φ(yd (t) − y(t)) ,
ω(t) = ωd (t) + k2 sign(vd (t))[cos φ(yd (t) − y(t)) − sin φ(xd (t) − x(t))]
+k3 (t)(φd (t) − φ(t)),

where xd (t), yd (t), and φd (t), t ≥ 0 present the desired trajectories of x and y posi-
tions and the heading angle, respectively, and vd (t) and ωd (t) present the desired
linear and angular velocities for the desired trajectory, respectively. The controller
gains are given as follows:

k1 (t) = k3 (t) = 2ζ ωd2 (t) + bvd2 (t), (17)
k2 (t) = b|vd (t)|, (18)

where ζ = 0.7 and b = 10 are obtained by trial and error. The portable robot control
input ω L (t) and ω R (t) are obtained as follows:

v(t) pω(t)
ω R (t) = + , (19)
rR 2r R
v(t) pω(t)
ω L (t) = − . (20)
rL 2r L

Three fault scenarios are considered in this section for a single mobile robot. Also,
the real-time simulation and implementation results for each scenarios are provided
and comparison between their performance is done. The desired trajectory for all
three scenarios is a circle with a radius of 0.5 m.
Scenario 1: 50% LOE fault in the left actuator at t = 50 s, i.e. K L = 0.5
Figure 7 shows the estimated left and right actuator gains for the real-time simu-
lation. It is obvious that K̂ L is correctly estimated at 55.64 s and K̂ R is remained
unchanged. Figure 8 shows the estimated values of K L and K R by the proposed joint
state and parameter estimation method implemented on Qbot-2. As shown in this
figure, the fault value K̂ L is correctly estimated at 63.44 s and the right actuator gain
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 21

K̂ L
0.5

0
0 10 20 30 40 50 60 70 80 90 100
Time(sec)

1
K̂ R

0.5

0
0 10 20 30 40 50 60 70 80 90 100
Time(sec)

Fig. 7 The estimated fault parameter corresponding to 50% LOE fault in the left wheel injected at
t = 50 s in the real-time simulation

K̂ R remained unchanged after a slight transient response at the fault injection time.
By comparing the real-time simulation and implementation results, it is observed
that the detection and isolation time in the real-time simulation is less than that in the
implementation due to an unavoidable error of dead reckoning and negligible surface
roughness. In other words, due to the surrounding environment, the robot is expe-
riencing some errors which result in slower fault detection and isolation compared
to the real-time simulation. Moreover, Fig. 9 shows that the desirable trajectory is
tracked in the presence of 50% loss of effectiveness fault in the left actuator fault.
The fault recovery time for this scenario is measured as t = 71.81 s. In most of
the scenarios, there is no difference between the tracked trajectory in the real-time
simulation and implementation. Accordingly, figures corresponding to the tracked
trajectory in the real-time simulation are not provided here.
Scenario 2: 50% LOE fault in the right actuator at t = 50 s, i.e. K R = 0.5
In the second scenario, 50% LOE fault is injected in the right wheel at t = 50 s, i.e.,
K R = 0.5. As shown in Figs. 10 and 11, the fault values are correctly estimated by the
proposed EKF scheme at t = 60.61 s and t = 62.5 s in the real-time simulation and
implementation, respectively. Figure 12 shows the trajectory tracking performance
of the robot in the presence of 50% LOE fault in the right actuator with the fault
recovery of 71.36 s.
Scenario 3: 30 and 80% LOE fault in the left and right wheels at t = 25 s and
t = 50 s, respectively
Figure 13 delineates the estimated K̂ L and K̂ R in the case of concurrent faults in the
real-time simulation. It is shown that K̂ L and K̂ R are estimated correctly at 31.83 s
22 P. Yazdjerdi and N. Meskin

K̂L
0.5

0
0 10 20 30 40 50 60 70 80 90 100
Time(sec)

1
K̂R

0.5

0
0 10 20 30 40 50 60 70 80 90 100
Time(sec)

Fig. 8 The estimated fault parameter corresponding to 50% LOE fault in the left wheel injected at
t = 50 s in the implementation

0.6
Tracked Trajectory
Desired Trajectory
0.4

0.2
y

-0.2

-0.4

-0.6
-0.6 -0.4 -0.2 0 0.2 0.4 0.6
x

Fig. 9 Desired versus tracked trajectory corresponding to 50% LOE fault injected in the left wheel
in the implementation
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 23

K̂L 0.5

0
0 10 20 30 40 50 60 70 80 90 100
Time(sec)

0.5
K̂R

0
0 10 20 30 40 50 60 70 80 90 100
Time(sec)

Fig. 10 The estimated fault parameter corresponding to 50% LOE fault in the right wheel injected
at t = 50 s in the real-time simulation

0.5

0
K̂ L

-0.5

-1
0 10 20 30 40 50 60 70 80 90 100
Time(sec)

0.6

0.4
K̂ R

0.2

0
0 10 20 30 40 50 60 70 80 90 100
Time(sec)

Fig. 11 The estimated fault values corresponding to 50% LOE fault in the right wheel injected at
t = 50 s in the implementation
24 P. Yazdjerdi and N. Meskin

0.6
Tracked Trajectory
Desired Trajectory
0.4

0.2
y

-0.2

-0.4

-0.6
-0.6 -0.4 -0.2 0 0.2 0.4 0.6
x

Fig. 12 Desired versus tracked trajectory corresponding to 50% LOE fault in the right wheel in
the implementation

and 55.72 s, respectively. From Fig. 14, it is obvious that in the implementation, the
numerical value of fault K̂ L and K̂ R are approximated correctly at t = 32.89 s and
t = 64.91 s, respectively. It is obvious that for both faults, the detection time in the
real-time simulation is less than that in the implementation. Moreover, Figs. 15 and
16 demonstrate that the desirable trajectory is tracked perfectly in the presence of
simultaneous multiplicative actuator LOE fault. The tracked trajectories do not have
any observable difference in the real-time simulation and implementation.

5.4 Fault Tolerant Control of Multiple Mobile Robot in

Leader–Follower Mission

In this section, the efficiency of the proposed FTC approach is verified by the real-time
simulation and implementation. The simulation and implementation is done for three
robots using a formation control with one leader and two followers. First, the detailed
of the leader–follower formation control of mobile robots is provided. It is assumed
that the leader controller is independent from the followers and is designed based
on the trajectory tracking controller presented in preceding section. Accordingly, the
follower’s controller design is explained in this part. As demonstrated in Fig. 17, the
corresponding coordinate of the leader in the follower’s local frame can be expressed
as follows [57]:
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 25

K̂ L
0.5

0
0 10 20 30 40 50 60 70 80 90 100
Time(sec)
1
K̂ R

0.5

0
0 10 20 30 40 50 60 70 80 90 100
Time(sec)

Fig. 13 The estimated fault parameter corresponding to 30% LOE fault in left wheel injected
at t = 25 s and 80% LOE fault in right wheel injected at t = 50 s, respectively in the real-time
simulation

1
K̂L

0.5

0
0 10 20 30 40 50 60 70 80 90 100
Time(sec)

1
K̂R

0.5

0
0 10 20 30 40 50 60 70 80 90 100
Time(sec)

Fig. 14 The estimated fault parameter corresponding to 30% LOE fault in left wheel injected at
t = 25 s and 80% LOE fault in right wheel injected at t = 50 s, respectively in the implementation
26 P. Yazdjerdi and N. Meskin

0.6
Tracked Trajectory
Desired Trajectory
0.4

0.2
y

-0.2

-0.4

-0.6
-0.6 -0.4 -0.2 0 0.2 0.4 0.6
x

Fig. 15 Desired versus tracked trajectory corresponding to 30 and 80% LOE fault in the left and
right wheels, respectively in the real-time simulation

0.6
Tracked Trajectory
Desired Trajectory
0.4

0.2
y

-0.2

-0.4

-0.6
-0.6 -0.4 -0.2 0 0.2 0.4 0.6
x

Fig. 16 Desired versus tracked trajectory corresponding to 30 and 80% LOE fault in the left and
right wheels, respectively in the implementation
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 27

Fig. 17 Relative coordinate

of leader and follower

⎡ ⎤ ⎡ 1⎤
ρ(t) ((x L (t) − x(t))2 + (y L (t) − y(t))2 ) 2
⎣ψ(t)⎦ = ⎣ arctan2( xyLL (t)−y(t) ) − φ(t) ⎦, (21)
(t)−x(t)
γ (t) φ L (t) − φ(t)

where (x L (t), y L (t), φ L (t)) and (x(t), y(t), φ(t)) are the Cartesian coordinates of the
leader and follower, (ρ(t), ψ(t)) denotes the leader polar coordinate in the follower’s
local reference frame, and γ (t) presents the deviation between the leader and the
follower heading angel. By differentiating (21) and using (7), the error dynamics is
expressed as below

ρ̇(t) = −v(t) cos ψ(t) + v L (t) cos(γ (t) − ψ(t)),

sin ψ(t) sin(γ (t) − ψ(t))
ψ̇(t) = v(t) − ω(t) + v L (t) , (22)
ρ(t) ρ(t)
γ̇ (t) = ω L (t) − ω(t),

where ω L (t) and v L (t) are the angular and linear velocities of the leader, respectively.
The goal of the follower is to keep ρ(t) and ψ(t) at their fixed values (ρd , ψd ).
The following controller is obtained based on the feedback linearization approach:

v L (t) cos(γ (t) − ψ(t))

v(t) = + v̄(t),
cos ψ(t)
(23)
v L (t) sin(γ (t) − ψ(t)) v(t) sin(ψ(t))
ω(t) = + + ω̄(t),
ρ(t) ρ(t)

where
28 P. Yazdjerdi and N. Meskin

ρ
v̄(t) = K ρp eρ (t) + K I eρ (τ )dτ, (24)

ψ
ω̄(t) = K ψp eψ (t) + K I eψ (τ )dτ, (25)

with eρ (t) = ρ(t) − ρd and eψ (t) = ψ(t) − ψd .

A circle with radius of 1.2 m is the desired trajectory for the leader. To conduct
the desirable formation, the relative coordinate of two followers are demonstrated
ρ
in Table 4. The proportional and integral gains for follower 1 are selected as K P1 =
ρ ψ ψ
0.15, K I1 = 0.005, K P1 = 0.15, and K I1 = 0.001. Similarly, the proportional and
ρ ρ ψ
integral gains for follower 2 are selected as K P2 = 0.15, K I2 = 0.005, K P2 = 0.15,
ψ
and K I2 = 0.001. Two scenarios are considered in this section to show the efficiency
of the proposed algorithm.
Scenario 1: Concurrent LOE fault in both actuators of the leader
In the first scenario, 50 and 30% LOE faults are injected in the left and right wheels
of the leader at t = 200 s and t = 350 s, respectively. Figure 18 shows that 50% LOE
in the left actuator is estimated at 202.6 s and 30% LOE fault in the right actuator
is estimated at 358.1 s in the real-time simulation. Also, as shown in Fig. 19, the
proposed FTC scheme can perfectly estimate both gains ( K̂ L , K̂ R ) in the leader at
t = 205.8 s and t = 360.01 s, respectively, in the implementation. Figures 20 and
21 verify that all three robots keep the desired formation while tracking a circular
trajectory in the presence of concurrent faults in the leader in the real-time simulation
and implementation, respectively.
Scenario 2: Concurrent LOE fault in both actuators of the both followers
In the second scenario, multiple faults are injected in both wheels of both followers as
follows: 50% LOE fault at t = 200 s in the right wheel of follower 1, 70% LOE fault at
t = 200 s in the left wheel of follower 2 , 30% LOE fault at t = 350 s in the left wheel
of follower 1, and 20% LOE fault at t = 350 s in the right wheel of the follower 2.
Figures 22 and 23 delineate that K̂ L and K̂ R are approximated properly in the right and
left wheels of the follower 1 and follower 2 in the real-time simulation, respectively.
Also, Figs. 24 and 25 delineate that fault parameters are properly approximated in
both followers in the implementation. Figures 26 and 27 show the error signal on
eρ (t) for follower 1 and follower 2, respectively. It is obvious from these figures that
the desired distance between both followers and the leader is kept constant in the
existence of multiple faults in both followers. Likewise, Figs. 28 and 29 show that the

Table 4 Desirable relative coordinates of followers

ρd (m) ψd
Follower 1 1 −π/4
Follower 2 0.8 π/4
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 29

Estimated left and right gain-Leader

0.6

0.4
K̂L
0.2

-0.2
0 50 100 150 200 250 300 350 400 450 500
time
0.4

0.2
K̂R

-0.2
0 50 100 150 200 250 300 350 400 450 500
time

Fig. 18 The estimated fault values corresponding to 50 and 30% LOE in the left and right wheel
injected at t = 200 and t = 350 s in the real-time simulation, respectively

Estimated Left and right gain-Leader

0.6

0.4
K̂L

0.2

-0.2
0 50 100 150 200 250 300 350 400 450 500
time
0.4

0.3
K̂R

0.2

0.1

0
0 50 100 150 200 250 300 350 400 450 500
time

Fig. 19 The estimated fault values corresponding to 50 and 30% LOE in the left and right wheel
injected at t = 200 and t = 350 s in the implementation, respectively
30 P. Yazdjerdi and N. Meskin

Trajectory Tracked
2
Leader
1.5 Follower1
Follower2

0.5
y

-0.5

-1

-1.5

-2
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
x

Fig. 20 The tracked trajectory corresponding to 50 and 30% LOE in the left and right actuator of
leader injected at t = 200 and t = 350 s in the real-time simulation, respectively

Trajectory Tracked
2
Leader
1.5 Follower1
Follower2

0.5
y

-0.5

-1

-1.5

-2
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
x

Fig. 21 The tracked trajectory corresponding to 50 and 30% LOE in the left and right wheels of
leader injected at t = 200 and t = 350 s in the implementation, respectively
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 31

Estimated left and right gain-Follower 1

0.4

0.2
K̂L

-0.2
0 50 100 150 200 250 300 350 400 450 500
time
0.6

0.4
K̂R

0.2

-0.2
0 50 100 150 200 250 300 350 400 450 500
time

Fig. 22 The estimated fault parameters for follower 1 corresponding to 50 and 30% loss of effec-
tiveness faults in the right and left wheels of follower 1 injected at t = 200 and t = 350 s in the
real-time simulation, respectively

Estimated left and right gain-Follower 2

1

0.5
K̂L

0
0 50 100 150 200 250 300 350 400 450 500
time
0.3

0.2

0.1
K̂R

-0.1
0 50 100 150 200 250 300 350 400 450 500
time

Fig. 23 The estimated fault parameters for follower 2 corresponding to 70 and 20% loss of effec-
tiveness faults in the left and right wheels of follower 2 injected at t = 200 and t = 350 s in the
real-time simulation, respectively
32 P. Yazdjerdi and N. Meskin

Estimated left and right gain-Follower 1

0.4

0.2
K̂L

-0.2
0 50 100 150 200 250 300 350 400 450 500
time
0.6

0.4
K̂R

0.2

-0.2
0 50 100 150 200 250 300 350 400 450 500
time

Fig. 24 The estimated fault parameters for follower 1 corresponding to 50 and 30% loss of effec-
tiveness faults in the right and left wheels of follower 1 injected at t = 200 and t = 350 s in the
implementation, respectively

Estimated left and right gain-Follower 2

0.8

0.6
K̂L

0.4

0.2

0
0 50 100 150 200 250 300 350 400 450 500
time
0.3

0.2
K̂R

0.1

0
0 50 100 150 200 250 300 350 400 450 500
time

Fig. 25 The estimated fault parameters for follower 2 corresponding to 70 and 20% loss of effec-
tiveness faults in the left and right wheels of follower 2 injected at t = 200 and t = 350 s in the
implementation, respectively
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 33

Error on ρ for Folower-1

0.015

0.01

0.005
eρ

-0.005

-0.01

-0.015
0 50 100 150 200 250 300 350 400 450 500
Time(sec)

Fig. 26 The error signal eρ (t) for follower 1 corresponding to 50 and 30% loss of effectiveness faults
in the right and left wheels of follower 1 injected at t = 200 and t = 350 s in the implementation,
respectively

Error on ρ for Folower-2

0.03

0.025

0.02

0.015

0.01
eρ

0.005

-0.005

-0.01

-0.015
0 50 100 150 200 250 300 350 400 450 500
Time(sec)

Fig. 27 The error signal eρ (t) for follower 2 corresponding to 70 and 20% loss of effectiveness faults
in the left and right wheels of follower 2 injected at t = 200 and t = 350 s in the implementation,
respectively
34 P. Yazdjerdi and N. Meskin

Error on ψ for Folower-1

0.15

0.1

0.05
eψ

-0.05

-0.1
0 50 100 150 200 250 300 350 400 450 500
Time(sec)

Fig. 28 The error signal eψ (t) for follower 1 corresponding to 50 and 30% loss of effectiveness
faults in the right and left wheels of follower 1 injected at t = 200 and t = 350 s in the implemen-
tation, respectively

Error on ψ for Folower-2

0.05

-0.05
eψ

-0.1

-0.15

-0.2
0 50 100 150 200 250 300 350 400 450 500
Time(sec)

Fig. 29 The error signal eψ (t) for follower 2 corresponding to 70 and 20% loss of effectiveness
faults in the left and right wheels of follower 2 injected at t = 200 and t = 350 s in the implemen-
tation, respectively
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 35

Trajectory Tracked
2
Leader
1.5 Follower1
Follower2
1

0.5

0
y

-0.5

-1

-1.5

-2

-2.5
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
x

Fig. 30 The tracked trajectory corresponding to fault in follower 1 and follower 2 in the imple-
mentation

desired ψ angle is achieved for both followers. Finally, Fig. 30 delineates the tracked
trajectory in the presence of faults in both followers. The maximum detection time
corresponds to 20% LOE fault on the right actuator of follower 2 which is 377.1 s
in the real-time simulation and 382 s in the implementation. The main reason of the
large detection time for 20% LOE fault in the right actuator is that right actuator has
lower speed than left actuator due to the clockwise motion of the robot and the fault
is very minor.

5.5 Fault Tolerant Obstacle Avoidance Controller for Single

Mobile Robot

The obstacle avoidance algorithm is designed to provide the robot more smooth
and an optimized movement in the presence of the obstacle. The control algorithm
switches from a go-to goal controller to an obstacle avoidance controller when an
obstacle is detected using a Kinect sensor mounted on Qbot-2. Furthermore, the
controller is enhanced by a fault tolerant feature using joint state and parameter esti-
mation. Hence, while the robot is moving from an initial point toward the desired
target, it can avoid a color specified object and tolerate the multiplicative loss of
effectiveness actuator fault at any time of the runtime. The effectiveness of obsta-
cle avoidance, fault tolerant controller is verified by the real-time simulation. Also,
36 P. Yazdjerdi and N. Meskin

extensive experiments are done on Qbot-2 to verify the effectiveness of the proposed
algorithm.

5.5.1 Go-to Goal Controller

To attract the robot toward a target point a simple proportional controller is designed
based on the position error [58, 59]. It is assumed that the target is presented as
a circle with radius r T and Cartesian coordinate (x T , yT ) as shown in Fig. 31. The
circular robot is assumed to be located at Cartesian coordinate (x, y) and radius r .
The kinematic equation of the differential drive mobile robot can be rewritten as
follows:
      
ẋ(t) cos φ(t) − sin φ(t) v(t) v(t)
= =M , (26)
ẏ(t) sin φ(t) cos φ(t) ω(t) ω(t)

such that matrix M is invertible. The error position between the robot and target is
calculated as follows:
   
ex (t) x(t) − x T (t)
= , (27)
e y (t) y(t) − yT (t)

and since the position of the target is assumed to be fixed, we have

   
ėx (t) ẋ(t)
= . (28)
ė y (t) ẏ(t)

Accordingly, using a simple proportional gain, the controller is designed as follows:

     
v(t) e (t) p1 0
= −P M −1 x , P= (29)
ω(t) e y (t) 0 p2

where p1 > 0 and p2 > 0. To prove the stability of the designed controller, the
following Lyapunov function is considered:

1 2
V (e(t)) = e (t), (30)
2

where e(t) = ex2 (t) + e2y (t) based on Pythagoras. Thus,

V̇ (e(t)) = e(t)ė(t) = ex ėx + e y ė y = ex ẋ + e y ẏ. (31)

Note that kinematic equation can be rewritten as follows:

Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 37

Fig. 31 Robot and target

configuration in a Cartesian
coordinate

       
ẋ(t) v(t) e (t) e (t)
=M = M × −PM−1 x = −P x . (32)
ẏ(t) ω(t) e y (t) e y (t)

Thus,

V̇ (e(t)) = − p1 ex2 − p2 e2y < 0, (33)

and the asymptotic stability is guaranteed.

5.5.2 Obstacle Avoidance Controller

Orbital obstacle avoidance algorithm is widely used in the literature to provide smooth
and fast navigation toward the desired destination. A real-time limit-cycle obstacle
avoidance algorithm is used in [60] for a soccer robot operating in a dynamically
changing environment. The limit-cycle principle is used in [61] to generate adaptive
elliptic trajectory based on the obstacle position information for flexible and smooth
operation of robot in unstructured environments. Orbital obstacle avoidance is also
studied in [62] using adaptive limit-cycle trajectories.
Given that the obstacle coordinate (xobst , yobst ) is known with respect to the robot
frame, and the robot is moving toward the desired target, the obstacle is detected if the
distance between the center of the robot to the center of the obstacle, denoted as dro ,
is less than r I + 0.25 (See Fig. 32). The value of 0.25 m is chosen by trial and error to
provide enough time for robot to be able to start the trajectory tracking and move on
the circle of influence. The circle of influence is defined with radius r I = r + ro +
Margin where r is the radius of the robot, ro is the radius of the obstacle, and Margin is
a small number to avoid any collision due to the system tolerances and provide safety
and reliability to the controller. As soon as the obstacle is detected, the controller
switches to the obstacle avoidance controller [62]. The obstacle avoidance controller
is designed such that the robot has to follow a limit-cycle vector field which is
38 P. Yazdjerdi and N. Meskin

described by two sets of differential equations for the clockwise and counterclockwise
trajectory motion, respectively, as follows [58, 60, 63]:
• Clockwise motion

Fig. 32 The used perception for the robot navigation

2 2

1.5 1.5

1 1

0.5 0.5
ys
ys

0 0

-0.5 -0.5

-1 -1

-1.5 -1.5

-2 -2
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2
xs xs

Clockwise. Counter-clockwise.
Fig. 33 Trajectory motions used for limit-cycles

Fig. 34 The frame

transformation
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 39


ẋs (t) = ys (t) + xs (t) rs2 − xs2 (t) − ys2 (t) ,

(34)
ẏs (t) = −xs (t) + ys (t) rs2 − xs2 (t) − ys2 (t) ,

• Counterclockwise motion


ẋs (t) = −ys (t) + xs (t) rs2 − xs2 (t) − ys2 (t) ,

(35)
ẏs (t) = xs (t) + ys (t) rs2 − xs2 (t) − ys2 (t) ,

where (xs (t), ys (t)) are the position of the robot with respect to the center of the
convergence circle characterized by radius rs of the limit-cycle. Figure 33 depicts
the trajectories formed by following equations with the radius rs = 1. It should be
noted that all points inside and outside of the circle, move toward the circumstance
of the circle.

Once the obstacle is detected, the primary step to define the limit-cycle trajectory
is to choose rs and the direction of motion. Toward this goal, it is required to define
a new frame in which the X o axis connects the center of the obstacle to the center
of the target and the Yo axis is perpendicular to X o as shown in Fig. 34. Moreover,
the coordinate position of the robot (x, y) is mapped into the new reference frame
(X o , Yo ) using the following transformation:
⎡ ⎤ ⎡ ⎤−1 ⎡ ⎤
xo cos α − sin α 0 xobst x
⎢ yo ⎥ ⎢ sin α cos α 0 yobst ⎥ ⎢y⎥
⎢ ⎥=⎢ ⎥ ⎢ ⎥, (36)
⎣0⎦ ⎣ 0 0 1 0 ⎦ ⎣0⎦
1 0 0 0 1 1

where α is the angle between X o axis and X axis as shown in Fig. 34. The radius rs =
r I − ζ in the beginning when x-coordinate of the obstacle xo ≤ 0 and rs = r I + ζ
otherwise where ζ  margin to guarantee that the robot do not move very close the
circle of influence. Note that if the x-coordinate of the robot is greater than zero in
the new reference frame (X o , Yo ), it means that the robot has passed more than half
of the obstacle and it can now switch from the obstacle avoidance to the go-to goal
controller.
On the other hand, the y-coordinate of the robot yo in the reference frame of the
obstacle provides the shortest trajectory in the length to reach the target. Accordingly,
if yo ≥ 0, then the clockwise motion is chosen, otherwise the counterclockwise tra-
jectory is tracked by the robot. Toward this goal, the limit-cycle set points provided
in Eqs. (34) and (35) are redefined as follows:
40 P. Yazdjerdi and N. Meskin


ẋs (t) = sign(yo (t))ys (t) + xs (t) rs2 − xs2 (t) − ys2 (t)

(37)
ẏs (t) = −sign(yo (t))xs (t) + ys (t) rs2 − xs2 (t) − ys2 (t)

To control the robot to move on the desired trajectory while obstacle avoidance mode
is activated, the trajectory tracking controller (Sect. 5.3) is used to move the robot on
the desired trajectory.
As mentioned earlier, if the distance between the robot and obstacle is less than
r I + 0.25, the obstacle avoidance controller is activated. Thus, it is desired to activate
the Kinect sensor up to the point that this condition is met. However, due to the usage
of image processing blocks to capture the depth and world images, there exists
an unavoidable delay which results in delayed sampled data from dead reckoning
and wrong estimation in EKF. Accordingly, the image processing is done prior to
implementing the controllers. In the other words, first the Cartesian position and the
radius of the obstacle are measured and then the calculated values are used by the
controller. Also, due to this limitation, when the obstacle is located furtherer than
3 m from the robot, the measured radius is not calculated correctly. It should be
emphasized that this is mainly due to the computational capability of Qbot-2.

5.5.3 Real-Time Simulation and Implementation Results

This section presents the real-time simulation and implementation results of orbital
obstacle avoidance and fault tolerant controller for a single mobile robot. In the
following experiments, the robot is desired to reach a target located at (5, 0.1) m,
Margin = 0.125 m, and ζ = 0.005 m. The obstacle position and size information
are given to the system as a constant number as it is not possible to use Kinect sensor
in the simulation. Figure 35 presents the clockwise movement of the simulated robot
in the presence of obstacle only. The initial Cartesian coordinates of the obstacle
are set as (2.06, 0.0254) m and the radius of the obstacle is measured as 0.1538 m.
Similarly, Fig. 36 presents the clockwise path tracked by Qbot-2 in the presence of an
obstacle. The radius of the obstacle is measured by Kinect sensor as 0.1538 m. As it
is shown in Figs. 35 and 36, the robot is located at a point which is above the center of
the obstacle. Accordingly, based on Fig. 34, yo has a positive value which results in a
clockwise movement of the robot. From Fig. 36, it is obvious that the robot reaches
the point (4.98, 0.15) m. Moreover, Figs. 37 and 38 show the estimated LOE gain for
each actuator in the real-time simulation and implementation, respectively. A small
transient is observed when the controller switches to the obstacle avoidance mode,
however it should be noted that it does not affect the path tracked by the robot. The
switching of the controller is done based on a generated flag from the algorithms
described earlier in this chapter to detect the obstacle. If the flag is one, the controller
is in the obstacle avoidance mode and if the flag = 0, the go-to goal controller is
activated. Figures 39 and 40 delineate the generated flag for the real-time simulation
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 41

0.6
Tracked path
Obstacle
0.4 Circle of influence

0.2
y o = 0.024

0
y

-0.2

-0.4

-0.6
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
x

Fig. 35 Obstacle avoidance with clockwise motion in the real-time simulation

and implementation. The obstacle avoidance mode was activated from t = 12.99 to
t = 27.92 s in Fig. 39. Similarly, Fig. 40 shows that the value of the flag start to be 1 at
t = 13.92 and it became 0 at t = 28.71 s which implies that obstacle avoidance mode
was activated for 14.93 and 14.79 s in the real-time simulation and implementation,
respectively.
On the other hand, Fig. 41 depicts the counterclockwise motion of the simulated
robot in the presence of an obstacle at point (2.15, 0.16) m with the radius of 0.158
m. It should be noted that the robot is located at negative yo when it switches to the
obstacle avoidance controller. Similarly, Fig. 42 shows the counterclockwise motion
of Qbot-2 due to the negative value of yo when the robot is located in a distance of one
meter from the obstacle which is the condition of activating the obstacle avoidance
controller. In this test, the radius of the robot is measured by Kinect sensor as 0.158
m. This figure shows that the robot reaches the point (4.98, 0.03) m. The amount
of error in the final position of the obstacle is negligible due to the error of dead
reckoning and by considering the size of the robot which has a radius of 0.175 m. By
comparing the real-time simulation and implementation results, it can be concluded
that in the simulation, the desired destination is accurately reached by the simulated
robot, however, in the implementation, small error exists due to dead reckoning
and by considering the size of the robot. Furthermore, the estimated LOE gain for
each actuator in the healthy mode is shown in Figs. 43 and 44 for the real-time
simulation and implementation, respectively. Switching of the controller from go-to
goal to obstacle avoidance mode and vice versa causes a negligible transient in the
detection of the gains which does not affect the tracked trajectory. By comparing the
estimated gains in the healthy mode for clockwise and counterclockwise motion, it
42 P. Yazdjerdi and N. Meskin

0.6
Tracked path
Obstacle
0.4 Circle of influence

0.2 y o = 0.05

0
y

-0.2

-0.4

-0.6
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
x

Fig. 36 Obstacle avoidance with clockwise motion in the implementation

Left wheel gain

0.9
Right wheel gain
0.8

0.7

0.6

0.5
Gain

0.4

0.3

0.2

0.1

-0.1
0 50 100 150
Time(sec)

Fig. 37 Estimated fault values corresponding to the healthy mode in the real-time simulation for
the clockwise motion
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 43

Left wheel gain

0.9
Right wheel gain
0.8

0.7

0.6

0.5
Gain

0.4

0.3

0.2

0.1

-0.1
0 50 100 150
Time(sec)

Fig. 38 Estimated fault values corresponding to the healthy mode in the implementation for the
clockwise motion

0.9

0.8

0.7

0.6
Flag

0.5

0.4

0.3

0.2

0.1

0
0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150
Time(sec)

Fig. 39 Obstacle avoidance and go-to goal flag corresponding to clockwise motion in the real-time
simulation
44 P. Yazdjerdi and N. Meskin

0.9

0.8

0.7

0.6
Flag

0.5

0.4

0.3

0.2

0.1

0
0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150
Time(sec)

Fig. 40 Obstacle avoidance and go-to goal flag corresponding to clockwise motion in the imple-
mentation

0.8
Tracked path
Obstacle
0.6 Circle of influence

0.4

y o = -0.13
y

0.2

-0.2

-0.4
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
x

Fig. 41 Obstacle avoidance with counterclockwise motion in the real-time simulation

can be concluded that due to the environment noises, the transient is larger in the
implementation and takes longer time to reach the desired value.
In addition, three scenarios are considered for the case of fault tolerant and obstacle
avoidance controller for a robot to reach the desired target in the presence of an
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 45

0.8
Tracked path
Obstacle
0.6 Circle of influence

0.4

y o = -0.12
0.2
y

-0.2

-0.4
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
x

Fig. 42 Obstacle avoidance with counterclockwise motion in the implementation

Left wheel gain

0.9
Right wheel gain
0.8

0.7

0.6

0.5
Gain

0.4

0.3

0.2

0.1

-0.1
0 50 100 150
Time(sec)

Fig. 43 Estimated fault values corresponding to the healthy mode in the real-time simulation for
the counterclockwise motion

obstacle and multiplicative loss of effectiveness actuator fault. A large transient is

observed in the value of the estimated gains by EKF when the controller switches
from the go-to goal to the obstacle avoidance or vice versa. To avoid this transient,
the EKF is designed to start its estimation after a few seconds.
46 P. Yazdjerdi and N. Meskin

0.9 Left wheel gain

Right wheel gain
0.8

0.7

0.6

0.5
Gain

0.4

0.3

0.2

0.1

-0.1
0 50 100 150
Time(sec)

Fig. 44 Estimated fault values corresponding to the healthy mode in the implementation for the
counterclockwise motion

Scenario 1: 40% LOE fault in the left actuator at t = 30, obstacle position:
(1.958, 0.0159)
In the first scenario, the obstacle is located at (1.958, 0.0159) m from the robot, the
radius of the obstacle is measured by Kinect sensor as 0.1516 m, and the target is set
to be at (9, 0.1) m from the robot. A 40% loss of effectiveness fault is occurred in
the left actuator at t = 30 s, i.e., K L = 0.4. Figure 45 shows that estimated gain for
the left actuator K̂ L is correctly estimated at 36.08 s in the real-time simulation. The
estimated left and right gain by the EKF is presented in Fig. 46 for implementation.
The estimated gain for the left actuator at t = 39.02 s is K̂ L = 0.443 which is within
10% of the real gain. The estimated gain for the right actuator is remained unchanged
in both real-time simulation and implementation. Figure 47 shows the path tracked
by the robot in the presence of an obstacle and left actuator LOE fault. It is obvious
that the robot roughly reaches the desired destination at (8.97, 0.13) m.
Scenario 2: 50% LOE fault in the right actuator at t = 5, obstacle position:
(4.93, 0.03)
In the second scenario, 50% LOE fault is occurred in the right actuator at t = 5 s
(i.e., K R = 0.5) before the robot reaches the obstacle which is located at (4.93, 0.03)
m from the robot. Figure 48 depicts that right actuator gain K̂ R is estimated correctly
at t = 14 s in the real-time simulation. Figure 49 shows that value of K̂ R = 0.55
which is within 10% of the real value at t = 16.26 s in the implementation. The left
actuator gain remained unchanged in the real-time simulation and implementation.
Note that since the fault is injected at the beginning of the runtime and the speed
of the robot is slow in the beginning, the detection time is larger than the previous
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 47

1
Left wheel gain
0.9 Right wheel gain

0.8

0.7

0.6
Gain

0.5

0.4

0.3

0.2

0.1

0
0 20 40 60 80 100 120 140 160 180 200
Time(sec)

Fig. 45 Estimated fault values corresponding to 40% LOE fault in the left actuator applied at
t = 30 s in the real-time simulation

1
Left wheel gain
0.9 Right wheel gain

0.8

0.7

0.6
Gain

0.5

0.4

0.3

0.2

0.1

0
0 20 40 60 80 100 120 140 160 180 200
Time(sec)

Fig. 46 Estimated fault values corresponding to 40% LOE in the left actuator applied at t = 30 s
in the implementation
48 P. Yazdjerdi and N. Meskin

1
Tracked path
Obstacle
Circle of influence

0.5
y

-0.5
0 1 2 3 4 5 6 7 8 9
x

Fig. 47 Tracked path in the presence of 40% LOE fault in the left actuator after obstacle avoidance
in the implementation

scenario. Moreover, it is obvious that the detection and isolation time is more in the
implementation due to unavoidable error caused by the surrounding environment.
Also, it is shown that estimated gains in the simulation are estimated smoothly than
that in the implementation (see Figs. 49 and 48). In the other words, the deviation
of the estimated gains in the implementation is much more than the deviation in the
simulation. Figure 50 demonstrates the path tracked by the robot and verifies that it
reaches (8.97, 0.06) m while the desired destination which is set to be at (9, 0.1) m
from the robot. As mentioned earlier, this error is negligible.
Scenario 3: 20% LOE fault in the right actuator at t = 17, obstacle position:
(2, −0.15)
Finally, in the third scenario, 20% LOE in the right actuator is occurred at t = 17 s
and an obstacle is located at (2, −0.15) m from the robot. The radius is measured by
Kinect sensor as 0.157 m. In this scenario, the fault is occurred while the robot is in
the obstacle avoidance mode. Figure 51 shows that the robot avoids the obstacle in
the presence of fault and roughly reaches the desired destination at point (8.98, 0.06).
As shown in Fig. 51, in the presence of the fault, the robot did not enter the circle
of influence, however, this is not considered in the FTC algorithm. Figure 52 shows
that the estimated value of K̂ R is 0.22 at t = 40.85 s. This delay in the detection
and compensation of the fault is due to the movement of the robot in the obstacle
avoidance mode and switching from one controller to the other. The estimated value
of K̂ L remains unchanged.
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 49

1
Left wheel gain
0.9 Right wheel gain

0.8

0.7

0.6
Gain

0.5

0.4

0.3

0.2

0.1

0
0 20 40 60 80 100 120 140 160 180 200
Time(sec)

Fig. 48 Estimated fault values corresponding to 50% LOE in the right actuator applied at t = 5 s
in the real-time simulation

1
Left wheel gain
0.9 Right wheel gain

0.8

0.7

0.6
Gain

0.5

0.4

0.3

0.2

0.1

0
0 20 40 60 80 100 120 140 160 180 200
Time(sec)

Fig. 49 Estimated fault values corresponding to 50% LOE in the right actuator applied at t = 5 s
in the implementation
50 P. Yazdjerdi and N. Meskin

0.5
Tracked path
0.4 Obstacle
Circle of influence
0.3

0.2

0.1
y

-0.1

-0.2

-0.3

-0.4

-0.5
0 1 2 3 4 5 6 7 8 9
x

Fig. 50 Tracked path in the presence of 50% LOE fault in the right actuator before the obstacle
avoidance activation in the implementation

0.8
Tracked path
Obstacle
0.6
Circle of influence

0.4

0.2
y

-0.2

-0.4

-0.6
0 1 2 3 4 5 6 7 8 9
x

Fig. 51 Tracked path in the presence of 20% LOE fault in the right actuator while obstacle avoidance
in the implementation
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 51

1
Left wheel gain
0.9 Right wheel gain

0.8

0.7

0.6
Gain

0.5

0.4

0.3

0.2

0.1

0
0 20 40 60 80 100 120 140 160 180 200
Time(sec)

Fig. 52 Estimated fault values corresponding to 20% LOE in the right wheel injected at t = 17 s
in the implementation

6 Conclusion and Future Works

In this chapter, a fault tolerant controller strategy is described for differential drive
mobile robots in the presence of LOE fault in actuators. A joint state and parameter
estimation strategy is used based on EKF to approximate parameters (actuator loss of
effectiveness) and the system states. The consequence of the estimated fault is com-
pensated in the controller for trajectory tracking goal of single robot and formation
control of multiple robots.
Moreover, a fault tolerant controller scheme with obstacle avoidance feature is
developed in this chapter for a single mobile robot in the presence of the LOE actuator
fault. A joint state and parameter estimation method is used to estimate the LOE gain
while robot moving toward the desired destination using the go-to goal controller. In
the presence of the obstacle, the controller switches to obstacle avoidance controller
to avoid collision considering the shortest path toward the desired destination and
tolerate the fault. The efficacy of the proposed algorithms is verified by real-time
simulation and implementation on Qbot-2. The comparison between the simulation
and implementation is done and it is shown that fault detection and isolation is faster
in the case of simulation.
The leading superiorities of the proposed FTC method compared to the literature
can be outlined as follows: (1) this method has ability of detection, isolation, and
identification of actuator faults while in [19, 21, 26, 30] and [31], only the detection
and/or isolation is possible; (2) in most of the former works except [45], all the
states of the robot are acquired to be able to detect, isolate, and identify faults;
52 P. Yazdjerdi and N. Meskin

however, the proposed method requires Cartesian coordinate of the mobile robot; (3)
the described FTC scheme is facile and simple to implement and does not require
reconfiguration/modification of the nominal controller while in [45], it is needed; (4)
fault detection, identification, and tolerance for loss of effectiveness faults in both
wheels is accomplished by only one extended Kalman filter while in [19], more than
filters are needed; (5) it is implementable by experiment to be verified while the
verification in the proposed FTC approaches in [40, 42, 45, 48, 53] are only done
by simulation; (6) in the case of FTC method described in this paper for multiple
mobile robots, each follower requires only to have the leader position rather than
the position information of other followers to be able to accomplish the desired task
while in [49] and [50], each robot requires to know the position of all other robots.
The main contribution of the proposed obstacle avoidance feature is the usage of
Kinect sensor to detect and avoid obstacles in the presence of LOE actuator fault
using depth image rather than path planning. Path planning and mapping of the
environment is well studied in the literature to detect edges and obstacles using
Kinect sensor [64]. However, this method brings complexity and time delay in the
implementation. Moreover, the fault detection approach in the presence of obstacle
is presented in [29]; however, this work provides fault compensation in the presence
of the obstacle.
The future directions for this chapter are outlined below
• Development of an actuator FTC algorithm for a team of mobile robots moving
from an initial point toward the desired destination in the presence of one or more
obstacles.
In many applications, a team of mobile robots is required to do a mission, thus
it is important to develop a FTC algorithm with the ability to avoid obstacles.
The main challenge in designing the multiple mobile robot system to cooperate
with each other and successfully accomplish a mission is developing a reliable
leader–follower formation control. Also, it is challenging to wisely develop an
obstacle avoidance algorithm for each one of them in the team to avoid any collision
with each other and with the obstacle and choose the shortest path to avoid the
obstacle. Toward this goal, robots should be allowed to move regardless of keeping
the formation if needed in the presence of the obstacle. However, their position
information should be transmitted to each other.
• Development of obstacle avoidance and fault tolerant controller such that it keeps
the robot out of the circle of influence to guarantee collision-free movement.
As shown in Sect. 5.5.3, the robot did not enter the circle of influence; however,
there is no guarantee in the controller to avoid entering the circle of influence in
the presence of fault. Accordingly, it is important to consider this issue by defining
a threshold in the controller.
• Development of obstacle avoidance and fault tolerant controller for sensor and
process faults in the single and multiple mobile robots.
Sensor and process faults are other types of fault that may occur in the system
rather than actuator faults. Thus, to have a safe and reliable operation, the controller
should be able to compensate for these types of faults using EKFs.
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 53

• Development of fault tolerant and obstacle avoidance controller which can avoid
actuator saturation in the presence of sensor, process, and actuator fault.
As mentioned earlier, due to the limitation of the robot, the angular and linear
velocities of the robot saturate when it is desired to move on some trajectories
(i.e., circles with a small radius). Thus, it is of a great importance to avoid the
saturation by considering the robot’s velocity threshold in the controller.

References

1. S. Kim, C.H. Kim, Y.C. Seo, S.H. Jung, G.S. Lee, B.S. Han, Development of tele-operated
mobile robot in nuclear power plants, in IFAC Proceedings Volumes, vol. 34, no. 4, (2001),
pp. 239–244; iFAC Workshop on Mobile Robot Technology, (Jejudo Island, Korea, 20–23 May
2001)
2. T.N. Khoon, P. Sebastian, A.B.S. Saman, Autonomous fire fighting mobile platform, in Inter-
national Symposium on Robotics and Intelligent Sensors 2012 (IRIS 2012), vol. 41(Supplement
C), (2012), pp. 1145–1153
3. B. Brumson, Robotics and energy cost reduction (2006). https://www.robotics.org/content-
detail.cfm/
4. Pintrest (2008) Senior care robot. https://www.pinterest.com/pin/61924563597475437/
5. J. Schweder, Mobile robots deliver material handling flexibility (2016). https://
www.automationworld.com/article/technologies/robotics/mobile-robots-deliver-material-
handling-flexibility
6. Yazdjerdi P, Meskin N (2018) Design and real-time implementation of actuator fault-tolerant
control for differential-drive mobile robots based on multiplemodel approach. J Systems &
Control Engineering 232(6):652–661
7. Kobishikar industrial designer, Transwheel delivery drone concept (2015). http://kobishikar.
wixsite.com/jaykob-shikar/transwheel
8. ETHZurich (2009–2010) Development of ballbot-rezero. http://www.rezero.ethz.ch/
9. Quanser (2017) Qbot-2 for quarc. https://www.quanser.com/products/qbot-2-quarc/
10. Mindstorms (2001) Trik3r. https://sites.google.com/site/gask3t/lego-ev3/my-models/trik3r
11. Sparkfun, Multi-chassis- 4wd kit (basic) (2008). https://www.sparkfun.com/products/12089
12. H. Durmus, E.O. Gunes, M. Kirci, B. Berk, The design of general purpose autonomous
agricultural mobile-robot, in Fourth International Conference on Agro-Geoinformatics (Agro-
geoinformatics) (2015), pp. 49–53
13. R. Shah, S. Ozcelik, R. Challoo, Design of a highly maneuverable mobile robot. Procedia
Computer Science 12(Supplement C), 170–175 (2012) (Complex Adaptive Systems 2012)
14. M. Dorigo, Swarm bots (2005). http://www.swarm-bots.org/
15. U.K. Farming, Advanced agricultrual mobile robot to help performing experiments at university
(2016). https://www.farminguk.com/News/Advanced
16. A.H. Reddy, B. Kalyan, C.S. Murthy, Mine rescue robot system—a review. Procedia Earth
Planet. Sci. 11(Supplement C), 457–462 (2015)
17. S. Ding, Model-Based Fault Diagnosis Techniques: Design Schemes (Springer, Algorithms
and Tools, 2013)
18. N. Meskin, K. Khorasani, Fault Detection and Isolation: Multi-Vehicle Unmanned Systems
(Springer, 2011)
19. S.I. Roumeliotis, G.S. Sukhatme, G.A. Bekey, Fault detection and identification in a mobile
robot using multiple-model estimation, in Proceedings 1998 IEEE International Conference
on Robotics and Automation, 1998, vol. 3 (1998), pp. 2223–2228
54 P. Yazdjerdi and N. Meskin

20. W.E. Dixon, I.D. Walker, D.M. Dawson, Fault detection for wheeled mobile robots with para-
metric uncertainty, in Proceedings 2001 IEEE/ASME International Conference on Advanced
Intelligent Mechatronics, 2001, vol. 2, (2001) pp. 1245–1250
21. E.N. Skoundrianos, S.G. Tzafestas, Finding fault- fault diagnosis on the wheels of a mobile
robot using local model neural networks. IEEE Robot. Autom. Mag. 11(3), 83–90 (2004)
22. V. Verma, Tractable particle filters for robot fault diagnosis. Ph.D. thesis, Stanford University,
2004
23. P. Sundvall, P. Jensfelt, Fault detection for mobile robots using redundant positioning systems,
in Proceedings 2006 IEEE International Conference on Robotics and Automation, 2006. ICRA
2006 (2006), pp. 3781–3786
24. A. Monteriu, P. Asthana, K. Valavanis, S. Longhi, Experimental validation of a real-time model-
based sensor fault detection and isolation system for unmanned ground vehicles, in 2006 14th
Mediterranean Conference on Control and Automation (2006), pp. 1–8
25. N. Sidek, N. Sarkar, Integrating actuator fault and wheel slippage detections within FDI frame-
work. WSEAS Trans. Syst. 6(2), 298 (2007)
26. B. Halder, N. Sarkar, Robust nonlinear analytic redundancy for fault detection and isolation in
mobile robot. Int. J. Autom. Comput. 4(2), 177–182 (2007)
27. F. Baghernezhad, K. Khorasani, A robust fault detection scheme with an application to mobile
robots by using adaptive thresholds generated with locally linear models, in 2013 IEEE Sym-
posium on Computational Intelligence in Control and Automation (CICA) (2013), pp. 9–16
28. N.B. Hoang, H.J. Kang, A model-based fault diagnosis scheme for wheeled mobile robots. Int.
J. Control, Autom. Syst. 12(3), 637–651 (2014)
29. A.G. Millard, J. Timmis, A.F.T. Winfield, Run-time detection of faults in autonomous mobile
robots based on the comparison of simulated and real robot behaviour, in 2014 IEEE/RSJ
International Conference on Intelligent Robots and Systems (2014), pp. 3720–3725
30. P.S. Pratama, A.V. Gulakari, Y.D. Setiawan, D.H. Kim, H.K. Kim, S.B. Kim, Trajectory track-
ing and fault detection algorithm for automatic guided vehicle based on multiple positioning
modules. Int. J. Control, Autom. Syst. 14(2), 400–410 (2016)
31. D. Stavrou, D.G. Eliades, C.G. Panayiotou, M.M. Polycarpou, Fault detection for service
mobile robots using model-based method. Auton. Robot. 40(2), 383–394 (2016)
32. P. Goel, G. Dedeoglu, S.I. Roumeliotis, G.S. Sukhatme, Fault detection and identification in a
mobile robot using multiple model estimation and neural network, in Proceedings 2000 ICRA
IEEE International Conference on Robotics and Automation, vol. 3, (2000), pp 2302–2309
33. M. Hashimoto, H. Kawashima, F. Oba, A multi-model based fault detection and diagnosis of
internal sensors for mobile robot, in Proceedings 2003 IEEE/RSJ International Conference on
Intelligent Robots and Systems, 2003. (IROS 2003), vol. 3 and 4, (2003) pp. 3787–3792
34. P. Tsarouhas, G. Fourlas, Reliability analysis of the mobile robot: a case study. Recent. Adv.
Mech. Eng. Mech., 151 (2014)
35. A.L. Christensen, R. O’Grady, M. Birattari, M. Dorigo, Automatic synthesis of fault detection
modules for mobile robots, in Proceedings of the Second NASA/ESA Conference on Adaptive
Hardware and Systems (IEEE Computer Society, Washington, DC, USA, AHS ’07, 2007), pp.
693–700
36. A.L. Christensen, R. O’Grady, M. Birattari, M. Dorigo, Fault detection in autonomous robots
based on fault injection and learning. Auton. Robot. 24(1), 49–67 (2008)
37. S. Oonk, F.J. Maldonado, Z. Li, K. Reichard, J. Pentzer, Extended Kalman filter for improved
navigation with fault awareness, in 2014 IEEE International Conference on Systems, Man, and
Cybernetics (SMC) (2014), pp. 2681–2686
38. M. Hashimoto, T. Ishii, K. Takahashi, Sensor fault detection and isolation for mobile robots in
a multi-robot team, in 2009 35th Annual Conference of IEEE on Industrial Electronics, IECON
’09 (2009), pp. 2348–2353
39. F. Arrichiello, A. Marino, F. Pierri, Observer-based decentralized fault detection and isolation
strategy for networked multirobot systems. IEEE Trans. Control Syst. Technol. 23(4), 1465–
1476 (2015)
Fault Tolerant Controller Schemes for Single and Multiple Mobile Robots 55

40. M. Ji, Z. Zhang, G. Biswas, N. Sarkar, Hybrid fault adaptive control of a wheeled mobile robot.
IEEE/ASME Trans. Mechatron. 8(2), 226–233 (2003)
41. Z. Li, Application of fault tolerant controller based on RBF neural networks for mobile robot,
in 2009 International Symposium on Intelligent Ubiquitous Computing and Education (2009),
pp. 531–534
42. A. Srebro, A self-tuning fuzzy PD controller for a wheeled mobile robot operating in the
presence of faults. Chall. Mod. Technol. 2(4), 11–20 (2011)
43. Z. Li, W. Jiang, Active fault-tolerant control for two-wheeled differential drive mobile robot
based on fault compensation method. Chin. Autom. Congr. (CAC) 2013, 359–363 (2013)
44. G. Fourlas, G.C. Karras, K.J. Kyriakopoulos, Fault tolerant control for a 4-wheel skid steering
mobile robot, in Safeprocess (2015), pp. 177–184
45. Y.H. Chang, C.I. Wu, C.Y. Yang, Adaptive output-feedback fault-tolerant tracking control for
mobile robots under partial loss of actuator effectiveness, in 2015 IEEE 54th Annual Conference
on Decision and Control (CDC), IEEE (2015), pp. 6306–6311
46. H. Olfa, B. Boumedyen, Z. Ahmed, A.M. Naceur, On fault tolerant control of mobile robot
based on fast adaptive fault estimation, in 2015 12th International Multi-Conference on Systems,
Signals & Devices (SSD), IEEE (2015), pp. 1–6
47. Q. Song, Z. Jiang, J.D. Han, Active-model-based fault tolerant control against actuator failures
for mobile robot, in 2006 6th World Congress on Intelligent Control and Automation, vol 1
(2006), pp. 1415–1420
48. A. Stancu, E. Codres, V. Puig, A fault hiding approach for the sliding mode fault-tolerant
control of a non-holonomic mobile robot, in 2016 3rd Conference on Control and Fault-
Tolerant Systems (SysTol) (2016), pp. 7–14
49. L. Cheng, Y.J. Wang, Fault tolerance for communication-based multirobot formation, in Pro-
ceedings of 2004 International Conference on Machine Learning and Cybernetics, vol. 1,
(2004), pp. 127–132
50. Y.H. Chang, W.S. Chan, C.Y. Yang, C.W. Tao, S.F. Su, Adaptive dynamic surface control for
fault-tolerant multi-robot systems, in 2013 International Conference on System Science and
Engineering (ICSSE) (2013), pp. 25–30
51. M.A. Kamel, X. Yu, Y. Zhang, Design of fault-tolerant cooperative control algorithm applied
to WMRs against actuator faults, in 2016 American Control Conference (ACC) (2016), pp.
7092–7097
52. P. Yazdjerdi, N. Meskin, Actuator fault tolerant control in a team of mobile robots, in 2018 15th
International Conference on Control, Automation, Robotics and Vision (ICARCV) (2018), pp.
1885–1890, https://doi.org/10.1109/ICARCV.2018.8581118
53. Z. Sun, Z. Yang, Application of joint parameter identification and state estimation to a fault-
tolerant robot system, in IECON 2011- 37th Annual Conference of the IEEE Industrial Elec-
tronics Society (2011), pp. 498–503
54. I. Hassanzadeh, M.A. Fallah, Design of augmented extended Kalman filter for real time simu-
lation of mobile robots using simulink, in 2009 6th International Symposium on Mechatronics
and its Applications (2009), pp. 1–6
55. S. Aslan, A.T. Cemgil, M. Şamil Aslan, B.U. Töreyin, A. Akın, Joint parameter and state
estimation of the hemodynamic model by iterative extended kalman smoother. Biomed. Signal
Process. Control 24, 47–62 (2016)
56. A. De Luca, G. Oriolo, M. Vendittelli, Control of wheeled mobile robots: An experimental
overview, in Ramsete (Springer, 2001), pp. 181–226
57. H.A. Poonawala, A.C. Satici, M.W. Spong, Leader-follower formation control of nonholonomic
wheeled mobile robots using only position measurements, in Control Conference (ASCC), 2013
9th Asian, IEEE (2013), pp. 1–6
58. L. Adouane, Hybrid and safe control architecture for mobile robot navigation, in 9th Conference
on Autonomous Robot Systems and Competitions (2009a)
59. P. Corke, Robotics, Vision and Control: Fundamental Algorithms in MATLAB® Second, Com-
pletely Revised, vol. 118 (Springer, 2017)
56 P. Yazdjerdi and N. Meskin

60. D.H. Kim, J.H. Kim, A real-time limit-cycle navigation method for fast mobile robots and its
application to robot soccer. Robot. Auton. Syst. 42(1), 17–30 (2003)
61. J. Vilca, L. Adouane, Y. Mezouar, Reactive navigation of a mobile robot using elliptic trajec-
tories and effective online obstacle detection. Gyroscopy Navig. 4(1), 14–25 (2013)
62. L. Adouane, Orbital obstacle avoidance algorithm for reliable and on-line mobile robot navi-
gation, in 9th Conference on Autonomous Robot Systems and Competitions (2009b)
63. M.S. Jie, J.H. Baek, Y.S. Hong, K.W. Lee, Real time obstacle avoidance for mobile robot
using limit-cycle and vector field method, in Knowledge-Based Intelligent Information and
Engineering Systems, ed. by B. Gabrys, R.J. Howlett, L.C. Jain (Springer, Berlin Heidelberg,
Berlin, Heidelberg, 2006), pp. 866–873
64. L. Jaiem, L. Lapierre, K. Godary-Dejean, D. Crestani, Fault tolerant autonomous robots using
mission performance guided resources allocation, in SysTol: Control and Fault-Tolerant Sys-
tems (2016)
Faults and Fault Detection Methods
in Electric Drives

Patrick D. Strankowski and Jarosław Guziński

Abstract The chapter presents a review of faults and fault detection methods in elec-
tric drives. Typical faults are presented that arises for the induction motor, which is
valued in the industry for its robust construction and cost-effective production. More-
over, a summary is presented of detectable faults in conjunction with the required
physical information that allows detection of specific faults. In order to address faults
of a complete drive system, characteristic failures of its mechanical parts are pre-
sented as well. Furthermore, the physical forces, which arise during specific faults
(i.e. centrifugal, kinematic) are presented along with dominant harmonics in the
frequency spectrum. These dominant harmonics are especially important for the
determination of a malfunction of the drive. The detection of the particular could be
performed with signal processing methods which are tabularly summarized for intro-
duction purposes. In order to cover a further industry interest, a cost-effectiveness
relation is presented, which describes whether a diagnostic system is appreciative
or not. Moreover, most important international standards regarding safety, health
for human and machinery are summarized that are required to be fulfilled in every
industrial application. The next subsection is dedicated to the presentation of fault
detection methods that include a review of conventional methods for monitoring the
conditions of the electric machine. That includes the monitoring of variables that
are based on electrical, chemical, mechanical and thermal changes in the induction
motor. The last subsection considers the fault detection methods, which are based on
utilization of mathematical models. In this kind of fault detection, the model descrip-
tion is utilized to identify changes in the drive system, which then can be used for
a potential fault identification procedure. Such mathematical constructs are mainly
based on observer, Kalman filters or neural networks. The chapter is concluded with
a short summary of the presented sections.

P. D. Strankowski (B) · J. Guziński

Faculty of Electrical and Control Engineering, Gdańsk University of Technology,
Narutowicza 11/12, 82-900 Gdańsk, Poland
e-mail: [email protected]; [email protected]
J. Guziński
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 57

H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_2
58 P. D. Strankowski and J. Guziński

Keywords Induction motor faults · Mechanical faults · Fault detection ·

Model-based fault detection · International standards

Nomenclature

MRAS Model Reference Adaptive System

FEM Finite Element Method
XRF X- Ray Fluorescence
MCSA Motor Current Signature Analysis
PD Power Density

1 Introduction

Adjustable speed AC drives are gaining popularity in recent decades as the pre-
ferred choice of electric energy conversion for industrial and transport applications.
Significant advances in the power electronics field allowed the implementation of
sophisticated control methods for electrical drives. The application of voltage source
inverters in conjunction with various control concepts, e.g. field-oriented control
(FOC) [1, 2], direct torque control (DTC) [3] and multiscalar control [4] provides a
great energy saving potential, increases the maximum motor efficiency usage, as well
as improves the drive system durability. Nevertheless, the reliability and efficiency of
a drive system depend not only on the control structure and inverter technology but
also on the condition of the machine and the mechanical system. For this reason, the
industry is interested in long-lasting and fault-tolerant drive systems, especially those
with high responsibility, and in single-unit production drives for special applications.
The following subsections cover the introduction and literature review of the areas
related to the fault detection of electric drives, especially the induction motor. First,
the introduction is given to induction motor faults and a description of signals that
serves as an indicator. Next, common mechanical faults are presented along with
the summary of physical forces that arises during specific types of forces. Since the
section also focuses on the industry application and fault detection possibility in
sensorless drives, the reader will be introduced to important international standards
regarding permitted vibrations, economic aspects of a fault detection application,
and fault detection methods.
Faults and Fault Detection Methods in Electric Drives 59

2 Induction Motor Faults

Faults in induction motor drives are mainly classified into four groups: bearing fail-
ures (≈41%), stator (≈37%), rotor (≈10%) and other faults (≈12%) [5–9]. Inaccu-
rate installation of the drive system, material fatigue, tolerances and an imprecise
manufacturing process may result in faults and, in the worst case, total failure of the
system. These faults are responsible for current and voltage deformations, increased
torque pulsations and/or increased losses, all this leading to the efficiency decrease
and excessive heating of the motor. Figure 1 presents the main classification of faults
in induction motors [7, 10, 11].
Bearing failures are mainly caused by wear out of bearings and have received
increasing attention in the last three decades. It was shown that the usage of voltage
inverters is a source of bearing currents [12]. Beyond that, bearings are subjected to
environmental factors, such as moisture or dust (contaminations) which accelerate
the degradation of the grease, as well as to the negligence of grease intervals or
faulty grease application [13]. Stator faults have mainly the form of stator insulation
damages caused by fatigued insulation, high temperatures, environmental influences
(dust and moisture) and/or loaded high current start-ups, as well as by faults in

Fig. 1 Classification of typical faults in induction motors [7, 10, 11]

60 P. D. Strankowski and J. Guziński

the manufacturing process. Rotor faults are mostly induced by broken rotor bars or
broken short-circuiting end rings, mainly resulting from high vibration level, material
fatigue and also manufacturing faults.
The category of other faults includes eccentric position of the rotor, excessive
clearances in bearings, bent rotor shaft and rotor unbalance. Each abovementioned
fault changes the machine performance, which is reflected in different physical drive
conditions in which the diagnosis takes place. Table 1 presents the responses of the
electric drive to common faults [10].
As can be interpreted from Table 1, there are several symptoms which allow to
diagnose common faults such as stator inner winding shorting, rotor cage damages or
static eccentricity, while specific faults such as bearing faults can be identified only
by one physical relation. For this reason, expert knowledge is indispensable in fault
identification. Sections 5 and 6 will present commonly used fault detection methods.

3 Mechanical Faults in Electric Drives

Mechanical faults occurring in the drive system depend on its mechanical struc-
ture. According to the classifications presented in the literature, the most frequent
mechanical faults can be divided into [14–16]:
• mechanical looseness (e.g. soft foot),
• faulty belt and chain drive,
• damaged or worn gears,
• rotating members out of balance,
• shaft misalignment and bent shafts,
• damages and excessive clearances in rolling bearing elements,
• faulty coupling elements (e.g. clutches),
• resonances (due to fatigued parts, such as vibroisolators or damping elements or
process changes stimulated by temperature or pressure).
Significant attention is paid to the faults in the mechanical transmission system,
where gear faults are the key element [17–26]. Typical gear tooth abnormalities are,
for instance, tooth breakage, root crack, chipping and damages on the tooth surface
[26, 27]. Each mechanical damage and/or abnormality causes changes in the vibration
behaviour of the complete drive system. The types and sources of forces that affect
the vibration spectrum are summarized in Table 2 [10].
The impact of these forces on the vibration spectrum was analyzed in [14].
Selected fault types are given in Table 3, which presents their nature and the expected
frequency in relation to the mechanical rotor speed frequency.
For reasons of safety and health for human [28–31] and machinery, the vibration
level of an electrical drive in the industry is limited by international standards, which
have to be fulfilled in all industrial applications.
Faults and Fault Detection Methods in Electric Drives 61

Table 1 Diagnostic signals

Diagnostic signal source Detectable faults
and detectable faults [10]
Vibration Loose motor fitting, rotor
eccentricity
Stator winding break, inaccurate
connection on stator terminals
Loose stator windings
Damaged bearing parts
Loose bearing housing
Inadequate bearing grease
Asymmetric load
Static eccentricity
Dynamic eccentricity
Rotor cage damages
Motor supply current Stator winding break, inaccurate
connection on stator terminals
Damaged bearing parts
Static eccentricity
Dynamic eccentricity
Rotor cage damages
Asymmetric supply
Stator inner winding shorting
Motor supply voltage Rotor cage damages
Stator inner winding shorting
Earth faults
Temperature Motor inner pollution
Cooling disturbances
Magnetic flux Stator winding break, inaccurate
connection on stator terminals
Static eccentricity
Dynamic eccentricity
Rotor cage damages
Asymmetric supply
Stator inner winding shorting

There are various standards according to condition monitoring of electrical drives,

such as thermography diagnostics [32], certification of personnel [33], signal process-
ing and data interpretation [34, 35], to mention only a few. The vibration condition
monitoring standards are collated in Table 4.
The importance of drive system vibrations is reflected in the standards which have
to be fulfilled and maintained by the industry. Measurements have to be done on both
stationary and rotary parts of the machine.
62 P. D. Strankowski and J. Guziński

Table 2 Forces caused by mechanical faults [10]

Type of force Source
Centrifugal Unbalances (e.g. rotor), eccentricity
Kinematic Uneven surface
Parametric Change of shaft stiffness, bearings
(due to e.g. temperature changes)
Friction Elements of rolling, slipping and pressing property
Strokes Damaged pressing surfaces

Table 3 Frequency harmonics related to faults [14]

Type of fault Dominant vibration frequency
Rotating members out of balance ωr
Shaft misalignment and deformation ωr , 2ωr , 3ωr , 4ωr
2 ωr , 3 ωr
1 1
Journal bearings loose in housing
Damaged or worn-out gears ωr · n teeth
Looseness in mechanical parts 2ωr
Faults in belt drive ωr , 2ωr , 3ωr , 4ωr

Table 4 International standards regarding vibrations of machines

Standard Description
ISO 20816–1:2016 Mechanical vibration—Measurement and evaluation of machine
vibration—Part 1: General guidelines
ISO 13373–1:2002 Condition monitoring and diagnostics of machines—Vibration
condition monitoring—Part 1: General procedures
ISO 13373–2:2016 Condition monitoring and diagnostics of machines—Vibration
condition monitoring—Part 2: Processing, analysis and presentation of
vibration data
ISO 21289:2008 Mechanical vibration and shock—Parameters to be specified for the
acquisition of vibration data
ISO 21940–31:2013 Mechanical vibration—Rotor balancing—Part 31: Susceptibility and
sensitivity of machines to unbalance
ISO 10816–3:2009 Mechanical vibration—Evaluation of machine vibration by
measurements on non-rotating parts—Part 3: Industrial machines with
nominal power above 15 kW and nominal speeds between 120 rpm and
15 000 rpm when measured in situ

4 Economic Aspects of Fault Detection in Electric Drives

In general, every industrial invention has to be economic-oriented to awake interest

and find application. However, every diagnostic system is connected with additional
expense, which might be not worthwhile in some drive systems, especially those
Faults and Fault Detection Methods in Electric Drives 63

representing the low power range (PM < 10 kW). This additional cost leads to certain
considerations that have to be made in order to implement a diagnostic system.
First, accurate planning of the intended diagnostics and its positive effects has to be
done. Next, the general positive effects of the diagnostic system have to be taken
into account and compared with the overall costs of the plan, including periodic
costs which are to be assessed in contrast to the effective efforts of the costs in
time. These efforts can be divided into applications that result in the improvement
of production quality and productivity, as well as the decrease of the resource and
labour consumption in production tasks [10]. Moreover, the diagnostic system can
take part in decreasing the intervals of maintenance and repair of the drive system.
In his monograph [10], Sw˛edrowski presented a pointer of effectiveness, in which
factor E > 1 means that the evaluated diagnostic system is appreciative. The goal
of optimization is to reach a diagnostic system that shows the maximum possible
factor E.


n 
l
oi + ui
i=1 i=1
E= (1)
JFE · S + kd

where
oi —saving effect within one year
ui —application effect within one year
n—number of saving effect types
l—number of application effects
J FE —all financial expenses needed to start the diagnostic task
s—average amortization rate of the diagnostic task
k d —annual operating costs of operating the diagnostics.
Considering an automatic fault detection procedure based on the values estimated in
the processor, from Eq. (1), it can be concluded that the sum of coefficients in the
numerator can reach a high value, while the denominator will reach a small value
when the development process of the fault detection procedure is excluded from the
costs of J. Consequently, the value of E will reach a high value, which underlines the
importance of such an implementation for fault detection, especially in low power
drives.

5 Conventional Methods of Fault Detection

In recent years, the interest in fault detection in electrical drives has increased. The
technology progress and the complexity of modern drives make the field of fault
detection desired by the industry. Early detection of abnormality in electrical motors
can prevent a total failure of the drive system. However, in order to detect faults, an
indicator is needed, of invasive or non-invasive type, to identify the actual condition
64 P. D. Strankowski and J. Guziński

of the machine, including changes of electrical properties (current, voltage or flux),

changes of physical behaviour (vibration, noise, torque or temperature), as well as
changes of chemical compounds. The access to these properties is possible with
sensors which measures relevant parameters.
Figure 2 illustrates the commonly used methods for condition monitoring of
electrical machines [7, 11, 36].
The motor current signature analysis (MCSA) is a very prominent and popular
method to detect abnormalities in electric machines. In this method, the stator current
of the machine is analyzed by means of signal processing methods. Flux measurement
can be also helpful to identify abnormal behaviour of the machine. However, the
required apparatus involves high costs and complex mounting. Consequently, the
flux measurement approach is undesired in the industry. Other electrical monitoring
opportunities include the motor voltage and motor power observations.
Another popular method is to measure the physical behaviour of the machine.
This includes the measurements of noise level change [37–40], and/or abnormalities
in the vibration characteristics, as well as the analysis of unsteady conditions in the
torque transmission system.
A further method which is commonly used in the industry is the machine temper-
ature analysis [41, 42]. There are also less frequently applied possibilities for fault
detection, such as chemical analysis of the coolant gas or other gases (ozone, carbon
monoxide, etc.) that might appear due to insulation degradation.
Each of the presented ways of fault identification serves as an indicator of unsteady
operation of the electrical machine. Signal processing and mathematical methods are
used to evaluate and interpret this indicator for potential fault detection. Common
techniques to identify unsteady operating conditions include the time averaging and
trend analysis, Fast and Short Time Fourier Transform, Hilbert Transform, Cepstrum
Analysis, Wavelet Transform, Bispectrum, Park Vector Pattern Recognition, Corre-
lation Analysis (Autocorrelation, Cross-Correlation), as well as Power Spectrum
Analysis and Spectral Kurtosis [11, 22, 36, 43–48].

Fig. 2 Methods of fault detection in electric drives [7, 11, 36]

Faults and Fault Detection Methods in Electric Drives 65

6 Model-Based Fault Detection Methods

Conventional fault detection methods have proved to be reliable for electrical drives
in the industry. However, the methods mentioned in Sect. 5 require in general special
equipment or sensors in order to detect different faults in electrical machines. This
additional equipment not only increases the cost of the complete drive system, but
it frequently makes the drive system more sensitive to disturbances and the fault
detection is done aside the operating machine.
The literature presents several methods of using mathematical models to identify
abnormalities in the investigated real object [25, 43, 49–61]. Certainly, it is possible
to adopt this method also to electrical drives to compare the ideal mathematical model
with the really existing drive unit. However, the accuracy of the model as well as the
abnormality detection depends on the simplifications assumed when designing the
mathematical model. Furthermore, in order to benefit from an online abnormality
detection possibility, a compromise is to be made between the calculation speed and
the model accuracy. The basic principle of such an approach is shown in Fig. 3.insert
The mathematical model is used as a parallel process to the real diagnosing system,
from which, the real system input variables are also brought to the mathematical
model. The concluded difference between the real drive and its mathematical model
provides an indicator of possible abnormality for the fault diagnosis. To implement
the mathematical model of a machine with low error impact, the following process
models are most frequently used to identify process changes [7, 50, 52, 62]:
• Neural networks [7, 63]—are not, in fact, a mathematical model of the process.
However, they are used to identify abnormalities of the process. The advantage of
applying such a network is that it does not require an exact mathematical model. In
this case, complete measurements of the real healthy process have to be provided in
order to perform the network training. All measurements are not always available,
which is the weak point of applying neural networks.
• State observers or estimators [64, 65], e.g. Luenberger, MRAS and extended
Kalman-Filter—are based on the mathematical interpretation of the process in
which the estimation error between the model and the process is brought to
the lowest possible level. The difficulty in this method is to determine specific
coefficients/gains that are responsible for estimation error reduction.
The literature presents several other ways of detecting abnormalities in the system,
e.g. parity equations [50, 52], Vienna Monitoring Method [61, 66] or fuzzy-based

Fig. 3 Schematic signal diagram of model-based fault detection

66 P. D. Strankowski and J. Guziński

method [67–69]. However, the application of observers will provide an insight into
the estimated variables as stator current, rotor flux, electromotive force, rotor angular
speed or load torque. Possible reuse of these variables for fault detection or fault
monitoring will enable additional life-expanding possibilities in the field of drive
systems used in the industry, e-mobility, and traction, as reference values or diagnostic
features for low-cost drives.

7 Conclusions

The monitoring of an electric drive is indispensable for a long-lasting drive operation.

In general, it could be said that up to now sophisticated drive monitoring and fault
detection procedures are used exclusively in high-cost machines that have to ensure
a reliable non-stop operation. The literature review presents that there are several
successful fault detection methods that found their application in the industry. The
most reliable fault detection procedures are up to now based on the measurement
of additional sensor signals that have to be evaluated by experts. Nevertheless, the
development in the field of µC-Technology with a simultaneous decrease in the
prices enable new additional capacities for low-cost drives. The actual trend shows
that the introduction of advanced mathematical model, into the software without any
further costs for the hardware, allows a fault detection possibility for low-cost drives,
and an additional source of information for drive system with high reliability.

Acknowledgements This work was supported by the National Science Centre of Poland grant no.
2015/19/N/ST7/03078.

References

1. F. Blaschke, A new method for the structural decoupling of A.C. induction machines, in Conf.
Rec. IFAC, (Düsseldorf Germany, 1971)
2. M. P. Kaźmierkowski, R. Krishnan, F. Blaabjerg, Control in Power Electronics: Selected
Problems (Academic press, 2002)
3. M. Depenbrock, Direct self-control (DSC) of inverter-fed induction machine. IEEE Trans.
Power Electron. 3(4), 420–429 (1988)
4. Z. Krzeminski, Nonlinear control of induction motor, in 10th IFAC Congress on Automatic
Control (1987), pp. 349–354
5. IEEE Motor Reliability Working Group, Report of large motor reliability survey of industrial
and commercial installations, part I. IEEE Trans. Ind. Appl. IA-21(4), 853–864 (1985)
6. IEEE Motor Reliability Working Group, Report of large motor reliability survey of industrial
and commercial installations. IEEE Trans. Ind. Appl. IA-21(4), 153–158 (1986)
7. C.T. Kowalski, Drive system diagnosis with induction motor and use of artificial intelligence—
Diagnostyka układów nap˛edowych z silnikiem indukcyjnym z zastosowaniem metod sztucznej
inteligencji—in Polish. Oficyna Wydawnictwa Politechniki Wrocławskiej (2013)
8. P. Zhang, Y. Du, T.G. Habetler, B. Lu, A survey of condition monitoring and protection methods
for medium-voltage induction motors. IEEE Trans. Ind. Appl. 47(1), 34–46 (2011)
Faults and Fault Detection Methods in Electric Drives 67

9. D.U. Campos-Delgado, D.R. Espinoza-Trejo, E. Palacios, Fault-tolerant control in variable

speed drives: a survey. IET Electr. Power Appl. 2(2), 121–134 (2008)
10. L. Sw˛edrowski, Measurement in squirrel cage induction motor diagnostic—Pomiary w diag-
nostyce silników indukcyjnych klatkowych—in Polish. Wydawnictwo Politechniki Gdańskiej
(2013)
11. S.A. Mortazavizadeh, S.M.G. Mousavi, A review on condition monitoring and diagnostic
techniques of rotating electrical machines. Phys. Sci. Int. J. 4(3), 310 (2014)
12. A. Romanenko, Study of Inverter-Induced Bearing Damage Monitoring in Variable-Speed-
Driven Motor Systems (Acta Univ, Lappeenrantaensis, 2017)
13. C. Radu, The most common causes of bearing failure and the importance of bearing lubrication.
RKB Tech. Rev. 1–7 (2010)
14. C.M. Harris, A.G. Piersol, Harris’ Shock and Vibration Handbook (McGraw-Hill, 2002)
15. M. Blodt, P. Granjon, B. Raison, J. Regnier, Mechanical fault detection in induction motor
drives through stator current monitoring-Theory and application examples, in Fault detection,
InTech (2010)
16. T.-J. Kang, C. Yang, Y. Park, D. Hyun, S.B. Lee, M. Teska, Electrical monitoring of mechanical
defects in induction motor-driven V-belt–pulley speed reduction couplings. IEEE Trans. Ind.
Appl. 54(3), 2255–2264 (2018)
17. D. He, R. Li, A new vibration signal processing method for gearbox fault detection, in 2011
IEEE Conference on Prognostics and Health Management (PHM) (2011), pp. 1–5
18. J. Arellano-Padilla, M. Sumner, C. Gerada, L. Jing, A novel approach to gearbox condition
monitoring by using drive rectifier input currents, in 13th European Conference on Power
Electronics and Applications, 2009. EPE’09 (2009), pp. 1–10
19. N. J. Wismer, Gearbox analysis using cepstrum analysis and comb liftering, Appl. Note Brüel
Kjaer Den. (1994)
20. N. Saravanan, K.I. Ramachandran, Incipient gear box fault diagnosis using discrete wavelet
transform (DWT) for feature extraction and classification using artificial neural network (ANN).
Expert Syst. Appl. 37(6), 4168–4181 (2010)
21. L. Gelman, N.H. Chandra, R. Kurosz, F. Pellicano, M. Barbieri, A. Zippo, Novel spectral
kurtosis technology for adaptive vibration condition monitoring of multi-stage gearboxes.
Insight—Non-Destr. Test. Cond. Monit. 58(8), 409–416 (2016)
22. L. Gelman, K. Solinski, B. Shaw, M. Vaidhianathasamy, Vibration diagnosis of a gearbox
by wavelet bicoherence technology. Insight—Non-Destr. Test. Cond. Monit. 59(8), 440–444
(2017)
23. R. Zhang, F. Gu, H. Mansaf, T. Wang, A.D. Ball, Gear wear monitoring by modulation signal
bispectrum based on motor current signal analysis. Mech. Syst. Signal Process. 94, 202–213
(2017)
24. C. Hu, W.A. Smith, R.B. Randall, Z. Peng, Development of a gear vibration indicator and its
application in gear wear monitoring. Mech. Syst. Signal Process. 76–77, 319–336 (2016)
25. J. Zhang, J.S. Dhupia, C.J. Gajanayake, Stator current analysis from electrical machines using
resonance residual technique to detect faults in planetary gearboxes. IEEE Trans. Ind. Electron.
62(9), 5709–5721 (2015)
26. S.H. Kia, H. Henao, G.-A. Capolino, Gear tooth surface damage fault detection using induction
machine stator current space vector analysis. IEEE Trans. Ind. Electron. 62(3), 1866–1878
(2015)
27. S.H. Kia, H. Henao, G.-A. Capolino, Fault index statistical study for gear fault detection using
stator current space vector analysis. IEEE Trans. Ind. Appl. 52(6), 4781–4788 (2016)
28. International Organization for Standardization, ISO 2631-1:1997 Mechanical vibration and
shock—Evaluation of human exposure to whole-body vibration—Part 1: General requirements
29. International Organization for Standardization, ISO 2631-2:2003 Mechanical vibration and
shock—Evaluation of human exposure to whole-body vibration—Part 2: Vibration in buildings
(1–80 Hz)
30. International Organization for Standardization, ISO 5349-1:2001 Mechanical vibration—Mea-
surement and evaluation of human exposure to hand-transmitted vibration—Part 1: General
requirements
68 P. D. Strankowski and J. Guziński

31. International Organization for Standardization, ISO 5349-2:2001 Mechanical vibration—Mea-

surement and evaluation of human exposure to hand-transmitted vibration—Part 2: Practical
guidance for measurement at the workplace
32. International Organization for Standardization, ISO 18434-1:2008 Condition monitoring and
diagnostics of machines—Thermography—Part 1: General procedures
33. International Organization for Standardization, ISO 18436-2:2014 Condition monitoring and
diagnostics of machines—Requirements for qualification and assessment of personnel—Part
2: Vibration condition monitoring and diagnostics
34. International Organization for Standardization, ISO 13379-1:2012 Condition monitoring and
diagnostics of machines—Data interpretation and diagnostics techniques—Part 1: General
guidelines
35. International Organization for Standardization, ISO 13374-1:2003 Condition monitoring and
diagnostics of machines—Data processing, communication and presentation—Part 1: General
guidelines
36. P.J. Tavner (ed.), Condition Monitoring of Rotating Electrical Machines (Institution of
Engineering and Technology, London, U.K, 2008)
37. D. Karkosiński, Vibroaccustic Occurances in Asynchron Squirrel Cage Motors—Zjawiska
wibroakustyczne w asynchronicznych silnikach klatkowych—in Polish. Wydawnictwo
Politechniki Gdańskej (2006)
38. F. Salazar-Villanueva, O.G. Ibarra-Manzano, Spectral analysis for identifying faults in induc-
tion motors by means of sound, in CONIELECOMP 2013, 23rd International Conference on
Electronics, Communications and Computing (2013), pp. 149–153
39. A. Glowacz, W. Glowacz, Z. Glowacz, J. Kozik, Early fault diagnosis of bearing and stator
faults of the single-phase induction motor using acoustic signals. Measurement 113, 1–9 (2018)
40. H. Akcay, E. Germen, Identification of acoustic spectra for fault detection in induction motors.
AFRICON 2013, 1–5 (2013)
41. G. Singh, T.C. Anil Kumar, V.N.A. Naikan, Induction motor inter turn fault detection using
infrared thermographic analysis. Infrared Phys. Technol. 77, 277–282 (2016)
42. D. Lopez-Perez, J. Antonino-Daviu, Application of infrared thermography to failure detection
in industrial induction motors: case stories. IEEE Trans. Ind. Appl. 53(3), 1901–1908 (2017)
43. S. Karmakar, S. Chattopadhyay, M. Mitra, S. Sengupta, Induction Motor Fault Diagnosis
(Springer Singapore, Singapore, 2016)
44. A.M. Da Silva, Induction Motor Fault Diagnostic and Monitoring Methods (Marquette
University, 2006)
45. T. Ghanbari, Autocorrelation function-based technique for stator turn-fault detection of
induction motor. IET Sci. Meas. Technol. 10(2), 100–110 (2016)
46. M. Riera-Guasp, J.A. Antonino-Daviu, G.-A. Capolino, Advances in electrical machine, power
electronic, and drive condition monitoring and fault detection: state of the art. IEEE Trans. Ind.
Electron. 62(3), 1746–1759 (2015)
47. J. Kim, S. Shin, S.B. Lee, K.N. Gyftakis, M. Drif, A.J.M. Cardoso, Power spectrum-based
detection of induction motor rotor faults for immunity to false alarms. IEEE Trans. Energy
Convers. 30(3), 1123–1132 (2015)
48. M. Kang, J.M. Kim, Reliable fault diagnosis of multiple induction motor defects using a 2-D
representation of Shannon wavelets. IEEE Trans. Magn. 50(10), 1–13 (2014)
49. O. Moseler, R. Isermann, Model-based fault detection for a brushless DC motor using parameter
estimation, in Proceedings of the 24th Annual Conference of the IEEE on Industrial Electronics
Society, 1998. IECON’98, vol. 4 (1998), pp. 1956–1960
50. R. Isermann, Fault-Diagnosis Systems: an Introduction from Fault Detection to Fault Tolerance
(Springer, Berlin; New York, 2006)
51. M. Blanke, J. Schröder (eds.), Diagnosis and Fault-Tolerant Control, 2nd edn. (Springer, Berlin;
New York, 2006)
52. R. Isermann, Fault-Diagnosis Applications (Springer, Berlin Heidelberg, 2011)
53. J. Chen, R.J. Patton, Robust Model-Based Fault Diagnosis for Dynamic Systems, vol. 3
(Springer, US, Boston, MA, 1999)
Faults and Fault Detection Methods in Electric Drives 69

54. A. Singh, B. Grant, R. DeFour, C. Sharma, S. Bahadoorsingh, A review of induction motor

fault modeling. Electr. Power Syst. Res. 133, 191–197 (2016)
55. P. Kołodziejek, Induction Motor Rotor Fault Identification with Sensorless Control System
for Diagnostic Purposes—Identyfikacja uszkodzeń wirnika silnika indukcyjnego z bezczu-
jnikowym układem regulacji dla celów diagnostycznych—in Polish, Gdańsk University of
Technology, Gdańsk (2009)
56. C.S. Kallesoe, V. Cocquempot, R. Izadi-Zamanabadi, Model based fault detection in a
centrifugal pump application. IEEE Trans. Control Syst. Technol. 14(2), 204–215 (2006)
57. K. Ohishi, K. Nakano, I. Miyashita, S. Yasukawa, Anti-slip control of electric motor coach based
on disturbance observer, in 1998 5th International Workshop on Advanced Motion Control,
1998. AMC’98-Coimbra (1998), pp. 580–585
58. J. Guzinski, M. Diguet, Z. Krzeminski, A. Lewicki, H. Abu-Rub, Application of speed and load
torque observers in high-speed train drive for diagnostic purposes. IEEE Trans. Ind. Electron.
56(1), 248–256 (2009)
59. S. Benmoussa, M.A. Djeziri, Remaining useful life estimation without needing for prior
knowledge of the degradation features. IET Sci. Meas. Technol. 11(8), 1071–1078 (2017)
60. H. Torkaman, E. Afjei, Sensorless method for eccentricity fault monitoring and diagnosis in
switched reluctance machines based on stator voltage signature. IEEE Trans. Magn. 49(2),
912–920 (2013)
61. C. Kral, F. Pirker, G. Pascoli, Model-based detection of rotor faults without rotor position
sensor—the sensorless vienna monitoring method. IEEE Trans. Ind. Appl. 41(3), 784–789
(2005)
62. D. Miljković, Fault detection methods: a literature survey, in MIPRO, 2011 Proceedings of the
34th International Convention (2011), pp. 750–755
63. C.T. Kowalski, Monitoring and fault diagnosis of induction motors with neural networks—
Monitorowanie i diagnostyka uszkodzeń silników indukcyjnych z wykorzystaniem sieci neu-
ronowych—in Polish,” Pr. Nauk. Inst. Masz. Nap˛edów Pomiarów Elektr. Politech. Wroc.
Monogr. vol. 57, nr 18 (2005)
64. J. Lunze, Control Systems II—Regelungstechnik 2—in German (Springer, Berlin Heidelberg,
Berlin, Heidelberg, 2010)
65. H. Unbehauen, Control Systems II—Regelungstechnik II—in German (Vieweg + Teubner,
Wiesbaden, 2009)
66. C. Kral, F. Pirker, G. Pascoli, H. Kapeller, Robust rotor fault detection by means of the vienna
monitoring method and a parameter tracking technique. IEEE Trans. Ind. Electron. 55(12),
4229–4237 (2008)
67. D.F. Leite, M.V. Arauijo, L. Secco, P. Costa, Induction Motors Modeling and Fuzzy Logic
Based Turn-To-Turn Fault Detection and Localization (2007), pp. 90–95
68. W. Xu-hong, H. Yi-gang, Fuzzy Model based On-line Stator Winding Turn Fault Detection for
Induction Motors (2006), pp. 2272–2276
69. A. Wolfram, R. Isermann, Fault detection of inverter-fed induction motors using a multi-model
approach based on neuro-fuzzy models, in 2001 European Control Conference (ECC) (2001),
pp. 1367–1372
Introduction to Condition Monitoring
of Wide Area Monitoring System

Papia Ray and Debani Prasad Mishra

Abstract The complication of existing electrical power systems is gradually increas-

ing. This is stimulating researchers and developers to suggest new answers capable of
addressing several challenges, particularly those related to power system operation.
Massive penetration of asynchronously connected renewable energy generation, the
generation connected over inverters is significantly changing the dynamics of modern
power systems. Lately, the wide area monitoring system (WAMS) helps the power
system operators to continuously analyze all the features of a large power network in
real time. Utilizing phasor measurement units (PMUs), information can be recorded
and monitored. Wide Area Monitoring Protection and Control (WAMPAC) enables
power systems to detect disturbances and improve knowledge of network behavior
under dynamic conditions, enabling system operators to maximize power flow and
network stability. This chapter will give a basic overview of wide area monitoring
and describe the components associated with it and its applications.

Keywords Phasor measurement unit · Wide area monitoring system · WAMPAC

1 Introduction

1.1 General Overview

High-quality power is an essential requirement for industries as well as residential

applications. The distribution system transmits power to end users of an electrical
system, and the occurrence of faults of any kind in such a distribution system has
a great effect on the economy and security of electrical systems. Major blackouts

P. Ray (B)
Department of Electrical Engineering, VSSUT, Burla, India
e-mail: [email protected]
D. P. Mishra
Department of Electrical Engineering, IIIT, Bhubaneswar, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 71

H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_3
72 P. Ray and D. P. Mishra

are mainly caused due to reasons like lack of wide area view, unawareness of situa-
tions, inaccurate and inappropriate model of the system and unavailability of high-
resolution time-synchronized data. For that there is a need to design an automatic
and self-healing system with good communication infrastructure to transmit power.
But the Supervisory Control And Data Acquisition System (SCADA) system used
in [1] most of the time cannot detect the transient condition quickly enough. With the
advancement of the synchrophasor technique, the above downside can be overcome.
A phasor measurement unit (PMU) gives power system quantities, such as voltage,
frequency, and synchronized phasors to deliver dependable data for power flow mon-
itoring or instantaneous power system control [2]. Data are achieved from several
stations and then guided to a control unit where these data are examined and control
signals are produced. Global positioning system (GPS) provides a common time ref-
erence for all achieved data to fulfill the needs of real-time control. PMU is recently
used for establishing smart grids to improve control and monitoring of it described
in [3]; PMU measurements are used for assessment of islanding detection, voltage
stability, oscillation monitoring, and state estimation [4]. PMUs are very expensive
to install because of the infrastructural requirements. It requires GPS signals, Ether-
nets, and other instruments, which make the setup expensive. Thus deployment of the
PMUs holds the foremost priority in order to reduce the overall cost of installation.
The main aim is to achieve full observability using less number of PMUs. PMUs at
important locations can be used to estimate the voltage and other phasors of other
buses where PMU is not placed. Previously many investigative methods are proposed
to find the optimal PMU placement (OPP) of PMU. Also, for different kinds of fail-
ures like PMU outage, transmission line outage is considered to make the system
immune. Previously, integer linear programming [5] was used to solve the placement
problems. But, in practice the existence of conventional measurements will make the
optimization problem nonlinear which is difficult to deal with. Different techniques
like Simulated annealing [6], Genetic algorithm [7, 8], SWAMP optimization [9],
and TLBO have been worked out to find the optimal place for PMUs. The different
methods are considered to find a method that will have less computational efficiency
and minimum PMU requirement.

1.2 Advantages of Phasor Measurement Units General

Overview

Accurate measurements are the basic requirement of any measurement system. Using
PMU, we can achieve precision and robustness in a measurement system. PMU
enables the power system to deal with severe faults as more accurate data can be
made available for the control and protection units. This also eliminates the chance
of getting bad data as we get a wide area view using the synchophasor technique.
Here GPS is used to time stamp the measurements that are useful to monitor complex
power systems. The measurement samples we get per second in PMU are more than
Introduction to Condition Monitoring … 73

that of any previously used techniques. Thus, transient faults of short duration can
be detected more accurately.
The followings are the advantages of the PMU measurements:
• Samples frequency is more.
• Wide area view can be obtained.
• Bad data can be eliminated.
• Timestamped data.
• Phasor angle can be found.

1.3 Problems in Condition Monitoring

The condition monitoring is prone to numerous errors that can affect the timing of
fault clearance and also affects the system security largely. Some problems are given
as follows:
• Lack of precise data,
• Unavailability of wide area view,
• Inaccessible timestamped data,
• Bad data elimination, and
• Unavailability of phasor angle measurements.
These problems make the condition monitoring difficult using the SCADA system
because of its low sampling rate and magnitude-based measurement. These demerits
introduce the PMU system for condition monitoring and monitoring in the power
system.

1.4 Literature Review

Many authors have done noteworthy research in the field of condition monitoring
using the phasor measurement unit. Some of the research work are discussed as
follows.
De La Ree et al. [2] proposed the basics of the synchrophasor measurements. This
paper describes the block diagram of a phasor measurement unit and discusses all
the constituents. It also gives the idea of wide area measurements and the technique
of monitoring power systems. The control and protection system accompanied by
the PMUs are also discussed in the paper. The out of step relay performance is
also considered in the measurement systems. The placement problem using integer
linear programming is carried out to find the important place for PMU installation.
The SCADA system is taken as a backup for the PMU measurements to ensure
robustness.
74 P. Ray and D. P. Mishra

Phadke et al. [10] provide the fundamentals of a PMU system and how to get
voltage and current phasors. It also deals with the frequency measurements and the
rate of change of frequency as the input to PMU changes. It also introduces the
application of GPS signals to the measurement system, thus enabling the system
to produce more precise data. It also limits the measurement error by comparing it
with the IEEE C37.118 [11]. The deviation of output signals needs to be within the
permissible limit decided by the IEEE standards.
Ashok Kumar and Karthikeyan [12] implement the PMU in a distribution network
to detect faults. Two cases are studied: one is a normal case and the other is a faulted
case. The phasor measurements like magnitude, phasor angle, and frequency of the
input signal were considered. The signals of different measurements are distinguished
between normal and faulted case. It is found that when the phasor angle crosses
the tolerance limit, there exists a fault in the system. The phasor angle deviation
can be treated as an input signal to the circuit breakers to operate quickly. This
quick operation can effectively eliminate detrimental effects on the reliability of the
whole system. Further, the PMU’s output error called the total vector error (TVE) is
compared with the standard given in the IEEE C37.118 standards. The performance
is found to be within the limits, which validates its effectiveness in power system
monitoring.
Funaki et al. [13] discusse Discrete Fourier Transform (DFT) used to sample the
measurements to the complex-valued function of frequency. The DFT converts the
time-domain input components to the frequency domain. In PMU, DFT is used to
find the fundamental frequency and the rate of change of frequency. The detailed
equations used in DFT are described and the method of calculating fundamental
frequency is given in the paper. An anti-aliasing filter with sample frequency greater
Nyquist frequency is used to eliminate the loss of measurements. Finally, the DFT
output is evaluated to satisfy the basic performance index.
Schweitzer et al. [14] deal with the control and protection unit application along
with PMU measurements. This paper shows different techniques to be used to trip the
circuit breakers. If the phasor angle deviation is found to be exceeding the tolerance
limits, this will indicate the existence of any fault in the system. The angle deviation
can be used to trip the circuit breakers.
Di Shi et al. [15] explain the errors that may persist in the synchophasor measure-
ments which may affect the monitoring system. A performance factor called total
vector error (TVE) is introduced which will provide the effectiveness of any PMU
system. The difference between the estimated signal and the measured signal will
decide the value of TVE. The TVE is compared with the IEEE standards to deliver
the desired performance. The limits for normal operating condition are found to be
1% for any measurement. Similarly, the presence of harmonics will also change the
TVE value for a PMU.
Bei Gou [5] delivers the idea of implementing PMUs at the strategic locations
and thereby reducing the cost of the monitoring wide area. The basic problem is
formulated using linear programming where the objective is to minimize the number
of PMUs. The constraints are decided so as to ensure at any condition that there
must be at least one PMU monitoring any given bus. Also the effects of the zero
Introduction to Condition Monitoring … 75

injection measurements (ZIB) are considered here which will reduce the number
of synchrophasor units. Different IEEE test systems are considered and placement
location is found out for two different cases: with ZIB and without ZIB measurement.

2 Methodology for Condition Monitoring

Power systems are the vast networks that provide power to the whole range of load
like industrial and residential loads. The basic requirement for any load is high-
quality power to ensure the effective utilization of the power. But due to lack of wide
area view, inaccurate monitoring data, and slower detection of abnormalities, the
system becomes unreliable and unsecure. The SCADA system used most of the time
cannot detect the transient condition quickly enough. Due to these inefficiencies, the
protection system fails to avoid transient faults. Phasor measurement unit (PMU)
measures the voltage and current waveform and produces time synchronized data
which consists of synchronized phasor, frequency of the signal, and rate of change
of frequency approximations. These data can help in acquiring current state of the
power system. Thus deployment of the PMUs holds the foremost priority in order to
reduce the overall cost of installation.

2.1 Synchronized Measurement Technology

PMU is used to measure the power system quantities and the measurements are
inspired by a concept called phasor. The details of the phasor measurements are
described below.

2.2 Definition of a Phasor

Any sinusoidal signal can be represented by a distinctive complex number known as

a phasor. Consider a sinusoidal signal as described in [5].

x(t) = X m cos(ω t + φ) (1)

Xm Xm
X = √ eiθ = √ (cosφ + jsinφ) (2)
2 2

where x(t) = input signal, w = frequency, ø = phase angle, X m = peak input.

76 P. Ray and D. P. Mishra

Fig. 1 Phasor representation of a sinusoidal signal [10]: a sinusoidal signal, b phasor representation

The signal and the phasor representation of it are given by (1) and (2) which are
demonstrated in Fig. 1. The phase angles are positive when measured in a counter-
clockwise direction from the real axis. If the signal remains stationary, the phase
will remain constant. For practical cases, the phasor representation may change if
the signal does not remain constant and frequency varies.

2.3 Phasor Measurement Concepts

Data window concept given in [6] for PMUs normally is one period of the funda-
mental frequency of signal. Generally frequency is not equal to its rated value, so
the frequency-tracking step given in [7] is used in PMU and therefore estimates the
fundamental frequency component period earlier than estimation of phasor. Clearly
harmonic or nonharmonic components may be present in the input signal. It is PMU’s
job to separate fundamental frequency components from the total signal. Commonly
to determine the phasor of input signal, samples are taken and DFT is applied as given
in [8, 9]. Anti-aliasing filters are to be applied to the signal before the data samples
are taken to eliminate the high-frequency noise from the signal. The anti-aliasing
filters are the low pass filters that pass the signals up to half of sampling frequency.
Let xk s have N samples between one period,
√ N −1
2
xk e− jk N
2π
X= (3)
N k=0

Then
√ RMS value obtained is given by dividing the peak fundamental frequency
by 2. Continuous phasor estimation is done and newer data samples are needed
to update the phasor estimation, for which an algorithm is adopted as described in
[2]. The algorithm is called non-recursive as it takes a fresh sample for each window
and does not consider the previous estimates. Here we have used PMU (PLL-based
Introduction to Condition Monitoring … 77

Fig. 2 PMU block diagram

positive sequence) as per the IEEE standards given in [8]. Tracking of frequency
and phase of a three-phase signal (abc) is done by means of an internal frequency
oscillator. The phase difference is kept at zero controlling oscillator frequency by
the control system. In Fig. 2, the magnitude and phase also called positive-sequence
components for a sinusoidal three-phase signal is calculated for a running window
of the nominal frequency by a closed-loop system of PLL.
The computation needs reference frame is given by the angle in radian which
varies in the range 0–2π, synchronized with a fundamental or positive sequence of
phase A’s zero crossing. The control system also tracks the angle of the input. As a
Fourier analysis is performed for running average estimation, correct magnitude and
phase can only be achieved after one cycle is completed. Initial input parameters hold
the output constant at the specific for initial cycle of simulation. The three outputs
of the PMU (PLL-based, positive-sequence) block return the magnitude (same units
as the input signal), the phase (in degrees relative to the PLL phase), and positive-
sequence component’s frequency. In the block sample, time (T s ) in seconds is a
function of the nominal frequency F n and the sampling rate N sr , as follows:

1
TS = (4)
Fn × Nsr

Finally, the reporting rate (Rt ) determines the span for which the event needs to
be reported. Here, the reporting rate factor decides the reporting rate as follows:

Rt = k × Ts (5)
78 P. Ray and D. P. Mishra

3 Applications and Benefits of Synchronized Measurement

Technology

Accurate measurements are the basic requirement of any measurement system. Using
PMU, we can achieve the precision and robustness in the measurement system. The
PMU enables the power system to deal with severe faults as more accurate data can
be made available for the control and protection units. This also eliminates the chance
of getting bad data as we get a wide area view using the synchophasor technique.
Here GPS is used to time stamp the measurements that are useful to monitor complex
power systems. The measurement samples we get per second in PMU are more than
that of any previously used techniques. Thus, transient faults of short duration can
be detected more accurately.
The following are the advantages of the PMU measurements:
• Samples frequency is more.
• Wide area view can be obtained.
• Bad data can be eliminated.
• Timestamped data.
• Phasor angle can be found.

4 Architecture of a Condition Monitoring System

The SCADA system used in power system monitoring is found to be inefficient

enough to detect transient fault at times. This makes PMU implementation more
important, as it gives more precise measurements of power systems. The SCADA
system previously gave not only magnitude measurement but also magnitude devia-
tion for less-severe fault . But using phasor angle deviation, less-severe faults can be
detected more effectively. PMU with positive sequence PLL is used to detect the fault
quickly and eliminate the effect of it to ensure system security and reliability. For
placement problem, the different IEEE test systems taken are applied with different
optimization algorithms. Algorithms like ILP, GA, and TLBO are used to minimize
the number of PMUs.

4.1 System Description

A distribution system is designed which consists of two generating stations. One

synchronous generator (thermal power plant type) with 30 KW installed capacity is
at bus1 whereas a solar photovoltaic (PV) plant produces power at the other end at
bus2. The PV plant produces a total of 24 KW of power. Two short transmission lines
are used to transmit power to the two load centers. One transformer (Y-Y) rated at
Introduction to Condition Monitoring … 79

208 V/11 KV is used to raise the voltage at line, where the other two transformers
(Y-Y) are used to step down to 208 V. There are two loads each rated at 25 KW
power at bus2 and bus3, respectively. The detailed layout is given in Fig. 3 and its
numerous parameters are specified in Table 1.
The PV module consists of numerous solar cells connected both in series and
parallel. The whole setup gives a regulated DC output like voltage and current. The
output changes with a change in irradiance level and temperature. The PV system
is capable of harvesting an optimal power output of 24 kW. A three-phase inverter
along with a series transformer is used to convert the DC output to the desired AC
voltage.

BUS1 PMU1 BUS4 BUS5 BUS2

PMU2

10KM
Transmission line
CB1 CB2 CB3
11KV/208V Inverter DC-DC
Converter PV array
Synchronous 208V/11KV TRANSORMER BUS7
With MPPT
Generator TRANSORMER

CB5
25KW AC
BUS6 BUS3
Load
PMU3

10KM CB4
Transmission line 11KV/208V
TRANSORMER 25KW AC Load

Fig. 3 System description for condition monitoring

Table 1 System specification

Parameters Specification
Rated voltage and frequency 208 V, 50 Hz
Generating station 30 KW
PV grid 24 KW
Total load 50 KW
No. of loads One at (bus2) 25 KW
One at (bus3) 25 KW
Number of transformers (Y-Y) Three (208 V/11 kV)
Number of distribution lines Two (10 km each)
Number of buses 7
Report rate factor of PMU 1
Sampling rate of PMU 64
80 P. Ray and D. P. Mishra

Fig. 4 MATLAB model of the system

Near the generation station, PMUs are installed to monitor the load sides. PMU
data are aligned and rearranged to the IEEE C37.118 format before transferring to
the PDC. The MATLAB/Simulink model is shown in Fig. 4.

5 Results and Discussion

In MATLAB 2018a, a Simulink model is modeled to study the performance of PMU

in condition monitoring applications. To study the system’s behavior on faulted
conditions, two cases are taken here: one normal condition, the other is a faulted
condition. In this test, at bus2 a three-phase fault is simulated. The fault is created at
0.8 s and was cleared at 1 s.

5.1 Performance of PMU1

The PMU installed at bus1 gives outputs like voltage and phase angle, which are
studied and compared in the below subsection.
Introduction to Condition Monitoring … 81

5.2 Test Under Normal Operating Conditions

The voltage and angle output of phasor were observed for healthy operating con-
ditions. Figures 5 and 6 give the outputs of PMU1 installed at bus1. The output of
PMU1 like magnitude of voltage and phase angle acts as the reference phasor for
calculating the difference in the phasor of the other buses.
From Fig. 5 it can be observed that the voltage recorded by PMU1 is 208 V which
is 1 p.u. in the normal operating condition and is within the stable limits.
From Fig. 6 it can be noticed that the phase rotates a total of 360°, i.e., from +
180° to − 180° and shows no deviation in the normal operating condition.

Fig. 5 PMU1 voltage per unit in normal operation

Fig. 6 PMU1 phase per unit in normal operation

82 P. Ray and D. P. Mishra

Fig. 7 PMU1 voltage per unit in fault operation

Fig. 8 PMU1 phase per unit in fault operation

5.3 Test Under Faulted State Condition

In this test PMU1 reads small deviation in voltage magnitudes, which shows that the
fault is not situated on bus1. The voltage dips to 0.83 p.u. in the fault condition at
0.8 s and recover to 1 p.u. at 1 s (Figs. 7, 8).
No abnormalities in the phasor angle suggest that there is no fault on bus1.

5.4 Performance of PMU2

PMU2 installed at bus2 gives the following results which can be used to differentiate
between fault and normal conditions.
Introduction to Condition Monitoring … 83

Fig. 9 PMU2 voltage per unit in normal operation

Fig. 10 PMU2 phase per unit in normal operation

5.5 Test Under Normal Operating Conditions

The voltage waveform is found to be stable near 1 per unit (p.u.). after 0.6 s under
normal operating conditions (Figs. 9 and 10).
The phase plot of PMU2 does not show any deviation under the normal case and
rotates a total of 360°.

5.6 Test Under Faulted State Conditions

A large phase deviation along with a huge drop in voltage of the bus2 suggests
there was a fault in the bus. Phase angle deviation exceeding more than 10° at any
instance would be referred to as fault in the bus. When the data is processed and
then transferred to the control center, then there exists a fault at the given bus; a trip
84 P. Ray and D. P. Mishra

Fig. 11 PMU2 voltage per unit in fault operation

Fig. 12 PMU2 phase per unit in fault operation

signal is sent to the CB to operate quickly to protect other healthy parts; the detailed
method is given in [14] (Figs. 11 and 12).
Fault is detected at time 0.8 s by PMU as the phase angle shows deviation which
exceeds the tolerance limit 10° for this chapter, violating which will trip the circuit
breakers.

5.7 Performance of PMU3

Here, the PMU3 outputs are studied to find the impact of fault on bus3.
Introduction to Condition Monitoring … 85

Fig. 13 PMU3 voltage per unit in normal operation

Fig. 14 PMU3 phase per unit in normal operation

5.8 Test Under Normal Operating Conditions

The phasor voltage magnitude is found to 1 p.u. under the healthy state of the system.
Also, the phasor’s phase angle is found to be within limits (Figs. 13 and 14).
In the load side, PMU3 shows phase to be in the normal state.

5.9 Test Under Faulted State Conditions

Fault created at bus2 has less impact on bus3. So, PMUs 3 read negligible angle
deviation and a little deviation in voltage magnitudes (Fig. 15).
The voltage dips to 0.82 p.u. in the fault condition at 0.8 s and recovers to 1 p.u.
at 1 s (Fig. 16).
It is clear that the fault is not located on bus3, as the phase does not violate the
limit. Thus no control action is needed here.
86 P. Ray and D. P. Mishra

Fig. 15 PMU3 voltage per unit in fault operation

Fig. 16 PMU3 voltage per unit in fault operation

5.10 Phase Deviation Normal to Fault State

Here, Fig. 17 shows that under healthy operating conditions the phasor angle is found
to have no deviation from the reference values; thus the angle deviation is found to
be negligible. But, in faulted case, i.e., from 0.8 s to 1 s the phasor angle changes
drastically which will give the indication of large catastrophes in their early stage.
Let the phasor difference limit be 10°, above which if the angle deviates the circuit
breakers may operate to clear the fault as given in Fig. 17. This will able to protect
itself from fault using real-time PMU measurements.
Introduction to Condition Monitoring … 87

Fig. 17 Phase deviation normal to fault state

6 PMU Measurement Evaluation

PMU reading may differ from the theoretical reading we get from the synchrophasor
representations of a signal. A factor called TVE is used to describe the error allowed
in PMU measurements given in [15]. IEEE sets the limits of TVE allowed. The
difference between an estimated phasor to the theoretical representation of sinusoidal
phasor is called TVE. Normalization is done by dividing the difference with the
theoretical value. TVE expression is given as

(X r (n) − X r (n))2 + (X i (n) − X i (n))2
TVE(n) = (6)
(X r (n) − X i (n))2

where X r (n) is the magnitude estimation and X i (n) is the phasor angle estimation
when the unit is under test, X r (n) and X i (n) are the sequences of theoretical values
of magnitude and angle at time instance (n). For voltage TVE, the equation is given
as below, where V is voltage phasor
− → 
→ −
 V Measured − V Ideal 
TVE = −  (7)
→ 
 V Ideal 

6.1 Test Under Nominal Frequency

In this test, 50 Hz is taken as a reference. Voltage magnitude and all other parameters
are taken as nominal values. TVE achieved in the case is within the 1% standard
88 P. Ray and D. P. Mishra

Fig. 18 TVE of voltage at

different buses

limit given by IEEE C37.118. Figure 18 shows the TVE bar chart. IEEE limits are
in blue, whereas orange block shows the TVE found here.
In PMU2 the TVE is found to be higher than the others which is due to the
harmonics that are there in the PV system due to the control mechanism. On the
other hand, the PMUs at bus1 and bus3 do experience a small TVE.

6.2 Placement of PMU

The installation cost of a PMU system is very high and thus the proper placement of
PMU is required. A placement technique that is described in [17] finds buses in which
PMU needs to be installed. But here we had used the deviation of phasor to decide
the placement. The phasor angle difference detected in PMU2 is more, so the PMU2
holds atmost priority for the stability of the system. In PMU2 the TVE is found to be
higher than the others which worsens the PMU2 result if more harmonics are there
in the system. Thus PMU should be implemented to bus2 to eliminate catastrophes.

7 Conclusion

The results of the condition monitoring part imply the effectiveness of using phasor
angle deviation in detecting the presence of fault. The fault introduced for the duration
will affect the phasor angle and will change beyond the tolerance limit. The difference
from normal operation will be used as a trip signal to circuit breakers. This enables
the system to avoid catastrophic faults. The performance of the synchrophasor used
here is found to be giving error below the requirement. TVE of voltage is found
to be less than 1% of the total PMU reading. The IEEE C37.118 is used to validate
synchrophasor unit used here. All the performance indices show that the effectiveness
Introduction to Condition Monitoring … 89

of the system is used to detect the fault in a wide area monitoring system. The next
chapter deals with the optimal placement problem formulation and implementation
of the test systems.

References

1. J. Marihart, Communications technology guidelines for EMS/ SCADA systems. IEEE Trans.
Power Deliv. 16(2) (2001)
2. A.J. De La Ree, V. Centeno, J.S. Thorp, A.G. Phadke, Synchronized phasor measurement
applications in power systems. IEEE Trans. Smart Grid 1(1), 20–27 (2010)
3. V.C. Gungor, D. Sahin, T. Kocak, S. Ergut, A survey on smart grid potential applications and
communication requirements. IEEE Trans. Ind. Inform. 9(1), 28–42 (2013)
4. A.R. Zarate, G.E. Mauersberger, O.J.M. Smith, State Estimation of Power Systems (Electronics
Research Laboratory, University of California, Berkeley, Memo. ERL-M248, 1968)
5. B. Gou, Optimal placement of PMUs by integer linear programming. IEEE Trans. Power Syst.
23(3), 1525–1526 (2008)
6. A. Sharma, S.K. Jain, A review and performance comparison of power system state estimation
techniques, in 2018 IEEE Innovative Smart Grid Technologies—Asia (ISGT Asia) (2018),
pp. 770–775
7. F.J. Marin, F. Garcia-Lagos, G. Joya, F. Sandoval, Genetic algorithms for optimal placement
of phasor measurement units in electrical networks. Electron. Lett. 39(19), 1403–1405 (2003)
8. G.N. Korres, N.M. Manousakis, T.C. Xygkis, J. Löfberg, Optimal phasor measurement unit
placement for numerical observability in the presence of conventional measurements using
semi-definite programming. Gener. Transm. Distrib. IET 9(15), 2427–2436 (2015)
9. R. Venkata Rao, V.D. Kalyankar, Parameter optimization of machining processes using a new
optimization algorithm. Mater. Manuf. Process. 27(9), 978–985 (2012)
10. A. Phadke, J. Thorp, M. Adamiak, A new measurement technique for tracking voltage phasors,
local system frequency, and rate of change of frequency, in IEEE Transactions on Power
Apparatus and Systems (1983), pp. 1025–1038
11. IEEE Standard C37.118.1-2011, IEEE Standard for Synchrophasor Measurements for Power
Systems
12. L. Ashok Kumar, S. Karthikeyan, Modeling of phasor measurement unit for wide area mon-
itoring and control of smart grids with distributed energy resources, in IEEE Conference on
Technologies for Sustainability (2016), pp 1–7
13. T. Funaki, K. Matsuura, S. Tanaka, Error estimation and correction of detected phase by real
time DFT. Trans. Inst. Electr. Eng. Jpn. 120-B(12), 1682–1690 (2000)
14. E.O. Schweitzer et al., Synchrophasor-based power system protection and control applications,
in Modern Electric Power Systems (2010)
15. Di Shi, Daniel J. Tylavsky, Nairn Logic, An adaptive method for detection and correction of
errors in pmu measurements. IEEE Trans. Smart Grid 3(4), 1575–1583 (2012)
Introduction to Condition Monitoring
of Electrical Systems

Atul J. Patil, Arush Singh and R. K. Jarial

Abstract Primarily, any electrical system is an assemblage of various electrical

components which are interconnected in a specified manner to perform a specialized
function. An electric system can be as simple as a torch on one hand and it can be as
abstruse as a spacecraft, on the other. Detailed coverage of all these electrical systems
is beyond the purview of this book and hence, a pragmatic review has been made
to put together developments related to monitoring techniques of electrical power
transmission systems here. This system performs a distinguishing role of ensuring
reliable power supply for the utilization by consumers for various applications. The
preponderance of faults can be minimized by monitoring certain critical parameters
and predicting a fault before it is going to actually happen. In this chapter, an elec-
tric power transmission system is branched into Power Transformers, Transmission
Lines, Power Transmission Towers and Protective Relays. The primarily focus has
been laid to highlight perspective view on the techniques which assist in the develop-
ment of health prognosis-based (or condition-based) maintenance tools and practices
to safeguard the health of these systems in large interest of utilities.

1 Introduction

Electrical systems suffer numerous unusual circumstances that can unfavourably

influence the health of interconnected electrical equipment. Basically, an electric
system is described as any combination of electrical elements which are intended
to bring out any electrical work. But, from electricians’ point of view, an electrical
system is a mixture of multiple kinds of equipment, for example, transformers, relays,
circuit breakers, towers, transmission lines, etc. that are associated with transmission
and distribution of electrical power [4–8].

A. J. Patil · A. Singh · R. K. Jarial (B)

Department of Electrical Engineering, NIT Hamirpur (HP), Hamirpur, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 91

H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_4
92 A. J. Patil et al.

Electrical systems’ condition monitoring is of prime importance for looking after

electrical systems. Condition monitoring (CM) of the electrical equipment is noth-
ing but observing different kinds of parameters of transformers, transmission lines,
relays, circuit breakers, etc. to minimize unwanted breakdowns and equipment fail-
ure. Also, it is evident that CM techniques improve the lifespan of the equipment and
its reliability [5]. Condition monitoring techniques include the manner of observing
the parameters governing the health of various electrical equipment for the identi-
fication of variations in these parameters which is actually a fault sign. The value
of condition monitoring grants programmed maintenance or other procedures to be
taken to prevent significant disasters and bypass its outcomes. Condition Monitoring
has an individual advantage in those conditions that would curtail normal lifespan
which can be addressed before they grow into a significant crash [3–9].

2 What Is Condition Monitoring (in Context of Electrical

Systems)?

Electrical equipment gets damaged/degraded throughout its operational life. If an

electrical equipment goes out of operation suddenly, it impacts system reliability and
leads to loss of revenue. Condition monitoring is primarily a maintenance strategy
that deals with equipment health monitoring by measuring the changes in certain
parameters which may lead to equipment failure. For condition-based maintenance
to be possible, it is necessary to have knowledge about the electrical equipment’s state
and its rate of variation within a period. Trend monitoring and condition checking
are the two main approaches adopted for condition monitoring in electrical systems
[5–9]. Condition monitoring can be defined as follows:
Condition monitoring is a technique used by the companies to protect their investment
which involves measurement of mechanical conditions of the machinery such as vibration,
temperature, and pressure and performance degradation over time while it is still in operation.

Traditionally, periodic maintenance strategy is employed in which equipment

maintenance is performed as per the given schedule. But this strategy has certain
drawbacks which are addressed by different kinds of condition-based maintenance
approaches [7–10]. Table 1 enlists the differences between periodic maintenance and
condition-based maintenance.

3 Why Condition Monitoring of Electrical Apparatus?

Condition monitoring is a significant weapon in revealing a robust, condition-based

maintenance approach or predictive maintenance strategy of electrical system equip-
ment. Condition monitoring begins with employing interconnected sensors that regu-
larly trans-receive data which is utilized to analyse any deviations in the performance
Introduction to Condition Monitoring of Electrical Systems 93

Table 1 Comparison criteria of periodic maintenance and condition-based maintenance [8, 9]

Criteria Periodic maintenance Condition-based
maintenance
Data collection Theory/principle For maintenance, it Here, maintenance time is
looks at failure time not fixed; it depends on a
which is user-based parameter that indicates
data equipment condition
Practical issues • Failure time data is To collect data.
sometimes high-expensive sensors
inaccessible are required
• Failure data is
complex, time
spending and
sometimes costly to
collect
Data Theory/principle Reliability theory Deterioration modelling
analysis/modelling Practical issues • Quixotic assumptions • A lot of data is needed
• Effective when an • Regular cleaning of
equipment is in its data is needed for
deteriorating state, waveform-type data,
i.e. high failure rates and it is a complex task
Decision process Theory/principle Optimization technique Failure forecast strategy
Practical issues • Challenging to Perception of failure
design and interpret boundaries may be biased
• Decision model is not
always constant

of that element while in operation, any divergence in performance from the standard
elements would symbolize the probability of wear and tear and deterioration [4–8].
Condition monitoring not only describes the equipment condition but also affords
evidence for the forecast of its remaining operational period. The report gathered by
condition monitoring is adapted to make a maintenance plan and suggest part repair
ere catastrophic breakdown. Thus, it plays an essential purpose in both avoiding
unplanned shutdowns and in the forecast of life prospects [1–10]. The demand for
condition monitoring can be paraphrased as follows:
• Minimizes chances of spontaneous shut down due to catastrophic failure,
• Minimizes time paid out on maintenance and overhaul,
• Minimizes overtime costs by proper scheduling,
• Minimizes the requirement of emergency spares by life assessment,
• Minimizes the chances of collateral damage to the system,
• Minimizes the cost of asset failures,
• Is essential for early fault detection,
• Improves the reliability of the equipment,
• Minimizes the chances of disruption of normal operation of the asset,
• Has optimal maintenance intervals than manufacturer recommendations,
• Has residual life assessment,
94 A. J. Patil et al.

• Has safer operating conditions.

4 Condition Monitoring of Power Transformer

Transformers are highly reliable assets in any electrical grid, however, most of the
in-service transformers complete their design life span. Extending the serviceable
lifetime is challenging in case of a transformer but it is a powerful utility approach
to enhance the life of any power grid infrastructure [6–10]. Transformer condition
monitoring is the method of data acquisition and processing related to multiple
parameters in order to anticipate probable breakdown by examining the variation
of certain parameters pertaining to transformers from their predetermined values.
Transformer failures could lead to power outages, loss of revenue and poor reliability
of the entire grid [1–7]. Some of the major reasons due to which failures can occur
in a power transformer are
• Dielectric breakdown,
• Winding distortion (due to short-circuit),
• Formation of local hot spots in transformer winding and magnetic circuit,
• Degradation of insulation,
• Lightning,
• Ineffective maintenance,
• Overloading,
• Breakdown of auxiliary equipment such as OLTCs and bushings.
CM techniques play an important role in transformer diagnostic analysis. A
detailed coverage of primarily employed qualitative tactics is covered in this section.
These techniques are subdivided as shown in Fig. 1.

4.1 Oil Infrared Spectroscopy Technique

Mineral oil in a transformer acts as a dielectric, cooling medium, arc quenching

medium, etc. Moisture ingress can negatively impact the properties of the oil and
hence deteriorate the reliability of oil-filled transformers. Thus, for the determination
of water content in transformer oil, oil infrared spectroscopy technique was devel-
oped. It is a chemical-free tactic analysis. Here we are using infrared beam for the
identification of water content [11, 12]. Three functional units are involved in this
tactic, which are mentioned in Table 2.
In this technique, a radioactive lightning source is used to transmit infrared rays
in the oil sample. After propagation through the oil sample, a detector (receiver)
is located to collect refracted rays from oil samples. In between light source and
detector, oil sample is placed [11, 12].
Introduction to Condition Monitoring of Electrical Systems 95

Fig. 1 Transformer monitoring techniques

Table 2 Functional unit of

Unit Function
oil infrared spectroscopy [11,
12] Light source Transmit radioactive light source
Oil sample Light passes from transformer oil sample
holder
Detector Detector receives emitted light rays from the
oil sample

Also, one must ensure that the sample container is free of dust and unwanted
foreign substances as they can impact the test results. The detector checks the wave-
length and the intensity of the beam. Pure water absorbs light rays of wavelength
~3400 cm−1 [11, 12].
If any kind of impurity or water particles are present in the oil samples, then
there is a deviation in the wavelength and intensity of the transmitted rays. This
arrangement is as represented in Fig. 2.

4.2 Oil Dielectric Breakdown Test

The determination of electrical stress that an insulating oil can endure without failure
is called a dielectric breakdown voltage. It is estimated by applying a voltage among
96 A. J. Patil et al.

Fig. 2 Infrared spectroscopy [11, 12]

two electrodes which are located in the oil under the designated circumstances. This
dielectric test estimates the voltage at which the oil failed to work as a dielectric
medium. This dielectric voltage suggests the quantity of other unwanted particles
(usually moisture) in the oil. For reliable operation of the transformer the breakdown
voltage of oil should be at least 27 kV. Virgin oil should have the least dielectric
strength of 45 kV (as per International Standards IEEE C57.106-2006) [3–10, 13,
26]. The BDV test circuit is represented in Fig. 3.
Mineral oil is used for the transformer oil which has huge dielectric strength and
other cooling properties. The transformer oil works for cooling purpose, health indi-
cator purpose, fault indicator purpose and as an insulating material. But transformer
oil breakdown strength is reduced by the appearance of moisture contents. Break-
down voltage is measured by keeping the two-electrode sphere in oil by maintaining
sufficient distances, usually 2.5 cm. The test cell is built with translucent glass. Two
polished spherical steel electrodes having diameter of 12.5 mm are generally sep-
arated from each other by a gap distance of 2.5 mm. The steel sphere is attached
to the high voltage (HV) side of the high voltage transformer. An autotransformer

Fig. 3 Dielectric breakdown test kit [13, 14]

Introduction to Condition Monitoring of Electrical Systems 97

is employed at the primary side of the transformer to supply the demanded voltage.
230 V AC and 50 Hz frequency are given by the primary of the autotransformer. Here,
for the safety-purpose circuit breaker is connected between the autotransformer and
the supply voltage. Least break down voltage (BDV) value indicates that the moisture
content and conducting substances are existing in the oil [13, 14].
Testing procedures for BDV Test:
(a) Collect the oil sample of the transformer oil in proper sampling tube;
(b) Pour the oil sample into the oil pot of the machine;
(c) After these, keep the pot open for 1/2 min for settling air vapours;
(d) Then maintain two test electrodes with distances of 2.5 mm with surrounding
oil samples;
(e) After maintaining the distance, apply a voltage across the electrode sphere but
here maintain the voltage increasing rate at 2 kV/s. After applying voltages,
monitor oil specimens through the glass;
(f) Increase this voltage until the breakdown does not occur between the two elec-
trodes. When a breakdown occurs, then those voltages are denoted as breakdown
voltages;
(g) Breakdown means that arc ignited by the electrode due to high voltage is applied.
So after this process, the testing device turns off the system automatically;
(h) Then report this voltage value in the notebook [13, 14].

4.3 Du-Nouy Ring Method Test

This is another powerful condition monitoring technique which is used for the oil
quality analysis. This measurement is for the computation of interfacial tension (IFT)
between oil and water using the ring method according to the international stan-
dard under non-equilibrium conditions. The approach is primarily essential in the
petroleum industry as it can be used to predict oil purity. In this test, water is taken
in a glass container and oil is suspended above it. A freely suspended platinum ring
is inserted inside the oil and then the container is slowly moved upwards. When the
ring is at the interface of water and oil, it gets raised upwards from the water to the oil
interface due to interfacial tension. This movement of platinum ring gives a measure
of IFT [15–19].
Due to the density variation of these two liquids, oil floats on the water surface.
Then a tiny amount of force is demanded to withdraw the ring from the water surface
because water has a greater surface tension than the oil. This force is used to calculate
the IFT (mN/m) between the oil and water. The instrument provides the computation
to be performed entirely automatically. A high interfacial tension content (40 mN/m)
means the inadequacy of unwanted polar contaminants in the hydrocarbon fluid
which indicates that the fluid is immiscible with water. The force exerted on the ring
is formulated as follows [15–19].
98 A. J. Patil et al.

Fig. 4 Du Nouy ring

method [15]

F = 4nRf

where
R is the ring radius.
4nf is gross molecular force along side ring periphery.
The Du Nouy ring method experiment kit is shown in Fig. 4.

4.4 Dissolved Gas Analysis (DGA) Test

DGA analysis is applied for the oil-immersed transformer. It performs a valuable

role in monitoring the current health condition of the transformer. Whenever an oil-
immersed transformer suffers from abnormal condition or faults, a combination of
gases are evolved. Whenever any serious fault occurs in the transformer, to detect
these faults Buchholz relay is used. But sometimes due to overstress, moisture con-
tent, hot spot temperature in the winding, corona loss, arcing, etc. these things cannot
be assessed by the Buchholz relay. So DGA analysis helps here to analyse those things
properly [3–7, 20].
In the DGA analysis, first take the sample oil from the transformer and using a
magnetic stirrer, extract these gases from the oil. Traditionally, gas chromatography
(GC) was employed for the measurement of the concentration of the gases evolved.
In GC, the evolved gases were separated into multiple columns of an instrument
called spectrophotometer and by observing the colour changes in dyes on the walls
of these columns, the concentration of dissolved gases was determined [4, 5, 19–21].
But nowadays, GC-based DGA instruments are replaced by portable DGA kits
working on the principle of photoacoustic spectroscopy (PAS). As per the principle
of PA, whenever an intermittent beam of light is irradiated onto a substance, a portion
Introduction to Condition Monitoring of Electrical Systems 99

of the radiation is absorbed and the substance emits acoustic waves of the same fre-
quency as the light pulse frequency due to pressure changes by electronic transitions.
DGA kits working on the principle of PAS consist of the following components [3–7,
19–21]:
A. Photoacoustic Test Cell
PA spectrometer used for DGA contains a photoacoustic cell that is used to contain
the sample in a leak-free atmosphere. A sample holder is present inside the cell on
which light is striking through a window. To avoid any signal attenuation, cell volume
is minimized.
B. Light Source
A xenon lamp of a few hundred watts power is employed as the white-light source.
C. Microphone
When light is incident on the sample, due to PA effect pressure waves are generated.
These waves are often in the acoustic frequency region. These waves are detected by
using microphones. The sensitivity of these microphones is ~10 mV/Pa [4–8, 19].
D. PC-Based Data Acquisition System
Input data once amplified is taken by the computer and is processed for conversion
into useful information after processing by a PC-based data acquisition system.
The concentration of dissolved gases depends on the type of fault. After this
process faults, interpretation is done with the help of Key Gas method, Rogers’s
Ratio method and Dual Triangular method. If the transformer suffers overheating,
then the DGA sample shows the high gas contribution of carbon oxides (monoxide
and dioxide). In some cases, methane and ethylene are detected at tremendous levels.
If this condition occurs, then necessary action must be required to overcome from
this winding problem. Some part of the oil comes into contact with the core and
winding insulation material so this causes decomposition of the oil. Hence proper
internal inspection of the transformer is required, whether the internal condition of
winding and winding insulation is okay or not [1–5, 19].

4.5 SFRA Test

Sweep frequency response analysis (SFRA) is primarily used for the determination
of any displacement in windings of the transformer. Transformer windings have
inductive and capacitive elements whose behaviour is dependent upon the supply
frequency. When the transformer is freshly manufactured, all such elements have
finite modelled values; but due to transportation, short circuit and vibrations, these
electrical parameter gets modified. A plot of these RLC parameter variations versus
frequency helps here to assess the electrical and mechanical condition of the winding.
Winding deformation is also caused by lightning or system switching. Due to these
100 A. J. Patil et al.

Fig. 5 SFRA kit [22, 23]

circumstances, periodical monitoring of winding is important. The impedance allow-

able change of the winding after fault occurrence should be less than 2% [22–27]. A
typical SFRA kit is presented in Fig. 5.
In the SFRA test, lv signal is applied at the input terminal of a bushing and
the voltage at both the input and output terminals of a particular phase is measured.
During this process, all the other winding terminals are disconnected from each other
and are grounded. Figure 5 shows the typical method of connection of test cables
along with their ground extensions for SFRA measurement. Both the capacitive
and inductive elements of a winding are depended on its geometry, so whenever
the winding is exposed to displacement or short-circuit faults, a variation in the
resonance frequency occurs. Therefore, the observed frequency response can indicate
the specific condition of the winding [1–5, 22–29].
Basically, the SFRA test is a comparative measurement method. These compar-
isons can be primarily of three types, namely time-based comparison, type-based
comparison and phase-to-phase-based comparison. Table 3 mentions the different
frequency sub-bands and the corresponding faulty component along with its failure
sensitivity [22–29].

4.6 Vibration-Based Monitoring

There are various old techniques that are available to monitor the transformer winding
condition but the latest vibration analysis also proves to be a powerful method to
observe the condition of the winding. Transformer vibrations may be due to core
vibration, winding vibration or due to on-load tap-changer vibration which occurs
due to loosened core segments or coil segments, shield problem, or a damaged bearing
in the oil cooling pumps or fans. It is a lot useful to observe the vibration signal to
predict the present health and residual life of the transformer. Transformer have a
Introduction to Condition Monitoring of Electrical Systems 101

Table 3 SFRA frequency sub-band [26]

Sr. no. Frequency of the Component Failure sensitivity
Sub-band
1 <2 kHz Main core bulk winding Core deformation, open
inductance circuits, short turns and
residual magnetism
2 2–20 kHz Bulk component shunt Bulk winding movement
impedances between winding and
clamping structure
3 20–400 kHz Main winding Deformation within the
main or tap winding
4 400–2 MHz Main winding, tap Movement of the winding,
winding and internal leads ground impedance
variations

unique vibration signature and hence, for vibration analysis a baseline variation tests
are compared with subsequent vibration tests. Vibration signals can be used for the
prediction of present health and determination of residual life of the transformer.
Excessive vibration if ignored may result in mechanical damage to the winding and
core [2–4, 30–33]. A typical vibration system is presented as shown in Fig. 6.
An accelerometer is used to measure the vibrations on the outside of the tank
wall. Measured vibration response results expose an impressive similarity between
current and vibration harmonics. After observing vibration response, compare it
with standard data, and this will indicate that the condition of winding and any shift
may highlight abnormal operation which signifies further analysis and investigations
[3–7, 30–34].

Fig. 6 Vibration analysis kit [31, 33]

102 A. J. Patil et al.

4.7 Sound-Level Monitoring

An audible sound is generated by the transformer due to core deformation arising

from magnetostriction of the core and electromagnetic forces acting on the winding.
Two things can be measured related to sound, i.e. sound intensity and sound pres-
sure. Sound intensity measurement is generally preferred over sound pressure-based
measurement systems. Vibration amplitude depends not only on the lamination flux
density but also on the magnetic properties of the core. In this technique, a generated
sound response is used for monitoring the condition of the transformer. The force
generated is directly proportional to the square of the current flowing in the winding
and the audible sound power is proportional to the square of the vibrational amplitude
[2–6, 33, 34].
Transformer sound is measured using a sound detector and the data obtained
is compared with standard data. The resultant output obtained by this comparison
is used to determine the current health condition. When transformer noise level
enhances, overloading comes into the picture. Thus, to mitigate any severe damage
to the transformer, several measurement techniques have been proposed. These tech-
niques may follow different approaches but the general procedure remains the same
which is given as follows [33, 34]:
• Monitoring noise level when a new transformer is loaded.
• Monitoring change in noise level overtime when the transformer has a specified
loading.
• Comparing relative noise level with the noise in the background.

4.8 Infrared Temperature Monitoring

Infrared monitoring is used to detect hot spot temperature forms inside the trans-
former due to winding failures. It is used for thermal imaging of hot spot and non-
contact-type localized temperature measurement. The self-emitted radiation from
the winding and core is measured using this technique and a corresponding electrical
signal is obtained as output. So whenever heat is produced on the winding, core
or any other portion, then it is shown as a reddish colour on that thermal imager
measurement device [1–6, 35, 36].
This device is used to measure the temperature at various positions of the equip-
ment. It is a non-contacting and online measurement device that cannot physically
touch the surface of the equipment. For measurement of the temperature of the wind-
ing, a thermal imager is generally used. This thermal imager can identify thermal
stress on the device. After investigation of the thermal image of the transformer, the
recommended precautionary action can be decided [4–8, 35, 36].
Introduction to Condition Monitoring of Electrical Systems 103

4.9 Wireless Monitoring System

Wireless transformer monitoring systems are one of the widely accepted, yet nascent
condition monitoring techniques. This technique curbs the need for site visits for
monitoring. In this technique, four stages are primarily involved and these are
mentioned in Table 4.
A generalized layout of this wireless monitoring system is shown in Fig. 7.

5 Overhead Transmission Lines

Overhead transmission lines are carriers of electrical energy right from the electric
power generating station switchyard to electric load substations. These transmission

Table 4 Wireless Monitoring Working Information [37]

Sr. no. Parameter Working
1 Data collection Sensors like infrared scanner are used to monitor the condition
of the transformer
2 Signal conditioning Amplifier is used to enhance the signal strength. Then amplified
signal is sent to signal transmitter through microcontroller
3 Data transmission Transmitter sends the data to a remote monitoring station
4 Monitoring Monitoring is done by examining the parameters of the
transformer by using this system. And if any unwanted
circumstances come into the picture, then take necessary action
according to fault severity

Fig. 7 Wireless monitoring system [37, 38]

104 A. J. Patil et al.

lines are made of conductors having a uniform cross-section and are circumvented
by air which acts as a dielectric.
Transmission line failure occurs mainly due to insulator failure, conductor cor-
rosion and mechanical damages. These failures may lead to repeated power outages
and unscheduled maintenance of the transmission line. To reduce the outage time
and maintenance cost and to enhance reliability and safety, electrical utilities are
tilting towards online condition monitoring techniques. Since a variety of condition
monitoring techniques are present for various failures, covering all of them is beyond
the purview of this book. The widely accepted condition monitoring techniques are
covered in this section.

5.1 Failure Modes of an Overhead Transmission Line

A brief description of the primary cause and effect of the failure modes of a
transmission line is given as follows.

5.1.1 Insulator Breakdown/Failure

Overhead transmission line primarily employs Pin, Suspension, Strain, Stay and
Shackle-type insulators depending upon the application and operational voltage.
These insulators restrict the flow of electrical energy by disallowing the free move-
ment of internal electric charges. But the electric field becomes capable of tearing
electrons away from atoms and employs them for conduction under the influence
of sufficiently high voltages [39]. This leads to insulator breakdown. This insulator
breakdown may occur due to the following:
i. Thermal cycling,
ii. Mechanical cycling,
iii. Abrasion (or removal of metal on conductor surface) from weathering,
iv. Electrothermal factors,
v. Torsional factors,
vi. Ionic motion under high voltages and
vii. Corrosion.
Insulator breakdown may lead to the declination of mechanical factors of safety
and flashover. The safe design of an insulator string must have a dry flashover of at
least three times the nominal operating voltage. Also, it must have a leakage path
of width at least two times the striking distance. The indicators of a forthcoming
flashover are increased by generation of corona and electrical discharges in the gap.
Introduction to Condition Monitoring of Electrical Systems 105

5.1.2 Conductor Corrosion

Aluminium conductor steel reinforced (ACSR) conductors are extensively used in

overhead transmission lines. The corrosion process in these conductors initiates with
strand breakage due to corrosion and results in line failure. The chemical process
of corrosion starts with the penetration attacks by an aqueous solution across the
aluminium strands. This solution contains chloride ions. If these attacks are success-
ful in removing the galvanized coating on the steel core, the steel substrate (iron) is
revealed. Galvanic cell is formed between the aluminium and iron due to this aqueous
solution [39].
Aluminium starts acting as an anode and corrodes to give aluminium hydroxide
as one of its products. The current-carrying capability of a line is decided by the
number of conducting strands. As the aluminium strands gets corroded the current
carrying capability and mechanical strength of the transmission line is significant
reduced.

5.1.3 Mechanical Damage

Mechanical damage to transmission lines occurs mainly due to Aeolian vibration

and icing. Aeolian vibrations occur due to wind-induced vortices on the conductor
which exerts a force on the conductor. These vibrations eventually build up and lock
into the resonant frequency (10–30 Hz). Depending on the type and intensity of
the vibrations, alternating bending stresses of different magnitude and duration are
applied to the conductor [39].
Severe stress reversals lead to strand breakages. Also, the sub-conductor oscilla-
tions may result in fretting between aluminium strands in the vicinity of the fittings.
Fretting failure is defined as the failure due to a slight periodic relative motion of two
surfaces in contact. The oscillations lead to flexing of the aluminium strands which
ultimately leads to fatigue failure.
Another form of mechanical damage is the damage due to the formation of an
aerofoil due to the accumulation of ice on the conductor. The lines begin to gallop and
the conductors of different phases start moving closer to each other which results in a
flashover. This results in localized damage of conductors at isolated points. Manually,
it is difficult to locate these isolated points.

5.2 Condition Monitoring of Transmission Line

Due to its localized nature, traditional detection techniques like visual inspection,
infrared inspection, ultrasonic detection and partial discharge detector cannot be used
to develop a remotely operated control station-based automatic means of instan-
taneous condition monitoring of the entire transmission line, viz., spanning over
hundreds of kilometres.
106 A. J. Patil et al.

These techniques involve the presence of humans constantly in the vicinity of

the line which is not possible at all times over the entire transmission line. The
major remote condition monitoring techniques which are employed in towers can be
classified as follows.

5.2.1 Insulator Condition Monitoring

Multiple Insulator condition monitoring automation techniques are presently being

used. Covering all of them in detail is beyond the purview of this book. Hence, one
of the widely accepted techniques is covered here. In this technique, a combination
of Machine learning techniques (MLTs) and Wavelet transform are used for insulator
condition monitoring [39, 40].
In this technique, the streaming video of the insulators is first acquired by either
surface vehicular patrolling (SVP) or by using remote terminal unit (RTU). The
insulators are then classified into three states, namely good, marginal or risky and
their corresponding feature vectors are extracted using Wavelet MRA. These feature
vectors are then used as training inputs for employing MLTs. These MLTs can classify
a new insulator image into any of the three aforementioned states [40]. A typical
layout is as shown in Fig. 8.
In order to determine the insulator’s health, fuzzy inference system (FIS) is
employed as the MLT. The degree of damage of the insulator in marginal or risky
states cannot be assessed by a deterministic mathematical model. But fuzzy sets the-
ory can be used to derive the correct result by formulating set of rules and membership
functions [39, 40].

Fig. 8 Schematic of power system insulator detection and condition monitoring automation system
[40]
Introduction to Condition Monitoring of Electrical Systems 107

5.2.2 Conductor Corrosion Monitoring

Multiple conductor corrosion monitoring techniques are presently being used. Cov-
ering all of them in detail is beyond the purview of this book. Hence, one of the widely
accepted techniques is covered here. In this technique, a non-destructive approach
is used in which an eddy current test instrument is employed. In this instrument, a
high-frequency current is applied to its field coil. This current generates a magnetic
field which in turn produces swirling eddy currents in the transmission line conduc-
tors. A transducer is used to detect the flow of these eddy currents. Corrosion leads
to surface defects in the various conductor strands. These defects decrease the flow
of eddy current and hence impact the impedance of the conductor [39, 40, 46].
One of the main advantages of this method is its capability to determine the loss
of aluminium cross-section along with the determination of the extent of conductor
corrosion. Common OHLCD and Fujikura’s detector are two common examples of
eddy current based corrosion monitoring instruments [39, 40].

5.2.3 Wind Monitoring

A transmission tower installed in coastal areas is prone to damage by hurricanes or

high-speed seasonal winds. Critical damages to the tower structure due to winds can
be detected using SHM in offshore areas but in coastal areas in order to enhance
reliability, dedicated wind monitoring systems are employed. Also, monitoring of
wind-induced vibrations is of prime importance for power system security [61].
A variety of solutions have been proposed for monitoring the effects of wind
but nowadays, Optical fibre based sensing systems are gaining attention. Since these
systems have relatively lower cost and due to its packaged structure, it can prove to be
a safe, robust and reliable monitoring system. Specifically, tilted fibre Bragg grating
(TFBG) based sensors which are a nascent species in the family of optical fibre based
sensor are employed for this purpose. It works on Bragg grating technology and it
allows single-point sensing [61].

5.2.4 Icing Monitoring

Transmission line icing often leads to accidents like unwanted tripping operations,
breaking of conductors, conductor galloping and insulator string flashover. Ice accre-
tions that are commonly observed on transmission lines can be subdivided into Glaze,
Hard Rime, Soft Rime, Hoar Frost and Snow and Sleet. Among these ice accretions,
glaze is the most dangerous one for transmission lines. The system used to effec-
tively monitor the icing on transmission line must take the icing information and
by real-time monitoring of the temperature and other parameters, the amount of ice
coating is determined. A typical icing monitoring unit consists of an array of sensors
for determining tension, deflection angle, wind speed, rainfalland sunshine intensity
[41–45].
108 A. J. Patil et al.

These sensors collect the necessary information and help in determining the con-
dition of icing in a region so that proper de-icing action can be done. Thickness,
density and type of icing can be obtained by using a combination of these sensors
and image processing techniques. Icing information of three or more than three base
towers is generally needed to develop the mechanical model of computing icing
coating and calculating the icing thickness accurately [41–45].

5.2.5 Other Solutions for Transmission Line CM

Recently, mobile robots are employed for line monitoring. Typically, these robots
use principles of line-suspended robotics to move along the overhead line while
performing line condition monitoring. These robots can perform various functions
like the condition assessment of ACSR conductors, detection of broken wires and
measurement of the cross-section of conductors [46].
Apart from line-suspended robots, aerial overhead conductor monitoring equip-
ment by an aerial vehicle has been used to monitor overhead lines. Unmanned Aerial
Vehicles (UAVs) allow visual inspection of the line using different types of cameras.
But these UAVs cannot be used widely as they have certain limitations/risks [46].
These limitations/risks are as follows:
i. Risk of UAV crash on live overhead lines,
ii. Limitations in flying UAVs imposed by aviation agencies,
iii. Fast-changing background and
iv. Blurring of the images due to hovering.

6 Condition Monitoring of Towers

Electric power transmission towers are mechanical structures that are used to tote
overhead lines right from the electric power generating station switchyard to electric
load substations while sustaining all kinds of natural calamities and induced distur-
bances. The shape, height and sturdiness of a tower are decided primarily on the basis
of geographical profile, power handling level and right-of-way (RoW) available with
an electrical utility.
Failure of a tower in a power corridor can be structural, foundational or electrical.
These failures are normally cascading in nature which means that adjacent towers
also get affected. Repairing damaged towers is a costly affair even if the cost of
litigation and revenue loss due to power disruption are kept aside. In this section,
the various failure modes of an electrical power transmission tower along with its
corresponding remote condition monitoring (CM) techniques are discussed.
Introduction to Condition Monitoring of Electrical Systems 109

6.1 Failure Modes of an Electrical Power Transmission

Tower

As mentioned earlier, the failure modes of a tower in a power corridor can be classified
as structural, foundational or electrical. A brief description of the primary cause and
effect of these failures is given below.

6.1.1 Structural Failures

Structural deformations in the electric power transmission tower structure can be due
to degraded structural connections. This degradation occurs mainly due to corrosion
in towers made of steel. Tower legs are set in concrete. The presence of any defects
or cracks in the concrete allows water and salts to seep through concrete and causes
corrosion of the tower leg [43, 50–54].
Due to corroded parts in the tower, the ability of the tower to withstand external
disturbances like the effect of wind and icing reduces significantly and may lead to
deformation or rotation of the tower structure [43, 50–54].

6.1.2 Foundational Failures

Foundational failures in the electric power transmission tower structure refer to vari-
ation in the tower settlement due to unequal compressive forces acting on the foun-
dation of the tower. Technically, tower settlement is defined as the total vertical
displacement at the foundation level of the tower structure due to the reduction of
volume air void ratio of the soil [43, 51–54].
The variation in tower settlement leads to dislocation of the bottom of the tower,
leading to displacement, inclination, cracking and subsidence of the tower. Also, it
leads to variation in the sag and stress acting on the transmission lines [43, 52, 53].

6.1.3 Electrical Failures

Electrical failures in the electric power transmission tower lead to the flow of leakage
current. Leakage current in the tower structure is defined as the current flowing across
the overhead insulator during over voltages into the ground. Electrical failures occur
mainly due to malfunctioning of these overhead insulators mounted on the tower
[43, 50–54].
Flow of small amounts of leakage currents is permissible but if this leakage current
is higher than a threshold, it can be hazardous for humans or animals in its vicinity.
Due to the increasing risk of loss of lives, tower designers have expressed concern
about possible hazards associated with these currents.
110 A. J. Patil et al.

Table 5 Summarize the various tower failures [43, 50–55]

Basis of difference Structural failure Foundational failure Electrical failure
Primary cause Corrosion Unequal compressive Insulator
forces malfunctioning
Effect on the tower Deformation and Displacement, Electrical shocks
structure rotation inclination, cracking
and subsidence
CM Technique Structural health Tower settlement Leakage current
monitoring (SHM) monitoring monitoring

Thus, in order to mitigate any permanent damage to the tower structure pertaining
to these failures, a dedicated tower maintenance strategy must be devised. But before
moving onto the various condition monitoring techniques, let us first summarize the
various tower failures along with its causes, effects and corresponding condition
monitoring techniques in Table 5 [43, 50–54].

6.2 Condition Monitoring Techniques of Tower

Any tower maintenance strategy involves routine recurring work of repairing and
replacement of damaged parts that keep the tower in such a condition that it can be
continuously used, at its designed capacity for its intended purpose. Traditionally,
health deterioration of tower structure due to corrosion is kept in check by visual
inspection of tower structure and painting or replacing the corroded part in the tower.
But, this process of visual inspection of towers is labour intensive and cumbersome.
In some areas, tower top patrolling practices using UAVs having ultra high defi-
nition cameras are used to identify corroded parts in the tower structure. But, UAVs
with top tower patrolling capabilities have certain limitations and cannot be accepted
as an all aspect solution to all the tower failures. These limitations are as follows:
• Absence of real-time condition monitoring and can only monitor towers periodi-
cally,
• Human operator is constantly in the loop, and
• UAVs themselves need regular maintenance and refuelling.
Remote condition monitoring techniques are employed widely for all-aspect
predictive maintenance of the tower. These techniques are discussed as follows.

6.2.1 Structural Health Monitoring (SHM)

SHM is used for the prognosis of structural deformations in the electric power trans-
mission tower structure. These deformations can be due to degraded structural con-
nections or bearing wear or a fatigue crack. In Lattice Steel Towers (LSTs), tower
Introduction to Condition Monitoring of Electrical Systems 111

Fig. 9 Overall SHM system [55]

legs are set in concrete. The presence of any defects or cracks in the concrete allows
water and salts to seep through concrete and causes corrosion of the tower leg. This
eventually leads to deformation of tower structure over a period [55–59].
Individual SHM units are mounted on all the towers in a power system network.
The electrical output of multiple units of these sensors is fed to a central module.
This module, after conditioning and converting the analog electrical output into
a digital signal, processes the signal using a microprocessor. For establishing an
interconnection between these modules for obtaining structural integrity information
of individual towers, generally CAN (Controlled Area Network) protocol is used.
The information provided by these modules is directed to a computer for taking
decisions. A generalized layout of an SHM-based health prognosis technique is as
shown in Fig. 9 [54–59].
The SHM technique employs an array of sensors for the prognosis of these defor-
mations. Brief information on these sensors along with their respective sampling per
second (SPS) rates is mentioned in Table 6.
A generalized layout of an SHM-based health prognosis technique is shown in
Fig. 9.

6.2.2 Tower Settlement Monitoring

Tower settlement is the dislocation of the bottom of the tower due to unequal com-
pressive forces acting on the foundation of the tower. This may lead to greatly varying
the sag and stress on the transmission lines and displacement, inclination, cracking
and subsidence of the tower. This deteriorates the capacity of towers to withstand
112 A. J. Patil et al.

Table 6 Sensors in a typical SHM unit [54–56]

Sensor Type Function SPS Location
Wilcoxon sensor Accelerometer Detect rotations 20 Two sensors at the
centre and three at
top
Extensometer Deformation sensor Detect structure 20 Four sensors at the
deformations bottom
LM 335 Temperature sensor Detect temperature 1 One sensor at the
variations centre
Anemometer Wind speed and Monitor wind speed 1 One sensor at top
direction sensor

Table 7 Components of vibration-based condition monitoring system [50]

Component Function
Acceleration sensor Collection of acceleration data when the tower is
vibrating
Excitation device Generate impulse excitation
Condition monitoring device (CMD) To monitor the acceleration of the tower when it is
vibrating
Communication module Mode of wireless communication (using Bluetooth and
4G wireless module)
Prognosis centre Prognosis of damage due to settlement variations

disturbances like the effect of wind and icing. Thus, to improve the reliability of tower
structure, the prognosis of variations in settlement of the tower is vital [43, 50–52].
Vibration-based condition monitoring systems are widely used to monitor tower
settlement. A typical vibration-based condition monitoring system can be subdivided
into five major components. These components along with their respective functions
are mentioned in Table 7.
A typical layout of vibration-based tower settlement is shown in Fig. 10.

6.2.3 Leakage Current Monitoring

During a lightning strike, currents with large magnitudes and a fast rise time start
flowing. Lightning strikes on power transmission towers induce high voltages in it
which eventually leads to insulation failure. Earthing of each tower of an electrical
transmission line is done and footing resistance of each tower is measured in dry
season before stringing the earth wire. The footing resistance of the tower should
never exceed 10 ohms. Higher the footing resistance, higher the chances of failure due
to lightning strikes. Thus, to lower the footing resistance on electrical towers either
pipe or counterpoise earthing is used for earthing of the electrical towers [53–55]. A
simple block diagram for leakage current monitoring is shown in Fig. 11.
Introduction to Condition Monitoring of Electrical Systems 113

Fig. 10 Typical layout of vibration-based tower settlement [50]

Fig. 11 Leakage current monitoring [53]

In the case of a pipe earthing system, a galvanized steel pipe (diameter 26 mm

and length 3 m) is buried vertically into the soil such that the top of the pipe is 1 m
below the ground level. Surroundings of the pipe are filled with layers of charcoal and
salt, which maintain the surrounding soil of the pipe moist. In case of counterpoise
earthing, galvanized wire (10.97 mm diameter, 25 m minimum length) is used. The
114 A. J. Patil et al.

galvanized wire is connected to the leg of the tower by a galvanized lug. The wire is
buried tangentially under the ground of minimum 1 m depth [53–55].
Apart from lightning strikes, it is also well known that ground faults on trans-
mission lines are the most recurring short-circuit faults. The impact of this fault is
cascading in nature due to which not only the tower on which fault has occurred but
also towers in the vicinity of the faulty tower acquire high potentials. If this voltage is
higher than a threshold, it can be hazardous for humans or animals in its vicinity. Due
to the increasing risk of loss of lives, tower designers have expressed concern about
possible hazards associated with these ground faults. Leakage current monitoring
is used for health prognosis of earthing of the tower. Information from voltage and
leakage current sensors placed on the tower is sent using a communication module
to a control station where the information is analysed, and the necessary actions are
taken accordingly [53–55].

7 Condition Monitoring of Relay

Protective relays are electrically operated switches which are designed to give a
trip signal to a circuit breaker during faults. The first protective relays employed
electromagnets for fault detection but these electromechanical relays were bulky,
energy-inefficient and provided only rudimentary indication of a fault. With the
advent of new technologies, relays are getting compact in size and are becoming
more energy efficient [62].
Types of relays, its construction and operating principle are beyond the purview
of this book and the sole focus will be on various monitoring techniques for relays.

7.1 Problems Pertaining to Testing of Protective Relays

Electromagnetic relays have a limited mechanical life due to the presence of moving
mechanical components. On the other hand, solid-state relays are having no moving
parts due to which they do not have limitations pertaining to the life of mechanical
components. But these solid-state relays must strictly adhere to operations within its
ratings provided by the manufacturer [63–65]. The various problems associated with
testing relays can be summarized in the following.

7.1.1 Problems Due to Loading Conditions

Operating lifetime of a relay is mentioned in the context of purely resistive loads.

This is due to the fact that the operating conditions for a purely resistive load can
be recurrently reproduced for testing. Thus, for defining the operational lifetime, a
power supply with a resistor connected in series is used for loading. A series resistor
Introduction to Condition Monitoring of Electrical Systems 115

is used for limiting the hot switch current. But, in practical applications, relays are
not operated with purely resistive loads. This means there is always an ambiguity in
determining the actual operational life of a relay [63–65].

7.1.2 Problems with Normal Life Testing

Normal life testing for assessment of the mechanical life of a relay is a sluggish
process and the time taken is so large that it makes the process impractical. This
can be understood by the following illustration: consider a relay that is designed to
perform a billion operations before failure. For normal life testing, even if the relay
is performing one operation per second, it takes—300,000 h for testing if the relay
operates for its designed lifetime. Thus, normal life testing for relays is practically
difficult [63–65].

7.1.3 Problems with Accelerated Life Testing

To reduce the time of testing of normal life testing routine, accelerated life testing was
proposed. Accelerating the rate of operation of a relay sharply reduces the time taken
for lifetime assessment of a relay. Accelerated life testing is useful for comparative
analysis of relays that use similar design technology and optimally decide the best
design. But in the context of lifetime assessment, test results of accelerated life
testing are pessimistic as the mechanical parts of the relay are not allowed to properly
dissipate the heat generated due to friction [63–65].

7.2 Tests by Type of Protective Relay

IS 2071-1962 has covered related standards for testing all classes of relays. In this
section, the tests performed on various types of relays are discussed in brief.

7.2.1 Directional Overcurrent Relay

Tests for a directional overcurrent relay are split into two parts, namely tests for
overcurrent and tests for a directional unit. The following tests are performed on a
typical directional overcurrent relay:
• Test for Minimum pickup,
• Test for Contact gap,
• Test for Clutch pressure and
• Test for Angle of maximum torque [66].
116 A. J. Patil et al.

7.2.2 Distance Relay

Tests for the four parameters mentioned in the case of directional overcurrent relays
are also performed in case of distance relays.

7.2.3 Differential Relay

Differential relays are inordinately sensitive devices and require elusive considera-
tions. The previous history is disregarded for any comparison in case of differential
relays and a maintenance test is needed to be properly performed. The two tests
primarily performed in case of differential relays are
• Test for minimum pickup and
• Checking differential slope and harmonic restraint [66].

7.2.4 Static Relay

The tests performed in case of testing of static relays are

• Impulse voltage withstand test,
• High frequency disturbance test,
• Damp heat test,
• Impact or shock test and
• Vibration test [66]
Apart from the tests, periodic inspection of protective relays is also performed for
proper maintenance of relays which are as follows.

7.3 Periodic Inspection of Protective Relays

Periodic inspection of protective relays involves the following checks:

• Relay connections are checked to be airtight. If loose connections are found, it is
an indicator of excessive vibrations.
• Relay must be checked for the absence of foreign materials like dust in relay gaps
as these foreign materials may deteriorate the relay’s operational reliability.
• Measurement of movable gaps is measured and compared with previous results.
If there is a significant variation in the length of the gap, it is considered to be an
indicator of excessive wear and tear.
• Measurement for alignment and wipe of all contacts [67].
Introduction to Condition Monitoring of Electrical Systems 117

8 Conclusion

This chapter summarizes the condition monitoring tactics that are employed for
health prognosis of an electrical system. Electrical system is modelled as an assem-
blage of equipment like transformer, relays, transmission lines and towers. The
chapter starts with defining condition monitoring in the context of electrical sys-
tems, highlighting its necessity and then moving on to providing a brief insight into
equipment-specific condition monitoring tactics. The cause of failure is identified by
assessing deviations in the parameters being monitored which allows a maintenance
engineer to optimize a maintenance strategy for the entire system. Finally, it can
be concluded that the key objectives of condition monitoring techniques are dam-
age assessment, maintenance cost reduction, equipment life enhancement, operator
safety enhancement, accident minimization and improvement of power reliability
and above all proper asset management.

References

1. S. Tenbohlen, F. Figel, On-line condition monitoring of power transformers, in IEEE Power

Engineering Society Winter Meeting, Vol. 3 (2000), pp. 2211–2216
2. P.A. Reddy, B.S. Rajpurohit, On-line monitoring of winding parameters for single-phase trans-
formers, in 2014 IEEE 6th India International Conference on Power Electronics (2014),
pp. 1–4
3. C. Roy, P. Suresh Babu, Condition monitoring approaches of a transformer. J. Electr. Eng.
204–215 (2017)
4. X. Zhang, E. Gockenbach, Asset-management of transformers based on condition monitoring
and standard diagnosis. IEEE Electr. Insul. Mag. 24(4), 26–40 (2008)
5. Y. Han, Y.H. Song, Condition monitoring techniques for electrical equipment—a literature
survey. IEEE Trans. Power Delivery 18, 4–13 (2003)
6. M. Schaefer, R. Kutzner, K. Feser, Condition monitoring system for power transformers, in
IEEE Conference on Power System Technology, Vol 3, pp. 1701–1705 (2000)
7. T. Leibfried, Online monitors keep transformers in service. IEEE Comput. Appl. Power 11,
36–42 (1998)
8. R. Ahmad, An overview of time-based and condition-based maintenance in industrial
application. Comput. Ind. Eng. 63(1), 135–149 (2012)
9. L. Li, X. Longjun, Y. Zhang Deng, G. Yafeng Bin, L. Fuchang, Condition assessment of power
transformers using a synthetic analysis method based on association rule and variable weight
coefficients. IEEE Trans. Dielectr. Electr. Insul. 20(6), 2052–2060 (2013)
10. A. Abu-Siada, S. Islam, A novel online technique to detect power transformer winding faults.
IEEE Trans. Power Delivery, 849–857 (2012)
11. Article Available, https://www.spectrosci.com/resource-center/lubrication-analysis/literature/
e-guides/guide-to-measuring-water-in-oil/
12. Article available, http://hpst.cz/sites/default/files/attachments/5991-0672en-voda-v-
turninovem-oleji.pdf
13. D. Jasper, D. Sarkar, N.K. Roy, Development of an automated break down voltage test set.
IEEE CATCON, 279–282 (2017)
14. Article available at http://www.electricalidea.com/testing-transformer-oil/
15. Article available at https://www.barnardhealth.us/food-emulsions/table-51.html
118 A. J. Patil et al.

16. ASTM, Standard Test Method for Interfacial Tension of Oil Against Water by the Ring Method,
D971–12 (2012)
17. Julian McClements, Interfacial properties and their characterization, Contemporary Food
Science (2004)
18. R.A. Ghunem, K. Assaleh, A.H. El-Hag, Artificial neural networks with stepwise regression
for predicting transformer oil furan content. IEEE Trans. Dielectr. Electr. Insul. 19, 414–420
(2012)
19. N.A. Bakar, A. Abu-Siada, S. Islam, A review of dissolved gas analysis measurement and
interpretation technique. IEEE Electr. Insulat. Mag. 30 (2014)
20. J.C. Dennison, on M. Trout, Transformer oil DGA monitoring technology study 2015, in
IEEE/PES Transmission and Distribution Conference and Exposition (T&D), pp. 1–5 (2016)
21. S. Koroglu, A case study on fault detection in power transformers using dissolved gas analysis
and electrical test methods. J. Electr. Syst. 12(3), 442–459 (2016)
22. A. Priti Gite, S. Sindekar, Investigating mechanical integrity in power transformer using sweep
frequency response analysis (Sfra), in IEEE International Conference on Electrical (2017),
pp. 1–6
23. K. Ludwikowski, K. Siodla, Investigation of transformer model winding deformation using
sweep frequency response analysis. IEEE Trans. Dielectr. Electr. Insul. 19(6), 1957–1961
(2012)
24. M. Brandt, M. Gutten, T. Koltunowicz, P. Zukowski, Analysis of winding fault in electric
machines by frequency method, IEEE Transaction on 2018 Elektro (2018), pp. 1–4
25. H.-T. Wu, C.-Q. Jiao, X. Cui, X.-F. Liu, J.-F. Ji, Transient electromagnetic disturbance induced
on the ports of intelligent component of electronic instrument transformer due to switching
operations in 500 kV GIS substations. IEEE Access 5104–5112 (2017)
26. G.M. Kennedy, A.J. McGrail, J.A. Lapworth, Transformer sweep frequency response analysis
(SFRA). Article Published in October 2007
27. M. Bagheri, M. Salay Naderi, T. Blackburn, Advanced transformer winding deformation diag-
nosis: moving from off-line to on-line. IEEE Trans. Dielectr. Electr. Insul. 19, 1860–1870
(2012)
28. M.A. Sathya, A.J. Thomas, S. Usa, Prediction of transformer winding displacement from
frequency response characteristics, in 2013 IEEE 1st International Conference on Condition
Assessment Techniques in Electrical Systems (CATCON) (2013), pp. 303–307
29. M. Bagheri, B. Phung, T. Blackburn, Transformer frequency response analysis: mathematical
and practical approach to interpret mid-frequency oscillations. IEEE Trans. Dielectr. Electr.
Insulat. 20(6) (2013)
30. M. Nafar, B. Bahmanifirouzi, M. Jabbari, Transformer monitoring by using vibration analysis.
Australian J. Basic Appl. Sci. 5(11), 984–990 (2011)
31. B.S. Munir, J.J. Smit, Evaluation of various transformations to extract characteristic param-
eters from vibration signal monitoring of power transformer, in 2011 Electrical Insulation
Conference (EIC) (2011), pp. 289–293
32. R. Rinaldi, Diagnosing winding and core condition of power transformer by vibration signal
analysis, in 2012 IEEE International Conference on Condition Monitoring and Diagnosis
(2012), pp. 429–432
33. L. Bowei, M. Hai, Z. Lixing, On-line monitoring of transformer vibration and noise based on
DC magnetic bias, in 2013 Fourth International Conference on Intelligent Systems Design and
Engineering Applications, pp. 412–416, November 2013
34. J. Karel, F. Monhart, Noise measurement diagnostics for large electric machines, in IEEE
International Conference on Diagnostics in Electrical Engineering (Diagnostika), pp. 1–4
(2018)
35. S. Goel, A. Akula, R. Ghosh, B.S. Surjan, Condition monitoring of transformer using oil and
winding temperature analysis, in IEEE Conference on Electrical, pp. 496–500 (2016)
36. S. Anoop, K. Ilango, A. Dhieep, C. Thomas, J. Jose, A. Kumar, P.R. Ajith, Thermal stress
monitoring and pre-fault detection system in power transformers using fibre optic technol-
ogy, in 2017 International Conference on Intelligent Computing, Instrumentation and Control
Technologies (ICICICT) (2017), pp. 886–891
Introduction to Condition Monitoring of Electrical Systems 119

37. P. Mercy, N. Uma Maheswari, S. Deepika Devi, V. Dhamodharan, Wireless protection and
monitoring of power transformer using PIC, IJCSMC, 4(3), 0634–640 (2015)
38. T.K. Roy, T.K. Roy, Implementation of IoT: smart maintenance for distribution transformer
using MQTT, in 2018 International Conference on Computer, Communication, Chemical,
Material and Electronic Engineering (IC4ME2) (2018), pp. 1–4
39. R.K. Aggarwal, A.T. Johns, J.A.S.B Jayasinghe, W. Su, An overview of the condition
monitoring of overhead lines. Electr. Power Syst. Res. (2000)
40. V.S. Murthy, D.K. Mohanta, S. Gupta, Power system insulator condition monitoring automation
using mean shift tracker-FIS combined approach. Int. J. Comput. Aided Eng. Technol. (2013)
41. S. Wang, X. Jiang, Progress in research on ice accretion on overhead transmission lines and its
influence on mechanical and insulating performance, in Springer Link Conference on Frontiers
of Electrical And Electronic Engineering (2012), pp. 326–336
42. X. Huang, X. Wei, A new on-line monitoring technology of transmission line conductor icing, in
2012 IEEE International Conference on Condition Monitoring And Diagnosis (2012), pp. 581–
585
43. W.Q. Jiang, H. Li, Z. Wang, On-line monitoring of transmission lines in mining cave-ins area,
in 2012 Asia-Pacific Power and Energy Engineering Conference, pp. 1–4 (2012)
44. K. Yu, L. Xie, Y. Liu, X. Zeng, Y. Gu, An online de-icing method based on single-phase
grounding fault current for distribution lines, in 2018 International Conference on Power
System Technology (POWERCON), pp. 3796–3800 (2018)
45. Y. Zhou, S. Niu, J. Lü, L. Zhou, Meteorological conditions of ice accretion based on real-time
observations of high voltage transmission line. Springer Link Conference on Chinese Science
Bulletin 57, 812–818 (2012)
46. C.F. Barbosa, F.E. Nallin, Corrosion detection robot for energized power lines, in Proceedings
of the 2014 3rd International Conference On Applied Robotics for the Power Industry (2014)
47. K.J. Stevens et al. Conductor Damage Inspection System For Overhead ACSR Power Cables
Cdis On Acsr”, In 2013 Seventh International Conference On Sensing Technology (Icst), 2013
48. V.T. Morgan, Effect of elevated temperature operation on the tensile strength of overhead
conductors. IEEE Trans. Power Delivery 11(1), 345–352 (1996)
49. Meng Zhang, Guifeng Zhao, Jie Li. Nonlinear Dynamic Analysis of High-Voltage Overhead
Transmission Lines, Shock and Vibration, Volume 2018, 35 pages 2018
50. Xinbo Huang, Yu Zhao, Long Zhao, A Vibration-Based Monitoring System For Transmission
Tower Settlement”, 2018 Condition Monitoring And Diagnosis (Cmd), Pp. 1–4, 2018
51. Tower Design, Document Available https://Powerline.Net.In/2017/10/03/Tower-Design/
52. Wang Zhang-Qi, Jiang Wen-Qiang, The Influence Of Tower Foundation Settlement On The Sag
And Stress Of Overhead Lines”, 2009 International Conference On Energy And Environment
Technology, pp. 340–343, 2009
53. N. Harid, A.C. Bogias, H. Griffiths, S. Robson, A. Haddad, A Wireless System For Monitoring
Leakage Current In Electrical Substation Equipment. IEEE Access 4, 2965–2975 (2016)
54. Jiang Xiang-Dong, Tang Yu-Liang, Lei Ying, “Wireless Sensor Networks In Structural
Health Monitoring Based On Zigbee Technology,” IEEE International Conference On Anti-
Counterfeiting, Security, And Identification In Communication (Asid), Pp. 449 – 452,
2009
55. Ataide R., Castro A., Sousa M., De Souza R.M., Klautau A., “Sensor Network For Structural
Health Monitoring: Online Monitoring Of Transmission Line Towers”, Conference On Xi
Microelectronics Students Forum, Dec 2011
56. G. Zhan-Feng, D. Yan-Liang, S. Mu-Biao, C. Biaoping, “Network Sensor And Its Applica-
tion In Structure Health Monitoring System”. First International Conference On Innovative
Computing, Information And Control 1, 68–71 (2006)
57. F. Necati Catbas, Melih Susoy, Dan M.Frangopo, “Structural Health Monitoring And Reliabil-
ity Estimation: Long Span Truss Bridge Application With Environmental Monitoring Data”,
Volume 30, Issue 9, Pp. 2347–2359, September 2008
58. Yu-liang, Tang, Luo Yu, Huang Lian-fen, Guo Jian, and Lei Ying. “Wireless sensor network for
on-line structural health monitoring”, 2012 7th International Conference on Computer Science
& Education (ICCSE), pp. 386–389, 2012
120 A. J. Patil et al.

59. Alex Myers, Md Anam Mahmud, Ahmed Abdelgawad, Kumar Yelamarthi. “Toward integrat-
ing Structural Health Monitoring with Internet of Things (IoT)”, 2016 IEEE International
Conference on Electro Information Technology (EIT), pp. 19–21, 2016
60. Yinggang Nan, Wenping Xie, Li Min, Shunshuo Cai, Xiaoyong Chen, Tuan Guo, “Optical
Fiber Sensing System For Online Monitoring Wind-Induced Vibration On Power Transmission
Tower Survey”, 2018 Asia Communications And Photonics Conference (Acp), pp.. 1–4, 2018
61. Yasui, H., Marukawa, H., Momomura, Y., Ohkuma, T., “Analytical Study On Wind-Induced
Vibration Of Power Transmission Towers”, Elsevier Conference On Journal Of Wind
Engineering And Industrial Aerodynamics, Pp. 431–441, 1999
62. F. Xiao, Z. Zhang, X. Yin, J. Ji, G. Chen, Study of maintenance strategy of relay protection sys-
tem based on condition Monitoring, in 2015 50th International Universities Power Engineering
Conference (UPEC) (2015)
63. Is 3231–1-1 (1986): Electrical Relays For Power System Protection, Part 1: General
Requirements, Section 1: Contact Performance [Etd 35: Power Systems Relays]
64. J.J. Kumm, M.S. Weber, D. Hou, E.O. Schweitzer, “Predicting The Optimum Routine Test
Interval For Protective Relays” IEEE Transactions on Power Delivery, Volume: 10, Issue: 2,
Apr 1995
65. D. Velaga Sreerama Murthy, Mohanta, Power system insulator condition monitoring automa-
tion using mean shift tracker-Fis combined approach. Int. J. Comput. Aided Eng. Technol. 5(1)
(2013)
66. R. Cimadevilla, I. Garcia, Improvements in the operation of a distance relay during resistive
faults, in 2014 67th Annual Conference for Protective Relay Engineers (2014), pp. 132–135
67. Y. Zheng, J. Cai, Z. Zhou, Research review and application prospect of secondary equipment
condition monitoring, in 2016 International Conference on Smart Grid and Clean Energy
Technologies (ICSGCE) (2016), pp. 290–295
Condition Monitoring, and Fault
Detection and Diagnostics of Wind
Energy Conversion System (WECS)

Tarkeshwar Mahto, Hasmat Malik and V. Mukherjee

Abstract Wind energy has a major contribution to the catalog of renewable energy
generation and also the most efficient energy solutions driver. As the requirement and
utilization for wind energy persist to develop exponentially, reduction in cost of O&M
and reliability growth are the prime concerns in the maintenance strategies of wind
energy conversion system (WECS). Failure of WECS results in wind turbine (WT)
shutdown or expenses maintenance that extensively cut down the yearly revenue. The
WECS failures result in frequent planned O&M scheduled to guarantee the reliability
of wind power (WP) generation. So, condition monitoring (CM) and fault detection
and diagnosis (FDD) methodology has been universally introduced for the premature
fault detection in order to reduce downtime period and increase generation. If at early
stages detestation and rectification are not performed, faults may lead to disastrous
state with large loss of revenue. Therefore, CM and FDD of WECS’s each equipment
(i.e., rotor, gearbox, drive trains, generators, and power electronics) is the need of
future research. This chapter introduces and provides an analysis of the current state
of CM and FDD for each key element in WECS. This work also introduces the survey
with numerous possible paybacks of CM for WECS.

Keywords Condition monitoring · Wind turbine · Gearbox · Bearing · Power

electronics converter · Generator

T. Mahto
Electrical and Electronics Engineering, Birla Institute of Technology Mesra, Ranchi 835215,
Jharkhand, India
e-mail: [email protected]
H. Malik (B)
BEARS, University Town, NUS Campus, Singapore, Singapore
e-mail: [email protected]
Division of ICE, NSUT Delhi, New Delhi, India
V. Mukherjee
Electrical Engineering, Indian Institute of Technology Dhanbad, Dhanbad 826004, Jarkhand, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 121

H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_5
122 T. Mahto et al.

1 Introduction

At present, the entire world is having a severe problem of energy shortage and the
challenge of air pollution is increasing. Thus, the growth of renewable energy is
of the highest priority. In the present scenario, WP generation is a well-established
and mature technology; it is growing fast and is found to be the most promising
form of nonconventional energy generation with being one of the most cost-effective
electricity sources [1]. WECS is an important part of the WP generation system. A
WECS is an arrangement that converts the wind’s KE into electrical energy. The
WECS consists of several key components, such as rotor (i.e., blade, and hub),
main shaft bearings, main shaft, gearbox, mechanical brake, generator, control and
power electronics systems, hydraulic and cooling systems, and yaw system. WTs
are generally positioned in difficult terrain and tough environment condition, such
as mountain islands, seashore, or offshore areas. Thus, the WECS is exposed to
faults frequently and its functioning gets extremely affected which sometimes even
paralysis the system, causing multitudinous loss. Therefore, the improvement in
reliability and security of WECS is of great practical significance [2–4].
A statistical data from Economic Co-operation and Development reveals that
the total renewable energy installation addition of 178 GW or 380 TWh in 2017
has created another history, which is accounting for more than two-thirds of net
global capacity growth in electricity for the first time [5]. The transformation in the
generation of electricity by different sources in 2016–17 is portrayed in Fig. 1 and the
average yearly global progress in renewables-based power production by technology
in the supportable development scenario is shown in Fig. 2. The progress of wind
and solar photovoltaic power in 2017 was astonishing; the largest share in overall
renewable energy growth of 36% was accounted by WP trailed by solar photovoltaic
with 27%, hydro power with 22%, and bioenergy with 12% [6]. Figure 3 displays the
statistical data as percentage of national electric demand for the electricity generated
via wind energy of international energy agency (IEA) wind member countries from
2010 to 2016; it may be noted that the share of wind energy generation has jumped
from 2.3 to 5.2% in the past 7 years [7–13].
With the rapid development in the WP generation sector, there is an increase
in the number of WT units which resulted in the increased rate of failure caused by
extreme working conditions. Figure 4 shows the statistical data accidents which were
documented and authenticated via official information and press reports between
2000 and June 2019 [14].
This increasing trend for the number of accidents reveals that more number of
turbines being installed thus report frequently occurring incidents. So in order to
overcome accidents, scheduled or periodic O&M is carried out which increases the
downtime of WT and additional cost is also required. Thus, a significant share of
wind farm investment is allocated for O&M of WTs [15–18]. The principal major
and minor contributors regarding O&M costs are access and repair of faults [19].
It has been reported that the cost of electricity from WT has two components, first
is capital cost repayments (with interest) and O&M costs is the second component.
Condition Monitoring, and Fault Detection … 123

Fig. 1 Change in electricity generation in the 2016–17 [6]

Fig. 2 Electricity generation

in 2017 [6]
124 T. Mahto et al.

Fig. 3 Average yearly global growth in renewable-based power production by technology [6–13]

Fig. 4 Wind turbine accident summary up to June 2019

Condition Monitoring, and Fault Detection … 125

The latter account for roughly 20% of the generation cost for onshore wind farms
and 15% of the generation cost for offshore wind farms [20]. Thus, it is compulsory
to perform CM and FDD of WT in order to diminish the cost of O&M, the cost of
electricity from wind farms, and improving the availability and reliability of WTs.
A large number of techniques have been projected to realize the CM and FDD
for these WT components. These studies are summarized as a review [21–28], as a
survey according to its investigating objects and are grouped according to different
measuring signals, like acoustic emission (AE), vibration, electricity, and temper-
ature. Elasha et al. [29] compared AE and vibration signals’ effectiveness for the
detection of gearbox’s bearing fault. It is noted that AE may be able to detect fault
more effectively. Along with reviews of different elements with numerous signals,
some work has also been carried out with different fault diagnosis methods for a
single component of WT, like rolling bearing [30], blades [31–33], power plant [34],
and gearbox [35]. Analyzing various components of WECS helps to recognize the
most vital elements of a WT. With this background, the most vital elements of WECS
may be identified optimally for maintenance.
The rest of the chapter has been structured as follows. In Sect. 2, WECS has been
described. Section 3 provides background information on CM & FDD of WECS.
Section 4 presents an overview on CM on WECS. And, Sects. 4.1–4.7 state about
existing systems and related challenges CM or health monitoring for WECSs rotor,
gearbox, main shaft, generator, and bearing for gearbox and generator and power
electronic converters, respectively. Finally, the conclusion is presented in Sect. 5.

2 Methodology

The most important key point and keyword in this study is “CM of WECS and its
application”. This keyword includes all possible applications of CM in the WECS
area. During literature survey, the scope of this study has been considered to limit
up to the important components (i.e., blades, hub, gearbox, main shaft, generator,
bearing, and power electronics converters), of the WECS. Nine digital libraries were
reconnoitered to achieve the target. IEEE, IET, ScienceDirect, Taylor Francis, ASME,
Springer, Wiley, Royal Society, SAGE, MDPI, IOP, and Web of Science provide the
most trustworthy and wide range manuscripts in the field of CM of WECS as shown
in Figs. 5 and 6. This study spread the awareness of the cost of equipment and
O&M which is improved by three basic filtrations and screening process, such as
(1) unrelated manuscripts being separated and then eliminated, (2) duplicate and
irrelevant manuscripts being separated and then deleted by checking the title and
abstract, and (3) full manuscript being reviewed carefully. All steps were applied to
each manuscript to avoid duplicity.
126 T. Mahto et al.

Fig. 5 Reliable sources for

study

Fig. 6 Component-wise
published literature

3 Brief Detail of WECS

Main components of WECS are rotor (which consists of WT blades and hub), gear-
box, main shaft, generator, bearing of gearbox and generator, power electronic con-
verters, and so on. The wind energy is arrested and converted into mechanical energy
Condition Monitoring, and Fault Detection … 127

Fig. 7 EESG based direct drive stand-alone WRSG schematic

by the rotor blades, then, the electrical generator is driven by this mechanical energy
in close aid with gearbox and main shaft connecting all the rotating bodies. The
electrical generator output is controlled through power electronic converters as per
the requirement and arrangements made to evacuate the generated power [36]. The
basic schematic arrangement of WECS is depicted in Fig. 7.
In an ideal state, the mechanical power obtained by the WT is accounted for by
Betz’ Law, as follows [37]:

Pwt = 0.5π R 2 ρv3 C p (λ, β) (1)

where Pwt = captured mechanical power by the WT, ρ = air density, v = wind
velocity, R = length of rotor blade; C p (λ, β) = conversion coefficient of wind energy
(which is a function of tip speed ratio λ and pitch angle β), Λ = ratio of tip speed
to wind speed, that is λ = Ωl (R/v), where Ωl = angular velocity of the WT rotor,
that is the low-speed axis.
Equation (1) shows that, when the speed of the wind is constant, mechanical power
captured by WT is only related to C p (λ, β). If the pitch angle β of the WT remains
unchanged, the conversion coefficient of wind energy C p is only related to the tip
speed ratio λ. For different types of WTs, an exclusive optimal tip speed ratio λ exists
to ensure the best conversion coefficient of wind energy C p and achieve maximum
wind energy capture.

4 Condition Monitoring of WECS

A large amount of data are being collected by CM arrangement which may be utilized
for different analysis and is the driving motivation behind the CM of WECS. For a
viable CM system of WECSs or WTs, the cost of data collection and analysis should
not overweigh the benefits provided through it and also for data storing, processing
128 T. Mahto et al.

and assessing. Though quantifying the CM benefits is complex and hard, researcher
pool is always in favor of CM of WTs.
In 2011, the NREL projected that the yearly cost of O&M per MW was estimated
at around $46,000 and $17,000 for an offshore and onshore wind project, respectively
[38], though CM has the potential to diminish costs for each outfitted WT undergoing
O&M. Unlike traditional power plants, the O&M cost is scattered over many differ-
ent but analogous arrangements (like WTs), and somewhat over a greater landscape
(like wind farm). Moreover, WTs are generally situated in isolated locations, which
requires money, planning, and time for any visit for inspection or O&M. Minimized
upkeep expenditure is a sure benefit of CM permitting for more proficient mainte-
nance practices. Currently, three O&M strategies are commonly implemented [27,
39] (1) time-based O&M (TBO&M, which is preventive), (2) failure-based O&M
(FBO&M, corrective), and (3) condition-based O&M (CBO&M, predictive) or CM.
A conservative O&M strategy involves a combination of TBO&M and FBO&M
strategies [40], but this methodology has room for improvement for WT O&M as
it operates in an unwise environmental condition. The modern O&M strategy is
shifting from TBO&M and FBO&M toward CBO&M [18, 28]. In 2015, Qiao et al.
have proposed a CBO&M strategy that tries to bargain an optimal point between
TBO&M and FBO&M, which may diminish inadvertent outage and preventable
repair actions [41, 42]. With this background, CM may be recognized as a source of
realizing greater accessibilities through dropping cost for O&M [43].
CM strategy empowers an alternate prognostic strategy, often discussed as
CBO&M, where operator executes O&M action when any element exhibits dam-
age signs. This method lessens the number of TBO&M essential and confirms that
the O&M action made is valuable. This decreases the costs of O&M in terms of
materials and labors both and also in advanced stock supervision. In addition to that,
if CM strategy delivers early warning about the failing component, the replacement
may be secured and the downtime is reduced and increases production.
As per the NREL report published in 2013, the cost of O&M may get reduced
by over $20 million annual solely for a 500 MW offshore wind farm by dropping
weather persuaded delay in accessing WTs by maintenance crews [44]. Caithness
Windfarm Information Forum (CWIF) of Scotland, which keeps a record of WT
accidents, logged that 29 blade failure events took place in 2013, in which turbine
blades part or entire blade has been broken away. Since 1990, 280 blade failure cases
have been recorded and 145 structural failure cases have been reported by CWIF
since 1980 [45]. An insurance provider for renewable energy, G-Cube, testified a
typical entitlement of $240,000 in case of blade damage and in case of foundation
damage an amount of $1.3 million in the year 2012. Likewise the claims for trans-
former, gearbox, and generator damage cases were also assessed. CM may provide
an indication for damages leading to failures and are repaired before the occurrence
of failure [46, 47]. Nilsson and Bertling [48] establish that upgrading 47% of reli-
ability centered maintenance measures into CBO&M events may recover the CM
strategy cost. Besnard and Bertling [49] noted that CM strategy for WT maintenance
was more feasible than that of visual examinations for crack opening proportions
Condition Monitoring, and Fault Detection … 129

and failure. Van Dam and Bond [50] projected that the cost of CM system may be
recuperated from CBO&M benefits, in 70% of its useful lifetime.
Finally, CM strategy of WECS may provide additional information for managing
the wind farm, such as (1) increase power generation: CM database may be utilized
for generating input and feedback signals for pitching of blades dynamically and,
for maximizing turbine output; (2) utilized for alleviating blade stresses [51–53];
(3) data sharing among WTs to conflict wake effects; (4) offer weather conditions
advance notice; (5) generation optimizing for the wind farm; (6) furnishe warning
in advance for WT which may be operational up till the maintenance gap which will
result in minimizing downtime and increases generation, like in [48] discovered that
the cost for CM system may be recovered if the operational period is increased by
0.43%; (7) quick verification of structural safety and also estimates expected lifetime
[54], which may be valuable in taking WTs back to operation after any man-made
or natural calamities; (8) automatic shutdown of WTs if threshold of stresses level
is reached, which assets in mitigating concerns related to both failure and public
safety; (9) ice development on blades may be detected [55]; (10) improve future
designees by recognize structural weaknesses from strain and dynamic responses,
which may lead to better consistency and minimizing material usages; (11) monitor
environmental conditions as that on a WT performance and life depends; (12) advance
the knowledge about wake effects in wind farm for researchers’ pool, a cause for
power loss [56] or for monitoring wildlife movement, which may generate turbine
speed reduction signal for an approaching flock of birds or bats, cultivating turbine
environmental-friendliness, etc. [57].

4.1 Rotor CM of WECS

In this study, rotor CM includes blade and hub CM. WT blades are accountable for
translating the winds horizontal force into rotational force that energies the WTs
generator. Blades may be longer than 40 m for the modern GW ratings and designed
with layers of GFRP material fused with resin. The GFRP compromises between
cost-effectiveness, weight, and strength which must be balanced by the designer. For
more long blades, CFRP is used for upturn strength and down tuning the weight; its
cost may be excessive. For extra bending stiffness, balsa wood is added in the shell
of the blades [58]. The structure of the blade consists of three parts: (1) an inner
shear web and (2) two outer shell halves. An “I beam” is created inside the shear
web, permitting extra strength with imitated material and weight. The blades are also
provided with a gel coat over the surface which offers defense from moisture and
minor impacts, which may cause the separation of the layers of blades [59, 60].
An article [61] about commercial CM schemes developed by LM WP and Smart
Fibres Ltd. using fiber Bragg grating sensors for quantifying the strain of the blade
was presented. These systems retain the potential to sense any type of impairment in
blades. The LM WP also states that blades’ lifespan may be increased by reducing
loads in the range of 10–30%.
130 T. Mahto et al.

Schroeder et al. [62] noted that the practices of a fiber-optic system were used in
blades for more than 2 years. The arrangement comprised six strain and one tempera-
ture sensors. Every sensor was fitted to the blades’ interior surface and was adept for
sensing strain characteristics during the operation course. This recorded strain con-
dition may be exploited for early fault identification and safety measures. Blanch and
Dutton [63] attached the sensors outside the blade along the length. The tests were
conducted with artificial loads and the turbine was maintained stationary. This test
method was not applicable to commercial practice. Though online experimentations
were conducted, outputs were indecisive. Yang et al. [64] state that, for harnessing
more and more wind energy, the overall size of the blades are increasing. This makes
it more vulnerable to faults, like cracks, rotor imbalance, hub and blades corrosion,
and severe aeroelastic deflections in bigger WTs.
In [65], the AE system has been utilized for field monitoring of WT blades.
AE sensors send singles to a wireless transmitter mounted in the hub from where
data is uploaded to the ethernet network or to the Internet. In 2014 Crabtree et al.
[66] indicates three fiber-optic-based CM systems for blades, which were skilled in
sensing acceleration, strain, displacement, and temperature. Sensors may be attached
at the time of manufacturing or may be bonded to working blades. The Crabtree report
lists numerous systems in which senses were attached to the exterior of the blade
and linked to the hub with wires and date in received wirelessly into nacelle and
transmitted. Sensors may be also planted inside blades for vibration, motion, and
path monitoring of the blades. Yang et al. [67] found that these systems are low-cost
methods but lack the precise analytical information from the data like information
that are enough to locate and classify the damage. In [68], the field test using an array
of sensors that comprised metal foil strain gauges, fiber-optic strain, accelerometers,
and temperature sensors; these sensors were installed inside the blade as a part of
the active aerodynamic control system through feedback from sensors. For housing
the sensors, inside the blade a mesh of copper was introduced in order to shield the
sensors from electrostatic discharge.
Malik et al. [69] stated that rotor asymmetry is mainly caused by incorrect pitch
angle of the blades designing and imbalance in mass of the blade. Exposure to varying
wind speed and aging are the causes of blades fatigue, which in the long term reduces
blade stiffness which finally drives to cracking. Roughness on surface of the blades is
triggered mainly through ice deposition and environmental pollution [70]. Any fault
in the rotor also brings change in the structure of blade material, mostly vibration and
AE sensors are used to diagnose these faults. AE sensors are capable of identifying
emerging change in structure by installing sensors on the surface of the blade [71].
Monitoring of the blade structure by using distributed temperature and strain
sensors has been carried out by many researchers, like W.Q. Jeffries et al. [72] used
vibration date and processed it with bicoherence (also known as bispectral coherency)
of electrical power for CM of WT blades. Shokrieh and Rafiee [73] carried out
a fatigue failure test using vibration signals in a full composite WT blade using
ANSYS emulation tool and predicted the average life of 21–33 years for a blade. In
[74], the authors have shown the experiment on modeling, optimization design, and
dynamic analysis for composite WT blade and emulated using ANSYS and FEM
Condition Monitoring, and Fault Detection … 131

analysis. Jüngert [75] used AE techniques, namely local resonance spectroscopy and
ultrasound-echo technique for damage detection in WT blades. The work focus on
fault created on a blade due to delamination and how this nondestructive technique
(NDT) is used to find the damaged area.
In [76], a WT imbalance faults investigation has been carried by generating a
twist at the blade in the form of pitch angle and loading weight; the work was
simulated using continuous wavelet transform. Besnard and Bertling [49] presented
an optimized condition-based maintenance method using benchmark data for WT
blades. This method is relevant to maintenance based on inspection-like ultrasonic,
thermography, visual or fiber-optic-based inspection.
Kusiak and Verma [77] have used SCADA for CM of pitch action in WTs blades.
The authors changed the asymmetry of the blade angle and classified it using a
neural network, genetic programming, PART, bagging, and K-nearest neighbor. They
attained the correctness of 74.7%, specificity of 80.5%, and sensitivity of 75.3%.
Similarly, Laouti et al. [78] have tapped radial basis function and support vector
machines for studding fault detection in WTs. Moreover, Kusiak et al. [79] have
used swarm intelligence and data mining with SCADA data for studding adaptive
control of a WT. They used particle swarm fuzzy algorithm as a fault classifier and
created it by pitch angle twist on blade.
Abouhnik and Albarbar [80] carried CM assessment of WT blades using vibration
measurements with ANSYS for the empirically decomposed characteristic. P. Mal-
hotra et al. [81] utilized NDT triaxial testing method and vibration signals to conduct
a design study of blade testing systems. Galappaththi et al. [82] have conducted an
analysis of defects for fatigue life and its effect on WT blades using ANSYS software
and FEM data. They predicted 20% fatigue life reduction in blades and 30% near the
root. In [83], the authors have used FEM analysis for finding WT blade efficiency and
power calculation. They stated that for tip speed ratio of 11.6, the power efficiency
will be 30%.
In [84], the effect of drag has been experimented using MATLAB by creating
twist on the blade for the performance study of efficient blade design. A. Staino et al.
[85] directed research using up and down method for edgewise vibrations in WT
blades for actuator control. Radicevic et al. [86] carried study about lightning strike
effect on WT blade at experimental and theoretical level along with up and down
method using a model of reduced size. Sayed et al. [87] have used computational
fluid dynamics, using finite volume method, and Reynolds-averaged Navier–Stokes
(RANS) equations on analyzing aerodynamic property for different profiles of WT
blade.
In [88], authors have studded the modern trend of blade designing for WT, with
propulsion, practical efficiency, theoretical maximum efficiency, HAWT, and blade
loads by Aerodyn. Bazilevs et al. [89] have proposed a process for precise estimation
of the prebent shape of WT blades. The process depends on aerodynamics simulation
that offers the wind loads on a strictly spinning rotor, followed by a series of structural
mechanics simulations to measure the stress-free prebent shape of the WT blade.
Frost et al. [90] carried out a work of incorporating contingency control with structural
health management by nonlinear high-fidelity simulation and have attained 90%
132 T. Mahto et al.

precision. In [91], the authors have prepared classification and detection of pitch
faults of WT using the RIPPER algorithm and SCADA data analysis which produced
87.05% accuracy in classifying fault on pitch angle.
In [92], WTs damage detection has been carried out by dynamic response
using vibration data. Continuous wavelet transforms has been used for transform-
ing dynamic responses to surfaces for accentuating the effect of wind or damage.
Yokoyama [93] carried a study on WT blades and recommended selected material for
blade lightning protection. In [94], FEM analysis has been carried out for determining
the dynamic characteristics of the WT blade for lag vibration.
In [95], a review of large WT blades based on acoustic, dielectric, thermal, ultra-
sonic, and fiber-optic monitoring methods and cure monitoring techniques for CM
has been carried out. They projected fiber Bragg gratings as a superior method for
monitoring when compared with other methods. Buckney et al. [96] have proposed
alternate structural layouts for blades of WT with respect to their design, reducing the
cost, and minimizing weight with the aim of improving the structural efficiency of a
WT blade. In [97], Chen et al. using SCADA with knowledge-based adaptive neuro-
fuzzy inference system (ANFIS) performed a laboratory work on WT diagnosis of
pitch faults with an aim of classification and attained 88.30% accuracy.
In [98], vibration signals have been used to study WT blade damages at trailing
edge due to their structure and location. The visual inspection method was used
to site the blade cracks (transverse and longitudinal cracks).The average life span
forecasted was to be 17–22 years for a blade of 100/300 kW WTs. Chou et al. [99]
have done WT blade failure analysis using vibration data and ANSYS under critical
wind loads. In [100] Bachynski et al. have conducted dynamic analysis during grid
loss, pitch actuator fault, and shutdown of floating WTs.
In [101], authors have prepared a study on inspecting, testing, and monitoring tech-
niques for a blade of WT. the survey comprises full-scale testing, mechanical property
testing, structural health monitoring, nondestructive testing/inspecting (X-ray imag-
ing and infrared thermography), and CM using vibration signals and acoustic. Bedon
et al. [102] studied evolutionary algorithm and blade element momentum theory for
a Darrieus vertical axis WTs optimization. In [103], annual energy production cal-
culation and blade element momentum theory have been utilized for optimizing the
slope of the liner radial profiles of the twist angle and blade chord for a fixed speed,
fixed-pitch WT blades through adopting linear radial profiles.
In [104], the author used MATLAB simulation and vibration data for estimation of
coefficient of WT power based on adaptive neuro-fuzzy approach. Dervilis et al. [105]
have used pattern recognition (like nonlinear principal component analysis, principal
component analysis (PCA), auto-associative neural network, radial basis functions,
and artificial neural network) on damage diagnosis with vibration signals for a WT
blade research. Kim et al. [106] carried a review on recognizing WT blade cracks
experiencing working loads with the technique of vibro-acoustic modulation. Bouzid
et al. [107] conducted a work using wireless sensor networks and acoustic source
localization for WT blade structural health monitoring and obtained an accuracy
rate of 92.02% in the work. Niezrecki et al. [108] presented a work on CM of WT
blade using the technique of acoustic beam forming. In [109], failure of adhesive
Condition Monitoring, and Fault Detection … 133

joints in WT blades has been studied using FEM and vibration signals with fracture
mechanics approach.
In [110], method for damage detection of WT blades based on mode shape differ-
ence curvature information and dynamics analysis has been carried out. Zangenberg
et al. [111] have recommended a new design for WT blades with fibrous composite
preform and different materials. Chen et al. [112] have made a suggestion about the
composite materials for WT blades after investigating a composite WT blade 52.3 m
glass/epoxy through visual inspection. In [113], a mathematical study about icing on
WT blade airfoil has been carried out. They have forecast maximum icing area ratio
and an icing rate at 80° of attack angle to be 21.8% and 6.6%, respectively.
In [114], dynamic mechanical analyzer and shear dynamic test has been used for
vibration damping in WT blades with star-shaped inclusions on macro-composites.
Zhou et al. [33] have done a structural testing review of WT blades on full-scale using
photogrammetry, digital image correlation, acoustic signal. and with X-Ray. Kumar
et al. [115] have carried a work on WT designing using vibration signals and FEA
for AL 2024 WT blade on free vibration analysis. In [116], composite rotor blades
cracking study has been carried out using cracking modeling matrix and attained the
precision of 92.4% with variational asymptotical beam sectional analysis framework
and geometrically exact beam theory.
In [117], the structural design of spar blades for 100-m biplane WT has been
carried out. Numerous designs for spar blades were structured in order to estimate
the biplane and monoplane spars. They determined that biplane spars have smaller
tip deflections and maximum root bending moments than monoplane spars by 25–
35% and 75%, respectively, for same length and mass spare. Pratumnopharat et al.
[118] have used vibration signals of horizontal axis WT blades with wavelet transform
based stress time history editing. They used time-correlated fatigue damage (TCFD),
Mexican hat wavelet (Mexh), Meyer wavelet (Meyr), Daubechies 30th order (DB30),
Morlet wavelet (Morl), and discrete Meyer wavelet (Dmey) as a diagnostic method
and obtained TCFD—89.82%, Morl—80.34%, Meyr—79.76%, Dmey—80.30%,
Mexh—79.23%, and DB30—80.81% accuracy.
In [119], FEM analysis has been used to compute the nonlinear aeroelastic behav-
ior of horizontal axis large WTs. Woude and Narasimhan [120] made a study on struc-
tural vibration isolation of WT using ANSYS depending on real and simulated data.
Zhang et al. [121] have done work about the impact of wind shear under fluid–struc-
ture interaction using ANSYS on vibration of geometrically nonlinear WT blade.
In [122], vibration data of a WT have been used for fuzzy gain-scheduled active
fault-tolerant control has been carried out.
In [123], an experimental assessment of air termination systems using the up and
down method for WT blades reported in IEC-61400-24 standard has been carried out.
Gómez et al. [124] have used composite materials and CAD/CAM techniques for
and manufacturing and designing WT blades of low capacity. In [125], radial basis
and polynomial function were considered as the kernel function for comparative
study on WT power coefficient approximation by soft computing methodologies.
With vibration signals, they gained a correlation coefficient of 0.997.
134 T. Mahto et al.

In [126], digital image correlation has been used for inspecting full field of a
50 utility WT blade subjected to cyclic loading and quasistatic. Mollineaux et al.
[127] have used the continuous wavelet transform as modeling techniques in order
to carry WT blade damage detection testing with standard data and autoregressive
moving average collected from accelerometer sensors with wire or wireless. Slot
et al. [128] have conducted a review and then advocated about the material and
lifetime estimation of blade with life coating models on leading-edge erosion of
blades. Zhang et al. [129] conducted performance evaluation based on design and
analysis of coated WT blade on jet-based laboratory equipment. Bitkina et al. [130]
performed theoretical and experimental investigation applying Cauchy equations and
Kirchhoff hypotheses for anisotropic multilayer composite panels for WT blade to
determine the state of stress-strain. Rajkumar et al. [131] have discussed WT blades
with low weight and high-velocity designing and material requirements. Benim et al.
[132] have carried a stability investigation of small WT by applying Mesh Adaptive
Direct Search optimization algorithm for optimizing the airfoil profiles under highly
unstable wind conditions. In [133], a work was conducted on improving joint design
using cohesive zone modeling for failure resistance of composite WT blade materials.
In [134], vibration signals have been used for CM of the WT blade by using
transformation algorithm on WT blade moment signals. Habibi et al. [135] have
used low-frequency forced vibrations and ultrasonic-guided high-power waves col-
lectively for de-icing WT blades. This either eliminates existing ice layer on the blade
surface or prevents ice accumulation (anti-icing). Rezaei et al. [136] have given a
reduced-order model for blade dynamics of WT and, Galerkin method and FEM
analysis for nonlinearity analysis. Li et al. [137] have used vibration data, sparse
reconstruction approach, lamb wave propagation, and growth sparse pursuit for car-
rying out research on growth of crack sparse pursuit for WT blade with the assistance
of CoSaMP algorithm for modeling. Xie et al. [138] introduced a novel folding blade
of WT using MATLAB for impactful control of power and have attained extreme
reduction of 51.1% associated with pitch control. Kumar et al. [139] have done a sim-
ulation work for management of generated power when functioning at above-rated
wind speed and maximization of power generated at below-rated speed through an
integral sliding mode. Liu et al. [140] have undertaken a cracked blade in rotor sys-
tem and studied the nonlinear vibration characteristics due to alternating loads using
FEM and vibration data. In [141], the effects of discrete water jets, impact distance,
substrate geometry, and water cushioning were considered for studying rain erosion
of WT blade coating properties using image processing.
Hoell and Omenzetter [142] have presented damage detection method for struc-
tural of WT blades using cross-correlations and vibration data established on time-
series demonstrations of dynamic responses, and as a diagnostic algorithm PCA
genetic programming has been considered. In [143], the FEM analysis has been
considered to study the consequence of local buckling in WT blades of delamina-
tion. They performed the simulation and experimentation by introducing flap-wise
bending of the blades.
Condition Monitoring, and Fault Detection … 135

4.2 Gearbox CM of WECS

In the drive train of WTs, gearboxes practice critical load transmitting, yaw and
pitch mechanisms [144], which leads to high level of reliability [145]. The WT drive
train comprises a single-stage planetary gearbox paired parallel with intermediate
and high-speed fixed-shaft gearboxes to provide a greater total gear ratio. Numerous
multistage planetary gearboxes are paired for pitch and yaw mechanism. Presently,
most of the existing WTs operate utilizing a fixed-gear ratio between the electri-
cal generator and the turbine rotor. A WT with a variable ratio gearbox permits a
constant-speed generator, to discretely modify rotor speed and to accomplish bet-
ter aerodynamic efficiency [146, 147]. Thus, for achieving greater input speed and
transmission ratio, the gearbox is generally fabricated as sun and planetary gear with
parallel gear structure. Since the gearbox operating hours are long in rough condi-
tions like wind gust, dust corrosion, or heavy load, this results in various failures of
the gearbox, like shaft misalignment, gear damage, shaft damage, shaft imbalance,
bearing damage, broken shaft, high oil temperature, poor lubrication, and leaking
oil.
In [148], an iterative atomic decomposition technique for fault diagnosis in the
planetary gearbox of WTs has been presented. Baoping [149] have worked on Shan-
non wavelet support vector machine and manifold learning for WT transmission
system fault diagnosis. Anmei [150] aimed early failure detection at planetary gear-
box based on failure mode effects and criticality analysis (FMECA) and ontology
analysis. Liu [151] focused on Fourier series analysis and robust feature extraction
algorithms for equally spaced planetary gearbox.
In [152], authors illustrate that the conformation of a local gear fault and its
location using correlated kurtosis and fast dynamic time warping techniques with
a time-domain method for planetary gearbox fault diagnose may be conducted. Lei
et al. [153] utilized two different approaches for diagnostic, namely the normalized
summation of positive amplitudes and filtered signal root mean square for various
spectrums between the healthy and unknown signals of planetary gearboxes. The
claim was for successful identification of cracked tooth and pitting gear faults. In
[154], the sideband levels and TSA have been used for identification of intermediate
stage gear faults in the intermediate stage of a high-speed WT compound gearbox.
In [155], the harmonics of planetary gearboxes for its fault detection were dis-
cussed based on kinematic modeling. Chaari et al. [156] have introduced two local
faults (i.e., cracks and tooth pitting) in a planetary gear set and studied its impact
using the vibration spectrum on gear set. Feng et al. [157] introduced a vibration
signal based model for fault diagnosis in sun and planet gears and ring undertaking
frequency and amplitude modulations. The model was also confirmed using indus-
trial data on laboratory. Bartelmus et al. [158] demonstrated a new model of CM
for planetary gearbox under changing conditions of external load. The technique
was based on the assumption that there exists a linear relationship between load and
vibration at a specific speed interval. It was revealed that the parameters of regression
fit for a faulty gearbox differ from a healthy gearbox on normal operating conditions.
136 T. Mahto et al.

Barszcz et al. [159] applied spectral kurtosis technique for detecting ring gear cracks
WT gearbox which was not detected by frequency analysis and standard time series.
Lei et al. [160] undertook the stochastic resonance mechanism and improved it by
utilizing ant colony optimization for the identification of chipped and broken gear
tooth in the second stage of multistage planetary gearbox were detected using the
Fourier spectrum was difficult. Feng et al. [161] have carried out fault diagnosis
for WT planetary gearbox considering empirical mode decomposition technique. In
[162], an upgradation in fault diagnosis of WT planetary gearbox was suggested
using adaptive time intervals for the case of nonstationary loads. Feng et al. 144]
unwind the novel planetary gearbox fault diagnostics method using vibration spec-
trum of the target gear via iterative atomic decomposition thresholding. In [163], an
adaptive Morlet wavelets spectrum has been utilized for discovery of tooth cracks in
WT gearboxes.
In [164], oil debris based CM and health assessment along with health manage-
ment for WT gearbox have been discussed. Dempsey [165] introduced a technique
based on fuzzy logic for sensing pitting damage levels in gears grounded by monitor-
ing oil debris. Comparative damage-level assessment for pitting damage in gearboxes
through vibration technique and oil debris monitoring was offered in [166], it was
stated that vibration and oil debris monitoring based CM could perceive the incidence
of damage but does not classify among gearbox and bearing damage exclusively. So,
a fuzzy logic-based method has been used for classifying damage detection in gears
and bearing by blending the diagnostic information from oil debris monitoring and
vibration sensing was offered in [167].
In [168], the detection of seeded gear fault based on AE was projected and also
the fact that the temperature of the lubrication may substantially affect the AE was
clarified. But, Qu et al. [169] carried out a comparative analysis between vibration and
AE data for gearbox tooth crack diagnostics. And discovered that for low rotational
speed, the AE technique is more sensitive and is less vulnerable to other mechanical
noise. Vicuña [170] studied the consequences of load variation; rotational speed and
lubricant temperature of planetary gearbox on AE was investigated.
In [171], the SCADA data has been used for behavioral modeling for sensing faults
in generator and gearbox based on neural networks trainer for online monitoring. It
was claimed that the technique was competent in identifying fault about 6 months
in advance than the actual fault occurrence. Kim et al. [172] implemented clustering
techniques and anomaly detection algorithms for detecting fault in gearbox using
historic data of SCADA. Loutas et al. [173] carried out a sensor smooth data collection
method utilizing vibration, AE and oil debris observing methods for estimating heath
of gearbox by utilizing fusion approach. Feng et al. [174] presented CMS and SCADA
signals based WT planetary gearbox diagnostic approach. The approach concluded
that from the temperatures of oil, bearing, nacelle, and the power output, prediction
of fault in gearbox may be conducted. [175].
Villa et al. [176] approached vibration study based research using angular resam-
pling in WTs under the influence of nonlinear speed fluctuation. Márquez et al. [27]
carried out a review on the CM of WT; they suggested numerous methods and tech-
niques in order to keep the gearbox and other parts of the WT healthy. Liu et al.
Condition Monitoring, and Fault Detection … 137

[177] presented a diagnosis approach for predicted the gear tooth crack in WT using
the local mean decomposition. Qu et al. [178] have conducted a gear fault detection
method using vibration signals with a new AE sensor.
In [179], ANFIS for a forecast for WT gearbox prognostics based on time-series
prediction and vibration analysis has been conducted. Crowley et al. [180] used
statistical analysis to learning the influence of dissimilar work scope schemes on
gearbox life cycle and repair costs of WT.
Yin et al. [181] pronounce the development of fault identifying straight using
processed data of robust residual generators. The benchmark model was used to
simulate faults in order to classify and detecte faults in yaw and pitch actuators,
generator, and stuck sensors. Zhu et al. [182] carried out research for planetary gears
of WT using FEM analysis and vibration signals on dynamical characteristics with
flexible pins. Antoniadou et al. [183] carried out classification of faults appearing
on WT gearbox due to varying load conditions using vibration research on a time–
frequency analysis approach from diverse techniques, such as Hilbert transform (HT),
empirical mode decomposition (EMD), MATLAB modeling, and Teager–Kaiser
energy operator (TKEO).

4.3 Main Shaft CM of WECS

Faults in the main shaft include misalignment, corrosion, coupling failure, and crack
[184]. The mentioned faults hinder the standard operation for the main shaft along-
side the other auxiliary systems associated with it, which finally affects the torque
transmitted through the drive train and may result in vibrations in particular at the
rotor, generator, and gearbox at specific characteristic frequencies [185]. Any mis-
alignment in shaft leads to variation in magnitude of the fundamental frequency of
the vibration in rotor, generator, and gearbox. Thus, the faults in the main shaft may
be analyzed and detected from torque, vibration, and electrical signals [184]. All the
analyses are carried out through the frequency analysis approaches like fast Fourier
transform.

4.4 Generator CM of WECS

The generator (electrical machine) is fitted inside the nacelle; the main shaft rotational
torque drives the generator, where the electrical energy is achieved from mechanical
energy. WT generators from the structure point are like squirrel cage induction gen-
erator, doubly fed induction generator (DFIG), and permanent magnet synchronous
generator (PMSG). WT generator operates in an electromagnetic atmosphere which
may lead them to certain usual faults like [186] overheating, vibration, bearing over-
heating, insulation damage, and abnormal noises; the generator faults are classified
into two categories: electrical and mechanical faults. The electrical faults include
138 T. Mahto et al.

open circuit, electrical imbalance, and stator/rotor insulation damage. And, bearing
failure, rotor mass imbalance, bent shaft, broken rotor bar, and air-gap eccentricity
are the main mechanical faults.
Yucai et al. [187] have used the electromotive force method for diagnosis of
short-circuit faults within the excitation system of turbo generator. Djurovic et al.
[184] used instantaneous power data and steady-state stator line current of DFIG for
CM based on a frequency spectrum. A comparative study between experimental and
simulation results of the current and power spectra were carried out. Marzebali et al.
[185] have presented the influence of torsional vibration on a planetary gearbox of
a wound-rotor induction generator with experiment and simulation results. It was
validated that with stator current spectrum, it is tough to recognize the planetary
gearbox mechanical characteristics of most frequencies. Yacai and Yonggang [187]
have developed a 2-D simulation model for a generator utilizing basic structural
parameters and electromagnetic principle applicable in generators. The generator’s
excitation current was found and two binary functions were utilized for the calculation
of electromotive force at no load expectations on the active power and excitation
current. It may be used for the overall health diagnosis for excitation winding and
real-time monitoring of the generator.

4.5 Bearing CM of WECS

Bearings are very crucial elements of gearbox and generators. The bearing alone
causes approximately 70% of gearbox and 50% of generator downtime (i.e., approx-
imately 21, 50, and 70% on small generators (P < 1 MW), medium generators (1 MW
< P < 2 MW), and large generators (P > 2 MW), respectively). Figure 8 depicts the
gearboxes, medium generators, and bearing downtime [188, 189]. Thus, it is vital to
use CM on bearings and to improve methods and sensors which may assist in early
fault diagnosis.
Dupuis et al. [164] conducted a fault diagnostics study on bearing and gear fail-
ures. Oil debris based diagnostic technique was applied to in-service WTs. Jayaswal
et al. [190] conducted a vibration analyses using FFT spectrum techniques on bear-
ings for detecting faults. Liang et al. [191] projected a method for fault detection
on bearing based on an intelligent calculus enhanced energy operator for identifying
the signature of defected bearing under multiple vibration interferences and strong
noise situation. Dalvand et al. [192] discussed a method based on the instantaneous
frequency with envelope analysis of vibration signal and calculus of condition indica-
tors such as crest and kurtosis feature to identify bearing faults. The effectiveness of
the work was demonstrated with experimental results. Zimroz et al. [193] carried out
a CM methodology utilizing parameters acquired from commercial diagnostic sys-
tems such as root mean square, peak-to-peak of vibration acceleration, and generator
power for bearings which are associated with the normal operating conditions. Ali
et al. [184] used empirical mode decomposition energy entropy for feature extraction
Condition Monitoring, and Fault Detection … 139

Fig. 8 Summary of drive train failures. Source https://www.windpowerengineering.com/how-

disruptive-technology-ischanging-condition-monitoring-systems-in-wind-turbines/

method of vibration signals. For classifying the bearing faults, the supreme signifi-
cant intrinsic mode functions were elected to train an artificial neural network. The
reliability of the tool in evaluating CM and bearing degradation was experimentally
presented. Azevedo et al. [30] review bearings CM of WT with numerous signals,
which reported about articles that reviewed fault detection methods projected for a
single WT element, like rolling bearing. Elasha et al. [29] equated the usefulness of
detecting bearing fault of gearbox by AE signals and vibration and found that earlier
indication of fault may be possible with AE.

4.6 Power Electronics Converters CM of WECS

With the rise in the power rating of WTs, the reliability issues for the power electronic
converter have increased. The literature states that about 20% failures in WECS
are caused by power electronic converters alone [21, 194]. The leading causes of
their failure are mainly due to their exposer to varying temperatures, humidity, and
vibration. [195]. In power electronic convertors of WECS the insulated gate bipolar
transistor (IGBT), joints and printed circuit board (PCB) are the major elements were
faults occurs normally as shown in Fig. 9. The IGBT fault includes bond short circuit,
wire liftoff, solder fatigue, cracks, and gate misfiring. Fault in capacitor comprises
short/open circuit, forms conductive paths, migration of electrode materials across
the dielectric, increased dissipation factor, and breakdown of dielectric. PCB faults
140 T. Mahto et al.

Fig. 9 Failure rate distribution of power electronic convertor for WECS

include corrosion or crack of traces, broken buried metal lines, board delamination,
cold-solder joints, and component misalignment [196].
In WTs, modern AC–DC–AC converter based WECSs are the widespread topol-
ogy for realizing two main goals: (1) tracking for extraction of maximum power and
(2) reactive power regulation between grid and WT.
Numerous partial- and full-scale topologies for power conversion of MW-level
WTs are being utilized for integrating WT output power, adaptive operation, and
control throughout the operating conditions (i.e., transient and steady state) [21].
The performance of WECS has improved with the advancement in power convert-
ers. It decreases WT’s mechanical stress and align WECS collectively for a controlled
generation though, at higher power density, converters become more volatile than
other elements in WECS as per the statistics for failure of WECS, in Sweden [197],
China [198], and Germany [199]. The control circuit and wind converter correspond
to a highest rate of malfunctioning within the WECS. Though downtime for electrical
systems is lesser than that of mechanical systems, the power converters and control
systems have resulted in comparatively high rate of failures. Figure 10 represents
the failure distribution statistic along with stress level for power converter [200, 201]
where semiconductor and DC-link capacitor have triggered 21 and 30% of the total
failures, respectively. The stress levels are created by several reasons listed as fol-
lows: (1) temperature steady-state and cyclical (TSS&C), (2) vibration and shock
(V&S), (3) humidity and moisture (H&M), and (4) contaminants and dust (C&D).
Power converter offers wide utilization in unified power quality correction,
adjustable-speed drive, grid interconnection with RES, hybrid electric vehicle or
electric vehicle, and energy storage system [202]. In industrial converters, faults may
occur because of overloading, aging, or unpredicted operating circumstances [203].
Though fault occurrence are usually similar in power electronics, failure phenomena

Fig. 10 Component failure rate and stress-level in power converter of wind turbine
Condition Monitoring, and Fault Detection … 141

and main failure reasons vary extensively for purposes like renewable energy, electric
vehicle, and motor drives convertors [204–209].
Kamel et al. [210] projected two methods of multiple fault diagnosis for power
converters involving either, boost chopper or uncontrolled three-phase rectifier or
single-phase inverter circuits. The current and voltage readings of the converter are
used to identify and locate the fault in utilizing spectrum analysis. In [211], the output
voltage of rectifier, estimated phase currents, and chopper average voltage values are
investigated to identify the analogous fault at modules with Fourier series expansion.
Karimi et al. [204] presented a method of detecting single switch fault for WECS of
WT. The process may classify the fault leg but unable to identify the faulty switch. In
order to perform, it needs an additional FPGA-based process module and pole voltage
detection circuit which finally increases the system complexity and cost. In response
to identifying various faults, Jlassi et al. [212] suggested a model-based method for
back-to-back (BTB) converter’s open-circuit fault diagnostic in PMSG-based WT.
The nature of current form factors (CFFs) gets affected due to open-circuit fault.
Thus, the difference between observed and measured CFFs is utilized to distinguish
the faulty leg. In [213], a method for detecting and localizing faults has been projected
for converter of WT. It embraces average value of current and load current analysis
for fault detection in single and double openswitch. The measurements in WECS
are affected by the deviation in control of real and reactive power. Freire et al. [200]
adopted the derivative of the current polarity, absolute value of current parks vector,
and errors in normalized current average for detection of open-circuit and switch-fault
detection on both side WT convertor system. In [214], author advocated a real-time
fault detection and localization technique for two-layer BTB (2L-BTB) converter.
The method uses the absolute average value of the normalized currents and the sum of
three-phase normalized currents. Additionally, this algorithm may also detect faults
in current sensor and power switch too.
Choi et al. [215] offered a technique for detection open-switch fault for a neutral-
point clamped grid-connected inverter in a single switch by utilizing the current
patterns. This technique is beneficial as additional sensors and complex calculations
are avoided in it. The open-switch fault may be sensed and located between half to
two fundamental periods, respectively. Lee et al. [216] projected an open-circuit fault
diagnostic method by undertaking both side open-switch fault together in converters.
It also offers two techniques for identifying the faulty converter side and detecting
faulty leg. It senses open-switch fault in both upper and lower sections of a three-layer
NPC BTB converter with predefined algorithm parameters.
A data mining based approach [217] has been suggested for diagnostics of capac-
itor’s multiple parametric faults in DFIG-based WECS. The method comprises auto-
adaptive dynamical clustering, dynamical feature generation, classifier learning, and
updating. In this approach, degradation of capacitor has been considered as an end-
less drift from normal functioning with time. It has been used for detecting and
monitoring multiple capacitor degradation, but no exact detection duration has been
distinct for simulation and experimental work. In [218], a way of open-switch fault
detection for three parallel converters in PMSG-based WTs was presented. In order
to classify and generate topographies along with computing historical data for the
142 T. Mahto et al.

three-phase current measurements, neural network (NN) and pattern recognition have
been utilized. The model-based diagnosis methods are adopted to detect open circuit
in modular multilevel converters (MMCs).
MMC fault may be detected with the Kalman filter by associating the measured
inner difference current and the estimated inner difference current [219]. This method
requires relatively long detection time and additional sensors too. To overcome these,
Liu et al. [220] unfolded a nonlinear adaptive observer for both fault detection and
localization for MMC’s short-circuit and open-circuit faults. The fault characteristics
are required in this method to design detection variables and generate diagnosis
features.

4.7 Research Gap Finding and Suggestion

The above subsections have briefed numerous research works regarding the CM and
fault diagnosis for WECS of WT through feature and fault condition determination.
However, there still exist a number of un-addressed issues in the existing literature.
Meanwhile, the authors would like to impart possible future research prospects for
WECS of WT CM and fault diagnosis.
(1) The operating modes of rotational speed and load condition always affect the
CM and fault diagnosis of WECS, and researchers pool have also proposed some
useful algorithms. However, it is still quite challenging to determine the fault
types, and also most of the studies reviewed above only focus on the detection
of the one fault type on certain components.
(2) Smaller section scholars pool carries researches that focuses on the root cause
of the faults. So, the authors advocate giving extra consideration for precise
CM and fault diagnosis. This may include aspects like exact dynamic modeling
for all fault types, creating more explicit demonstrative features, and building a
more accurate classifier.
(3) The authors would like to suggest that there still exists a substantial research gap
in building cost-effective, online CM techniques for WECS. With the advance-
ment in power electronics, communication technology, and sensor accuracy,
reliable online CM for WTs will be viable in the near future.
(4) The prevailing CM detects and identifies WECS’s faults on the basis of a con-
siderable deviation of system states. The closed-loop controls have not been
effectively investigated, which checks the sensitivity and accuracy of CM and
fault detection methods.
(5) The financial capability and applicability to any CM method is an essential
consideration. The authors believe that there could be more papers remarking
and reciting financial and practical attributes of their practices for adaptation on
WTs easily and fastly.
Condition Monitoring, and Fault Detection … 143

(6) Future work may undertake to employ the online real-time function of the
present WT CM technique based on cloud computing, and cultivate an online
real-time CM device based on the existing method.

5 Conclusion

It has been observed that the O&M cost is having a high share in WT or wind farms.
In order to reduce the O&M cost of WT or wind farms, CM or fault diagnostics of
different components is very essential in order to detect any incipient faults in any
component to avoid potential failure threat. This chapter presents a comprehensive
review of common faults, signals, and signal processing methods for CM and fault
diagnosis for WECS of WT. Each technique’s advantages and limitations have been
focused too. The review found that there is still a noteworthy research gap in devel-
oping comprehensive, cost-effective, CM techniques for WECS of WT or wind farm
exists. In future, these techniques may also be combined with each other to provide
a better result for the CM and fault diagnostics and for carrying out the best CM and
fault diagnosis method for the WECS of WT.

References

1. A.B. Attya, J. Dominguez-Garcia, O. Anaya-Lara, A review on frequency support provision by

wind power plants: Current and future challenges. Renew. Sustain. Energy Rev. 81, 2071–2087
(2018). https://doi.org/10.1016/j.rser.2017.06.016
2. Z. Yang, C. Yi, A survey of fault diagnosis for onshore grid-connected converter in wind
energy conversion systems. Renew. Sustain. Energy Rev. 66, 345–359 (2016). https://doi.org/
10.1016/j.rser.2016.08.006
3. H. Jun et al., An overview of condition monitoring and fault diagnostic for wind energy
conversion system. Transactions of China Electrotechnical 28(4), 261–272 (2013). http://en.
cnki.com.cn/Article_en/CJFDTotal-DGJS201304035.htm
4. H. Malik, S. Mishra, FAST and simulink based simulation investigation of wind turbine faults.
Int. J. Renew. Energy Technol. 8(3/4), 286–304 (2017). https://doi.org/10.1504/IJRET.2017.
088970
5. IEA, Renewables 2018. https://www.iea.org/renewables2018
6. IEA Global Energy & CO2 Status Report 2017. Accessed on 10 Aug 2019. https://webstore.
iea.org/global-energy-co2-status-report-2017
7. I.E.A. Wind, IEA wind annual report 2010, 2011. Accessed on 10 Aug 2019.
https://community.ieawind.org/HigherLogic/System/DownloadDocumentFile.ashx?
DocumentFileKey=822d2af3-4078-922e-5146-eec3a17cae6d&forceDialog=0
8. I.E.A. Wind, IEA wind annual report 2011, 2012. Accessed on 10 Aug 2019.
https://community.ieawind.org/HigherLogic/System/DownloadDocumentFile.ashx?
DocumentFileKey=133ea3d2-bff1-8800-3dcf-a08f27f6fc55&forceDialog=0
9. I.E.A. Wind, IEA wind annual report 2012, 2013. Accessed on 10 Aug 2019.
https://community.ieawind.org/HigherLogic/System/DownloadDocumentFile.ashx?
DocumentFileKey=133ea3d2-bff1-8800-3dcf-a08f27f6fc55&forceDialog=0
144 T. Mahto et al.

10. I.E.A. Wind, IEA wind annual report 2013, 2014. Accessed on 10 Aug 2019. https://
nachhaltigwirtschaften.at/resources/iea_pdf/reports/iea_wind_annual_report_2013.pdf
11. I.E.A. Wind, IEA wind annual report 2014, 2015. Accessed 10 Aug 2019. https://
nachhaltigwirtschaften.at/resources/iea_pdf/reports/iea_wind_annual_report_2014.pdf
12. I.E.A. Wind, IEA wind annual report 2015, 2016. Accessed 10 Aug 2019. http://www.rse-
web.it/notizie/IEA-publishes-Wind-TCP-2015-Annual-Report.page
13. I.E.A. Wind, IEA wind annual report 2016, 2017. Accessed 10 Aug 2019. https://community.
ieawind.org/blogs/cezanne-murphy-levesque/2017/10/24/iea-wind-tcp-2016-annual-report
14. Caithness Windfarm Information Forum, Summary of WT Accident data to 30 June
2019, 2019. Accessed on 10 Aug 201919 at http://www.caithnesswindfarms.co.uk/
AccidentStatistics.htm
15. W. Musial, R. Ram, Large-scale offshore wind power in the United States: Assessment of
opportunities and barriers. No. NREL/TP-500-40745. National Renewable Energy Lab.
(NREL), Golden, CO (United States) (2010). https://www.nrel.gov/docs/fy10osti/40745.pdf
16. D. McMillan, W.A. Graham, Quantification of condition monitoring benefit for offshore wind
turbines. Wind Eng. 31(4), 267–285 (2007). https://doi.org/10.1260/030952407783123060
17. T.W. Verbruggen, Wind turbine operation and maintenance based on condition monitoring
(Final Report Energy Research Center, The Netherlands, 2003). https://publicaties.ecn.nl/
PdfFetch.aspx?nr=ECN-C--03-047
18. M.D. Reder, E. Gonzalez, J.J. Melero, Wind turbine failures-tackling current problems in
failure data analysis. J. Phys. Conf. Ser. 753(7) (2016). https://iopscience.iop.org/article/10.
1088/1742-6596/753/7/072027/pdf
19. R. Martin et al., Sensitivity analysis of offshore wind farm operation and maintenance cost
and availability. Renewable Energy 85, 1226–1236 (2016). https://doi.org/10.1016/j.renene.
2015.07.078
20. A. Joshuva, V. Sugumaran, Fault diagnostic methods for wind turbine: a review. ARPN J.
Eng. Appl. Sci. 11(7), 4654–4668 (2016). https://www.arpnjournals.com/jeas/archive.htm
21. W. Qiao, D. Lu, A survey on wind turbine condition monitoring and fault diagnosis—Part
I: Components and subsystems. IEEE Trans. Industr. Electron. 62(10), 6536–6545 (2016).
https://doi.org/10.1109/TIE.2015.2422112
22. M. Luengo and A. Kolios, “Failure mode identification and end of life scenarios of offshore
wind turbines: a review”, Energies, vol. 8, no.8, pp. 8339-8354, 2015. Accessed on 10-08-2019
at https://www.mdpi.com/1996-1073/8/8/8339/pdf
23. W.Y. Liu et al., The structure healthy condition monitoring and fault diagnosis methods in
wind turbines: A review. Renew. Sustain. Energy Rev. 44, 466–472 (2015). https://doi.org/
10.1016/j.rser.2014.12.005
24. Z. Hameed et al., Condition monitoring and fault detection of wind turbines and related
algorithms: A review. Renew. Sustain. Energy Rev. 13(1), 1–39 (2009). https://doi.org/10.
1016/j.rser.2007.05.008
25. A. Salem, A. Abu-Siada, and S. Islam, “A review of condition monitoring techniques of the
wind turbines gearbox and rotor”, International Journal of Electrical Energy, Vol. 2, No. 1,
2014. Doi:https://espace.curtin.edu.au/handle/20.500.11937/38325
26. W. Qiao, D. Lu, A survey on wind turbine condition monitoring and fault diagnosis—Part
II: Signals and signal processing methods. IEEE Trans. Industr. Electron. 62(10), 6546–6557
(2015). https://doi.org/10.1109/TIE.2015.2422394
27. F.P.G. Márquez et al., Condition monitoring of wind turbines: Techniques and methods.
Renewable Energy 46, 169–178 (2012). https://doi.org/10.1016/j.renene.2012.03.003
28. P. Tchakoua, et al., “Wind turbine condition monitoring: State-of-the-art review, new trends,
and future challenges”, Energies, vol. 7, no.4, pp. 2595-2630, 2014. <https://www.mdpi.com/
1996-1073/7/4/2595>
29. F. Elasha et al., A comparative study of the effectiveness of vibration and acoustic emission
in diagnosing a defective bearing in a planetry gearbox. Appl. Acoust. 115, 181–195 (2017).
https://doi.org/10.1016/j.apacoust.2016.07.026
Condition Monitoring, and Fault Detection … 145

30. H.D.M. de Azevedo, A.M. Araújo, N. Bouchonneau, A review of wind turbine bearing con-
dition monitoring: State of the art and challenges. Renew. Sustain. Energy Rev. 56, 368–379
(2016). https://doi.org/10.1016/j.rser.2015.11.032
31. R. Yang, Y. He, H. Zhang, Progress and trends in nondestructive testing and evaluation for
wind turbine composite blade. Renew. Sustain. Energy Rev. 60, 1225–1250 (2016). https://
doi.org/10.1016/j.rser.2016.02.026
32. M.A. Drewry, G.A. Georgiou, A review of NDT techniques for wind turbines. Insight-Non-
Destructive Testing and Condition Monitoring 49, 137–141 (2007). https://doi.org/10.1784/
insi.2007.49.3.137
33. H.F. Zhou et al., A review of full-scale structural testing of wind turbine blades. Renew.
Sustain. Energy Rev. 33, 177–187 (2014). https://doi.org/10.1016/j.rser.2014.01.087
34. P. Li, J. Eickmeyer, and O. Niggemann, “Data driven condition monitoring of wind power
plants using cluster analysis”, in Proc. IEEE International Conference on Cyber-Enabled
Distributed Computing and Knowledge Discovery, 2015. https://doi.org/10.1109/cyberc.
2015.16
35. J.P. Salameh et al., Gearbox condition monitoring in wind turbines: A review. Mech. Syst.
Signal Process. 111, 251–264 (2018). https://doi.org/10.1016/j.ymssp.2018.03.052
36. J. Carroll, A. McDonald, D. McMillan, Reliability comparison of wind turbines with DFIG
and PMG drive trains. IEEE Trans. Energy Convers. 30(2), 663–670 (2014). https://doi.org/
10.1109/TEC.2014.2367243
37. Y.B. Qu, H.H. Song, Energy-based coordinated control of wind energy conversion system
with DFIG. Int. J. Control 84(12), 2035–2045 (2011). https://doi.org/10.1080/00207179.2011.
631588
38. Tyler Stehly et al., “2017 Cost of Wind Energy Review”, Technical report, National Renewable
Energy Laboratory, 2017. <https://www.nrel.gov/docs/fy18osti/72167.pdf>
39. Y. Sinha, J.A. Steel, A progressive study into offshore wind farm maintenance optimisation
using risk based failure analysis. Renew. Sustain. Energy Rev. 42, 735–742 (2015). https://
doi.org/10.1016/j.rser.2014.10.087
40. I. Dinwoodie, et al., “Analysis of offshore wind turbine operation and maintenance using a
novel time domain meteo-ocean modeling approach”, Proceedings of the ASME Turbo Expo.
6. https://doi.org/10.1115/gt2012-68985, 2012. https://doi.org/10.1115/gt2012-68985
41. Z. Tian et al., Condition based maintenance optimization for wind power generation systems
under continuous monitoring. Renewable Energy 36(5), 1502–1509 (2011). https://doi.org/
10.1016/j.renene.2010.10.028
42. W. Qiao, P. Zhang, M.Y. Chow, Condition monitoring, diagnosis, prognosis, and health
management for wind energy conversion systems. IEEE Trans. Industr. Electron. 62(10),
6533–6535 (2015). https://doi.org/10.1109/TIE.2015.2464785
43. S. Sheng, Improving component reliability through performance and condition monitor-
ing data analysis”, NREL-Wind Farm Data Management & Analysis North America, 2015.
<https://www.nrel.gov/docs/fy15osti/64027.pdf>
44. B. Maples et al., Installation, operation, and maintenance strategies to reduce the cost of
offshore wind energy. Technical Report of NREL (2013). https://www.nrel.gov/docs/fy13osti/
57403.pdf>
45. Summary of wind turbine accident data to 31 December 2013. Technical report, Caithness
Windfarm Information Forum, Accessed on 10-08-2019 at http://www.caithnesswindfarms.
co.uk/AccidentStatistics.htm
46. GCube Top 5 US Wind Energy Insurance Claims Report (2013). http://www.gcube-insurance.
com/news/gcube-top-5-us-wind-energy-insurance-claims-report/>
47. Christopher A. Walford, Wind turbine reliability: understanding and minimizing wind turbine
operation and maintenance costs. Sandia Report, Sandia National Laboratories (2006). https://
energy.sandia.gov/wp-content/gallery/uploads/SAND-2006-1100.pdf
48. J. Nilsson, L. Bertling, Maintenance management of wind power systems using condition mon-
itoring systems—life cycle cost analysis for two case studies. IEEE Trans. Energy Convers.
22(1), 223–229 (2007). https://doi.org/10.1109/TEC.2006.889623
146 T. Mahto et al.

49. F. Besnard, L. Bertling, An approach for condition-based maintenance optimization applied

to wind turbine blades. IEEE Trans. Sustain. Energy 1(2), 77–83 (2010). https://doi.org/10.
1109/TSTE.2010.2049452
50. J.V. Dam, L. Bond, Economics of online structural health monitoring of wind turbines: Cost
benefit analysis, in AIP Conference Proceedings, vol. 1650, no. 1. AIP (2015). https://doi.
org/10.1063/1.4914695
51. SmartScan: Optical Blade Loads Monitoring For Wind Turbines. Technical Report, Smart
Fibres; 2010. https://www.smartfibres.com/products/smartscan
52. L.W.M.M. Rademakers, T.W. Verbruggen, Fibre optic blade load monitoring (FOBM), Tech-
nical report, Energy Research Centre of the Netherlands (2011). https://fbgs.com/wp-content/
uploads/2019/03/ECN_Flyer-b-11-0081_pdf.pdf
53. L.W.M.M. Rademakers, Fibre optic load monitoring of wind turbines, Technical report,
Energy Research Centre of the Netherlands (2012). https://fbgs.com/wp-content/uploads/
2019/03/Leaflet_FOBM_WTB_2012-11-30.pdf
54. C.R. Farrar, K. Worden, An introduction to structural health monitoring. Philos. Trans. Royal
Society A Math. Phys. Eng. Sci. 365(1851), 303–315 (2006). https://doi.org/10.1098/rsta.
2006.1928
55. M.C. Homola, P.J. Nicklasson, P.A. Sundsbo, Ice sensors for wind turbines. Cold Reg. Sci.
Technol. 46(2), 125–131 (2006). https://doi.org/10.1016/j.coldregions.2006.06.005
56. R.J. Barthelmie et al., Modelling and measurements of power losses and turbulence intensity
in wind turbine wakes at Middelgrunden offshore wind farm. Wind Energy Int. J. Prog. Appl.
Wind Power Conversion Technol. 10(6), 517–528 (2007). https://doi.org/10.1002/we.238
57. R. May et al., Mitigating wind-turbine induced avian mortality: sensory, aerodynamic and
cognitive constraints and options. Renew. Sustain. Energy Rev. 42, 170–181 (2015). https://
doi.org/10.1016/j.rser.2014.10.002
58. O. Guillermin, Developing composite wind blades that will stand the test of time (2011).
Accessed on 10-08-2019 http://winncad.blogspot.com/2011/02/developing-composite-wind-
blades-that.html
59. B. Lu et al., A review of recent advances in wind turbine condition monitoring and fault
diagnosis. Power Electron. Mach. Wind Appl. (2009). https://doi.org/10.1109/PEMWA.2009.
5208325
60. H. Malik, Wavelet and Hilbert huang transform based wind turbine imbalance fault classifi-
cation model using k-nearest neighbor algorithm (in Press). Int. J. Renew. Energy Technol.
9(1/2) (2018). https://doi.org/10.1504/IJRET.2018.090105
61. Managing the wind, Reducing kilowatt-hour costs with condition monitoring. Refocus 6(3),
48–51 (2005). https://doi.org/10.1016/S1471-0846(05)70402-9
62. K. Schroeder et al., A fibre Bragg grating sensor system monitors operational load in a wind
turbine rotor blade. Meas. Sci. Technol. 17(5), 1167–1172 (2006). https://doi.org/10.1088/
0957-0233/17/5/S39TI
63. M.J. Blanch, A.G. Dutton, Acoustic emission monitoring of field tests of an operating wind
turbine, damage assessment of structures, in 5th International Conference on Damage Assess-
ment of Structures (DAMAS-2003), Southampton, UK (2003). Doi:https://doi.org/10.4028/
www.scientific.net/KEM.245-246
64. H. Chengbing, F. Xinxin, Institutions function and failure statistic and analysis of wind turbine.
Phys. Procedia 24, 25–30 (2012). https://doi.org/10.1016/j.phpro.2012.02.005
65. D. Papasalouros et al. “Acoustic Emission Monitoring of Composite Blade of NM48/750
NEG-MICON Wind Turbine”, Journal of Acoustic Emission, vol. 31, no.1, 2013. < https://
www.ndt.net/article/jae/papers/31-036.pdf>
66. C.J. Crabtree, D. Zappalá, P.J. Tavner, Survey of commercially available condition monitoring
systems for wind turbines (2014). http://dro.dur.ac.uk/12497/
67. W. Yang et al., Wind turbine condition monitoring: technical and commercial challenges.
Wind Energy 17(5), 673–693 (2014). https://doi.org/10.1002/we.1508
68. J.C. Berg, et al., SMART wind turbine rotor: design and field test, Sandia Report SAND2014-
0681, Sandia National Laboratories, Albuquerque, NM (2014). https://pdfs.semanticscholar.
org/13de/468c4521e8e67b799e488c11923bfb934007.pdf
Condition Monitoring, and Fault Detection … 147

69. H. Malik, S. Mishra, Artificial neural network and empirical mode decomposition based
imbalance fault diagnosis of wind turbine using TurbSim, FAST and Simulink. IET Renew.
Power Gener. 11(6), 889–902 (2016). https://doi.org/10.1049/iet-rpg.2015.0382
70. J. Zeng et al., Wind turbine fault detection and isolation using support vector machine and a
residual-based method, in IEEE Proceedings of. American Control Conference (2013). https://
doi.org/10.1109/ACC.2013.6580398
71. J. Helander et al., Compressive sensing techniques for mm-wave nondestructive testing of
composite panels. Trans. Antennas Propagat. 65(10), 5523–5531 (2017). https://doi.org/10.
1109/TAP.2017.2738034
72. W.Q. Jeffries, J.A. Chambers, D.G. Infield, Experience with bicoherence of electrical power
for condition monitoring of wind turbine blades. Proc. Vision Image Signal Process. 145(3),
141–148 (1998). https://doi.org/10.1049/ip-vis:19982013
73. M.M. Shokrieh, R. Rafiee, Simulation of fatigue failure in a full composite wind turbine blade.
Compos. Struct. 74(3), 332–342 (2006). https://doi.org/10.1016/j.compstruct.2005.04.027
74. F. Song, Y. Ni, Z. Tan, Optimization design, modeling and dynamic analysis for compos-
ite wind turbine blade. Procedia Engineering 16, 369–375 (2011). https://doi.org/10.1016/j.
proeng.2011.08.1097
75. A. Jüngert, Damage detection in wind turbine blades using two different acoustic techniques.
NDT Database J. (NDT) (2008). https://www.ndt.net/article/v13n12/juengert.pdf
76. X. Gong, W. Qiao, Simulation investigation of wind turbine imbalance faults, in Proceedings
of IEEE International Conference on Power System Technology (2010). https://doi.org/10.
1109/powercon.2010.5666455
77. A. Kusiak, A. Verma, A data-driven approach for monitoring blade pitch faults in wind
turbines. IEEE Transactions on Sustainable Energy 2(1), 87–96 (2010). https://doi.org/10.
1109/TSTE.2010.2066585
78. N. Laouti, N. Sheibat-Othman, S. Othman, Support vector machines for fault detection in wind
turbines. IFAC Proceedings 44(1), 7067–7072 (2011). https://doi.org/10.3182/20110828-6-
IT-1002.02560
79. A. Kusiak, A. Verma, A data-mining approach to monitoring wind turbines. IEEE Transactions
on Sustainable Energy 3(1), 150–157 (2011). https://doi.org/10.1109/TSTE.2011.2163177
80. A. Abouhnik, A. Albarbar, Wind turbine blades condition assessment based on vibration
measurements and the level of an empirically decomposed feature. Energy Convers. Manag.
64, 606–613 (2012). https://doi.org/10.1016/j.enconman.2012.06.008
81. P. Malhotra et al., A review and design study of blade testing systems for utility-scale wind
turbines. Renew. Sustain. Energy Rev. 16(1), 284–292 (2012). https://doi.org/10.1016/j.rser.
2011.07.154
82. U.I.K Galappaththi et al., Analysis of the effect of defects for fatigue life of composite wind
turbine blades using FEM data, in International Conference on Renewable Energies and
Power Quality, March. 2012. http://www.icrepq.com/icrepq%2712/845-galappaththi.pdf
83. A. Sarkar, D.K. Behera, Wind turbine blade efficiency and power calculation with electrical
analogy. Int. J. Scientif. Res. Publ. 2(2), 1–5 (2012). http://www.ijsrp.org/research_paper_
feb2012/ijsrp-feb-2012-06.pdf
84. A.H. Almukhtar, Effect of drag on the performance for an efficient wind turbine blade design.
Energy Procedia 18, 404–415 (2012). https://doi.org/10.1016/j.egypro.2012.05.052
85. A. Staino, B. Basu, S.R.K. Nielsen, Actuator control of edgewise vibrations in wind turbine
blades. J. Sound Vib. 331(6), 1233–1256 (2012). https://doi.org/10.1016/j.jsv.2011.11.003
86. B.M. Radičević et al., Impact of wind turbine blade rotation on the lightning strike incidence–
A theoretical and experimental study using a reduced-size model. Energy 45(1), 644–654
(2012). https://doi.org/10.1016/j.energy.2012.07.032
87. M.A. Sayed, H.A. Kandil, A. Shaltot, Aerodynamic analysis of different wind-turbine-blade
profiles using finite-volume method. Energy Convers. Manag. 64, 541–550 (2012). https://
doi.org/10.1016/j.enconman.2012.05.030
88. P.J. Schubel, R.J. Crossley, Wind turbine blade design. Energies, 5(9), 3425–3449 (2012).
<https://www.mdpi.com/1996-1073/5/9/3425/pdf>
148 T. Mahto et al.

89. Y. Bazilevs et al., A computational procedure for prebending of wind turbine blades. Int. J.
Numer. Meth. Eng. 89(3), 323–336 (2012). https://doi.org/10.1002/nme.3244
90. S.A. Frost, K. Goebel, L. Obrecht, Integrating structural health management with contingency
control for wind turbines. Int. J. Prognostics Health Manage. 4(9), 11–20 (2013). https://www.
phmsociety.org/node/984>
91. J.L. Godwin, P. Matthews, Classification and detection of wind turbine pitch faults through
SCADA data analysis. IJPHM Special Issue on Wind Turbine PHM 90 (2013). https://www.
phmsociety.org/node/982
92. D.D. Chase et al., Detection of damage in operating wind turbines by signature distances. Int.
J. Prognostics Health Manage. 4(2), 101–114 (2013). https://www.phmsociety.org/node/981
93. S. Yokoyama, Lightning protection of wind turbine blades. Electric Power Syst. Res. 94, 3–9
(2013). https://doi.org/10.1016/j.epsr.2012.07.017
94. Y. Li et al., Dynamic characteristics of lag vibration of a wind turbine blade. Acta Mech.
Solida Sin. 26(6), 592–602 (2013). https://doi.org/10.1016/S0894-9166(14)60004-5
95. P.J. Schubel et al., Review of structural health and cure monitoring techniques for large wind
turbine blades. Renewable Energy 51, 113–123 (2013). https://doi.org/10.1016/j.renene.2012.
08.072
96. N. Buckney et al., Structural efficiency of a wind turbine blade. Thin-Walled Structures 67,
144–154 (2013). https://doi.org/10.1016/j.tws.2013.02.010
97. B. Chen, P.C. Matthews, P.J. Tavner, Wind turbine pitch faults prognosis using a-priori
knowledge-based ANFIS. Expert Syst. Appl. 40(17), 6863–6876 (2013). https://doi.org/10.
1016/j.eswa.2013.06.018
98. S. Ataya, M.M.Z. Ahmed, Damages of wind turbine blade trailing edge: forms, location, and
root causes. Eng. Fail. Anal. 35, 480–488 (2013). https://doi.org/10.1016/j.engfailanal.2013.
05.011
99. J.S. Chou et al., Failure analysis of wind turbine blade under critical wind loads. Eng. Fail.
Anal. 27, 99–118 (2013). https://doi.org/10.1016/j.engfailanal.2012.08.002
100. E.E. Bachynski et al., Dynamic analysis of floating wind turbines during pitch actuator fault,
grid loss, and shutdown. Energy Procedia 35, 210–222 (2013). https://doi.org/10.1016/j.
egypro.2013.07.174
101. B. Yang, D. Sun, Testing, inspecting and monitoring technologies for wind turbine blades: a
survey. Renew. Sustain. Energy Rev. 22, 515–526 (2013). https://doi.org/10.1016/j.rser.2012.
12.056
102. G. Bedon, M.R. Castelli, E. Benini, Optimization of a Darrieus vertical-axis wind turbine
using blade element–momentum theory and evolutionary algorithm. Renewable Energy 59,
184–192 (2013). https://doi.org/10.1016/j.renene.2013.03.023
103. X. Liu, L. Wang, X. Tang, Optimized linearization of chord and twist angle profiles for fixed-
pitch fixed-speed wind turbine blades. Renewable Energy 57, 111–119 (2013). https://doi.
org/10.1016/j.renene.2013.01.036
104. D. Petković, Ž. Ćojbašič, V. Nikolić, Adaptive neuro-fuzzy approach for wind turbine power
coefficient estimation. Renew. Sustain. Energy Rev. 28, 191–195 (2013). https://doi.org/10.
1016/j.rser.2013.07.049
105. N. Dervilis et al., On damage diagnosis for a wind turbine blade using pattern recognition. J.
Sound Vib. 333(6), 1833–1850 (2014). https://doi.org/10.1016/j.jsv.2013.11.015
106. S. Kim et al., Crack detection technique for operating wind turbine blades using vibro-
acoustic modulation. Struct. Health Monitor. 13(6), 660–670 (2014). https://doi.org/10.1177/
1475921714553732
107. O.M. Bouzid et al., Structural health monitoring of wind turbine blades: acoustic source
localization using wireless sensor networks. J. Sensors 2015, 1–11 (2014). https://doi.org/10.
1155/2015/139695
108. C. Niezrecki et al., Wind turbine blade health monitoring using acoustic beamform-
ing techniques. J. Acoust. Soc. Am. 135(4), 2392–2393 (2014). https://doi.org/10.1121/1.
4877915
Condition Monitoring, and Fault Detection … 149

109. Y.M. Ji, K.S. Han, Fracture mechanics approach for failure of adhesive joints in wind turbine
blades. Renewable Energy 65, 23–28 (2014). https://doi.org/10.1016/j.renene.2013.07.004
110. Y. Wang, M. Liang, J. Xiang, Damage detection method for wind turbine blades based on
dynamics analysis and mode shape difference curvature information. Mech. Syst. Signal
Process. 48(1–2), 351–367 (2014). https://doi.org/10.1016/j.ymssp.2014.03.006
111. J. Zangenberg, P. Brøndsted, M. Koefoed, Design of a fibrous composite preform for wind
turbine rotor blades. Mater. Des. 56, 635–641 (2014). https://doi.org/10.1016/j.matdes.2013.
11.036
112. X. Chen et al., Preliminary failure investigation of a 52.3 m glass/epoxy composite wind
turbine blade. Eng. Fail. Anal. 44, 345–350 (2014). https://doi.org/10.1016/j.engfailanal.2014.
05.024
113. Y. Li et al., A wind tunnel experimental study of icing on wind turbine blade airfoil. Energy
Convers. Manag. 85, 591–595 (2014). https://doi.org/10.1016/j.enconman.2014.05.026
114. F. Agnese, F. Scarpa, Macro-composites with star-shaped inclusions for vibration damp-
ing in wind turbine blades. Compos. Struct. 108, 978–986 (2014). https://doi.org/10.1016/j.
compstruct.2013.10.030
115. A. Kumar et al., Free vibration analysis of Al 2024 wind turbine blade designed for Uttarak-
hand region based on FEA. Procedia Technol. 14, 336–347 (2014). https://doi.org/10.1016/j.
protcy.2014.08.044
116. H. Pollayi, W. Yu, Modeling matrix cracking in composite rotor blades within VABS
framework. Compos. Struct. 110, 62–76 (2014). https://doi.org/10.1016/j.compstruct.2013.
11.012
117. P. Roth-Johnson, R.E. Wirz, E. Lin, Structural design of spars for 100-m biplane wind turbine
blades. Renewable Energy 71, 133–155 (2014). https://doi.org/10.1016/j.renene.2014.05.030
118. P. Pratumnopharat, P.S. Leung, R.S. Court, Wavelet transform-based stress-time history edit-
ing of horizontal axis wind turbine blades. Renewable Energy 63, 558–575 (2014). https://
doi.org/10.1016/j.renene.2013.10.017
119. C.G. Gebhardt, B.A. Roccia, Non-linear aeroelasticity: an approach to compute the response of
three-blade large-scale horizontal-axis wind turbines. Renewable Energy 66, 495–514 (2014).
https://doi.org/10.1016/j.renene.2013.12.040
120. C.V. der Woude, S, Narasimhan, A study on vibration isolation for wind turbine structures.
Eng. Struct. 60, 223–234 (2014). https://doi.org/10.1016/j.engstruct.2013.12.028
121. J. Zhang et al., The influence of wind shear on vibration of geometrically nonlinear wind
turbine blade under fluid–structure interaction. Ocean Eng. 84, 14–19 (2014). https://doi.org/
10.1016/j.oceaneng.2014.03.017
122. H. Badihi, Y. Zhang, H. Hong, Fuzzy gain-scheduled active fault-tolerant control of a wind
turbine. J. Franklin Inst. 351(7), 3677–3706 (2014). https://doi.org/10.1016/j.jfranklin.2013.
05.007
123. A.M. Abd-Elhady, N.A. Sabiha, M.A. Izzularab, Experimental evaluation of air-termination
systems for wind turbine blades. Electr. Power Syst. Res. 107, 133–143 (2014). https://doi.
org/10.1016/j.epsr.2013.10.002
124. U.E.Y. Gómez et al., Design and manufacturing of wind turbine blades of low capacity using
CAD/CAM techniques and composite materials. Energy Procedia 57, 682–690 (2014). https://
doi.org/10.1016/j.egypro.2014.10.223
125. S. Shamshirband et al., Support vector regression methodology for wind turbine reaction
torque prediction with power-split hydrostatic continuous variable transmission. Energy 67,
623–630 (2014). https://doi.org/10.1016/j.energy.2014.01.111
126. P. Poozesh et al. Full field inspection of a utility scale wind turbine blade using digital image
correlation, in Proceedings of Industrial and Commercial Applications of Smart Structures
Technologies, Vol. 7979, pp. 2891–2960 (2014). https://doi.org/10.1117/12.879930
127. M. Mollineaux et al., Damage detection methods on wind turbine blade testing with wired
and wireless accelerometer sensors, in European Workshop on Structural Health Monitoring
(2014). https://hal.inria.fr/hal-01022037
150 T. Mahto et al.

128. H.M. Slot et al., Leading edge erosion of coated wind turbine blades: Review of coating
life models. Renewable Energy 80, 837–848 (2015). https://doi.org/10.1016/j.renene.2015.
02.036
129. S. Zhang et al., Erosion of wind turbine blade coatings–design and analysis of jet-based lab-
oratory equipment for performance evaluation. Prog. Org. Coat. 78, 103–115 (2015). https://
doi.org/10.1016/j.porgcoat.2014.09.016
130. O. Bitkina, K.W. Kang, J.H. Lee, Experimental and theoretical analysis of the stress–strain
state of anisotropic multilayer composite panels for wind turbine blade. Renewable Energy
79, 219–226 (2015). https://doi.org/10.1016/j.renene.2014.11.004
131. P. Rajkumar et al., Increasing the impact testing of wind turbine blade with wire string. Int.
J. Appl. Eng. Res. 10(2), 1730–1732 (2015)
132. A.C. Benim, M. Diederich, M. Nikbay, Optimization of airfoil profiles for small wind turbines,
in 8th ICCHMT, Istanbul (2015). https://www.researchgate.net/publication/277953865_
OPTIMIZATION_OF_AIRFOIL_PROFILES_FOR_SMALL_WIND_TURBINES
133. O. Al-Khudairi, H. Ghasemnejad, To improve failure resistance in joint design of composite
wind turbine blade materials. Renew. Energy 81, 936–951 (2015). https://doi.org/10.1016/j.
renene.2015.04.015
134. J.K. Lee et al., Transformation algorithm of wind turbine blade moment signals for blade
condition monitoring. Renew. Energy 79, 209–218 (2015). https://doi.org/10.1016/j.renene.
2014.11.030
135. H. Habibi et al., A dual de-icing system for wind turbine blades combining high-power
ultrasonic guided waves and low-frequency forced vibrations. Renew. Energy 83, 859–870
(2015). https://doi.org/10.1016/j.renene.2015.05.025
136. M.M. Rezaei et al., Development of a reduced order model for nonlinear analysis of the wind
turbine blade dynamics. Renew. Energy 76, 264–282 (2015). https://doi.org/10.1016/j.renene.
2014.11.021
137. X. Li et al., Crack growth sparse pursuit for wind turbine blade. Smart Mater. Struct. 24(1),
1–8 (2014). https://doi.org/10.1088/0964-1726/24/1/015002
138. W. Xie, P. Zeng, L. Lei, A novel folding blade of wind turbine rotor for effective power control.
Energy Convers. Manag. 101, 52–65 (2015). https://doi.org/10.1016/j.enconman.2015.05.037
139. R.S. kumar, D. Jena, Validation of an integral sliding mode control for optimal control of a
three blade variable speed variable pitch wind turbine. Int. J. Electr. Power Energy Syst. 69,
421–429 (2015). https://doi.org/10.1016/j.ijepes.2015.01.031
140. C. Liu, D. Jiang, F. Chu, Influence of alternating loads on nonlinear vibration characteristics
of cracked blade in rotor system. J. Sound Vib. 353, 205–219 (2015). https://doi.org/10.1016/
j.jsv.2015.05.007
141. S. Zhang et al., Rain erosion of wind turbine blade coatings using discrete water jets: Effects
of water cushioning, substrate geometry, impact distance, and coating properties. Wear 328,
140–148 (2015). https://doi.org/10.1016/j.wear.2015.01.079
142. S. Hoell, P. Omenzetter, Structural damage detection in wind turbine blades based on time
series representations of dynamic responses, Smart Materials and Nondestructive Evaluation
for Energy Systems, 9439 (2015). https://doi.org/10.1117/12/2083533
143. P.U. Haselbach, R.D. Bitsche, K. Branner, The effect of delaminations on local buckling in
wind turbine blades. Renewable Energy 85, 295–305 (2016). https://doi.org/10.1016/j.renene.
2015.06.053
144. R. Sharma et al., Feature extraction using EMD and classifier through artificial neural networks
for gearbox fault diagnosis, Book chapter in Applications of Artificial Intelligence Techniques
in Engineering, Advances in Intelligent Systems and Computing, vol. 697 (2018),pp. 309–317.
https://doi.org/10.1007/978-981-13-1822-1_28
145. S. Sheng, J. Keller, C. Glinsky, Gearbox reliability collaborative update, NREL,
National Renewable Energy Laboratory (2013) https://digital.library.unt.edu/ark:/67531/
metadc831120/
146. T. Igarashi, H. Hamada, Studies on the vibration and sound of defective rolling bearings:
First report: Vibration of ball bearings with one defect. Bulletin of JSME 25(204), 994–1001
(1982). https://doi.org/10.1299/jsme1958.25.994
Condition Monitoring, and Fault Detection … 151

147. J.F. Hall, D. Chen, Performance of a 100 kW wind turbine with a Variable Ratio Gearbox.
Renewable Energy 44, 261–266 (2012). https://doi.org/10.1016/j.renene.2012.01.094
148. Z. Feng, M. Liang, Complex signal analysis for wind turbine planetary gearbox fault diagnosis
via iterative atomic decomposition thresholding. J. Sound Vib. 333(20), 5196–5211 (2014).
https://doi.org/10.1016/j.jsv.2014.05.029
149. B. Tang et al., Fault diagnosis for a wind turbine transmission system based on manifold
learning and Shannon wavelet support vector machine. Renewable Energy 62, 1–9 (2014).
https://doi.org/10.1016/j.renene.2013.06.025
150. A. Zhou, D. Yu, W. Zhang, A research on intelligent fault diagnosis of wind turbines based
on ontology and FMECA. Adv. Eng. Inform. 29(1), 115–125 (2015). https://doi.org/10.1016/
j.aei.2014.10.001
151. L. Hong, J.S. Dhupia, S. Sheng, An explanation of frequency features enabling detection of
faults in equally spaced planetary gearbox. Mech. Mach. Theory 73, 169–183 (2014). https://
doi.org/10.1016/j.mechmachtheory.2013.10.014
152. L. Hong, J.S. Dhupia, A time domain approach to diagnose gearbox fault based on measured
vibration signals. J. Sound Vib. 333(7), 2164–2180 (2014). https://doi.org/10.1016/j.jsv.2013.
11.033
153. Y. Lei et al., Fault detection of planetary gearboxes using new diagnostic parameters. Meas.
Sci. Technol. 23(5), 1–11 (2012). https://doi.org/10.1088/0957-0233/23/5/055605
154. D. Zappalá et al., Side-band algorithm for automatic wind turbine gearbox fault detection and
diagnosis. IET Renew. Power Gener. 8(4), 380–389 (2014). https://doi.org/10.1049/iet-rpg.
2013.0177
155. M. Mosher, Understanding vibration spectra of planetary gear systems for fault detection, in
International Design Engineering Technical Conferences and Computers and Information in
Engineering Conference. American Society of Mechanical Engineers (2003). https://doi.org/
10.1115/DETC2003/PTG-48082
156. F. Chaari, T. Fakhfakh, M. Haddar, Analytical investigation on the effect of gear teeth faults
on the dynamic response of a planetary gear set. Noise Vibration Worldwide 37(8), 9–17
(2006). https://doi.org/10.1260/095745606778600857
157. Z. Feng, M.J. Zuo, Vibration signal models for fault diagnosis of planetary gearboxes. J.
Sound Vib. 331(22), 4919–4939 (2012). https://doi.org/10.1016/j.jsv.2012.05.039
158. W. Bartelmus, R. Zimroz, Vibration condition monitoring of planetary gearbox under varying
external load. Mech. Syst. Signal Process. 23(1), 246–257 (2009). https://doi.org/10.1016/j.
ymssp.2008.03.016
159. T. Barszcz, R.B. Randall, Application of spectral kurtosis for detection of a tooth crack in
the planetary gear of a wind turbine. Mech. Syst. Signal Process. 23(4), 1352–1365 (2009).
https://doi.org/10.1016/j.ymssp.2008.07.019
160. Y. Lei et al., Planetary gearbox fault diagnosis using an adaptive stochastic resonance method.
Mech. Syst. Signal Process. 38(1), 113–124 (2013). https://doi.org/10.1016/j.ymssp.2012.
06.021
161. Z. Feng et al., Fault diagnosis for wind turbine planetary gearboxes via demodulation analysis
based on ensemble empirical mode decomposition and energy separation. Renewable Energy
47, 112–126 (2012). https://doi.org/10.1016/j.renene.2012.04.019
162. Z. Feng, M. Liang, Fault diagnosis of wind turbine planetary gearbox under nonstationary
conditions via adaptive optimal kernel time–frequency analysis. Renew. Energy 66, 468–477
(2014). https://doi.org/10.1016/j.renene.2013.12.047
163. X. Yao et al. Wind turbine gearbox fault diagnosis using adaptive morlet wavelet spectrum,
in International Conference on Intelligent Computation Technology and Automation, vol. 2
(2009). https://doi.org/10.1109/icicta.2009.375
164. R. Dupuis, Application of oil debris monitoring for wind turbine gearbox prognostics and
health management, in Annual Conference of the Prognostics and Health Management Society
(2010). https://www.phmsociety.org/events/conference/phm/10/proceedings/Application-of-
Oil-Debris-Monitoring-for-Wind-Turbine-Gearbox-PHM
152 T. Mahto et al.

165. P.J. Dempsey, Gear damage detection using oil debris analysis, in International Congress and
Exhibition on Condition Monitoring and Diagnostics Engineering Management, Manchester,
UK (2001). https://doi.org/10.1016/b978-008044036-1/50050-0
166. P.J. Dempsey, A.A. Afjeh, Integrating oil debris and vibration gear damage detection tech-
nologies using fuzzy logic. J. Am. Helicopter Soc. 49(2), 109–116 (2004). https://doi.org/10.
4050/JAHS.49.109
167. T. Toutountzakis, C.K. Tan, D. Mba, Application of acoustic emission to seeded gear fault
detection. NDT and E Int. 38(1), 27–36 (2005). https://doi.org/10.1016/j.ndteint.2004.06.008
168. Y. Qu et al., Gearbox tooth cut fault diagnostics using acoustic emission and vibration
sensors—a comparative study. Sensors 14(1), 1372–1393 (2014). https://doi.org/10.3390/
s140101372
169. C.M. Vicuña, Effects of operating conditions on the Acoustic Emissions (AE) from plane-
tary gearboxes. Appl. Acoust. 77, 150–158 (2014). https://doi.org/10.1016/j.apacoust.2013.
04.017
170. A.S.A.E. Zaher et al., Online wind turbine fault detection through automated SCADA data
analysis wind energy. Int. J. Prog. Appl. Wind Power Conversion Technol. 12(6), 574–593
(2009). https://doi.org/10.1002/we.319
171. K. Kim et al., Use of SCADA data for failure detection in wind turbines, No. NREL/CP-
5000-51653. National Renewable Energy Lab. (NREL), Golden, CO (United States) (2011).
https://www.nrel.gov/docs/fy12osti/51653.pdf
172. T.H. Loutas et al., The combined use of vibration, acoustic emission and oil debris on-line
monitoring towards a more effective condition monitoring of rotating machinery. Mech. Syst.
Signal Process. 25(4), 1339–1352 (2011). https://doi.org/10.1016/j.ymssp.2010.11.007
173. Y. Feng et al., Use of SCADA and CMS signals for failure detection and diagnosis of a wind
turbine gearbox, in European Wind Energy Conference and Exhibition (2011). http://eprints.
whiterose.ac.uk/83334/
174. A. Salem, A. Abu-Siada, S. Islam, Improved condition monitoring technique for wind turbine
gearbox and shaft stress detection. IET Sci. Meas. Technol. 11(4), 431–437 (2017). https://
doi.org/10.1049/iet-smt.2016.0338
175. L.F. Villa et al., Angular resampling for vibration analysis in wind turbines under non-linear
speed fluctuation. Mech. Syst. Signal Process. 25(6), 2157–2168 (2011). https://doi.org/10.
1016/j.ymssp.2011.01.022
176. W.Y. Liu et al., A new wind turbine fault diagnosis method based on the local mean
decomposition. Renew. Energy 48, 411–415 (2012). https://doi.org/10.1016/j.renene.2012.
05.018
177. Y. Qu et al., A new acoustic emission sensor based gear fault detection approach. Int. J.
Prognost. Health Manage. 4, 32–45 (2013). https://www.phmsociety.org/node/980
178. S. Hussain, H.A. Gabbar, Vibration analysis and time series prediction for wind turbine gear-
box prognostics. Int. J. Prognost. Health Manag. 4, 69–79 (2013). https://www.phmsociety.
org/node/965 >
179. A. Lesmerises, D. Crowley, Effect of different workscope strategies on wind turbine gearbox
life cycle repair costs, IJPHM Special Issue on Wind Turbine PHM, 64 (2013). https://core.
ac.uk/reader/26122960 >
180. S. Yin, G. Wang, H.R. Karimi, Data-driven design of robust fault detection system for wind
turbines. Mechatronics 24(4), 298–306 (2014). https://doi.org/10.1016/j.mechatronics.2013.
11.009
181. C. Zhu et al., Research on dynamical characteristics of wind turbine gearboxes with flexible
pins. Renew. Energy 68, 724–732 (2014). https://doi.org/10.1016/j.renene.2014.02.047
182. I. Antoniadou et al., A time–frequency analysis approach for condition monitoring of a wind
turbine gearbox under varying load conditions. Mech. Syst. Signal Process, 64, 188–216
(2015). https://doi.org/10.1016/j.ymssp.2015.03.003
183. O. Blancke et al., Strategic fault diagnosis approach for hydro generator shaft current dis-
charges, in International Conference on Electrical Machines (ICEM) (2016). pp. 2346–2351.
https://doi.org/10.1109/icelmach.2016.7732849
Condition Monitoring, and Fault Detection … 153

184. M.H. Marzebali et al., Planetary gearbox torsional vibration effects on wound-rotor induction
generator electrical signatures. IEEE Trans. Ind. Appl. 52(6), 4770–4780 (2016). https://doi.
org/10.1109/TIA.2016.2600599
185. A.B. Attya, T. Hartkopf, Penetration impact of wind farms equipped with frequency variations
ride through algorithm on power system frequency response. Int. J. Electr. Power Energy Syst.
40(1), 94–103 (2012). https://doi.org/10.1016/j.ijepes.2012.02.009
186. W. Yucai, L. Yonggang, Diagnosis of short circuit faults within turbo generator excitation
winding based on the expected electromotive force method. IEEE Trans. Energy Convers.
31(2), 706–713 (2016). https://doi.org/10.1109/TEC.2016.2521422
187. S. Djurovic et al., Condition monitoring artefacts for detecting winding faults in wind
turbine DFIGs, in European Wind Energy Conference and Exhibition (2009), pp. 491–
512. Accessed https://www.researchgate.net/profile/PJ_Tavner/publication/228553600_
Condition_Monitoring_Artefacts_for_Detecting_Winding_Faults_in_Wind_Turbine_
DFIGs/links/0fcfd50e602e8ee1ed000000/Condition-Monitoring-Artefacts-for-Detecting-
Winding-Faults-in-Wind-Turbine-DFIGs.pdf
188. M. Whittle, Wind Turbine Generator Reliability: An Exploration of the Root Causes of Gen-
erator Bearing Failures, Durham University (2013). Available at Durham E-Theses. http://
etheses.dur.ac.uk/9422/
189. P. Jayaswal, B. Agrawal, New trends in wind turbine condition monitoring system. J. Emerg.
Trends Eng. Development 3, 133–148 (2011)
190. M. Liang, H. Faghidi, Intelligent bearing fault detection by enhanced energy operator. Expert
Syst. Appl. 41(16), 7223–7234 (2014). https://doi.org/10.1016/j.eswa.2014.05.026
191. F. Dalvand et al., Time-domain bearing condition monitoring in induction motors using instan-
taneous frequency of motor voltage. Smart Grid Conference (SGC) (2014). https://doi.org/
10.1109/SGC.2014.7150708
192. R. Zimroz et al., Diagnostics of bearings in presence of strong operating conditions non-
stationarity—a procedure of load-dependent features processing with application to wind
turbine bearings. Mech. Syst. Signal Process. 46(1), 16–27 (2014). https://doi.org/10.1016/j.
ymssp.2013.09.010
193. J.B. Ali et al., Application of empirical mode decomposition and artificial neural network
for automatic bearing fault diagnosis based on vibration signals. Appl. Acoustics 89, 16–27
(2015). https://doi.org/10.1016/j.apacoust.2014.08.016
194. H. Wang, M. Liserre, F. Blaabjerg, Toward reliable power electronics: challenges design tools
and opportunities. IEEE Ind. Electron. Mag. 7(2), 17–26 (2013). https://doi.org/10.1109/MIE.
2013.2252958
195. L. Ran et al., Condition monitoring of power electronics for offshore wind. IET Eng. Technol.
Reference 1(1), 1–10 (2014). https://doi.org/10.1049/etr.2014.0004
196. J. Ribrant, L. Bertling, Survey of failures in wind power systems with focus on Swedish wind
power plants during 1997–2005. IEEE Power Eng. Soc. General Meeting (2007). https://doi.
org/10.1109/PES.2007.386112
197. Y. Lin et al., Fault analysis of wind turbines in China. Renew. Sustain. Energy Rev. 55,
482–490 (2016). https://doi.org/10.1016/j.rser.2015.10.149
198. B. Hahn, M. Durstewitz, K. Rohrig, Reliability of wind turbines. Wind Energy. Springer,
Berlin, Heidelberg, 329–332 (2007). https://doi.org/10.1007/978-3-540-33866-6_62
199. N.M.A. Freire, J.O. Estima, A.J.M. Cardoso, Open-circuit fault diagnosis in PMSG drives for
wind turbine applications. IEEE Trans. Industr. Electron. 60(9), 3957–3967 (2012). https://
doi.org/10.1109/TIE.2012.2207655
200. N.M.A. Freire, J.O. Estima, A.J.M. Cardoso, Converters fault-diagnosis in PMSG drives for
wind turbine applications. Annual Conference on IEEE Industrial Electronics Society (2010).
https://doi.org/10.1109/IECON.2010.5674992
201. Y. Song, B. Wang, Survey on reliability of power electronic systems. IEEE Trans. Power
Electron. 28(1), 591–604 (2012). https://doi.org/10.1109/TPEL.2012.2192503
202. M. Alavi, D. Wang, M. Luo, Short-circuit fault diagnosis for three-phase inverters based on
voltage-space patterns. IEEE Trans. Industr. Electron. 61(10), 5558–5569 (2012). https://doi.
org/10.1109/TIE.2013.2297298
154 T. Mahto et al.

203. S. Karimi et al., FPGA-based real-time power converter failure diagnosis for wind energy
conversion systems. IEEE Trans. Industr. Electron. 55(12), 4299–4308 (2008). https://doi.
org/10.1109/TIE.2008.2005244
204. E. Ribeiro, A.J.M. Cardoso, C. Boccaletti, Open-circuit fault diagnosis in interleaved DC–DC
converters. IEEE Trans. Power Electron. 29(6), 3091–3102 (2013). https://doi.org/10.1109/
TPEL.2013.2272381
205. T. Geyer, S. Schroder, Reliability considerations and fault-handling strategies for multi-MW
modular drive systems. IEEE Trans. Ind. Appl. 46(6), 2442–2451 (2010). https://doi.org/10.
1109/ECCE.2009.5316422
206. S. Yang et al., An industry-based survey of reliability in power electronic converters. IEEE
Trans. Ind. Appl. 47(3), 1441–1451 (2011). https://doi.org/10.1109/ECCE.2009.5316356
207. H. Wang, F. Blaabjerg, Reliability of capacitors for DC-link applications in power electronic
converters - An overview. IEEE Trans. Ind. Appl. 50(5), 3569–3578 (2014). https://doi.org/
10.1109/TIA.2014.2308357
208. M. Sami, R.J. Patton, Wind turbine power maximisation based on adaptive sensor fault tolerant
sliding mode control. Mediterranean Conference on Control & Automation (MED) (2012).
https://doi.org/10.1109/MED.2012.6265799
209. T. Kamel, Y. Biletskiy, L. Chang, Real-time diagnosis for open-circuited and unbalance faults
in electronic converters connected to residential wind systems. IEEE Trans. Industr. Electron.
63(3), 1781–1792 (2015). https://doi.org/10.1109/TIE.2015.2498903
210. T. Kamel, Y. Biletskiy, L. Chang, Fault diagnoses for industrial grid-connected converters
in the power distribution systems. IEEE Trans. Industr. Electron. 62(10), 6496–6507 (2015).
https://doi.org/10.1109/TIE.2015.2420627
211. I. Jlassi et al., Multiple open-circuit faults diagnosis in back-to-back converters of PMSG
drives for wind turbine systems. IEEE Trans. Power Electron. 30(5), 2689–2702 (2014).
https://doi.org/10.1109/TPEL.2014.2342506
212. P. Duan et al., Open-switch fault diagnosis and system reconfiguration of doubly fed wind
power converter used in a microgrid. IEEE Trans. Power Electron. 26(3), 816–821 (2010).
https://doi.org/10.1109/TPEL.2010.2095470
213. N.M.A. Freire, J.O. Estima, A.J.M. Cardoso, A new approach for current sensor fault diagnosis
in PMSG drives for wind energy conversion systems. IEEE Trans. Ind. Appl. 50(2), 1206–1214
(2013). https://doi.org/10.1109/TIA.2013.2271992
214. U.M. Choi et al., Method for detecting an open-switch fault in a grid-connected NPC inverter
system. IEEE Trans. Power Electron. 27(6), 2726–2739 (2011). https://doi.org/10.1109/
TPEL.2011.2178435
215. J.S. Lee, K.B. Lee, F. Blaabjerg, Open-switch fault detection method of a back-to-back con-
verter using NPC topology for wind turbine systems. IEEE Trans. Ind. Appl. 51(1), 325–335
(2014). https://doi.org/10.1109/TIA.2014.2327151
216. H. Toubakh, M. Sayed-Mouchaweh, Hybrid dynamic classifier for drift-like fault diagnosis in
a class of hybrid dynamic systems: Application to wind turbine converters. Neurocomputing
171, 1496–1516 (2016). https://doi.org/10.1016/j.neucom.2015.07.073
217. Y.J. Ko et al., Fault diagnosis of three-parallel voltage-source converter for a high-power wind
turbine. IET Power Electronics 5(7), 1058–1067 (2012). https://doi.org/10.1049/iet-pel.2011.
0109
218. F. Deng et al., Fault detection and localization method for modular multilevel converters.
IEEE Trans. Power Electron. 30(5), 2721–2732 (2014). https://doi.org/10.1109/TPEL.2014.
2348194
219. H. Liu et al., Online fault identification based on an adaptive observer for modular multilevel
converters applied to wind power generation systems. Energies 8(7), 7140–7160 (2015).
https://doi.org/10.3390/en8077140
220. H. Malik, S. Mishra, Application of GEP to investigate the imbalance faults in direct-drive
wind turbine using generator current signals. IET Renew. Power Gener. 12(3), 279–291 (2018).
https://doi.org/10.1049/iet-rpg.2016.0689
Condition Monitoring, Fault Detection
and Diagnosis (FDD) of Photovoltaic
System and Its Approaches

Omkar Singh, Amit Kumar Yadav and Anjan Kumar Ray

Abstract Accuracy in measurements of fault is necessary for improving the relia-

bility of PV module. This chapter provides an introduction about various techniques
dedicated to identify and categorize the faults along with different types of model-
ing. This chapter also deals with the faults and its level of severity and can affect
the working of the PV module. Different types of computational techniques have
been explained in detail. Finally, a recommendation has been presented in form of
the research gap of the study.

Keywords Faults · PV · ANN · Conventional · Renewable

1 Introduction

1.1 General Overview

The day by day increasing demand in energy is one of a major problem for the
world. Since the nonrenewable energy (RE) sources are exhausting at a faster rate,
it is necessary to optimum for new conventional forms of energy. Sun is one of the
free best sources of RE and it supplies a colossal amount of energy to the earth in
the form of solar energy. We have been extracting this solar energy from a very long
time, but still we are not successful in extracting the energy at our full potential.
Hence, it becomes a necessary task to enhance the efficiency and also the quality
of the power that is extracted from solar energy. As we use photovoltaic
 (PV) cells
to extract this energy, in order to increase the efficiency η pv and quality of the
power extracted many methods are proposed and are in use and this can be done
by either at the material level also by proper monitoring and control of the PV cell.
Incessant monitoring of PV cell’s condition will give us the information and data
regarding the faults and condition of the module, which is very helpful in reducing

O. Singh · A. K. Yadav (B) · A. K. Ray

Electrical and Electronics Engineering Department, National Institute of Technology Sikkim,
Ravangla Barfung Block, South Sikkim 737139, Sikkim, India
e-mail: [email protected]
© Springer Nature Singapore Pte Ltd. 2020 155
H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_6
156 O. Singh et al.

these abnormalities and we can achieve better output from the existing PV modules.
Solar energy is widely used across the globe to generate electricity. Solar energy
is harnessed by the use of photovoltaic modules. Thus, these PV modules can be
used in any place where we receive sunlight. Thus, these PV modules gave us a
flexibility to generate electricity in nay remote location or space, where we do not
have transmission or distribution from the grid, these PV modules can be called as
standalone power sources. As PV modules have widely penetrated in every sector
and area, may it be space exploration, electric vehicle, smart homes, street lightning,
etc. However, till date the most PV modules have efficiency less than 18% [1].
Though research is still progressing on to increase the module η pv , still most of the
advanced PV modules are having efficiency less than 22%. Most of the PV modules
are having a lifetime <25 years [2]. Most of the PV module is subjected to harsh
weather conditions, this leads to their early degradation. Therefore, it is essential to
monitor the operating condition of the PV module and recognize the faults associated
to PV modules.

2 Methodology

2.1 PV Modeling

A conventional PV model is modeled by single and Double Diode (DD). Though both
the models can be used for the study, in most of the studies single diode model is used.
For areas with low irradiance levels, double diode model is used. The additional diode
used accounts for the losses associated with minority charge carrier recombination in
the depletion zone of the semiconductor diode. However, mostly one diode model is
widely considered in the analysis. One diode model comprises five key parameters,
whereas two-diode model comprises seven parameters. Both the diode models are
widely used in PV modeling. However, two-diode model is more exact compared
to one diode model but since DD model consists of more parameters. Thus, in DD
model the calculation complexity increases. Hence in most of the literature one diode
model is used for analysis.

2.1.1 Single Diode (SD) PV Model

SD PV modeling is the most simple and widely used modeling method for PV
module. SD model contains parameters, i.e., series resistance (Rs ), shunt resistance
(Rsh ), reverse saturation current (I 0 ), ideality factor (η), photo generated current (I ph ),
as presented in Fig. 1. The equations describing SD model are given from Eqs. (1)
to (3).
Condition Monitoring, Fault Detection and Diagnosis … 157

Fig. 1 Single diode PV model

VL + I L Rs
I L = Iph − I − (1)
Rsh
 
VL + I L Rs
Id = I0 exp −1 (2)
ηVt
Ns kT
Vt = (3)
q

where
k Boltzmann’s constant.
q electron charge.
T module temperature (K).
Ns Number of cells connected in series.
Vt Thermal voltage.

2.1.2 DD PV Model

The DD model is similar to SD model, but it consists of two diodes D1 and D2.
Thus, all the parameters are same as one diode model. Only we have to calculate the
ideality factor (η2 ) and reverse saturation current (I02 ) of diode D2. The DD model is
presented in Fig. 2, and the equations for two-diode model are given in Eqs. (4)–(6)

VL + I L Rs
I L = I ph − Id1 − Id2 − (4)
Rsh
 
VL + I L Rs
Id1 = I01 exp −1 (5)
η1 Vt1
158 O. Singh et al.

Fig. 2 Double diode PV model

 
VL + I L Rs
Id2 = I02 exp −1 (6)
η2 Vt2

I 01 is reverse saturation current diode 1, I 02 is reverse saturation current diode 2, η1

is ideality factor of diode 1, η2 is ideality factor of diode 2.

2.2 Faults in PV System

A PV system may be subjected to very harsh weather conditions and if proper

maintenance and monitoring are not done, it may undergo many faults (Figs. 3,
4 and 5).
Encapsulation failure is mainly due to hot and humid conditions and many
external factors which cause delamination and discoloration between the active
cells and encapsulant the various external factors that cause these failures are salt
accumulations, moisture penetration, and environmental factors [2].

Fig. 3 Faults classification

Condition Monitoring, Fault Detection and Diagnosis … 159

Fig. 4 ANN-based fault detection technique

Fig. 5 Thermography using

infrared camera

Delamination occurs due to adhesion loss between the cells and encapsulating
polymer and between the front glass causes this fault which may lead to the following
problems [3]:
• Increased reflection of light, a significant reduction in the net solar energy utilized
by the module and leads to reduced efficiency.
• Water penetration inside the PV cell which may lead to corrosion and short circuit
faults.
160 O. Singh et al.

• When delamination occurs at the edges of the module this leads to several major
faults like electrical risks related to the module.
• Shading is due to some obstruction that prevents the solar light to an incident on
the PV module, it can be classified into two types [6]:
• Static shading—it is due to the objects that stick to the PV module surface like bird
stools, leaves and accumulated dust, these masking requires human intervention.
• Dynamic shading—it occurs due to proximate trees, establishments, etc. that cause
a reduction in the net irradiance on the PV module.
Module fault mainly comprises three faults, i.e., Micro-cracks, Snail tracks, and
Hotspots [17].
Micro-crack is always present in the PV module and is unavoidable these cracks
are introduced due to mechanical strains during lamination, welding, and transporta-
tion [18]. Initially, micro-cracks do not affect the power output much but with aging
the cracks are increased and may cause complete separation of the PV cell which
causes significant reduction in power output [19].
Snail tracks are due to creation of silver carbonate (Ag2 CO3 ) which causes fading
of the silver grid. The tracks are mainly witnessed in edges of the cells, however,
many complex elements are being used in cell constructions, due to these elements
the snail tracks may increase [23]. Peng [24] revealed that average power reduction
of 1.4% in faded modules, degradation in average power is 0.56% per year after
deducting the initial light induced.
Hotspots are also considered as degradation of PV cells because they cause
increased heating of one or more cells which may further cause adhesion loss [25].
Shading can also lead to hotspots as the covered region experiences reverse biasing
and lead to dissipation thereby releasing heat and forming hotspot.

2.3 Techniques for Fault Detection

2.3.1 Artificial Neural Networks

In this method, multilayer neural network is used to determine the fault. The param-
eter values, such as irradiance, temperature are obtained and then compared with the
real measured values to determine any type of fault since neural networks are used
computation is very fast and is accurate thus the response being fast, the fault can be
detected and cleared within a very small time [2].

2.3.2 Thermography to Detect Failures Due to Environmental Factors

These faults can be detected by the following processes [2]:

• X-ray tomography.
• Thermography
Condition Monitoring, Fault Detection and Diagnosis … 161

Fig. 6 Thermography of healthy PV module

Fig. 7 Thermography of unhealthy PV module

• Ultrasonic scanners.
In literature [4] the thermography is done using an infrared camera which can
sense the module(s) surface temperature (Figs. 6 and 7).
From Fig. 4 it is seen that PV module operating temperature is maintained and
under the standard test conditions (STC), the PV conditions are captured by an
infrared camera which can sense the heat radiations from the module and the obtained
results are then studied.
Figure 5 shows the thermography results of a healthy PV module from which it
can be clearly seen that there are almost negligible cold spots which indicate that the
module is working properly.
It can be seen from Fig. 6 that there are many cold spots that indicate the PV model
is not working properly and suffers from delamination causing reduced power output.
Discoloration also results due to loss of Ethylene Vinyl Acetate (EVA), it causes
a variation in PV module material color (mostly yellow or brown), which changes
the light-capturing capacity of the cell surface [2].
162 O. Singh et al.

2.3.3 Detection of Shading Faults and Array Faults

In literature [6] the authors have proposed a method in which under all irradiation
conditions the three parameters are studied array current, array voltage, irradiance.
Based on the above parameters they have given three status to the PV module:
1. Normal operation.
2. Partial Shading.
3. Fault.
A parameter is given which is given as

VPV × IPV
γ = (7)
G
where
VPV is instant voltage of array.
IPV is instant current of array.
G is the irradiance at any instant, on unshaded portion of array.
S.I unit of is m2
Array losses (L array ) is the variance between instantaneous expected power and
the actual power.
  
G
L array = Pm − (VPV × IPV ) (8)
G0

by obtaining the ratio the author has given various threshold values from which we
can determine the various faults like line to line faults, mismatch faults, etc (Tables 1
and 2).
As a consequence of partial shading, the power productivity of PV module reduces
from the peak power value [7]. Partial shading causes more than one peak values
in characteristics of the PV module so global maximum point is obtained in the
literature [8–10] or by changing the array arrangement [11, 12].

Table 1 Condition of PV
Sl no. Larray (W) Condition
module
1. <50 Normal operation
2. 50–100 Partial shading
3. >100 Faults

Table 2 Array fault detection

Sl no. γ (m2 ) Fault
1 0.2 Line to line (one module mismatch)
2 0.31 Line to line (two modules mismatch)
Condition Monitoring, Fault Detection and Diagnosis … 163

In [7] authors have obtained the deviation of actual values from the expected
values, However, the onset values of deviation are not discussed, The onset values are
given in [13] but the short circuit and shading faults are not segregated. In literature
[14] the authors have used a decision tree algorithm to detect the faults and used
experimental data like array operating temperature, array voltage, irradiance, and
operating voltage are used as training sets. But including all faults is difficult in this
algorithm. ANN with three layers is proposed in [15] to detect the faults in PV module
along with shading faults. In literature [16] the authors have proposed a Probabilistic
Neural Network model to obtain the faults and it does not need the onset values,
however, this method does not include faults due to partial shading.

2.3.4 Detection of Module Faults

Macro Paggi gave a mathematical approach to analyze micro-cracks in

[20]. M. Kontges gave the threshold value of micro-cracks on a sole PV cell, the
condition is given by the Eq. (9):

ainactive i SC − iMPP
< (9)
atotal i SC

where atotal = area of solar cell, aactive = defective cell’s active area, iMPP = current
at maximum power point, i SC = short circuit current. Ainactive can be around 8%
of the total area and can cause significant power loss [22]. Shading can also lead to
hotspots as the shaded region becomes reverse biased and causes dissipation in the
form of heat and hotspot is created [26] shown in Fig. 5. The shading rate at which
hotspot is very severe is given by the condition:

IMP − I100
1− × 100 (10)
ISC

where
IMP current at maximum power without any cell shading.
I100 break point current at 100% shading.
ISC current at short circuit condition.

2.3.5 Neuro-Fuzzy Classifier for Detection of Fault

Reliability of PV module increases by detection of faults and classifying them. Most

of the faults are obtained by MPPT tracking and when maximum power point is
violated, it shows the occurrence of a fault [31]. PV modules are equipped with
series overcurrent protection devices to detect overcurrent faults [32]. But faults like
shading [33] and resistance increase [34] are not accounted as the PV characteristics
164 O. Singh et al.

Fig. 8 Fault fuzzy classifier

are nonlinear [35, 36]. In literature [37] observing and remote fault detection done by
satellite observation are done. In [38] the authors have used Kalman filter to detect
reduction of power in PV module. A fault detection rule is proposed in [39]. Loss
of power is used as a detection tool in [13, 40]. Several algorithms implementing
intelligent techniques are proposed with statistical approach [42–45]. A clustering
method of classification is proposed in [41]. BNN and Regression polynomial is
used in [43]. In [44] GB-SSL method is used to detect and classify faults. To design
controllers fuzzy logic is implemented in [45] and which is further used to design
classifier [46]. Neuro-fuzzy classifier is developed in [30] to identify and categorize
faults, these faults here are mainly bypass diode faults, series losses, and blocking
diode faults (Fig. 8).

3 Identified Research Gaps

1. PV performance is highly dependent on environmental conditions, but most of the

available literature are restricted to a given location. Hence, we need a generalized
technique to identify fault at any required location.
2. Most of the literature are limited to some specific faults, i.e., maximum of seven
faults are addressed. Here more faults can be taken into account.
3. Most of the literature focus on post fault analysis, but a real-time fault detection
method can be proposed.

4 Conclusions

In this chapter, a review on different methods of condition monitoring (CM) of PV

systems is presented. The CM is necessary to ensure and enhance the efficiency,
security, and dependability of PV modules. Since due to stationary parts the mainte-
nance is easy, however, modules are exposed to harsh weather conditions and faults.
Since the modules keep on working during daytime, it is difficult to locate faults and
disconnect the modules working under fault. As a result, this leads to a long duration
Condition Monitoring, Fault Detection and Diagnosis … 165

and prolonged unrecognized fault which results in gradual degradation of module.

Thus without condition monitoring, it will become an impossible task to isolate the
faults and this will result in a significant power loss and economy for large PV plants.

References

1. D.S. Pillai, N. Rajasekar, Metaheuristic algorithms for PV parameter identification: A compre-

hensive review with an application to threshold setting for fault detection in PV systems. Solar
Energy Research Cell (SERC), School of Electrical Engineering (SELECT), VIT University,
Vellore, India
2. Fault detection and monitoring systems for photovoltaic installations: A review Asma Triki-
Lahiania, Afef Bennani-Ben Abdelghania, b, Ilhem Slama-Belkhodjaa a Université de Tunis
El Manar, Ecole Nationale d Ingénieurs de Tunis, Laboratoire des Systèmes Electriques (LR
11 ES 15), 1002, Tunis, Tunisia
3. Degradations of silicon photovoltaic modules: A literature review Ababacar Ndiaye a, Abde´rafi
Charki b, Abdessamad Kobi b, Cheikh M.F. Ke´be´ a, Pape A. Ndiaye a, Vincent Samboua
a Centre International de Formation et de Recherche en Energie Solaire (CIFRES), Ecole
Supe´rieure Polytechnique-UCAD, BP 5085, Dakar-Fann, Senegal b,University of Angers-
ISTIA-LASQUO, 62 Avenue Notre Dame du Lac, 49000 Angers, France Received 29 October
2012; received in revised form 5 April 2013; accepted 8 July 2013
4. L. Cristaldi et al., Diagnostic architecture: aprocedure based on tha analysis of the failure causes
applied to PV plants. Measurement 67, 99–107 (2015)
5. Encapsulant Adhesion to Surface Metallization on Photovoltaic Cells Jared Tracy, Nick Bosco,
and Reinhold Dauskardt
6. R. Hariharan, M. Chakkarapani, G. Saravana Ilango, Member, IEEE, C. Nagamani, Senior
Member, IEEE, A Method to Detect Photovoltaic Array Faults and Partial Shading in PV
Systems
7. D. Stellbogen, Use of PV circuit simulation for fault detection in PV array fields, in Proceedings
of Conference Record 23rd IEEE Photovoltaic Spec. Conference, Louisville, KY, USA, May
10–14 (1993), pp. 1302–1307
8. H. Patel, V. Agarwal, Maximum power point tracking scheme for PV systems operating under
partially shaded conditions. IEEE Trans. Ind. Electron. 55(4), 1689–1698 (2008)
9. J. Young-Hyok et al., A real maximum power point tracking method for mismatching com-
pensation in PV array under partially shaded conditions. IEEE Trans. Power Electron. 26(4),
1001–1009 (2011)
10. M. Boztepe et al., Global MPPT scheme for photovoltaic string inverters based on restricted
voltage window search algorithm. IEEE Trans. Ind. Electron. 61(7), 3302–3312 (2014)
11. G. V.-Quesada, F. G.-Gispert, R. P.-L´opez, M. R.-Lumbreras, and A. C.-Roca, “Electrical
PV array reconfiguration strategy for energy extraction improvement in grid-connected PV
systems,” IEEE Trans. Ind. Electron., vol. 56, no. 11, pp. 4319–4331, Nov. 2009
12. M.Z.S. El-Dein, M. Kazerani, M.M.A. Salama, Optimal photovoltaic array reconfiguration to
reduce partial shading losses. IEEE Trans. Sustain. Energy 4(1), 145–153 (2013)
13. S. Silverstre, A. Chouder, E. Karatepe, Automatic fault detection in grid connected PV systems.
Sol. Energy 94, 119–127 (2013)
14. ] Y. Zhao et al., “Decision tree-based fault detection and classification in solar photovoltaic
arrays,” in Proc. 27th Annu. IEEE Appl. Power Electron.Conf. Expo., Orlando, FL, USA, Feb.
5–9, 2012, pp. 93–99
15. Z. Li, Y.Wang, D. Zhou, and C.Wu, “An intelligent method for fault diagnosis in photovoltaic
array,” in Proc. Int. Conf. Syst. Simul. Sci. Comput.,2012, pp. 10–16
16. M.N. Akram, S. Lotfifard, Modeling and health monitoring of DC side of photovoltaic array.
IEEE Trans. Sustain. Energy 6(4), 1245–1253 (2015)
166 O. Singh et al.

17. Fault Analysis and Detection Techniques of Solar Cells and PV Modules Ishtiak Ahmed Karim
Department of Electrical and Electronic Engineering Ahsanullah University o{Science and
Technology, Dhaka, Bangladesh
18. Kajari-Schroder, S., Iris Kunze, Ulrich Eitner, and M. Kontges.”Spatial and directional dis-
tribution of cracks in silicon PV modules after uniform mechanical loads.” In Photovoltaic
Specialists Conference (PVSC), 2011 37th IEEE, pp. 833–837. IEEE, 2011
19. Kontges, M., S. Kajari-Schroder, I. Kunze, and U. Jahn. “Crack statistic of crystalline silicon
photovoltaic modules.” 26th EU-PVSEC (2011): 3290–3294
20. ] Paggi, Marco, Mauro Corrado, and Maria Alejandra Rodriguez. “A multi-physics and
multi-scale numerical approach to microcracking and power-loss in photovoltaic modules.”
Composite Structures (2012)
21. Kontges, M., 1. Kunze, S. Kajari-Schroder, X. Breitenmoser, and B.Bjofneklett. “Quantifying
the risk of power loss in PV modules Due to micro cracks.” Proc. 25th EU PVSEC (2010):
3745
22. Kontges, M., I. I (unze, S. Kajari-Schri:ider, X. Breitenmoser, and B.Bjorneklett. “The risk of
power loss in crystalline silicon based photovoltaic modules due to micro-cracks”Solar Energy
Materials and Solar Cells 95, no. 4 (2011): 1131–1137
23. Sylke Meyer, Susanne Richter, Sebastian Timmel, Marcus Glaser, Martina Werner, Sina Swa-
tek, Christian Hagendorf, Snail trails: root cause analysis and test procedures. Energy Procedia
38, 498–505 (2013)
24. Peng, Anming Hu, Wenda Zheng, Peter Su, David He, Ken D.Oakes, Albert Fu et al. “Mi-
croscopy study of snail trail phenomenon on photovoltaic modules.” RSC Advances 2, no. 30
(2012): 11359–11365
25. G. Acciani, O. Falcone, S. Vergura, Defects in poly-Silicon and amorphous Silicon solar cells
(In IntI. Conf. on Renewable Energy and Power Quality, Spain, 2010)
26. Herrmann, W., W. Wiesner, and W. Vaassen. “Hot spot investigations on PV modules-new
concepts for a test standard and consequences for module design with respect to bypass diodes.”
In Photovoltaic Specialists Conference, 1997., Conference Record of the Twenty Sixth IEEE,
pp. 1129–1132. IEEE, 1997
27. Qasem, Hassan, Thomas R. Betts, and Ralph Gottschalg. “Soiling correction model for
long term energy prediction in photovoltaic modules.”In Photovoltaic Specialists Conference
(PVSC), 2012 38th IEEE, pp. 003397–00340 l. IEEE, 2012
28. Kumar, E. Suresh, Bijan Sarkar, and D. K. Behera. “Soiling and Dust Impact on the Effi-
ciency and the Maximum Power Point in the Photovoltaic Modules.” International Journal of
Engineering 2, no. 2 (2013)
29. Mani, Monto, and Rohit Pillai. “Impact of dust on solar photovoltaic (PV) performance:
research status, challenges and recommendations. “Renewable and Sustainable Energy Reviews
14, no. 9 (2010): 3124–3131
30. A.Belaout, F. Krim, A. Mellit, “Neuro-Fuzzy Classifier for Fault Detection and Classification
in Photovoltaic Module”
31. Y. Zhao, J. de Palma, J. Mosesian, R. Lyons, B. Lehman, Line line fault analysis and protection
challenges in solar photovoltaic arrays. IEEE Trans. Ind. Electron. 60, 3784–3795 (2013)
32. Article 690 - Solar Photovoltaic Systems, NFPA70, National Electrical Code, 2014
33. P.S. Rao, G.S. Ilango, C. Nagamani, Maximum Power from PV Arrays Using a Fixed Con-
figuration under Different Shading Conditions. IEEE Journal of Photovoltaics 4(2), 679–686
(2014)
34. S. Spataru, D. Sera, T. Kerekes, R. Teodorescu, “Detection of increased series losses in PV
arrays using Fuzzy Inference Systems,” Photovoltaic Specialists Conference (PVSC), 2012
38th IEEE, pp. 464– 469, 2012
35. S. Jing Jun and L. Kay-Soon, “Photovoltaic model identification using particle swarm optimiza-
tion with inverse barrier constraint,” IEEE Trans. Power Electron., vol. 27, no. 9, pp. 3975–3983,
2012
36. S. Harb, R.S. Balog, “Reliability of candidate photovoltaic module integrated- inverter (PV-
MII) topologies—A usage model approach”,IEEE Trans. Power Electron. 28(6), 3019–3027
(2013)
Condition Monitoring, Fault Detection and Diagnosis … 167

37. A. Drews, A.C. de Keizer, H.G. Beyer, E. Lorenz, J. Betcke, W.G.J.H.M. van Sark, W. Hey-
denreich, E. Wiemken, S. Stettler, P. Toggweiler, S. Bofinger, M. Schneider, G. Heilscher, D.
Heinemann, Monitoring and remote failure detection of grid-connected PV systems based on
satellite observations. Sol. Energy 81, 548–564 (2007)
38. K. Byung-Kwan, K. Seung-Tak, B. Sun-Ho, P. Jung-Wook, Diagnosis of output power lowering
in a PV array by using the Kalman-filter algorithm. IEEE Trans. Energy Convers. 27(4), 885–
894 (2012)
39. ] Y. Zhao, B. Lehman, R. Ball, J. Mosesian, and J.-F. de Palma, “Outlier detection rules for
fault detection in solar photovoltaic arrays,” in Proc.28th IEEE Appl. Power Electron. Conf.,
pp. 2913–2920, 2013
40. A. Chouder, S. Silvestre, Automatic supervision and fault detection of PV systems based on
power losses analysis. Energy Convers. Manag. 51, 1929–1937 (2010)
41. W.A. Omran, M. Kazerani, M.M.A. Salama, A clustering-based method for quantifying the
effects of large on-grid PV systems. IEEE Trans. Power Del. 25(4), 2617–2625 (2010)
42. Y. Zhao, L. Yang, B. Lehman, J. F. de Palma, J. Mosesian, and R.Lyons, “Decision tree-based
fault detection and classification in solar photovoltaic arrays,” in Proc. 27th IEEE Annu. Appl.
Power Electron.Conf., pp. 93–99, 2012
43. A. Massi Pavan, A. Mellit, D. De Pieri, and S. A. Kalogirou, “A comparison between BNN
and regression polynomial methods for the evaluation of the effect of soiling in large scale
photovoltaic plants,”Appl. Energy, vol. 108, pp. 392–401, 2013
44. Y. Zhao, R. Ball, J. Mosesian, J.F de Palma, and B. Lehman, “Graph-Based Semi-supervised
Learning for Fault Detection and Classification in Solar Photovoltaic Arrays” IEEE Trans.
Power Electron., vol. 30, NO.5, 2015
45. A.G. Pipe, B. Carse, Further experiments in Fuzzy Classifier Systems for mobile robot control,”
Intelligent Control. 2003 IEEE International Symposium on IEEE conference publications,
pp. 805–810, 2003
46. L.I. Kuncheva, “How good are fuzzy if-then classifier?”, Systems, man, and Cybernet.—Part
B: Cybernet. IEEE Trans. on 30(4), 501–509 (2000)
Introduction to Condition Monitoring
of PV System

Peeyush Kala, Puneet Joshi, Sanjay Agrawal, Lokesh K. Yadav

and Medha Joshi

Abstract The pace of installation of photovoltaic (PV) systems is growing over the
last two decades worldwide. The PV systems have been installed in extreme climatic
and environmental conditions which often degrade their performance. In order to
improve the performance of PV systems, duly maintenance and detection of faults
are required which prompt for the need for condition monitoring. The purpose of
condition monitoring of PV systems is to acquire information about power outages, to
maintain good power quality, protection from islanding, evaluation of performance,
identification of faulty modules, calculation of efficiency of modules, and compli-
ance to grid standards for connecting PV to the grid. The condition monitoring of
a PV system requires a substantial amount of measured data and software tools.
Currently, the operation and maintenance of most of the PV plants are performed
by manual and dispersive means. According to studies, condition monitoring using
these means is not only quite difficult to be performed but also not able to improve
the efficiency up to desired level. Hence, it is desirable to implement novel inspection
methods and analysis tools for efficient condition monitoring and accurate module
defect detection in PV systems. At present, there is a considerable interest given to the
storage, dispatchability, and management of photovoltaic (PV) power in real time.
So, there is need to manage. Researchers suggest that automation of the monitor-
ing techniques, real-time processing of information, and Unmanned Aerial System

P. Kala
Department of Electrical Engineering, Women Institute of Technology, Dehradun 248007, India
e-mail: [email protected]
P. Joshi (B) · S. Agrawal · L. K. Yadav
EED, REC, Ambedkarnagar 224122, India
e-mail: [email protected]
S. Agrawal
e-mail: [email protected]
L. K. Yadav
e-mail: [email protected]
M. Joshi
EED, SLSET Group of Institutions, Kichha, India
e-mail: [email protected]
© Springer Nature Singapore Pte Ltd. 2020 169
H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_7
170 P. Kala et al.

(UAS) technology can play an important role in the future. Apart from the commer-
cial applications, photovoltaic (PV) systems in residential premises are growing day
by day and represent a predominant component around the globe. As the PV systems
are major sources in microgrids and distributed generation (DGs), reliable islanding
detection algorithm is required to determine the electrical grid status and operation
in the grid-connected inverter. The issues related to performance of PV systems have
the strong dependency on many factors such as insolation intensity, ambient temper-
ature, cell temperature, air velocity, humidity, cloudiness, and pollution. Therefore,
in this chapter, conditioning monitoring of PV systems is explained.

Keywords Photovoltaic · System · Condition monitoring · Islanding protection ·

Grid

1 Introduction

The PV systems have experienced fast growth in various grid-connected and stand-
alone applications. However, the performance of PV systems degrades with time
due to various environmental factors as well as due to external and internal faults in
the system. The environmental factors include solar irradiance, high temperatures,
partial shading, high moisture, ultraviolet (UV) irradiation, wind, snow, hailstorms,
sand, salt, etc. The faults can be grid outage, islanding problem, faulty modules, and
failure of batteries. These performance issues have led to the idea of monitoring of
PV systems. The monitoring of installed photovoltaic (PV) system is a new area of
research.

1.1 Need for Condition Monitoring in PV Systems

PV systems are affected by various climatic and environmental conditions such as

continuous cycles of temperature, humidity, irradiation, mechanical stress, soiling,
moisture which degrade their performance. Figure 1 shows the various issues and
their classification that hinders the optimum performance of the installed PV systems.

1.2 The Need for AC Side Monitoring

PV arrays require effective monitoring on the AC side of the inverter to help the utility
company maintain power quality, obtain information about outages, and adhere to
standards for connecting PV to the grid.
Introduction to Condition Monitoring of PV System 171

Fig. 1 Photovoltaic systems issues

1.2.1 Prevention of Outages

The power output of PV systems, unlike traditional fossil fuel-based generation

schemes, depends on factors such as solar irradiance, cloud cover, and temperature
[1]. The power output of the array might drop or increase drastically at any instant
in time depending on external factors. Sudden drops in output power from large
PV plants could lead to power outages. As PV achieves higher grid penetration, a
centralized monitoring system becomes essential. Such a system would allow greater
control over the contribution of specific PV plant to the grid. If the PV plant with a
monitoring system has reduced output, controllers would be able to supplement the
power with traditional systems such as a storage battery or a diesel-electric plant.
172 P. Kala et al.

1.2.2 Islanding Protection

The addition of arrays to the power grid poses a safety risk known as islanding.
Islanding occurs when a fault in the wider electrical grid causes a power outage,
but a PV array does not turn off. This might result in the PV array supplying the
local loads resulting in the formation of a small “island” in the grid. This island of
energized wires poses a threat to repair technicians seeking to fix the fault. Modern
inverters are, for the most part, equipped with anti-islanding features which detect
when the AC grid goes down and shut off the inverter. In addition to this safeguard,
many electrical companies require that residential arrays connected to their grid
come equipped with a manual disconnect switch, to be opened by a power company
technician before beginning work on nearby parts of the grid.
The IEEE standard 1547 provides guidelines for interconnecting a distributed gen-
eration resource (such as PV) with power systems. The standard provides guidelines
such as bounds on voltage and frequency distortion while laying out response require-
ments to abnormalities. The guidelines mandate that the PV array interconnection
system must detect unintentional islanding and cease to energize the electric power
system. The standard mandates that if the aggregate of the distributed resources at
a single point of common coupling (point where a local electric power system is
connected to an area electric power system) equals or exceeds 250 kVA, there must
be provisions to monitor the connection status, power output, and voltage at the point
of distributed resource connection.

1.2.3 Improve Power Quality

Power utility companies are responsible for ensuring power quality. Significant devi-
ation in the magnitude, frequency, or purity of waveforms results in power quality
problems. The utility has no control over the amount of current a load might draw.
Power quality is maintained by keeping the voltage and frequency within certain lim-
its. The output of PV arrays varies significantly over time resulting in potential power
quality problems. Monitoring and control of PV arrays can be used to improve power
quality by changing the outputs such as voltage and frequency of a PV array. The
availability of monitoring information from various distributed sources in a given
area can provide better insight for the electric power company. This information pro-
vides avenues for better understanding of a given area and means to improve power
quality. As more PV is added to the grid, power quality becomes a serious concern.

1.2.4 Performance Evaluation

Effective monitoring of a PV array can be used to evaluate the performance of the

array. For instance, the voltage and currents generated by the array can be used to
determine the annual energy production. This can then be used to determine the
cost per watt of the PV system. The electrical outputs such as power generated can
Introduction to Condition Monitoring of PV System 173

be compared to the weather data using PV models to determine the efficiency of

operation of the system.

1.3 The Need for DC Side Monitoring

Monitoring on the AC side of the array cannot be used to determine the location of
the faults, if any, within a PV array. PV arrays require monitoring on the DC side
to identify faulty modules, calculate the efficiency of operation of the modules and
determine long-term and short-term trends of the system.

1.3.1 Identify Faults

Methods of fault detection that rely only on human operators are not accurate and
might result in faults remaining unnoticed. Automated monitoring systems must be
used to identify the presence of faults in an array or subarray. Monitoring on the
AC side cannot detect faults that affect only a few modules in the array. The loss of
power due to such a fault is indistinguishable from measurement errors and seasonal
variations. Hence, monitoring is required on the DC side to detect faults effectively.
For instance, [2] suggests an Ad hoc method that utilizes information obtained at
the end of each string to identify faults that reduce the string output by over 20%.
Automated monitoring systems can use sophisticated signal processing algorithms to
identify the exact location of the fault within the array. A signal processing algorithm
that utilizes measurements from each module to detect ground fault and arc fault
locations at over 99% accuracy is described in [3].

1.3.2 Evaluating Trends

The outputs collected on the DC side can be compared to the historical data to deter-
mine the long-term trends. For instance, module data collected from previous years
can be compared with that of the current year to calculate the module degradation
factor. The output of the modules can be compared to weather data to update PV
models.

1.3.3 Literature Review and Research Gap Findings

Lindig et al. [2] investigated the performance of degradation models of PV mod-

ules and systems. Several statistical models were chosen for the calculation of the
performance loss using as case study two PV systems, installed in Bolzano/Italy.
Simple linear regression (SLR), classical seasonal-decomposition, seasonal- and
trend-decomposition using loss (STL), Holt–Winters exponential smoothing, and
174 P. Kala et al.

autoregressive integrated moving average (ARIMA) were discussed. The perfor-

mance loss results show that SLR produces results with highest uncertainties. In
comparison, STL and ARIMA perform with the highest accuracy, whereby STL is
favored because of its easier implementation. On the other hand, if monitoring data
at PV module level are available in controlled conditions, analytical models can be
applied. In this study, various analytical models depending on different degradations
modes were discussed. A comparison study is carried out for models proposed for
corrosion. The results of the models agree in explanation of experimental obser-
vations, a big difference in degradation prediction was observed. Finally, a model
proposed for potential induced degradation was applied to simulate the degradation
of PV systems maximum power in three climatic zones: alpine (Zugspitze, Germany),
maritime (GranCanaria, Spain), and arid (Negev, Israel). As expected, a more severe
degradation is predicted for arid climates.
Pigazo et al. [3] proposed a reliable islanding detection algorithm (passive or
active) for distributed power generation systems (DPGSs). This paper proposed a
time–frequency detection algorithm based on monitoring the DPGS output power
considering the influence of the PWM, the output LCL filter, and the employed current
controller. Wavelet analysis is applied to obtain time localization of the islanding
condition. Simulation and experimental results show the performance of the proposed
detection algorithm in comparison with existing methods.
In [1], a microcontroller-based power management system (PMS) was designed
for the online operation of an experimental low-voltage microgrid equipped with
a battery storage system and two power supplies: a kilowatt (kW)-class proton
exchange membrane (PEM) fuel cell (FC) and a photovoltaic (PV) module emulator,
both connected to a low-voltage AC node. The connections of the energy sources
to the common AC bus make use of power inverters with specific functionalities.
The AC node feeds electric active and reactive load emulators able to reproduce
programmable profiles. The automatic PMS provides the microgrid monitoring and
the FC power scheduling in both grid-connected and islanded operating conditions.
Experimental investigation was performed for the microgrid in islanded conditions.
In [4], the use of infrared thermography was proposed to determine the operational
status of photovoltaic solar systems on a large scale using flight. Solar thermography
is the use of an infrared camera to inspect photovoltaic solar systems for problems
that can cause damage to the cells, loss of efficiency, and fire hazards. The difficulty
of finding and tracking down problems and potential issues that arise in these instal-
lations are monumental. Therefore, infrared thermography is a valuable technology
that can be used on a wider scale to pinpoint problems in these large arrays.
A new system, PV on-time was developed [5] to supervise the operating mode of
a Grid-Connected Utility-Scale PV Power Plant. This system presents architecture
of acquisition devices, including wireless sensors distributed around the plant, which
measure the required information. It is also equipped with a high-precision protocol
for synchronizing all data acquisition equipment, something that is necessary for
correctly establishing relationships among events in the plant. Moreover, a system
for monitoring and supervising all of the distributed devices, as well as for the
real-time treatment of all the registered information, is presented.
Introduction to Condition Monitoring of PV System 175

A new in situ moisture monitoring technique for photovoltaic (PV) modules was
proposed in [6] using small digital humidity and temperature sensors. The sensors
were embedded in three different ethylene-vinyl-acetate (EVA) stacks and proved to
be resistant to lamination conditions. The fact that they are in direct contact with EVA
does not affect their performance, since their saturated relative humidity (RH) reading
is proportional to the external RH in the air. By exposing the sensors to elevated
temperature and RH conditions, water vapor transmission rate of the backsheet and
diffusion coefficient of the EVA can be determined. Obtained coefficients agree
with reference values within their measurement uncertainties. Besides determining
material moisture ingress properties, this monitoring technique is also applicable for
long-term outdoor PV module monitoring. It shall provide a valuable location and
installation-specific information of RH and temperature stress conditions, especially
as feedback information to manufacturers of materials and PV modules.
In [7], the maintenance of PV systems is described which is largely based on
periodic inspections (PIs), with a few PV systems manufacturers and maintenance
service providers now starting to offer services based on continuous monitoring
(CM). Profitability of CM versus PI has been widely studied in the literature, with
respect to conventional power systems as well as wind turbines. No such studies exist
on residential PV systems, particularly from the viewpoint of the service provider.
Our work addresses this gap. We first develop a stochastic model of a residential PV
installation, accounting for its degradation and failures. We then develop a model
of the costs and benefits for both the system owner and the maintenance service
provider, under both PI and CM. We conclude by analyzing the profitability of the
two maintenance schemes using real data from the Japanese market.
Vergura et al. [8] deals with the problem of supervising and monitoring a photo-
voltaic (PV) plant. First, an offline descriptive and inferential statistical procedure for
evaluating the goodness of system performance is presented. Then, an online infer-
ential algorithm for real-time monitoring and fault detection is introduced. The two
methodologies utilize the energy output of inverters as input data and are valid for
both Gaussian and non-normal distribution of data. The procedures have been tested
on a real PV installation, and results are reported for the case of a grid-connected PV
plant in Italy for which one PV module over 132 resulted in being badly connected.
Unmanned Aerial Vehicles (UAVs) are becoming a reliable and useful instru-
ment in monitoring and diagnostic services in the energy field [9]. Operation and
maintenance are a crucial factor for PV plant inspection and control activities which
currently are quite hard to be performed due to manual and dispersive procedures.
The automation of the monitoring techniques will become in future an essential task
in this market and Unmanned Aerial System (UAS) technology can play a role in
it. This paper proposes a novel, comprehensive and innovative approach in order to
automate the inspection and prognostic procedure by means of designing a control
system in order to provide accurate information on operating conditions of PV plants.
The system is able to perform PV systems’ monitoring, diagnosis, defects and failures
reconnaissance, data processing and to propose remedial actions. In current exper-
imental research, only IR sensor imaging was examined to evaluate thermographic
behavior of the PV modules and to build the related image processing algorithm.
176 P. Kala et al.

However, it is possible to easily extend such a procedure in order to manage a larger

amount of data. First results of this work have proven that the proposed procedure is
very promising being fast, cost-effective and adaptable to large PV plants which can
be controlled during their entire lifetime.
Dhoke et al. [10] considers the problems of condition monitoring and fault detec-
tion in an existing solar photovoltaic (PV) plant in Australia. A PV prediction model is
proposed to accurately model the PV plant output. This model is then used with three
condition monitoring and fault detection methods. The considered methods involve
a comparison of measured and modeled voltage and current ratios with appropri-
ate thresholds and adjacent string values. The procedure to calculate thresholds is
described. The proposed PV model and the condition monitoring approaches are col-
lectively used with real PV data. As a result, the string disconnection and bypassed
module faults are detected. It is also found that the string level monitoring is ideally
suited for reliable condition monitoring and fault detection especially for large PV
plants.
A health monitoring method for photovoltaic (PV) systems based on probabilistic
neural network (PNN) was proposed in [11] that detects and classifies short and
open circuit faults in real time. To implement and validate the proposed method
in computer programs, a new approach for modeling PV systems was proposed
in this work that only requires information from manufacturer’s datasheet reported
under normal operating cell temperature (NOCT) conditions and standard-operating
test conditions (STCs). The proposed model precisely represents characteristics of
PV systems at different temperatures, as the temperature dependency of parameters
such as ideality factor, series resistance, and thermal voltage is considered in the
proposed model. Although this model can be applied to a variety of applications, it is
specifically used to test and validate the performance of the proposed fault detection
and classification method.
Ahmad et al. [12] discuss the need to interconnect low-voltage solar Photovoltaics
(PV) with DC system, a DC-DC boost converter is required. To minimize the switch-
ing frequency oscillations in solar PV voltage, Aluminum Electrolytic Capacitors
(AECs) are connected between solar PV and the converter. Operational life of AECs
depends on electrical and environmental parameters and is less than that of solar
PV panels. Equivalent Series Resistance (ESR) of AEC increases with life, thereby
increasing the voltage ripple across PV terminals. This results in less power extrac-
tion from solar PV. This paper proposes a method for health monitoring of AEC using
the parameters measured for maximum power point tracking (MPPT) (PV voltage
and current). The proposed technique is applicable for both Continuous Conduction
Mode (CCM) and Discontinuous Conduction Mode (DCM) of operation of the con-
verter. Effect of temperature on ESR is incorporated for accurate determination of
the health of capacitor. Key advantage is that no additional current or voltage sensor
is required to implement this technique, thereby offering a low-cost solution. Fur-
ther, the proposed technique does not require significant processing capabilities and
is implemented in the same digital controller used for MPPT. Detailed simulation
studies are carried out and results are included in this paper. A laboratory prototype of
Introduction to Condition Monitoring of PV System 177

DC-DC boost converter is developed for experimentation. Results are in agreement

to that obtained from simulation studies.
The efficient condition monitoring and accurate module defect detection in large-
scale photovoltaic (PV) farms demand for novel inspection method and analysis
tools. Li et al. [13] has presented a deep learning-based solution for defect pattern
recognition by the use of aerial images obtained from unmanned aerial vehicles
(UAVs). The convolutional neural network (CNN) is used in the machine learning
process to classify various forms of module defects. Such supervised learning process
can extract a range of deep features of operating PV modules. It significantly improves
the efficiency and accuracy of asset inspection and health assessment for large-scale
PV farms in comparison with the conventional solutions. The proposed algorithmic
solution is extensively evaluated from different aspects, and the numerical result
clearly demonstrates its effectiveness for efficient defect detection of PV modules.
Ahmed et al. [14] presented and discussed the performance evaluation, moni-
toring, and analysis of a stand-alone Photovoltaic (PV) system. This experimental
investigation achieves its objectives using Laboratory Virtual Instrument Engineer-
ing Workbench (LabVIEW) interface capabilities. The installed PV generator system
operates in extreme and severe weather conditions. The local area is characterized
by a high level of irradiation, high humidity variations, very high ambient temper-
ature, and frequent dust storms. Two PV panels of 150 W rating each are used.
Each panel feeds a variable resistive load separately and simultaneously. The con-
structed LabVIEW engine displays environmental parameters and the associated
electrical variables (voltage, current, and power) on dedicated graph windows con-
tinuously in a real-time manner. The current versus voltage (I–V) and power versus
voltage (PV) characteristics are taken for any chosen panel, displayed and stored
accurately. LabVIEW allows appending any characteristics for the sake of compar-
ison, investigation and research purposes. LabVIEW has shown high performance
in communicating with several devices simultaneously and high capability of dis-
playing several variables behavior at a time. The designed virtual instrument (VI)
filters are programmed to execute different tasks on a priority basis. The developed
system reveals and stores the pronounced impact of dust, the variation of temperature
and the irradiation changes on the PV panel power output behavior. The monitoring
results are very satisfactory. The online data display in multi-scale window frame is
very informative. The online efficiency evaluation is very useful for system operation
and analysis. The developed system was found to be very supportive of research and
educational purposes.
The monitoring of the solar panels on periodic time is important to improve the
durability functioning of the solar system [15]. To obtain the best solar capacity of
PV systems is feasible through an advance and automatic supervisory technique.
This paper presents advance automatic monitoring and fault detection technique in
Photovoltaic (PV) system. The performance and efficiency of Photovoltaic system
depend on many parameters like incident angle, shading, aging, dirt, Temperature,
MPPT error. This automatic monitoring system required data-acquisition system
to collect all the data from PV system. The performance of PV module degraded
with aging effect and dirt condition. But there are also other factors which affect
178 P. Kala et al.

the performance of module like partial shadows, increase in series resistance and
shunting of cells. For enhancing the operability of PV system we need to be automatic
supervision and conditioning monitoring. In this paper, we have given a review on
different monitoring techniques of PV system. The monitoring of the solar panels on
periodic time is important to improve the durability functioning of solar system. To
obtain the best solar capacity of photovoltaic (PV) is feasible through an advance and
automatic supervisory technique. This paper presents advance automatic monitoring
and fault detection technique in Photovoltaic (PV) system.
Stonier and Lehman [16] presented an intelligent-based fault tolerant system for
a solar Photovoltaic (PV) inverter. Artificial Neural Network (ANN) based con-
troller is used to monitor, detect, and diagnose the faults in solar PV panels, battery,
semiconductor switches, and inverters. The Cascaded Multilevel Inverter (CMI) is
connected across the combination of solar PV panel and battery for DC-AC conver-
sion. The major advantage of the proposed topology is that it can deliver power from
source to the load even under fault and partial shaded conditions. The paper also
reviews on various faults in the solar PV energy conversion process and provides
suitable solutions for each circumstance. Simulations are undertaken in MATLAB
2016a/Simulink and experimental investigation is carried out for a 3kWp solar PV
system. The results have proven that the system is capable to deliver power in spite
of faulty environments. The comparison and discussion of the results were made to
show the effectiveness of the proposed system.
Kolodenny et al. [17] have identified representative metrics, methods to calcu-
late the metrics, and estimation of the metric uncertainty level. They recommended
following four performance metrics:
1. Power Performance Index (PPI) of actual instantaneous kW AC power output
divided by expected instantaneous kW AC power output. Metrics and methods to
assess the performance of existing systems to aid bankability of PV asset class.
Determining and evaluating system performance based on actual weather and
actual system characteristics is critical to developing creditability for PV as an
asset class.
2. Performance Ratio with temperature corrected final yield using weighted-average
cell temperature (CPR).
3. Energy Performance Index (EPI-SAM) of actual kWh AC energy divided by
expected kWh AC energy as determined from an accepted PV model, such as
SAM, using actual climate data and assumed derate factors.
4. Energy Performance Index (EPI-REGRESSION) of actual kWh AC energy
divided by expected kWh AC energy as determined from a polynomial regression
equation having coefficients determined from actual operating and climate data
collected during the model “training” period.
Some conclusions of this study show how the above four metrics are applicable
for the following performance assessment objectives:
1. Monitoring of a specific PV system to identify degraded performance and need
for condition-based maintenance. Recommendations, including varied levels of
uncertainty, are to use EPI-SAM or EPI-Regression or CPR.
Introduction to Condition Monitoring of PV System 179

Table 1 Performance matrices

Metric Purpose Method Uncertainty
Performance ratio Maintenance (kWh/Rated High—15 to 20%
kWDC )/(kWh/1000)
CPR—Temperature Maintenance [(kWh/(Rated Moderate—10 to
corrected PR kWDc *Temp 15%
corr.)]/[kWhsun /1000]
EPI—Energy Perf. Maintenance, Actual kWhAC /calc. Moderate—10 to
Index SAM model commissioning, SAM kWhAC 15% (Model
Financial dependent)
EPI—Energy Perf. Maintenance, Actual kWhAC /Calc. Low—5 to 10%
Index regression model commissioning, regression kWhAC
Financial
kWh production Maintenance Compare AC kWh High—15 to 20%
period to period
Yield Financial kWh per DC watt Low—5 to 10%
power
PPI—Power Commissioning, Measured kW power Low to
Performance Index troubleshooting output versus moderate—5 to
calculated kW power 15% (model and
expected measurement
dependent

2. Commissioning of a new system, re-commissioning, or assessment after major

maintenance and to set a baseline for future performance measurements and
comparisons. Recommendation is to use PPI and EPI metrics.
3. Determination of specific industry parameters, such as Yield or Performance
Ratio, to allow comparison of systems in different geographic locations for design
validation or investment decisions. In some cases, depending on the objective,
combinations of these metrics are most useful (Table 1).

1.4 Objectives of Performance Assessment

The prime objective of a performance assessment is to detect changes in system

performance; usually decreases in performance, to allow the system owner to inves-
tigate and potentially perform cost-effective maintenance. This can be done best on
a relative scale where the specific performance of the system is compared to itself
which reduces adverse effects of modeling input assumptions and uncertainty.
The objectives for performance assessment can be summarized as
• How is my system, or a portion of my system, performing currently in comparison
to how I expect it to perform at this point in its life?
180 P. Kala et al.

• How is my system performing for both the short term and long term in comparison
to how it is capable of performing with its given design, site location, and baseline
performance?
• How is my system performing over an assessment period in comparison to other,
similar systems in similar climates?
• How is my system performing compared to the last assessment periods? This
trending model is useful for maintenance objectives.
• How can I develop metrics in support of accurate prediction of future energy yield
and ROI for reliable investment assessment?
• During commissioning, what metrics should be used to set a baseline for future
performance assessments?
Performance metrics can first be divided into instantaneous, short-term, and long-
term assessment periods. Various degradation mechanisms and intermittent anoma-
lies develop and occur over long-term periods so both periods are needed to complete
an assessment. Instantaneous output is based on power and is denoted by kW (power).
A long-term assessment period, such as weekly, monthly, or annually is based on
energy and yield, and is denoted by kWh (energy).

2 Methods of Monitoring

Several systems have been developed to monitor PV arrays and modules. We look
into some of these systems currently being used and mainly emphasizing systems
that perform intra-array monitoring. The usability of each system varies depending
on their targeted capabilities. A monitoring system used to evaluate the performance
of PV array installed on a building. The monitoring system measures voltage, cur-
rent, and power at the AC output of the inverter. It measures the solar intensity using
two photo diode sensors and taking their average. Module internal temperatures
are obtained using temperature sensors. The measurements are done over 20-min
intervals. The data can be used for both short-term purposes such as monitoring the
system’s status, and long-term purposes such as monitoring the deterioration of com-
ponents. Comparison of the different strings helps evaluate the effect of shading on
the array. The correlation of output power with temperature can be used to determine
the effect of module temperature on output power for the same irradiance. These
measurements can be used to evaluate the annual energy production and cost of elec-
tricity produced by the array. Measurements are transferred to a computer enabling
internet access to the data. The authors mention the use of commercially available
data acquisition systems to transfer the data to the computer. Here, the monitoring
system does not consider module level measurements and communication systems
for such an arrangement.
Mohame et al. [18] proposed an approach that uses modern informatics tools such
as XML to analyze the acquired data. Their goal is to analyze the performance of
a PV system of any type and size. They proposed a protocol called PV mark up
Introduction to Condition Monitoring of PV System 181

Fig. 2 Schematic diagram for data access

language (developed from XML) to automatically access, extract, and use data from
several sensors systems/database sources. The system collects and classifies the data.
Depending on whether the user is a technician or plant owner or scientist, different
views are presented containing only the most relevant information. Figure 2 shows
the schematic diagram for data access system. A PV logger system constantly collects
information about the state of the array (such as voltages, currents, and irradiance) and
updates it in a database. The system diagnostics retrieves the data and detects possible
failures. This information is then updated in the database. The different users can
query the database and obtain information they are interested in. The owner can view
the overall system health and output of the array while the technician can view the
PV system parameters. The scientist’s view provides access to all the data collected
including the PV system models. This work provides a comprehensive method to
store and retrieve PV array data and can be used with different current-voltage sensor
location and weather data. However, it does not describe a complete PV monitoring
system which includes sensing, data communication, and user interface systems.
A simple and cost-effective method of monitoring a PV power station using a
GUI built in NI LabVIEW is presented in [19]. The setup consists of current and
voltage sensors connected to a micro controller through an analog multiplexer. Also
included are sensors for measuring the module surface temperature and irradiance.
The collected data can be used for both monitoring and control. The micro controller
is interfaced to a laptop through a serial port where the data can be viewed and
analyzed using a GUI designed using LabVIEW.
182 P. Kala et al.

3 Monitoring System Considerations

Each monitoring system of PV array requires data acquisition, data management,

and graphical user interface. These different aspects of a PV monitoring have to be
combined in a single automated system. The block diagram for a PV array with an
automated monitoring system is shown in Fig. 2. The PV modules are connected
to the inverter to form the PV array. Sensors are placed to collect the PV currents
and voltages from the DC input to inverter, end of each string and/or individual
modules. The measured parameters are transmitted to the central server through a
communication channel. The communication channels used can be ethernet, wireless,
or power line. A weather station records the irradiance and temperature values and
transmits them to the central server. The central server maintains a database of the
array outputs and weather data. It runs PV models to calculate expected output values
and fault detection algorithms to detect the presence of faults in the array. The array
operator has access to the expected and actual output and fault information through an
easy to use GUI. There are several aspects to be considered for an effective monitoring
and control system for PV arrays. These include the parameters to be measured,
the placement of sensors for measuring these parameters, communication systems
that can transfer the measurements to a central database, algorithms to manage and
process the collected data, and user interfaces to visualize the state of the PV arrays.
This section discusses these aspects briefly.

4 Sensor Measurements

PV array characteristics can be determined from the measured parameters such as

voltage, current, and irradiance. The measurements can be made at different locations
in an array, from the inverter to individual modules.

4.1 Inverter Level Measurements

Most currently used PV monitoring systems record current and voltage measurements
at the inverter to analyze the performance of the array. Using the data for a typical
year at the site and models such as the Sandia performance model [1], the expected
output power of the array can be calculated. These can be compared with the actual
power output by the inverter to determine whether the array is operating without any
faults. For example, using monthly averaged energy at the inverter as a metric, loss of
energy in the array was identified and rectified in [1]. Inverter faults are characterized
by large variation between expected and actual AC power while actual and expected
DC power remain similar. Array faults are characterized by significant variation in
expected and actual DC power.
Introduction to Condition Monitoring of PV System 183

4.2 String/Module Level Measurement

Although measurements at the inverter level enable us to identify the overall health
of the system, they cannot in most cases detect non-catastrophic faults. They do not
provide enough information to identify the location and nature of faults immediately.
Identifying and correcting a fault or underperforming component in the array still
involves taking field measurements by technicians.
Higher resolution data acquired by means of additional sensors installed on indi-
vidual modules or strings provide more precise understanding of the array. These sen-
sors can measure module or string currents and voltages. These additional readings
help to identify the modules that need to be inspected in case of a fault.
It is possible to measure the temperature, wind speed, and irradiance at the instal-
lation site. Approximate estimates of the module cell temperature can be estimated
from the atmospheric temperature and wind speed. Significant differences between
the predicted outputs and actual measurements can be utilized in an automated fault
detection setup. The measurements at string level provide a more definitive descrip-
tion of the performance of an array when compared to only inverter level measure-
ments. However, they are far from perfect. Cell temperatures calculated from atmo-
spheric temperature do not reflect the exact temperatures of the cells required for
use in the PV models. This can be due to manufacturing process variations, wiring
effects and other factors such as high resistance connections. Even cells within a
single module might have variations in their temperatures. Better estimates of cell
temperature may be acquired by sensors on the back surface of modules. These
module level measurements can then be used in conjunction with the string level and
inverter level measurements. The different measurement levels for a PV array and
their capabilities are summarized in Table 2. As the number of measurement points
increases, the automatic monitoring system becomes more effective in localizing the
faults.

Table 2 Measurement levels

Level I Level II Level III
in a PV array starting at the
level of the inverter Inverter String Module
Array real/reactive Currents/voltages Current and
output to grid in a string voltage in a
module
Array performance Faults can be Faults can be
localized within a localized within a
string string
Significant faults
in the array
Inverter faults
184 P. Kala et al.

4.3 Measurement Resolution Considerations

Measurement of currents and voltages at the module level or string level provides
better analysis and control capabilities. But this requires effective methods to transmit
the sensor information to a central server for processing. The central server can be
located at the site of the PV array and analysis or data transfer to the grid level
monitoring systems can be performed from this server. The data rate requirements
for monitoring systems scale linearly with the size of the array, and inversely with
the sampling period of the measurement sensors. To avoid clutter associated with the
transfer of data through wires, wireless technologies can be used. Individual sensors
would require relatively low data bandwidth to transfer the measurements to the
server. The power consumption for the data transfer must be low to enable extended
use of the sensors in the field without battery replacements. The requirements of
low power and low data rate make wireless technologies such as variants of IEEE
802.15.4.2006 (Zigbee, Wireless HART) and in some cases Ultra-wideband (UWB)
practical. Data communication for PV systems to transfer the data from the sensor
can also utilize power line communications (PLC) which would utilize the existing
infrastructure to transfer the data across power lines to the inverter. Any data transfer
mechanism can be used to transmit this data collected at a single point to the server.
There are two types of communication systems is given below.

4.3.1 Power Line Communications

To avoid extra wiring, or interference with existing devices, PLC technologies can
be used for communications. PLC is widely used both over high-voltage and low-
voltage lines. However, there is no adopted IEEE standard, especially in the context
of PV arrays. Communication of data on the DC generation side of the PV array can
make use of protocols such as RS 232. The PLC channel introduces large amounts
of time and frequency variation in the medium of propagation. The time-variation
of the communication medium is due to sudden electrical load changes which alter
the overall impedance of the system. The frequency variation is due to the dispersive
nature of the medium which causes reflections of the transmitted waveform.

4.3.2 Wireless Technologies for Communication

Using wireless technologies to transmit the data from the sensors to the server can
be considered as an alternative to PLC. ZigBee is a wireless standard which is an
extension of the IEEE 802.15.4.2006 standard focusing on low-cost, low-power com-
munication between devices located within a typical range of 10–75 m. The data rates
are in the order of kilobits per second, which is sufficient to transmit the array state to
the server or the gateway sensor. ZigBee utilizes significantly lower power compared
to standards such as Wi-Fi and Bluetooth. Ultra wideband (UWB) communication
Introduction to Condition Monitoring of PV System 185

can be used if the distance between the transmitter and receiver is extremely small
(around 10 m). UWB communications consume low power and transmit information
without affecting other systems using the same frequency band. However, UWB can
handle much higher bandwidth than ZigBee. UWB communications do not have any
security protocols whereas ZigBee uses encryption. These wireless technologies for
data communication of the PV system suffer from some of the same problems as
that of PLC. These might include interference from other wireless devices, signal
distortions in the form of time dispersion and signal fading.
There is no standard that dictates the usage of these technologies in the PV area.
It is advantageous to have PLC communication over ZigBee when the data rates are
low (in the order of hundreds of bits per second) and the coverage requirements of
the network are more than the range of ZigBee (say over 100 m). ZigBee is better
suited for data rates in the order of kilobits per second and a smaller coverage area
[20].

4.3.3 User Interface

Once the sensors and the communication systems to the server are set up, there is
a need for algorithms and user interfaces that enable the user to make sense of the
data and take appropriate actions. Simple and easy to use GUIs need to be developed
for visualizing solar array data. Such GUIs can provide information on the overall
health of the system. They are linked to sophisticated data processing algorithms
and models of PV systems and can be used to evaluate the PV array performance,
detect failures, etc. The GUI must provide a real-time update of the current state of
operation of the array. This includes the AC/DC power output generated by the array,
faults if any detected and the status of the battery.
It can include other information such as real and reactive power output to the grid,
total power output during the day/week/month and the efficiency of array operation.
The cost of power generated can be displayed. This is especially useful for residential
PV arrays as owners can keep track of the effective amount earned from their array
after energy consumption for their residence. The GUI can provide comprehensive
tools to analyze the array as a whole or individual modules, strings, etc. Fault detection
algorithms can be used to display the nature and location of any faulty modules. The
GUI can provide tools to plot/analyze the array and module output against weather
data such as irradiance. This can be done for the current day or across several days
using stored data and can be used to study the short-term and long-term trends in the
array.
1. The GUI should be connected to other systems such as audible alarms to alert the
operator when faults occur. They can be connected to emergency systems such
as fire or flood alert systems as required. They can be integrated with perimeter
security systems such as video cameras to provide a single point of reference for
the human operator. If required, a PV monitoring system GUI must be able to
enforce data security based on access privileges.
186 P. Kala et al.

2. It can be used to generate automated reports regarding the performance of the

array. The GUI must, if required, be capable of handling multiple arrays within
a site.

5 Conclusions

This chapter discussed the need for monitoring systems for photovoltaic arrays. Such
monitoring systems should be able to provide the system operator with information
regarding the array’s operation. The level of detail collected from a PV array deter-
mines the accuracy and capabilities of the monitoring system. Small scale PV arrays
generally do not require high-resolution data collection whereas large-scale arrays
do. The data collected from the sensors must be transmitted to a central server for
processing using wireless or power line channels. Wireless channels are suited for
higher data rate and smaller coverage areas while power line communication can
handle large coverage areas at lower data rates. The central server must contain algo-
rithms used for processing the data collected from the array and a user interface that
lets the operator visualize and take actions based on the data. This chapter presented
an overview of PV array monitoring system. The existing methods of monitoring are
discussed. The design considerations of a monitoring system that include the location
and placement of sensors, the type of communication system to be employed, and
the algorithms used for processing the data collected from the array are explained.
The requirements of a GUI for PV system are discussed. The decreasing cost of
monitoring and control technology has created an opportunity to update PV array
management practices. Automated systems must be developed to detect and respond
to faults with little or no human involvement. Automated fault detection has the
potential to dramatically reduce the meantime to repair of PV arrays, in some cases
from weeks to less than a day. This increases array availability, and therefore power
output. The addition of control circuitry to a PV array also allows dynamic reconfig-
uration of the array topology, either to mitigate a fault condition or to optimize the
array size to provide a better yield.

References

1. B. Belvedere, M. Bianchi, A. Borghetti, C.A. Nucci, M. Paolone, A. Peretto, A microcontroller-

based power management system for standalone microgrids with hybrid power supply. IEEE
Trans. Sustain. Energy 3(3), 422–431 (2012)
2. S. Lindig, I. Kaaya, K. Weiß, D. Moser, M. Topic, Review of statistical and analytical degra-
dation models for photovoltaic modules and systems as well as related improvements. IEEE J.
Photovolt. 8(6), 1773–1786 (2018)
3. A. Pigazo, M. Liserre, R.A. Mastromauro, V.M. Moreno, A. Dell’Aquila, Wavelet-based island-
ing detection in grid-connected PV systems. IEEE Trans. Industr. Electron. 56(11), 4445–4455
(2009)
Introduction to Condition Monitoring of PV System 187

4. H. Denio, Aerial solar Thermography and condition monitoring of photovoltaic systems, in

2012 38th IEEE Photovoltaic Specialists Conference, Austin, TX, 2012, pp. 000613–000618
5. I.M. Moreno-garcia, E.J. Palacios-garcia, V. Pallares-lopez, I. Santiago, M.J. Gonzalez-
redondo, M. Varo-martinez, R.J. Real-calvo, Real-time monitoring system for a utility-scale.
Sensors (Basel) 16(6), 1–25 (2016)
6. M. Jankovec, F. Galliano, E. Annigoni, H.Y. Li, F. Sculati-meillaud, C. Ballif, M. Topiˇ, In-Situ
Monitoring of Moisture Ingress in PV Modules Using Digital Humidity Sensors, vol. 6, no. 5,
pp. 1152–1159 (2016)
7. T. Mukai, M. Tomasella, A.K. Parlikad, N. Abe, Y. Ueda, The competitiveness of continuous
monitoring of residential PV systems: a model and insights from the Japanese market 5(4),
1176–1183 (2014)
8. S. Vergura, G. Acciani, V. Amoruso, G.E. Patrono, F. Vacca, Descriptive and inferential statistics
for supervising and monitoring the operation of PV plants 56(11), 4456–4464 (2009)
9. M. Aghaei, F. Grimaccia, C.A. Gonano, S. Leva, Innovative automated control system for PV
fields inspection and remote control. IEEE Trans. Industr. Electron. 62(11), 7287–7296 (2015)
10. A. Dhoke, R. Sharma, T.K. Saha, Condition monitoring of a large-scale PV power plant in
Australia, in 2016 IEEE Power and Energy Society General Meeting (PESGM), Boston, pp. 1–5
(2016)
11. M.N. Akram, S. Lotfifard, Modeling and health monitoring of DC side of photovoltaic array.
IEEE Trans. Sustain. Energy 6(4), 1245–1253 (2015)
12. M.W. Ahmad, N. Agarwal, S. Anand, Online monitoring technique for aluminum electrolytic
capacitor in solar PV-based DC system. IEEE Trans. Industr. Electron. 63(11), 7059–7066
(2016)
13. X. Li, Q. Yang, Z. Lou, W. Yan, Deep learning based module defect analysis for large-scale
photovoltaic farms. IEEE Trans. Energy Convers. 34(1), 520–529 (2019)
14. C.B. Ahmed*, M. Kassas, S.E. Ahmed, PV-standalone monitoring system performance using
LabVIEW. Int. J. Smart Grid Clean Energy 3(1), 44–50 (2014)
15. A. Tyagi, M. Dubey, D.S. Gawre, Advance monitoring of electrical and environmental param-
eters of PV system: a review, in 2018 International Conference on Sustainable Energy,
Electronics, and Computing Systems (SEEMS), Greater Noida, India, pp. 1–5 (2018)
16. A.A. Stonier, B. Lehman, An intelligent-based fault-tolerant system for solar-fed cascaded
multilevel inverters. IEEE Trans. Energy Convers. 33(3), 1047–1057 (2018)
17. W. Kolodenny, M. Prorok, T. Zdanowicz, N. Pearsall, R. Gottschalg, Applying modern infor-
matics technologies to monitoring photovoltaic (PV) modules and systems, in 2008 33rd IEEE
Photovoltaic Specialists Conference, San Diego, CA, USA, pp. 1–5
18. M. Zahran, Y. Atia, A. Alhosseen, I. El-Sayed, Wired and wireless remote control of PV system.
WSEAS Trans. Syst. Control 5, 656–666 (2010)
19. J. Lee, Y. Su, C. Shen, A comparative study of wireless protocols: Bluetooth, UWB, ZigBee,
and Wi-Fi, in IECON 2007 – 33rd Annual Conference of the IEEE Industrial Electronics
Society, Taipei (2007), pp. 46–51
20. N. Baker, ZigBee and Bluetooth strengths and weaknesses for industrial applications. Comput.
Control Eng. J. 16(2), 20–25 (2005)
Soft Computing Methods and Its
Applications in Condition Monitoring
of DGS—A Review

Puneet Joshi, Sanjay Agrawal, Lokesh K. Yadav, Medha Joshi, Vikas Patel
and Peeyush Kala

Abstract Soft computing is the use of conjecture, uncertainty, approximation, and

ambiguity to compute complex mathematics for which no known algorithm exists
that computes a precise result in polynomial time. Condition monitoring (CM) of a
system is the practice of observing a set of features defining health of it, such that
a deviation is symptomatic of a developing failure. CM of any man-made system is
the backbone of predictive maintenance. CM framework comprises of data acqui-
sition, data evaluation, feature extraction, judgment and finally diagnosis. DGS are
modular, decentralized, and flexible, which are dispersed near load, albeit having
capacities of less than or equal to ten megawatts (MW). DGS typically comprise
diverse sources, like, wind, solar, hydro, biomass, biogas, geothermal, and storage
components. DGS and CM together play a vital role in transcending conventional
distribution system into a grid with advanced sensitivity, automation, practicability,
and reliability. Each source can be operated either as grid-connected or stand-alone,
thus, the major challenges for CM are to monitor sources simultaneously to enhance
the reliability and operational efficiency of the grid. For a reliable system, compre-
hensive and continuous monitoring of operating parameters is necessary. It needs a
variety of communication protocols so that it can receive and handle parameters in

P. Joshi · S. Agrawal (B) · L. K. Yadav · V. Patel

EED, REC, Ambedkarnagar, Akbarpur 224122, India
e-mail: [email protected]
P. Joshi
e-mail: [email protected]
L. K. Yadav
e-mail: [email protected]
V. Patel
e-mail: [email protected]
M. Joshi
EED, SLSET Group of Institutions, Kichha, India
e-mail: [email protected]
P. Kala
Department of Electrical Engineering, Women Institute of Technology, Dehradun 248007, India
e-mail: [email protected]
© Springer Nature Singapore Pte Ltd. 2020 189
H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_8
190 P. Joshi et al.

different formats. CM system requires a data storage system for analysis of acci-
dents, damages, incipient faults reported during condition assessment of DGS. It
should be able to manage and remotely control the operation of the whole system.
Thus, this comprehensive survey, is felt obligatory with a motivation to reflect the
state-of-the-art developments, over the years and review the synthesis and applica-
tion methodology of various soft computing methods that are prevalent in CM and
diagnostics of DGS.

Keywords Artificial neural networks · Fuzzy logic · Genetic algorithms ·

Evolutionary algorithms · MetaHeuristics algorithms

Nomenclature

ANN Artificial Neural Networks

FL Fuzzy Logic
GA Genetic Algorithms
PSO Particle Swarm Optimization
PV Photovoltaic
SC Soft Computing
CM Condition Monitoring
DGS Distributed generation systems
GPS Global positioning system
WT Wavelet Transform
SCADA Supervisory Control and Data Acquisition System
HELM Hierarchical Extreme Learning Machine
SOM Self Organizing Map
DGA Dissolved Gas Analysis
STFT Short Time Fourier Transform
FALCON Fuzzy Adaptive Learning Control/Decision Network
WVD Wigner Ville Distribution
RBF Radial Basis Function
KF Kalman Filter
PCC Point of Common Coupling
FT Fourier Transform
FEM Finite Element Model
BP Back Propagation
1DCNN One-dimensional convolutional neural network
EMD Empirical Mode Decomposition
Soft Computing Methods and Its Applications … 191

1 Introduction

The conventional power system consists of large-scale interconnected network

between synchronous generating units, transmitting units and distribution units. With
the advent of renewable energy sources (RES), issue of power deficit is solved some
extent, but it also brings issues which rises more severe operational complexity and
uncertainty. Renewable sources are intermittent, dispersed and located at distribution
side, results in bidirectional power flow.
For a robust and steadfast operation of a power system, the parameters, i.e., MW,
MVAr, bus voltage, phase angle, and frequency, must always be monitored. Till date,
SCADA systems are widely employed for monitoring power systems which pro-
vides asynchronous measurements and is unable to measure phase angle. Phadke
[1] has proposed the concept of GPS based synchronized phasor measurement. Sub-
sequently, wide area measurement systems along with synchrophasor measurement
have emerged as a promising technology for the future. Therefore, to preserve the
reliability of the smart grid and efficient usage of the RES, power system CM plays
a key role in accomplishing operational flexibility.
Figure 1 depicts the general approach for CM of a system. The first step is data
collection and storage. For data storage in binary and floating-point formats, IEEE
C37.110 is widely used for the time sequence phasor measurement in WAMS. IEEE
C37.118 provides the standard for phasor data transmission.
Another very important aspect of data is its quality quantified as preciseness, legal-
ity and validity, reliability and firmness, timeliness and suitability, completeness and

ELECTRICAL
MACHINE

ROTATING
TRANSFORMERS
MACHINES

• Insulation ELECTRICAL MECHANICAL

• Spring
• Bearings
• Shaft
• Drive Mechanism • Stator/Rotor
• Solid State • Air Gap
Windings
• Eccentricity
• Stator/Rotor Core
• Bent Shaft
• Slip Ring and
• Bearing and Gear
Brushes
Box
• Rotor Bars

Fig. 1 Representation of faults in electrical Machines

192 P. Joshi et al.

extensiveness, availability and, granularity and uniqueness. To maintain the qual-

ity, data is primarily collected in either time, modal, frequency, or time–frequency
domain [2].
In time domain the signal is recorded in the form of data with time stamping
information. In time domain, kurtosis, RMS, variance, and means are usually used
to extract the meaningful information from raw data.
Structural faults are spotted by accurately measuring and examining the small
magnitude shifts in the natural frequency components. However, due to low magni-
tudes, small shifts are hard to detect [3] and, therefore, high-frequency modes of local
responses can be used but model overlap; and sensitivity to environmental conditions
is high in high-frequency modes in compression to the low-frequency modes.
Damping, a function of both frequency and mode, is the amount of energy dis-
sipated in a system and an indicative of structural behavior. Usually, an undamaged
structure has lower damping as compared to the damaged or a deteriorating structure
[4]. Commonly, it is applied to detect faults in composite materials.
Mode shapes are the function of damping, mass, stiffness, and boundary condi-
tions at different natural frequencies of the structure. Each “mode parameter” consists
of a resonant/modal frequency, modal damping, and a mode shape [5]. Mode shapes
are adequate for identification and detection of large faults [3]. However, they are
hard to compute, vulnerable to modal analysis noises. Thus, it never considers out-
side of the frequency bandwidth modes, and solely applied to lightly damped and
linear structures.
FT can be utilized to transform measured excitation and response into the fre-
quency domain as a ratio of the response to the excitation at every frequency called
the frequency response. It is best suited for structures having high damping and modal
density. One measurement provides enough data that comprises out-of-frequency-
bandwidth modes. Modal identification errors are also eliminated. Nevertheless, they
are difficult to implement as they contain trivial data not required for fault detection.
Also, no methodology exists for the frequency bandwidth of interest calculation.
Some faults transform linear structures to nonlinear, e.g., cracks caused by fatigue
failures, are analyzed using time–frequency approaches. These faults generate a
nonstationary signal. Thus, time–frequency approaches, e.g., the STFT, WT, and
WVD have been applied. Feng et al. [6] have enlisted 20 such methods applied for
machinery fault diagnosis.
Data analysis involves the process of extracting key features from redundant data.
Feature selection [5–8] an integral part of data analysis is the process of reducing
the dimensionality of collected data such that the useful information is extracted and
quantified. Signal processing technique is strong candidate for feature extraction.
Based on the extracted features the decision making for the condition diagnosis
depends.
CM methods can be categorized broadly into correlation-based method, the FEM
Updating method, and SC methods. In recent times, application of computational
approaches for prognosis of the electrical systems has been increased.
Soft Computing Methods and Its Applications … 193

This chapter has been structured as follows. In second section, prevalent applica-
tions of the SC approaches for CM are elaborated. Concluding remarks and summary
of the article has been elaborated in Sect. 3.

2 Soft Computing

SC is a domain of computer science pertaining with designing of systems that exhibit

some traits that we usually associate with human intelligence. Soft computing meth-
ods compromise with accuracy and precision to find a robust and low-cost solution.
Some of the popular soft computing approaches are ANN, FL, GA, Expert Systems,
etc.

2.1 Artificial Neural Networks

ANNs are the simplified representation of a biological nervous system. It imitates

the memorization and the generalization capability through a massively parallel and
highly interconnected network of neurons. ANN learns through various mechanisms
like supervised, unsupervised or reinforced learning. Some of the recognized NN
systems include BP, SOM, RBF, State Vector Machines, etc.
Recently, research subjected to health assessment of electrical machines has
picked up a remarkable pace. Its manufacturers and consumers are keenly inter-
ested in software-based diagnostic features to improve scalability and reliability.
In addition to identifying specific harmonic components of current, speed, torque,
noise, vibration signals, etc., are also examined along with their frequency spectrum.
In some of the appliances, combination of thermal measurements, chemical analysis,
etc., along with conventional techniques are also used to determine the behavior and
severity of the fault.
The main faults of rotating electrical machines can usually be categorized as (1)
stator windings faults, (2) broken rotor bar or fractured rotor end rings, (3) static
and/or dynamic air gap irregularities, (4) eccentricity, (5) rotor field winding faults,
and (6) bearing and gearbox failures [9]. Particularly, faults associated with bearing,
stator, broken bars, and end ring are the most prevalent faults and can be detected
by identifying: (1) unbalance in line currents and air gap voltages, (2) increment in
torque pulsation, (3) decrement in mean value of torque, (4) increment in losses,
and (5) immoderate heating. Numerous invasive and noninvasive CM techniques
have been introduced by several researchers. Benbouzid et al. [10] provide a good
literature survey of MCSA techniques.
Of late, ANN has been used for modeling, protection, and control of electrical
apparatus. Over the years, one of the major applications of ANN is fault detection
and classification. There are several review articles published on CM for machines,
out of these, the most appropriate is [11] dealing with large machines, [12, 13] related
194 P. Joshi et al.

to induction drives, [14] exploring winding and insulation problems; [15] aimed at
electrical equipment monitoring, and diagnostics. Figure 1 summarizes some of the
CM issues in various equipments.
Induction motors are the workhorse of modern industry. The occurrence of
machine faults results in variation in electrical and mechanical forces developed
in a machine. These forces mostly affect the bearings, and health condition of bear-
ings depends upon the type and intensity of the fault. The alteration in the vibrations
under normal operations give rise to vibration harmonics. Induced harmonics upset
the entire frequency spectrum of the vibration signal, but small number of compo-
nents play significant role in the diagnosis of machine faults. Defects monitoring is
most effective technique for finding the health condition of bearings [16].
Phase unbalancing, short circuiting of windings, broken rotor bars, and eccentric
running cause harmonics in line currents, vibrations in the structure and leakage
fluxes in induction drives [14]. Penman and Yin [17] used Kohonen Feature Maps
[18] for identifying faults in induction motors. They also indicated the superiority
of unsupervised learning over MultiLayered Perceptrons (MLP) that require a pri-
ori training data. Dybkowski and Klimkowski [19] demonstrated a stator current
fault detection algorithm for the vector control of an induction drive. Sometimes
abnormalities like complete or partial loss of sensor signals, locking of actuators,
short circuits, and sudden disconnections of system components occur that may lead
to system failure. Thus, for prevention of the system various techniques have been
devised in [20–22].
In gearbox, teeth breakage, wear, plastic flow, spalling, scuffing, and deterioration
of the roller of bearing are the defect conditions which usually reported at the time of
gearbox health monitoring or scheduled maintenance. Wu et al. [23] demonstrated
that the learning of features can be done by the 1-DCNN. In this type, network
learning is done from the raw vibration signals which results accurate fault diagnosis
system for fixed-shaft and planetary gearboxes. Malik et al. [24] utilized vibration
signals of gearbox and uses the EMD for feature extraction and fault identification
is done through these features by using ANN. In [25] with the advent in technology,
sensor data collection has become faster and voluminous. Cloud computing has been
employed by [26] for the big data processing for CM of the gearbox in a wind turbine
using (HELM) algorithm [27].
Lithium-ion battery has become essential for an energy storage system for dif-
ferent renewable sources and are promptly used for various applications especially
electric vehicles (EVs). Determination of state-of-health (SoH) is essential for CM
of battery-based applications. Thus, to study the various parameters of the battery
such as charging and discharging voltage, life cycle, temperature, etc. accurate bat-
tery modeling is required. Empirical modeling of the battery can be achieved using
genetic programming, FL, KF, ANN, etc. Sarmah et al. [28], Kheirandish et al. [29]
proposed a kernel function based SVM model for the commercial proton exchange
membrane fuel cell. Sorrentino et al. [30] employed a recurrent and a simple ANN
with a statistical model to incorporate the nonlinearity of the cell. Shao et al. [31]
used ANN for fault detection in fuel cells. Morando et al. [32] employed ANOVA
method for aging prediction using echo state networks for a fuel cell.
Soft Computing Methods and Its Applications … 195

Power transformers vibrate under normal operating conditions due to magne-

tostriction. However, alarming situations arise due to loosening of the core clamping
bolts, repeated switching of transformers, external short circuit, rapidly fluctuating
loads. Vibration in a transformers may result in damage to the winding insulation
[33]. Siva Sarma and Kalyani [34] stated that the DGA methods lack the range and
thus, proposed a novel BP-ANN based approach for CM of power transformers. Two
separate networks were trained for gas ratios obtained by Rogers method and IEC
method.
The radiometric sensor data provided by a radiometric sensor mounted on an
unmanned aerial vehicle was analyzed using ANN for dust detection on the PV
panels. García Márquez and Segovia Ramírez [35] Condition-based maintenance
(CBM) is much more efficient approach to health monitoring compared to time-based
methods. Three approaches, viz., Logical Analysis of Data, ANN and Proportional
Hazard Models were compared and ANN was concluded to be the most favored [36–
38] demonstrated the use of ANN to maintain power quality in distribution systems by
operating ultra-rapid switching devices [39]. Levenberg–Marquard-based multilayer
perceptron for determining the energy and exergy efficacies for thermodynamically
optimizing a district geothermal heating system. Abd-Elkader et al. [40] proposed
a novel passive approach for islanding detection for wind DGS using ANN. The
symmetrical components of second harmonics of current and voltage measurements
were extracted using Fourier transform and fed to ANN. Finally, the ANN was
employed to pinpoint the abnormality at the PCC to differentiate between power
quality and islanding events.

2.2 Fuzzy Logic(FL)

FL is expression of human intelligence and natural language semantics empowered

actions to be made based on the imprecise information. It is an extension of the
classical set theory and uses linguistic variables to represent real–world inputs and
outputs. The real-world data is fuzzified using input membership functions and the
output is inferred using the If-Then rules. The output is defuzzied back to the real-
world form using the output membership functions.
Singh and Reisner [41] used FL to determine the SoH of lead acid batteries. Zenati
et al. [42, 43] employed FL to estimate the SoC and SoH of Li-ion cells. [44, 45]
employed FL to detect and diagnose fault in a PWM-VSI induction drive. FL-based
approach is fed with the output inverter current to diagnose misfiring of pulses by
identifying a shift in the Concordia reference frame [46].
Stator faults of an IM can be detected by feeding the RMS value of phase currents
to a FIS [47]. Merabet et al. [48] detected the short/open circuiting of the stator
windings in a DFIG using FL. Benbouzid and Nejjari [49] employed an asymmetrical
FIS for identifying the stator faults using the stator line currents. Akar and Cankaya
[50] used FL along with stator current analysis for detecting crack in rotor bars of a
SCIG.
196 P. Joshi et al.

Awadallah and Morcos [51] studied the [52] for the identification and diagnosis of
a faulty PM-BLDC due to open-switch conditions. Daubechies (db2) mother wavelet
was employed to extract diagnostic indices from the DC link current signal. Finally,
a zero-order-Sugeno ANFIS model has been trained from the wavelet data. Vicente
et al. [53], Mini and Ushakumari [54] proposed use of FL for spotting wrecked rotor
bars, eccentricities, and inter-turn short circuit of stator windings by monitoring the
motor phase currents. Mini and Ushakumari [54] proposed two novel approaches (1)
FALCON, (2) ANFIS, for fault diagnosis problems in induction motors is used in
which, current, speed and torque is taken as inputs and friction as output. Lindh et al.
[55] employed FL and a multivariate statistical fault classifier to detect the cyclic
bearing faults in a rotating machine. Qualitative features have been extracted from
an envelope spectrum of the acceleration frame of a motor. Mahalanobis distance-
based statistical classification method was used to generate the quantitative measure
[55]. Both the qualitative and the quantitative measures were used as inputs to the
FIS to evaluate the probability of a fault. Higher order statistical (HOS) analysis is
used for feature extraction from the vibration data which yields bispectrum, power
spectral density, and bicoherence signatures [56]. This data is then reduced and
parametrized and fed to the FIS for diagnosis of the machine fault. The results
of FL are compared with ANN to prove its superiority [57]. ANFIS and its noise
model constituent represent a mth-order hidden Markov model and it utilized by
high-order particle filter to predict the probability density function of the fault [58].
Goode and Chow [57] utilized ANFIS to detect bearing and insulation faults in a
motor. Machine health can be forecasted using ANFIS and Bayesian networks [59,
60]. The trained model prognoses the fault state with time. The model is trained
based on a probability distribution function of ANFIS predictions and actual data.
Bayesian estimator (Particle Filtering) is used to enhance the degree of confidence
in the estimations.
Incipient faults in power transformers are detected typically by using DGA
method. There are plentiful procedures established for investigating faults using
the dissolved gases such as key Gas contents, Roger gas ratio, Doernenburg, IEC gas
ratio, and Duval Triangle. However, these methods fail when two or more types of
faults occur at the same time.
Abu-Elanien et al. [61] used FL approach to develop a health index that provides
an indication of a 69 kV or lesser rating transformer’s propinquity to end of life.
The approach relies on the oil analysis methods like the furan and the DGA method.
Gases librated from oil and paper decompositions are dissolved in transformer oil.
According to IEC-599, normal conditions are prevalent if dissolved gases are within
permissible limits as indicated in Table 1. Abu-Siada et al. [62]. Carbon monoxide

Table 1 Normal values of Dissolved gases in Transformer Oil as per IEC-599

GAS H2 CH4 C2 H6 C2 H4 C2 H2 CO2 CO
PPM 100 120 65 50 35 2500 350
Soft Computing Methods and Its Applications … 197

(CO) and carbon dioxide (CO2 ) are mainly produced due to decomposition of paper
and rest due to oil decomposition.
Abu-Siada et al. [63], Abu-Siada and Hmood [64] incorporated all DGA methods
like the Key Gas Method, Roger Ratio Method, and Duval Triangle Method in one
FIS system.
This limitation was mitigated using a FIS based on the Rogers, IEC, and Duval
method. Paper tensile strength is determined by calculating the degree of polymer-
ization (DP). However, furan measurements are preferred over DP as it is unfeasible
to cut paper insulation for sampling. Also, DP has a very strong correlation with
the furan content in the oil. Due to aging of the transformer paper furans dissolve
in the oil. Thus, by measuring its concentration in the oil paper’s conditions can be
assessed. ASTM D 5837 standard is a widespread guideline to measure the furan
concentrations in the transformer oil. Abu-Siada et al. [65] developed a furan con-
tent measurement approach using FL based on the ultraviolet to visible (UV-Vis)
spectroscopic analysis.
The asset management aids in determining the retirement/replacement or reloca-
tion of a transformer. A precise asset management provides assurance for longer life
and maintaining its rated output without hampering its practical integrity. IEEE Std.
C57.91-2011 and IEC Standard 60076-7 provide detailed modeling of transformer’s
remnant life as a function of the hot spot temperature inside the equipment. FL is
used for health management of transformers build on the basis of its remnant life
and aging rate [66, 67] (Table 2).
Flores et al. [68] demonstrated the use of two Type-2 FIS for CM of power
transformers. First FIS applied Duval’s triangle and gas ratios method to obtain the
results of DGA (IEC 60599). The second FIS evaluates the condition based on the
paper-oil system. FIS output is linguistic variables easily interpretable by the utility
personnel.
Tomsovic [69], Momoh et al. [70] have discussed the applications of FL for
protection, monitoring, and control in power systems. Rubio et al. [71], Flores et al.
[72] used FL for controlling solar power plant. Holmes and Ray [73] have used FL to
improve reliability of a fossil fuel-based power plant. Holbert and Lin [74] applied
FL to monitor the level of steam generation in a nuclear power plant.
FL has been widely employed for power plant monitoring and control. The FL-
based diagnostic monitoring through data acquired from instrumentation is described

Table 2 Parameters for asset

Class Indicative parameter
management and remnant life
decision model Remnant life estimation Furan, H2 O, O2 , Working
temperature rise
Paper insulation Furan, CO, CO2
Faults severity CH4 , C2 H2 , C2 H4 , C2 H6
Impurity in oil Interfacial tension
Relative accelerated Aging H2 O, O2 , operating
severity temperature rise
198 P. Joshi et al.

in [75] for accessing the system status and the probable requirement for equip-
ments upkeep of a nuclear power plant. Baggenstoss [76] proposed the maintenance
scheduling of offshore wind systems. The approach used a hierarchical FIS-based
maintenance advisor and optimizer that handled the overall maintenance activities
and operational variations of a group of substations in a smart grid. The optimizer
initially generated the maintenance plan for every substation fulfilling multiple objec-
tives related to the design or average operational conditions. However, during oper-
ation, problems may arise due to aging, control operations, changing weather load
fluctuations, etc. Therefore, the maintenance advisor, at each substation, shall assess
the reliability situation and inform the optimizer accordingly which in turn may
reevaluate the optimization objectives. The Hidden Markov Model (HMM) can be
employed to identify plant accidents by its spatial and temporal data. The HMM was
trained using maximum likelihood algorithm and a Baum–Welch estimator [77]. The
system classified eight different accidents and normality using 22 input variables [78,
79] proposed an ANFIS predictor for monitoring and prognosis of machine through
a monitoring index. The proposed approach was evaluated through benchmark data
and compared with Support Vector Regression (SVR). Wu et al. [80] combined the
multiscale entropy method, Gray relation Analysis (GRA) and FL for predicting
faults in stator windings and optimal settings for running an induction drive. Current
and vibration data at various operating speeds were generated for normal and faulty
motor and denoised using WT. Andruşcă et al. [81] identified faulty gears using WT
for feature extraction from vibration signal data and feeding its output to the FL.
Castillo and Melin [82] demonstrated the use of FL for health monitoring of power
circuit breakers. Type-2 FL and Fractal Theory were combined for plant monitoring
and maintenance. Fractal dimension is extracted from the time series data of the
measured parameters. The fractal dimensions are then fed as input to the FL designed
to identify certain patterns in the measured data. Castillo et al. [83], Atanassov [84]
described the use of intuitionistic FIS [85] for monitoring a nonlinear dynamic plant.
In [86], WT and ANFIS-based bearing fault diagnosis systems are designed to
identify the inner race, outer race, and ball defects of ball bearings. The WT utilized
the accelerometer signals, measuring the vibrations data, and generated feature vec-
tors and fed it to the ANFIS for diagnostic classification. Chen et al. [87] used ANFIS
for the prognosis of bearings using vibration data and proved it to be better than the
Radial Basis Function Networks. Razik et al. [88] discussed the implementation of
ANFIS for fault diagnosis in railway track circuits.

2.3 Optimization Techniques

In [89], induction motor rotor broken bar diagnostic system is developed for balanced
and unbalanced supply. In this approach, spectrum of stator current is adopted for
the diagnosis of rotor fault. Genetic algorithm is adopted to find out the magnitude of
spectral lines. Authors have designed a FL approach based on amplitude of spectral
lines which determines the state of motor (abnormal connection, load level, rotor
Soft Computing Methods and Its Applications … 199

fault) and gives the information about rotor fault severity. Authors have performed
several experimental tests to prove the performance under different load levels and
fault severities. Tang and Wu [90] genetic algorithm is applied to extract risk attributes
from the WT output of line currents of faulty induction motors.
In [91] provides a comprehensive insight on the use of computational intelligence
for CM of power transformers. Kanović et al. [92] presented a comprehensive review
on the various uses of PSO in power systems. Cerrada et al. [93] presented a fault
identification approach for an electrical machine which is based on fine-tuning of
the SVM through PSO. Chen et al. [94] proposed a combination of GA and random
forest algorithm for fault diagnosis in spur gears. Gao and Ovaska [95] designed
WT-SVM and immune-GA-based gearbox fault diagnosis system.
In [96] used Elman Neural Network (ENN) for fault detection of gearboxes. The
ENN was trained using GA to enhance its approximation capability. Peimankar et al.
[97] demonstrated the use of PSO for determination of health of Li-Ion batteries. Chou
and Ghaboussi [98] proposed a binary multi-objective PSO for fault classification in
power transformers.
In [99] used GA to identify structural damage. Kang et al. [100] presented a novel
combination of ANN, PSO, and time–frequency representation for induction drive
fault classification. The time–frequency data is represented such that the separability
between classes is maximized. Optimality of feature vector is achieved using PSO.
Finally, the classifies is implanted using ANN. Kang et al. [100] proposed immunity
enhanced PSO for crack detection in structures.

3 Summary

Apart from those enlisted above, some other typical uses of SC methods are [101]
illustrated the various prevalent approaches for CM of wind turbines. Brezak et al.
[102] estimated tool wear using a fuzzy classifier. Azadegan et al. [103] reviewed the
use of FL in manufacturing industry. Çelik and Bayir [104] used FL to identify faults
in IC engines. FL can also be used to measure groundwater vulnerability [105]. FL
is also been used to find an energy-efficient cluster in WSN [106]. Tumer and Bajwa
[107] presented the various soft computing methods used for health monitoring of
aircraft engines. Thus, this chapter provides some global applications of SC methods
in CM of DGS have been outlined as introduction.

References

1. A.G. Phadke, Synchronized phasor measurements in power systems. IEEE Comput. Appl.
Power 6(2), 10–15 (1993)
2. T. Marwala, Condition monitoring using computational intelligence methods: applications in
mechanical and electrical systems (Springer, 2012)
200 P. Joshi et al.

3. T. Marwala. Fault identification using neural networks and vibration data, Doctoral
dissertation, University of Cambridge, 2001
4. C. Williams, O.S. Salawu, Damping as a damage indication parameter. in Proceedings- Spie
The International Society For Optical Engineering, vol. 3089, no. pt2, (1997), pp. 1531-1536
5. P. Guillaume, Modal Analysis (Vrije University, Brussels)
6. Z. Feng, M. Liang, F. Chu, Recent advances in time—frequency analysis methods for machin-
ery fault diagnosis: a review with application examples. Mech. Syst. Signal Process. 38(1),
165–205 (2013)
7. B. Xue, M. Zhang, W.N. Browne, X. Yao, A survey on evolutionary computation approaches
to feature selection. IEEE Trans. Evol. Comput. 20(4), 606–626 (2016)
8. G. Chandrashekar, F. Sahin, A survey on feature selection methods. Comput. Electr. Eng.
40(1), 16–28 (2014)
9. P. Vas, Parameter Estimation, Condition Monitoring, and Diagnosis of Electrical Machines,
Monographs in Electrical. Electronics. Engineering (Oxford University Press, 1993)
10. E. Benbouzid, Hachemi, E. Hachemi, A review of induction motors signature analysis as a
medium for faults detection. IEEE Trans. Ind. Electron. 47(5), 984–993 (2000)
11. P.J. Tavner, B.G. Gaydon, D.M. Ward, Monitoring generators and large motors, in IEE
Proceedings B—Electric Power Applications, vol. 133, no. 3 (1986), pp. 169–180
12. G.K. Singh, A. Saleh, A. Kazzaz, Induction machine drive condition monitoring and
diagnostic research-a survey. Electr. Power Syst. Res. 64(64), 145–158 (2003)
13. A. Siddique, G.S. Yadava, B. Singh, A review of stator fault monitoring techniques of induction
motors. IEEE Trans. Energy Convers. 20(1), 106–114 (2005)
14. G. Stone; E.A. Boulter; I. Culbert; H. Dhirani, Electrical Insulation for Rotating Machines,
Design, Evaluation, Aging, Testing and Repair (Wiley, IEEE Press, 2014)
15. S. Nandi, H.A. Toliyat, X. Li, Condition monitoring and fault diagnosis of electrical motors—a
review. IEEE Trans. Energy Convers. 20(4), 719–729 (2005)
16. A. Sharma, R. Jigyasu, L. Mathew, S. Chatterji, Bearing fault diagnosis using frequency
domain features and artificial neural networks, in Information and Communication Technology
for Intelligent Systems. Smart Innovation, Systems and Technologies, ed. by S. Satapathy, A.
Joshi, vol 107 (Springer, Singapore, 2019)
17. J. Penman, C.M. Yin, Feasibility of using unsupervised learning, artificial neural networks
for the condition monitoring of electrical machines, in IEE Proceedings - Electric Power
Applications, vol. 141, no. 6 (1994), pp. 317–322
18. T. Kohonen, Springer series in information sciences, in Self-Organization and Associative
Memory, vol. 8, 3rd edn. (Springer, Berlin Heidelberg, 2012)
19. M. Dybkowski, K. Klimkowski, Artificial neural network application for current sensors fault
detection in the vector controlled induction motor drive. Sens. (Basel) 19(3), 1–16 (2019)
20. R. Isermann, Fault-Diagnosis Applications, Model-Based Condition Monitoring: Actuators,
Drives, Machinery, Plants, Sensors, and Fault-Tolerant Systems (Springer, Berlin Heidelberg,
2011)
21. M.R. Boukhari, Proprioceptive sensors fault tolerant control strategy for an autonomous
vehicle. Sens. (Basel). 18(6), 1–24 (2018)
22. F. Hu, D. Luo, C. Luo, Z. Long, G. Wu, Cascaded robust fault-tolerant predictive control for
PMSM drives. Energies 11(11), 1–17 (2018)
23. C. Wu, P. Jiang, C. Ding, F. Feng, T. Chen, Intelligent fault diagnosis of rotating machinery
based on one-dimensional convolutional neural network. Comput. Ind. 108, 53–61 (2019)
24. H. Malik, Y. Pandya, A. Parashar, Feature extraction using EMD and classifier through arti-
ficial neural networks for gearbox fault diagnosis, in Applications of Artificial Intelligence
Techniques in Engineering. Advances in Intelligent Systems and Computing, ed. by H. Malik,
S. Srivastava, Y. Sood, A. Ahmad, vol. 697 (Springer, Singapore 2019)
25. H. Malik, R. Sharma, EMD and ANN based intelligent fault diagnosis model for transmission
line. J. Intell. Fuzzy Syst. 32, 3043–3050 (2017)
26. P. Qian, D. Zhang, X. Tian, Y. Si, L. Li, A novel wind turbine condition monitoring method
based on cloud computing. Renew. Energy 135, 390–398 (2019)
Soft Computing Methods and Its Applications … 201

27. J. Tang, C. Deng, G. Huang, Extreme Learning Machine for Multilayer Perceptron. IEEE
Trans. Neural Netw. Learn. Syst. 27(4), 809–821 (2016)
28. S.B. Sarmah, P. Kalita, A. Garg, X.D. Niu, X.W. Zhang, X. Peng, D. Bhattacharjee, A review
of state of health estimation of energy storage systems: challenges and possible solutions for
futuristic applications of Li-Ion battery packs in electric vehicles. J. Electrochem. Energy
Convers. Storage 16(4), 040801 (2019)
29. A. Kheirandish, N. Shafiabady, M. Dahari, I. Dino, M.S. Kazemi, Modeling of commercial
proton exchange membrane fuel cell using support vector machine. Int. J. Hydrog. Energy
41(26), 11351–11358 (2016)
30. M. Sorrentino, D. Marra, C. Pianese, M. Guida, F. Postiglione, K. Wang, On the use of neural
networks and statistical tools for nonlinear modeling and on-field diagnosis of solid oxide
fuel cell stacks. Energy Procedia 45, 298–307 (2014)
31. M. Shao, X. Zhu, H. Cao, H. Shen, An artificial neural network ensemble method for fault
diagnosis of proton exchange membrane fuel cell system. Energy 67, 268–275 (2014)
32. S. Morando, S. Jemei, D. Hissel, R. Gouriveau, N. Zerhouni, ANOVA method applied to
proton exchange membrane fuel cell ageing forecasting using an echo state network. Math.
Comput. Simul. 131, 283–294 (2017)
33. C. Booth, J.R. McDonald, The use of artificial neural networks for condition monitoring of
electrical power transformers. Neurocomputing 23(1–3), 97–109 (1998)
34. D.V.S.S. Siva Sarma, G.N.S. Kalyani, ANN approach for condition monitoring of power
transformers using DGA, in 2004 IEEE Region 10 Conference TENCON 2004, vol. 3 (Chiang
Mai, 2004), pp. 444–447
35. F.P. García Márquez, I. Segovia Ramírez, Condition monitoring system for solar power
plants with radiometric and thermographic sensors embedded in unmanned aerial vehicles.
Measurement 139, 152–162 (2019)
36. H. Lo, A. Ghasemi, C. Diallo, Comparative study on logical analysis of data (LAD), artificial
neural networks (ANN), and proportional hazards model (PHM) for maintenance prognostics.
J. Qual. Maint. Eng. 25(1), 2–24 (2019)
37. E. Siami-irdemoosa, S.R. Dindarloo, Prediction of fuel consumption of mining dump trucks
: A neural networks approach. Appl. Energy 151, 77–84 (2015)
38. B. Cannas, G. Celli, M. Marchesi, F. Pilo, Neural Networks for Power System Condition
Monitoring and Protection, vol. 23 (1998), pp. 111–123
39. A. Keçebaş, I. Yabanova, Thermal monitoring and optimization of geothermal district heating
systems using artificial neural network: a case study. Energy Build. 50, 339–346 (2012)
40. A.G. Abd-Elkader, D.F. Allam, E. Tageldin, Islanding detection method for DFIG wind
turbines using artificial neural networks. Int. J. Electr. Power Energy Syst. 62, 335–343 (2014)
41. P. Singh, D. Reisner, Fuzzy logic-based state-of-health determination of lead acid batteries,
in 24th Annual International Telecommunications Energy Conference (Montreal, Quebec,
Canada, 2002), pp. 583–590
42. A. Zenati, P. Desprez, H. Razik S. Rael, Impedance measurements combined with the fuzzy
logic methodology to assess the SOC and SOH of lithium-ion cells, in 2010 IEEE Vehicle
Power and Propulsion Conference (Lille, 2010), pp. 1–6
43. A. Zenati, P. Desprez, H. Razik, Estimation of the SOC and the SOH of li-ion batteries, by
combining impedance measurements with the fuzzy logic inference, in IECON 2010—36th
Annual Conference on IEEE Industrial Electronics Society (Glendale, AZ, 2010), pp. 1773–
1778
44. F. Zidani, D. Diallo, M.E.H. Benbouzid, R. Nait-Said, A fuzzy-based approach for the diagno-
sis of fault modes in a voltage-fed PWM inverter induction motor drive. IEEE Trans. Industr.
Electron. 55(2), 586–593 (2008)
45. P. Jayaswal, S. Verma, A. Wadhwani, Application of ANN, fuzzy logic and wavelet transform
in machine fault diagnosis using vibration signal analysis. J. Qual. Maint. Eng. 16(2), 190–213
(2015)
46. K. Bhardwaj, A. Agarawal, Fault diagnosis of three phase induction motor using fuzzy logic
controller and sequence analyzer. MIT Int. J. Electr. Instrum. Eng. 2(2), 112–118 (2012)
202 P. Joshi et al.

47. H.F. Azgomi, J. Poshtan, Induction motor stator fault detection via fuzzy logic, in 2013 21st
Iranian Conference on Electrical Engineering (ICEE) (Mashhad, 2013), pp. 1–5
48. H. Merabet, T. Bahi, N. Halem, Condition monitoring and fault detection in wind turbine
based on DFIG by the fuzzy logic. Energy Procedia 74, 518–528 (2015)
49. M.E.H. Benbouzid, H. Nejjari, A simple fuzzy logic approach for induction motors stator
condition monitoring, in IEMDC 2001. IEEE International Electric Machines and Drives
Conference (Cat. No.01EX485) (Cambridge, MA, USA, 2001), pp. 634–639
50. M. Akar, I. Cankaya, Broken rotor bar fault detection in inverter-fed squirrel cage induction
motors using stator current analysis and Fuzzy Logic. Turkısh J. Electr. Eng. Comput. Scı.
20, 1077–1089 (2012)
51. M.A. Awadallah, M.M. Morcos, Automatic diagnosis and location of open-switch fault in
brushless DC motor drives using wavelets and neuro-fuzzy systems. IEEE Trans. Energy
Convers. 21(1), 104–111 (2006)
52. J.R. Jang, ANFIS: adaptive-network-based fuzzy inference system. IEEE Trans. Syst., Man,
Cybern. 23(3), 665–685 (1993)
53. P. Vicente, J. Rodrı, M. Negrea, A. Arkkio, A simplified scheme for induction motor condition
monitoring. Mech. Syst. Signal Process. 22(5), 1216–1236 (2008)
54. V. P. Mini, S. Ushakumari, Incipient fault detection and diagnosis of induction motor using
fuzzy logic. 2011 IEEE Recent. Adv. Intell. Comput. Syst. 675–681, (Trivandrum, 2011),
55. T. Lindh, J. Ahola, P. Spatenka, A. Rautiainen, Automatic bearing fault classification com-
bining statistical classification and fuzzy logic. in 4th Nordic Workshop in Power & Industrial
Electronics (2004)
56. I. Lasurt, A.F. Stronach, J. Penman, A fuzzy logic approach to the interpretation of higher order
spectra applied to fault diagnosis in electrical machines, in PeachFuzz 2000. 19th International
Conference of the North American Fuzzy Information Processing Society— NAFIPS (Cat.
No.00TH8500) (Atlanta, GA, USA, 2000), pp. 158–162
57. P.V. Goode, M. Chow, Using a neural/fuzzy system to extract heuristic knowledge of incipient
faults in induction motors : part i-methodology. IEEE Trans. Ind. Electron. 42(2), 131–138
(1995)
58. C. Chen, B. Zhang, G. Vachtsevanos, M. Orchard, Machine condition prediction based on
adaptive neuro-fuzzy and high-order particle filtering. IEEE Trans. Industr. Electron. 58(9),
4353–4364 (2011)
59. N. Friedman, D. Geiger, M. Goldszmidt, Bayesian network classifiers. Mach. Learn. 29(2-3),
131–163 (1997)
60. C. Chen, B. Zhang, G. Vachtsevanos, Prediction of machine health condition using neuro-
fuzzy and bayesian algorithms. IEEE Trans. Instrum. Meas. 61(2), 297–306 (2012)
61. A.E.B. Abu-Elanien, M.M.A. Salama, M. Ibrahim, Calculation of a health index for oil-
immersed transformers rated under 69 kV using fuzzy logic. IEEE Trans. Power Deliv. 27(4),
2029–2036 (2012)
62. A. Abu-Siada, M. Arshad, S. Islam, Fuzzy logic approach to identify transformer criticality
using dissolved gas analysis (Providence, RI, IEEE PES General Meeting, 2010), pp. 1–5
63. A. Abu-Siada, S. Hmood, S. Islam, A new fuzzy logic approach for consistent interpretation
of dissolved gas-in-oil analysis. IEEE Trans. Dielectr. Electr. Insul. 20(6), 2343–2349 (2013)
64. A. Abu-Siada, S. Hmood, A new fuzzy logic approach to identify power transformer criticality
using dissolved gas-in-oil analysis. Int. J. Electr. Power Energy Syst. 67, 401–408 (2015)
65. A. Abu-Siada, S.P. Lai, S.M. Islam, A novel fuzzy-logic approach for furan estimation in
transformer oil. IEEE Trans. Power Deliv. 27(2), 469–474 (2012)
66. M. Arshad, S.M. Islam, A. Khaliq, Fuzzy logic approach in power transformers management
and decision making. IEEE Trans. Dielectr. Electr. Insul. 21(5), 2343–2354 (2014)
67. N.A. Bakar, A. Abu-Siada, Fuzzy logic approach for transformer remnant life prediction and
asset management decision. IEEE Trans. Dielectr. Electr. Insul. 23(5), 3199–3208 (2016)
68. W.C. Flores, E.E. Mombello, J.A. Jardini, G. Rattá, A.M. Corvo, Expert systems with appli-
cations expert system for the assessment of power transformer insulation condition based on
type-2 fuzzy logic systems. Expert. Syst. Appl. Int. J. 38(7), 8119–8127 (2011)
Soft Computing Methods and Its Applications … 203

69. K. Tomsovic, Fuzzy systems applications to power systems, Chapter V in Evolutionary

Techniques and Fuzzy Logic in Power Systems (2012), pp. 41–52
70. J.A. Momoh, X.W. Ma, K. Tomsovic, Overview and literature survey of fuzzy set theory in
power systems. IEEE Trans. Power Syst. 10(3), 1676–1690 (1995)
71. F.R. Rubio, M. Berenguel, E.F. Camacho, Fuzzy logic control of a solar power plant. IEEE
Trans. Fuzzy Syst. 3(4), 459–468 (1995)
72. A. Flores, D. Saez, J. Araya, M. Berenguel, A. Cipriano, Fuzzy predictive control of a solar
power plant. IEEE Trans. Fuzzy Syst. 13(1), 58–68 (2005)
73. M. Holmes, A. Ray, Fuzzy damage-mitigating control of a fossil power plant. IEEE Trans.
Control Syst. Technol. 9(1), 140–147 (2001)
74. K.E. Holbert, K. Lin, Nuclear power plant instrumentation fault detection using fuzzy logic.
Sci. Technol. Nucl. Install. 2012, 1–11 (2012)
75. C.S. Chang, Z. Wang, F. Yang, W.W. Tan, Hierarchical fuzzy logic system for implementing
maintenance schedules of offshore power systems. IEEE Trans. Smart Grid 3(1), 3–11 (2012)
76. P.M. Baggenstoss, A modified Baum-Welch algorithm for hidden Markov models with
multiple observation spaces. IEEE Trans. Speech Audio Process. 9(4), 411–416 (2001)
77. K.C. Kwon, J.H. Kim, Accident identification in nuclear power plants using hidden Markov
models. Eng. Appl. Artif. Intell. 12(4), 491–501 (1999)
78. B. Samanta, C. Nataraj, Prognostics of machine condition using soft computing. Robot.
Comput. Integr. Manuf. 24(6), 816–823 (2008)
79. M.U. Ahmed, D.P. Mandic, Multivariate multiscale entropy : A tool for complexity analysis
of multichannel data. Phys. Rev. E 84(6), 061918 (2011)
80. J.D. Wu, C.C. Hsu, G.Z. Wu. Fault gear identification and classification using discrete wavelet
transform and adaptive neuro-fuzzy inference. Expert Syst. Appl. 36(3), 6244–6255 (2009)
81. M. Andruşcă, M. Adam, A. Baraboi, A. Dragomir, A. Munteanu, Using fuzzy logic for
diagnosis of technical condition of power circuit breakers, in 2014 International Conference
and Exposition on Electrical and Power Engineering (EPE) (Iasi, 2014), pp. 268–273
82. O. Castillo, P. Melin, A hybrid fuzzy-fractal approach for time series analysis and prediction
and its applications to plant monitoring, in Power Plant Surveillance and Diagnostics, ed. by
D. Ruan, P.F. Fantoni (Berlin, Heidelberg, Power Systems. Springer, 2002), pp. 209–219
83. O. Castillo, A. Alanis, M. Garcia, H. Arias, An intuitionistic fuzzy system for time series
analysis in plant monitoring and diagnosis. Appl. Soft Comput. 7(4), 1227–1233 (2007)
84. K.T. Atanassov, On Intuitionistic Fuzzy Sets Theory, vol. 283 (Springer, 2012)
85. X. Lou, K.A. Loparo, Bearing fault diagnosis based on wavelet transform and fuzzy inference.
Mech. Syst. Signal Process. 18(5), 1077–1095 (2004)
86. F. Zhao, J. Chen, L. Guo, X. Li, Neuro-fuzzy based condition prediction of bearing health. J.
Vib. Control 15(7), 1079–1091 (2009)
87. J. Chen, C. Roberts, P. Weston, Fault detection and diagnosis for railway track circuits using
neuro-fuzzy systems. Control. Eng. Pract. 16(5), 585–596 (2008)
88. H. Razik, M.B.d.R. Correa, E.R.C. da Silva, A novel monitoring of load level and broken
bar fault severity applied to squirrel-cage induction motors using a genetic algorithm. IEEE
Trans. Ind. Electron. 56(11), 4615–4626 (2009)
89. L. Cristaldi, M. Lazzaroni, A. Monti, F. Ponci, F.E. Zocchi, A genetic algorithm for fault
identification in electrical drives: a comparison with neuro-fuzzy computation, in Proceed-
ings of the 21st IEEE Instrumentation and Measurement Technology Conference (IEEE Cat.
No.04CH37510), vol. 2 (Como, 2004), pp. 1454–1459
90. W.H. Tang, Q.H. Wu, Condition Monitoring And Assessment of Power Transformers Using
Computational Intelligence (Springer Science & Business Media, 2011)
91. M.R. AlRashidi, M.E. El-Hawary, A survey of particle swarm optimization applications in
electric power systems. IEEE Trans. Evol. Comput. 13(4), 913–918 (2009)
92. Ž. Kanović, M.R. Rapaić, Z.D. Jeličić, Generalized particle swarm optimization algorithm—
theoretical and empirical analysis with application in fault detection. Appl. Math. Comput.
217(24), 10175–10186 (2011)
204 P. Joshi et al.

93. M. Cerrada, G. Zurita, D. Cabrera, R.V. Sánchez, M. Artés, C. Li, Fault diagnosis in spur gears
based on genetic algorithm and random forest. Mech. Syst. Signal Process. 70–71, 87–103
(2016)
94. F. Chen, B. Tang, R. Chen, A novel fault diagnosis model for gearbox based on wavelet
support vector machine with immune genetic algorithm. Meas. J. Int. Meas. Confed. 46(1),
220–232 (2013)
95. X.Z. Gao, S.J. Ovaska, Genetic algorithm training of Elman neural network in motor fault
detection. Neural Comput. Appl. 11(1), 37–44 (2002)
96. B. Long, W. Xian, L. Jiang, Z. Liu, An improved autoregressive model by particle swarm
optimization for prognostics of lithium-ion batteries. Microelectron. Reliab. 53(6), 821–831
(2013)
97. A. Peimankar, S.J. Weddell, T. Jalal, A.C. Lapthorn, Evolutionary multi-objective fault
diagnosis of power transformers. Swarm Evol. Comput. 36, 62–75 (2017)
98. J.H. Chou, J. Ghaboussi, Genetic algorithm in structural damage detection. Comput. Struct.
79(14), 1335–1353 (2001)
99. A. Medoued, A. Lebaroud, A. Laifa, D. Sayad, Classification of induction machine faults
using time frequency representation and particle swarm optimization. J. Electr. Eng. Technol.
9(1), 170–177 (2014)
100. F. Kang, J.J. Li, Q. Xu, Damage detection based on improved particle swarm optimization
using vibration data. Appl. Soft Comput. J. 12(8), 2329–2335 (2012)
101. F.P. García Márquez, A.M. Tobias, J.M. Pinar Pérez, M. Papaelias, Condition monitoring of
wind turbines: Techniques and methods. Renew. Energy 46, 169–178 (2012)
102. D. Brezak, D. Majetic, T. Udiljak, Tool wear estimation using an analytic fuzzy classifier and
support vector machines. J. Intell. Manuf. 23(3), 797–809 (2012)
103. A. Azadegan, L. Porobic, S. Ghazinoory, P. Samouei, A. Saman, Fuzzy logic in manufacturing
: A review of literature and a specialized application. Int. J. Prod. Econ. 132(2), 258–270 (2011)
104. M.B. Çelik, R. Bayir, Fault detection in internal combustion engines using fuzzy logic,
in Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile
Engineering, vol. 22, no. 5 (2007), pp. 579–587
105. R.C.M. Nobre, O.C.R. Filho, W.J. Mansur, M.M.M. Nobre, C.A.N. Cosenza, Groundwater
vulnerability and risk mapping using GIS, modeling and a fuzzy logic tool. J. Contam. Hydrol.
94(3–4), 277–292 (2007)
106. A.K. Singh, S. Goutele, S. Verma, N. Purohit, An Energy Efficient Approach for Clustering
in WSN using Fuzzy Logic, vol. 44 (2012), pp. 8–12
107. I. Tumer, A. Bajwa, A survey of aircraft engine health monitoring systems, A. Engine, and
H. Monitoring, 35thAIAA/ASME/SAE/ASEE Joint Propulsion Conference and Exhibit (Los
Angeles, California, 1999), pp. 1–8
Introduction to Conditioning Monitoring
of Mechanical Systems

Akhilesh Kumar Choudhary and Dilshad Ahmad Khan

Abstract This chapter deals with the impact of condition monitoring in mechanical
system. Over the past few decades condition monitoring of the mechanical system is
most interesting area for researchers. The extensively used mechanical equipments
in severe conditions are subjected to failure and progressive deteriorations from their
initial condition. For this suitable parameter has to be selected that indicates real-
time internal conditions and fault incidence of the mechanical system. Therefore, to
early detection of failure that may cause system shutdown, continuous monitoring
of the system is more important aspect that reduces downtime of the systems and
it also saves money. In mechanical system fault occurs due to imbalance of forces,
fault in gearing system, bearing fault, loading conditions on shaft, and improper
machine selection for particular work. To detect the fault at early stage, the sys-
tem parameters like performance, vibrations, noise, temperature, pressure, wear at
movable, and lubrication monitoring are required. In this regard, for monitoring the
conditions of system parameter several sensors have been used to detect real-time
conditions. The sensors signal is acquired in terms of electrical signal which has
been processed by signal processing techniques and analyzed. In-process sensing
and control of parameters are required to smooth operation of system. The use and
benefits of sensor/transducers and advanced signal processing are having their own
strengths and limitations. This chapter is intended to provide fundamentals of various
condition monitoring techniques as well as signal processing methods with applica-
tion in mechanical system. It also describes the various transducers used to provide
the corresponding signals for condition monitoring, has been discussed. This chapter
also intended that applications of condition monitoring in various mechanical system
like identification of bearing defects, Vibration monitoring of rotary machine, cutting
tool monitoring, machine tool monitoring, power plant monitoring, vehicles, guiding

A. K. Choudhary · D. Ahmad Khan (B)

Mechanical Engineering Department, National Institute of Technology, Hamirpur 177005,
Himachal Pradesh, India
e-mail: [email protected]
A. K. Choudhary
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 205

H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_9
206 A. K. Choudhary and D. Ahmad Khan

rails, 3D printing, identification of gears defects, lubricant condition monitoring, rail-

way crossing, wind turbines, surface texture condition monitoring, grinding wheel
condition monitoring, etc.

Keywords Condition monitoring · Mechanical system · Rotary machine · Cutting

tool · Machine tool · IC engine · Vibrations

1 Introduction

Mechanical systems are commonly used in industrial applications. The extensively

used mechanical equipments in severe conditions are subjected to failure and progres-
sive deteriorations from their initial condition. The damage and fault in mechanical
components can take place at any time and any conditions which cause severe harmful
results in terms of brake down or production delay. The prognosis of unexpected fault
and brake down of mechanical equipments at early stage of operation are important.
In order to avoid unexpected failure, condition monitoring process is involved, which
helps in early detection of fault and reduces downtime of the systems and it also saves
money. The fault on mechanical system can be classified as fault occurs due to imbal-
ance of forces, fault in gearing system, bearing fault, loading conditions on shaft and
improper machine selection for particular work. Therefore, many researchers have
been focused on failures and fault diagnosis techniques for monitoring the health of
mechanical systems. Researchers have been developed many techniques like visual
inspections and sensors-based conditioning monitoring of mechanical systems. In
this regard, various methods have been established for condition monitoring like
lubricant analysis, thermography, performance analysis, wear debris (Ferrography),
and vibration analysis. But immediate response of signal changes in vibration sig-
nature has great advantages over other methods. The vibration signals are facili-
tated intermittent and permanent monitoring of the systems with greater flexibility
of signal processing. So vibration measurements are used broadly in machine con-
dition monitoring. For example machine housing vibrations are absolute vibration
and vibration between a shaft and the housing are relative vibration can be easily
identified using different vibration transducers which can be recorded and analyzed.
So it is most important to recognized and understand the information provided by
the vibration sensors. The time-domain analysis and frequency-domain analysis are
the common classical methods for condition monitoring. In most of the research,
the researchers also used the Fourier transform for the analysis. However, nowadays
applications of these methods are not sufficient for the development of profound con-
dition monitoring of the system. In this regard, several techniques, like Time Fourier
Transform, Hilbert–Huang Transform, Short Time Fourier Transform, Wigner–Ville
Distribution, wavelet transform, Cyclostationary Analysis, Spectral Correlation, and
Spectrum Analysis, are used for condition monitoring of the systems [1–5]. As the
results of that these methods are employed in various mechanical system to monitor
the condition like identification of bearing defects, monitoring of vibration of rotary
Introduction to Conditioning Monitoring of Mechanical Systems 207

machine, cutting tool monitoring, machine tool monitoring, power plant monitoring,
vehicles, guiding rails, 3D printing, identification of gears defects, lubricant condition
monitoring, railway crossing, wind turbines, surface texture condition monitoring,
grinding wheel condition monitoring, etc.
Recently plenty of excellent work has been introduced in the aspects of condi-
tion monitoring of mechanical equipments. Martin del Campo and Sandin [1] have
investigated the rolling-element bearings of rotating machinery which essential for
load carrying and to reduce the friction between moving components of rotating
machines, using a vibration signature. In this investigation dictionary distance and
signal fidelity-based heuristics have been used to fault detections and apart from this
vibration and acoustic emission signals are analyzed by time time-propagated fea-
tures and sparse approximation of the signals. In the condition monitoring process
comprises sensors signal acquisition and analysis of those sensors for detection of
abnormal behaviors and faults of the system. Teng et al. [2] has been Multi-sensors
for condition monitoring of rotary machines and suggested that multi-sensors-based
monitoring policy can effectively detect changes in system behavior status during
continuous operations.
For condition monitoring of multi-sensors architecture, automatic change detec-
tion was implemented with Autoregressive Integrated Moving Average (ARIMA)
regression and multidimensional time series analysis process. Bustos et al. [3] have
been used a graphical depiction of state configurations and chromogram of bands of
frequency for condition monitoring of the system. The signals of vibrations from
different mechanical components have been acquired and recorded at dedicated
database. The recorded data processed in the time domain, frequency domain and/or
time–frequency domains, before that data has been extracted in different groups and
power spectrum density of data were computed then average spectra were taken out.
Figure 1 represents condition diagnostic procedure for mechanical system.

2 Need of Condition Monitoring

Mainly condition monitoring deals with the maintenance aspects of the systems it
can be machine, plant or any mechanical systems. In order to smooth function of the
mechanical system, the present and past condition of the machine has to be known.
On the basis of condition monitoring analysis, appropriate maintenance or corrective
actions would be taken so that the system able to perform on their original design
functions. The unpredictable fault of mechanical system results in eventual loss to
the industry and affects the overall operations of the system. With help of condi-
tion monitoring techniques system can run without any significant breakdown and
it will be in proper conditions. In this aspect, different sensors have been installed
all around the mechanical system so that real-time information about the system’s
health condition can be monitored, analyzed, and decisions to be made regarding
smooth function of the system. Figure 2 shows fault detection system of mechanical
systems. In the mechanical system have of rotating parts and components, while
208 A. K. Choudhary and D. Ahmad Khan

Fig. 1 Basic condition

monitoring procedure of
mechanical system

operation, mechanical system provides information or signals in terms of vibra-

tions, wear, noise, pressure, temperature, lubricant conditions [5–8]. In condition
monitoring, these informations are collected through installed sensors and trans-
ducers then analyzed by appropriate signal processing and data analysis methods,
in this regards signal conditioning, amplifiers, and analog-to-digital converters are
involved. Nowadays, in the aspect of fault diagnosis of mechanical system, the new
and advanced technology of condition monitoring are promising tools for mechani-
cal system maintenance. Enhancement of technology and improve analysis methods
facilitate the system to get more accurate, predictable, and decision-making capabil-
ities. Improvements in sensors technology like self-calibration, signal conditioner,
instrumentation techniques, data handling, and data processing methods are making
systems more reliable and self-decision maker [9–11]. Figure 3 shows the various
stages of condition monitoring and faults detection of mechanical systems.
Introduction to Conditioning Monitoring of Mechanical Systems 209

Fig. 2 Fault detections

procedure of mechanical
systems

3 Condition Monitoring Methodology

There are several methods to identify faults on mechanical systems like observa-
tion and sensor-based monitoring. The observation or visual inspection technique
is simplest and most cost-effective for monitoring the condition of the system. In
this technique’s inspection is done by independently unaided eyes. Due to cheap and
easy to employ, this technique has advantage that immediate indication of system
conditions and provide fundamental scheme for condition monitoring. This tech-
nique evolves to identification of deviations in systems parts, parameter signals, and
assessment of specific features of the system components. The visual monitoring
involves an assessment of system behaviors in qualitative and quantitative manner
and active monitoring of equipment operational functions abnormal behavior from
normal operating conditions. In sensors-based monitoring method, the sensors and
210 A. K. Choudhary and D. Ahmad Khan

Fix threshold

Limit checking
Adaptive
threshold

Change detection method

Detection with
single signal Trend checking

Correlation

Signal model based Spectrum analysis

Wavelet analysis
Fault detection
Method
Param. Estimation

Neural network

Process Model based State observer

Detection with State estimation

multiple signals and
models
Paraty equation

Multi-variant data analysis Principle component analysis

Fig. 3 Various stages of condition monitoring and fault detection of mechanical system

computer system being used for active monitoring of system health. During operation
machine components provides information or signal in terms of sound, vibrations,
pressure, temperature, and wear. By installing sensors in system components this
information will be acquired by data acquisition and information logging, these sig-
nals are analyzed by suitable algorithms in computer system. The obtained signals
are usually analog in order to get appropriate these signals are converted into digital
signals which are efficiently analyzed by software [9–17]. Figures 4 and 5 show

Fig. 4 Condition monitoring methodology of mechanical system

Introduction to Conditioning Monitoring of Mechanical Systems 211

PaƩern
RecogniƟon Decision Tables

Bayes Classifier
ClassificaƟon StaƟsƟcal
Methods ClassificaƟon
Decision Tree

ApproximaƟon
Method Polynomial Classifier

Density Based
Geometrical Classifier
Methods
Fault Diagnosis
Method Fuzzy Classifier
ArƟficial
Intelligence
methods
Neural Network Classifier

Binary Reasoning Predicate Logic

Inference
Methods

Fuzzy Logic
Approximate
Reasoning
Neural Networks

Fig. 5 Fault diagnosis methodology of mechanical system

the condition monitoring and fault diagnosis methodology of any mechanical sys-
tem. Figure 6 elicited a signal model-based fault detection condition monitoring
methodology.

4 In the Condition Monitoring Techniques Various Fault

Diagnosis Methods Are Involves

Performance-based condition monitoring

Vibration signature-based condition monitoring
Noise monitoring
Pressure monitoring
Temperature monitoring
Wear monitoring
Lubrications monitoring.
212 A. K. Choudhary and D. Ahmad Khan

Bandpass Filtering Decision Tables

Bayes Classifier

Fourier Analysis

Decision Tree
Periodic Signals

ApproximaƟon
Method Polynomial Classifier

Density Based
Geometrical Classifier
Methods

Signal Model-
Based Fault StochasƟc Signals
DetecƟon Fuzzy Classifier
ArƟficial
Intelligence
methods
Neural Network Classifier

Binary Reasoning Predicate Logic

Non-staƟonary Approximate
Signals Reasoning

Fuzzy Logic
Approximate
Reasoning
Neural Networks

Fig. 6 Signal model-based fault detection condition monitoring methodology

4.1 Performance-Based Condition Monitoring

Performance of the system is a major criterion for system operation as well as it

monitoring is also an important parameter. Deterioration in system condition affects
the system performance and results in increased energy uses by the system. The key
process parameters can deviate from their actual performance with several reasons
like pressure fluctuations, leaks across the machine casing and rotating shaft, cracks
in machine parts, unbalance of forces, bearing faults and upstream–downstream of
flow rate. The loss of performance can be monitored by process parameters several
aspects like vibrations, acoustics emissions, pressure temperature, flow, speed, etc. of
the system [4, 6, 11, 15]. In performance monitoring basic parameters are taken into
account like mass, length, temperature, force, pressure, work, energy, power, rate of
flow, basic and specific rate of fuel consumption, torque, heat transfer rate, potential
difference, electric current, electric charge, electric resistance, electric capacitance
and inductance. These system parameters are measured in qualitative and quantitative
manners using plant instrumentations, data acquisition systems and data processing
units specifically designed to identify problems within process parameters. Some
Introduction to Conditioning Monitoring of Mechanical Systems 213

time these practical parameters can measure repeatedly to proven performance of

the system. Simultaneously performance data, spreadsheets, charts, and other data
can be stored using some software for future reference [1, 2, 18, 19, 20].

4.2 Vibration Signature-Based Condition Monitoring

In general, every machine is consisting of a prime mover, shaft and linkage, etc.
and machines have several rotating parts. The vibrations are due to events that are
occurred by machine operation such as combustion processes in IC engine, fluid flow
in turbine pump, etc. Normally, good condition machines are also generate vibrations.
Such vibrations are directly associated with intermittent events of machine opera-
tions such as rotating shafts, loading condition, meshing gear teeth, rotating electric
fields, and so on. The faults on these rotating parts have detected by vibration mon-
itoring [4, 16]. The vibration of machine elements which need to be measured has
been acquired in terms of signals using transducers and process to get meaning full
information for machine health conditions. The vibration of machine components
should be measured in all three mutually perpendicular directions. For this, triaxial
accelerometer should be used to acquire vibration signals at any location of machines,
to measurement of vibrations in all tree directions [4, 5, 16, 18]. In order to acquire
vibration signals in case of rotating shaft, transducer must be placed close to the
support of bearing. The vibration signals are related to rotating speed of the rotat-
ing components, rated power, foundation types, load conditions and other special
features which needs to be noticed while signal acquiring and recording because all
the predominant frequencies in the vibration signal and spectra are related to these
parameters [4, 5]. The vibration spectra can be acquired form the machines with
help of installing transducer at different locations in machine for analysis purposes,
While placing transducer, location and sensitivity of the transducer are important
and user should know about the amplitude of machine components on which corre-
sponding signal voltage level is acquired. These acquired signals are processed by
suitable signal processing techniques. Before signal analysis, an analog-to-digital
conversion system has been used. There different types of vibration are present in
mechanical systems like absolute vibration, relative vibration, and torsional vibra-
tions. Here absolute vibration can measured against a stationary point in space like
vibration of machine housing, relative vibration can measure against relative position
of the transducer with machine components like shaft and the housing which sepa-
rated by fluid film or journal bearing where transducer have been placed nearby the
bearing. The torsional vibration can measure against the angular velocity fluctuation
of machine components like rotor discs and gears. Using vibration signature sev-
eral machine fault can be identified like Misalignment, Cracked Shaft, Eccentricity,
Unbalanced Shaft, Bowed and Bent Shaft, Looseness of Bearing, and Gear related
Fault [10, 19, 20, 21].
214 A. K. Choudhary and D. Ahmad Khan

4.3 Lubrication Oil Condition Monitoring

Lubricating oils are used to reduce friction between mating surfaces. The lubricating
oil carries a product of deterioration which is generated by surface friction and these
products cause further deterioration in parts. The rate of machine parts deterioration
strongly depends on the operating condition of the machines parts, so lubricating
oil must be changed before machine parts start deterioration and its protective prop-
erties of the oil will lose. An unnecessary oil change affects the environment and
also causes economical effects. In order to avoid unnecessary change of the oil it
is important to know about the actual condition such as physical and chemical of
the lubricating oil. The qualitative and quantitative analysis of product deterioration
helps for early detection of machine failures. This information assesses the extent
and source of the deterioration within the machine itself. The oil used for lubrica-
tion carries different contamination products, which has to be developed by surface
friction like Aluminum, Antimony, Boron, Chromate, Copper, Iron, Lead, Silicon,
Silver, and Tin. These products cause lubrication degradation which results in loss
of lubrication properties in lubricating oil [9, 17, 22, 23]. It is an early alarm of
fault occurrence in machine components [19, 24]. Some research showed that con-
dition monitoring of lubricating oil provides warnings earlier for malfunction of
machine than other condition monitoring techniques. The viscosity of the lubricat-
ing oil, level of additives mixed in oil, moisture, specific gravity, oxidation related
properties, metallic particles, and wear debris analysis are used to analyze the lubri-
cating oil condition which shows the condition and purity of oil itself deteriorated
and suitability to fulfillment of its basic function of reduction in friction along with
wear between the machine parts. The lubricating oil monitoring can be divided into
several categories like chip detectors, In this technique is used to analyzed the chip
quantity, type, shape, size, and so on. For this, filters and magnetic plugs are used to
retain chips and other debris in lubricating oil. Spectrographic oil analysis procedure
(SOAP) is used to analyze spectrographic chemical components in the oil; sample
of lubricant has been collected at regular intervals then analyzed for wear debris,
viscosity, degradation, contaminants and additives in the oil. The ferrography is a
microscopic investigation of lubricating oil for analysis of debris present in the oil
which is retained by magnetically. The Electron spin resonance spectroscopy indi-
cates the long-level free radicals present in used oil, which contained contamination
in terms of wear products. Gas chromatography is used to detect very low levels
of added tracer elements of lubricating oil. Gradient elution chromatography is an
alternative technique of gas chromatography, which also divides products and com-
ponents of contamination presents in lubricating oil. The photoluminescence is used
to detect the presence of anti-oxidants and oxidation products in used oil. The dilu-
tion of oil and their acidity, pollution, anti-oxidants and their products can analyzed
by thin-layer chromatography and visible absorptionmetry methods for monitoring
the oil condition [17, 19, 22, 25].
Introduction to Conditioning Monitoring of Mechanical Systems 215

4.4 Acoustic Emission (AE) and Noise Monitoring

Acoustics are emitted sound waves of the machine and its requiring medium for
their propagation. It can be a longitudinal, transverse, bending, shear in nature. The
study of sound generation, propagation, and reception are important parameters for
monitoring of the system. The undesirable sound is called as noise which can affect
the human eyes also. The sound waves propagating are longitudinal in nature and in
solids they are usually shear or bending [4, 6, 16]. The particular frequency range of
sound can hear by the normal human beings, which is known as audible frequency
and its range is 20 Hz–20 kHz. In condition monitoring technique Acoustic Emis-
sion Testing (AET) has been used to identify the faults in machine components,
when machines enter under fault conditions, while it operations, an excessive noise
and vibrations are produced. The acoustic measurements or acoustic monitoring of
system produce advance warnings for future faults. This can be determined by per-
forming a detailed analysis of acoustic and noise signals produced by machines [25].
This monitoring system is passive means of scanning in which holding a microphone
or other acoustic sensor nearby the noise source and record various phenomena of
acoustic and noise. In mechanical system, the acoustic emission includes the fric-
tion, hitting of parts or impacting, cyclic fatigue, turbulence, cavitations, leakage and
material loss in components, etc., and this acoustic waves are propagate by surface of
the system. Most of the conventionally acoustic and noise monitoring system involve
by handheld instruments which can be fixed or locate at many point on the system to
sense the acoustic and noise signals for fault detections. In the purpose of acquiring
acoustic signals, AE (acoustic emission) sensor can be used, the sensing element
of the sensor produce charge when it is subjected to mechanical stress and respond
directly to stress waves acting on the system components. Usually AE sensors has
inbuilt electronics system comprises with filters and signal conditioning unit and
some more electronics components. The acoustic emission and noise monitoring
methods are widely used for rotating machines like shaft bearing, housing, and other
rotating parts [16, 18, 25].

4.5 Pressure Monitoring

Pressure is defined as force per unit area; it can be static and dynamic. In static
pressure, there is no motion on working medium while in dynamic pressure working
medium has motion and produces changes in the force which applied to its surround-
ings. Measurement of accurate dynamic pressure is very difficult, for these the total
and static pressures should be measured and then subtracted static pressures from it
to obtain the dynamic pressure [26]. These pressures will helps to determine the fluid
velocities and flow rates in dynamic systems. There are three methods for measuring
pressure (a) absolute pressure (b) gauge Pressure and (c) differential pressure. The
absolute pressure is measured with respect to reference to the pressure in a vacuum,
216 A. K. Choudhary and D. Ahmad Khan

while gauge and differential pressures are measured with respect to reference pres-
sure of another pressure like ambient atmospheric pressure [4]. In mechanical system
pressure measurements plays important role in monitoring the health of the system to
identify pressure drop, leakage, working medium velocity fluids flow rates and total
pressure difference in a system. For different measuring conditions, different types of
pressure sensors can be used these can be bridge-based, capacitive type, piezoelec-
tric material based and Optical type, depends upon ranges, and materials used and
requirements. These sensors convert pressures into some intermediate form, such as
deflection and displacement. This intermediate form converts the sensor signals into
electrical signals. For getting accurate reading of pressure sensor, the location or
pressure tap holes and probes must be even and free from any burrs or obstructions,
because they cause disturbances in the flow which will give incorrect interpretation
of measurement, these locations must have included any obstruction or change such
as bend, valves, flow splits, pumps, fans, etc. [25–27].

4.6 Temperature Monitoring

The sensation of hotness and coldness of the body is defined as thermometry which is
defined as temperature. With the help of temperature, the body behavior can be easily
identified , if it is normal then body behaves normally while if it is fluctuating which
indicates that faults being accompanied in the body. For example if human body
temperature rises this indicates the illness. Similarly in industrial system unexpected
increase or decrease in temperature shows the fault on the system. There are various
examples of defect occurrence unforeseen by an increase or decrease in temperature.
In fact that, before any physical damage accurses in the system, the temperature
starts fluctuation. For example: overheating of bearing and electrical components.
The temperature measurement facilitates to detection of fault before a catastrophic
failure occurs in the system, consequently well inform to scheduled maintenance
programmers [25, 28].

4.7 Wear Monitoring

The operation of system and mechanical equipment is concern with the transfer of
energy. Due to wear process in machine parts while operation, energy is dissipated.
This energy causes the component degradation and results in breakdown of the system
and deterioration in machine parts. The mechanical wear in system or equipment is
closely associated with dynamic behavior of the system. For monitoring of wear
in machine parts, wear debris generation analysis are taken into account. While
relatively moving machine parts and load-carrying surfaces like bearing, shafts, gears,
cam, etc. are in contact then debris is generated and it has direct consequences for
wear in machine parts. With the help of wear particles and wear debris status of the
Introduction to Conditioning Monitoring of Mechanical Systems 217

surfaces can be identified and it also provides the significant information about the
conditions at which wear debris is generated. In wear process several activities occur
while two surfaces are in relative motion like microcutting, local heating, molecular
interaction, plastic and elastoplastic deformation, Surface fatigue and Oxidation.
Because of these activities actual form of wear is present in contact surfaces like
abrasive wear, fatigue wear, adhesive wear, and cavitation wear. To monitoring of
wear of mechanical parts, the physical structure of contacted surface and operating
environment should be studied and some parameters like load-bearing surfaces, wear-
rate variations with time, wear surface temperature and wear product should be
monitored. The most common wear monitoring techniques are oil analysis and wear
debris particle analysis. In which contamination of oil, life deterioration of oil and
corrosion in parts are monitored [9, 17, 29].

5 Condition Monitoring Methodology

The condition monitoring methods can be classified into two main types: 1. Direct
monitoring method and 2. Indirect monitoring methods.

5.1 Direct Monitoring Method

In direct condition monitoring methods defined as direct measurements of the process

parameters of the system, which include, vision system, direct observation, compar-
ison with previously available database, graphical signature, and observation. In this
method, operator or human observation is required who can take decision because of
previous experiences. The actual geometric, sound, noise changes and other abnor-
mality can be directly observed. However, direct methods are less beneficial because
of their inaccuracy. It needed more experience and practices for observation [4].

5.2 Indirect Monitoring Method

In indirect condition monitoring method, the system will be equipped with differ-
ent sensors and actuators. The indirect method is based on the real-time parameter
measurement of the system. This method includes different measurable parameters
like temperature, pressure, vibration, sound and acoustic, surface roughness, forces,
torque and current, etc. To measurement of these parameters, different sensors have
been installed at different locations and system parameter will be sensed by the
sensors and with the help of data acquisition system, data will be sent to personal
218 A. K. Choudhary and D. Ahmad Khan

computers for store and analysis purposes. The sensors signal output of measurable
parameters indicates the real-time situations of the system to predict the condition
of the system [4].

6 Sensors for Condition Monitoring

In condition monitoring system deployment of sensors is the key part for the mea-
surement of different parameters of the system. On the other hand, selection of proper
sensors for the particular parameter measurement is also important. Sensors are prov-
ing real-time information of parameters to be measured. For correct measurement
sensor placement and particular condition, where you wish to get signal from the
sensors are also important [29, 30]. Here some sensors are discussed for monitoring
the several important parameters.

6.1 Force Sensor

The force can be measured by various sensors like load cells, Proof ring transducer,
Strain gauges, and Strain gauge (Wheatstone) bridges. From the load cells compres-
sive force can be measured by interpretation of stress or strain using Wheatstone
bridge circuit Force may be measured directly by a variety of methods as also, may
specific force. The proof ring transducer basically works on the principle of deforma-
tion (compressive and tensile conditions of load) on the ring which can be sensed by
the small inductive displacement transducer. For any strain and linear deformation
strain gauge can be used. In the strain gauge resistance of coil are proportionally
change with respect to strain or deformation induced in the structure. The strain
gauge (Wheatstone) bridges work on the principle of Wheatstone bridge circuit, in
which full bridge, half bridge, and Quarter bridge can be prepared. Due to small
changes in strain or deformation, the Wheatstone bridge circuit can detect in terms
of small changes circuit resistance [31, 32].

6.2 Temperature Sensor

In mechanical system and power plants, the heat transfer rates and system temperature
are the most important parameters for performance monitoring. If temperature gradi-
ents of system are not in define range then performance of system will be degraded.
For the temperature monitoring various sensors are available, these sensors will col-
lect the information about the temperature from the particular locations then convert
this information in suitable form and send it to observer or device. The most common
temperature sensors are thermometers, Resistance Temperature Detector (RTD) or a
Introduction to Conditioning Monitoring of Mechanical Systems 219

thermocouple, Thermistors, Infrared sensors, and Semiconductors. A thermometer

contains glass tube which is filled with mercury or alcohol and its volume is propor-
tional to the temperature. While temperature increases its volume are also increases
and shown in calibrated scale with précis unit (Fahrenheit, Kelvin, or Celsius). The
thermocouple is constructed with two dissimilar metals and these metals are work
on the principle of thermo-electric effect. When the junction of these metals is in
contact with cold or hot body, the thermal difference will be present and voltage will
be induced at other end. Some frequent types of thermocouples like K, J, T, R, E, S,
N, and B are available in premade forms [31–33]. The RTD sensor is constructed with
platinum, nickel, and copper metals and it has wide range of operation (−270 °C to
+850 °C). the semiconductor sensors are also known as IC temperature sensor. The
various types of semiconductor sensors are available like voltage output temperature
sensor, resistance output silicon temperature sensor, diode temperature sensor, cur-
rent output temperature sensor, and digital output temperature sensor. The Infrared
sensors are working on the principle of emitting or detecting Infrared radiation. It is
a non-contacting type temperature sensor and detects the temperature by measuring
the radiations [31, 32].

6.3 Vibration Transducers

The vibration transducers are used for measuring displacement, velocity, and accel-
eration. However, it can be measured relative vibrations rather than absolute vibra-
tions. The most commonly used vibrations transducers are velocity and acceleration
transducers which can measure displacement, velocity, and acceleration component,
which work on the principle of electromagnetic, electrodynamic, capacitive, piezo-
electric or strain gauge. The vibrations of the system can be measured in different
modes—peak-to-peak, peak, average, RMS and it can be measured in time domain or
frequency domain. The commonly used sensors are accelerometer, velocity sensors,
piezoelectric-based sensor. The accelerometer comprises with piezoelectric disc and
it can be installed directly on the surface of vibratory system. The accelerometer can
be classified as capacitive accelerometers, piezoelectric accelerometers, and strain
gauge accelerometers. In capacitive accelerometers, diaphragm supported seismic
mass is attached which acts as moving electrode, when vibration force is applied the
capacitance will be change and give output for measurement. In this sensor piezo-
electric disc will deflect, which produce voltage as output and it is proportional to
the rate of acceleration of the system [4, 16, 18]. The strain gauge accelerometer,
Wheatstone bridge circuit is used. The small displacement will be converted into
change in resistance which can be sensed as output of the Wheatstone bridge which
is proportional to the acceleration. The velocity sensors are consisting of permanent
magnet surrounding a moving coil which used to measure oscillatory or angular
velocity of the system. When this permanent magnet attached with vibratory system,
the emf will induce in the coil winding and this emf is proportional to velocity of
system [9, 33].
220 A. K. Choudhary and D. Ahmad Khan

6.4 Sound Monitoring

The human operator ears can perceive the defects by sudden sound changes from the
system when system components are loosened due to wear, loose fastenings and other
broken or loose parts. Such abnormal sound can be identifying using sound sensors
like Microphones. There are three most commonly microphones for capturing the
sound signals are piezoelectric-based microphone, moving coil-based microphone
and condenser-based microphone. The high acoustical impedance of condenser-
based microphone is suitable for measuring the sound pressure. The microphone
will be fitted with probes and used at the location where sound will be measured
[31, 32, 34].

7 Analysis of Sensor Signals

7.1 Sensor Signal Signatures Type

In condition monitoring the sensors signal are indicated the real status of the
machine/process and it also facilitates to identify the state and to take decision for
the condition of the machine. The sensor signal nature will help to take decision for
machine health [4]. For example abnormal sound and vibrations from the machine are
indicating that machine is not in normal condition. Hence, the sensors are generating
different types of signal signatures like periodic signals through periodic events like
oscillation of imbalance force or misalignments and periodic operational episodes.
Another sensor signal signature is transient signal, which in results of impact or other
unexpected events [33]. These signals can be recognized and characterize by the dif-
ferent sets of frequencies which can be predetermined. While acquiring the signal
from the sensor some signal may get distorted by external noise or transmission path
line which results the signals are complex than actual signals. The sensor signal can
be processed by time domain, frequency domain, and time–frequency-domain signal
processing methods [4, 16, 18].

7.2 Time-Domain Analysis

In machinery condition monitoring, a decision on the condition of a machine is

dependent on the nature of its signals. For example, if the machine produces excessive
noise and vibration, one gets a clue that something is not normal with the machine.
These signals are usually time-varying and need to be understood because they
carry information from the machine to the decision system. In time-domain analysis
sensors are providing a signal of response as a function of time. For example, x
signal is generated by the sensor which is in time domain and it has a variation with
Introduction to Conditioning Monitoring of Mechanical Systems 221

respect to time so x(t) is a function of time and this real signal is comprises with sum
of many sine waves with different pulses, phases, and amplitude and the smoothing
may be done to reveal for any underlying trends then the series are representing
by time-domain model [4, 16]. There are many conditions when the data cannot be
present by mathematical models. In this time the data smoothing will reduce the
variability of data and evoked to any underlying information. This variability in data
is present due to noise and variation in signal capturing process or monitoring the
process. The time-domain response can be divided into two parts transient response
and steady-state response. The statistical analysis also can be done in time-domain
signals. It includes the probability distribution of the data as mean, mode, median
variance, skew, kurtosis, and standard deviation [4, 16, 18].

7.3 Frequency-Domain Analysis

In the frequency domain analysis, a signal is characterized by their frequency con-

tent. Basically, the time-domain signal is transformed into a frequency domain. In
frequency-domain analysis three transformations can be used for signal like Fourier
series, Fourier transform, discrete Fourier transform. The Fast Fourier transform
(FFT) is the most popular method for frequency-domain analysis [4]. The advantage
of frequency-domain analysis over time-domain analysis is that Fourier representa-
tion of signals has an ability to identify the signal components and isolate certain
frequency components without any problem. For the nonstationary signals time–fre-
quency domain analysis can be done. Normally wavelet transforms techniques are
used for feature extraction in the time–frequency domain. The advantage of wavelet
transformation is that at the same time localization of a signal can be identified for
time domain and the frequency domain and it takes less time for processing. The
Hilbert and Hilbert-Huang Transform can be also used to identify the information
about instantaneous phase frequency and amplitude of the signal [16].

8 Applications of Condition Monitoring in Mechanical

Systems

The condition monitoring techniques can employ on various mechanical systems.

Some mechanical systems are discussed here.
222 A. K. Choudhary and D. Ahmad Khan

8.1 Cutting Tools Condition Monitoring

Nowadays unmanned manufacturing concept is most vital area for modern manufac-
turing industries and it is attracting the many researchers for real-time monitoring of
manufacturing process. The online cutting tool condition monitoring system is con-
taining measurement of tool wear, tool life, cutting force, and dimensional accuracy.
Due to regular operation, the chance of gradual failure of cutting tool increases which
directly affects the tool life and tool gets worn out, which is undesirable in machining
process. The cutting process using worn tool increases the friction between work-
piece and tool which causes more power consumption and rough surface finishing
or poor surface quality of the workpiece [9]. The dimensional inaccuracy and inad-
equate vibrational effect on the system are also introduced in machining process.
Subsequently, worn tools may cause unpredictable faults on the machine compo-
nents, damages to the machine tool, production delay and machine breakdown at
any time, which is economically not feasible. Hence to avoid the cutting tool failure,
and progressive wear on a cutting tool, an effective continuous cutting tool moni-
toring system is needed and worn tool can be changed before failure. The real-time
cutting tool condition monitoring system helps us to avoid unexpected failure in
machine components and downtime of the machine. For these various parameters
can be monitored like force, stress, vibration, chip formation, quality of surface fin-
ish, torque Motor current/power and sound of the machine. The cutting tool condition
monitoring system is consisting of the different types of sensors (strain gauge, load
cell, accelerometer, microphone, thermocouple, etc.), signal condition unit and anal-
ysis unit [4, 5, 8]. With the help of the monitoring system various parameters can be
measure like cutting force, vibration, temperature, acoustic emission, etc. To measure
the cutting tool parameters, sensor must place near or close to the target position,
generally, closer to the tooltip is the most suitable location for sensor placing. The
signal conditioning unit received information from the sensors and converted into
suitable formats like feature extraction, pattern recognition for analysis purpose and
results will be displayed in the monitor. The estimation of tool life can do with the
help of suitable decision-making techniques.

8.2 Machine Tools Condition Monitoring

The effective condition monitoring system and maintenance prediction are the per-
petual demands of modern industries. The machine infrastructure is including several
drive systems, control systems, sensors and actuators, and other electronic system,
which delivered a large amount of data and processing. The signals from the sensors
are used for the controlling purpose as well as monitoring purpose and information
will correlate with normal state of the machine and current state of the machine com-
ponents or time-dependent conditions. In modern production system, machine tool
Introduction to Conditioning Monitoring of Mechanical Systems 223

cost is major factor for production cost. The failure in machine tool increases produc-
tion cost and loss, both. For effective maintenance of machine tool there is need for
earlier warning system to prevent the breakdown conditions. The system will indicate
the actual condition of machine and provide warning for faults occurrence. To assess
a real-time condition of the machine tool and their component it is important to get
information about the components and which can classify into different categories
like normal, critical and worn out. The mechanical components in machine tools usu-
ally cause damage and deteriorate the machine performance, consequence in higher
production losses, therefore part quality during machining process should be sig-
nificantly monitored and this is directly related to machine performance. Therefore,
critical monitoring of machine tools components is required for different machines
like lathe, milling, boring, planning, etc. So, to determine the health of the machine
tool (healthy or faulty condition) an efficient monitoring system will be required
which recorded and analyze the machine performance. The system will consist of
sensors and data acquisition system which acquired the signal from the machine
for several parameters like speed, torque, sound, noise, pressure, temperature, loads,
dynamic forces, motion, etc. The analysis of different machine parameters at various
operating conditions is very complex and it is hampered for machine tool operation.
For example, the machine speed is varied according to machine operations and it is
not necessary, most recent measured data will be directly compared by previous ana-
lyzed data or records. Except, when the operating conditions, environment, and data
analysis and processing methods are independent from the other machine parame-
ters. The major disadvantage of machine tool condition monitoring is complexity and
additional cost. The machine tool condition monitoring system will increase the pro-
duction cost because use of different/multiple sensors and signal conditioning units
will be increased the cost and this will be added in production cost. However, this cost
will be varied for system to system. So now a day’s industries are looking for reliable
and low-cost monitoring system to reduce the production cost. Another process of
data analysis for parameters is mathematical modeling, through which parametric
effect and estimation can be done using suitable analysis techniques [2, 4, 5, 8, 9].

8.3 IC Engine Condition Monitoring

The Internal combustion engine (IC engine) is a heart of prime movers and equip-
ments like vehicle, power earth mover, agriculture equipments, ship, etc. it consists
with cylinder, linner, piston, fuel supply mechanism, cooling system, lubrication
system, exhaust system, and heat exchange system. The condition monitoring of
is valuable procedures/technique to keep engine health in good condition and it is
comprised of several measures like performance monitoring, combustion, emission,
and efficiency monitoring. The various fault or failure mode are related to engine
performance like incorrect valve timing, incorrect valve clearance, faulty fuel injec-
tor/calibration, faulty injectors, incorrect valve or fuel pump timing, cooling system
fault and coolant leakage, contaminated fuel, Pre-combustion chamber loose, air
224 A. K. Choudhary and D. Ahmad Khan

entering fuel system fault, fuel pump fault, exhaust system fault, mechanical fault
with engine, gasket fault, combustion loss, cracked cylinder head or cylinder block,
overheating of the engine, wear in parts, noise, and vibrations, thermal overload,
power loss, emission change, etc. [10]. In order to develop of an efficient condi-
tion monitoring system for IC engine, it is very much essential to know the process
parameters and failure mechanism of the engine and there are several parameters
to be monitored like pressure measurement which includes combustion pressure,
fuel injection pressure, air intake pressure, and exhaust pressure and oil pressure. To
monitoring of pressure at different locations pressure transducers can be used and
acquired signals will be analyzed for getting real-time information. The noise and
vibration on the engine are also major measurands for condition monitoring sys-
tems and fault diagnosis; it includes the measurement of displacement, velocity, and
accelerations. Various kinds of accelerometers can be used for the measurement of
engine vibrations [2, 10, 11]. The piezoelectric-based accelerometers are commonly
used for this purpose. The measurement of oil contamination and wear particles play
vital role in fault diagnosis systems for these different wear and oil contamination
techniques are used. The oil conditions and fuel are directly related to engine exhaust
system so velocity, density, acid number, and contamination are major parameters
to be monitored in the engine. The temperature gradient of the engine is also an
important. The thermocouple, optical thermometer, and other techniques are the
sensors for temperature measurement. The rotational angle and speed of the engine
can be measured by encoders and load can be measured by load cell. The crankshaft
angle measurement helps to measure engine torque. All the signals of the sensors
are acquired by data acquisition system and store into the personal computer then
analyze by using suitable signal processing techniques [10–12, 16, 22].

8.4 Power Plant Condition Monitoring

Power plant equipments are costlier in nature and they crucial in nature for mainte-
nance. A lot of money is required for the maintenance of power plants. The well-
maintained power plant system enhances the service life and reduced the downtime
of the system. To reduce the maintenance cost of the power plant, continuous con-
dition monitoring is required for both process performance and structure of power
plant and it proves to economical run of power generating units. In power plant
monitoring system conventional and innovative methods like sensors and instrumen-
tations, in situ, modeling tools for estimating the parameters (components damage,
pressure, temperature, heat exchange, combustion process, transmission process etc.
[2] and other convention testing methods can be used. There are various faults has
been notices in power plant like imbalance of rotor, bearing faults, overloading, loss
of alignment, looseness in mechanical assembly etc. The dynamic imbalance of the
system causes various malfunctions and deterioration in mechanical parts directly
reflected in the efficiency of power plants. For example, Due to looseness assem-
bly of bearing origins a dynamic imbalance in rotor and start vibrating and these
Introduction to Conditioning Monitoring of Mechanical Systems 225

vibration signatures will be stemming from the power plant structure. For condition
monitoring of power plants, vibrational signature will be utilized to predict the faults
and rectify the power plant’s problems. For vibration measurement seismometer and
accelerometer sensor are used and analog signal from the vibration sensor has been
used to determine vibrational amplitude and peak-to-peak values. In case of boiler
power plant there is a distributed control system for controlling the basic opera-
tional parameter like combustion, heat exchange, pressure, temperature, mass flow,
chemical properties of used water for steam generation. For the monitoring purpose
various modeling-based faults diagnostic systems are available to monitor the boiler
power plant and provides guidance to optimization of performance and corrective
measures. Due to various component present in power plant monitoring of individ-
ual components are crucial task. So first basic operations monitoring is important
so that potentially harmful deviations can be observed and prevent the failure. For
this operational experience is required to ensure that adopted monitoring methods
are cost and performance effective [16, 22, 28, 33].

8.5 Mechanical System Condition Monitoring

Determination of abnormal behavior and operational conditions and faults prediction

are important to estimate the overall efficiency of the mechanical system. To increase
the machine reliability and availability and knowledge of actual machine condi-
tions mechanical condition monitoring is needed. Initially for preliminary decision-
making of the condition of system, fault diagnostics in a selected part can be done.
Normally mechanical system must be operated in predefined parameters like speed,
load, temperature, pressure, etc. [2, 4, 7]. The system failure is explained as inca-
pability of a machine to do its required function or role and it can be identified
based on operating conditions, performance, and specifications. The general, causes
of mechanical system failure are improper assembly of components, improper main-
tenance, deficits in the original design, overloading during work, improper operation
while processing and prescribe materials. The potential advantage of mechanical sys-
tem monitoring is ensuring that system is available and reliable for utilization. The
condition monitoring system provides extended operational life and improves effi-
ciency of the system however, it increases monitoring equipment costs, operation and
training cost which is elicited as disadvantages. There are many factors in the system
which can be measured for condition monitoring like forcing factors of the mechan-
ical system which includes unbalance system parts, misalignment, looseness, rubs
(excessive looseness or whirl), resonances, structural vibrations, damaged or worn
rolling-element bearings, wear in parts, worn gears, etc. Figure 7 demonstrated vari-
ous parameters to be monitored for condition monitoring of mechanical systems and
Fig. 8 shows various parameters to be monitored for condition monitoring of electrical
systems. The system parameter can be sensed by different transducer which can inter-
pret physical quantity into electrical signal output and output signals are proportional
to the measured parameters. The selection, placing, mounting, and using of sensors
226 A. K. Choudhary and D. Ahmad Khan

Pressure Monitoring

Temperature Monitoring

Mechanical Equipment
Or VibraƟon Monitoring
Systems

Wear/Corrosion Monitoring

Chemical Analysis

Fig. 7 Various parameters which need monitoring for condition monitoring of mechanical systems

Moisture Monitoring

Temperature Monitoring

Oil TesƟng

MagneƟc Balance Test

Electrical Equipment Capacitance & Ten

Or
Systems
Winding Resistance Test

Sweep Frequency Response Analysis

Dissolved Gas Analysis

IR & PI Measurement

Turn Radio Measurement

Fig. 8 Various parameters which need monitoring for condition monitoring of electrical systems

or transducers is also important so that correct information of the parameters will

be gain and understood for their characteristics. Normally, for mechanical system,
vibrations signatures are used to diagnose the fault condition in the system. There are
three main types of vibration can be used for this purpose i. accelerometer, ii. velocity
sensor and iii. Displacement sensor [2, 4, 16]. Now a day’s microprocessor-based
vibration sensor is available which are capable to be measured, store and transferred
the data. These microprocessor-based sensors are accomplished with time-domain
Introduction to Conditioning Monitoring of Mechanical Systems 227

display and frequency-domain display of vibrational waveform data. Usually, vibra-

tional data is displayed in time domain, frequency domain, and modal domain. The
time-domain data refers to displayed data as time functions which include wave-
form analysis, time synchronous, averaging, probability density functions, etc. and
frequency-domain data refers to as functions of frequency in which time-domain
data will be processed by applying Fourier transform techniques of FFT algorithm.
The frequency domain data allows to represent time-domain data into frequency
spectrum which is more sensitive to change in the signals with the changing of
system conditions. The specific condition of the system also imposes to change in
vibrational signature, for example, Normal rolling-element bearings produce little
vibration signals which can be traced by vibrational signature when faults occurred
in the bearing the vibrational signature will be changed. In fact, gradually faults will
be developed in the bearing and vibrational signature will be changed accordingly.
For getting accurately detection and diagnosis of rolling-element-bearing defects,
placement of vibrational sensor is important which detects small changes in physical
condition of bearing. In case of gears, there are many significant forces are acting
while meshing of gears. Beyond the gears robust design, gear teeth are getting wear
while their continuous operation and loading, unloading conditions. The properly
designed and meshed gears produce less vibration whereas damaged or faulty gears
produce excessive vibration. To detect the actual status of gears meshing and gear
mesh frequency their vibrations patterns will be analyzed and critical information
will be obtained. Figure 9 shows condition monitoring of Rotating Machines.

RotaƟng Machine

Without CondiƟon Monitoring With CondiƟon Monitoring

Unexpected Shutdowns Data Monitoring

Motor failure Avoid Unnecessary Shutdowns

Heavy Economic Loss Healthy Motor

Fig. 9 Comparison of conditions of rotating machine with and without condition monitoring
228 A. K. Choudhary and D. Ahmad Khan

9 Conclusion

The various aspects of condition monitoring for mechanical systems, have been dis-
cussed like condition monitoring parameters, different types of sensors, sensors loca-
tions and installations, signal conditioning units, data collection methods, and data
analysis techniques. Simultaneously, procedure for detection, diagnosis and progno-
sis for condition monitoring have been revealed in this article. From the extensive
discussion it can be concluded that in condition monitoring first of clearly identify
the parameters to be monitored and select the sensors carefully for measuring the
desired physical quantity. Correspondingly signal conditioning unit and data acqui-
sition system will also be select with suitable data processing techniques for data
analysis. Mostly vibrations monitoring system are used to monitor the condition of
the mechanical systems. The vibrations signal-based condition monitoring facilitate
to determine actual faults on the mechanical systems and allows to estimate remain-
ing life, time required for maintenance, schedule for parts replacement if needed,
etc. Using the condition monitoring techniques various faults can also be identi-
fied which include, unbalance of the system, misalignment of parts, mechanical
looseness, improper assembly, resonances, structural vibrations, and other faults. A
common type of condition monitoring system can be developed for common group
of machines like pumps, compressors, reciprocating machines, turbines. The condi-
tion monitoring techniques facilitate to identify the faults occurrence in the systems
before failure of breakdown and it helps to reduce maintenance cost and idea to
prepare schedule for maintenance.

References

1. S. Martin-del-Campo, Sandin, Online feature learning for condition monitoring of rotating

machinery. Eng. Appl. Artif. Intell. 64, 187–196 (2017)
2. T. Wang et al., Multi-sensors-based condition monitoring of rotary machines: an approach of
multidimensional time-series analysis. Measurement 134, 326–335 (2019)
3. A. Bustos et al., Condition monitoring of critical mechanical elements through graphical rep-
resentation of state configurations and chromogram of bands of frequency. Measurement 135,
71–82 (2019)
4. R.B. Randall, Vibration-Based Condition Monitoring: Industrial, Aerospace and Automotive
Applications (Wiley, United Kingdom, 2011)
5. A. Davies, Handbook of Condition Monitoring Techniques and Methodology (Springer Science
+ Business Media, Dordrecht UK, 1998)
6. R. Isermann, Fault-Diagnosis Applications Model-Based Condition Monitoring: Actua-
tors, Drives, Machinery, Plants, Sensors, and Fault-Tolerant Systems (Springer, Heidelberg,
Dordrecht, London, New York, 2011)
7. T.R. Lin, A. Tan, I. Howard, J. Pan, P. Crosby, J. Mathew, Development of a diagnostic tool
for condition monitoring of rotating machinery, in ICOMS Asset Management Conference
Proceedings (Asset Management Council Limited, Gold Coast Convention and Exhibition
Centre, Queensland, 2011), pp. 1–9
8. L. Wang, R.X. Gao, Condition Monitoring and Control for Intelligent Manufacturing (Springer,
London, Dordrecht, Heidelberg, New York, 2006)
Introduction to Conditioning Monitoring of Mechanical Systems 229

9. K. Nikolai et al., On-line condition monitoring in industrial lubrication and tribology. Appl.
Cond. Monit. (2018). https://doi.org/10.1007/978-3-319-61134-1
10. S. Bari (ed.), Diesel Engine—Combustion, Emissions and Condition Monitoring (Techaneza-
Trdine 9, 51000 Rijeka, Croatia, 2013). http://dx.doi.org/10.5772/2782
11. T. Marwala, Condition Monitoring Using Computational Intelligence Methods Applications
in Mechanical and Electrical Systems (Springer, London, Dordrecht, Heidelberg, New York,
2012)
12. B.A. Shenoi, Introduction to digital signal processing and filter design (Wiley, Hoboken, New
Jersey, 2006)
13. T. Fakhfakh, W. Bartelmus, F. Chaari, R. Zimroz, M. Haddar, Condition monitoring of machin-
ery in non-stationary operations, in Proceedings of the Second International Conference “Con-
dition Monitoring of Machinery in Non-Stationnary Operations” CMMNO’2012 (Springer,
Heidelberg, New York, Dordrecht, London, 2012)
14. R.A. Collacott, Mechanical Fault Diagnosis and Condition Monitoring (Chapman and Hall,
Wiley, London, New York, 1977)
15. G.O. Chandroth, A.J.C. Sharkey, N.E. Sharkey, Cylinder pressures and vibration in inter-
nal combustion engine condition monitoring (1999). https://www.researchgate.net/publication/
2613893
16. T. Barszcz, Vibration-Based Condition Monitoring of Wind Turbines (Springer Nature,
Switzerland, AG, 2019)
17. J. Zhu et al., Lubrication oil condition monitoring and remaining useful life prediction with
particle filtering, in International Conference of Prognostic and Health Management (2013),
pp. 1–15
18. K. Jafarian, D. Morteza, Z. Honarkar, Vibration analysis for fault detection of automobile engine
using PCA technique, in 2016 4th International Conference on Control, Instrumentation, and
Automation (ICCIA) 27-28 January 2016 (Qazvin Islamic Azad University, Qazvin, Iran, 2016)
19. H.A. Toliyat, S. Nandi, S. Choi, H. Meshgin-kelk, Electric Machines Modeling, Condition
Monitoring, and Fault Diagnosis (CRC Press, Taylor & Francis Group, 2013)
20. A.C. Starr, R. Rao (ed.), Condition monitoring and diagnostic engineering management (coma-
dem 2001), in Proceedings of the 14’” International Congress 4–6 September 2001 (Elsevier
Science Ltd, Manchester, UK, 2001)
21. R.S. Beebe, Predictive Maintenance of Pumps Using Condition Monitoring (Elsevier Science
& Technology Books UK, 2004)
22. A. Agoston, C.O. Tsch, B. Jakoby, Viscosity sensors for engine oil condition monitoring—
application and interpretation of results. Sens. Actuators A 121, 327–332 (2005)
23. R. Bogue, Sensors for condition monitoring: a review of technologies and applications. Sens.
Rev. 33/4, 295–299 (2013)
24. M. Tuckwell, Condition Monitoring: Why Now? As Cost of Operation Reduces, Condition
Monitoring is Moving into New Sectors and New Applications (Sagentia Ltd., Harston Mill,
Harston, Cambridge, CB22 7GG, UK, 2015)
25. C.K. Mechefske, Machine Condition Monitoring and Fault Diagnostics, Vibration and Shock
Handbook (2012)
26. A. Verma, S. Srivastava, Review on condition monitoring techniques oil analysis. Thermogr.
Vib. Anal. 3(7), 18–25 (2014) (erpublications.com)
27. D.N. Brown, Machine-Condition Monitoring Using Vibration Analysis, A Case Study from a
Nuclear Power-Plant. http://www.bksv.com/doc/BO0209.pdf. Accessed 20 May 2019
28. M. Kande et al., Rotating electrical machine condition monitoring automation—a review,
machines, 5, 24 (2017). https://doi.org/10.3390/machines5040024
29. Y.-C. Chou, L. Yao, Automatic diagnostic system of electrical equipment using infrared ther-
mography, in International Conference of Soft Computing & Pattern Recognition (2009),
pp. 155–160
30. M. Picazo-Rodenas, Use of infrared thermography for computation of heating curves and
preliminary failure detection in induction motors, in XXth International Conference on electric
machine (ICEM) (2012), pp. 525–531
230 A. K. Choudhary and D. Ahmad Khan

31. P. Bowman, Identification for the global environment, in Sensor Based Condition Monitoring,
Building Radio Frequency (2009), Accessed 13 May 2019
32. Bogue R. Robert, Sensors for condition monitoring: a review of technologies and applications.
Sens. Rev. 33(4), 295–299 (2013)
33. J.A. Francisco et. al., A smart sensor for the condition monitoring of industrial rotating
machinery (2015). https://doi.org/10.1109/icsens.2012.6411586
34. G. Dalpiaz et al., Advances in condition monitoring of machinery in non-stationary operations,
in Proceedings of the Third International Conference on Condition Monitoring of Machinery
in Non-Stationary Operations CMMNO 2013 (Springer, Berlin, Heidelberg, 2014)
Novel Application of Artificial Neural
Network Techniques for Prediction of Air
Pollutants Using Stochastic Variables
for Health Monitoring: A Review

Vibha Yadav and Satyendra Nath

Abstract Due to population growth and urbanization, air pollutant (AP) in environ-
ment is increasing day by day which creates a lot of health problem. AP data provide
information about quality of air and health risk in surrounding which is important for
environmental management. In this chapter, a review is made on air pollutant pre-
diction using ANN techniques which are dependent on types of prediction intervals,
i.e. monthly, daily and hourly. It is found that influence of different input variables,
training algorithm and architecture changes prediction accuracy of ANN models.

Keywords Air quality · Air pollution · Artificial neural network · Air quality
prediction · Meteorological variables

Nomenclature

ANN Artificial Neural Network

PCA Principle Component Analysis
AP Air Pollutant
PM Particulate Matters
CO2 Carbon di oxide
RBF Radial Basis Function
CO Carbon Monoxide
R Regression
D Day
RH Relative humidity
H Hour
RSPM Respirable Suspended PM
LM Levenberg Marquardt

V. Yadav (B) · S. Nath

Department of Environmental Sciences and NRM, College of Forestry, Sam Higginbottom
University of Agriculture, Technology and Sciences, Allahabad 211007, Uttar Pradesh, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 231

H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_10
232 V. Yadav and S. Nath

RMSE Root Mean Square Error

MSE Mean Square Error
SR Solar Radiation
M Month
SO2 Sulphur di oxide
MLR Multiple Linear Regression
T Temperature
NO Nitrogen Oxide
TSP Total Suspended Particles
NO2 Nitrogen di oxide
WD Wind Direction
P Pressure
WS Wind Speed

1 Introduction

Air pollutants (AP) means the presence of gaseous substance (including noise), liquid
and solid to such a concentration in atmosphere causes injuries to human beings,
creatures, plants, etc. It originates from indoor or outdoor sources. Many domestic
activities like cooking, burning of biofuels and heating produce indoor air pollutants.
Detergents and household cleaners, aerosols sprays, wax, paints are few chemicals
release into home. Outdoor AP are of two types mainly (i) primary, (ii) secondary.
Primary AP are discharged directly in atmosphere resulting from industrial activities
and fuel combustion whereas secondary pollutants are emitted by interaction of
primary pollutants or by reaction in atmosphere [1]. The life cycle of air pollutants
is shown in Fig. 1. Different air pollutants are Suspended particulate matter (PM),
Oxides of Nitrogen (NO), Sulphur dioxide (SO2 ), Lead (Pb), Chlorine, Fluorine,
Peroxyacetyl nitrate, CO2 , O3 and Carbon monoxide (CO), etc. The top ten AP
gases and their effects on health of human and environment are shown in Table 1

Fig. 1 Life cycle of air

pollutants
Novel Application of Artificial Neural Network Techniques … 233

Table 1 Effects and sources of different air pollutants

S. no. Air pollutant Sources Effects
1. Sulphur dioxide (SO2 ) Impure forms of Fossil Formation of Smog
fuels Causes Acid rain
Coal-fired power plant Plant damage
Water damage
Corrosion
Health problems
2. Carbon monoxide (CO) Highly dangerous gas Fatigue in healthy people
Car exhausts Heart disease
Gas boiler Impaired vision
Fuel-burning appliances
3. Nitrogen oxide (NO) Indirect result of Formation of Ozone,
combustion Smog and Acid rain.
Vehicle engines and power Reducing the carrying
plants capacity of blood.
Decreased the functioning
of thyroid gland,
Causes asthma
4. Volatile Organic Content of chemicals like Chronic effects on people
Compounds paints waxes and varnishes
(VOCs)
5. Particulates Sooty deposits in air Make building blackish
Soot particle less than 10 Causes breathing
microns difficulties
means (10 millionths of a Decrease in lung
meter) fluctuation
6. Ozone (O3 ) Trioxygen key ingredients Induction of respiratory
of Smog disorder and throat
infection
7. Chlorofluorocarbons Refrigerators and aerosols Greenhouse effect
(CFCs) cans Global Warming
8. Unburned hydrocarbons Incomplete burn of Release carbon monoxide
petroleum in air and contributing to
and other fuels smog
9. Lead and heavy metals Toxic compounds Permanent brain damage,
dispersed in air either in Stomach discomfort,
solid or liquid form Constipation,
Such things as exhaust Loss of Appetite
fumes and flyash Insomnia

[2]. The worldwide distribution of PM2.5 derived by National Aeronautics and Space
Administration (NASA) satellite is shown in Fig. 2 which presents average value of
PM2.5 between 2001 and 2006. Air Quality (AQ) Index is used to communicate status
of air quality to people which are easy to understand. Based on ambient concentration
and health impact (health break point) AQI categorization is shown in Table 2 [3].
234 V. Yadav and S. Nath

Fig. 2 Satellite derived PM2.5 [4]

The rapid growth of population and change in modern lifestyle substantially

increases the generation of air pollutants resulting in contamination of air, water and
land resources. Therefore, it is a big challenge to provide prior information about air
quality status which can be achieved by prediction models. This is helpful for taking
early preventive measures to control air pollution and formation of policies.
In this study, various methods of different types of air pollutants are reviewed.
Based on this, research gaps are identified which is important for researchers as well
as industry.

2 Artificial Neural Network for Air Quality Prediction

Air pollutants and meteorological variables vary from site to site therefore its predic-
tion for different time intervals with ANN are challenging tasks. In the past decade,
air quality prediction by ANN has been performed by [5] and multivariate adaptive
regression splines [6, 7]. ANN consists of three layers (Fig. 3).
The standard ANN model for prediction is shown in Fig. 4 [8]. ANN is trained
by data and due to its generalization properties, it is used for modelling and forecast-
ing. It constitutes processing units associated with learning rule. The mathematical
behaviour is shown in Eq. 1.
⎛ ⎞
n
S = f⎝ ej wj⎠ (1)
j=1

where ej is input, wj is weight, f activation function, n number of inputs and S is

output.
For prediction, ANN in MATLAB is developed by nftool and its implementation
is shown in Fig. 5 [9]. Hidden layer neurons selection is given in [10, 11].
Table 2 Air quality index of eight air pollutant
AQI category AQI Concentration range
PM10 PM2.5 NO2 O3 CO SO2 NH3 Pb
Good 0–50 0–50 0–30 0–40 0–50 0–1 0–40 0–200 0–0.5
Satisfactory 51–100 51–100 31–60 41–80 51–100 1.1–2 41–80 201–400 0.5–1.0
Moderately polluted 101–200 101–250 61–90 81–180 101–168 2.1–10 81–380 401–800 1.1–2.0
Poor 201–300 251–350 91–120 181–280 169–208 10–17 381–800 801–1200 2.1–3.0
Very poor 301–400 351–430 121–250 281–400 209–748 17–34 801–1600 1200–1800 3.1–3.5
Novel Application of Artificial Neural Network Techniques …

Severe 401–500 430 + 250+ 400+ 748+ 34+ 1600+ 1800+ 3.5+
235
236 V. Yadav and S. Nath

Fig. 3 Architecture of ANN

AQ model provides relationship between input and meteorology, emission and

AQ as shown in Fig. 6. The categorization of model given by Moscardini is shown
in Fig. 7. The different types of air quality models are given by [12] in Fig. 8.

2.1 Monthly Air Pollutant Prediction

The air pollutant becomes an important environmental concern in urban area so pre-
diction of air pollutant concentration becomes a major issue for places where measur-
ing instrument is not installed. Corani [13] proposed Feedforward neural networks
(FFNNs), lazy learning (LL) and pruned neural networks (PNNs) for prediction of
ozone and PM10 for Milan. The input variables for models are SR, WS, rain, T and
RH. The PNN method is found to be superior in comparison to FNN and LL.
Tecer [14] used ANN to predict monthly SO2 and PM for Turkey. The ANN
model utilized cloudiness, pressure, RH, maximum and minimum value of WS and
T, average T, WD as inputs and results shows that ANN model effectively predicts
AP. The R value varies from 81.7 to 89.9% and concentrations of SO2 and PM are
maximum in winter season.
Xie et al. [15] developed four ANN models to predict indoor air quality. The
input to ANN models is six indoor air pollutants (CO2 , PM2.5 , formaldehyde, volatile
organic compounds, bacteria and fungi) and three indoor air comfort (air velocity,
RH, T ). The RMSE for testing ANN models varies from 7.7 to 8.9.
Mechgoug et al. [16] used optimized neural network (NN) to forecast monthly
value of O3 and CO. Genetic algorithm (GA) is used to optimize NN structure
Novel Application of Artificial Neural Network Techniques … 237

Fig. 4 Standard air pollutant prediction pipeline

and prediction error was minimum during testing phase. The MSE is found to be
2.0916 × 10−17 .
Azid et al. [17] utilized principal component analysis (PCA) and ANN to predict
the air pollutant index (API) for Malaysia. The input variables such as CO, NO2 ,
CH4 , WD, PM10 , humidity, O3 and T are used for prediction. The R value for the
best model is 74.9%.
Hossain et al. [18] utilized 12 meteorological and 6 traffic variables to predict
the monthly values of CO, PM10 and PM2.5 for Dhaka. The predicted and measured
values are closed to each other. The RMSE values vary from 1.89 to 3.39.
Kottur et al. [19] proposed Kriging and ANN for predicting AP at Mumbai sites.
The inputs are latitude, longitude, month, RH, WD, WS, rainfall, cloud octal, max-
imum and minimum value of T, pressure and output variables are SOx , NOx and
RSPM. R value is found to be more than 70%.
238 V. Yadav and S. Nath

Fig. 5 nftool implementation

Fig. 6 Air quality model

Fig. 7 Model categories

Novel Application of Artificial Neural Network Techniques … 239

Fig. 8 Air quality models

2.2 Daily Air Pollution Prediction

The daily air pollution (DAP) is an important parameter in environmental applica-

tions especially for information about health problem due to air pollutant. The DAP
prediction has been given by number of researchers and are shown in Table 3.

2.3 Hourly Air Pollutant Prediction

In this section, we discussed hourly AP prediction by various researchers. Sohn et al.

[25] used ANN model for 1–6 h ahead air pollutant prediction in Seoul, Korea. Input
data consist of AP concentration that consists of ozone, SO2 , NO2 , total hydrocarbon
(THC), CH4 , CO and NO. ANN model accuracy is found in acceptable range.
Elangasinghe et al. [26] utilized input variables as RH, D, M, H, SR, T, WS and
WD to predict hourly, daily and monthly value of NO2 . It is found that suitable
selection of inputs improves the accuracy of ANN model. The ANN model performs
better than linear regression model based on the same input parameters.
Lu et al. [27] utilized PCA and RBF for hourly AP prediction. PCA is used to
orthogonalize and reduce variables which then used in RBF to forecast respirable
suspended particulates in Hong Kong. The hybrid PCA-RBF gives better results than
only RBF only.
Challoner et al. [28] proposed ANN and personal-exposure activity location model
(PALM) for prediction of hourly NO2 and PM2.5 . ANN model is used to model
relationship between outdoor and indoor building. The ANN–PALM approach is
used to predict indoor AP for cities in Dublin. The R value is found to be more than
90%.
Russo et al. [29] used eigen variables for input variables reduction which increases
ANN model accuracy. The input data consists of hourly value of NO2 , NO and CO.
The prediction accuracy is improved by input variable selection.
240 V. Yadav and S. Nath

Table 3 Daily air quality prediction by ANN

Study Location Types of ANN type Input Highlights
air parameters
pollutant
Moustris Athens, NO2 , MLP P, SR Statistical
et al. [20] Greece SO2 , CO significance
and O3 level of p <
0.01
He et al. Lanzhou, SO2 , NO2 MLP Metrological R value
[21] northwestern and PM10 variables varies from
China 71 to 83%
Ding et al. Hong Kong NO2 , Feedforward Different FFAN give
[22] NOX , ANN (FFANN) combinations better results
SO2 , O3 backpropagation of AP and than
and PM2.5 meteorological regression
variables model
Prasad Howrah Daily ANFIS models Different RMSE varies
et al. [23] value of combinations from 0.12 to
NO2 , of P, T, RH, 0.48
SO2 , O3 dew point,
CO and visibility, WS
PM10 and
precipitation
Yadav and Varanasi, Daily PCA-ANN SR, vertical MAPE is
Nath [24] India value of model wind speed, 9.88%
PM10 AP, PM2.5 ,
NO, benzene

Syafei et al. [30] used WS, solar radiation (SR), H and T to predict 30-min ahead
of NO2 , PM10 and O3 . Faster WS reduces the concentration of NO2 and increases
ozone. The holiday and weekend day is also important for pollutant concentration.
Lu et al. [31] predicted hourly value of RSP, NO2 , NOx using RBF and PCA-RBF
for Hong Kong. The PCA-RBF and RBF give better results than simple network.
PCA-RBF learn faster than RBF. The air pollutant prediction by various authors at
different locations are shown in Table 4.

3 Identified Research Gap from Literature Survey

1. To develop ANN models for air quality prediction researchers have used air
pollutant data of one to four years or five years and maybe hourly data. Therefore
comparative study can be done about number of years air pollutant data required
for ANN model development with less prediction error.
Novel Application of Artificial Neural Network Techniques … 241

Table 4 Air pollutant prediction for short time scale

Study Location Types of ANN type and Input Highlights
air algorithm parameters
pollutant
Gardner and London Hourly Multilayer Different RMSE
Dorling [32] value of perceptron (MLP) combinations varies from
hourly neural networks of Low cloud 0.6 to 6.0
NOx and amount,
NO2 Visibility,
Vapour
Pressure,
cloud, WS and
temperature
Song et al. Yangtze Hourly Adaptive AP MAPE
[33] River and 12 h neuro-fuzzy (ANF) varies from
Delta average model 7.62 to
(YRD) value of 10.40%
region of PM10 ,
China PM2.5
Kukkonena Central Hourly Neural network Temporal R values
et al. [34] Helsinki NO2 and (NN), Deterministic variables, varies from
PM10 modelling system concentration 32 to 73%
(DET) and Linear variables and for different
statistical model meteorological models and
variables years
Hooyberghsa Belgium Forecast ANN model WS, Cloud R is close to
et al. [35] of one day cover, WD, T 1
ahead and D
PM10
Niska et al. Helsinki Forecast MLP and Numerical Temporal, R value
[36] of weather prediction concentration varies from
nitrogen model and 23 to 73%
dioxide meteorological
(NO2 ) and data
PM2.5
Perez and Santiago, Forecast Three-layer Meteorological ANN with
Reyes [37] Chile of PM10 feedforward NN variables relevant
inputs are
accurate
Mishra et al. Delhi Forecast Neuro-Fuzzy (NF), SO2 , NO2 , O3 , ANN, NF
[38] of PM2.5 ANN and MLR CO, previous and MLR
hour’s have R
concentration value of 53,
of PM2.5 , T, 72 and 25%
RH, WS
Elangasinghe New PM2.5 ANN WS, WD, R value for
et al. [39] Zealand and PM10 SR, T, RH PM2.5 and
PM10 are
79 and 69%
(continued)
242 V. Yadav and S. Nath

Table 4 (continued)
Study Location Types of ANN type and Input Highlights
air algorithm parameters
pollutant
Singh et al. Lucknow SO2, PCA, Ensemble T, RH, WS, R value for
[40] NO2, learning approaches SS, network is
RSPM Evaporation more than
and SPM 82.5%
Nagendra Delhi CO Three different ANN SH, visibility, RMSE
and Khare models with different T, WS, WD varies from
[41] input combinations. and humidity 0.36 to 1.71
Turias et al. Campo CO, SO2 MLP Pollutant ANN
[42] de and Concentrations models give
Gibraltar suspended superior
particulate results than
matter conventional
(SPM) models
Ragosta et al. Basilicata Hourly Back Propagation Atmospheric Statistical
[43] region of value of Network pressure, RH analysis by
southern CO, SO2 , and WS PCA is
Italy NO2 , O3 useful for
and PM10 relevant
input
selection
Dunea et al. Urban Hourly Wavelet–feedforward Decompose Forecasting
[44] area of value of neural networks and component of accuracy
Romania O3 , NO2 , Feedforward neural AP depends on
PM10 and networks precisions of
PM2.5 measured
data. R
value varies
from 84 to
93%
Azid et al. Malaysia CO, O3 , MLP feedforward AP PCA
[45] SO2 , artificial neural concentration identifies
NO2 , network CH4 ,
PM10 , NmHC,
CH4 , THC, O3
NmHC, and PM10
THC as
significant
inputs
Arhami et al. Tehran NOx, MLP WS, RH, WS, RH,
[46] NO2 , NO, T, and WD T, and WD
O3 , CO are
and PM10 influencing
parameters
Novel Application of Artificial Neural Network Techniques … 243

2. Due to change in neurons hidden layer one by one is time-consuming for MAPE
evaluation. Therefore a suitable method based on optimization technique should
be developed at which there is minimum error.
3. Application of ANN models for different locations worldwide should be done for
investigating the best ANN model which will be able for an efficient air quality
prediction in any location all over the world.
4. For improving the accuracy of air quality prediction, the effect of different input
variable such as AT, RH, WS, WD, AP, SH and, etc. should be considered.
5. A comparative study can be done by using different ANN trained models with
the help of optimization techniques like evolutionary algorithm.

4 Conclusion

In this study, the literature review of monthly, daily and hourly AP prediction are
provided and research gaps are identified which are useful for researchers working
in AP prediction field. ANN prediction accuracy is found to be accurate by most of
the authors and it is dependent on most influencing inputs, learning algorithm and
different models. To predict different AP, meteorological variables can be used as
input and using input variable selection techniques accuracy can be increased.

Acknowledgements We would like to thank the Department of Science and Technology (DST),
New Delhi-110016 India for providing inspire fellowship with Ref. No. DST/INSPIRE Fellow-
ship/2016/IF160676.

References

1. Niharika, M. Venkatdri, A novel sir quality prediction model using artificial neural networks.
Int. J. Eng. Res. Technol. 3(2), 2278–0181
2. Niharika, M, Venkatadri, P.S. Rao, A survey on air quality forecasting techniques. Int. J.
Comput. Sci. Inform. Technol. 5(1), 103–7
3. F.I. Khan, R. Sadiq, Risk-based prioritization of air pollution monitoring using fuzzy synthetic
evaluation technique. Environ. Monit. Assess. 105, 261–283 (2005)
4. https://www.nasa.gov/topics/earth/features/health-sapping.html. [Accessedon 7 July 2017]
5. F. Benvenuto, A. Marani, Neural networks for environmental problems: data quality control
and air pollution nowcasting. Global Nes. Int. J. 2(3), 281–292
6. P.J.G. Nieto, J.C.A. Antón, J.A.V. Vilán, E. García-Gonzalo, Air quality modeling in the Oviedo
urban area (NW Spain) by using multivariate adaptive regression splines. Environ. Sci. Pollut.
Res. 22, 6642–6659 (2015)
7. M. Sarkhosh, J.B. Ghasemi, M. Ayati, A quantitative structure—property relationship of gas
chromatographic/mass spectrometric retention data of 85 volatile organic compounds as air
pollutant materials by multivariate methods. Chem. Cent. J. 6, 54–61 (2012)
244 V. Yadav and S. Nath

8. A. Qazi, H. Fayaz, A. Wadi, R.G. Raj, N.A. Rahim, W.A. Khan, The artificial neural network
for solar radiation prediction and designing solar systems: a systematic literature review. J.
Clean. Prod. 104, 1–12 (2015)
9. A.K. Yadav, H. Malik, S.S. Chandel, Selection of most relevant input parameters using WEKA
for artificial neural network based solar radiation prediction models. Renew. Sustain. Energy
Rev. 31, 509–519 (2014)
10. S.K.H. Chow, E.W.M. Lee, D.H.W. Li, Short-term prediction of photovoltaic energy generation
by intelligent approach. Energy Build. 55, 660–667 (2012)
11. M. Frederick, Neuroshell 2 Manual, Ward Systems Group Inc. (1996)
12. E. Weber, Air Pollution Assessment Modelling Methodology, NATO, Challenges of Modern
Society, vol. 2 (Plenum Press, 1982)
13. G. Corani, Air quality prediction in Milan: feed-forward neural networks, pruned neural
networks and lazy learning. Ecol. Model. 185(513–29) (2005)
14. L.H. Tecer, Prediction of SO2 and PM concentrations in a coastal mining area (Zonguldak,
Turkey) using an artificial neural network. J. Environ. Stud. 16(2), 633–638 (2007)
15. H. Xie, F. Ma, Q. Bai, Prediction of indoor air quality using artificial neural networks, in Fifth
International Conference on Natural Computation (2009), pp. 415–18
16. R. Mechgoug, T.A. Ahmed, L. Cherroun, Optimization of neural predictor for air pollution, in
World Congress on Engineering (2012), pp. 12078–958
17. A. Azid, H. Juahir, T.M. Latif, M.S. Zain, R.M. Osman, Feed-forward artificial neural network
model for air pollutant index prediction in the Southern Region of Peninsular Malaysia. Energy
J. Environ. Prot. 7(4), 1–10 (2013)
18. A.M.K. Hossain, 2014 Predictive ability improved neural network models to stimulate pollutant
dispersion. Int. J. Atmos. Sci. Res. Article No. 141923, 1–12
19. V.S, Kottur, S.S. Mantha, An integrated model using artificial neural network (ANN) and
Kriging for forecasting air pollutants using meteorological data. Int. J. Adv. Res. Comput.
Commun. Eng. 4(1), (2015) ISSN 2319-5940
20. K.P. Moustris, I.C. Ziomas, A.G. Paliatsos, 3-Day-Ahead forecasting of regional pollution
index for the pollutants NO2 , CO, SO2 , and O3 using artificial neural networks in Athens,
Greece. Water Air Soil Pollut. 209, 29–43 (2010)
21. J. He, Y. Yu, Y. Xie, H. Mao, L. Wu, N. Liu, S. Zhao, Numerical model-based artificial neural
network model and its application for quantifying impact factors of urban air quality. Water
Air Soil Pollut. 227(235) (2016)
22. W. Ding, J. Zhang, Y. Leung, Prediction of air pollutant concentration based on sparse response
back-propagation training feedforward neural networks. Environ. Sci. Pollut. Res. 23, 19481–
19494 (2016)
23. K. Prasad, A.K. Gorai, P. Goyal, Development of ANFIS models for air quality forecasting
and input optimization for reducing the computational cost and time. Atmos. Environ. 128,
246–262 (2016)
24. V. Yadav, S. Nath, Novel hybrid model for daily prediction of PM10 using principal component
analysis and artificial neural network. Int. J. Environ. Sci. Technol. https://doi.org/10.1007/
s13762-018-1999-x
25. H.S. Sohn, C.S. Oh, K.Y. Yeo, Prediction of air pollutants by using an artificial neural network.
J. Chem. Eng. 16(3), 82–87 (1999)
26. M.A. Elangasinghe, N. Singhal, K.N. Dirks, J.A. Salmond, S. Samarasinghe, Complex time
series analysis of PM10 and PM2.5 for a coastal site using artificial neural network modeling
and k-means clustering. Atmos. Environ. 94(106–16) (2014)
27. Z.W. Lu, W.J. Wang, Y.H. Fan, T.Y.A. Leung, B.Z. Xu, K.C.J. Wong, Prediction of pollutant
levels in Causeway Bay area of Hong Kong using an improved neural network model. J.
Environ. Eng. 128(22), 1147–1157 (2001)
28. A. Challoner, F. Pilla, L. Gill, Prediction of indoor air exposure from outdoor air quality using
an artificial neural network model for inner city commercial buildings. Int. J. Environ. Res.
Public Health 12, 15233–15253 (2015)
Novel Application of Artificial Neural Network Techniques … 245

29. A. Russo, F. Raichel, G.P. Lind, Air quality prediction using optimal neural networks with
stochastic variables. J. Atmos. Environ. 79, 822–830 (2013)
30. D.A. Syafei, A. Fujwara, J. Zhang, Prediction model of air pollutant levels using linear model
with component analysis. Int. J. Environ. Sci. Dev. 6(7), 1–7 (2015)
31. W.Z. Lu, W.J. Wang, X.K. Wang, Z.B. Xu, A.Y.T. Leung, Using improved neural network
model to analyze RSP, NOx and NO2 levels in urban air in Mong Kok, Hong Kong. Environ.
Monitor. Assess. 87, 235–254 (2003)
32. M.W. Gardner, S.R. Dorling, Neural network modelling and prediction of hourly NOx and
NO2 concentrations in urban air in London. Atmos. Environ. 33, 709–719 (1993)
33. Y. Song, S. Qin, J. Qu, F. Liu, The forecasting research of early warning systems for atmospheric
pollutants: a case in Yangtze river delta region. Atmos. Environ. 118, 58–69 (2015)
34. J. Kukkonena, L. Partanena, A. Karppinena, J. Ruuskanenb, H. Junninenb, M. Kolehmainenb,
H. Niskab, S. Dorling, T. Chattertonc, R. Foxalld, G. Cawleyd, Extensive evaluation of neural
network models for the prediction of NO2 and PM10 concentrations, compared with a determin-
istic modelling system and measurements in central Helsinki. Atmos. Environ. 37, 4539–4550
(2003)
35. J. Hooyberghsa, C. Mensinka, G. Dumontb, F. Fierensb, O. Brasseurc, A neural network
forecast for daily average PM10 concentrations in Belgium. Atmos. Environ. 39, 3279–3289
(2005)
36. H. Niska, M. Rantama, T. Hiltunen, A. Karppinen, J. Kukkonen, J. Ruuskanen, M.
Kolehmainen, Evaluation of an integrated modelling system containing a multi-layer percep-
tron model and the numerical weather prediction model HIRLAM for the forecasting of urban
airborne pollutant concentrations. Atmos. Environ. 39, 6524–6536 (2005)
37. P. Perez, J. Reyes, An integrated neural network model for PM10 forecasting. Atmos. Environ.
(40), 2845–51 (2006)
38. D. Mishra, P. Goyal, A. Upadhyay, Artificial intelligence based approach to forecast PM2.5
during haze episodes: a case study of Delhi, India. Atmos. Environ. 102, 239–248 (2015)
39. M.A. Elangasinghe, N. Singhal, K.N. Dirks, A. Jennifer, J.A. Salmond, Development of
an ANN–based air pollution forecasting system with explicit knowledge through sensitivity
analysis. Atmos. Pollut. Res. 5, 696–708 (2014)
40. K.P. Singh, S. Gupta, P. Rai, Identifying pollution sources and predicting urban air quality
using ensemble learning methods. Atmos. Environ. 80, 426–437 (2013)
41. S.M.S. Nagendra, E.M. Khare, Modelling urban air quality using artificial neural network.
Clean Technol. Environ. Policy 7, 116–126 (2005)
42. I.J. Turias, F.J. González, M.L. Martin, P.L. Galindo, Prediction models of CO, SPM and
SO2 concentrations in the Campo de Gibraltar Region, Spain: a multiple comparison strategy.
Environ. Monit. Assess. 143, 131–146 (2008)
43. M. Ragosta, M.D. Emilio, G.A. Giorgio, Input strategy analysis for an air quality data modelling
procedure at a local scale based on neural network. Environ. Monit. Assess. 187(307) (2015)
44. D. Dunea, A. Pohoata, S. Iordache, Using wavelet–feedforward neural networks to improve
air pollution forecasting in urban environments. Environ. Monit. Assess. 187(477) (2015)
45. A. Azid, H. Juahir, M.E. Toriman, M.K.A. Kamarudin, A.S.M. Saudi, C.N.C. Hasnam, N.
Azlina, A. Aziz, F. Azaman, M.T. Latif, S.F.M. Zainuddin, M.R. Osman, M. Yamin, Prediction
of the level of air pollution using principal component analysis and artificial neural network
techniques: a case study in Malaysia. Water Air Soil Pollut. 225(2063) (2014)
46. M. Arhami, N. Kamali, M.M. Rajabi, Predicting hourly air pollutant levels using artificial
neural networks coupled with uncertainty analysis by Monte Carlo simulations. Environ. Sci.
Pollut. Res. 20, 4777–4789 (2013)
Application of Soft Computing in Condition
Monitoring and Diagnostics of Electrical
and Mechanical Systems
Intelligent System for Monitoring
and Recognising the Type of Domestic
Harmonic Loads

Pullabhatla Srikanth and Chiranjib Koley

Abstract Nonlinear loads that are found in daily life have been accurately analysed
and recognised using a new inverse property methodology of Stockwell transform.
The method uses the magnitude of dilated, translated and inverted S-Transform (ST)
window varying with time and hence identification of nonlinear loads is very accurate.
Waveforms of currents and voltages obtained through practical setup are used as the
inputs for analysis and recognition of the type of nonlinear load. The results have been
compared with the existing techniques and other variations of ST methodology. This
novel method outperforms all the other signature identification techniques, thereby
showing the pre-eminence. Later, a fuzzy decision box has been used with the indices
developed using the proposed technique as the input. With this fuzzy logic design,
different nonlinear loads that are in domestic household use have been analysed and
recognised. The final results using a combination of Fuzzy decisions and proposed
techniques have been found 100% accurate and efficient.

Keywords Customers · Fuzzy · Harmonics · Nonlinear loads · Stockwell

transform

1 Introduction

Owing to the rapid progress in the power sector, most of the customers are concerned
about the effective utilisation of their power appliances. However, stepping into
the next generation, unawareness in customers about the quality of power being
supplied is prevailing massively. Though there is a considerable development in mass
media and telecasting techniques, commercialisation has become the cornerstones

P. Srikanth (B)
Directorate of Civil Works and Estates, CCE (R&D), South Secunderabad, Telangana, India
e-mail: [email protected]
C. Koley
Department of Electrical and Electronics Engineering, NIT Durgapur, Durgapur 713 209, West
Bengal, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 249

H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_11
250 P. Srikanth and C. Koley

of these organisations. Consequently, many customers are still in darkness regarding

power quality. Since late 1980s, power quality (PQ) has gained enormous interest
in the researchers, power sector and utilities. However, PQ issues have failed to
fetch interest in the customers. Throughout the world, there are defined standards of
power production [1]. Even there are parameters that define the characteristics of the
power supply. A failure or maloperation of customer equipment is majorly due to
PQ problems occurring in voltage, current or frequency, thereby deviating from the
defined standards. The utilities state that the customers need awareness about PQ.
The statement can be convinced by the results obtained regarding the survey and
analysis of power quality problems. The survey brings out to notice that in customer
perception 17% of the PQ problems are by the utility and 12% is by the customers,
whereas from the utility perception, it is 25% by customers and 1% by utilities [2, 3].
From such statistics and the expert view, i.e. from the utility, it can be concluded that
the customers are playing a significant role in PQ problem manifestation. However,
it is a difficult task to make every customer to be known about the PQ problems.
Further, it is also challenging to identify and analyse the source and type of power
quality problem in a real-time power system. This information gives the distribution
and transmission power companies to take speedy remedial actions and generate
statistical reports. Vast research has been done on finding out from which side of
utility and customer the PQ problem arises.
Consequently, there is a need to analyse and identify various PQ problems. There
is a considerable advance in this direction, especially using time-frequency reso-
lution techniques, viz., Gabor-Wigner transform (GWT), Wavelet transform (WT),
S-transform (ST), etc. [4–12]. However, these signal processing techniques were not
explored in the recognition and monitoring of household nonlinear loads. Researchers
have used spectrogram and wavelet for nonlinear load identification and its analy-
sis [13–16]. To further contribute to the research domain, monitoring, analysis and
recognition of the household nonlinear using S-transform have been taken up in this
chapter. The Stockwell Transform technique is used to produce time-frequency sig-
natures for a given signal. A novel technique derived utilising the inverse properties
of ST has been proposed to provide signatures in the time domain [17, 18]. Conse-
quently, indices have been derived for providing the input to the fuzzy logic algorithm
for recognising the type of domestic harmonic injecting load [19]. All the test signals
obtained are real time, live recorded [20], and the plots are obtained using MAT-
LAB® [21]. Further sections provide the details about the proposed methodology,
case studies, analysis along with results and conclusions at the end.
Intelligent System for Monitoring and Recognising … 251

2 Background of S-Transform and Proposed Methodology

2.1 Continuous S-Transform (CST)

As proposed by Stockwell, the CST can be represented as a phase-corrected contin-

uous wavelet transform (CWT). As per the standard definition, CWT is defined as
shown in Eq. (1)

∞
B(t2 , f 2 ) = p(t1 ) ∗ b(t1 − t2 , f 2 )dt1 (1)
−∞

where B is the replicated output w.r.t input wavelet. The resolution is governed
by the wavelet width-defining dilation. Therefore, by multiplying the CWT with
‘exp(i2π f 1 t2 )’ which is the phase correction factor, the S-Transform can be obtained

S(t2 , f 1 ) = exp(i2π f 1 t2 ) ∗ B(t2 , f 2 ) (2)

In this case, the definition of mother wavelet is shown in Eq. (3).

  2 2 
| f1| t1 f 1
B(t1 , f 1 ) = √ ∗ exp − ∗ exp(−i2π f 1 t1 ) (3)
2π 2

Further, in the case of Eq. (3), the frequency of the dilation factor differs inversely.
The Continuous Stockwell Transform (CST) is therefore described as

+∞    
| f1| (t1 − t2 )2 f 12
S(t2 , f 1 ) = p(t1 ) √ . exp − . exp(−(i2π f 1 t1 )) . dt1 (4)
2π 2
−∞

where Gaussian window size is governed by the frequency

1
ρ= (5)
| f1|

2.2 Discrete S-Transform (DST)

The DST is defined as the shift operation on the Fourier spectrum for any local
spectra and is shown in Eq. (6),
252 P. Srikanth and C. Koley

N −1 
 
n m+n −2π 2 m 2 i2πmk
S[kT, ]= P .e n2 e N for n = 0 (6)
NT k=0
N T

where k, n, m varies from 0 to N−1 with Sampling time ‘T ’ and Number of samples
as ‘N’.
It is known that S-Transform does localisation of both phase and amplitude spec-
trum. The Gaussian window defined in ST varies with the input data frequency and
time. Due to such a definition, ST maintains good time and frequency resolution.
Stockwell Transform can be effectively applied to all types of power system distur-
bances due to the non-stationary nature; hence; ST magnitude matrix (STM) is used
to analyse the power system harmonic producing nonlinear loads. ST frequencies
and time both are generated as a single column and row matrices, respectively. Rows
of STM contains the information about ST amplitude versus frequencies, and the
columns contain the information about the variation of ST amplitude with the vari-
ation in time. The magnitude (z-axis) versus time (x-axis) or frequency (y-axis) plot
is obtained using Eq. (6).

2.3 Proposed Methodology

The generalised S-transform is totally invertible as it satisfies the condition shown

in Eq. (7) [17].

+∞
g(t2 − t1 , f 1 ) . dt1 = 1 (7)
−∞

Therefore, by averaging the time of the S-spectrum S(t2 , f 1 ) a spectrum U ( f 1 )

can surely be yielded, which is expressed as Eq. (8).

+∞ +∞ +∞

−i2π f 1 t1
S(t2 , f 1 )dt2 = u(t1 )e g(t2 − t1 , f 1 )dt2 dt1 = U ( f 1 ) (8)
−∞ −∞ −∞

It can be concluded that to obtain an unfiltered inverse S-Transform (IST), one-

time inverse Fourier Transform operation will be sufficient. In this chapter, all the
algorithms are run using MATLAB and Eq. (8) is derived using the Inverse Fourier
Transform function [17]. The final Inverse ST is described as an inverse Fourier
Transform of ST Matrix, as shown in Eq. (9).

INST = INFT[S(t2 , f 1 )] (9)

Intelligent System for Monitoring and Recognising … 253

Here, the matrix S(t2 , f 1 ) has the dimensions same as that of the resultant matrix
obtained in Eq. (9). It is interesting to observe that when the absolute value of
S(t2 , f 1 ) is plotted against time samples, a group of Gaussian windows with dilation
and translation are seen in Fig. 1b. Subsequently, a plot between the maximum values
of these translated and dilated Gaussian windows versus ST time can be obtained as
shown in Fig. 1c. In the proposed chapter for recognition of nonlinear loads, unique
signatures are generated using the plot shown in Fig. 1c. Further, parameter extraction
has been done using these values to feed the fuzzy identification box. The proposed
methodology is termed as Maximum values of Unfiltered Inverse ST (MUNIST) as
presented in [18]. Summarily, Eq. (10) represents the proposed MUNIST technique.

M(t) = max(INFT[S(t2 , f 1 )]) (10)

The word ‘max’ relates to the highest value of the internal matrix in Eq. 10
and is the MUNIST matrix obtained in the time domain. From Fig. 1b it can be

Fig. 1 a Swell event considered for testing, b inverse ST Amplitude versus time plot, c plot between
IST Gaussian windows maximum values shown in (b) versus time
254 P. Srikanth and C. Koley

inferred that the magnitude of the Gaussian window varies with the change in signal
amplitude. It can be visualised that for the duration 0.2–0.28 s in the input signal
(represented by dotted lines) the Gaussian window increases its amplitude along with
the variation in swell signal. The count of Gaussian windows is defined and governed
by sampling rate. This number is 250 in the present case, which is half the number
of samples. For a given longer duration, the number of Gaussian windows will be
produced, leading to unclear plots, as shown in Fig. 1b. Hence, the plot between the
MUNIST time samples and maximum values of each Gaussian window of Fig. 1b
can be utilised as a technique for generating unique signatures. DST technique also
produces signatures that are similar to Fig. 1c; however, it carries few disadvantages
as explained in further sections.

3 Practical Setup

In order to perform the analysis and recognition of various household nonlinear

loads, the variety and availability of the type of nonlinear loads need to be decided.
In this regard, with a glance into the literature, it can be recalled that all electronic run
loads are the cause of nonlinearity. Further, aged out malfunctioning equipment is
also responsible for such nonlinear behaviour. Exploring through various buildings
nearby, it is found the equipment like Printers, Monitors, CPUs, Fluorescent lamps
and many such electronic run loads are found in daily use. Hence, these equip-
ments are considered for analysis and identification. Table 1 shows the equipment
considered along with the PQ Meters used for recording the required data.
Except for the fluorescent lamp, the other equipment is in service from long-
time whereas the fluorescent lamp is in use for the first time. For verification, only
essential number of observations is recorded, and during testing minimum power
has been spent for the required number of samples. Few measurements have been

Table 1 Equipment details along with recording device

S. no Name of the equipment Technical data Units
1 Fluke power quality analyzer 115-230 V, 1000 Vrms, 400A, 1 No.
4 M/5pF
2 Central processing unit Pentium 4, IBM Thinkcenter 1 No.
3 Polychrome monitor CRT monitor with 1280 × 1024 1 No.
resolution at 50 Hz—Model E74 IBM
4 Black and white laser printer XEROX, 2.5 A, 240 V, 50 Hz 1 No.
5 Fluorescent lamp Havells, 15 W, 220–240 V, 115 mA, 1 No.
50 Hz
6 T8 tube light 40 W, 220–240 V, 50 Hz, Crompton 1 No.
Intelligent System for Monitoring and Recognising … 255

repeated to record right signals for analysis. Fluke® Power Quality Analyser (PQA)
is used for recording all the measurements, and its specifications are given in Table 1
[20].

4 Case Studies and Discussion

Domestic nonlinear loads that injects harmonics into the residential power system
and which are commonly used in domestic purposes are considered in this proposed
work. The technical details, along with the make and quantity, are mentioned in
Table 1. All the techniques, as explained in Sects. 2.1, 2.2 and 2.3 have been used
for comparison. The sampling rate of the data is set at 7.5 kHz per cycle, and all
the equipment are of 50 Hz rated frequency. In total per case study, 300 samples are
obtained for analysis of two cycles of data.
The logged waveforms data by the Fluke 435® PQ meter are transferred to the
MATLAB domain and are then plotted for better readability as provided in Figs. 2a–
6a. The output of the ST matrix of the CST algorithm for the input signals at 2a–6a
is plotted in Fig. 2b–6b. Some valuable information is obtained by searching the
maximum value of the whole STM, which is done through an iterative algorithm.
Using such information, i.e. variation of ST matrix column versus time Figs. 2c–6c
are drawn whereas 2d–6d show the variation of the maximum value of ST magnitude
with the frequency of the STM. This plot is included just to show the composition of
harmonics in the recorded signal. Figure 2e–4e represent the magnitude versus time
output generated by the proposed MUNIST algorithm, as discussed in Sect. 2.

4.1 Case 1—Central Processing Unit

Personal computers have become an integral part of any organisation as on today.

Apart from that it is also a household device in this generation. All the components
of a Desktop System, namely, CPU, Monitor and the Printer are analysed and studied
separately. The rated specifications are given in Table 1. CPU draws a non-sinusoidal
current waveform, and it is due to the pulsed switching of resistors, inductors and
capacitors along with Integrated Circuits. The non-sinusoidal current is generated
due to electronic switching operations. The active and reactive power rating is less
for a CPU; however, it injects a considerable harmonics into the 1-phase system. To
get an exact waveform during its regular usage, the waveforms are recorded only
with an optimal count.
The logged data of waveforms along with the acquired results from the examina-
tion of CPU current is demonstrated in Fig. 2. In Fig. 2a the maximum value of CPU
current is 0.8 A. It is evident from the figure and clarified above, the captured signal
is non-sinusoidal, which indicates the presence of harmonics. There exists a medium
frequency burst in the sign at time t = 0.006 s and a high-frequency harmonic causing
256 P. Srikanth and C. Koley

Fig. 2 Results obtained from analysis of recorded current of a CPU

a drop t = 0.004 s. Such irregularity is monotonous in the entire series and altogether
eight (08) uncommon recurrences are found in this waveform.
As mentioned earlier in this chapter, both the time and frequency resolution are
good using continuous ST, which is visible from the contour plot and it is shown in
Fig. 2b. Four numbers of contours are seen at the recorded signal’s high-frequency
range. Also present in medium frequency range are ST contours that give the current
Intelligent System for Monitoring and Recognising … 257

waveform a typical signature. Figure 2c shows the DST plot between the row with
the highest ST magnitude value and time samples. Due to the variation in the input
signal, it is expected that there will be a variation in the DST peak values [8–11].
However, the plot obtained is a straight line with a value of 0.2, which is constant
for the complete time interval.
Figure 2d is the plot between maximum values of ST matrix and frequency. This
plot is used for any signal to define the dominant frequencies. The frequency corre-
sponding to the Gaussian window’s maximum values provides arbitrary frequency
data corresponding to all the four peaks, as seen in Fig. 2d. It is seen that the funda-
mental frequency element of 50 Hz transformed using CST has a value of 0.0067. The
successive peaks in the ST domain are 3rd, 5th, 9th and 13th harmonics with a cor-
responding frequency value of 0.02, 0.033, 0.06 and 0.086, indicating the existence
of odd harmonics.
The last figure, i.e. Fig. 2e shows the plot derived using the MUNIST method,
which is a time-domain signature and distinctive. For every positive and negative
peak in the input signal of Fig. 2a, the MUNIST signature has a conical shape. Like
DST, the MUNIST signature is unique and varying. The significance of the proposed
technique is justified in further case studies.

4.2 Case 2—Flickering Tube Light

A flicker is defined as irregular lighting of a tube light, which is usually easy to

perceive by a human. In order to study flicker, a near to outage T8 tube light flickering
due to its ageing is considered. The voltage recorded at the input terminals is then used
to map in the S-transform time–frequency domain. During flicker, many significant
frequency components get clubbed besides the variation of voltage, which is analysed
and explained further.
As seen in Fig. 3a, there exists a disturbance in the recorded voltage for the
complete two cycles. This data is useful to recognise the type of nonlinear load.
At the moment t = 0.009 s and t = 0.027 s, there is a voltage drop with a chirp
voltage variation. These differences correspond to the existence of elements of high-
frequency. Further, after time t = 0.028 s, there is a voltage interruption.
The time–frequency variation after performing ST of the waveform is visible in
the contour plot in Fig. 3b. Two contours in the high-frequency range corresponding
to the peaks in the flickering voltage are visible with two more contours in the medium
frequency range. The rest of the plot constitutes sag and interruption.
The sudden magnitude change in the signal can be seen from the plot of DST
amplitude-time shown in Fig. 3c. Two peaks corresponding to the flickered magnitude
are seen in this plot. The highest value in the ST domain has reached 77, which
corresponds to 230 V measured voltage across the tube light. The magnitude of DST
is seen to vary with the peak value of the signal.
The plot, shown in Fig. 3d, represents the frequencies of the peaks, which are
numerically non-integral versions of 50 Hz frequency. This indicates the existence
258 P. Srikanth and C. Koley

Fig. 3 Results obtained from analysis of recorded voltage of a flickering tube light

of inter-harmonics in the signal and the presence is very substantial [3, 13]. Cor-
responding to the peaks of the ST plot shown in Fig. 4d ST frequency values are
0.00667, 0.01667, 0.02667 and 0.04333. These values indicate that flicker is the sum-
mation of both fundamental and fractional components of fundamental frequency and
which corresponds to frequencies of 2.5, 4 and 6.5 times the fundamental in this case.
Intelligent System for Monitoring and Recognising … 259

Fig. 4 Results obtained from analysis of recorded current of a CFL

The plot corresponds to row with ST matrix maximum values, and the information
about the high frequencies is not available. MUNIST plot shown in Fig. 3e is having
similar spikes as of DST in Fig. 3c. However, the variation is typical and closer to
identification. For a small dip at t = 0.008 s in the current, the MUNIST has produced
260 P. Srikanth and C. Koley

a spike. Again at t = 0.025 s there is noise in the input signal, and corresponding to
that MUNIST plot has changed the amplitude dynamically.

4.3 Case 3—Fluorescent Lamp

Since the power consumption of tube lights is more, fluorescent lamps have replaced
almost all around the world and which is a revolution in the world platform. These
lamps contain a ballast at the bottom, which is the primary source for injection
of current harmonics. A lamp with rated capacity, shown in Table 1, S No. 5 was
considered to witness the type of harmonics present in its current.
The CFL lamp input current as recorded by the PQA is shown in Fig. 4a. As
per the design and literature of the CFL, the current non-sinusoidal [3] with a peak
current of 0.3366 A. It is interesting to note that there are also oscillations that
correspond to additional low and high-frequency components. The current waveform
recorded using the PQA is shown in Fig. 4a. As expected, the current drawn is non-
sinusoidal with a maximum current of 0.3366 A. It is also interesting to see the
existence of oscillations, which corresponds lower to the higher range of components
of frequency.
Figure 4b shows the contour plot generated using ST of the signal. The plot shows
that there are many frequencies in the signal. The contours are distributed from lower
to higher frequency ranges in the ST contour plot, thereby generating a contour plot
at precisely every peak of the waveform of Fig. 4a. The waveform is therefore proven
to contain varied frequencies.
The DST technique, as applied to the input current has produced a straight line
shown in Fig. 4c. A similar plot was obtained for CPU in Fig. 2c previously. However,
here the maximum value is 0.05, which is due to decrease in the current drawn.
The plot, shown in Fig. 4d, contains many peaks at different frequencies. It is seen
that the first three peaks are present at values of 0.0067, 0.02 and 0.033 on the
frequency axis. These values correspond to fundamental frequency, 3rd and 5th
harmonic components, respectively, present in the signal. The next two peaks are
occurring at 0.05667 and 0.1233 on the frequency axis, which is non-integral again,
as seen in the previous case. These frequencies are the normalised values obtained in
the ST domain. In real time these correspond to 8.5 and 18.5 times the fundamental.
The last two peaks correspond to 35th and 59th harmonic with values of 0.236 and
0.393 in ST domain. The presence of number of peaks gives an idea that the signal
is a sum of many signals with different frequencies ranging from low to high.
The proposed MUNIST plot is shown in Fig. 4e. The variation in the MUNIST
magnitude is beneficial as compared to DST. It can be observed that the spikes in the
MUNIST plot are of increasing amplitude up to 250th sample, which is due to small
changes in input amplitude. The DST has failed to show the variations except that its
maximum value has reached 0.05. The uniqueness of MUNIST is still maintained.
Figures 5 and 6 show the similar plots obtained for a Polychrome CRT Monitor and
Printer. The CST and DST plots are almost similar in both the cases leading to lack
Intelligent System for Monitoring and Recognising … 261

Fig. 5 Comparison of ST, DST and MUNIST algorithm for the recorded current of a polychrome
monitor
262 P. Srikanth and C. Koley

Fig. 6 Comparison of ST, DST and MUNIST algorithm for the recorded current of a printer

of discrimination. In such cases, it is difficult to classify the type of nonlinear loads.

However, it is evident from these figures that the signatures of MUNIST algorithm
are unique for all the cases considered in this study. Though the signatures obtained
by the MUNIST algorithm are unique compared to DST algorithm an untrained
person cannot directly predict the type of load. Hence, for better identification, the
Intelligent System for Monitoring and Recognising … 263

fuzzy logic machine has been added to the output. The details about the fuzzy logic
design and the consecutive outputs are discussed in the next section.

5 Parameter Extraction

To recognise different nonlinear loads, it is necessary to extract few parameters.

There is no necessity that these parameters shall be associated with the recorded
current waveforms only and hence the MUNIST values acquired from different case
studies in Sects. 4.1–4.3 and Figs. 5e, 6e are used. From this, it can be observed
that the behaviour of maximum value and mean values in MUPLOT are unique and
non-overlapping. Hence, these variations in value have been used for designing the
parameters presented in Table 2, thereby making the identification process simple
and discriminative.
Since the voltage rating of all nonlinear loads is 230 V, a multiplication factor of
230 is considered in the calculation of AM. The identification of nonlinear loads is
performed based on the above-specified variables, and the Fuzzy Decision Box is
also included for automating the process. For each type of nonlinear load, the range
of values derived for parameter AM is shown in Table 3.
In Table 3, the following notations have been used: CPU = Central Processing
Unit, MON = Monitor, PRIN = Printer.
FLOR = Fluorescent Lamp, FLIK = Flickering Tube Light.
It is visible that for any of the considered case studies, the values obtained from
the defined or extracted parameters are not overlapped. To verify the robustness of
the suggested method, case studies other than those mentioned in Chapter 4 are also

Table 2 Extracted parameters using the MUNIST method

Designed parameter Parameter description
MUNISTMAG-SINE Magnitude calculated using MUNIST algorithm for a sine
wave of 1per-unit, 50 Hz as input. The MUNIST magnitude
for this case is 0.00868
MUNISTMAG-LOAD Magnitude obtained using MUNIST algorithm for any input
current of 1 per-unit, 50 Hz as input
AM (Avg. MUNIST magnitude) 230 * (Avg. of [MUNISTMAG-SINE —MUNISTMAG-LOAD ])

Table 3 Type of nonlinear

Type of load AM
load and corresponding
values of extracted FLIK <−20
parameters obtained using FLOR −10 to −20
MUNIST algorithm
PRIN −3 to −10
MON 0 and −3
CPU 5–0
264 P. Srikanth and C. Koley

included. The rule sets and fuzzy variables have now been intended to design based
on these values. The next section discusses the design and outcomes obtained from
the fuzzy variables.

6 Design of Fuzzy Logic Decision Box

Custom designed fuzzy parameters and respective details are as follows. Figure 7
displays both input and output Membership Functions (MFs) considered for identifi-
cation of type of nonlinear load [19]. Most of the input and output MFs are triangular,
and they are considered so for the reason that the range of values is accurate and highly
discriminative. The AM input includes only one trapezoidal feature representing the
<−20 value range, as shown in Fig. 7a.
Figure 8 shows the broad FIS framework or the Fuzzy decision box layout. For
identifying or recognising the type of nonlinear through fuzzy logic, a rule base has
been generated, as shown in Table 4. Based on this rule base the outcomes were
acquired and presented in Table 5. The AM values are achieved by taking an average

Fig. 7 Membership functions of fuzzy parameters AM and NON-LIN-LOAD-TYPE

Fig. 8 Fuzzy Inference System (FIS) named ‘nonlinload’ designed for recognising the nonlinear
load type
Intelligent System for Monitoring and Recognising … 265

Table 4 Fuzzy rule base created for identifying the nonlinear load type
Name Rules
Rule-1 If (AM is M5) then (NON-LIN-LOAD-TYPE is CPU) (1)
Rule-2 If (AM is M4) then (NON-LIN-LOAD-TYPE is MON) (1)
Rule-3 If (AM is M3) then (NON-LIN-LOAD-TYPE is PRIN) (1)
Rule-4 If (AM is M2) then (NON-LIN-LOAD-TYPE is FLOR) (1)
Rule-5 If (AM is M1) then (NON-LIN-LOAD-TYPE is FLIK) (1)

Table 5 Values of AM obtained for various nonlinear loads

Type of load AM (mean of 25 recordings) Correct number of fuzzy Accuracy (%)
decisions produced out of 25
test runs
CPU 1.8 25 100
MON −2.7 25 100
PRIN −6.84 25 100
FLOR −10.38 25 100
FLIK −27.54 25 100

of 25 recordings of each case study with an interval of two cycles, i.e. 40 ms for each
case. These number of recordings are done to test the robustness of the technique for
repeated applications.
All types of nonlinear loads, as mentioned in Table 3 have been considered for
analysis and recognition. It is visible from Table 5 that the extracted parameters from
the proposed MUNIST technique have produced discriminative values for different
types of nonlinear loads without any overlap or confusion. Till now many of the
techniques proposed for nonlinear identification have some measures of overlap.
Here, it is possible to achieve highly discriminative values and ranges because of
unique extracted parameters. For better accuracy and fastness, the proposed algorithm
when fed to a fuzzy logic has lead to 100% accuracy. Such a system can finally
be integrated with any Power Quality Monitoring system leading to integrated PQ
monitoring. Thus, the combination of MUNIST and Fuzzy Logic can be used as
a robust tool for analysing and recognising various types of domestic household
nonlinear loads.

7 Conclusions

The chapter proposes complete assessment of different kinds of harmonic loads

to identify them using various signal processing methods. The method Stockwell
Transform gives a full understanding of the time, and frequency characteristics of
nonlinear household loads. It can further be extended to identify other kinds of loads
266 P. Srikanth and C. Koley

also. The findings provided in Figs. 2–6 shows that the behaviour of nonlinear loads
is unique, and hence it is feasible to identify them accurately. However, there is lack
of ease in visual identification, and it has been observed that the ST signatures are
overlapping for the case of Monitor and Printer. Hence, in order to automatically
identify and distinguish such type of loads, a new method (MUNIST) has been
mentioned in Sect. 2 and is utilised in the present work.
Based on properties of inverse of S-Transform is in time domain and the variations
of the magnitude of the domestic nonlinear loads under investigation, i.e. Figures 2e–
6e are unique compared to the DST. It is proved that for better pictorial or visual
observation the proposed MUNIST technique can be used as the uniqueness and
distinctive variation is seen even for the case of Monitor and Printer. For automat-
ing the identification process, a Fuzzy logic has been added. The outcome of such
combination of DSP technique with fuzzy machine is highly robust and accurate.
The method can be further used for industrial nonlinear loads also, thus, leading to
design of an Integrated Power Quality Monitoring System for both Transmission and
Distribution utility companies.

Acknowledgements The first author is thankful to Shri Ashwani Chandel, Professor, EED, NIT
Hamirpur for providing an opportunity to record waveforms of various equipment using Fluke PQ
Meter during Post Graduation.

References

1. IEEE Recommended Practice and Requirements for Harmonic Control in Electric Power
Systems—IEEE 519-2014, IEEE Power Engineering Society, Mar 2014
2. M.H.J. Bollen, I.Y.-H. Gu, Signal Processing of Power Quality Disturbances (IEEE Press
Series on Power Engineering, IEEE Press, Wiley Inc. Publication, 2006)
3. F.C. De La Rosa, Harmonics and Power Systems (CRC Press, Taylor & Francis Group, 2006)
4. J. Barros, R.I. Diego, M. de Apraiz, Applications of wavelet transform for analysis of harmonic
distortion in power systems: a review. IEEE Trans. Instrum. Meas. 61(10), 2604–2611 (2012)
5. R.G. Stockwell, S-Transform analysis of gravity wave activity from a small scale network.
Ph.D. dissertation, Department of Physics, Western Ontario University, London, Ontario, 1999
6. R.G. Stockwell, A basis for efficient representation of the S-transform. Digit. Signal Process.
17(1), 371–393 (2007)
7. S. Ventosa, C. Simon, M. Schimmel, J.J. Dañobeitia, A. Mànuel, The S-Transform from a
wavelet point of view. IEEE Trans. Signal Process. 44, 2771–77 (2008)
8. J. Yi, J. Peng, Power quality disturbances denoising using modified S-transform, in 2009
International Conference on Sustainable Power Generation and Supply (Nanjing, 2009), pp. 1–
5
9. P.K. Dash, M.V. Chilukuri, Hybrid S-transform and Kalman filtering approach for detection and
measurement of short duration disturbances in power networks. IEEE Trans. Instrum. Meas.
53(2), 588–596 (2004)
10. C. Venkatesh, D.V.S.S. Siva Sarma, M. Sydulu, Detection of voltage sag/swell and harmonics
using discrete S-transform, in TENCON 2008--2008 IEEE Region 10 Conference (Hyderabad,
2008), pp. 1–6
11. S. Kaewarsa, Classification of power quality disturbances using S-transform based artifi-
cial neural networks, in 2009 IEEE International Conference on Intelligent Computing and
Intelligent Systems (Shanghai, 2009), pp. 566–570
Intelligent System for Monitoring and Recognising … 267

12. S.H. Cho, G. Jang, S.H. Kwon, Time-Frequency analyses of power-quality disturbances via
the Gabor–Wigner transform. IEEE Trans. Power Deliv. 25, 494–99 (2010)
13. V.K. Sharma, M. Doja, M.N. Ibraheem, M.A. Khan, Power quality assessment and harmonic,
comparison of typical nonlinear electronic loads, in International Conference on Industrial
Technology (2000), pp. 729–734
14. H.O. Aintablian, H.W. Hill, Harmonic currents generated by personal computers and their
effects on the distribution system neutral current, in IEEE (1993), pp. 1483–89
15. E.O.A. Larsson, C.M. Lundmark, M.H.J. Bollen, Measurement of current taken by fluorescent
lights in the frequency range 2–150 kHz, in IEEE (2006), pp. 1–6
16. S.K. Rönnberg, M. Wahlberg, M.H.J. Bollen, M. Lundmark, Equipment currents in the
frequency range 9–95 kHz, measured in a realistic environment, in IEEE (2008), pp. 1–8
17. C. Simon, S. Ventosa, M. Schimmel, A. Heldring, J.J. Dañobeitia, J. Gallart, A. Mànuel, The S-
transform and its inverses: side effects of discretizing and filtering. IEEE Trans. Signal Process.
55, 4928–37 (2007)
18. P. Srikanth, A.K. Chandel, A.K. Naik, Identification of power quality events using inverse
properties of S-transform, in IEEE PES Power System Conference and Exposition (2011),
pp. 1–7
19. A. Banshwar, A.K. Chandel, Identification of harmonic sources using fuzzy logic. IEEE PEDES
2010, 1–6 (2010)
20. Fluke 435 A—Power quality analyser
21. MATLAB/SIMULINK 7.6
Fault Analysis of Variable Frequency
Transformer (VFT) for Power Transfer
In-Between Synchronous Grids

Farhad Ilahi Bakhsh, Dheeraj Kumar Khatod and Hasmat Malik

Abstract VFT is the latest power transmission technology. It can transfer power
within synchronous, in-between synchronous and in-between asynchronous grids.
This chapter discusses the application of the VFT as controllable, bidirectional AC
link for power flow control in-between synchronous grids under steady state and
fault conditions. For the study of steady state analysis, a complete VFT simulation
model has been developed in MATLAB/Simulink. A sequence of studies on power
transfer in-between synchronous grids is carried out with the developed model. The
power flow control with respect to applied torque is achieved. The VFT concept
and its application in-between synchronous grids are verified by the obtained results.
Moreover, LLLG fault analysis has been carried out at three locations under different
conditions in order to understand the effect of LLLG fault on VFT and how VFT
reacts to mitigate the effect of the fault.

Keywords Fault analysis · Synchronous grids · Power transfer · Steady state

analysis · Variable frequency transformer (VFT)

F. I. Bakhsh (B)
Department of Electrical Engineering, National Institute of Technology, Srinagar, Hazratbal,
Srinagar 190006, Jammu and Kashmir, India
e-mail: [email protected]
D. K. Khatod
Department of Electrical Engineering, Indian Institute of Technology, Roorkee, Roorkee 247667,
India
e-mail: [email protected]
H. Malik
BEARS, University Town, NUS Campus, College Town, Singapore
e-mail: [email protected]
Division of ICE, NSUT Delhi, New Delhi, India
BERS, NUS, Singapore, Singapore

© Springer Nature Singapore Pte Ltd. 2020 269

H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_12
270 F. I. Bakhsh et al.

Nomenclature

VFT Variable Frequency Transformer

LLLG Triple Line to Ground
PAR Phase Angle Regulator
HVDC High Voltage Direct Current
WRIM Wound Rotor Induction Machine
DDM Dc Drive Motor

1 Introduction

VFT is a power transferring device which transfers electrical power in-between

power systems or grids. The power flow through VFT is bidirectional and control-
lable. First time in world, the VFT was developed by GE and practically installed
in Hydro-Quebec’s Langlois substation for exchanging the power in-between the
Quebec (Canada) grid and New York (USA) grid [1–3].
The construction of VFT is identical to conventional asynchronous machine. The
two different electrical networks are linked to each other via VFT [4, 5]. The various
applications of VFT and its comparison with phase angle regulator (PAR) and HVDC
system are discussed in [5–7]. The simulations of the various applications of VFT
are presented in [8, 9].
The response of VFT to grid disturbances by simulating events like three-phase
faults, single phase to ground faults and islanding is shown in [10]. A simulation study
under PSCAD/EMTDC environment on Kundur’s four-machine two-area system in
order to realize the VFT’s ability in suppressing the power oscillations, improving the
power system stability and preventing faults from spreading is discussed in [11]. In
[12], a model of the VFT has been presented using MATLAB/Simulink to represent
its dynamic characteristics. Based on this model, they conducted some simulations
to study VFT’s characteristics under fault conditions and frequency disturbances.
With the help of simulation results, they showed the VFT’s outstanding capability in
improving power stability and preventing faults from spreading into the neighboring
systems. Further, a simulation study of VFT connected between two asynchronous
grids under the MATLAB environment to analyze its impact on the networks under
fault condition is carried out in [13].
In this chapter, the application of VFT as an AC link or tie for power transfer
in-between synchronous grids under steady state and fault conditions is discussed.
A digital simulation model of VFT and its control system are developed under MAT-
LAB/Simulink environment. A sequence of studies on power transfer under steady
state and fault conditions are carried out with this model. This chapter has been orga-
nized in V sections. Section 2 of the chapter describes the VFT modeling. Section 3
discusses the steady state analysis of VFT for power transfer in-between synchronous
Fault Analysis of Variable Frequency Transformer (VFT) … 271

grids. Section 4 presents the LLLG fault analysis of VFT during power transfer in-
between synchronous grids under MATLAB/Simulink environment. Finally, Sect. 5
concludes the chapter.

2 VFT Modeling

The VFT is modeled as a doubly-fed WRIM whose rotor is mechanically coupled

with the DDM as shown in Fig. 1. The detailed modeling of VFT is done based on
the technique presented in [14, 15].
The power transfer through the VFT is given by [15]:
 
Vs × Vr
PV F T = × Sin[θs − (θr + θr s )] (1)
X sr

The final power and torque equations of the DC drive motor for power flow from
grid #1 to grid #2 is given as follows [16]:

PD = Pr × [1 − ( f s / fr )] (2)

TD = Nr × ψa × Ir (3)

The final power and torque equations of the DC drive motor for power flow from
grid #2 to grid #1 is given as follows [16]:

Fig. 1 The VFT model representation

272 F. I. Bakhsh et al.

PD = Ps × [1 − ( fr / f s )] (4)

TD = −Ns × ψa × Is (5)

3 Steady State Analysis of VFT

3.1 VFT Simulation Model for Steady State Analysis

For steady state analysis, VFT is represented as a doubly-fed WRIM. The torque of
doubly-fed WRIM is controlled by DDM as shown in Fig. 2. The doubly-fed WRIM
is having ratings of three phase, 4 kW, 400 V, 50 Hz. The grids #1 and #2 are simulated
with three-phase sources 1 and 2, respectively. The stator windings of doubly-fed
WRIM is connected to the three-phase source 1. The rotor windings of doubly-fed
WRIM is connected to the three-phase source 2. The power flow in-between grids
are measured by three-phase stator and three-phase rotor sides active power meters,
respectively. The DDM is simulated with signal builder which gives output signal
as mechanical torque Tm to the WRIM. Here, the grid #1 is having rating of 400 V
(L-L) and 50 Hz whereas grid #2 is having rating of 385 V (L-L) and 50 Hz. The
simulated model (as shown in Fig. 2) is used to analyze the power flow in-between
these synchronous grids using VFT.

Fig. 2 Simulation model for steady state analysis

Fault Analysis of Variable Frequency Transformer (VFT) … 273

Table 1 Simulation results

S. no. T D (Nm) Ps (W) Pr (W)
for power transfer in-between
synchronous grids 1 25 4122 −3603
2 20 3256 −2940
3 15 2418 −2246
4 10 1610 −1525
5 5 830.8 −777.0
6 0 79.13 0.980
7 −5 −646.5 809.3
8 −10 −1341 1645
9 −15 −2007 2508
10 −20 −2642 3403
11 −25 −3246 4327

3.2 Simulation Results of Steady State Analysis

The obtained simulation results are presented in Table 1.

It is quite obvious from Table 1 that when T D = 0 Nm, the power transfer in-
between the synchronous grids is zero, but there is power flow through grid #1 (Ps =
79.13 W) and grid #2 (Pr = 0.98 W). This shows that at this condition, VFT is taking
power from both the grids. As torque is applied in positive direction, grid #1 starts
suppling power to grid #2. Further, with the increase in the value of torque from 0
to 25 Nm the power transfers through VFT also increases. Now, as the direction of
applied torque is reversed, grid #2 starts suppling power to grid #1. Moreover, with
the increase in magnitude of torque from 0 to −25 Nm the power flow via VFT again
increases. The variation of power flow through grids (#1 and #2) with the applied
torque is achieved in Table 1 is shown in Fig. 3.
From Fig. 3 it is evident that the power flow in-between the synchronous grids is
bidirectional and directly proportional to the torque applied by DDM.

4 Fault Analysis of VFT

4.1 VFT Simulation Model for Fault Analysis

For fault analysis of VFT in MATLAB, a LLLG fault is created for 5 cycles, i.e., 0.1
from 1.5 to 1.6 s of simulation time at three different locations as shown in Fig. 4.
At each location the LLLG fault analysis is done for two possible cases, i.e., power
flow from grid #1 to grid #2 and power flow from grid #2 to grid #1.
274 F. I. Bakhsh et al.

Fig. 3 The power flow through grids (#1 and #2) with the applied torque

Fig. 4 Simulation model for LLLG fault analysis

4.2 Simulation Results of Fault Analysis

4.2.1 Fault at Rotor Terminals i.e. Location 1

Grid #1 to Grid #2 Power Flow

A LLLG fault is simulated in MATLAB/Simulink at location 1, i.e., rotor terminals

of WRIM. The value of applied torque is 25 Nm. Since the applied torque is positive,
Fault Analysis of Variable Frequency Transformer (VFT) … 275

Fig. 5 Plots showing different waveforms under normal operating condition at T D = 25 Nm

thus grid #1 is supplying power to grid #2. The fault occurs at 1.5 s and remained for
around 0.1 s (five cycles), i.e., till 1.6 s. Basically, here we are trying to find out what
happens to the system during LLLG fault. Hence, after five cycles the fault mitigates
and system comes to normal condition. The different waveforms obtained under
normal condition and during LLLG fault are shown in Figs. 5 and 6, respectively.
It is clear from Fig. 5 when 25 Nm torque is applied at the rotor of the doubly-fed
WRIM, the power flow through grid #1 is having positive value (Ps = 4.122 kW)
and power flow through grid #2 having negative value (Pr = −3.603 kW) which
means grid #1 is delivering power to grid #2. Further, the mechanical power from
the DDM is negative (PD = −413.6 W) which indicates that the DDM is supplying
power to the rotor of WRIM. Since both grids are having same frequency thus the
speed of the rotor of WRIM is zero. Moreover, the waveforms of rotor current, rotor
voltage, stator current, and stator voltage is almost sinusoidal. Hence, the harmonics
are negligible.
From Fig. 6, it is evident that when LLLG fault occurs at the rotor terminals of
WRIM (t = 1.5 s) the voltage at the rotor side of WRIM becomes zero but the voltage
at the stator side remains almost constant, i.e., due to LLLG fault at rotor terminals
276 F. I. Bakhsh et al.

Fig. 6 Plots showing different waveforms under LLLG fault condition at location 1 when T D =
25 Nm

there is small effect on the stator voltage. This is because as the LLLG fault occurs
at the rotor side of WRIM, it shorts the rotor terminals. Now the WRIM will work
as simple induction motor whose stator windings are supplied by grid #1 and rotor
windings are short circuited. Hence, at this condition
(a) The speed of WRIM increases from zero to 1440 rpm (rated speed) and remains
rated speed until fault is removed (t = 1.6 s).
(b) The power at rotor side changes from negative value (Pr = −3.603 kW) to
positive value (Pr = 200 W) which means that rotor is not supplying power
to grid #2 but it is taking power from grid #1. It remains positive until fault is
removed.
(c) Due to small effect on the stator voltage there is slight variation in DC motor
torque.
(d) Due to increase in speed of WRIM the speed of DC drive motor also increases
as both are mechanically coupled. Hence, this will result in increase in the
mechanical power of DC drive motor which results in slight increase of the
stator side power.
Fault Analysis of Variable Frequency Transformer (VFT) … 277

(e) The fault current or rotor side current shoots almost 25 times the value of rotor
side current at normal condition and remains until fault is removed but the stator
side current shoots to almost 7 times the value of stator side current at normal
condition and dies down to normal value at t = 1.57 s.
(f) Due to increase in stator current the stator power also increases and comes to
normal value at t = 1.57 s.
From here it can be concluded that as the LLLG fault occurs at the rotor terminals
of the WRIM the speed of the WRIM rotor will adjust itself to mitigate the effects
of the fault and thus provides sufficient time for switch gears to actuate. At t = 1.6 s
the fault is removed. After t = 1.6 s the transient period starts and each variable will
reach its steady state value after certain time.

Grid #2 to Grid #1 Power Flow

Here everything is same as in case 4.2.1.1 only the direction of applied torque is
reversed, i.e., −25 Nm. Since the applied torque is negative, hence grid #2 will
supply power to grid #1. The different waveforms obtained under normal condition
and during LLLG fault are shown in Figs. 7 and 8, respectively.
It is clear from Fig. 7 when a torque of −25 Nm is applied at the rotor of the WRIM,
the power flow through grid #1 is negative (Ps = −3.246 kW) and power flow through
grid #2 is positive (Pr = 4.327 kW) which means now grid #2 is delivering power
to grid #1. The value of mechanical power from the DC drive motor is −359.2 W
which means that at this condition also, the DDM is supplying power to the rotor of
the WRIM.
From Fig. 8 it is evident that variation in all variable are same as in case 4.2.1.1
only difference is that power at rotor side reduces from 4.327 kW to 200 W (instead
of changing from negative value to positive value) during fault period and the stator
side power changes from negative value (−3.246 kW) to positive value (20 kW) for
short duration. Moreover, the stator side current and power dies down to permissible
value at t = 1.55 s. Hence, from here it can conclude that the effect of fault on stator
side, in this case, is lesser than the previous case.

4.2.2 Fault at Stator Terminals i.e. Location 2

Grid #1 to Grid #2 Power Flow

In this case, all quantities are same as in case 4.2.1.1 only the location of fault is
different. The fault occurs at location 2, i.e., at the stator terminals of WRIM.
From Fig. 9, it is clear that when LLLG fault occurs at the stator terminals of
WRIM (at t = 1.5 s) the voltage at the stator side of WRIM becomes zero but there
is small effect on the rotor voltage. This happens because when the LLLG fault
occurs at the stator side, it shorts the stator terminals of WRIM. Now, the WRIM
278 F. I. Bakhsh et al.

Fig. 7 Plots showing different waveforms under normal operating condition at T D = −25 Nm

will again operate as simple induction motor whose (instead of stator windings) rotor
windings are supplied by grid #2 and stator windings are short circuited. Hence, at
this condition:
(a) The speed of WRIM increases from zero to −1440 rpm (negative sign shows
reverse direction of rotation of WRIM) and remains rated speed until fault is
mitigated (t = 1.6 s).
(b) The power at stator side decreases from 4.122 kW to a value of 200 W (no load
power of WRIM) which means that instead of working as VFT (i.e., supplying
power to grid #2) it is working as induction machine (i.e., taking power from
grid #2). The stator side power remains 200 W until fault is omitted.
(c) Since the stator terminals are shorted, there will be large effect on the DC motor
torque.
(d) The stator side current or fault current increases almost 20 times the value of
stator side current at normal condition and remains until fault is removed but
the rotor side current increases to almost 4 times the value of rotor current at
normal condition and settles down to normal value at t = 1.55 s (two and half
cycles).
Fault Analysis of Variable Frequency Transformer (VFT) … 279

Fig. 8 Plots showing different waveforms under LLLG fault condition at location 1
when T D = −25 Nm

(e) Due to increase in rotor current the rotor power also increases and comes to
normal value at t = 1.55 s.
(f) The speed of DC drive motor also increases due to increase in speed of WRIM
(mechanically coupled). Hence, causes increase the mechanical power of DC
drive motor.
From here it can be concluded that as the LLLG fault occurs at the stator terminals
of the WRIM the WRIM rotor will again adjusts itself to mitigate the effects of the
fault and thus provides sufficient operating time for switch gears to actuate. At t
= 1.6 s the fault is removed. After t = 1.6 s the transient period starts and each
variable will reach its steady state value after certain time depending on individual
time constant.
280 F. I. Bakhsh et al.

Fig. 9 Plots showing different waveforms under LLLG fault condition at location 2 when T D =
25 Nm

Grid #2 to Grid #1 Power Flow

In this case, all quantities are same as in case 4.2.1.2 only the location of fault is
changed from 1 to 2.
Figure 10 shows that the variation in all variable are almost same as in case 4.2.2.1
except that the power at stator side changes from −3.246 kW to 200 W during fault
period and the rotor side power increases from 4.327 to 20 kW for short duration.
Moreover, rotor side increase in current and power settles down to normal value at t
= 1.57 s. Hence, from here we can conclude that the effect of fault on rotor side, in
this case, is more than the case 4.2.2.1.
Fault Analysis of Variable Frequency Transformer (VFT) … 281

Fig. 10 Plots showing different waveforms under LLLG fault condition at location 2 when T D =
−25 Nm

4.2.3 Fault just Before DDM Connection Point i.e. Location 3

Grid #1 to Grid #2 Power Flow

In this case, all conditions are same as in case 4.2.2.1 only the location of fault
is different. Here the fault occurs at location 3, i.e., just before DC drive motor
connection point.
From Fig. 11 it is evident that the variation of all waveforms during fault is almost
same as in case 4.2.2.1 except the stator side current. The stator side current shoots
to almost 6 times (instead of 20 times) the normal value of stator side current for
short period (instead of full fault period) and comes down to within rated value at t
= 1.53 s. Thus, in this case the stator current is safer than cases 4.2.1.1 and 4.2.2.1.
282 F. I. Bakhsh et al.

Fig. 11 Plots showing different waveforms under LLLG fault condition at location 3 when T D =
25 Nm

Grid #2 to Grid #1 Power Flow

Figure 12 shows that the variation of all waveforms during fault is almost same as in
case 4.2.2.2 except the waveform of stator side current. Here, the stator side current
increases to almost 4 times (instead of 20 times) for short period (instead of full
fault period) and settles down at t = 1.58 s. Hence, from the results we can conclude
that the effect of fault on stator side, in this case, is lesser than the case 4.2.2.2. A
comparison of LLLG fault under all conditions is presented in Table 2.
In Table 2, a comparative analysis of LLLG fault under all discussed conditions is
presented. Here negligible means very small effect on the variable, small (S) stands
for little change which is within permissible limit of the variable, large (L) indicates
heavy change but within permissible limit or comes within permissible limit after
short period of time, severe shows very large change which is not within permissible
limit and may cause damage to the system. Hence, protection is required. From
Table 2 it is clear that as LLLG fault occurs at location 1 under two cases only the
rotor current becomes severe. Hence, over current protection is required at rotor side
only. Further, when LLLG fault occurs at location 2 under two cases only the stator
Fault Analysis of Variable Frequency Transformer (VFT) … 283

Fig. 12 Plot showing different waveforms under LLLG fault condition at location 3 when T D =
−25 Nm

side current becomes severe. Hence, over current protection is required at stator side
only. Moreover, when LLLG fault occurs at location 3 no change in variable is severe.
Hence, no protection is required. Thus, from Figs. 5, 6, 7, 8, 9, 10, 11, and 12 and
Table 2, it can be concluded that the system requires overcurrent protection at stator
and rotor sides of VFT under LLLG fault conditions.

5 Conclusion

The chapter presents the steady state and fault analysis of VFT. For this, a MAT-
LAB/Simulink simulation model has been developed. The simulation model is used
for steady state analysis of VFT for power flow control in-between the synchronous
grids. The simulation results are obtained using the simulation model is presented in
Table 1 and shown in Fig. 3. From the steady state simulation results it is evident that
VFT is viable technology for achieving real power transfer in-between synchronous
grids. Moreover, the power transfer through VFT is bidirectional, controllable, and
 284

Table 2 Results for LLLG fault

LLLG LLLG Effect
Fault Direction DC drive Power Stator side Stator side Stator side WRIM Rotor side Rotor side Rotor side
of DDM motor from the voltage, V s current, I s power, Ps Speed voltage, V r current, I r power, Pr
torque, T D torque, T D DDM, PD
(Nm)
Location 1 Positive Negligible L S L L L L Severe L
Negative Negligible S S L L L L Severe L
Location 2 Positive L L L Severe L L S L L
Negative L L L Severe L L S L L
Location 3 Positive L L L L L L S L L
Negative L L L L L L S L L
L stands for Large and S stands for Small
F. I. Bakhsh et al.
Fault Analysis of Variable Frequency Transformer (VFT) … 285

directly proportional to the torque applied at the rotor of the VFT. Thus, VFT concept
discussed in literature and its application in-between synchronous grids are verified
by simulation results. Further, a simulation model has been developed for LLLG
fault analysis. The LLLG fault analysis has been done at three locations under posi-
tive and negative DC motor torque conditions for understanding the effect of LLLG
fault on VFT while transferring the power from grid #1 to grid #2 and vice-versa.
From the simulation results obtained in Figs. 6, 8, 9, 10, 11, and 12 and Table 2 it
can be concluded that when the LLLG fault occurs, the speed of the WRIM rotor
will adjusts itself to mitigate the effects of the fault and thus provides sufficient time
for switch gears to actuate. Moreover, it can be further concluded that the system
requires overcurrent protection at both stator and rotor sides of VFT under LLLG
fault conditions which are generally provided in each machine. Hence, no additional
protection is required in VFT under LLLG fault condition.

References

1. E. Larsen, R. Piwko, D. McLaren, D. McNabb, M. Granger, M. Dusseault, L-P. Rollin,

J. Primeau, Variable Frequency Transformer—A New Alternative for Asynchronous Power
Transfer, Canada Power, Toronto, Ontario, Canada, September 28–30 (2004)
2. P. Doyon, D. McLaren, M. White, Y .Li, P. Truman, E. Larsen, C. Wegner, E. Pratico, R. Piwko,
Development of a 100 MW Variable Frequency Transformer, Canada Power, Toronto, Ontario,
Canada, September 28–30
3. F.I. Bakhsh, Application of Variable Frequency Transformer (VFT) for Integration of Wind
Energy System, Ph.D. Thesis, IIT Roorkee, India (2017)
4. F.I. Bakhsh, D.K. Khatod, A novel method for grid integration of synchronous generator based
wind energy generation system, in IEEE International Conference on Power Electronics, Drives
and Energy Systems (PEDES), IIT Bombay, India, 16–19 Dec (2014)
5. F.I. Bakhsh, VFT application for asynchronous power transfer. Int. J. Control Syst. Instrument.
USA, 4(1), 1–7 (2013)
6. F.I. Bakhsh, D.K. Khatod, Application of variable frequency transformer (VFT) for grid inter-
connection of PMSG based wind energy generation system, in Sustainable Energy Technologies
and Assessments, Elsevier, vol. 8, pp. 172–180
7. F.I. Bakhsh, D.K. Khatod, A new synchronous generator based wind energy conversion system
feeding an isolated load through variable frequency transformer, Renew. Energy, Elsevier, 86,
106–116 (2016)
8. F.I. Bakhsh, D.K. Khatod, Digital simulation of VFT applications between grid networks, in
Fourth International Joint Conference on Advances in Engineering and Technology, NCR,
India (2013), pp. 60–74, 13–14 Dec
9. B. Bagen, D. Jacobson, G. Lane, H. Turanli, Evaluation of the performance of back-to-
back HVDC converter and variable frequency transformer for power flow control in a weak
interconnection (IEEE PES General Meeting, Tampa, FL, 2007)
10. D. Nadeau, A 100-MW variable frequency transformer (VFT) on the Hydro-Québec transén-
ergie network—the behavior during disturbance, in IEEE Power Engineering Society General
Meeting, Tampa, FL, 24–28 June (2007)
11. R. Yuan, Y. Chen, G. Chen, Y. Sheng, Simulation model and characteristics of variable fre-
quency transformers used for grid interconnection, in IEEE Power & Energy Society General
Meeting (PES ‘09), Calgary, AB, 26–30 July (2009)
286 F. I. Bakhsh et al.

12. A. Hossam-El-Din, M.A. Abdullah, M. Ibrahim, A novel model to study the vFT perfor-
mance when controlling power transfer between weak and strong AC grids using MAT-
LAB/SIMULINK, in IEEE International Energy Conference and Exhibition (EnergyCon),
Manama, 18–22 Dec (2010)
13. K. Thesika, R. Rajasekaran, Evaluation of the performance of variable frequency transformer
for controllability in power transfer between asynchronous grids. Int. J. Commun. Eng. 03(02),
No. 3, 32–39
14. F.I. Bakhsh, M. Irshad, M.S.J. Asghar, Modeling and simulation of variable frequency trans-
former for power transfer in-between grid networks, in Indian International Conference on
Power Electronics (IICPE), Delhi, India, 28–30 Jan
15. A. Merkhouf, P. Doyon, S. Upadhyay, Variable frequency transformer—concept and elec-
tromagnetic design evaluation, in IEEE Transactions on Energy Conversion, vol. 23, no. 4,
December, pp. 989–996 (2008)
16. F.I. Bakhsh, D.K. Khatod, Closed loop control analysis of variable frequency transformer
(VFT), in IEEE Students’ Conference on Electrical, Electronics and Computer Sciences,
Bhopal, India, 1–2 March (2014)
Performance Enhancement
and Extension of DGA-Based
Transformer Fault Diagnosis Methods
Using Soft-Computing Techniques

Shakeb Ahmad Khan, M. Ajmal Khan, Shufali Ashraf Wani

and Md. Danish Equbal

Abstract Incipient faults are the slowly growing faults found within oil-immersed
power transformers. These anomalies develop due to the deteriorating transformer
insulating system. Various types of electrical, mechanical and environmental stress;
manufacturing defects, in addition to, normal ageing phenomenon contribute to the
degrading health of transformer insulation. The degradation results in the release of
gases that remain dissolved in oil. The concentration of these dissolved gases has
direct relationship with incipient transformer faults. Persistence of these faults within
the transformers can lead to irreversible insulation damage. Hence transformers need
to be put out of service, and consequently customers have to face interruptions in
the power supply. The recognition of these faults in incipient stages can expand
the service life of transformers and decrease these unplanned interruptions. DGA-
based transformer condition monitoring is the widely accepted practice for reliable
and economic functioning of power system. Several DGA interpretation methods are
provided by experts to classify incipient faults. However, these interpretation methods
have significant limitations such as wrong and uncertain diagnoses. The inaccuracy
and uncertainty are mainly due to unidentified and missing codes, boundary problems
and presence of concurrent incipient faults or multiple faults. There is lot of scope
to address above-mentioned limitations of the DGA-based methods using different

S. A. Khan (B) · S. A. Wani

Department of Electrical Engineering, Jamia Millia Islamia (A Central University),
New Delhi 110025, India
e-mail: [email protected]
S. A. Wani
e-mail: [email protected]
M. A. Khan
Department of Physics, College of Science, King Khalid University, Abha 61421,
KSA, Saudi Arabia
e-mail: [email protected]
Md. D. Equbal
Department of Electrical Engineering, Galgotias College of Engineering & Technology,
Greater Noida 201306, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 287

H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_13
288 S. A. Khan et al.

soft-computing techniques. These techniques can improve the performance of DGA

methods when used alone or as a hybrid (more than one). Further, it is possible
to enhance the performance of DGA-based methods by combining more than one
soft-computing aided DGA methods, and by energy weighting of concentration of
fault gases. A comprehensive study of application of soft-computing techniques in
DGA-based transformer condition monitoring is presented in the article.

Keywords DGA · Transformer incipient fault · ANFIS · FIS · FIM · Energy

weighting · Boudary condition · Mutiple fault diagnosis

Nomenclature

DGA Dissolved gas analysis

ANN Artificial neural network
FIS Fuzzy inference system
ANFIS Adaptive neuro-fuzzy inference system
SVM Support vector machine
GEP Gene expression programming
EWDGA Energy weighted DGA
EWMEA Energy weighted Methane, Ethane, and Acetylene
FIM Fault interpretation matrix

1 Introduction

1.1 General Overview

Power transformers play a pivotal role in the transmission and distribution functions
of the power system. The health of the costly equipment is essential to provide
uninterrupted power supply to end users and thereby reduce the expenses of power
industry [1]. Their health can be directly monitored from the status of the insulation
system. Insulation health is directly associated with the presence of incipient faults.
These faults may turn catastrophic when proper attention is not given to them. During
the early days, insulation status was examined manually, and decision was made
based on human expertise [2, 3]. With the evolution of DGA, reliable identification
of incipient faults is possible. Using DGA, different incipient faults that can be
identified are broadly classified as partial discharges, thermal and electrical faults.
Partial discharges or corona lead to formation of air bubbles due to high voltage. They
can rupture the insulation locally. Thermal faults or pyrolysis are mainly developed
due to overheating of insulation. They result in the formation of hot-spots. Lastly,
Performance Enhancement and Extension of DGA-Based … 289

electrical faults or arcing arise due to short circuit between normally insulated points.
These faults can perforate solid insulation. The increasing severity order of these
faults is corona, pyrolysis and arcing.
Incipient faults generate various gases by weakening transformer insulation. Some
of the important gases released are hydrocarbon gases such as acetylene (C2 H2 ),
ethylene (C2 H4 ), ethane (C2 H6 ), and methane (CH4 ) and hydrogen (H2 ). Besides
carbon oxides (CO and CO2 ) and non-fault gases, oxygen and nitrogen (O2 and N2 )
are also released. C2 H2 and H2 are key gases released during arcing; C2 H4 , C2 H6,
CH4 are the key gases released due to pyrolysis while as due to corona mainly H2 is
released. The concentration of these key gases is correlated to the type of incipient
faults using various DGA-based interpretation methods.
The established conventional interpretation method is broadly divided into ratio-
based methods and graphical methods. Ratio-based DGA methods include Rogers’,
IEC and Doernenburg methods [4–6]. In these methods, different combinations of
key gas ratios are assigned in the ranges as per DGA standards (IEEE C57.104-2008
and IEC 60599-2007). Each range is given a code, and different code combinations
define the type of fault. In graphical methods such as Duval’s triangle, fault data is
traced on the triangular map by plotting the percentage concentration of key gases.
The map is divided into zones corresponding to a different type of incipient fault.
Thus the location of fault data on the map defines the fault. Although the conven-
tional methods have straightforward implementation and are economical, but suffer
from unavoidable shortcomings, such as wrong and unresolved diagnosis. These
need to be overcome for a reliable fault diagnosis [6]. The literature shows many
applications of various soft-computing techniques to address these limitations and
algorithms to facilitate the DGA-based diagnostic process. In this chapter, some of
recently reported very useful works on improving the incipient fault diagnosis by the
interpreting standards through the use of soft-computing techniques, namely, fuzzy,
ANN and ANFIS and others are summarised.

1.2 Literature Review and Research Gap Findings

The bulk of the research work in DGA-based fault diagnosis is related to handling
the existing limitations and improves the performance of DGA data interpretation
methods. Application of soft-computing techniques has immensely improved the
performance of DGA-based transformer monitoring and diagnosis systems. Fuzzy
logic [7–9], ANNs [10–13], SVM [14, 15], wavelet transforms [16], GEP [17] and
neuro-fuzzy systems [18] are vastly studied for application in fault diagnosis through
DGA. These methods boost the diagnostic capability of both single-handedly and
also when used along with optimisation algorithm [19]. In [20] ANFIS and fuzzy
variants of IEC, Rogers and Doernenburg methods are developed and compared.
For testing purpose, the public database is used [21]. Both techniques enhance the
diagnosis ability by obviating the limitations of conventional methods. However,
290 S. A. Khan et al.

ANFIS-based models show greater accuracy and can be used in the development of
reliable fault diagnosis system.
Researchers are also working on developing composite fault diagnosis methods
by integrating various conventional methods [10]. Some indigenous methods are also
developed where new gas ratios are added, or new combinations of gas ratios are
used in incipient fault classification [21, 22]. The work presented in [23] signifies the
boundary issues of Duval triangle method and their impact on the performance of the
method. The authors proposed merging of IEC and Duval’s triangle to address this
issue. The composite method enables Duval’s triangle to differentiate between normal
and faulty conditions as well. Before this, Duval triangle was never used for fault
detection but only for classification. The composite method uses diagnosis principles
of IEC method to identify “no-fault” condition and boundary faults. Duval triangle
rules are used for detection of dominant faults. Wani et al. [24] present the fault
diagnosis matrix (FIM), formulated using conventional DGA methods. IEC, Rogers
and Duval triangle methods are used in its formulation. FIM-based diagnosis system
is based on the principle of integrating conventional methods. The fault classification
is carried in two stages. In the first stage, fuzzy-based conventional methods provide
the fault classification. Then outputs of first stage are balanced so that these methods
can classify equal number of faults. The balanced output is fed to FIM which makes
the decision based on the sensitivity of the conventional method for each fault type.
FIM-based decision-making ensures the removal of contradiction in diagnosis at
lesser complexity. Results indicate the performance of the two-stage system is far
better than its constituent fuzzified conventional methods.
Accurate fault classification is important for monitoring the condition of trans-
formers with different types of faults. However, the fleet of transformers with a similar
type of fault priorities can be decided based on severity of the developed fault. The
severity of the fault is associated with the energy content of the fault gases generated
while fault is developed. In gas concentration-based DGA equal weightage is given
to all gases while classifying the faults whereas in Energy weighted Dissolved Gas
Analysis (EWDGA) gases are weighed according to their energy content so that they
can be used in fault classification and severity determination as well. The effect of
energy weighing on the performance of DGA-based diagnosis is studied in [25]. Two
ANN models are developed and compared, one is based on gas concentration-based
DGA, and the other one uses EWDGA. In [26] EWDGA is applied to Duval triangle
method to evaluate the severity of faults classified using this method. Duval trian-
gle requires only three inputs (Methane Ethylene and Acetylene) for classification
of incipient faults. Hence in Duval triangle Energy Weighted Methane Ethylene and
Acetylene (EWMEA) is used for severity evaluation. The comparison of unweighted
and weighted MEA combinations shows the impact of energy weighing on the sever-
ity of faults classified by this method. The work also presents the fuzzy-based severity
model for Duval triangle so that it can be simultaneously used in classification and
severity evaluation.
Incipient faults developing in transformers can either be single faults or multi-
faults. Conventional DGA interpretation methods are known to classify single faults
and cannot be used for multi-fault detection. This limitation, along with others, is
Performance Enhancement and Extension of DGA-Based … 291

responsible for their average diagnostic performance. Multiple fault detection by

up-gradation of conventional Duval triangle method is given in [27]. The classifica-
tion scheme uses diagnostic rules of Duval triangle for detecting single faults and
extends it for detection of multiple faults in case fault data point is plotted on bound-
aries. Optimal FIS and neural network configurations of the classification scheme
are developed. The performance analysis shows that both configurations have good
single and multiple recognition and isolation capabilities.
Not only the classification but also location of incipient faults is equally impor-
tant. It is because paper faults can have detrimental effects. Hence ANFIS model
using IEC-599 standard is developed for simultaneously classifying the fault and
identifying its location [28]. Thus, it helps in assessing the condition of liquid as
well as paper insulation of the transformer.
Although voluminous research is carried out to improve and extend the diagnos-
tic capabilities of conventional DGA method, there is still much scope to improve
and extend the capabilities of the methods using different intelligent techniques.
Furthermore, new methods for condition monitoring of oil-immersed transformers
are needed. Combined diagnostic techniques can provide useful tools to experts for
detecting transformer faults with higher accuracy. Merging various DGA techniques
to enhance the diagnostic reliability and capability should be a giant leap in new age
of DGA-based monitoring and maintenance of transformers.

2 Methodology

DGA is currently the most widely used tool in monitoring transformer condition
through incipient fault detection. DGA-based data interpretation standards and meth-
ods have certain limitations. The methodology of some of the recently reported very
interesting works to improve the performance and extend the capabilities of the DGA
data interpretation methods using different intelligent techniques is discussed in this
section.

2.1 Using Fuzzy Logic and ANFIS Models for DGA-Based

Transformer Fault Diagnosis

An adaptive neuro-fuzzy system (ANFIS) model is constructed using IEC599 stan-

dard for fault identification and location. Faults in the transformer can cause either the
deterioration of liquid insulation, paper insulation or both. So the developed ANFIS
model is trained to simultaneously identify and locate the incipient fault.
DGA-based condition monitoring is further implemented using fuzzy logic and
ANFIS soft-computing techniques. Fuzzy- and ANFIS-based implementation of IEC
599, Rogers’ ratio and Doernenburg’s methods are carried out to overcome their
292 S. A. Khan et al.

limitations. The comparative study is carried out to find the best model for fault
diagnosis.

2.1.1 Incipient Fault Identification and Location Using ANFIS-Based

Model [28]

ANFIS model is trained according to IEC-599 standard for incipient fault detection.
Hence inputs to the model are ratios X1 = C2 H2 /C2 H4 , X2 = CH4 /H2 and X3 =
C2 H4 /C2 H6 calculated using five key gases. The model can detect eight different
incipient faults and also distinguish between fault and normal operating conditions.
The output of the ANFIS model is binary. So for identifying fault, a separate model
is built for each type. Each ANFIS model uses the same inputs [X1 X2 X3]. The
output of each model can be either the fault type for which it is trained or no-fault
condition. The structure of ANFIS model for fault identification is given in Fig. 1.
The training of ANFIS network requires an appropriate dataset. So each model
is trained according to the input pattern for a fault given by IEC-559 standard. The
training dataset for each model contains 300 samples. The models are rigorously
trained and show the training error of less than 0.001.
Location of incipient fault is as important as its identification. The transformer
insulation system is comprised of liquid and solid insulation. Mineral oil is the most
common liquid insulation whereas cellulosic paper is the preferred solid insulation.
Deterioration of solid insulation is of greater concern because it is wound over the
transformer core, the high electric field area. The main by-products of deteriorat-
ing cellulose paper are carbon oxides. So timely assessment of paper insulation is
important as its degradation can have a detrimental impact.
Another ANFIS model is designed for locating the faults within the transformer
and consequently analysing the health of paper insulation. The model is also trained
according to the IEC-599 standard. The input used for fault location ANFIS model
is CO2 /CO. According to IEC standard, if this ratio is less than 3 or greater than 11,

Fig. 1 The ANFIS architecture for transformer incipient fault identification

Performance Enhancement and Extension of DGA-Based … 293

faults are likely existing in paper insulation. The output of the model can be either
1 indicating that faults are existing in paper insulation or 0 indicating that there is
no involvement of paper. A dataset of 200 samples is used to train the fault locating
ANFIS model, and after appropriate training, error is limited to 0.001.
ANFIS and fuzzy logic-based implementation of IEC 599 standard, Rogers
ratio method and Doernenburg’s method are carried out and their performance is
compared [20].

2.1.2 Comparative Study of Fuzzy and ANFIS Models for DGA-Based

Transformer Fault Diagnosis [20]

A comparative analysis of two soft-computing techniques, which are ANFIS and

fuzzy, is carried out after their application for DGA-based transformer monitor-
ing. First ANFIS and fuzzy counterparts of IEC 599 standard, Rogers ratio method
and Doernenburg’s method are developed. Then their performance is compared
based on their ability to overcome the limitation of conventional DGA methods.
For comparative analysis all models are tested with the same database.

DGA-Based Transformer Fault Diagnosis Using Fuzzy Logic

Fuzzy logic implementation of the conventional ratio-based DGA methods is car-

ried out for comparing it with ANFIS implementation of the same methods. Fuzzy
models of IEC 599 standard, Rogers’ ratio method and Doernenburg’s methods are
constructed. As all methods are ratio based, their fuzzy models use gas ratios as
inputs. In IEC standard-599 and Rogers’ ratio methods (X1 = C2 H2 /C2 H4 , X2 =
CH4 /H2 , X3 = C2 H4 /C2 H6 ) serve as inputs. In Doernenburg method, absolute con-
centration of key gases and ratios (X1 = CH4 /H2 , X2 = C2 H2 /C2 H4 , X3 = C2 H2 /CH4 ,
X4 = C2 H6 /C2 H2 ) serve as the inputs. The inputs of each method are fuzzified using
membership functions. The membership functions are comprised of linguistic vari-
ables 0, 1 and 2. These correspond to a range defined for each input in different
fault conditions. The rules are the combinations of linguistic variables of different
inputs analogous to code combinations in IEC method. The output variable of the
model has nine membership functions showing eight fault conditions and one no-fault
condition.
The schematic of the fuzzy logic model used for transformer incipient fault diag-
nosis is given in Fig. 2. The fuzzy models for all three methods consist of a rule base
defined by the code combination used to define fault. The codes are the crisp values
fuzzified using membership functions. The membership function of each code is the
predefined range.
294 S. A. Khan et al.

Fuzzy Inference System

Transformer Fault Fuzzy Membership Defining of Fuzzy Rule

Database covering all Functions of gas Ratios Base Using the Input Gas Defuzzyfication of FIS
Diagnosed Fault Type
fault types as defined by Standard Ratios– Fault Type output
Mapping

Fig. 2 FIS model for transformer incipient fault identification

ANFIS

Training of separate
Preparation of ANFIS Model Testing
Initialization of ANFIS Networks to Output in the form of
Training data for all Using the External
ANFIS Model collectively identify Diagnosed Fault Type
fault types Fault Database
all fault types

Database of Known
Fault Types

Fig. 3 ANFIS Model for transformer incipient fault identification

DGA-Based Transformer Fault Diagnosis Using ANFIS

ANFIS models need to be trained with appropriate fault data before it can be tested to
fault identification. Training must be done according to some standard that provides
guidelines for fault detection. Since ANFIS is a binary classifier, separate ANFIS
models are required to be designed for each type of incipient fault. Different ANFIS
models are trained according to IEC, Rogers’ ratio and Doernenberg ratio methods
for all incipient faults. ANFIS implementation of these methods is carried out to
overcome their shortcomings.
ANFIS implementation of the IEC standard requires the construction of eight
ANFIS models. This is because this standard can detect eight fault types. Each
model uses three inputs, and they are gas ratios (X1, X2 and X3) and output of each
model is either the fault type or no-fault. Each model is trained with the dataset of
300 samples and training is done according to diagnostic procedure of each fault
defined in IEC standard. After proper training of the model for appropriate epoch,
error is found to be less than 0.001. The testing of models shows that their learning
capability after rigorous training is such that ANFIS implementation can successfully
address limitations of its conventional counterpart. The schematic of the incipient
fault diagnosis of a transformer using ANFIS is given in Fig. 3.
Similarly, ANFIS models are designed for Rogers’ and Doenenburg ratio methods
to achieve better performing counterparts of conventional methods. The inputs to all
models are gas ratios and training is done using generated dataset, as per diagnostic
procedure defined in Rogers’ and Doenenburg ratio methods.
Performance Enhancement and Extension of DGA-Based … 295

2.2 Composite DGA Method for Diagnosing Incipient

Boundary and Dominant Faults [23]

Incipient boundary faults are one of the major concerns of accurate Duval triangle
methods. While plotting the fault data point on the triangular map of Duval, if a data
point is located inside the zones of the triangle, the fault is said to be an incipient
dominant fault. If the data point is located on the triangular or zonal boundaries, fault
is said to be incipient boundary fault. Duval triangle fairly accurate while detecting
dominant faults, however, its performance degrades in case of incipient boundary
fault. The issue becomes more critical in case fault data is located on the interface
of electrical and thermal faults. In order to improve the performance of user-friendly
Duval triangle method, an algorithm is proposed for amalgamation of Duval triangle
and IEC method. For the working of algorithm five inputs are required. Three inputs
of Duval triangle and two additional inputs (H2 and C2 H6 ) for evaluating IEC ratios.
The working of algorithm is such that initially IEC method differentiates between
faulty and normal operating conditions. In case transformer is faulty and data point
is located within fault zone, Duval triangle classifies the dominant fault. When data
point is located on boundary IEC method comes to rescue and classifies the fault.
The working of algorithm is given in flow chart as shown in Fig. 4.
For the implementation of Duval triangle to classify incipient faults, triangular
coordinates are converted to Cartesian coordinates using mathematical basics of
similar triangles. The Cartesian coordinates are used to obtain intercept form straight
line equation (a = mx + ny + c). In this equation a, c m, n represent the constant
values. The following steps are used for the classification of faults when software
implementation of the method is carried out.
1. Unsymmetrical polygon-shaped fault zones of Duval triangle are converted to
symmetrical sub-triangles.
2. Each vertex of sub-triangle has two coordinates. Using these equations are
obtained for each side. Hence three equations are used to model each sub-triangle.
3. When coordinates of fault data point (x and y) are substituted in these equations,
the value of ‘a’ (LHS of the given equation) for three equations of each sub-
triangle is calculated. This value can be either positive or negative.
4. The logic is such that for any sub-triangle if two values of ‘a’ are negative and
one is positive then the point is located in that sub-triangle. The zone to which
the sub-triangle belongs is the probable fault.
5. If the value of ‘a’ is zero for any equation, then boundary faults are present. For
the classification of these faults, diagnostic rules of IEC method are used.
However, it is important to note that practically the value of ‘a’ is never zero but
a very small value close to zero. It is known as a threshold value and this value is
calculated for each equation by back-calculation. The following steps are used in the
classification of boundary incipient faults by the IEC method.
If the fault data point is located on the boundary between DT zone and other fault
zones (D1, D2, T1, T2, T3 or PD and DT), then IEC rules for an adjacent zone are
296 S. A. Khan et al.

Start

Input Test Data

Determine the Gas

Ratios

Yes ‘NO FAULT’

Transformer Under
Condition
Normal Operation
Satisfied
?

No
End
Calculate the
Cartesian
Coordinates

Convert fault
zones in Duval
Triangle into
minimum Nos. of
Triangle

Generate
Threshold value of
ith side of the
triangle

Determine the
Compare with IEC Boundary Yes Is the Threshold
value satisfied by
gas ratio ranges condition of the ith
the coordinates?
Triangle

i=i+1
No
Indicate the No Is there any Yes Check the zone
Stop Transformer Fault Violation of IEC
Conditions
Type limits?

Yes Is Inequality No
Condition, a<0, b<0,
c>0, Satisfied ?

Fig. 4 Flowchart for extended Duval triangle method

verified. If the rules are satisfied adjacent zone is the probable fault otherwise DT is
the fault.
If fault data point lies on the boundary between fault zones excluding DT (D1-D2,
T1-T2, T2-T3, T1-PD), then gas ratios of two fault zones with similar ranges are
redundant, and gas ratios with un-identical ranges choose the type of incipient fault.
For example, in case of D1-D2, C2 H2 /C2 H4 and C2 H4 /C2 H6 are used in classification,
and CH4 /H2 is ignored because for both faults it shows the same range. In case of
T1-T2 and T2-T3, C2 H2 /C2 H4 and CH4 /H2 are redundant ratios and C2 H4 /C2 H6 is
used in classification. For T1-PD, all three ratios have dissimilar ranges and hence
C2 H4 /C2 H6 , CH4 /H2 ,and C2 H2 /C2 H4 are used in classification.
Consideration effect of energy weighing of concentration of dissolved fault gases
in DGA-based diagnostics is an important step towards accuracy and reliability. In the
Performance Enhancement and Extension of DGA-Based … 297

next section, the significance of energy weighting in fault diagnosis and measurement
of fault severity is investigated.

2.3 Energy Weighting of Fault Gases and Severity

of Diagnosed Fault [23, 25]

The conventional DGA methods are dependable incipient fault classification

schemes. However, they do not have any provision for predicting the severity of
these faults. It is because they give equal weightage to each gas during classifica-
tion. The severity of fault is particularly important if immediate corrective action
is to be taken for transformers with similar faults. Hence, to study the severity of
incipient faults EWDGA is used. In EWDGA, each fault gas is weighed according
to the energy content associated with the particular gas. To evaluate the severity of
faults diagnosed by Duval triangle, the concept of EWMEA is derived. It is this
method uses only three fault gases to determine the type of fault. These gases are
methane, ethane and acetylene. The comparison of energy weighted and unweighted
MEA shows the effect of energy weighing on Duval triangle-based diagnosis. Fur-
ther using EWMEA, FIS of Duval triangle is used for fault classification and severity
prediction simultaneously.

2.3.1 Estimation of Weighing Factors

Information regarding the energy content of the fault gases whose generation led
to the development of fault is required for severity estimation. This information is
obtained from thermodynamics behind the formation of each fault gas. Thermody-
namically, enthalpy of formation shows (H°) each fault gas is different and the
addition of this information would aid the severity estimation. Hence weighing fac-
tors for each fault gas are evaluated by normalising the enthalpies of their formation.
Table 1 shows fault gases, their H° and normalised (H°). The values of H° are
obtained from the model involving the decomposition of n-octane into various fault
gases. N-octane is chosen as a model compound because its properties are similar

Table 1 Standard and normalised enthalpies of fault gases

Transformer fault gas Enthalpy of Formation (Hf °) in Relative enthalpy (normalised
KJ/Mol values of Hf °)
C2 H2 278.3 1.00
H2 128.5 0.46
C2 H4 104.1 0.37
C2 H6 93.5 0.33
CH4 77.7 0.27
298 S. A. Khan et al.

to constituents of transformer oil whose decomposition would generate same gases.

The table shows that maximum energy is released during the formation of acetylene
and hence is given maximum weightage for estimating severity by EWDGA.
For estimating severity, thermodynamic information is to be added to each fault
gas. So, fault gases are multiplied by their normalised energy contents, hence nor-
malised H° act as weighing factors in EWDGA. These multiplication factors are
proportional to energy released during generation of fault gas. Hence, it is worth to
investigate the effect of energy weighing on DGA-based diagnostics.

2.3.2 Severity Model for Duval Triangle

The severity model for Duval triangle is obtained by limiting EWDGA to EWMEA.
Hence the model has five inputs, i.e. unweighted concentrations of CH4 , C2 H4 and
C2H2 for fault classification and weighted concentration of C2 H4 and C2 H2 for
severity prediction. The outputs of the model are fault type and simultaneous severity
as represented in Fig. 5.
In Duval triangle severity model, H° of the fault gases is normalised with respect
to CH4 (least H° among all), so weighing factors of both C2 H4 and C2 H2 are greater
than one. EWCH4 and CH4 both have the same values because the weighing fac-
tor of CH4 is one. Hence to avoid redundancy, EWC2 H4 and EWC2 H2 are used to
evaluate severity. The inputs of the severity model are fuzzified using trapezoidal
membership functions defined by linguistic variables. Each linguistic variable of
unweighted inputs represents the region inside Duval’s triangle. The linguistic vari-
ables of weighted input are symmetrically divided over the range of 0–100. The
two outputs of the model are the fault name and respective severity. The triangu-
lar membership functions are used for fuzzifying outputs. Eight linguistic variables
(D1, D2, DT, T1, T2, T3, PD, mf) are used in fuzzification of fault name and three
variables for fuzzyfying severity. mf and DT represent multiple faults and severity
is not calculated for them. 108 rules are used for generating FIS of severity model.

Input
CH4 Membership
Function

Input
C2H4 Membership
Function

Output Membership FAULT TYPE

Input
Fuzzy IF-Then Function
C2H2 Membership
Function Rules Output Membership SEVERITY
Function

Weighting Input OUTPUT of FIS

C2H4 Membership
Factor Function
EWC2H4

Weighting Input
C2H2 Factor Membership
Function
EWC2H2

Fig. 5 FIS based fault and severity model based on Duval triangle
Performance Enhancement and Extension of DGA-Based … 299

For classifying faults, diagnostic rules of Duval triangle are used, and for predicting
severity, rules are written in such a way that greater the concentration of acetylene
more severe is the fault.

2.4 Effect of Energy Weighting of Fault Gases

on DGA-Based Fault Diagnosis [26]

One major diagnostic attributes missing in gas concentration-based DGA is the infor-
mation regarding the energy released when these gases are generated. The thermody-
namic decomposition model shows that maximum energy is released when acetylene
is generated, and minimum energy is released when methane is generated. Other gases
have in-between energies. Since acetylene and hydrogen are key gases released dur-
ing electrical arcing faults, hence they are high energy faults. Methane and ethane
with lesser H° are the key gases in thermal faults. Hence such faults are low energy
faults. In the case of corona, only hydrogen is the key gas, these faults are also low
energy faults.
To demonstrate the effect of energy weighing on DGA diagnostics, of BackProp-
agation Artificial Neural Network (BP-ANN) is used. The inputs to the model are
suitably weighed fault gas concentrations and output is the fault type. The model is
trained using energy weighted fault gas concentration, and faults are diagnosed using
IEC-599standard. The diagnostic capability of the model is tested using the public
database of Duval and other literature. The model is compared with its unweighted
ANN counterpart to show the improvement in diagnosis due to energy weighing.
Figure 6 shows the flow diagram of diagnosis procedure of both the models.
There is a lot of published literature on the ANN-based implementation of DGA.
This work is carried out with the objective of comparing the diagnostic power of
DGA and EWDGA. For this individual models are constructed gas concentration-
based and energy weighted DGA in the neural domain. Each model uses five fault
gases (H2 , CH4 , C2 H6 , C2 H4 and C2 H2 ) as inputs and nine outputs which include
eight fault types and No-fault condition as well. The gas concentration-based ANN
model is multilayer network trained with BP algorithm. The model used sigmoid
activation function. A dataset of 250 samples corresponding to the output fault types
is generated for training the network. The optimal configuration of the network
showing least mean square error was obtained when the number of hidden layer
neurons was changed to 35. The testing of the model was carried out using published
database.
The ANN model for EWDGA was constructed with the same architecture as its
unweighted counterpart. The multilayer model is comprised of input, output and
hidden layers and uses sigmoid activation function. The inputs to the model are
weighted concentrations of five key gases. The weighed are obtained from the nor-
malised energy content of each gas. The nine outputs are same as that of unweighted
ANN. Based on hit and trial approach, number of neurons corresponding to least
300 S. A. Khan et al.

Gas Conc. of Known

fault types from Fault
EWANN Model Database

Development of
Training dataset ANN Model

Energy weighing Energy Development of

of gases conc. with Weighing Training dataset
relative weights of Gases

ANN model
EWANN model
EWANN Model ANN Model Training using Key
Training
Fault Gas Conc.

No Fault Condition
Y
N N
Fault Type Fault Type

Comparison of Fault Type with

fault condition diagnosed by IEC-
599 standard

Fault
Diagnosis

Fig. 6 The ANN- and EWANN-based transformer fault diagnosis process

mean square error are 40. The training data is the weighted counterpart of dataset
used for training unweighted ANN. The performance of the EWDA ANN model was
determined by using same testing database as used in ANN for gas concentration-
based DGA. Multiple incipient fault detection is one of the pertinent issues in DGA
base transformer incipient fault detection. Conventional methods may be extended
to address the issue.

2.5 Multiple Incipient Fault Detection Using DGA [27]

Multiple faults are concurrently existing faults in transformers. These faults can lead
to changing endpoints of defined range for any ratio or oscillating location of fault
data point at the interface region of fault zones in ratio-based and graphical methods,
respectively. As per known thermodynamic information, the energy content of such
faults would lie between two single faults of low and high severity. Therefore, such
faults need to be looked after to improve the reliability of transformer monitoring.
The established DGA interpretation techniques have no provision for detecting these
concurrent faults, and hence their performance is low when the fault lies on the
boundary or multiple faults occur. In this work, the objective of improvement in
Performance Enhancement and Extension of DGA-Based … 301

diagnostic performance of conventional methods through multiple fault detection is

accomplished. Multiple fault detection by extended version of Duval triangle when
fault data lies in DT zone or on the boundary between two fault zones is achieved.
Extension of Duval triangle for detecting multi-fault cases is significant move as
Duval triangle shows higher diagnostic capability when used for classifying single
faults. Hence extended Duval triangle can prove to be a potential low complexity
interpretation technique of DGA-based single and multi-fault diagnosis. Some of the
objectives of high-performance DGA-based transformer monitoring achieved in this
work are as follows:
• Addressing boundary issues of conventional methods by correlating them with
multiple incipient faults.
• Analysing Duval triangle for both multiple and single fault detection and
classification.
• Optimizsed fuzzy and neural versions of extended Duval triangle for industrial
applications.

2.5.1 Diagnosis of Multiple Faults Using Duval Triangle

DGA literature shows that multiple faults and boundary issues are interrelated. Hence,
addressing boundary issues by detection of multiple faults can not only improve
the diagnosis capability of conventional DGA interpretation methods but also add
multiple fault information feature to them. Among all conventional methods, Duval
triangle is the better performing method; therefore, it is a suitable choice for upgrading
to multi-fault detection.
Conventional Duval triangle has seven fault zones. D1, D2, T1, T2, T3, PD cor-
respond to dominant incipient faults existing in transformers. Seventh is a mixture
of fault zone or DT. It separates electrical and thermal faults. So the location of data
point in this zone or on the boundary of this zone is the result of concurrent electrical
and thermal fault. However, interface that separates electrical faults from each other
is D1-D2 boundary. Along D1-D2 boundary, CH4 and C2 H2 concentrations vary, but
C2 H4 concentration is constant at 23%.
Similarly, interface separating thermal faults among each other is T1-T2 and T2-
T3 boundaries. Along these boundaries, again CH4 and C2 H2 concentrations vary
but C2 H4 concentration is constant at 20% and 50%, respectively. Another boundary
between electrical and thermal fault is PD-T1. Along this boundary C2 H4 and C2 H2
vary and CH4 is constant at 98%. So if fault data point is plotted on these boundaries
or DT zone, multiple faults can be detected and classified according to the name of
the boundary on which it is located. This concept is used for extending conventional
Duval triangle for multi-fault diagnosis. Although conventional version of Duval
triangle can detect multiple faults through boundary condition or DT zone but its
extended counterpart can differentiate between concurrent faults and classify them
as well. The concentration range of gases for diagnosing multiple faults by extended
Duval’s triangle is given in Table 2.
302 S. A. Khan et al.

Table 2 Fault gas concentration limits for various multiple faults diagnosed by extended Duval
triangle
Fault combinations %CH4 %C2 H4 %C2 H2
D1-D2 0 < CH4 < 64 C2 H4 = 23 13 < C2 H2 < 80
T1-T2 46 < CH4 < 80 C2 H4 = 20 0 < C2 H2 < 4
76 < CH4 < 98 C2 H4 = 20 0 < C2 H2 < 4
T2-T3 0 < CH4 < 50 C2 H4 = 50 0 < C2 H2 < 15
T1-PD CH4 = 98 0 < C2 H4 < 2 0 < C2 H2 < 2
DT 0 < CH4 < 50 40 < C2 H4 < 100 15 < C2 H2 < 29
37 < CH4 < 96 0 < C2 H4 < 50 4 < C2 H2 < 13
37 < CH4 < 57 40 < C2 H4 < 50 13 < C2 H2 < 15
31 < CH4 < 57 40 < C2 H4 < 57 15 < C2 H2 < 29

Table 2 shows the range of key gases that are varying along boundaries in extended
Duval triangle is the same as that of dominant faults. For example, in the case of D1-
D2 fault, CH4 and C2 H2 have same concentration range as dominant fault D1 and D2
in conventional triangle. Only, the range of C2 H4 is fixed which would be otherwise
varying in the conventional version of Duval triangle. Also, since boundary adjacent
to DT zone and dominant fault zone is treated as DT fault only to avoid redundancy.
The flow chart showing the implementation methodology of the proposed scheme
is shown in Fig. 7. For multi-fault detection, the extended counterpart of Duval
triangle is used. It provides a comprehensive diagnostic strategy for detecting single
and multi-faults. Since input dimensions of conventional and extended Duval triangle
are the same, lesser complexity and better performance are some of the features of this
scheme. The implementation of this scheme is carried in fuzzy and neural domains
using MATLAB.
FIS architecture of extended Duval triangle is constructed using the procedure
summarised in Table 2. Inputs to the FIS are key gases of conventional Dual triangle.
They are fuzzified using trapezoidal membership functions. Membership function of
each input is defined by eleven linguistic variables. The range of these variables is
obtained from the symmetrical triangular regions traced by all inputs within the asym-
metrical fault zone. The triangular membership function is used for output variable.
Output membership functions are defined by nine linguistic variables corresponding
to single and multiple incipient fault types. FIS detects single faults if data point is
located well within fault zone and multiple faults if located on boundary or in DT
zone.
FIS for extended Duval triangle maps input key gases into output single-multiple
fault using a set of 350 rules. For detecting single faults, 206 rules are required, and
for multi-faults, 144 are required. The implication operator is AND as all the inputs
are important for diagnosing faults. Equal weightage is assigned to all rules as all
contribute equally in diagnosing the fault. So operation of rules is commutative and
OR is used as aggregation operator.
Performance Enhancement and Extension of DGA-Based … 303

Development of Dissolved Gas

Fuzzy Inference Analysis Feature ANN Training
System Extraction

Cross Validation of
Developed Soft
computing Models

Testing Optimal FIS Optimal Testing

Data Model ANN Model Data

Single Fault
Exists

Yes

Is the Fault
Point Inside
Zone?

No

Fault Point Lies on

the Boundary

Multiple Fault
Condition

Fig. 7 Flow chart for multiple fault detection

In the neural domain, the architecture of extended Duval triangle is comprised of

a multilayer network. It has input layer with sigmoid transfer function, output layer
with linear transfer function and in-between hidden layer. The network is trained
using Levenberg–Marquardt algorithm according to diagnostic principle of extended
Duval triangle shown in Table 2. Training is such that single faults are diagnosed
when fault data is located inside any zone (excluding DT), and multiple faults are
diagnosed if it lies on boundary or DT zone. The input vector [PD D1 D2 DT LTF
HTF] of the ANN network is comprised of percentage of three key gas of Duval
triangle output vector is comprised of six elements corresponding to zones of Duval
triangle. If any element of output vector excluding DT is one, ANN diagnosis single
fault and if two elements or DT is one, multiple faults are diagnosed.
For constructing ANN architecture, training and testing databank is required. The
utilised databank consists of 430 samples, and the ratio of 60:40 is maintained for
training and testing purpose. So 259 are used in training the network, and 171 are
304 S. A. Khan et al.

used for testing purpose. The entire training database consists of simulated data, and
testing data is an amalgamation of simulated and literature acquired samples. 122
out of 171 are obtained from the literature.

2.5.2 F1 Scoring and Fault Confusion Matrix

The performance of these fuzzy and neural models is analysed using F1 scores. It is
equal to twice the parallel combination of sensitivity and accuracy. It is evaluated by
constructing a confusion matrix. The confusion matrix shows positive and negative
diagnosis cases of binary classifier. For example, if output of model containing ther-
mal and electrical faults is considered, and confusion matrix is obtained for same.
The matrix shows correctly diagnosed thermal faults and electrical faults and also
gives the details of misdiagnosed faults and missed diagnosis of model. If any ther-
mal fault is predicted as electrical, it is considered as misdiagnosed electrical fault
and missed diagnosis for thermal fault. Thus matrix gives us idea how the model is
confused while diagnosing faults. F1 scores of both FIS and ANN implementation
of extended Duval triangle are obtained to analyse their learning capabilities in case
of single and dominant faults as well. A smart diagnosis method is discussed in the
next section.

2.6 Smart Diagnosis with Fault Interpretation Matrix (FIM)

[24]

A reliable DGA diagnosis is essential for proper maintenance of transformers. For

this, a smart diagnosis system is proposed, which combines the advantages of conven-
tional DGA methods and removes the contradictions if any. The proposed integrated
system has two stages. They are performance enhancement and decision-making
stages. The intermediate between the two is the normalisation stage. Performance
enhancement of three conventional methods Rogers’ ratio, IEC and Duval trian-
gle is carried using fuzzy implementation. Normalisation is carried out to balance
the output of FIS of three methods. Decision-making is carried by FIM. The main
advantage of the proposed integrated system is that it improves the resolution of fault
classification by classifying faults based on their origin and severity rating as well.
Above all, this improvement in resolution is achieved at lower complexity as only
three methods are used.
The methodology of the smart diagnosis system is depicted through the block
diagram of the developed system shown in Fig. 8. The figure shows that output of
FIS of Rogers’ ratio, IEC and Duval triangle is input of the normalisation stage. This
stage is required as each of considered methods classify different number of faults.
To integrate them, the outputs need to be balanced. That is, each method should
classify same number of faults before FIM can decide the probable fault.
Performance Enhancement and Extension of DGA-Based … 305

Fuzzy Models of Normalization of

Rogers Ratio, IEC Fault Types of each Final Decision
FIM Based Diagnosis
and Duval Triangle Method into Same on Fault Type
Methods Number of Faults

Fig. 8 Block diagram of the FIM-based system

2.6.1 Fuzzy Logic Implementation of DGA-Based Fault Diagnosis

FIS of the Rogers’ ratio and IEC method uses key gas ratios as input while the Duval
triangle uses percentage of relative concentration of key gases as input. The input
of each FIS is fuzzified using trapezoidal membership functions. The rule base of
Rogers’, IEC and Duval comprises 6, 11 and 106 rules, respectively, to map inputs
to output incipient fault type. The rules of Rogers and IEC method are obtained
from the code combinations used in the methods. For Duval triangle according to its
diagnosis procedure utilised during manual plotting of fault data point. The output
space of Rogers’, IEC and Duval’s method comprise linguistic variables defined by
triangular membership functions. To integrate these methods, the output stage needs
to be balanced. So incipient faults diagnosed by the integrated method are categorized
as low and high temperature thermal faults, low and high energy electrical faults so
that their severity can also be predicted. Besides, there are partial discharges and
undiagnosed faults also. Also it is important to note that fault detection is carried
only by Rogers’ and IEC ratio methods whereas Duval triangle is used only for
classification. Hence output space of Duval triangle only shows linguistic variables
for the type of fault and ‘no-fault’ linguistic variable is absent.

2.6.2 Performance Evaluation of Fuzzy Models

The performance of fuzzy models of conventional DGA methods is evaluated on the

basis of accuracy, sensitivity, consistency and reliability. A fault databank comprising
515 samples is obtained from DGA literature. The accuracy of conventional Rogers’,
IEC and Duval methods are 55.10%, 58.25% and 76.09%, respectively. Their fuzzy
counterparts show the accuracies of 71.8%, 72.4% and 81.08%, respectively. The
consistencies of conventional Rogers’, IEC and Duval methods are 49.49%, 53.86%
and 71.75%, respectively. Their fuzzy counterparts show the consistencies of 68.06%,
69.42% and 76.66%, respectively. Similarly, reliabilities of conventional Rogers’,
IEC and Duval methods are 52.29%, 56.05% and 73.92%, respectively. Their fuzzy
counterparts show the reliabilities of 69.57%, 70.91% and 78.87%, respectively. The
performance analysis shows that fuzzy logic implementation has improved all the
performance indices of Rogers’, IEC and Duval methods. Also, Duval triangle is the
most accurate, consistent and reliable of the three.
306 S. A. Khan et al.

2.6.3 Formulation and Working of FIM

FIM is the decision-making stage of the integrated model. It is formulated after

analysing sensitivity performance index of the fuzzy models of Rogers’, IEC and
Duval methods. The testing dataset contains 515 fault cases comprising 34 No-fault,
74 LEF, 116 HEF, 106 LTF, 131 HTF and 54 PD fault cases. The sensitivity analysis
shows that
1. Both Fuzzy Rogers’ and Fuzzy IEC correctly predict 20 out of 34 ‘No fault
cases’. This makes their sensitivity for ‘No fault’ conditions as 58.82%. Duval
triangle does not have provision for detecting No-fault.
2. Out of 74 ‘LEF’ cases, Fuzzy IEC correctly predicted 45 cases and Fuzzy Duval
51 cases. The sensitivities of Fuzzy IEC and Fuzzy Duval are 60.8% and 68.9%,
respectively. Rogers’ method does not have a provision for detecting LEF.
3. Considered fault dataset has 116 ‘HEF’ cases. Fuzzy Rogers’ correctly diagnoses
85 of them, Fuzzy IEC diagnoses 79 correctly, and Fuzzy Duval shows 103 correct
diagnoses. This makes their respective sensitivities 73.2, 68.1 and 88.7%.
4. For 106 LTF cases, Fuzzy Rogers’, Fuzzy IEC and Fuzzy Duval, respectively,
diagnose 71, 89 and 90 cases. Hence, their respective sensitivities are 66.98, 83.9
and 84.9%.
5. 131 HTF cases are included in the dataset. Out of 131, 110 are correctly diagnosed
by Fuzzy Rogers’, 105 correct diagnoses for HTF are shown by Fuzzy IEC, 119
cases are correctly diagnosed by Fuzzy Duval. This makes the sensitivity of
Fuzzy Rogers’ 83.9%, Fuzzy IEC 80.1% and Fuzzy Duval 90.83%.
6. Finally, 54 PD fault cases are also present in the dataset. Fuzzy Rogers’, Fuzzy
IEC and Fuzzy Duval, respectively, diagnose 31, 35 and 27 of them correctly.
Hence respective sensitivities are 57.4, 64.81 and 50%.
From points 1–6, it can be concluded that
• ‘Rogers and IEC’ methods can be used in No-fault condition detection.
• ‘Duval and IEC’ methods can be used in LEF detection.
• ‘Rogers and Duval’ methods can be used in HEF detection.
• ‘Duval and IEC’ methods are better alternatives for LTF detection.
• ‘Rogers and Duval’ methods are better alternatives for HTF detection.
• ‘Rogers and IEC’ methods can be used in PD fault detection.
FIM consists of rows that correspond to the type of incipient faults and columns
which correspond to the DGA method used as shown in Table 3. 1’s represent the
priority assigned to each method after sensitivity analysis, for diagnosing a particular
fault. Don’t care condition is shown by ×. These elements do not participate in
decision-making.It is important to note that only after fuzzy implementation structure
of FIM is constructed.
Performance Enhancement and Extension of DGA-Based … 307

Table 3 Integration of DGA methods using FIM

Rogers ratio method IEC method Duval triangle method
LEF × 1 1
HEF 1 × 1
LTF × 1 1
HTF 1 × 1
PD 1 1 ×
NF 1 1 ×

The workflow of FIM is given in flow chart in Fig. 9. Initially, the absolute
concentration of five key gases is given as input to the proposed model. Then input
ratios of fuzzy models of Rogers and IEC methods; relative % concentrations for
fuzzy Duval are calculated. Each of the fuzzy models maps these inputs into output
fault type along with performance enhancement function. The balanced output from

Start

Key Gas
Concentrations (H2,
CH4, C2H4, C2H6, C2H2)

Determine Gas ratios and

Relative Gas Concentrations

Fuzzy Models

Rogers IEC Duval

Ratio Method Triangle
Method Method

FIS Models Fault Vector

Yes Does the Fault Vector No

Have Same Fault Types
?

Fault having the maximum

All Faults are Two of the Faults sensitivity is taken as the most
identical are identical likely fault type

Repeatedly occurring Yes No

Are Both Faults Find their Priority Fault with higher sensitivity is
Fault type is the
at high priority levels the probable fault Type
prevailing fault condition
level?
FIM

Fig. 9 Working of FIM

308 S. A. Khan et al.

these methods is used as input to FIM. [Y Rogers , Y IEC , Y Duval ]T is the input vector of
FIM. It heuristically chooses the best among three.
For decision-making, FIM checks two conditions, repeatability and priority of
input element. There are seven possibilities using these conditions: all three are
similar, two are similar, or none is similar and all high priority, two with higher
priority or only one with high priority. If all are similar, then fault is repeating
element. If two are similar and at high priority, then the fault is repeating element. If
two similar elements have different priorities, then sensitivity comparison is made
with third element and decision is made based on sensitivity. The last possibility is
that there is no repeating element, so decision is made by comparing sensitivity of
all three elements.

3 Results and Discussions

Results of the methods discussed in the previous section are explained and analysed
in this section.

3.1 Performance Analysis of ANFIS Model for Identification

and Location of Incipient Faults of Transformer [28]

The identification and location of faults through DGA-based ANFIS model is verified
using the reported database. The model explained in Sect. 2.1.1 is implemented using
MATLAB. The fault data of ten transformers shows the type and location of the fault
are given in Table 4. The fault existing in these transformers and its location is
obtained using IEC-599 standard.
ANFIS results of same fault data are shown in Table 5. The table shows that
ANFIS model correctly diagnoses and locates the fault in all nine transformers.
ANFIS model uses diagnosis procedure of IEC-599 standard for diagnosing and
locating the fault. The output vector of ANFIS model has ten elements. The first
element helps to differentiate between normal operation and fault transformers. The
last element of the vector helps to locate the faults. F1 to F8 positions represent
different incipient faults. The element or elements of high values denote fault and its
location.
For example, visual inspection shows the sixth transformer is faulty with electrical
discharges in paper. IEC standard correctly diagnosis the fault with involvement of
paper. ANFIS model constructed using diagnosis procedure of IEC standard also
shows correct diagnosis and location. After visually inspecting third transformer,
hot spot in paper insulation was found. However, using IEC standard CO2 /CO ratio
Table 4 Testing data for ten faulty transformers
Transformer fault sl. no. Key gas concentration (in ppm) Fault identified by Fault Located in Paper
H2 CH4 C2 H2 C2 H4 C2 H6 CO CO2 inspection Insulation (By Inspection)

1 7600 1230 1560 836 318 4970 4080 Discharges of low energy Yes
2 1450 940 61 322 211 2420 3560 Thermal fault < 700 °C Yes
3 960 4000 6 1560 1290 15800 50300 Thermal fault < 700 °C Yes
4 24700 61000 1560 42100 26300 14400 30400 Thermal fault < 700 °C Yes
5 90 28 32 31 8 1380 11700 Discharges of high energy No
6 210 22 7 6 6 19 74 Discharges of low energy No
Performance Enhancement and Extension of DGA-Based …

7 48 610 – 10 29 1900 970 Thermal fault < 700 °C Yes

8 1230 163 692 233 27 130 115 Discharges of low energy Yes
9 480 1075 – 1132 298 464 1000 Thermal fault < 700 °C Yes
10 12705 23498 5188 34257 6047 4004 8539 Thermal fault > 700 °C Yes
309
310 S. A. Khan et al.

Table 5 Result of ANFIS model using IEC standard

Transformer As per IEC-599 ANFIS output
fault Sl. No. Standard
Fault Fault Output fault vector Fault Fault located
type location in type in paper
paper insulation
insulation
1 F3 Yes [0.0079 0.0020 F3 Yes
0.1692 0.4624
0.13960.1781 0.0016
0.0130 0.0086
1.0010]
2 F5 Yes [0.1429 0.0387 F5 Yes
0.0579 0.1266
0.0279 0.51370.0082
0.0530 0.0250
0.9750]
3 F7 No [0.0952 −0.0209 F7 No
0.0250 −0.0434 −
0.0053 0.1864
0.1397 0.5631
0.0428 0.0668]
4 F7 Yes [0.1089 0.0168 F7 Yes
0.0381 0.0513
0.0026 0.4046
0.0347 0.5676
0.0707 1.0010]
5 F4 No [−0.0204 0.0067 − F4 No
0.0276 0.3738
0.54330.299−0.0014
0.0321 0.0549
0.0196]
6 F2 No [0.1065 0.0755 F2 No
0.9299 0.1430 −
0.0035 0.14050.0061
0.0109 0.0018
0.0338]
7 F6 Yes [0.0251 0.2014 F6 Yes
0.0032 0.1153
0.0028 0.0148
0.8568 0.3725 −
0.1815 1.0010]
(continued)
Performance Enhancement and Extension of DGA-Based … 311

Table 5 (continued)
Transformer As per IEC-599 ANFIS output
fault Sl. No. Standard
Fault Fault Output fault vector Fault Fault located
type location in type in paper
paper insulation
insulation
8 F3 Yes [0.0027 −0.0030 F3 Yes
0.0639 0.5277
0.2386 −0.0298 −
0.0015 −0.0037
0.0270 1.0010]
9 F8 Yes [−0.0368 0.0004 − F8 Yes
0.0053 0.1385
0.0441 0.3497 −
0.0259 0.1854
0.5961 1.0010]
10 F8 Yes [−0.0824 −0.0044 F8 Yes
−0.0014 0.1407
0.0371 0.1867 −
0.0626 0.1328
0.4576 0.9750]

exceeds 3, so the involvement of paper is ruled out. ANFIS method also produces
output in agreement with IEC standard. So proper training with appropriate database
is required to improve the potential of ANFIS-based DGA and remove diagnostic
discrepancies if any.

3.2 Comparison of Performance of Fuzzy Logic and ANFIS

Models [20]

The diagnostic abilities of fuzzy and ANFIS counterparts of Rogers’, IEC and Doer-
nenburg ratio methods are evaluated using a database of 100 fault samples from
reported literature. The ANFIS and fuzzy-based implementations to obviate the
shortcomings of conventional versions of these methods are explained in Sect. 2.1.2.
Table 6 shows the fault gas concentration of five transformers.
Table 7 shows the diagnostic comparison of fuzzy and ANFIS models for five
faulty transformers. The output of fuzzy and ANFIS models is compared with the
diagnostic results of IEC standard. The output of ANFIS model is the vector whose
elements correspond to various types of incipient fault and no-fault condition. The
output is decided by the position of element with the highest value (close to 1).
The output of fuzzy model is the defuzzified value. The output fault is predicted by
comparing the value with range assigned to fault type in the output space of fuzzy
312 S. A. Khan et al.

Table 6 Gas concentrations of faulty transformers

Transformer fault Sl. No. Key gas concentration (in ppm)
H2 CH4 C2 H2 C2 H4 C2 H6 CO CO2
1 1270 3450 8 1390 520 483 44500
2 60 10 4 4 4 780 7600
3 420 250 800 530 41 300 751
4 1450 940 61 322 211 2420 3560
5 400 940 24 820 210 390 1700

Table 7 Diagnosis by Fuzzy and ANFIS models using IEC standard

Transformer fault Fault as per IEC Fuzzy output ANFIS output
Sl. No. method Output fault vector Fault type
1 F5 F5 [0.0084 0.0047 F5
0.0138 0.0904
0.0117 0.8571
0.0121 0.2007
0.2037]
2 F3 F5 [0.1254 0.0211 F3
0.1529 0.5458 −
0.0013 0.1581
0.0068 0.0155
0.0026]
3 F4 F4 [0.0171 −0.0051 F4
0.0269 0.2546
0.9386 −0.2087
0.0113 −0.0229
0.1976]
4 F5 F3 [0.1429 0.0387 F5
0.0579 0.1266
0.0279 0.5137
0.0082 0.0530
0.0250]
5 F8 F8 [0.0005 −0.0014 − F8
0.0025 0.0001 −
0.1506 0.0998
0.2169 −0.2844
1.1494]

model. The table shows that ANFIS model has correctly diagnosed faults of all
five transformers. The fuzzy model predicted the faults of second and fourth trans-
former incorrectly. Transformer 2 is diagnosed with low energy arcing fault (F3) by
IEC standard. The output of ANFIS model is in agreement with the standard but
Performance Enhancement and Extension of DGA-Based … 313

Table 8 Results of ANFIS and fuzzy models based on DGA methods

IEC Method Roger’s Ratio Doernenburg
Method Ratio Method
ANFIS Fuzzy ANFIS Fuzzy ANFIS Fuzzy
No. of Faulty cases tested 100 100 100 100 100 100
No. of correct diagnosis 98 95 97 84 99 95
No. of Incorrect diagnosis/no 02 05 03 16 01 05
diagnosis
Accuracy (%) 98 95 97 84 99 95

fuzzy model misdiagnoses it as low-temperature thermal fault (F5). Similarly trans-

former 4 is diagnosed with low energy discharge (F3) by fuzzy model, but diagnosis
is expected to be the low-temperature thermal fault.
The results of entire testing data are summarised in Table 8 which indicates the
superiority of ANFIS-based diagnosis.

3.3 Testing of Composite DGA Method to Resolve Boundary

Issue in Duval Triangle [23, 25]

A composite model comprising Duval triangle and IEC standard is used to resolve
the boundary issue in Duval triangle method. The complete methodology of the
proposed model is discussed in Sects. 2.2 and 2.3. The testing of this model is
carried out using dataset of 364 faults. The fault dataset is obtained by combining
the fault samples from literature and field. Since the composite method is developed
to address boundary issues in Duval triangle, fault data lying at boundaries is also
included. The details of fault identification by composite method and its comparison
with the constituent methods (IEC and Duval) are given in Table 9. It shows the
performance of each method in detecting the dominant fault.
Table 10 shows the comparison of the composite method and conventional in
detecting the dominant fault and boundary fault. Dataset of 364 faults comprises 343
dominant faults and 21 boundary faults. The composite method and Duval triangle
method can resolve same number of dominant faults as both use the same rules
for detecting these faults. Boundary faults are detected by IEC and Duval triangle is
unable to detect these. Since composite method uses rules of IEC method for detecting
boundary faults, it can also detect such faults. Hence performance of composite
method is better than its constituent methods.
FIS of the severity model based on Duval triangle can diagnose the fault and
determine its severity. The severity of diagnosed faults is predicted on the basis of
energy weighting of fault gases. The results of fuzzy-based severity model are shown
in Table 11. The inputs to the FIS are unweighted concentration of methane, ethylene
314 S. A. Khan et al.

Table 9 Performance comparison of the composite method with conventional methods

Fault type Total IEC method Duval triangle Composite method
no. of method
fault Correct Accuracy Correct Accuracy Correct Accuracy
cases diagnosis (%) diagnosis (%) diagnosis (%)
Partial 13 6 46.1 12 92.3 12 92.3
discharges
Discharges 52 33 63.4 43 82.6 46 88.4
of low
energy
Discharges 63 56 88.8 58 92 60 95.2
of high
energy
Thermal 39 15 38.4 28 71.7 29 74.3
faults < 700
°C
Thermal 45 33 73.3 42 93.3 42 93.3
faults > 700
°C

Table 10 Summarized comparison of composite method and the conventional method

IEC method Duval Triangle method Composite method
Dominant faults 343 217 320 320
Boundary fault points 21 21 0 21
Total no. of fault cases 364 238 320 341
Accuracy (%) 65.38 87.9 93.6

and acetylene along with the weighted concentration of ethylene and acetylene. FIS
generates two defuzzified numbers at the output for fault type and respective severity.
The values are interpreted from the output space of FIS.

3.4 Performance Comparison of Energy Weighted ANN DGA

Model [26]

To study the effect of energy weighting on DGA-based diagnostics, the performance

of IEC-599-based ANNs are compared when these networks are trained with energy
weighted and non-weighted fault data. The methodology is explained in Sect. 2.4.
Both the models are tested with 100 pre-known fault samples. The output obtained
from both the models is compared with pre-known fault for the sample. The fault
type corresponding to each sample of testing data is obtained according to the IEC
standard to ensure reliability in performance analysis. A test case of fault samples
Table 11 Fault diagnosis and severity estimation results of the fuzzy model
Unweighted Weighted Defuzzified output Result
CH4 C2 H4 C2 H2 EWC2 H4 EWC2 H2 Fault Severity Actual fault Estimated fault Severity
32.8 21.3 45.9 12.7 72.8 1 4.5 D1 D1 Medium
7.73 26.4 65.8 49.2 49.2 3 7.5 D2 D2 High
52.9 47 0.06 54.24 0.18 12 1.5 T2 T2 Low
10.3 89.5 0.22 91.5 0.61 14 7.5 T3 T3 High
99.08 0.45 0.455 0.60 1.60 7.5 1.5 PD PD Low
Performance Enhancement and Extension of DGA-Based …

16.23 34.55 49.21 19.39 73.80 3 4.5 D2 D2 Medium

95.59 2.73 1.66 3.47 5.67 10 1.5 T1 T1 Low
315
316 S. A. Khan et al.

Table 12 Transformer DGA data and fault type

Transformer Key gas concentration (in ppm) Known
fault Sl. No. H2 CH4 C2 H2 C2 H4 C2 H6 CO CO2 fault
type
1 6454 2313 6432 2159 121 3628 225 F4
2 305 100 541 161 33 440 3700 F3
3 1230 163 692 233 27 130 115 F4
4 33046 619 – 2 58 51 1 F1
5 796 999 31 1599 234 389 1334 F8
6 34 21 56 49 4 95 315 F4
7 960 4000 6 1560 1290 15800 50300 F7
8 6 2990 67 26076 29990 6 26 F6
9 2500 10500 6 13500 4790 530 2310 F7
10 300 700 36 1700 280 760 9250 F8
11 37800 1740 8 8 249 56 197 F2
12 1450 940 61 322 211 2420 3560 F5

from 12 transformers along with the key gas concentration is given in Table 12. The
fault type is pre-known as diagnosed by IEC standard.
The results of ANN model for fault diagnosis are trained according to IEC standard
for 12 faulty cases are shown in Table 13. The output vector of ANN has nine
elements F0 to F8. They represent eight fault conditions and one no-fault condition.
For predicting the fault using ANN, the element with the highest value is found.
Then the position of this element corresponds to fault type. The table shows that ANN
model makes seven correct diagnoses and five incorrect diagnoses for the considered
test data. In the first transformer existing fault is high energy density discharge but
ANN predicts it as low energy discharge. Position of F3 in vector shows low energy
discharge and position of F4 in vector shows high energy discharge.
Table 14 shows the results of the energy weighted counterpart of ANN model
for the same 12 fault cases. From the table, it can be seen that performance energy
weighted model is better as compared to unweighted ANN model. It makes only
two wrong diagnoses. The pattern of output vector is similar to expected pattern of
IEC standard for 10 out of 12 faults. Transformers 6 and 9 high energy discharges
and high temperature thermal faults are wrongly diagnosed as low energy discharges
and medium temperature thermal fault by energy weighted model. The summarised
results of gas concentration ANN model and energy weighed ANN model are given
in Table 15. For the test database of 100 faults, performance accuracy of ANN-based
DGA is improved from 53–86% when inputs are energy weighed. The comparative
study was carried with the purpose of investigating the effect of energy weighting
on DGA diagnostics.
Performance Enhancement and Extension of DGA-Based … 317

Table 13 Results of ANN model without energy weighing

Transformer Known fault type Output fault vector Diagnosed fault type
fault Sl. No.
1 F4 [−0.0404 0.08271 −0.0526 0.9734 F3
−0.1149 −0.1658 −0.0359 −
0.0925 0.38773]
2 F3 [−0.0762 0.0226 −0.0713 0.6737 F3
0.5940 −0.0615 0.1426 −0.1509 −
0.0498]
3 F4 [0.0427 0.0171 0.0048 0.6568 F4
0.7702 −0.0932 0.0198 −0.1627
−0.0694]
4 F1 [0.0164 0.7246 0.6362 −0.0024 − F1
0.0124 −0.0761 0.0005 0.0058 −
0.0034]
5 F8 [0.0803 0.0787 −0.0626 0.0492 F7
0.0854 0.1232 −0.1995 0.4234
0.4102]
6 F4 [0.1009 −0.02108 0.0199 0.3654 F3
0.2181 0.0895 0.00563 0.1124
0.1052]
7 F7 [−0.0424 −0.0193 0.03401 − F7
0.1647 −0.1372 0.06917 0.4217
0.5885 0.2435]
8 F6 [−0.3156 1.2610 0.4428 −0.1179 F6
−0.1162 0.7736 1.3143 −0.1134
−0.2148]
9 F7 [−0.4483 −0.2852 0.6743 0.1692 F6
−0.2551 −0.0914 0.74133 0.4687
−0.0938]
10 F8 [0.0799 0.0633 −0.0806 0.0552 F7
0.1024 0.1618 −0.2088 0.4136
0.4039]
11 F2 [0.0057 0.7975 0.6003 −0.0314 F1
0.0057 −0.0131 −0.0009 0.0286 −
0.0019]
12 F5 [0.0767 0.0613 0.0257 0.2357 F5
0.1357 0.4673 0.0489 0.1964
0.1412]

3.5 Multiple Incipient Fault Detection Using Extended Duval

Triangle Method [27]

The ANN and Fuzzy logic-based models are developed which are capable of diag-
nosing the single fault based on Duval triangle and multiple faults when fault data
318 S. A. Khan et al.

Table 14 Fault diagnosis results of EWANN model

Transformer fault Sl. Known fault type Output fault vector Diagnosed fault type
No.
1 F4 [0.1269 −0.8596 0.8805 F4
0.2297 1.0370 −0.0888
0.0399 0.0783 −0.4534]
2 F3 [−0.1120 −0.0186 − F3
0.0221 0.7825 0.1238
0.2457 −0.1617 0.0110
0.1467]
3 F4 [−0.0040 −0.1416 0.1335 F4
0.5823 0.6867 0.1613 −
0.0365 −0.0796 0.0952]
4 F1 [0.0963 0.7147 0.5022 F1
0.0939 −0.1277 0.0323 −
0.0803 0.0892 −0.0224]
5 F8 [−0.0139 −0.4319 0.4054 F8
−0.2383 0.3353 −0.0469
−0.0558 0.4479 0.5991]
6 F4 [0.0747 −0.0447 0.0490 F3
0.4777 0.2550 −0.0197
0.0177 0.0936 0.0956]
7 F7 [0.1130 −0.0774 0.0484 − F7
0.1531 0.1961 0.0158
0.0968 0.9010 −0.1388]
8 F6 [0.2972 0.1660 −1.2663 F6
0.4912 −1.1672 −0.1778
1.6503 −1.1333 1.3404]
9 F7 [0.9883 0.0187 −0.0882 − F8
0.8844 −2.1782 1.3297 −
0.4382 0.3872 1.8337]
10 F8 [−0.0253 −0.4486 0.3776 F8
−0.1577 0.3745 −0.1340
0.0475 0.3970 0.5701]
11 F2 [−0.1077 0.6849 0.8751 F2
0.0688 −0.0767 0.0217
0.1064 0.0602 −0.1254]
12 F5 [0.1403 −0.0355 0.0964 − F5
0.0233 0.3321 0.5118 −
0.1032 0.2271 0.1383]

corresponds to boundary of any two adjacent fault zones of the Duval triangle. Imple-
mentation of the multiple fault detection scheme is explained in Sect. 2.5. Both FIS
and neural models are tested with 171 fault cases. The testing data is amalgamation of
published and simulated data. Simulated data is added to the testing database so that
model can adapt to new fault cases. It is used to test generalisation capability of the
Performance Enhancement and Extension of DGA-Based … 319

Table 15 Summary of transformer fault diagnosis results

Diagnosis by ANN model Diagnosis by EWANN model
No. of faulty cases tested 100 100
No. of correct diagnosis 53 86
No. of wrong diagnosis 47 14
Accuracy (%) 53 86

model. The database limited to published literature and field cannot ensure whether
the model works for that data or any transformer type. The comparison of ANN and
fuzzy models for detection of multi-faults are shown in Table 16. Sample data for
multi-faults comprises DT, HTF-LTF and LTF-PD. Both the models correctly pre-
dicted three multi-fault cases but failed to recognise LTF-PD. Detailed performance
analysis of the model is carried out by calculating their F1 scores and is comparison
made thereafter.
F1 scores of both models are evaluated by obtaining confusion matrices from
the testing database. It shows the fault identification and isolation capacity of the

Table 16 Performance of FIS and ANN models for detecting multi-faults

S. Input vector Actual FIS ANN FIS &
no. Expected Obtained Expected Obtained ANN
value value value value prediction

1 [0.51 0.38 DT 7 6.99 [0 0 0 1 0 [0.0985 DT

0.11] 0] 0.055 −
0.033
0.769
0.210
0.029]
2 [0.480.51 HTF-LTF 17 16.99 [0 0 0 0 1 [0.006 HTF-LTF
0.01] 1] 0.231 −
0.005
−0.098
1.081
0.782]
3 [0.550.330.12] DT 7 6.99 [0 0 0 1 0 [0.042 DT
0] 0.032
0.030
0.955
0.187 −
0.009]
4 [0.140.840.02] PD-LTF 15 10.99 [1 0 0 0 1 [0.021 HTF
0] 0.034 −
0.078
0.022
1.087
0.018]
320 S. A. Khan et al.

Table 17 Fault diagnosis capability of FIS and neural model including multiple faults
Fault Type % F1 score of FIS model % F1 score of ANN model
PD 98.5 96.6
D1 98.5 89.2
D2 99 98.5
LTF 95 96
HTF 96 95
DT 98 99
D1-D2 98.9 97.2
LTF-HTF 98.5 92
LTF-PD

95 85
F1(fault,remaining)
f1 = 9 97.4 94.2

Table 18 Summary of comparison of ANN and fuzzy models with Duval triangle method
Method Single faults cases Multiple faults cases Total fault types
117 54 171
Conventional duval triangle 96 20 (DT zone) 117
FIS of proposed method 99 51 150
ANN implementation of the 97 47 144
proposed method

models. The score of each incipient (single or multi) fault is calculated with respect
to other faults, and f1 score of the model is obtained by taking average of F1 scores
of all incipient faults. Table 17 shows the individual scores of fault type for both the
models. From the table, it can be found that f1 score of FIS is 97.4% and ANN is
94.2%.
The summary of the performance of ANN and fuzzy-based multi-fault scheme
with conventional Duval triangle method is shown in Table 18. The table clearly
shows that extension of Duval triangle for multi-fault detection improves its
diagnostic capability.

3.6 Testing of FIM-Based Smart DGA Data Interpretation

Method [24]

A two-stage smart DGA data interpretation method is proposed. The first stage
comprises fuzzy implementation of Rogers’ ratio, IEC and Duval triangle methods.
The output of the first stage is fed to the FIM stage which makes decision regarding
the fault existing in transformer and removes the contradiction of first stage classifiers
if any. Details of complete integrated method are discussed in Sect. 2.6.
Performance Enhancement and Extension of DGA-Based … 321

3.6.1 Performance Evaluation of Integrated FIM-Based Diagnosis

Approach

The performance of FIM is evaluated using the same database which is used for
testing its constituents. The details of different incipient faults diagnosed by FIM
are depicted by the confusion matrix given in Table 19. The rows of matrix give
the details of faults diagnosed by FIM whereas columns provide the information
about actual faults existing in transformer. The diagonal of the matrix represents
the correctly diagnosed fault cases by FIM. The first row and column of the matrix
represent performance of FIM for No-Fault cases. The fault database has 34 No-fault
samples. The first row depicts that FIM correctly diagnoses 20 No-fault cases, and
remaining 14 are misdiagnosed. 5, 2, 1, 4 and 2 are misdiagnosed low energy faults,
low-temperature faults, high temperature faults, partial discharges and undiagnosed
cases faults, respectively. The first column depicts that two partial discharge cases
are missed diagnosis. It means two partial discharges are incorrectly diagnosed as
normal operating condition.
The performance indices used for diagnosing the performance of FIM are accu-
racy, sensitivity, consistency and reliability. The details of the indices are given in
Table 20. The sensitivity of FIM is highest (93.12%) for HTF and lowest for no-fault

Table 19 Confusion matrix for FIM

Fault type No. of samples Fault classification of samples
NF LEF HEF LTF HTF PD UD
NF 34 20 5 0 2 1 4 2
LEF 74 0 54 12 7 0 0 1
HEF 116 0 8 106 0 0 2 0
LTF 106 0 0 0 94 12 0 0
HTF 131 0 0 0 9 122 0 0
PD 54 2 0 3 2 0 47 0

Table 20 Performance indices of FIM

Fault type No. of samples Correct diagnosis Incorrect diagnosis Sensitivity
NF 34 20 14 58.82%
LEF 74 54 20 72.97%
HEF 116 106 10 91.3%
LTF 106 94 12 88.67%
HTF 131 122 9 93.12%
PD 54 47 7 87.03%
Total 515 443 64 Consistency
Accuracy 86% 13.9% 81.98%
322 S. A. Khan et al.

Fig. 10 Performance
comparison of FIM with
other methods

(58.82%). It is because the volume of key gases released for this fault is large and
hence maximum chances of rules of HEF to get satisfied. Therefore, sensitivity of
HEF is also high (91.3%). FIM shows sensitivity of 88.6%, 72.97% and 87.03% for
LTF, LEF and PD, respectively. These figures show that sensitivity of FIM for all
incipient faults is increased as compared to fuzzy models of its constituents. How-
ever, for normally operating conditions its sensitivity is same as fuzzy Rogers’ and
fuzzy IEC. Since these methods had same sensitivity of 58.82% and Duval triangle
cannot differentiate between normal and faulty condition, hence even after integra-
tion sensitivity for No-fault condition does change. The resulting consistency and
accuracy of FIM are calculated as 81.98% and 86%.
Further, the reliability of FIM comes out to be 83.99%. The diagnostic ability of
FIM is also compared with SFDA mentioned in literature for same test data. Figure 10
shows the comparison of FIM with other methods based on reliability index. The
figure depicts that FIM is the most reliable DGA-based diagnostic procedure among
all.

4 Conclusions

Application of soft-computing methods to enhance the performance and to extend

the diagnostic capability of conventional DGA methods is discussed in this article.
Different soft-computing technique-based implementations of conventional DGA
methods is presented. Fault identification and location are carried out using DGA-
based ANFIS trained in accordance with IEC standard IEC-599. Duval triangle-
based severity model is proposed and the fuzzy logic implementation of the same is
carried out. Severity information is added by considering the energy released during
generation of various fault gases. Developed FIS can predict the incipient fault and
estimate the severity of the fault as well. In conventional gas concentration-based
DGA equal weightage is given to fault gas while diagnosing. However, the amount
of energy released is different when different gases are generated. In EWDGA, fault
gas is given weightage according to its energy content. So EWDGA-based ANN
model is developed to explore the effect of energy weighing on DGA diagnosis.
Performance Enhancement and Extension of DGA-Based … 323

Further, multi-fault diagnosis is proposed by extended version of Duval triangle.

The relation between boundary faults data and multi-faults is explored. Extended
Duval triangle is implemented using fuzzy logic and neural networks and has ability
to diagnose single and multiple incipient faults. Lastly, DGA methods are integrated
so that the advantages of individual methods can be combined and their limitations
are covered. For this FIM using fuzzy counterparts of DGA methods is presented.
FIM predicts the incipient fault after removing the contradictions in the outputs
of fuzzy models. FIM-based integration results in twofold increase in diagnostic
capability of DGA. First, fuzzy logic improves the performance of conventional
method. Then, FIM chooses the best alternative among contradictory outputs of first
stage and predicts the fault. Therefore, augmentation of conventional methods with
suitable soft-computing techniques improves the capability and reliability of DGA
diagnosis.

Acknowledgements Support from the research project reference no. R.G.P. 2/24/40 under
Deanship of Scientific Research, King Khalid University, Abha, Saudi Arabia is gratefully
acknowledged.

References

1. M. Duval, Dissolved Gas Analysis: It Can Save Your Transformer”. IEEE Electr. Insul. Mag.
5(6), 22–27 (1989)
2. R. Liao, H. Zheng, S. Grzybowski, L. Yang, Y. Zhang, Y. Liao, An Integrated Decision-Making
Model for Condition Assessment of Power Transformers Using Fuzzy Approach and Evidential
Reasoning. IEEE Trans. Power Delivery 26(2), 1111–1118 (2011)
3. W.H. Tang, K. Spurgeon, Q.H. Wu, Z.J. Richardson, An evidential reasoning approach to
transformer condition assessments. IEEE Trans. Power Del. 19(4), 1696–1703 (2004)
4. Interpretation of the Analysis of Gases in Transformers and Other Oil-Filled Electrical
Equipment in Service, IEC Publ. 599, 1978
5. Guide to the Interpretation of Dissolved and Free Gases Analysis, IEC Publ. 60599, 2007
6. IEEE Guide for the Interpretation of Gases Generated in Oil-Immersed Transformers, IEEE
Std C57.104-2008, 2009
7. Y. Huang, H. Sun, Dissolved gas analysis of mineral oil for power transformer fault diagnosis
using fuzzy logic. IEEE Trans. Dielectr. Electr. Insul. 20(3), 974–981 (2013)
8. Q. Su, C. Mi, L. Lai, P. Austin, A fuzzy dissolved gas analysis method for the diagnosis of
multiple incipient faults in a transformer. IEEE Trans. Power Syst. 15(2), 593–598 (2000)
9. M. Nooril, R. Effatnejad, P. Hajihosseini, Using dissolved gas analysis results to detect and
isolate the internal faults of power transformers by applying a fuzzy logic method. IET Gen.
Trans. Distr. 11(10), 2721–2729 (2017)
10. S. Ghoneim, I. Taha, N. Elkalashy, An Integrated ANN-based proactive fault diagnostic scheme
for power transformer using dissolved gas analysis. IEEE Trans. Dielectr. Electr. Insul. 23,
1838–1845 (2016)
11. M. Yang, L. Hu, Intelligent fault types diagnostic system for dissolved gas analysis of oil-
immersed power transformer. IEEE Trans. Dielectr. Electr. Insul. 20, 2317–2324 (2013)
12. S. Souahlia, K. Bacha, A. Chaari, MLP neural network-based decision for power transformers
fault diagnosis using an improved combination of Rogers and Doernenburg ratios DGA. Int.
J. Electr. Power Energy Syst. 43(1), 1346–1353 (2012)
324 S. A. Khan et al.

13. A. Akbari, A. Setayeshmehr, H. Borsi, E. Gockenbach, Intelligent agent-based system using

dissolved gas analysis to detect incipient faults in power transformers. IEEE Electr. Insul. Mag.
26(6), 27–40 (2010)
14. K. Bacha, S. Souahlia, M. Gossa, Power transformer fault diagnosis based on dissolved gas
analysis by support vector machine. Electr. Power Syst. Res. 83, 73–79 (2012)
15. H. Zhengab, Y. Zhangaf, J. Liu, H. Wei, J. Zhao, R. Liao, A novel model based on wavelet
LS-SVM integrated improved PSO algorithm for forecasting of dissolved gas contents in power
transformers. Electr. Power Syst. Res. 155, 196–205 (2018)
16. W. Chen, C. Pan, Y. Yun, Y. Liu, Wavelet networks in power transformers diagnosis using
dissolved gas analysis. IEEE Trans. Power Deliv. 24, 187–194 (2009)
17. A. Abu-Siada, S. Islam, A new approach to identify power transformer criticality and asset
management decision based on dissolved gas-in-oil analysis. IEEE Trans. Dielectr. Electr.
Insul. 19, 1007–1012 (2012)
18. R. Naresh, V. Sharma, M. Vashisth, An integrated neural fuzzy approach for fault diagnosis of
transformers. IEEE Trans. Power Delivery 23(4), 2017–2024 (2008)
19. T. Kari, W. Gao, D. Zhao et al., An integrated method of ANFIS and Dempster-Shafer theory
for fault diagnosis of power transformer. IEEE Trans. Dielectr. Electr. Insul. 25(1), 360–371
(2018)
20. Shakeb A. Khan, Equbal Danish, Islam Tarikul, A comprehensive comparative study of DGA
based transformer fault diagnosis using fuzzy logic and ANFIS models. IEEE Trans. Dielectr.
Electr. Insul. 22(1), 590–596 (2015)
21. S.J. Lee, Y.M. Kim, H.D. Seo et al., New methods of DGA diagnosis using IEC TC 10 and
related databases part 2: application of relative content of fault gases. IEEE Trans. Dielectr.
Electr. Insul. 20(2), 691–696 (2013)
22. S.W. Kim, S.J. Kim, H.D. Seo et al., New methods of DGA diagnosis using IEC TC 10 and
related databases part 1: application of gas-ratio combinations. IEEE Trans. Dielectr. Electr.
Insul. 20(2), 685–690 (2013)
23. A. Wani Shufali, A. Khan Shakeb, G. Dawal, N. Manzar, Diagnosis of incipient dominant
and boundary faults using composite DGA method. Int. Trans. Electr. Energy Syst. (Wiley),
27:e2421 (2017). https://doi.org/10.1002/etep.2421
24. S.A. Wani, S.A. Khan, G. Prashal, D. Gupta, Smart diagnosis of incipient faults using dissolved
gas analysis-based fault interpretation matrix (FIM). Arab. J. Sci. Eng. (2019). https://doi.org/
10.1007/s13369-019-03739-4
25. Wani SA, Farooque MU, Khan SA, Gupta D, Khan MA. (2015) Fault severity determination in
transformers using dissolved gas analysis (DGA) Proceeding of 12th IEEE India International
Conference (INDICON 2015)
26. E. Danish, K.A. Shakeb, I. Tarikul, Transformer incipient fault diagnosis on the basis of energy-
weighted DGA using an artificial neural network, Turk. J. Elec. Eng. Comp. Sci. 26, 77–88
(2018)
27. S.A. Wani, D. Gupta, U. Farooque, S.A. Khan, Multiple incipient fault classification approach
for enhancing the accuracy of Dissolved Gas Analysis (DGA), IET Sci. Measurem Technol.
1–9. https://doi.org/10.1049/iet-smt.2018.5135
28. S.A. Khan, M.D. Equbal, T. Islam, ANFIS based identification and location of paper insulation
faults of an oil-immersed transformer, in Proceedings of 6th IEEE Power India International
Conference, PIICON (2014)
29. R.A. Hooshmand, M. Parastegari, Z. Forghani, Adaptive Neuro-Fuzzy Inference System
Approach for Simultaneous Diagnosis of the Type and Location of Faults in Power Trans-
formers. IEEE Electr. Insul. Mag. 28(5), 32–42 (2012)
30. N. Bakar, A. Abu-Siada, S. Islam, A review of dissolved gas analysis measurement and
interpretation techniques. IEEE Electr. Insul. Mag. 30(3), 39–49 (2014)
31. F. Jakob, P. Noble, J.J. Dukarm, A thermodynamic approach to evaluation of the severity of
transformer faults. IEEE T Power Deliver 27, 554–559 (2012)
32. F. Jakob, J.J. Dukarm, Thermodynamic Estimation of transformer fault severity. IEEE T Power
Deliver 30, 1941–1948 (2015)
Big-Data Analytics Based Energy
Analysis and Monitoring for Multi-storey
Hospital Buildings: Case Study

Nuzhat Fatema, Hasmat Malik and Atif Iqbal

Abstract Big-Data analytics based energy analysis and monitoring for multistory
buildings plays an important role in proper management and control of operation of
the hospital. In this study, hospitals of 16 time zones of three construction categories
(Pre-1980 construction, Post-1980 construction, and new construction categories)
have been analyzed, which includes hospital summary (details about form, fabric,
and HVAC), hospital’s zone summary (details about area, volume, lighting, and
occupants), and location summary (details about utility costs, energy end use, and
peak energy demand). The energy consumption in hospital has been analyzed as
per the HVAC consumption, utility costs, energy end uses and intensities, monthly
peak energy demand, sources of energy, water consumption, and emissions of gases.
Emissions of gases analysis deal with the occupancy analysis in the hospital. Energy
end uses and intensities based analysis and monitoring deals with the analysis of con-
sumed electricity energy (kWh) (in heating, cooling, interior and exterior lighting,
interior and exterior equipment, fan, pump, heat rejection, humidification, heat recov-
ery, water system, refrigeration, generators, and total end uses) analysis, gas analysis
(MJ), and purchase cooling and heating (MJ) analysis. Presented energy analysis
and monitoring provides the better management compatibility of the hospital under
perturbing operating condition as well.

Keywords Hospital building · Energy analytics · Big-Data · Monitoring · Climate

zones

N. Fatema (B)
Singapore Polyclinic, Singapore, Singapore
e-mail: [email protected]
H. Malik
BEARS, University Town, NUS Campus, Singapore, Singapore
e-mail: [email protected]
Division of ICE, NSUT Delhi, New Delhi, India
A. Iqbal
Department of Electrical Engineering, Qatar University, Doha, Qatar
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 325

H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_14
326 N. Fatema et al.

Nomenclature

kWh Kilowatt hour

MJ Autoregressive integrated moving average
HVAC Heating, ventilation, and air conditioning
VAV Variable air volume
CAV Constant air volume
COP Coefficient of Performance

1 Introduction

According to Building Energy Data book 2011 [1], 19% of world energy is consumed
by the US in the following sectors such as transportation (29%), commercial building
(19%), residential building (22%), and industrial (30%) as shown in Fig. 1. So,
building consumes nearly 41% energy of the total US energy consumption. The
consumed energy in US building sector is shared by petroleum (5%), coal (35%),
natural gas (34%), nuclear (16%), and renewable energy sources (9%).
According to Building Energy Efficiency R&D Roadmap, Singapore [2], 31%
of total electrical energy of Singapore is consumed in the building in the following
subsectors like cooling (60%), lighting (15%), ventilation (10%), lift and escalator
(10%), and others (5%) as shown in Fig. 2. As per BCA Building Energy Bench-
marking Report (BEBM) 2018 [3], nine different types of buildings (i.e., office build-
ings (32.94%), hotel buildings (8.96%), retail buildings (21%), mixed developments

Fig. 1 Energy consumption in buildings (commercial and residential) in the United States [1]
Big-Data Analytics Based Energy Analysis … 327

Fig. 2 Electricity consumption in the building sector in Singapore [2]

Fig. 3 Electricity consumption in the building sector in India [6]

buildings (10.63%), healthcare facilities building (6.82%), educational institutions

(10.92%), civic, community, and cultural institutions (1.88%), sports and recreation
centers (2.57%), and transport facility buildings (4.31%)) are included for the major
consumption, the electrical energy in Singapore. Here, in all kinds of commercial
buildings, cooling system (ACMV—Air Conditioning and Mechanical Ventilation)
consume more electrical energy of the total building in the following sectors chiller
(55%), fans (35%), cooling towers (5%), and pumps (5%).
India is a fast-growing country in the world in electricity consumption and it
secures rank third after China and USA [4]. The building energy consumption in
India is increasing by an average of 2.7% per year between 2015 and 2040, which is
more than twice the global average increase [5]. According to Energy Conservation
Building Directive—2018 (based on Energy Conservation Building Code (ECBC)
2017) [6] (Fig. 3), 33% of total electricity is consumed in buildings such as commer-
cial buildings (9%), domestic buildings (24%), and other buildings (rest buildings)
328 N. Fatema et al.

Fig. 4 Factors influencing energy utilization in a building

(67%). The uses of this 33% of electricity in buildings are consumed by lighting
(59%), HVAC system/air condition (31%), and other applications (10%).
After analyzing the above examples, it is concluded that energy consumption in
the building has been increased at epidemic level in the USA, Singapore, India, and
as well as in the world. Due to enhancement of energy uses, several factors are influ-
enced within the building. As per the International Energy Agency (IEA) project
report on total energy use in buildings analysis and evaluation methods [7], six fac-
tors are influenced by the total energy used in the building. These six factors are
climate, building envelope, building equipment, operation and maintenance, occu-
pant behavior, and indoor environmental conditions as shown in Fig. 4. First, three
factors are related to building and next, three factors are related to the user/occupant
of the building.
In this chapter, energy uses within the two hospital/healthcare buildings with a total
floor area of (241351 + 40946 = 282297 ft 2 ) has been analyzed using data-driven
approaches. Here, several analysis at subcomponents level have been performed,
which include electricity uses, gas uses, water consumption, carbon emission, and
scheduling (i.e., scheduling of light, occupancy, equipment, cooling set point, heating
set point, etc.).
The organization of the chapter is as follows: Sect. 2 represents the list of critical
components within the building as per its uses. Section 3 represents the detail expla-
nation of the data set used in the study. Detailed analysis of energy consumption in
the hospital buildings is represented in Sect. 4.
Big-Data Analytics Based Energy Analysis … 329

2 Critical Components Identification Within a Building

In the area of facility management within the building, there are some equip-
ment/assets that are critical to the healthy operation of the building. The failure
probability of these critical equipment has a high risk that must be diagnosed. There
are different types of critical equipment available in the building including secu-
rity systems, safety systems, energy management, utility supplies, material egress,
environmental controls, cooling systems, hazardous substances handling, etc. In clas-
sification of equipment being critical, maintenance operation metrics for reliability
must be considered. Expected RUL (Remaining Useful Life), MTTR (Mean Time
to Repair), and MTBF (Mean Time between Failures) are essential metrics for reli-
ability that must be considered for criticality level identification. RUL defines the
expected life, MTBF determines the elapsed time between healthy operation to fail-
ure condition, and MTTR evaluates the necessary average time needed to repair
faulty equipment. The data generation related to MTTR and MTBF can be obtained
from maintenance management system. In the way of criticality identification of
equipment, three basic steps need to be followed as given below:
Step #1: First, identify the inherent issue (i.e., safety as inherent issue or violation
of any governmental regulation, etc.) for failure. This can be identified by observing
a five-step asset criticality calculation record sheet as shown in [8].
Step #2: Key performance indicators (KPIs) (i.e., need of a special tool, complexity
of repair, lead time for repair, etc.) and its degree needs to be analyzed.
Step #3: evaluation of FMEA (Failure Mode Effect and Analysis) for each equip-
ment operation. Equipment is highly critical if equipment’s failure would generate
a failure in high level of priority that is known as Risk Priority Number (RPN). If
an equipment achieves the above average critical score (critical benchmark), then
equipment would be treated as critical. A list of key components within the building
is shown in Fig. 5.

3 Data Set Used for Study

The dataset used in this study is created by the US Department of Energy (DOE) in
collaboration with three of its national laboratories. Generally, this data set is devel-
oped from commercial reference buildings (formally known as commercial building
benchmark models). These reference buildings are available in three categories as
(1) new constructed (Post-2004), (2) existing constructed after 1980 (Post-1980),
and (3) existing constructed before 1980 (Pre-1980). DOE created the data from
330 N. Fatema et al.

Fig. 5 Key components within the building

16 reference buildings of 16 different locations (including all USA climate zones).

Reference buildings include approximately 70% commercial buildings. Out of 16
reference buildings, two buildings are related to healthcare domain (i.e., one five-
story hospital for inpatient and one three-story for outpatient healthcare center).
Hospital and outpatient healthcare center has a total floor area of 241351 ft 2 and
40946 ft 2 , respectively and are located at time zone 5B (Boulder, Colorado) and 6A
(Minneapolis, Minnesota), respectively. Brief details of both hospital (IPD) and OPD
have been given in Tables 1 and 2, respectively.

4 Analysis of Energy Consumption in Hospital Building

A big-data of energy consumption by hospital buildings (i.e., 16 hospital building

of IPD and 16 buildings of OPD center) has been analyzed and explained into two
case studies. Case study #1 belongs to the energy analytics of hospital buildings
Big-Data Analytics Based Energy Analysis … 331

Table 1 Building information of hospital used as IPD

Building information Value for new Value for post 1980 Value for pre-1980
points construction construction construction
Principle building Healthcare, inpatient Healthcare, inpatient Healthcare, inpatient
use
Total floor area 22,422 (m2 ) 22,422 (m2 ) 22,422 (m2 )
Building shape Rectangle Rectangle Rectangle
Number of floors 5 + basement 5 + basement 5 + basement
Window to wall ratio 0.146 0.146 0.146
Floor to ceiling 4.3 (m) 4.3 (m) 4.3 (m)
height
Roof type IEAD IEAD IEAD
Window dimension 845.4 (m2 ) total area 845.4 (m2 ) total area 845.4 (m2 ) total area
Foundation type Basement Basement Basement
Air barrier system Infiltration (ACH) = Infiltration (ACH) = Infiltration (ACH) =
0.07 0.27 0.27
HVAC system
System type MZ-VAV MZ-VAV MZ-VAV
Heating type Boiler Boiler Boiler
Cooling type Chiller, water cooled Chiller, water cooled Chiller, water cooled
Fan control CAV, VAV CAV, VAV CAV, VAV
Thermal efficiency 80 (%) 78 (%) 80 (%)
Temperature setpoint 60 (°C) 60 (°C) 60 (°C)
Water consumption 4037.61 (m3 ) 4037.61 (m3 ) 4037.61 (m3 )

used as IPD whereas, energy analytics of OPD buildings has been mentioned in case
study #2. In the dataset, consumption of electricity, gas, and water along with equip-
ment’s scheduling has been recorded by the DOE. In this chapter, only electricity
consumption has been focused as key energy of the building and its detail analysis
of consumption has been explained in subsequent sections.

4.1 Case Study #1 (Hospital Type—Inpatient)

Here, electricity consumption by three different categories of 16 IPD buildings of

16-different locations has been analyzed and represented in Fig. 6a–l. Figure 6a
represents the total yearly electricity consumption (in kWh) by each IPD building.
332 N. Fatema et al.

Table 2 Building information of hospital used as OPD

Building information Value for new Value for Post 1980 Value for Pre-1980
points construction construction construction
Principle building Healthcare, Healthcare, Healthcare,
use outpatient outpatient outpatient
Total floor area 3804 (m2 ) 3804 (m2 ) 3804 (m2 )
Building shape Irregular Irregular Irregular
Number of floors 3 3 3
Window to wall ratio 0.20 0.20 0.20
Floor to ceiling 3.05 (m) 3.05 (m) 3.05 (m)
height
Roof type IEAD IEAD IEAD
Window dimension 308.3 (m2 ) total area 308.3 (m2 ) total area 308.3 (m2 ) total area
Foundation type Mass floor Mass floor Mass floor
Air barrier system Infiltration (ft3 /min Infiltration (ft3 /min Infiltration (ft3 /min
per ft2 of exterior per ft2 of exterior per ft2 of exterior
wall) = 0.28 wall) = 0.93 wall) = 0.93
HVAC system
System type Package-VAV with Package-VAV with Package-VAV with
reheat reheat reheat
Heating type Boiler Boiler Boiler
Cooling type DX Cooling Coil DX Cooling Coil DX Cooling Coil
Fan control VAV VAV VAV
Thermal efficiency 80 (%) 78 (%) 80 (%)
Temperature set point 60 (°C) 60 (°C) 60 (°C)
Water consumption 240.48 (m3 ) 240.48 (m3 ) 240.48 (m3 )
Insulation entirely above deck (IEAD)

Maximum energy is consumed by pre-1980 constructed building where newly con-

structed buildings consume least energy. Equipment-wise energy consumption has
been represented in Fig. 6b–k and the energy saving in percentage for each building
has been represented in Fig. 6l. The evaluation of energy saving is performed for
buildings of two categories (post-1980 constructed and new constructed building) by
assuming base consumption of energy by pre-1980 constructed buildings. Maximum
energy saving is done by newly constructed buildings as compared with post-1980
constructed buildings while scheduling of equipment of each building is same for
both categories. In IPD case, the maximum energy saving (reduction in %) is done
by lighting system which is shown in Table 3. In Table 3, green, pink, and white
colors are representing the level of saving. Pink color shows extra consumption and
white color represents the change in energy saving is very negligible for some cases.
Big-Data Analytics Based Energy Analysis … 333

Fig. 6 a–l Electricity consumption (kWh) by IPD building’s equipment

334 N. Fatema et al.

Fig. 6 (continued)

After analyzing Table 3, it is concluded that consumption of total end uses elec-
tricity has been reduced from 19.75% to 24.34% by the newly constructed buildings
as compared with pre-1980 constructed buildings. Here, the total end uses electric-
ity means is the sum of consumed electricity by heating, cooling, interior lighting,
exterior lighting, interior equipment, exterior equipment, fans, pumps, heat rejection,
humidification, heat recovery, water system, refrigeration, and generators. The saving
in consumption mainly depends upon the scheduling of the subsystem of the build-
ing. Here, scheduling for lighting and occupancy during weekdays, saturday, and
other holidays play a major role in saving of electricity. Some kinds of schedulings
are represented in Fig. 7a, b.
Critical heating set point scheduling and critical cooling set point scheduling
remain fixed throughout the 24 h at 21.1 °C and 22.2 °C, respectively. Scheduling
for light, occupancy, heating set point, and cooling set point is varied as per the
occupancy level. Scheduling for some other equipment remains constant.
Table 3 Based on pre-1980 construction, representation of energy saving (%) by IPD buildings
Miami Houston Phoenix Atlanta Los Angeles Las Vegas San Francisco Baltimore
Cooling-1980 −4.1279 −2.9832 −2.4002 −1.8941 −6.2332 −0.7299 −3.8099 −0.1119
21.0110 21.5794 22.3243 23.8267 21.8599 24.710 19.632 24.1461
Interior lighting 1.43E-07 1.43E-07 1.43E-07 1.43E-07 1.43E-07 1.43E-07 1.43E-07 1.43E-07
50.2721 50.2721 50.2721 50.2721 50.2721 50.2721 50.2721 50.2721
Exterior lighting 0 1.22E-06 −1.22E-06 1.22E-06 −2.44E-06 −3.67E-06 −1.22E-06 2.44E-06
26.0225 26.0228 26.0216 26.0206 26.0223 26.0203 26.0206 26.0219
Interior equipment −2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07
−2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07
Exterior equipment 4.81E-07 4.8E-07 4.8E-07 4.8E-07 4.8E-07 4.8E-07 4.8E-07 4.8E-07
4.8E-07 4.8E-07 4.8E-07 4.8E-07 4.8E-07 4.8E-07 4.8E-07 4.8E-07
Big-Data Analytics Based Energy Analysis …

Fans 5.5073 6.8172 5.8398 8.7970 2.424 8.5785 5.5971 8.9095

17.7982 18.5121 21.7412 21.2837 17.1697 24.8627 10.1513 19.2155
Pumps 2.7226 3.6354914 3.8479 4.7104 0.3579 5.8545 2.7742 5.1479
14.3441 14.7797 16.4803 17.1517 15.5876 19.4354 11.6895 15.8351
Heat rejection 1.8986 2.8769 3.24740 4.0210 −0.2059 4.9776 2.1221 5.0774
14.4412 14.9139 15.708 17.5799 15.4778 18.7221 12.3718 17.4098
Humidification 7.4813 7.4768 6.0787 9.1048 2.7653 8.271 7.0525 9.0829
7.4813 7.4768 6.0787 9.1048 2.7653 8.271 7.0525 9.0829
Refrigeration −0.3764 −0.4781 −0.5018 −0.3327 0.0052 −0.4547 −0.1024 −0.3034
−0.4399 −0.5951 −0.7096 −0.3433 −0.2355 −0.3710 −0.0809 −0.1950
Total end uses electricity −1.2728 −0.5447 −0.1091 0.3619 −2.4026 1.3407 −0.8611 1.2894
(continued)
335
Table 3 (continued)
336

Miami Houston Phoenix Atlanta Los Angeles Las Vegas San Francisco Baltimore
21.3926 21.7765 22.6385 23.2836 21.9573 24.338 20.2579 23.187
Albuquerque Seattle Chicago Boulder Minneapolis Helena Duluth Fairbanks
Cooling-1980 −1.2867 −1.2769 −5.127 −5.0305 −4.9376 −5.2089 −7.2765 −10.3567
24.7409 23.0465 20.3323 21.5081 20.0774 20.3632 18.3173 18.9838
Interior lighting 1.43E-07 1.43E-07 1.43E-07 1.43E-07 1.43E-07 1.43E-07 1.43E-07 1.43E-07
50.2721 50.2721 50.2721 50.2721 50.2721 50.2721 50.2721 50.2721
Exterior lighting 1.22E-06 2.45E-06 2.45E-06 −4.89E-06 2.45E-06 3.67E-06 3.67E-06 0
26.0228 26.0220336 26.0209 26.0223 26.0209 26.0222 26.0202 26.0224
Interior equipment −2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07
−2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07 −2.45E-07
Exterior equipment 4.8E-07 4.8E-07 4.8E-07 4.8E-07 4.8E-07 4.8E-07 4.8E-07 4.8E-07
4.8E-07 4.8E-07 4.8E-07 4.8E-07 4.8E-07 4.8E-07 4.8E-07 4.8E-07
Fans 7.4881 6.9584 1.1704 0.8788 1.1657 0.8123 −2.6576 −7.5338
23.87 18.9211 13.5582 17.6591 13.7159 16.1628 10.7680 12.2569
Pumps 4.8856 4.7528 0.5692 0.8054 0.6039 0.5823 −0.9436 −3.9213
18.3481 16.1963 12.1116 13.9624 11.9093 12.6305 10.2418 11.4521
Heat rejection 4.1570 4.0848 0.0558 0.283 0.2068 0.0321 −1.6937 −4.9339
18.6813 15.9495 13.1358 14.8046 12.8521 13.5906 11.2078 12.2955
Humidification 7.1917 8.4027 1.3886 0.8323 1.5414 1.2429 −1.5166 −4.4664
7.1917 8.4027 1.3886 0.8323 1.5414 1.2429 −1.5166 −4.4664
Refrigeration −0.663 −0.3698 −0.0165 −0.2440 0.0944 0.1521 0.1085 0.0757
(continued)
N. Fatema et al.
Table 3 (continued)
Miami Houston Phoenix Atlanta Los Angeles Las Vegas San Francisco Baltimore
−0.3858 −0.1601 0.03855 −0.1109 0.0833 0.2478 0.1199 0.2271
Total end uses 1.0182 0.5192 −1.8178 −1.4805 −1.5737 −1.4416 −2.7249 −3.7847
24.2014 22.7274 20.7716 21.9572 20.7328 21.3586 19.7445 20.4995
Big-Data Analytics Based Energy Analysis …
337
338 N. Fatema et al.

Fig. 7 a–b Scheduling of IPD building’s equipment/occupancy

4.2 Case Study #2 (Hospital Type—Outpatient Center)

Here, electricity consumption by three different categories of OPD buildings has

been analyzed and represented in Fig. 8a–j. Figure 8a represents the total yearly
electricity consumption (in kWh) by each OPD building. In this case, maximum
energy is consumed by pre-1980 constructed OPD buildings where newly constructed
OPD buildings consume least energy. Equipment-wise energy consumption has been
represented in Fig. 8b–i and the energy saving in percentage for each OPD building
of second and third category has been shown in Fig. 8j. The evaluation of energy
saving is performed for buildings of second and third category (second = post-1980
constructed and third = new constructed building) by assuming base consumption
of energy by first category (pre-1980 constructed buildings). Here, maximum energy
saving is also done by newly constructed buildings as compared with post-1980
constructed buildings while scheduling of equipment of each building is same for
both categories. The detailed information of component-wise energy saving is listed
in Table 4. The consumption of total end uses electricity has been reduced from 14.81
to 20.77% by the newly constructed buildings as compared to pre-1980 constructed
buildings. Details of scheduling of lighting system according to the occupancy has
been represented in Fig. 9.
Big-Data Analytics Based Energy Analysis … 339

Fig. 8 a–j Electricity consumption (kWh) by OPD building’s equipments

340 N. Fatema et al.

Fig. 8 (continued)

5 Conclusion

In this study, a detailed statics for energy consumption by building sector has been pre-
sented and compared with other sectors. The consumed electricity at sub-component
level in the building has also been explained in detail. The factors influencing the
energy consumption in a building have been extracted and represented for a better
understanding of the researchers. After this, method for critical component iden-
tification has been presented to identify the critical equipment within/outside the
building. Based on open access dataset provided by DOE (USA), a critical analysis
related to electricity consumption by two types of hospital buildings (i.e., IPD and
OPD) has been presented. Major components of the building have been considered
for the analysis. As per available dataset, maximum reduction (in consumed %) of
electricity is seen in lighting system.
Table 4 Based on pre-1980 construction, representation of energy saving by OPD buildings
Miami Houston Phoenix Atlanta Los Angeles Las Vegas San Francisco Baltimore
Cooling-1980 −12.33643 −2.191886 −3.314799 0.6830158 −5.134306 −0.357508 −1.038956 1.5327002
20.432861 21.65013 23.200528 24.219862 25.747588 26.909428 20.106142 22.501246
Interior lighting 0 0 0 0 0 0 0 0
29.310258 29.310258 29.310258 29.310258 29.310258 29.310258 29.310258 29.310258
Exterior lighting 0 0 0 0 0 0 0 0
16.848498 16.847317 16.846725 16.849253 16.850642 16.850683 16.849336 16.848193
Fans −2.906994 7.8923283 5.8131228 10.009321 2.6637804 8.9422859 8.6331834 10.635381
18.311931 19.150843 21.12489 21.387615 20.75103 23.384866 16.943612 19.121592
Pumps 1.6099356 5.0022946 2.9525653 5.8397465 2.114061 5.1741294 5.8128079 5.8458813
10.073597 10.18816 12.633107 11.498416 11.897738 11.840796 9.3103448 10.008857
Big-Data Analytics Based Energy Analysis …

Heating −24.36441 35.864911 29.030727 40.737096 5.9218377 38.478213 39.112768 45.832651

73.78178 59.890816 62.676474 66.779887 76.655728 71.046229 58.60205 56.754544
Humidification −4.159132 8.2386016 6.0018553 9.6634738 2.9848943 8.7323301 9.7372488 10.377561
−4.159132 8.2386016 6.0018553 9.6634738 2.9848943 8.7323301 9.7372488 10.377561
Total end uses −5.563142 1.647492 0.8391936 4.2924631 −0.737601 3.5400777 3.4900593 5.8999733
15.763994 17.11248 18.3414 19.305873 17.91097 20.774172 15.799257 18.453125
Albuquerque Seattle Chicago Boulder Minneapolis Helena Duluth Fairbanks
Cooling-1980 1.644988 −1.154272 −6.255315 −6.15192 −5.802209 −6.055596 −9.389811 −13.1802
23.297759 21.757777 19.517284 17.020642 19.132327 16.876211 14.692819 18.551184
Interior lighting 0 0 0 0 0 0 0 0
29.310258 29.310258 29.310258 29.310258 29.310258 29.310258 29.310258 29.310258
(continued)
341
Table 4 (continued)
342

Miami Houston Phoenix Atlanta Los Angeles Las Vegas San Francisco Baltimore
Exterior lighting 0 0 0 0 0 0 0 0
16.849377 16.849114 16.849418 16.848275 16.845784 16.845824 16.847806 16.84577
Fans 8.7275185 8.8906151 4.2210718 3.8201799 4.58942 4.4171767 1.7677677 −1.240027
17.878993 19.219789 15.049358 13.889232 14.510681 14.128419 11.731648 12.813996
Pumps 4.534005 5.6264775 1.9329897 1.6034985 1.9818332 2.1538462 0.6686168 0.3681885
9.3702771 9.4562648 6.7869416 6.7055394 7.8447564 8.1318681 6.5190138 9.5729013
Heating 44.114662 41.628881 29.661889 37.684705 31.867572 33.081262 33.923272 31.021086
59.185888 58.679881 49.613553 52.84784 42.733447 43.610862 44.634455 51.382757
Humidification 8.5236854 9.270656 4.9913901 3.7452654 4.8728297 4.9038993 1.8619713 −0.830783
8.5236854 9.270656 4.9913901 3.7452654 4.8728297 4.9038993 1.8619713 −0.830783
Total end uses 4.8665435 5.2168613 1.9669619 2.0022955 2.7560439 2.8013546 2.6963167 3.0912932
17.710445 18.210787 16.165163 14.978818 15.672632 14.999387 14.812096 18.639535
N. Fatema et al.
Big-Data Analytics Based Energy Analysis … 343

Fig. 9 Scheduling of OPD building’s equipment/occupancy

References

1. Building Energy Data Book, U.S. Department of Energy (September, 2019). https://ieer.org/
resource/energy-issues/2011-buildings-energy-data-book/
2. Building Energy Efficiency R&D Roadmap, Singapore. https://www.nccs.gov.sg/docs/default-
source/default-document-library/building-energy-efficiency-r-and-d-roadmap.pdf
3. BCA building energy benchmarking report (BEBM) 2018, p. 5. Accessed 01 Sept 2019 https://
www.bca.gov.sg/GreenMark/others/BCA_BEBR_Abridged_FA_2018.pdf
4. Electricity Statistics, Detailed, Comprehensive Annual Data on Electricity and Heat. Accessed
01 Sept 2019 https://www.iea.org/statistics/electricity/
5. Independent Statistics & Analysis, U.S. Energy Information Administration. Accessed 01 Sept
2019 https://www.eia.gov/todayinenergy/detail.php?id=33252
6. Energy Conservation Building Directive—2018. Accessed 01 Sept 2019 https://www.grihaindia.
org/sites/default/files/pdf/ECBC-Code.pdf
7. The International Energy Agency (IEA) project report 2016, total energy use in buildings: analy-
sis and evaluation methods. Accessed 01 Sept 2019 https://www.iea-ebc.org/Data/publications/
EBC_PSR_Annex53.pdf
8. Whole Building Design Guide: A program of the national institute of building science.
Accessed 24 Sept 2019. https://www.wbdg.org/resources/critical-equipment-identification-and-
maintenance
Maximum Power Extraction
and Monitoring from Wind Power
Generation System Using Intelligent
Controllers

Aabid Hussain Sheikh and Farhad Ilahi Bakhsh

Abstract There has been a growing interest in wind energy in latest years as it is
a budding source for generating electricity with a negligible environmental impact.
With the development of aerodynamic models, wind turbines are easily accessible,
capable of capturing hundreds of kilowatts of energy. They generate a significant
quantity of energy when such WECS are incorporated into the grid. Variable Speed
Generation (VSG) strategies were used to maximize energy extraction from the wind.
While variable velocity operation requires greater complication and original invest-
ment, these disadvantages are offset by increasing energy capture and improving
flexibility in control. A growing number of variable velocity WECS systems are
therefore being suggested as wind technology advances. Variable Velocity Gener-
ation (VSG) approaches differ the turbine’s rotational speed by regulating either a
mechanical or electrical parameter to keep a steady turbine tip-speed ratio. This is
called a MPPT controller. Maximum power point region is the region in which wind
speed is 5–12 m/s and in this region the pitch angle control is done. The monitor-
ing was when the pitch angle controller was designed for MPPT using Fuzzy Logic
controller for a 50 Hz system connected to a grid and compared the performance
with conventional PI controller. The proposed designed control system topologies
designed will regulate pitch command; therefore, power extraction framed can be
controlled and maximized. The results show that the proposed Fuzzy logic pitch angle
control approach had better performance and improved by 10% from the controllers
such as PI and therefore, the efficiency of the system gets improved.

Keywords Wind power · WECS · Variable speed generation (VSG) · Fuzzy logic
controller · MPPT

A. Hussain Sheikh · F. Ilahi Bakhsh (B)

Department of Electrical Engineering, National Institute of Technology Srinagar, Hazratbal,
Srinagar 190006, India
e-mail: [email protected]
A. Hussain Sheikh
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 345

H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_15
346 A. Hussain Sheikh and F. Ilahi Bakhsh

Nomenclature

WECS Wind Energy Conversion Systems

KW Kilowatt
VSG Variable Speed Generation
MPPT Maximum Power Point Tracking
WPGS Wind Power Generation System
DFIG Doubly Fed Induction Generator
SCIG Squirrel-Cage Induction Generators
WTG Wind Turbine Generator
MPP Maximum Power Point
ANN Artificial Neural Networks
WT Wind Turbine

1 Introduction

1.1 Overview

From recent past years, the priority concerning the long-run health and effects con-
cerning environment of standard electricity generation has been growing. Earth sur-
face is not exposed to an equivalent reasonably heat intensity and it is well mirrored
within the movement of huge atmospheric state on the face of the planet. This has
been necessitated as a result of the current usage wants an energy infrastructure, with
an eye on sound of energy from renewable supply with additional property, because
the gap between supply and demand expands year after year, technologist and sci-
entists are on the lookout to find a new source of energy. Though our traditional
nonrenewable sources help us lot to meet out our energy needs, they have their own
limitations and are considered harmful to the environment in the context of pollution
it contributes to our environment. We are forced to have a second thought and tap-
ping of energy from these available sources vigorously. So our attention is directed
towards nuclear energy but the bitter experience that the Russian experienced in the
case of Chernobyl and the recent nuclear accident that took place in Japan put a check
on our endeavor to harness energy from this source also. This report is according to
Monterey Institute of International Studies, May 2010. So we are finally left with the
option of wind energy as it is a seasonal one still we are upset in our onward journey
to harness energy, we find a silver line in the dark, by converting some drawbacks
into advantage we can meet our energy need boldly in future. WECS converts the
kinetic energy available in wind into electrical energy.
WECS has witnessed tremendous development over the previous ten years and
acknowledged as an economically competitive means of electricity generation and
environmentally friendly [1, 2]. As the depth of wind energy into the power system
Maximum Power Extraction and Monitoring … 347

increases, the effects of wind farms on power system operation will become an
increasingly important factor in their profitability and management [3, 4]. Due to the
environmental benefits of wind power, technological advancement, and government
incentives, wind power generation has experienced a big growth from the past ten
years. Worldwide enhancement of wind energy has been expanded rapidly since
early 1990s. Wind energy system has become a cheap, reliable and competitive
means of generating electrical power due to extensive R&D efforts over the past ten
years. Modern WTs life span is now within the range of 20–25 years, which can
be compared to many other conventional power generation technologies. Through
technological growth, enhanced manufacturing rates and the use of bigger turbines,
wind power expenses have continued to decrease.
According to the world wind energy report, in the year 2008, the total contribution
of energy from wind power system is around 9590 MW in India. According to India’s
MNRE, the total installed wind energy capacity could reach the target of 18,420 MW
by the end of 2012. Wind Power in India reported, Tamil Nadu has an installation
capacity of wind energy around 7,150 MW as on January 2013, which is 41% that
of the country. The objective is to ensure that wind energy is the most cost-effective
source of energy on the market by 2030. This is not a small achievement while
considering the rapid advancement in the dimensions of WTGs and in the field of
power electronics and their functionality in the extraction of wind power. On one
side there is a growing demand for green energy and on the other side the rising cost
of turbine and the compulsion one feels in supplying green electricity to the grid,
create a new situation where one should do something to improve the power output.
Earlier WTs were directly linked to the grid using SCIG. Recently, DFIG has evolved
to variable velocity.
Wind turbine controllability is becoming increasingly essential as the turbine
power level rises. DFIG control is more complex than conventional induction
machine control. To regulate the DFIG, a power electronic converter controls the
current of its rotor. A popular technique to regulate the current of the rotor is through
field-oriented (vector) control. Several vector control systems were suggested for the
DFIG. One popular way is by changing the pitch angle of the WT. For linear process
systems, conventional P, PI and PID controllers are efficient. Due to the variable
parameters of the wind scheme, the wind generation scheme is an extremely non-
linear method as it involves power electronic equipment. The nonlinear controller
is therefore essential to control the conversion scheme for wind energy. The objec-
tive of this job is to obtain peak wind energy and discover the optimization strategy
for maximum effectiveness. In order to obtain the above solutions, adaptively opti-
mized intelligent algorithms are proposed for extraction of maximum energy from
the WECS.
This chapter discusses the maximum power extraction and monitoring from
WPGS using intelligent controllers. The region of MPP is the region in which wind
speed is around 5–12 m/s. The WT is operated in this region at its maximum power
output condition using pitch angle control. For achieving maximum power condition,
the power coefficient of the WT is maintained at its maximum value by varying the
rotor speed with respect to the wind speed. The pitch angle is controlled using two
348 A. Hussain Sheikh and F. Ilahi Bakhsh

controllers, fuzzy logic controller, and conventional PI controller. The monitoring is

done when the output from both the controllers is seen and compared and the end
result is seen that the maximum output is obtained when the pitch angle is controlled
using fuzzy logic controller instead of conventional PI controller. This chapter has
been organized into three sections.
Section 1 (current section) gives an overview, classification and different modes of
operation of WECS; presents a brief idea of literature review on the integration of
WECS.
Section 2 This section shows the proposed WECS and its control techniques that
have been created with the evolution of WECS over the past decade.
Section 3 presents the results and discussions. The output from both the controllers
(PI and fuzzy) was obtained and compared.
Section 4 concludes the work presented in this chapter.
Section 5 presents the suggestion for the extension of presented work in future.

1.2 Literature Review

The literature indicates that a lot of efforts are taken to enhance the control perfor-
mance of the WPGS under various conditions. The modern control system theory
has achieved much great importance, nowadays; because of its great impact in per-
formance and thereby affecting the dynamic response of WECSs during the last two
decades such as optimal control, adaptive control, and intelligent soft computing
methods.
MPPT controls have recently been discussed in [5, 6], where wind speeds are
estimated to be MPPT or the MPP is obtained without the information of wind
speed. Paper [7] proposed a work which focuses on the enhancement of maximum
wind power extraction, using hill-climbing algorithms for inverter-based variable
speed WPGSs. Paper [8] suggested a job that suggested and applied a solar and
diesel-wind hybrid generation scheme. With more effectiveness and more stability
and better transient, even under disruption, an effective peak power sharing and
extraction method is effectively proved among energy sources using neural networks.
Paper [9] suggested an interface consisting primarily of a dynamic wind conver-
sion system MPPT control and a semi-controlled single-stage rectifier with embedded
control. In [10], a fuzzy-logic-based control has been used to obtain the MPPT. The
Fuzzy control scheme executes either an incremental or decremented search for a
specific wind velocity for the WTG shaft velocity along the path to boost wind pro-
duction until the device settles at full power. A long search time is required to locate
MPP in the event of change in wind speed.
ANNs are a renowned tool to implement nonlinear time-varying input-output
mapping. To solve the process disadvantages in fuzzy [11] and Li et al. [12] suggest
a wind speed prediction technique based on multilayer perceptron neural network
(MLPNN) for a tiny direct-drive WTG scheme. Fast and smooth speed estimation is
Maximum Power Extraction and Monitoring … 349

made by calculating generator electrical power but this method hit by one drawback
such as black-box property of ANN.
Literature [13] proposed an adaptive maximum algorithm for the extraction of
wind power from variable speeds. With density impacts, the suggested control algo-
rithm is evaluated in the MPP monitoring technique based on the standard hill climb
search. Consideration is given to the memory characteristics and the original Tip
Speed Ratio (TSR) to put the algorithm into practice. The control system is conducted
in three distinct steps in accordance with a generator’s power curve to enhance moni-
toring efficiency. Simulations and experimental outcomes have verified the feasibility
and efficiency of the suggested MPPT approach.

2 Proposed WECS Design

2.1 Introduction

The traits of the wind energy resources are useful from different angles regarding
wind energy utilization. Suitable site selection and economic viability assessment
are two major aspects to be considered before establishing a windmill. The elec-
trical energy output of the WT is highly or totally reliant on the wind speed. As a
renewable/natural resource, the wind velocity will vary with time more abruptly. The
significance of this adaptability becomes serious since it is highly enlarged by the
cubic relation of the available energy.

2.2 Wind Energy Conversion Systems

A WECS is a structure that converts the wind energy to the electrical energy. This
conversion of electricity occurs as follows in two respects. Under the action of the
wind stream, the extraction unit called the WT rotor turns, collecting a mechanical
energy. The rotor moves an electrical rotating machine (generator) which produces
electrical energy. Various researches have done to enhance the efficiency, perfor-
mance characteristics, etc., of WECS. Figure 1 shows the DFIG connected with
WT.
So as to fully understand the ability of variable-speed generation to harvest the
available energy over a broad range of wind speeds, all system components must
be intended to function with high effectiveness throughout the entire working spec-
trum. Because a typical wind generation scheme can most of the moment function
at a fraction of the rated power, achieving high light load effectiveness is critical to
lowering the energy cost (COE) in these schemes. This requirement separates the
350 A. Hussain Sheikh and F. Ilahi Bakhsh

Fig. 1 DFIG with wind turbine

implementation of wind power from almost all other apps where light load effec-
tiveness is generally significantly lower than other factors, such as complete load
effectiveness.

2.3 Control Methods

Many distinct control techniques have been created with the progress of WECS over
the past decades. The control techniques created for WECS are usually split into
two main classifications: methods of constant velocity and techniques of variable
velocity. We cannot control the turbine shaft velocity in constant-speed turbines.
Constant speed control is a simple and cost-effective technique, but variable velocity
provides the following benefits:
(a) Maximum power monitoring for the use of maximum wind energy,
(b) It has low mechanical stress,
(c) Fewer variations in electrical power.
The following will briefly describe these advantages. Using shaft speed control,
greater energy is gained. The energy obtained from a real WT variable-speed system
having a 34 m-diameter rotor is compared with a constant-speed WT at separate wind
speeds and the outcomes are shown in Fig. 2. Figure 2 demonstrates that a system
of variable speed produces more energy than the scheme of constant speed.
There are some fluctuations associated with mechanical part or electrical part of
WT during turbine operation. The mechanical components changes include present
fluctuations triggered by the passing blades through the tower and different present
amplitudes created by changing wind speeds. The electrical converter causes changes
linked to the electrical components, such as harmonics in voltage. You can conquer
the voltage harmonics by selecting the correct filter. However, electrical components
Maximum Power Extraction and Monitoring … 351

Fig. 2 Powers produced by

a variable speed WT and a
constant speed WT at
different wind speeds

can not cancel them owing to the big moment limit of mechanical component changes.
One alternative is a variable-speed WT that can considerably lessen the mechanical
component disturbance.

2.4 Why DFIG Is Preferred?

Nowadays, almost all industries use DFIG to generate power because of the advan-
tages it offers, compared to other types of generators. Although large-scale WTs such
as constant speed WTs and variable velocity WTs have their own merits, these merits
are defeated by the demerits shown by this WT in converting wind energy. A three-
phase asynchronous generator is used in a fixed-speed WT. The total power produced
by the generator is directly given to the grid. Both the generator rotation velocity
and the WT rotor rotation rate are also fixed in this form. So the variation in wind
speed is directly reflected in the torque that has been accumulated at the WT rotor in
the event of turbine rotor speed is fixed. So every time wind gust gets strengthened,
the force or torque changes accordingly, which may finally lead to fluctuation in the
power output generated by the WTG. In the power network to which it is attached,
there is so excellent instability. In the case of variable WTs, the rotation velocity of
the WT rotor may be changed according to different wind speeds. In such WTs, this
deprives us of the option of using asynchronous generators.
In the whole process, the variation in the wind gust, causes more variation in
the torque which in turn, causes some damage to components (gearbox). While
summarizing all the disadvantages that one faces in using the above large scale
WT we prefer DFIG to other large scales WT. In DFIG, by making appropriate
changes, we can drive energy from the generator stator without any change in its
amplitude and frequency in the amplitude and frequency of the ac currents fed into
the generator rotor windings, although differing wind speeds have a major impact on
352 A. Hussain Sheikh and F. Ilahi Bakhsh

the generator rotor velocity as well as the generator stator velocity. The above result
can be reproduced in the case of variable wind speed turbines provided we use a
three-phase asynchronous generator and power electronics. But the drawback is the
size of the power electronics converter should be huge enough to produce 100% of
the generator output.
However, when we incorporate DFIG in variable speed WT, the size of power
electronics can be reduced to 30% of the size that we are supposed to use in the
case of variable speed WT. The main advantages of the Power Electronic Convertor
(PEC) are that the one would enjoy in DFIG are, only 30% of the generator rated
power is fed into generator rotor winding, so reduction is possible in the cost of
power electronic device and power losses in these devices.

2.4.1 DFIG Principle

DFIG is a variable velocity induction machine and it is a wound rotor induction

machine that finds growing applications in the Megawatt spectrum in particular. The
generator’s stator and rotor winding are linked to the grid in the DFIG. Both stator
and rotor are three-phase insulator winding in this model and the stator is supplied
with the required pole design, linked to the grid via three-phase transformer. But a
set of slip rings and brushes connect the rotor to the grid via an internal stationary
loop. Through these parts it is possible to either inject or absorb the regulated rotor
present from the rotor winding.
In a normal situation, it does not generate electricity if the wind speed is underrated
speeds (about 1–3 m/s). The WT will begin operation at wind speeds between 2.5
and 5 m/s calling as “cut in wind velocity.” The wind speed range about 12–15 m/s
is called nominal wind speed. The WT will be stopped above 25 m/s [14].
To solve this issue, torque controller is applied when wind speed is underrated
speed and pitch controller is used when wind speed is overhead rated wind speed.
This above idea is integrated in DFIG as shown in Fig. 3.
During the energy conversion process, power is fed into the grid from both rotor
and stator winding and so this is called “DFIG”. Unlike the WRIG of the type B
where, most probably, the rotor power is dissipated as heat energy, but in type C
WPPS, power in the rotor circuit recovered, treated, transferred and this is fed into
the grid through a partial scale back to back PEC (instated of passive resistors as in
the WRIG) [15].
The power production characteristic depends not only on the induction machine
but also on the two PWM converters as well as how they are controlled. In the super
synchronize speed mode (negative slip) of the DIFG, i.e., when the wind speed goes
above the rated wind speed, power is delivered to the grid directly by stator. At the
same time, power from rotor is also supplied to grid through the power electronic
converter, i.e., passive resistance is inserted. In the face of increasing wind speed,
i.e., above the rate wind speed power is limited to the rated value and additional
power is spilled away by pitching the rotor blade out of the wind.
Maximum Power Extraction and Monitoring … 353

Coupling DFIG
Grid Transformer

Gear box
Partial Scale Power Converter

WT
AC AC
LPF
DC DC

Filter

Control System

Pitch Angle

Fig. 3 Control block diagram of DFIG WTs

In the second case, when wind speed gets below rated wind power that is DIFG
is under the influence sub-synchronize speeds (Positive Slip) PEC enable supply
power at constant value of 50 Hz without any fluctuation because the PEC has a
bidirectional power flow and insertion negative resistance in the PEC by borrowing
power from the line.

2.4.2 DFIG Model

Therefore, the rotor power is absorbed in the sub-synchronous operation from the
grid (Positive Slip) and is fed to the gird (Negative Slip) in case of over-synchronous
operation. In these (both) cases, the stator supplies power to the grid. In Eq. 2.1:

Pgrid
Pstator = (2.1)
1−s

Protor = −s · Pstator

Ωh
S=− being the slip.
ωs

As it is mentioned above, the slip is defined clearly by measuring the speed relation
between rotor flux vector and stator flux vector [16]. In the case of super synchronous
operation, the rotor speed is much higher than that of rotational speed stator flux. The
rotor speed is lesser than that rotation speed stator flux in the case sub-synchronous
operation. The basic point is, throughout process only the rotor rotation flux vector
is adjusted to relative speed of stator flux vector and attempts are made to capture
354 A. Hussain Sheikh and F. Ilahi Bakhsh

maximum power despite where in speed. The rotor slip(s) can be negative or positive
because, owing to the bidirectional nature of the power electronics converter, the
rotor energy may be negative or positive.
The DFIG model is the same as the asynchronous wound rotor machine model
with the exception of the following:
(a) Consider only positive sequence and eliminate the negative sequence.
(b) An input trip has been added. If it is high, the induction generator gets detached
from the rotor converter and the grid.
The application of DFIG in the WECS gives the following lucrative gains:
(a) Via the current of the rotor, the reactive and active powers can be controlled
individualistically.
(b) The DFIG is capable of generating reactive power provided by the grid-side
converters. This converter usually works under a steady unit power factor and
is not reactive in the grid, so that a certain quantity of reactive power can be
produced or consumed by the DFIG. Thus in weak distribution grids, the voltage
is controlled.
(c) The convertor size is obtained by the speed range of the machine and not by the
total power of the generator. If the velocity range is regulated between +30%
of the nominal velocity, the slip range is equivalent to the value of 30% of the
generator’s nominal energy. By the increased performance of the system and
the economical optimization the selected speed range is decided.

2.5 Proposed Simulation Model of WT

A sophisticated system that converts the wind energy into mechanical energy and
then into electrical energy is the conversion scheme for wind energy. In terms of
basic factors such as WT shape and size, wind speed, WT performance or moment
is defined. A WTs dynamic model must comprise parameters that define the WT’s
conduct. Using this defined WT, the efficiency of the WT can be controlled to achieve
the required features. In respect to WPGS, turbines having various characteristics
which show a significant role in electrical power generation. The main components
of a WT system are described in Fig. 4, including a gearbox, WT rotor, power
electronics system, generator, and transformer. WTs transform wind energy into
mechanical energy at the time of high wind speed. The controlling processes are
carried out by active pitch or stall control.
This chapter presents an advanced control system as shown in Fig. 5 that inte-
grate models, based on an intelligent controller to maximize the power of a DFIG
WT. The control scheme of the rotor side converter is organized in a generic way
with intelligent controllers. In terms of MPPT, Fuzzy Logic is used to regulate the
velocity error signal. In addition, another controller is used to generate the rotor
current component’s reference signal to regulate the generator’s reactive power. The
Maximum Power Extraction and Monitoring … 355

Fig. 4 Components of a WT system

Fig. 5 Control of WECS

control system requires the measurement of the stator voltage, stator currents, and
the mechanical rotor position.
For variable speed operation of the WT the use of power electronics converter
paves the way and finally led to more extraction of power. A control technique is
intended to obtain peak energy from the WT during variable speed operation and
to provide steady grid voltage and frequency. A WT’s output energy is maximized.
Two other controllers are created for the WT system based on the transfer function
methods. First, a straightforward strategy is taken based on DC’s system decoupling.
The controller is closely intended for the scheme, but an unstable closed-loop system
outcome from the final controller. Consequently, this instability results in cross-
channel interactions.
The controller bandwidth should be reduced in each channel to reduce the impact
of cross-channel interactions. By trial and error, the closed-loop system becomes
stable and the controller gain reduces. Reducing the benefit from the controller,
however, improves the settling time of the scheme. Second, a sequential loop closure
technique is used to design a Multiple Input Multiple Output (MIMO) controller at
a certain frequency without ignoring the cross-channel interactions. It is possible to
356 A. Hussain Sheikh and F. Ilahi Bakhsh

convert the transfer function of the scheme under research into a triangular form by
ignoring a weak cross-channel coupling.
The design of the SLC controller starts with the design of a SISO controller (single
input single output) for the channel that has no interactions with other channels.
Then, while taking into account the effect of the previously designed channel as a
disturbance, the design will continue to be the next channel.
The first channel transfer function is approximated by a second-order scheme
for simplicity in controller layout. A PI controller is intended to meet the design
requirements of the second-order scheme. The second loop controller is intended in
such a manner that there is a large profit in the closed-loop system at low frequencies
resulting in excellent monitoring and disruptive dismissal. In addition, the closed
loop with second loop controllers has low gain at elevated frequencies, resulting in
robustness of stability and rejection of sensor noise. The energy coefficient, Cp is a
function of rotor blade pitch angle q and tip velocity ratio, which is the proportion of
blade tip velocity to rotor upstream wind speed. The energy coefficient calculation
needs the use of blade geometry and blade element theory understanding. These
complicated problems are usually regarded empirically.

3 Results and Discussions

A PIPAC was intended to control the pitch angle blade for varying speeds in order
to enhance the operational efficiency of the power generation scheme to obtain max-
imum energy in the required quality and value. Since the frequency of the WECS
output voltage is directly proportional to the velocity of the asynchronous generator,
it is possible to adjust the frequency at the same time. The voltage, current and rotor
speed of the WECS with respect to time are shown in Figs. 6, 7 and 8 respectively.
Figure 6 represents the voltage response of the WECS in pu for wind speed of
10 m/s. In the figure, it is clearly shown that the sub transient comes first followed
by transient phase which, is followed by steady-state phase. In steady-state phase,
the voltage variations are lesser compared to the other two phases.
Figure 7 represents the current response of the WECS in pu for wind speed of
10 m/s. In the figure it is clearly shown that there are three phases of the current
response that are sub-transient phase followed by transient phase which is followed
by steady-state phase. The sub-transient phase lasts for few cycles (2–3) followed
by transient (10–15) cycles followed by steady-state.
Figure 8 represents the rotor response of the WECS in pu for wind speed of 10 m/s.
The rotor gets stabilized after some time.
The designed PIPAC system for pitch angle control effectively controls the system
and maximizes the power efficiency put in Fig. 9. Also Figs. 10 and 11 show the
PI-based wind power through which results are obtained (Figs. 12 and 13).
Maximum Power Extraction and Monitoring … 357

Fig. 6 Voltage response of the WECS in pu for 10 m/s

Fig. 7 Current response of the WECS in pu for 10 m/s

When fuzzy controller is applied to alter the pitch angle degree with respect to the
wind speed and rotor speed, the maximum power output is attained. Like that way,
maximum power output is measured for every controller and a comparative study is
done to find out the optimum controller. The designed FPAC system for pitch angle
control effectively controls the system and maximizes the power efficiency.
358 A. Hussain Sheikh and F. Ilahi Bakhsh

Fig. 8 Rotor speed response of the WECS in pu for 10 m/s

Fig. 9 Graphical result using PIPAC

3.1 Comparative Analysis

The performance of fuzzy logic algorithms is equated and analyzed with the exist-
ing PI controller and the obtained simulation results under various wind speed are
displayed in Fig. 14.
Maximum Power Extraction and Monitoring … 359

Fig. 10 PI-based wind power

Fig. 11 Stabilization of output after certain time

Fig. 12 General Simulink circuit diagram

360 A. Hussain Sheikh and F. Ilahi Bakhsh

Fig. 13 The wind farm model

Fig. 14 Comparative analyses of graphical results using PIPAC and FPAC

Maximum Power Extraction and Monitoring … 361

4 Conclusion

This chapter is mainly concerned with the MPPT in WECS using intelligent con-
trollers based pitch angle control. The coordinated application is done at 50 Hz power
system network with a grid-connected wind energy system. The effectiveness of the
proposed control schemes in maximizing the power by adjusting the pitch angle has
been verified for the proposed control scheme Fuzzy Logic and monitoring was done
when the performance of fuzzy logic controller was compared with the performance
of conventional PI controller.
The Fuzzy controller based Pitch Angle Controller (FPAC) is designed with two
inputs to get better control actions. The FPAC exhibits better performance compared
to the conventional Proportional Integral based Pitch Angle Controller (PIPAC) in
tracking maximum power in WECS. But traditional Fuzzy suffers from the curse of
complex rule dimensionality, causing delayed control actions and trial/error design
procedure.

5 Future Work

There are a number of issues that are still to be addressed in the WECS and their
effect on improving the control performance in the grid.
(a) Intelligent controller can be effectively applied at the Grid side and Rotor side
control of the WECS for better performance.
(b) The tuning of controllers can be done using various evolutionary algorithms for
optimized control and thus enhancing the WECS performance.

References

1. F.I. Bakhsh, D.K. Khatod, A new synchronous generator based wind energy conversion system
feeding an isolated load through variable frequency transformer. Renew. Energy 86, 106–116
(2016)
2. F.I. Bakhsh, M.M. Shees, M.S.J. Asghar, Performance of wound rotor induction generators with
the combination of input voltage and slip power control. Russian Elect. Eng. 85(6), 403–417
(2014)
3. F.I. Bakhsh, D.K. Khatod, A novel method for grid integration of synchronous generator based
wind energy generation system, in IEEE International Conference on PEDES, IIT Bombay,
India (2014), pp. 16–19
4. F.I. Bakhsh, D.K. Khatod, Application of variable frequency transformer (VFT) for grid inter-
connection of PMSG based wind energy generation system. Sustain. Energy Tech. Assessments
8, 172–180 (2014)
5. A.G. Abo-Khalil, Dong-Choon Lee, MPPT control of wind generation systems based on
estimated wind speed using SVR. IEEE Trans. Indust. Electron. 55(3), 1489–1490 (2008)
6. T. Ackermann, Wind Power in Power Systems, Wiley (2005)
362 A. Hussain Sheikh and F. Ilahi Bakhsh

7. V. Agarwal, R.K. Aggarwal, P. Patidar, C. Patki, A novel scheme for rapid tracking of maximum
power point in wind energy generation systems. IEEE Trans. Energy Conv. 25(1), 228–236
(2010)
8. V. Akhniatov, Analysis of Dynamic Behavior of Electric Power Systems with Large Amounts
of Wind Power, Ph.D. Thesis, Electric Power Engineering, Technical University of Denmark
(2003)
9. K.H. Ang, G. Chong, Y. Li, PID control system analysis, design and technology. IEEE Trans.
Control Syst. Tech. 13(4), 559–576 (2005)
10. K.J. Astrom, T. Hagglund, The future PID control. Control Eng. Practice 9, 1163–1175 (2001)
11. M. Azzouz, A.-L. Elshafei, H. Emara, Evaluation of fuzzy-based maximum power-tracking in
wind energy conversion systems. IET Renewable Power Gen. 5(6), 422–430 (2011)
12. S.M. Barakati, M. Kazerani, J.D. Aplevich, Maximum power tracking control for a wind turbine
system including a matrix converter. IEEE Trans. Energy Conv. 24(3), 705–713 (2009)
13. B. Beltran, T. A-Ali, M.E.H. Benbouzid, Sliding mode power control of variable-speed wind
energy conversion systems, IEEE Trans. Energy Conv. 23(2), 55 1–558 (2008)
14. S.N. Bhadra, D. Kastha, S.B. Banerjee, Wind Electrical System, Oxford University Press (2011)
15. S. Bhowrnik, R. Spec, J.H.R. Enslin, Performance optimization for doubly fed wind power
generation systems. IEEE Trans. Industr. Appl. 35(4), 949–958 (1999)
16. Y.-L. Binwu, S. Navidzargari, S. Ir-kouro, Power Conversion and Control of Wind Energy
System (John Wiley and Sons Publication, Wiley IEEE, 2011), ISBN-978-1-118-02900-8
Novel Applications of Soft Computing
Techniques for Comparative Analysis
of Maximum Power Point Tracking
in Solar Photo-Voltaic System Under
Perturb Conditions

Mohammad Junaid Khan, Lini Mathew and Amit Kumar Yadav

Abstract Maximum Power Point Tracking (MPPT) is a control technique for the
dynamic optimization of power production by a Solar Photo-voltaic (SPV) module.
MPPT is useful not only in SPV applications but also whenever the power produced
using the power source must be maximized regardless of some unmanageable param-
eters. In this chapter, different MPPT techniques such as Perturb and Observe (P&O),
Fuzzy Logic (FL), Artificial Neural Network (ANN), and Adaptive Neuro-Fuzzy
Inference System (ANFIS) have been designed in MATLABTM /SIMULINKTM envi-
ronment for proposed SPV system. Simulation results have been carried out to record
the tracking performance of MPPT controllers by introducing changes in the solar
radiations. It is found that the proposed model using ANFIS-based MPPT controller
gives better performance as compared to other specified MPPT controllers.

Keywords MPPT · SPV · FL · ANN and ANFIS

1 Introduction

During recent years environmental concern for emission of greenhouse gas and cost
of energy give rise to power generation using Renewable Energy (RE) sources. The
availability of RE sources is different for various sites so it is necessary to carry out
its techno-economic analysis of different hybrid RE sources [1]. SPV framework

M. Junaid Khan
Electrical and Electronics Engineering Department, Mewat Engineering College, Nuh 122107,
Haryana, India
e-mail: [email protected]
L. Mathew
Department of Electrical Engineering, National Institute of Technical Teachers Training and
Research, Chandigarh 160019, India
A. K. Yadav (B)
Electrical and Electronics Engineering Department, National Institute of Technology, Sikkim
737139, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 363

H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_16
364 M. Junaid Khan et al.

SPV Boost Converter

Load

Voltage Duty
Current
Sensor Cycle
Sensor

Different MPPT
Controllers

Fig. 1 Block diagram of proposed SPV framework

output power depends upon Solar Radiation (SR), load impedance and also up to a
certain extent on the solar module temperature. The DC-DC power converters are
used to enhance the performance of renewable energy systems. As SR and ambi-
ent temperatures are dynamic in nature so the controller is essentially needed for
determining the operating point of the renewable modules. The efficient conversion
of renewable energy into electrical energy can be possible using MPPT controller.
Conventional and Artificial Intelligence (AI) MPPT techniques have been described
in [2].
Triki et al. presented two novel MPPT techniques for a standalone SPV framework
[3]. Simulation results established that the novel methods provided better perfor-
mance as compared to conventional P&O MPPT technique which fails under quick
vary in SR. Raju et al. proposed an MPPT controller for SPV system, assessed results
of a specified model connecting with boost converter and P&O MPPT technique [4].
It gives a simplified SPV system, with a simplified MPPT control method used
for higher power outcome. MATLABTM /SIMULINKTM platform has been used for
modeling of SPV with DC-DC power converter and the different MPPT controllers.
This research chapter has been organized in the following manner. Description of
the proposed SPV system has been discussed in Sect. 2, proposed system using P&O,
FL, ANN, and ANFIS-based MPPT controllers have been described in Sects. 3–
6, respectively. In Sect. 7, Comparative analysis of various MPPT controllers for
proposed SPV system in MATLAB/SIMULINK has been provided and Sect. 8,
conclusions.
Novel Applications of Soft Computing Techniques … 365

2 Depiction of Proposed SPV Framework

In Fig. 1, the suggested SPV scheme comprises of the SPV array, a boost converter,
load, and various MPPT controllers. The SPV framework is used as an important
source and generates energy at DC load from the source. Finally, use P&O, FL,
ANN, and ANFIS to compute the Maximum Power Point (MPP) of the specified
modules. MPPT technique is used to change the boost converter duty cycle to match
the impedance of the load with the impedance of the source.

3 Proposed SPV Framework Using P&O-Based MPPT

Technique

P&O measures the current and voltage of the framework and calculates the resulting
power [P1(k)]. After the small voltage change [V = V(k) − V(k − 1)] or perter-
bation of the DC-DC power converter’s duty ratio (D) is in the same direction, the
consequent power [P2(k)] is intended. If [P2(k)] is higher than [P1(k)], the pertur-
bation will be in the same direction, else direction of perturbation will be inverted.
MPP is achieved using this P&O technique. This method creates steady-state oscil-
lations around the MPP. These changes generated are liable for the loss of energy
and therefore trigger an efficiency decrease. The P&O control process flowchart is
shown in Fig. 2 [5].
Although P&O control technique is easy and frequently used in SPV renewable
energy systems, the major limitation of this MPPT method is the fluctuations around
the MPP during fast variation of ecological conditions.

3.1 Modeling and Simulation of P&O MPPT Controller

Specifications of SPV panel and a boost converter for simulation analysis using
P&O-based MPPT controller have been given in Tables 1 and 2, respectively.
Model of SPV system with a boost converter, resistive load, and P&O-based
MPPT controller has been designed using MATLABTM /SIMULINKTM as illustrated
in Fig. 3.
Figure 4 shows the subsystem of P&O MPPT Controller for SPV system.
MATLABTM /SIMULATIONTM platform is used for computing the MPP based on
P&O algorithm and subsystem has been developed based on the algorithm as illus-
trated in Fig. 4. The algorithm receives voltage and current data for computing duty
ratio (D).
Pulse Width Modulation (PWM) subsystem comprises a comparator and triangu-
lar waveform generator blocks whose frequency can be set appropriately as shown
in Fig. 4. The test pattern of input variation for SPV is shown in Fig. 5.
366 M. Junaid Khan et al.

Start

Measure V(k), I(k)

P(k)=V(k)* I(k)
∆P(k)=P(k)-P(k-1)

Yes No
P(k)-P(k-1)>0

V(k)-V(k-1)<0 V(k)-V(k-1)<0
No Yes

Yes
No

Increase duty- Decrease duty- Decrease duty- Increase duty-

rao by ∆D rao by ∆D rao by ∆D rao by ∆D

Send to IGBT

Fig. 2 Flowchart of P&O MPPT technique

The variation in test patterns for solar radiations (in Watt/m2 ), used for simulation
analysis is shown in Fig. 5. It has the step change in radiation, from 800 to 600 W/m2
that exists for 0.5 s, with ramp increase introduced at 1 s. Since the environmental
condition is never constant, a combination of step and ramp signal is selected as the
test pattern for simulation of SPV system.

3.2 Simulation Results and Discussions

Simulation has been carried out using discrete Powergui with a sample time of 50 *
10−6 s and with solver Ordinary Differential Equation (ODE) 45. MPPT controller
has been enabled at 0.1 s so as to provide sufficient initialization time to stabilize the
output of the SPV system. The complete simulation time has been considered as 5 s.
The boost converter’s input power and voltage for the SPV framework are illus-
trated in Figs. 6 and 7, respectively. As the response of the framework, depends on
the radiation intensity, both the waveforms have no specific characteristics and may
vary with change in radiation intensity.
SPV system output Power and voltage with boost converter and P&O technique,
for variation in radiation intensity, with the duty ratio perturbation of 0.001, have
been shown in Figs. 8 and 9. The initial rise of the response is due to the circuit
Novel Applications of Soft Computing Techniques … 367

Table 1 Specifications of
Parameter Value
Vikram SPV array
Isc (in A) 8.1
Voc (in V) 44
Imp (in A) 7.8
Vmp (in V) 34.7
Coefficient of temperature of Voc (% °C) −0.36
Coefficient of temperature of Isc (% °C) 0.025
Light generated current IL (A) 8.2
Saturation current of diode IO (A) 2.5 * 10−10
Ideality factor of diode 0.98
Shunt resistance (Rp in ohms) 3126.56
Series resistance (Rs in ohms) 0.52
Electron charge (q) 1.602 * 10−19 C
Boltzman’s constant (k) 1.38 * 10−23
No. of series-connected solar cells in each 72
array (ns )
No. of series-connected panels per string 2
(Nss )
No. of parallel-connected arrays per 1
strings (Npp)
Maximum power (W) 70.66

Table 2 Boost converter

Parameter Value
parameters
Input voltage (Vin ) ≈88 V
Inductor (L in mH) 3 mH
Capacitor (Co in µF) 470 µF
Switching frequency (fs ) 100 kHz
Duty ratio (D) Variable
Load 100 

components, and then it responds as per the input provided to the SPV model, as
shown in Fig. 5. Fluctuations are observed due to P&O MPPT control method.
The values of input power and voltage of the boost converter for the SPV frame-
work as illustrated in Figs. 6 and 7 are about 190 W and 85 V, respectively, and the
values of output power and voltage with P&O technique values are 200 W and 140 V
as illustrated in Figs. 8 and 9, respectively.
368 M. Junaid Khan et al.

Fig. 3 Modeling of SPV system with P&O MPPT controller

Fig. 4 Subsystem of P&O-based MPPT controller

Fig. 5 Test pattern of solar radiation (in W/m2 ) used for simulation analysis
Novel Applications of Soft Computing Techniques … 369

Fig. 6 Response of input power of a boost converter for SPV framework

Fig. 7 Response of input voltage of a boost converter for SPV framework

Fig. 8 Response of output power of a boost converter with P&O technique for SPV framework
370 M. Junaid Khan et al.

Fig. 9 Response of output voltage of a boost converter with P&O technique for SPV framework

4 Proposed SPV System Using FL-Based MPPT Technique

FL-based MPPT technique has become more common in latest years compared to
conventional MPPT methods due to very good and rapid response, no over-shooting,
and reduce fluctuations under quick changes in radiation and temperature. For its
design, the FL-based MPPT method does not require the right SPV framework model
[5]. Two inputs and one output were expected for most of the investigator, FL con-
troller. The input variables are taken as (i) the E(k) error and (ii) the changes in error
E(k). It is provided by [6]

I I
E(k) = + (1)
V V

E(k) = E(k) − E(k − 1) (2)

where
I is the output current of RE system,
I is the change in current for two consecutive instants given by I (k) − I (k − 1)
V is the output voltage of SPV array
V is the change in voltage for two consecutive instants given by V(k) − V(k −
1)
The fuzzy inference can be done using one of the different available methods.
Mamdani’s technique of FL controller was mostly used and the technique of defuzzi-
fication of the center of gravity is used to calculate the yield (change in duty ratio,
D (k)). Figure 10 has shown the system of such a MPPT method.
Novel Applications of Soft Computing Techniques … 371

Fig. 10 Block diagram of FL-based MPPT

4.1 Modeling and Analysis of FL-Based MPPT Controller

for SPV System

FL-based MPPT algorithm, applied to an SPV system with a boost converter and
resistive load, has been developed and simulated in MATLABTM /SIMULINKTM plat-
form. This is shown in Fig. 11. In the model of the FL controller, the primary control
variables should be identified and the sets describing the values of each linguistic
variable and their range should be determined. The calculation and transformation
of input signal of FL tool into linguistic variables.
The subsystem of the FL controller is intended as revealed in Fig. 12. Figure 13
illustrates the input and output of the FL-based MPPT controller.
In order to develop membership functions (MF), human intelligence, and com-
prehension are essential for the intuition to be achieved. It must be well known to
understand the issues and expertise connected with the linguistic variable. The trian-
gular MF is used in the intuition technique to fuzzify the crisp value into membership

Fig. 11 Modeling of SPV system with FL controller

372 M. Junaid Khan et al.

Fig. 12 Sub-system for FL-based MPPT controller

Fig. 13 FL controller
scheme with inputs and
output

value. The variables such as error (Err) and error change (Err,ate) are fuzzified by
triangular MF into membership value in the FL tool. The range of (−5, 5) is used
for input variables and (−0.2, 0.2) for output variable of SPV system.
Seven fuzzy sets are regarded for MF, these variables are described as fundamental
fuzzy sets in terms of language factors, for example, Negative Small (NS), Negative
Medium (NM), Negative Big (NB), Zero (ZE), Positive Small (PS), Positive Medium
(PM), and Positive Big (PB).
As shown in Fig. 14, seven triangular MF were assigned for input and output
factors. Mamdani’s fuzzification technique and centroid defuzzification technique
were used to implement the FL MPPT algorithm. The rule base has been considered
as specified in Table 3.

4.2 Simulation Results and Discussions

All the models explained in the previous sections in which FL-based MPPT controller
has been incorporated in the SPV framework (as illustrated in Fig. 11) have been
simulated and the results are obtained.
The simulation parameters considered are as given in Tables 1 and 2. The test
patterns of input variation employed are as illustrated in Fig. 5. The input power and
Novel Applications of Soft Computing Techniques … 373

Fig. 14 MF for SPV system: a E(k), b E(k), and c D(k)

Table 3 Rule base for FL controller for SPV framework

Change in duty ratio Error (Err)
(delta-D) NB NM NS ZE PS PM PB
NB ZE ZE NM NS PM PM PB
NM ZE ZE NS ZE PM PS PB
Error NS ZE ZE ZE ZE PM PS PB
Rate ZE NM NB ZE NM PM PS PB
(Err.ate) PS NM BB ZE NM ZE ZE ZE
PM NM NB PS NM ZE ZE ZE
PB NM NB PM NM ZE PS ZE
374 M. Junaid Khan et al.

Fig. 15 Power response of power boost converter with FL controller for SPV framework

Fig. 16 Voltage response of a power boost converter with FL controller for SPV system

voltage of DC-DC power converters obtained are as shown in Figs. 6 and 7. Output
power and output voltage of the SPV system with FL controller have been obtained.
Figure 15 and 16 illustrate the output power and output voltage of the SPV system
with a boost converter and FL controller, to variation in radiation intensity. The stable
output power is 230 W and output voltage is 150 V. When compared to the output
power response through P&O technique as shown in Figs. 8 and 15 show reduction
in fluctuations, and thus FL-based MPPT controller provides better performance.

5 Proposed System Using ANN-Based MPPT Controller

For nonlinear and complex systems, ANN architecture provides the suitable solution.
The backpropagation network among various types of neural networks is more widely
spread and useful. ANN controller operates like a black box model and detailed
information is not required about the framework [7].
Novel Applications of Soft Computing Techniques … 375

Fig. 17 ANN model

In this chapter, two-layer feedforward ANN architecture has been used for gener-
ating duty ratio in order to track MPP. The input region receives external data (voltage
and current). The hidden layer (first layer) contains the hidden neurons which receive
data from the input region and sends them to the output layer to the second layer.
Figure 17 shows this scheme.
The layer of input consists of voltage and current collected from a real-time
system under varying radiation and temperature. Ten hidden layer nodes are used in
this neural network. The relationship between the inputs and outputs are given in [8]


2
n1j = wji xi + b1j , j = 1 − 10 (3)
i=1

where wji is the weighting on the connection from the ith input unit, b1j represents
the bias for the neurons of the hidden layer. The neuron output in the hidden layer is
determined by
 2 

a1j = f1 wji xi , f1 (x) = purelin(x) = x (4)
i=2

The net input to the output layer neurons can be written as follows:


1
n2k = wkj a1j + b2k , k = 1 − 4 (5)
i=1

where wkj is the weight on the connections from the jth input unit and b2k represents
the bias for the second layer neurons. The output of the second layer a2k is the network
output of interest and these outputs are labeled as yk and given by
⎛ ⎞

10
a2k = yk = f2 ⎝ wkj + b1k ⎠, f2 (x) = purelin(x) = x (6)
j=1
376 M. Junaid Khan et al.

Function fitting is the method of training a neural network to generate an related

set of target outputs on a set of inputs. It forms a generalization of the input–output
relationship once the neural network has fitted the information and can be used to
produce outputs for inputs on which it has not even been trained. During practice,
training data were provided to the network and the network is modified based on
its error. Validation information has been referred to as generalizing the measured
network and stopping training when generalization stops improving. Neural network
data testing offers an autonomous find of network performance during and after
training.

5.1 Experimental Setup of SPV Emulator

SPV emulator is a power supply that is programmable to emulate solar panels. The
emulator reacts to changes in load circumstances with the quick transient response
and retains the output on the selected panel’s I-V features for a specified ambient
situation. It is a versatile tool intended to emulate solar panel production from various
producers, variations due to daytime, seasonal impact, and distinct assembly places.
SPV emulator uses an internal algorithm to adjust the Voc and Isc , therefore, as
to match the solar panel selected by the user based on ambient temperatures, solar
radiation levels, etc. The power output by solar panels depends on lot of parameters.
These parameters are dynamic and interdependent, making it a complex process to
predict the response of the SPV framework. SPV emulator takes into account weather
conditions at the given moment of year, sun position at the given moment of day,
panel place, panel placement, panel technology, and panel manufacturer to estimate
solar panel response.
The emulator then ensures that the response matches the response of the real panel
under all within-range load circumstances. The user is presented with a complete set
of information with tables and plots on user applications. The emulator is capable
of storing up to 25 I-V curves into its memory, with a programmed time interval
range of 1 hour. It can emulate the I-V curve for the complete day for SPV inverter
testing or dynamic I-V curve transient testing. SPV emulator was taken data from the
database [9]. SPV emulator is illustrated in Fig. 18 and its specifications are specified
in Table 4.

5.2 Modeling and Analysis of ANN-Based MPPT

Block diagram representation of ANN-based MPPT is shown in Fig. 19. The inputs
of ANN model can be current, voltage, and/or ecological data such as temperature
and solar radiation. The output of the ANN model is the duty ratio (D) used in order
to control the DC-DC power converter for MPP.
Novel Applications of Soft Computing Techniques … 377

Fig. 18 Laboratory setup of SPV emulator

Table 4 Specifications of SPV emulator

I/O Parameter Value
Input Supply voltage 230 V AC, 50 Hz
Isc per channel 0–8 A
Voc per channel 0–50 V DC
Maximum output power per channel 400 W
Output Maximum channels in series/parallel 4
Maximum absolute voltage at output 200 V DC
Voltage range of slew rate 0.01–2 V/ms
Current range of slew rate 0.01–1 A/ms or Inf
Mode of operation Fix, table, simulator and programming
modes
Connector at output Banana type
Operating environment Indoor use

Fig. 19 ANN-based MPPT Hidden

Layer
controller Input
Layer
Output
Layer

I
i
Wij
j
378 M. Junaid Khan et al.

To train the neural network with the input data of the SPV framework this can
be yield from experimental setup [5]. The ANN system can be used for tracking of
MPP in SPV framework after training it with the input data of proposed model. A
neural network is designed through two-layer feedforward and ten sigmoid hidden
neurons using the ANN toolbox of MATLABTM /SIMULINKTM . The neural network
is trained by experimental data set using Levenberg–Marquardt backpropagation
algorithm. A total of 1000 real-time samples were gathered, experimental setup of
SPV system (as shown in Fig. 18) out of which about 70%, viz., 700 data samples
were used to train the neural network although remaining data samples were used
for testing and validation purposes. These data samples (which are of voltage and
current) are used to train a ANN for SPV model.
ANN-based MPPT controller was presented in MATLABTM /SIMULINKTM as
shown in Figs. 20 and 21. This was used as a subsystem in the model of SPV
framework with boost converter and MPPT techniques. This is shown in Fig. 22.
ANN system has been specified based on Mean Square Error (MSE) and regres-
sion in Table 5. Regression provides the correlation between goals and outputs while
MSE is the average squared difference between the outputs and the targets. 1’s
regression value relates to a close relationship while 0 relates to a random connection.

Fig. 20 Sub-system for ANN-based MPPT controller

Fig. 21 Details of the subsystem of ANN controller

Novel Applications of Soft Computing Techniques … 379

Fig. 22 Modeling of SPV system with ANN-based MPPT controller

Table 5 MSE and regression

Process Samples MSE Regression
values by data process in
neural network model Training 700 1.04727 * 10−5 1.66035 * 10−1
Validating 150 1.23945 * 10−5 1.33384 * 10−1
Testing 150 1.01458 * 10−5 8.61696 * 10−2

5.3 Simulation Results and Discussions

The models described in the previous section, in which the SPV system incorporated
with the ANN-based MPPT controller, have been simulated and results obtained.
The simulation parameters and the test patterns of input variables are the same
as given in previous sections. The responses of input power and voltage of DC-DC
converters for SPV system are also same as in previous section. The responses of
output power and voltage with variations in radiations and load using ANN controllers
for SPV system have been shown in Figs. 23 and 24. These responses are very much

Fig. 23 Response of output power of a boost converter with ANN controller for SPV system
380 M. Junaid Khan et al.

Fig. 24 Response of output voltage of a boost converter with ANN controller for SPV system

similar to the output responses obtained when FL-based MPPT controller has been
connected as in Figs. 15 and 16.
Comparative analysis of waveforms of output power and output voltage of the
SPV system under consideration with P&O-based MPPT controller, FL-based MPPT
controller and with ANN-based MPPT controller has been carried out. No overshoot
and oscillation free tracking have been observed. It is inferred that ANN-based MPPT
controller behaves very much similar to FL-based MPPT controller, but illustrates
better performance than P&O-based MPPT controller.
The stable output power and output voltage values of the SPV framework with a
boost converter and ANN technique are obtained as 240 W and 155 V, respectively.

6 Proposed System Using ANFIS-Based MPPT Controller

ANFIS scheme consists of the FL controller and the neural network. The FL controller
is capable of converting the linguistic variables into numerical values, including the
fuzzy rule base and MF [5]. However, it can become a difficult job to deliver optimum
fuzzy laws and MF that depend on system behavior. Neural network method is a
prevalent method used to represent nonlinear features. ANFIS method overcomes
the disadvantages of individual techniques of control [10].
Using an ANFIS-based MPPT controller, MPP can be monitored with greater
precision as compared to the P&O method. A typical ANFIS-based approach for
MPPT in RE systems has been illustrated in Fig. 25. The ANFIS algorithm receives
voltage and current as inputs and generates duty ratio to drive the switch of the
DC-DC converter to monitor the MPP. ANFIS controller has six layers as shown in
Fig. 25.
Novel Applications of Soft Computing Techniques … 381

Fig. 25 ANFIS-based MPPT scheme

Fig. 26 Subsystem for ANFIS MPPT controller

6.1 Modeling and Analysis of ANFIS-Based MPPT

ANFIS controller was created with two inputs in the current work: (i) voltage and (ii)
current and one output of the duty ratio. It is necessary to generate fuzzy rule-base
based on the Sugeno inference model in this controller. The ANFIS subsystem has
been shown in Fig. 26. Structure of the ANFIS model has been depicted in Fig. 27.
The training data has kept the same as that of ANN MPPT controller. MATLAB’s
anfisedit tool was used to develop the ANFIS controller in the fuzzification layer
with two neurons in layer 1 and 14 neurons. Modeling of SPV system using ANFIS
as illustrated in Fig. 28 has been carried out.

6.2 Simulation Results and Discussions

The investigator designed and simulated the models of SPV with ANFIS-based
MPPT controller and obtained the results. The simulation parameters and the test
382 M. Junaid Khan et al.

Fig. 27 ANFIS model structure

Fig. 28 Modeling of SPV system with ANFIS-based MPPT controller

patterns of input variation for SPV system are the same as been used in the other
cases also. The responses of input power and input voltage of DC-DC converters for
SPV system are also same as in Sect. 4.2.
Figures 29 and 30 show the responses of output power and output voltage with
variations in radiations and load using ANFIS-based MPPT controllers for SPV
system. The investigator carried out comparative analysis of the output power and
voltage waveforms of the proposed system under consideration with P&O, FL, ANN,
and ANFIS controllers.
Figures 29 and 30 are almost similar to the waveforms obtained when the ANN-
based MPPT controller has been employed. No overshoot, no oscillation stable
Novel Applications of Soft Computing Techniques … 383

Fig. 29 Output power response of a power boost converter with ANFIS controller for SPV system

Fig. 30 Output voltage response of power boost converter with ANFIS technique for SPV system

response are characteristics of these waveforms, showing that ANFIS controller

provides better performance.
The stable output voltages of the SPV system with ANFIS controller is 155 V. In
steady-state condition, the output power created by SPV system is 250 W.

7 Comparative Analysis of Various MPPT Controllers

for Proposed SPV System

Simulations have been carried out to compare MPPT performances of different MPPT
controllers presented in the earlier sections. Test patterns of the solar radiation have
been used for the simulations as shown in Fig. 5.
The response of output power of the SPV system with P&O, FL, ANN, and ANFIS
MPPT controllers have been drawn combining Figs. 8, 15, 23, and 29, in one figure
for comparison purpose. This is shown in Fig. 31.
From Fig. 31 it can be seen that the response of FL controller is faster than P&O.
The response with P&O MPPT controller shows sustained oscillations even of small
magnitude.
384 M. Junaid Khan et al.

300

250

Power (W) 200

150

P&O
100 Fuzzy Logic
Neural Network
ANFIS
50

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (Sec)

Fig. 31 Response of SPV power with P&O, FL, ANN, and ANFIS MPPT controllers to the variation
in radiations

Figure 31 shows that the ANFIS controller provides fast and better performance
as compared to other MPPT controllers. It is also clearly seen that the steady-state
response is achieved fast and better results obtained using ANFIS controller as
compared to other specified MPPT systems.

8 Conclusions

Different kinds of MPPT techniques, viz., P&O, FL, ANN, and ANFIS controllers
for SPV system have been offered in this chapter. Modeling of these controllers for
the SPV system with boost converter, and resistive load has been implemented on
MATLABTM /SIMULINKTM environment. Simulation has been carried out to record
tracking performance of MPPT controllers introducing step and ramp changes in
radiations.
In steady-state, the power produced by SPV system using P&O, FL, ANN, and
ANFIS MPPT controllers are 200 W, 230 W, 240 W, and 250 W, respectively. It
is very clear that the power production of SPV system using ANFIS MPPT con-
troller
is higher than

the SPV system using P&O-based

MPPT controller by 25%
(250−200) (250−230)
× 100 and 8.7% × 100 than the SPV system using FL-based
200
230

MPPT controller and 4.2% (250−240) 240

× 100 than the SPV system using ANN
MPPT controller. Thus, the SPV renewable energy system using ANFIS-based MPPT
controller is found to be a better option. The steady-state oscillations were found to
be greatly reduced using FL MPPT as compared to the P&O. Better tracking has
been achieved for step and ramp changes in radiations. Finally, it is observed that the
ANFIS MPPT controller gives better response for changes in radiation as compared
to other specified MPPT controllers.
The proposed MPPT controllers for a grid-connected hybrid system with a
multilevel inverter can also be carried out.
Novel Applications of Soft Computing Techniques … 385

References

1. M.J. Khan, A.K. Yadav, L. Mathew, Techno economic feasibility analysis of different combina-
tions of PV-Wind-Diesel-Battery hybrid system for telecommunication applications in different
cities of Punjab, India. Renew. Sustain. Energy Rev. 76, 577–607 (2017)
2. M.J. Khan, S. Chatterji, L. Mathew, A. Sharma, A survey of various maximum power point
tracking techniques used in solar photovoltaic system. Excel India Publishers New Delhi (2014),
pp. 283–288
3. Y. Triki, A. Bechouche, H. Seddiki, D.O. Abdeslam, P. Wira, ADALINE based maximum
power point tracking methods for stand-alone PV systems control, in 2018 IEEE International
Conference on Industrial Technology (ICIT) (2018, February) (IEEE), pp. 880–885
4. J. Raju, B.T. Kumar, Optimized maximum power point tracking technique for the simula-
tion of photovoltaic system sourced for higher output power, in Proceedings of International
Conference of Innovations in Electrical, Electronics, Instrumentation and Media Technology
(ICEEIMT), pp. 225–228 Coimbatore
5. D.S. Karanjkar, S. Chatterji, A. Kumar, S.L. Shimi, Fuzzy adaptive proportional-integral-
derivative controller with dynamic set-point adjustment for maximum power point tracking in
solar photovoltaic system. Syst. Sci. Control Eng. Open Access J. 2(1), 562–582 (2014)
6. M.M. Algazar, H.A. El-Halim, M.E.E.K. Salem, Maximum power point tracking using fuzzy
logic control. Int. J. Electr. Power Energy Syst. 39(1), 21–28 (2012)
7. P.C. Chen, P.Y. Chen, Y.H. Liu, J.H. Chen, Y.F. Luo, A comparative study on maximum power
point tracking techniques for photovoltaic generation systems operating under fast changing
environments. Sol. Energy 119, 261–276 (2015)
8. Y.H. Chiu, Y.F. Luo, J.W. Huang, Y.H. Liu, An ANN-based maximum power point tracking
method for fast changing environments, in The 6th International Conference on Soft Computing
and Intelligent Systems, and The 13th International Symposium on Advanced Intelligence
Systems (IEEE, 2012, November), pp. 715–720
9. Solar Photovoltaic (SPV) Emulator as Database of I–V characteristics, http://pvemulator.
igrenenergi.com/pvemulator. Accessed 2014
10. H. Abu-Rub, A. Iqbal, S.M. Ahmed, Adaptive neuro-fuzzy inference system-based maximum
power point tracking of solar PV modules for fast varying solar radiations. Int. J. Sustain.
Energ. 31(6), 383–398 (2012)
Novel Application of Relief Algorithm
in Cascade ANN Model for Prognosis
of Photovoltaic Maximum Power Under
Sunny Outdoor Condition of Sikkim
India: A Case Study

Vinay Kumar Reddy Chimmula, Amit Kumar Yadav and Hasmat Malik

Abstract In photovoltaic (PV) modules the manufacturer provides rating under

standard test conditions (STC). But STC hardly occur under outdoor conditions so it
is important to investigate PV power by experimental analysis. It is found that exper-
imental analysis of PV modules’ maximum power under outdoor conditions remains
a major research area. For this measurement of 74 Wp, PV module is performed
under outdoor conditions at National Institute of Technology, Sikkim, India. To find
the most influencing variables for PV power prediction relief attribute evaluator is
implemented and cascade ANN models are used to predict maximum power under
sunny outdoor condition. It is found by Relief algorithm that Temperature, Solar
Radiation, Short-Circuit Current, Open-Circuit Voltage are relevant variables and
humidity is a less influencing variable. Cascade ANN model which utilizes open-
circuit voltage and short-circuit current has least root mean square error of 0.17 and
SVRM utilizes solar radiation that has least RMSE of 0.40, showing these variables
can be used for prediction of maximum power. This study is useful for PV installation
for providing prior knowledge of maximum power.

Keywords Cascade ANN · Relief algorithm · Maximum power · Photovoltaic

module

V. K. Reddy Chimmula · A. K. Yadav

Electrical and Electronics Engineering Department, National Institute of Technology Sikkim,
Sikkim, India
e-mail: [email protected]
H. Malik (B)
BEARS, University Town, NUS Campus, Singapore, Singapore
e-mail: [email protected]
Division of ICE, NSUT, Delhi, India

© Springer Nature Singapore Pte Ltd. 2020 387

H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_17
388 V. K. Reddy Chimmula et al.

1 Introduction

The performance characteristics of PV modules are needed in order to model their

annual performance. Information available from manufacturers is typically limited
to temperature coefficients, short-circuit current Isc, open-circuit voltage Voc, and
the maximum power Pmax at rating conditions. This information, while useful for
comparing PV module performance at rating conditions, is inadequate for predicting
annual performance under typical operating conditions. It has been shown that the
relative performance ranking at real conditions may not be in agreement with the
annual performance ranking at standard conditions. Discrepancies between the rated
and annual rankings are attributed to the large annual variations, which are experi-
enced in incident angle, solar spectrum, operating temperature, and solar irradiance
when compared with the typical rating conditions of 25 °C cell temperature, a pre-
scribed solar spectrum, 1000 W/m2 irradiance level, and zero angle of incidence. For
the optimum design of PV power systems, it is desirable to measure their long-term
performances at the site of installation.
The reduction of output power in PV modules can be attributed to many factors.
These factors are the mismatch effects, wind speed, weathering conditions, and shad-
ows. The main reasons for this reduction are the mismatch between the PV modules,
which is caused by differences in the working temperature of the PV modules, the
inclination and orientation angles, and differences in the electrical characteristics
resulting from the manufacturing process. The output power of the complete PV
system is lower than the sum of the power of each PV module.
The performance of PV system is influenced by solar radiation and temperature
therefore to design an accurate PV system electrical characteristic model under real
operating condition is necessary. In this aspect various authors estimate PV module
electrical characteristics. Bahgat et al. [1] used ANN to predict maximum power
and normal operating power of UNI-POWER US-64 solar PV module for clear and
cloudy days. The network utilized solar radiation and PV module surface tempera-
ture as inputs for prediction. Almonacid et al. [2] used ANN technique for predicting
electrical characteristics of Si-crystalline PV modules. It is found that ANN method
generates voltage-current curve of the PV module for low value of solar radiation.
It is suggested that ANN is a useful tool to determine PV module power under out-
door conditions proving useful for photovoltaic system installations. In another study
Almonacid et al. [3] developed ANN model for predicting maximum power of high
concentrator photovoltaic (HCPV) modules at south of Spain. The input parameters
to model are direct normal irradiance (DNI), air mass, precipitable water, tempera-
ture, and wind speed (WS). The ANN model gives better prediction except for DNI
value below 400 W/m2 and wind speed above 3 m/s. Sulaiman et al. [4] incorpo-
rate partial evolutionary programming (EP) and artificial neural network (ANN) for
output power prediction of 42 kWp grid-connected PV system installed at roof of
Quadrangle Building, University of New South Wales, Australia. The network used
solar radiation, wind speed, and ambient temperature as inputs and PV power as out-
put. The optimum momentum, hidden layer neurons, and learning rate for training
Novel Application of Relief Algorithm in Cascade ANN … 389

are calculated by EP. The results show that ANN with EP gives better prediction than
only ANN. Bonanno et al. [5] used radial basis function neural network (RBFNN)
to predict electrical characteristics of PV module. The input variables to RBFNN are
solar radiation and temperature. The back propagation algorithm is used for training
the RBFNN.
Mellit and Pavan [6] used two ANN models (ANN-1, ANN-2) for predicting
power of 20 kWp grid-connected photovoltaic at Trieste (Italy). The solar radiation
and air temperature are used as input variables for ANN-1 model, and for ANN-2
model solar radiation is used as the input variable. The Root Mean Square Error
(RMSE) in ANN-1 model varies from 0.00375 to 0.00398 and RMSE for ANN-2
model varies from 0.00442 to 0.00478, showing air temperature affect has less effect
on prediction accuracy of the model.
Mellit et al. [7] developed ANN models for estimating the power of 50 Wp Si-
polycrystalline PV module in cloudy and sunny days. The comparative analysis
shows that ANN model estimate power better than polynomial regression, multiple
linear regression, analytical and one-diode models. Brano et al. [8] used air tem-
perature, solar cell temperature, solar irradiance, wind speed, open-circuit voltage,
short-circuit current as input variables to different ANN models for predicting max-
imum power of PV modules. Based on literature survey it is found that prediction
accuracy of PV maximum power is affected by different input variables and ANN
models. Therefore, the objectives are (1) To use relief attribute evaluation for the
selection of relevant input variables for PV module maximum power prediction,
(2) To develop cascade ANN for prediction of maximum PV power under outdoor
condition.

2 Methodology

2.1 Experimental Setup

A 74 Wp Si multi crystalline PV module is used in this study. The electrical charac-

teristics of PV module at STC given by the manufacturer Maharishi Solar, India are
shown in Table 1. The outdoor experimental setup for carrying out outdoor measure-
ments of a 74 Wp PV module is shown in Figs. 1 and 2. The experimental setup is

Table 1 Specification of PV
Module type SS 74P
module
Rated peak power (Pmax ) 74 W
Rated voltage (V mp ) 17.86 V
Rated current (I mp ) 4.19 A
Open-circuit voltage (Voc ) 21.89 V
Short-circuit current (Isc ) 4.48 A
390 V. K. Reddy Chimmula et al.

Fig. 1 Experimental setup

Fig. 2 DHT11 sensor

developed at Electrical and Electronics Engineering Department, National Institute

of Technology Sikkim. DHT11 Sensor is used to measure temperature and humidity.
Novel Application of Relief Algorithm in Cascade ANN … 391

Fig. 3 Experimental setup of measuring temperature and humidity

2.2 Arduino and DHT Sensor

2.2.1 Arduino-Based Implementation

Arduino is an open-source electronics platform. Arduino is designed to make elec-

tronics more accessible to everyone like artists, designers, hobbyists, and anyone
interested in creating interactive objects or environments. In this chapter, we used
arduino for measuring temperature and humidity, and the experimental setup is shown
in Fig. 3. Arduino attracts both academia and industry because of its simplicity and
simple code. For storing code it has 32 kb flash memory. The Arduino Uno board
is a microcontroller based on the ATmega328. It has 14 digital input/output pins in
which 6 can be used as PWM outputs: a 16 MHz ceramic resonator, an ICSP header,
a USB connection, 6 analog inputs, a power jack and a reset button.

2.2.2 DHT 11 Sensor

Measurement and control of temperature and relative humidity finds applications in

numerous areas. The use of smart sensors greatly simplifies the design and complexity
of experiment setup which in turn reduces the overall cost. These smart sensors
are capable of measuring both temperature and relative humidity on the same time.
Moreover, they will provide fully calibrated digital outputs, surprisingly these sensors
are relatively lot cheaper than the other sensors.
The DHT11 sensor comes in a single row 4-pin package and operates from 3.5 to
5.5 V power supply. It can measure temperature from 0 to 50 °C with an accuracy of
±2 °C and relative humidity ranging from 20 to 95% with an accuracy of ±5%. The
392 V. K. Reddy Chimmula et al.

Fig. 4 DHT 11 sensor

sensor provides fully calibrated digital outputs for the two measurements. Physical
appearance of DHT 11 sensor is shown in Fig. 4.

2.3 Database of Experimental Readings

The I-V characteristics were taken based on the developed setup are shown in Table 2.

2.4 Relief Attribute Evaluator

Relief is an algorithm developed by Kira and Rendell in 1992 that takes a filter-method
approach to feature selection that is notably sensitive to feature interactions. It was
originally designed for application to binary classification problems with discrete or
numerical features. Relief calculates a feature score for each feature which can then be
applied to rank and select the top scoring features for feature selection. Alternatively,
these scores may be applied as feature weights to guide downstream modeling. Relief
feature scoring is based on the identification of feature value differences between
the nearest neighbor instance pairs. If a feature value difference is observed in a
neighboring instance pair with the same class (a“hit”), the feature score decreases.
Alternatively, if a feature value difference is observed in a neighboring instance pair
with different class values (a “miss”), the feature score increases. The original Relief
algorithm has since inspired a family of Relief-based feature selection algorithms
(RBAs), including the Relief algorithm Beyond the original Relief algorithm, RBAs
have been adapted to (1) perform more reliably in noisy problems, (2) generalize to
Novel Application of Relief Algorithm in Cascade ANN … 393

Table 2 Sunny condition

S. Date Time Temperature Lux Humidity I sc V oc Pmax
no readings
1 02-03-18 10.05 17 642 48 3.4 19.68 47.49
2 02-03-18 10.15 18 614 47 3.38 19.69 43.84
3 02-03-18 10.26 18 634.2 44 3.46 19.66 39.51
4 03-03-18 10.15 18.5 640.2 41 3.464 19.7 39.56
5 03-03-18 10.27 20 798.4 41.5 4.35 19.58 57.13
6 03-03-18 10.38 19 795.8 42 4.15 19.52 54.82
7 03-03-18 10.51 19 898.8 42 4.3 19.68 51.26
8 03-03-18 11.06 20.5 853.2 41 4.41 19.63 54.84
9 04-03-18 10.25 20 777.4 41.2 4.43 19.62 46.37
10 04-03-18 10.36 20 841 41 4.13 19.69 55.12
11 04-03-18 10.46 20.8 834.6 42 4.08 19.63 57.03
12 04-03-18 10.58 21 819.4 41 4.02 19.59 51.12
13 04-03-18 11.08 19 818 41.4 3.921 19.58 39.78
14 04-03-18 11.18 19 811.6 42 3.7 19.58 29.35
15 06-03-18 10 18.5 811.8 42.3 3.87 19.56 53.99
16 06-03-18 10.11 17 760 42 3.443 19.51 43.29
17 06-03-18 10.21 17 785 42.7 3.42 19.51 23.52
18 06-03-18 10.31 18 792 42.4 3.45 19.49 61.95
19 07-03-18 10.15 18.5 753 42.1 3.88 19.72 43.52
20 07-03-18 10.25 18.5 773 41 3.69 19.5 38.77
21 07-03-18 10.36 16.5 684.6 41.9 3.68 19.61 45.47
22 07-03-18 10.5 17.5 774 41.6 4.1 19.77 50.77
23 07-03-18 11.01 17 742 41 4.26 19.59 39.31
24 07-03-18 11.16 15.5 772 39.9 3.84 18.97 42.52
25 08-03-18 10.34 17.5 748 40 4.26 19.97 45.06
26 08-03-18 10.45 19 796 41 4.096 19.66 35.50
27 08-03-18 10.55 20 768 39 4.397 19.96 48.86
28 08-03-18 11.05 17 626 40.2 3.489 18.97 35.51
29 08-03-18 11.16 19 724 40 4.29 19.89 48.95
30 11-03-18 10.07 18 744 43 4.097 19.698 45.96
31 11-03-18 10.17 17.5 672 42.1 4.197 19.798 41.38
32 11-03-18 10.28 17 744 42 4.239 19.509 37.27
33 11-03-18 10.41 17.5 715 44 4.197 19.397 44.08
34 11-03-18 10.51 18 744 43.7 4.199 19.399 41.39
394 V. K. Reddy Chimmula et al.

multi-class problems (3) generalize to numerical outcome (i.e., regression) problems,

and (4) to make them robust to incomplete (i.e., missing) data.
To date, the development of RBA variants and extensions has focused on four
areas: (1) improving performance of the “core” Relief algorithm, i.e., examining
strategies for neighbor selection and instance weighting, (2) improving scalability
of the “core” Relief algorithm to larger feature spaces through iterative approaches,
(3) methods for flexibly adapting Relief to different data types, and (4) improving
Relief run efficiency.
Their strengths are that they are not dependent on heuristics, they run in low-order
polynomial time, and they are noise-tolerant and robust to feature interactions, as well
as being applicable for binary or continuous data; however, it does not discriminate
between redundant features, and low numbers of training instances fool the algorithm.

2.5 Artificial Neural Network

The term “Neural” is derived from the human (animal) nervous system’s basic func-
tional unit “neuron” or nerve cells which are present in the brain and the other parts of
the human (animal) body. Artificial neural networks are the biologically inspired sim-
ulations performed on the computer to perform certain specific tasks like clustering,
classification, pattern recognition, etc.
Artificial neural networks, in general — is a biologically inspired network of arti-
ficial neurons configured to perform specific tasks. We will start with understanding
the formulation of a simple hidden layer neural network. A simple neural network
can be represented as shown in the Fig. 5.

Fig. 5 ANN
Novel Application of Relief Algorithm in Cascade ANN … 395

Fig. 6 ANN block diagram

The linkages between nodes are the most crucial finding in an ANN. The only
known values in the above diagram are the inputs. The inputs as I 1 , I 2, and I 3 ; hidden
states as H 1 , H 2 , H 3, and H 4 ; outputs as O1 and O2 . The weights of the linkages
can be denoted with the following notation, W (I 1 H 1 ) is the weight of the linkage
between I 1 and H 1 nodes (Fig. 6).

2.6 Cascade Artificial Neural Network (CANN)

Cascade correlation neural networks are “self-organizing” networks. The network

begins with only input and output neurons. During the training process, neurons are
selected from a pool of candidates and added to the hidden layer.
A cascade correlation network consists of a cascade architecture, in which hidden
neurons are added to the network one at a time and do not change after they have
been added. As new neurons are added to the hidden layer, the learning algorithm
attempts to maximize the magnitude of the correlation between the new neuron’s

Fig. 7 Network with only

input and output layer
neurons with no hidden layer
neurons
396 V. K. Reddy Chimmula et al.

Fig. 8 After the addition of

the first hidden neuron,
network structure

output and the residual error of the network which we are trying to minimize (Figs. 7
and 8).
Input Layer: A vector of predictor variable values (x 1 …x p ) is presented to the
input layer. The input neurons perform no action on the values other than distributing
them to the neurons in the hidden and output layers.
Hidden Layer: Arriving at a neuron in the hidden layer, the value from each input
neuron is multiplied by a weight, and the resulting weighted values are added together
producing a combined value. The weighted sum is fed into a transfer function, which
outputs a value. The outputs from the hidden layer are distributed to the output layer.
Output Layer: In this, there is only a single neuron in the output layer. Each
output neuron receives values from all of the input neurons and all of the hidden
layer neurons. Each value presented to an output neuron is multiplied by a weight,
and the resulting weighted values are added together producing a combined value.
The weighted sum is fed into a transfer function, which outputs a value. The y values
are the outputs of the network.

3 Results and Discussions

3.1 Relevant Input Variable Selection Using Relief Attribute

Evaluator

The relevant input variable selection is a vital part for cascade ANN model in the
prediction of maximum power of PV Module. The variables Temperature (T), Solar
Radiation (S.R), Humidity (H), Short-Circuit Current (Isc), Open-Circuit Voltage
(Voc) are used as input and maximum power (Pmax) as the target. The relief value
attribute is implemented using Matlab 2017 and its rank is given in Table 3. It is found
that humidity is having lower rank showing less influencing variable. In order to check
the authentication of the relief attribute evaluator for relevant input variable selection,
five cascade artificial neural networks (CANN-1, CANN-2, CANN-3, CANN-4, and
Novel Application of Relief Algorithm in Cascade ANN … 397

Table 3 Rank of input

S. no Input variables Ranks
variables using relief attribute
evaluator 1 Ambiant temperatue (T) 0.4633
2 Solar radiation (S.R) 0.3531
3 Humidity (H) −0.0361
4 Short-circuit current (I oc ) 0.3451
5 Open-circuit voltage (Voc ) 0.2215

CANN-5) are developed and its prediction accuracy is compared with conventional
models. The input Variables to CANN models are shown in Table 4 and its prediction
accuracy is given in Table 5.
The performance plot of ANN-1, ANN-2, ANN-3, ANN-4, and ANN-5 are shown
in Figs. 9, 10, 11, 12, and 13. The performance plot shows that mean square error
(MSE) has decreasing pattern with epochs, showing networks as minimum error
during training. The epochs are a measure of number of times of training data used
once to update the weights on bias of neural networks.
The error histogram plot of CANN-1, CANN-2, CANN-3, CANN-4, and CANN-
5 are shown in Figs. 14, 15, 16, 17, and 18. The comparison between measured and
predicted maximum power by CANN-1, CANN-2, CANN-3, CANN-4, and CANN-
5 are shown in Figs. 19, 20, 21, 22 and 23 showing that predicted value is close to
measured value.

Table 4 Cascade ANN model input variables

Models Input variables
CANN-1 Temperature, solar radiation, short-circuit current, open-circuit voltage
CANN-2 Temperature, solar radiation
CANN-3 Short-circuit current, open-circuit voltage
CANN-4 Solar radiation, short-circuit current, open-circuit voltage
ANN-5 Solar radiation, temperature, open-circuit voltage

Table 5 ANN model accuracy

Models Input variables RMSE training RMSE testing
CANN-1 Temperature, solar radiation, short-circuit 0.39 0.19
current, open-circuit voltage
CANN-2 Temperature, solar radiation 0.35 0.22
CANN-3 Short-circuit current, open-circuit voltage 0.32 0.17
CANN-4 Solar radiation, short-circuit current, 0.44 0.28
open-circuit voltage
CANN-5 Solar radiation, temperature, open-circuit 0.55 0.28
voltage
398 V. K. Reddy Chimmula et al.

Fig. 9 Performance plot of CANN-1

Fig. 10 Performance plot of CANN-2

Novel Application of Relief Algorithm in Cascade ANN … 399

Fig. 11 Performance plot of CANN-3

Fig. 12 Performance plot of CANN-4

400 V. K. Reddy Chimmula et al.

Fig. 13 Performance plot of CANN-5

Fig. 14 Error histogram plot of CANN-1

Novel Application of Relief Algorithm in Cascade ANN … 401

Fig. 15 Error histogram plot of CANN-2

Fig. 16 Error histogram plot of CANN-3

402 V. K. Reddy Chimmula et al.

Fig. 17 Error histogram plot of CANN-4

Fig. 18 Error histogram plot of CANN-5

Novel Application of Relief Algorithm in Cascade ANN … 403

Fig. 19 Comparison between measured and predicted maximum power for CANN-1

Fig. 20 Comparison between measured and predicted maximum power for CANN-2

Fig. 21 Comparison between measured and predicted maximum power for CANN-3
404 V. K. Reddy Chimmula et al.

Fig. 22 Comparison between measured and predicted maximum power for CANN-4

Fig. 23 Comparison between measured and predicted maximum power for CANN-5
Novel Application of Relief Algorithm in Cascade ANN … 405

4 Conclusions

In this study, an overview of ANN methods for PV module power prediction is

provided. Based on the research gap experimental analysis for 74Wp PV module
under outdoor conditions and relevant variable selection using Relief algorithm is
presented. The electrical characteristics of 74Wp PV module are performed under
sunny and partially cloudy condition. Relief algorithm shows that Temperature, Solar
Radiation, Short-Circuit Current, Open-Circuit Voltage are relevant variables and
humidity is a less influencing variable for Cascade ANN-based prediction of PV
module maximum power. Based on relevant variables cascade ANN models are
developed. Cascade ANN model is used to predict maximum power under sunny
outdoor condition. Cascade ANN model which utilizes open-circuit voltage and
short-circuit current has least root mean square error of 0.17, showing these variables
can be used for prediction of maximum power.
Future work is focused on the influence of dust on electrical characteristics of PV
module under outdoor condition.

References

1. A.B.G. Bahgat, N.H. Helwa, G.E. Ahamd, E.T.E. Shenawy, Estimation of the maximum power
and normal operating power of a photovoltaic module by neural networks. Renew. Energy 29,
443–457 (2004)
2. F. Almonacid, C. Rus, L. Hontoria, M. Fuentes, G. Nofuentes, Characterisation of Si-crystalline
PV modules by artificial neural networks. Renew. Energy 34, 941–949 (2009)
3. F. Almonacid, E.F. Fernández, P. Rodrigo, P.J. Pérez-Higueras, C. Rus-Casas, Estimating the
maximum power of a High Concentrator Photovoltaic (HCPV) module using an artificial neural
network. Energy 53, 165–172 (2013)
4. S.I. Sulaiman, T.K.A. Rahman, I. Musirin, Partial evolutionary ANN for output prediction of a
grid-connected photovoltaic system. Int. J. Comput. Electr. Eng. 1(1), 40–45 (2009)
5. F. Bonanno, G. Capizzi, G. Graditi, C. Napoli, G.M. Tina, A radial basis function neural network
based approach for the electrical characteristics estimation of a photovoltaic module. Appl.
Energy 97, 956–961 (2012)
6. A. Mellit, A.M. Pavan, Performance prediction of 20 kWp grid-connected photovoltaic plant at
Trieste (Italy) using artificial neural network. Energy Convers. Manag. 51, 2431–2441 (2010)
7. A. Mellit, S. Sağlam, S.A. Kalogirou, Artificial neural network-based model for estimating the
produced power of a photovoltaic module. Renew. Energy 60, 71–78 (2013)
8. V.L. Brano, G. Giuseppina Ciulla, M.D. Falco, Artificial neural networks to predict the power
output of a PV panel. Int. J. Photo Energy 1, 12 (2014)
Data-Driven Intelligent Model for Sale
Price Prediction and Monitoring
of a Building

Nuzhat Fatema, Hasmat Malik and Atif Iqbal

Abstract The construction cost forecasting and monitoring plays an important role
in a building condition assessment. The construction cost of a building (CCB) not
only depends on the method of construction, equipment, labor, and material but also
depends on type, scheduling, project locality, and project duration, etc. Moreover,
abrupt variations in economic indices and attributes (i.e., WPI: wholesale price,
liquidity, building services, etc.) are reasons for cost variation and deviate the CCB,
which are not possible to monitor and/or identify in easy way during the current
economic scenario. So, these indices may be snubbed in CCB. In this chapter, a
data-driven intelligent model for sale price monitoring and detection of a building is
presented which may be utilized in hospital planning. For the implementation of the
proposed approach, the cost’s data of construction for 372 buildings of three to nine
stories have been utilized. The recorded dataset has physical, financial, and economic
variables and indices of real sites. The proposed approach includes the comparative
analysis of conventional statistical and advanced soft computing techniques. The
obtained results show that monitor and/or identification of CCB is higher in case of
soft computing technique than statistical method.

Keywords Building · Construction cost · Monitoring · Identification · Prediction ·

ANN · WEKA · Machine learning · Linear regression

N. Fatema (B)
Singapore Polyclinic, Singapore, Singapore
e-mail: [email protected]
H. Malik
BEARS, University Town, NUS Campus, Singapore, Singapore
e-mail: [email protected]
Division of ICE, NSUT, Delhi, India
A. Iqbal
Department of Electrical Engineering, Qatar University, Doha, Qatar
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 407

H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_18
408 N. Fatema et al.

Nomenclature

ANN Artificial neural network

ARIMA Autoregressive integrated moving average
CCB Construction cost of building
EV&I Economic variables and indexes/indices
$ Dollar

1 Introduction

Estimation of the sale price of a building is the method of bringing value to a property
in the real market [1]. Accurate and qualitative building construction costs estimation
is a tedious job in the field of construction management, which is usually done by
skilled construction cost estimators may lead to some bias [2, 3]. It is necessary
to overcome this bias by evaluating the factors which interfere with the building
cost estimation phenomenon [4]. Efficient building cost estimation is not only cost-
effective but it also decreases the thread of project budget overrun [5].
Building cost estimation is directly related to the construction project evaluation
which has five phases, i.e., planning phase, designing phase, bidding phase, letting
phase, and construction phase as shown in Fig. 1. Each phase requires budgeting;
therefore, cost estimation is done at each phase. At planning phase, the agency eval-
uates size, scope, type, and location of the project. Public and stake holder’s opinions

Fig. 1 Project cost evaluation phases

Data-Driven Intelligent Model for Sale Price Prediction … 409

are also considered at this phase. If all these factors are in favor, then the project gets
selected for designing phase where project design is approved for bidding phase.
Using different bidding methods, building project is awarded to efficient bidders.
The final contract is given at letting phase and hereafter, actual building construction
begins.
If building cost estimation up to phase 4 (letting phase) approximately equals to
the actual construction cost of the building, then it can be stated that project runs
smoothly; on the other hand, if it exceeds the actual construction cost, the project is
overrun and if it falls short of underrun accordingly.
Generally, different cost estimating models have been developed such as arith-
metically [6], regression analysis [7], quantity-based approach [8–10], Case-based
reasoning and genetic algorithms [11], conceptual cost estimates [12], and by risk
consistency analysis at design phase [13]. Eventually, all these conventional meth-
ods have many disadvantages [14]. Moreover, regression approach lacks in defined
mathematical form. Therefore, various artificial intelligence applications are utilized
by researchers for building cost estimation at planning phase [15].
AI-based model was proposed using preexisting knowledge to build up a knowl-
edge base assessment system [16]. A proficient model was presented for decision-
making criteria utilized for different cost estimating packages, mentioned in [17].
An analytic hierarchy procedure was utilized for developing a case-based reasoning
model which integrates experience [18]. A case-based reasoning model was applied
for estimating construction duration and different costs at early design phase [19]. A
neural network model was proposed for highway construction costs estimation [20].
Highway construction cost estimator was employed using NNs and also, recognized
noise in data [1].
Most of AI methods are straightway implemented without explanation of math-
ematical relation. So, it is very difficult to understand by a new researcher. In this
chapter, a step-by-step mathematical implementation for ANN approach has been
presented and validated by using simulated database.
The organization of the chapter is comprised into five sections as follows: Sect. 1
represents the introduction and Sect. 2 shows the used database and its attribute’s
detail. The methodology for the study has been shown in Sect. 3, which includes
proposed approach along with mathematical modeling of MLP-based NN. Section 4
represents the results and discussion. Finally, Sect. 5 shows the conclusion and future
scope of the study.

2 Brief Detail of Dataset

Dataset used in this study is available at open-source data center [21], which includes
372 buildings data of 103 predictor variables for two predictable attributes. The
two predictable attributes are “actual sale price” and “actual construction cost” in
Rial (IRR—Irianian currency) of the building, whereas, all predictor variables are
410 N. Fatema et al.

created/grouped as per six subcategories (i.e., project physical and financial variables,
and five-type economic variables and indices-EV&I) as shown in Fig. 2.
The EV&I of five different time lags are represented. 1–5 time lag number
represents the number of time resolutions before the start of the project.

Fig. 2 Predictor (103) and predictable (01) variables representation

Data-Driven Intelligent Model for Sale Price Prediction … 411

3 Methodology

3.1 Proposed Approach

The proposed approach algorithm for sale price monitoring of a building is

represented in Fig. 3, which includes seven simple steps as mention below:
Step #1: measurement of multivariable dataset of buildings

Fig. 3 Flowchart for proposed approach

412 N. Fatema et al.

Step #2: training and testing file formation of the dataset

Step #3: design the artificial/machine learning model
Step #4: set the optimal parameters of the AI/ML model
Step #5: train the model
Step #6: test the model
Step #7: save the AI/ML model for future prospective
For an example, the artificial neural network is designed and implemented along
with other machine learning approaches, i.e., linear regression (LR) and decision
table (DT). Numbers of neuron at input and output layers are equal to the number
of input and output variables, respectively. Moreover, number of neurons at hidden
layer is evaluated as per Eq. (1) [22]:
 
N o. I nput V ariables + N o. Out put V ariables 
N H idden Layer = + N o. Samples ± 10%
2
(1)

The brief detail of AI/ML approaches used in this study is explained in the
subsequent sections.

3.2 Implementation of Artificial Neural Network (ANN)

The mathematical modeling for multilayer perceptron (MLP) neural network has
been represented in four basic steps as given below [22–26]:
Step #1: Assume some basic assumptions:
Let us assume χ = [x1 , x2 , . . . , xn ]T = an input variables with n linear nodes and
a single bias node,
L = sigmoidal neurons a bias node in the hidden layer, and Oj sigmoidal neurons
in the output layer.
β I = [wiIh ](n+1)×q weights between the input layer and the hidden layer
β H = [whHj ](q+1)× p weights between the hidden layer and the output layer
The MLP-based NN architecture is shown in Fig. 3, which is as per [22–26],
Step #2: Mathematical modeling at input Layer:

S(xi ) = xi , i = 1, 2, . . . , n; (2)

S(x0 ) = 1. (3)

where S(·) denotes the activation function of a neuron.

Step #3: Mathematical modeling at hidden layer: The evaluation at the Lth neuron
of this layer is
Data-Driven Intelligent Model for Sale Price Prediction … 413


n
Computation at input site of hidden layer neuron; z h = wiIh S(xi ), h = 1, 2, . . . , q (4)
i=0
1
Computation at output site of hidden layer neuron; S(z h ) = , h = 1, 2, . . . , q
1 + e−z h
(5)

S(z 0 ) = 1. (6)

Step #4: Mathematical modeling at output layer: The evaluation at the jth neuron
of this layer is

q

Computation at input site of hidden layer neuron; y j = whHj S(z h ), j = 1, 2, . . . , p
h=0
(7)
1
Computation at output site of output layer neuron; S(y j ) = , j = 1, 2, . . . , p
1 + e−y j
(8)

If there are multiple outputs, then p will exist.

4 Results and Discussion

4.1 Training and Testing Dataset Formation

Collected dataset of 372 buildings with 103 attributes has been divided into two parts,
i.e., training and testing data part. Training data file includes 70% and testing data
file includes 30% samples of the total available samples. Both training and testing
data files are in separate files and are different from each other. During training
phase of the MLP NN, only training dataset is utilized and after successful training
with acceptable model’s performance, testing is performed to validate the model
performance. Then, the prepared model can be used for future prospective to predict
the sale price of the building.

4.2 Sale Price Interpretation and Monitoring

An MLP neural network for sale price prediction and monitoring has been designed
with its optimal parameters as mentioned below in Table 1.
The graphical results for sale price prediction using MLP NN have been repre-
sented in Figs. 4, 5, 6 and 7. Figure 4 shows the performance plot for MLP model
during training, testing, and validation phase. Error histogram shows the available
 414

Table 1 Optimal parameters of MLP NN

Parameter name Value
Number of input layer neurons 103
Number of hidden layer neurons 10 or selected as per Eq. (1)
Number of output layer neurons 1
Learning algorithm Levenberg–Marquardt
Set number of epoch (max) 1000
Learning rate 0.5
Momentum constant 0.3
Number of training data samples 280
Number of testing data samples 92
Bias value at layer 1 (b1) [−0.28260298992075467; −1.103883310691816; 1.121859900686782; −1.2089256256918919; −
0.063069727123333932;0.29507169256276622;0.53429006754056907; 0.41064570638280212;
2.2137558415187364; 1.2900418148692734]
Bias value at layer2 (b2) −0.43191233146634084
Weight at layer1* * big matrix, available on request
Weight at layer2 LW2_1 = [0.013802120993594843 −0.33378909097986481 0.3244380781339184 0.096493500759377807
0.17814059109081137 0.3451447825577228 0.11246626372349791 −0.13169098789195643 0.057632768893189593
−0.050034411805829543]
N. Fatema et al.
Data-Driven Intelligent Model for Sale Price Prediction … 415

Fig. 4 Plot performance

Fig. 5 Plot error histogram

Fig. 6 Plot train state

416 N. Fatema et al.

Fig. 7 a–d Plot regression

error in the model for training, testing, and validation datasets (Fig. 5). Plot train state
is shown in Fig. 6. It includes the training history of gradient method, momentum
value, and validation checks for each training epoch. The regression plot for training
phase is shown in Fig. 7. It includes the internal information during training phase
which shows the overall performance of the model. In this case, training performance
R is equal to 0.99845 which shows that model’s performance during training phase
is very closer to target dataset. Figure 8 shows the further validation for the trained
proposed model. It shows the testing phase performance. The testing phase perfor-
mance R is equal to 0.94506 which shows that prediction model is acceptable for
prediction of an unknown dataset for the future prospective.
The result comparison for each sample with the actual value and predicted value
by MLP model has been represented in Fig. 9(a) and Fig. 9(b) for training and
testing phase, respectively, which shows a very close relationship between actual
and predicted value.
Moreover, for further fair cross-validation of the performance of the MLP NN
model, a decision table (DT) based machine learning and linear regression (LR)
approach have been implemented by using the same training data sample and have
Data-Driven Intelligent Model for Sale Price Prediction … 417

Fig. 8 Testing phase plot for a Error histogram and b Regression plot

Fig. 9 Result comparison using MLP method for a Training phase of 280 data samples b Testing
phase of 92 data samples

been tested with testing dataset of 92 samples. Both approaches have been imple-
mented in WEKA platform. For more understanding for WEKA implementation,
reader may refer [27, 28]. The detailed results obtained from RL and DT have
been represented in Table 2 and its graphical representation in Fig. 10 and Fig. 11,
respectively, for both training and testing phases.
The comparative result analysis for both DT and RL with MLP NN has been
represented in Table 3. After analyzing Table 3, it is concluded that MLP NN has

Table 2 Performance evaluation representation of DT and LR model

S. no. Performance Performance of RL Performance of DT
indices Training phase Testing phase Training phase Testing phase
1 Correlation 0.9868 0.8781 0.9364 0.8808
coefficient
2 Total number 280 92 280 92
of instances
418 N. Fatema et al.

Fig. 10 Result comparison using RL method for a Training phase of 280 data samples b Testing
phase of 92 data samples

Fig. 11 Result comparison using DT method for a Training phase of 280 data samples and b Testing
phase of 92 data samples

Table 3 Performance evaluation comparison

Sample number Sale price prediction
Actual ANN DT RL
Sample #1 2200 2191.669 1191.705 1942.658
Sample #2 5000 4937.016 4210 4479.743
Sample #3 1200 1232.991 485.505 1229.168
Sample #4 165 204.172 485.505 168.561
Sample #5 5500 5421.771 5683.333 6163.792
Sample #6 5200 5233.55 5683.333 5879.924
Sample #7 5800 6382.297 5200 6475.101
Sample #8 4600 4576.318 5200 5117.548
Sample #9 870 876.95 1191.705 869.329
Sample #10 1700 1701.958 1925.714 1745.374
Data-Driven Intelligent Model for Sale Price Prediction … 419

Fig. 12 Distribution using 100 bins with kernel smoothing function fit

efficient performance for sale price prediction of the building. Moreover, a distribu-
tion function of predicted value by each approach has been analyzed as shown in
Fig. 12 for training and testing phase performance. Distribution function using 100
bins with kernel smoothing function fit has been analyzed, which shows that further
validation of the proposed MLP approach is comparable to others and can be used
for future prospective.

5 Conclusions

In this chapter, a data-driven intelligent model for sale price estimation and mon-
itoring has been presented using MLP NN for a building. A standard open-access
database from UCI machine learning repository has been utilized for validation of
proposed approach. First, a method for project cost evaluation has been presented,
then MLP-based NN has been implemented which has seven simple steps. For the
fair validation of the performance of proposed approach, two WEKA based machine
learning approaches (i.e., LR and DT) have been implemented using the same train-
ing and testing dataset. Obtained results from LR and DT have been compared with
the proposed approach, which shows the effectiveness of the MLP NN for the sale
price estimation and monitoring.
The future scope of the study is to implement this approach in the real site and
also, enhance the performance accuracy by pruning the number of input variables.

Acknowledgements Authors are very thankful to Mohammad Hossein Rafiei et al. [22] for pro-
viding open-access dataset at UCI machine learning repository. Authors are also very thankful to
the reviewers for providing valuable suggestions in this regard.
420 N. Fatema et al.

References

1. Hoijat Adelil, Wu Mingyang, Regularization neural network for construction cost estimation.
J Construction Engineering and Management 124(1), 18–24 (1998)
2. Hoijat Adelil, Asim Karim, Scheduling/cost optimization and neural dynamics model for
construction. Journal of Construction Engineering and Management 123(4), 450–458 (1997)
3. Olga Perminova et al., Defining uncertainty in projects – a new perspective. Int. J. Project
Manage. 26, 73–79 (2008)
4. Dinu Philip et al., Artificial Neural Network Model for Cost Estimation: City of Edmonton’s
Water and Sewer Installation Services. Journal of construction engineering and management
136(7), 745–756 (2009)
5. Abhishek Bhargava et al., Predicting Cost Escalation Pathways and Deviation Severities of
Infrastructure Projects Using Risk-Based Econometric Models and Monte Carlo Simulation.
Computer-Aided Civil and Infrastructure Engineering 32, 620–640 (2017)
6. S. Singh et al., Cost model for reinforced concrete beam and slab structures in building. Journal
of Construction Engineering and Management 116(1), 54–67 (1990)
7. P.A. Bowen, P.J. Edwards, Cost modeling and price forecasting; practice and theory in
perspective. Construction Management and Economics 3, 199–215 (1985)
8. Sanders et al, “Preliminary estimating models for infrastructure projects”, Cost Engineering,
vol. 34(8), pp. 7–13, 1992
9. K. Kim et al., Preliminary cost estimation model using case-based reasoning and genetic
algorithms. Journal of Computing in Civil Engineering 24(6), 499–505 (2010)
10. Sharma et al., Preliminary cost estimation models for water treatment plants. Journal of
Infrastructure Systems 19(4), 451–464 (2013)
11. Asmar et al., New approach to developing conceptual cost estimates for highway projects.
Journal of Construction Engineering and Management 137(11), 942–949 (2011)
12. Stoy et al., Drivers for cost estimating in early design: case study of residential construction.
Journal of Construction Engineering and Management 134(1), 32–39 (2008)
13. Accuracy in design phase cost estimating through risk contingency analysis: a theoretical
exploration, in Proceedings of the 2014 Construction Research Congress, Atlanta, pp. 1478–87,
May 19–21, 2014
14. Min-Yuan Cheng et al., Conceptual cost estimates using evolutionary fuzzy hybrid neural
network for projects in construction industry. Expert Syst. Appl. 37, 4224–4231 (2010)
15. Mohamed and Celik, An integrated knowledge-based system for alternative design and
materials selection and cost estimating. Expert Syst. Appl. 14(3), 329–339 (1998)
16. Serpell et al., Towards a knowledge-based assessment of conceptual cost estimates. Building
Research and Information 32(2), 157–164 (2004)
17. D. Arditi, K. Suh, Expert system for cost estimating software selection. Cost Engineering 33(6),
9–19 (1991)
18. S.H. An, G.H. Kim, K.I. Kang, A case-based reasoning cost estimating model using experience
by analytic hierarchy process. Build. Environ. 42(7), 2573–2579 (2007)
19. Yau and Yang, Case-based reasoning in construction management. Computer-Aided Civil and
Infrastructure Engineering 13(2), 143–150 (1998)
20. C.G. Wilmot, B. Mei, Neural network modeling of highway construction Costs. Journal of
Construction Engineering and Management 124(3), 210–218 (2005)
21. Rafiei, M.H. and Adeli, H., “A Novel Machine Learning Model for Estimation of Sale Prices
of Real Estate Units”, ASCE, Journal of Construction Engineering & Management, 142(2),
pp. 2015. 04015066–1 to 10. https://doi.org/10.1061/(asce)co.1943-7862.0001047
22. H. Malik, S. Mishra, Artificial Neural Network and Empirical Mode Decomposition Based
Imbalance Fault Diagnosis of Wind Turbine Using TurbSim, FAST and Simulink. IET Renew.
Power Gener. 11(6), 889–902 (2017). https://doi.org/10.1049/iet-rpg.2015.0382
23. H. Malik, R. Sharma, EMD and ANN based Intelligent Fault Diagnosis Model for Transmission
Line. Journal of Intelligent & Fuzzy Systems 32(4), 3043–3050 (2017). https://doi.org/10.3233/
JIFS-169247
Data-Driven Intelligent Model for Sale Price Prediction … 421

24. A. Azeem, N. Fatema, H. Malik, k-NN and ANN Based Deterministic and Probabilistic Wind
Speed Forecasting Intelligent Approach. Journal of Intelligent & Fuzzy Systems 35(5), 5021–
5031 (2018). https://doi.org/10.3233/JIFS-169786
25. AK.Shah and A.Yadav and H.Malik, “EMD and ANN Based Intelligent Model for Bearing
Fault Diagnosis”, Journal of Intelligent & Fuzzy Systems, vol. 35, no. 5, pp. 5391–5402, 2018,
2018. https://doi.org/10.3233/jifs-169821
26. A.K. Yadav, H. Malik, S.S. Chandel, Application of Rapid Miner In ANN Based Prediction Of
Solar Radiation for Assessment of Solar Energy Resource Potential of 76 Sites in Northwestern
India. Renew. Sustain. Energy Rev. 52, 1093–1106 (2015). https://doi.org/10.1016/j.rser.2015.
07.156
27. H. Malik, S. Mishra, A.P. Mittal, Selection of Most Relevant Input Parameters Using Waikato
Environment for Knowledge Analysis for Gene Expression Programming Based Power Trans-
former Fault Diagnosis. International Journal of Electric Power Components and Systems
42(16), 1849–1862 (2014). https://doi.org/10.1080/15325008.2014.956952
28. A.K. Yadav, H. Malik, S.S. Chandel, Selection of Most Relevant Input Parameters Using WEKA
for Artificial Neural Network Based Solar Radiation Prediction Models. Renew. Sustain.
Energy Rev. 31, 509–519 (2014). https://doi.org/10.1016/j.rser.2013.12.008
Artificial Neural Networks Based
Condition Monitoring of Air Pollutants
for Allahabad Cities in India

Vibha Yadav and Satyendra Nath

Abstract Air pollution produces major influences on human health and environ-
ment. Therefore, it is necessary to forecast air pollutant (AP) data to provide prior
information about its concentration in environment for health monitoring. In this
study, six ANN models are developed to forecast PM10 , NO2 , and SO2 at 1 month
ahead for two sites Laxmi Talkie, Bharat Yantra Nigam located at Allahabad, India.
For this monthly average time series, data of PM10 , SO2 , and NO2 during 2013–2017
are taken from Uttar Pradesh Pollution Control Board (UPCB) Lucknow. ANN mod-
els are developed with Levenberg–Marquardt (LM) algorithm. The root mean square
error for testing varies from 0.1714 to 0.2477, showing forecasted and measured air
pollutant concentrations, are close to each other.

Keywords PM10 · SO2 · NO2 · Forecasting · ANN

1 Introduction

Most urban areas in India especially experienced the severe effect of air pollution
rapidly. These airborne pollutants are PM10 , SO2 , CO, and NO2 which are of great
interest due to harmful health impacts. Rapid growth of population, industrialization,
vehicular sources, and many anthropogenic activities results to increase in concen-
tration of air pollutant. These pollutants arise either in form of primary pollutant or
may be secondary pollutant. SO2 is an irritating and colorless gas affecting healthy
peoples and causing effects like reduced visibility and acid deposition. NO2 is an

V. Yadav · S. Nath (B)

Department of Environmental Sciences and NRM, College of Forestry, Sam Higginbottom
University of Agriculture, Technology and Sciences, Allahabad 211007, Uttar Pradesh, India
e-mail: [email protected]
V. Yadav
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 423

H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_19
424 V. Yadav and S. Nath

irritating and reddish-brown gas that can damage lung and involve in the deposition
of acid. Suspended particulate matter suspended in atmosphere causes many lung
and reproductive disorders. Carbon monoxide is a colorless and odorless gas, nega-
tively affects the health, poisonous for animals, and destroys cells and tissues causing
oxygen deficient. The release of air pollutant directly or indirectly affects the plants
and animal, therefore, the accurate estimation of its concentration is important. Pre-
diction of air quality is needed to asses these types of emissions to take preventive
measures. Different practices and model developments for air quality management
have been done by various authors [1–6] and some are shown in Table 1.
From the above discussion, ANN is used to predict air pollutant for different sites
but for Indian sites, it remains a major research gap. In this study, six ANN models
are developed to forecast PM10 , SO2 , and NO2 at 1 month ahead for two sites Laxmi
Talkie, Bharat Yantra Nigam located at Allahabad, India.

Table 1 AP predictions
Study Types of air ANN type Input Highlights
pollutant parameters
Elminir and PM10 , NO2 , CO MLP WS, RH, T, and Accuracy varies
Galil [7] WD from 96 to
99.6%
Benjamin et al. NOx NN RH, T, air Mean square
[8] velocity, and error is 0.00223
rainfall (RF)
Golizadeh and CO MLP Meteorological R is 84.17%
Darand [9] variables (MV)
Mahmoudzadeh CO MLP and hybrid MV MSE varies
et al. [10] model from 0.0120 to
0.1924
Skrzypski and PM10 RBF and MLP WS, RF, SR, Accuracy is 3.7
Szakiel [11] vertical wind and 1.9% in
velocity, T and MLP and RBF
AP
Sahin et al. [12] SO2 and PM10 Cellular neural MV CNN model
network (CNN) predict SO2
better than PM10
Sonaje et al. Respirable MLP Four previous R is found to be
[13] suspended days RSPM 90.2%
particulate
matter (RSPM)
Artificial Neural Networks Based Condition Monitoring … 425

Table 2 Measured PM10 for Laxmi Talkie of different years

Months 2013 2014 2015 2016 2017
Jan (1) 348 308 248.5 295 203
Feb (2) 296 323 294 331 272
March (3) 258 261 231 269 292
Apr (4) 243 230 261 233 255
May (5) 205 282 263 289 220
June (6) 262 232 293 293 206
July (7) 215 267 266 147 137
Aug (8) 210 257 289 220 183
Sep (9) 224 252 268 224 146
Oct (10) 228 237 261 206 187
Nov (11) 255 268 262 296 266
Dec (12) 300 252 302 201 357
Average 253.7 264.0 269.8 250.3 227
Max 348 323 302 331 357
Min 205 230 231 147 137

2 Methodology

2.1 Data Collection

The selected area for the present study is Allahabad, District of Uttar Pradesh. Alla-
habad District is situated at 25° 28 N latitudes 81°50 E. at the meeting point of the
three sacred rivers of Ganges, Yamuna, and Saraswati and located in southeastern
of Uttar Pradesh. Monthly average time series data of PM10 , SO2 , and NO2 during
2013–2017 for Laxmi Talkie and Bharat Yantra Nigam, Allahabad, U.P India was
taken from Uttar Pradesh Pollution Control Board (UPCB) Lucknow. The values of
PM10 , SO2 , and NO2 for Laxmi Talkie are shown in Table 2, Table 3, and Table 4,
respectively. The values of PM10 , SO2 , and NO2 for Bharat Yantra Nigam are shown
in Table 5, Table 6 and Table 7, respectively.

2.2 Artificial Neural Network

ANN models are used for prediction [7]. Mainly, it incorporates three layers namely
input, hidden, and output. The training algorithm is Levenberg–Marquardt (LM).
426 V. Yadav and S. Nath

Table 3 Measured SO2 for Laxmi Talkie of different years

Months 2013 2014 2015 2016 2017
1 7.8 7.8 4.2 4.3 5
2 5.2 6.9 3.8 3.9 5
3 5.4 4.6 4.4 2.6 9.7
4 5.6 4.8 3.7 2.3 6.9
5 5.6 4.2 4 4.5 3.6
6 4.5 4 6 4 5
7 4.5 2.6 5 2.5 2.4
8 4.7 3 3 3.8 2.7
9 3.9 2.1 1.9 2.5 4.4
10 3 2.2 3.3 3.7 4.1
11 6.9 3.6 6.1 6.7 5.7
12 5.2 3.9 4.5 4.1 5.2
Average 5.1 3.9 4.2 3.74 4.9
Max 7.8 6.9 6.1 6.7 9.7
Min 3 2.1 1.9 2.3 2.4

Table 4 Measured NO2 for Laxmi Talkie of different years

Months 2013 2014 2015 2016 2017
1 44.1 32.5 39.8 46.3 52.2
2 39.3 28.3 42.5 49.3 45.1
3 41.1 28 35.2 36 39.1
4 38.6 28 33.4 31 34.2
5 40.2 23.7 16 27.6 30.9
6 39.4 27.5 24 24 30.4
7 20.3 24.5 20 24.2 26
8 13.4 23.8 19 31.2 22.9
9 15.9 22.7 20.4 30.5 24.7
10 26.1 42.3 19.8 39.8 27.4
11 25.1 42.5 41.4 59.3 38
12 38.4 40 40.6 73..4 34.8
Average 31.9 30.3 29.3 36.2 33.8
Max 44.1 42.5 42.5 59.3 52.2
Min 13.4 22.7 16 24 22.9
Artificial Neural Networks Based Condition Monitoring … 427

Table 5 Measured PM10 for Bharat Yantra Nigam of different years

Months 2013 2014 2015 2016 2017
1 247 274 216.4 238 157
2 252 314 236 253 199
3 205 246 203 227 234
4 198 198 234 199 210
5 176 240 233 242 169
6 217 209 254 246 156
7 186 246 239 121 121
8 206 242 235 147 147
9 218 214 228 114 114
10 269 219 233 159 159
11 236 231 230 207 199
12 256 206 260 190 233
Average 222 236 233.5 195.2 174.8
Max 269 314 260 253 234
Min 176 198 203 114 114

Table 6 Measured SO2 for Bharat Yantra Nigam of different years

Months 2013 2014 2015 2016 2017
1 5.4 3.4 2.2 3 4
2 3.5 5.2 3.7 3.1 4.2
3 4.3 3.6 3.5 1.9 7.6
4 4.5 3.7 3 1.6 4.5
5 4.1 3.3 2 3.6 3.8
6 2.9 3.2 5 3 4.1
7 3.2 1.6 4 2.2 2.2
8 4.1 2.4 3 3.2 2.3
9 3.4 1.9 1.9 2.4 5
10 1.4 1.4 4.2 2.7 4.1
11 5.4 3 5 5.6 4.7
12 4.1 2.1 3.9 3.2 4
Average 3.9 2.9 3.5 2.9 4.2
Max 5.4 5.2 5 5.6 7.6
Min 1.4 1.4 1.9 1.6 2.2
428 V. Yadav and S. Nath

Table 7 Measured NO2 for Bharat Yantra Nigam of different years

Months 2013 2014 2015 2016 2017
1 36 23.1 38.4 33.3 37.9
2 32.2 22.7 33.9 36.1 34.6
3 35.7 22.8 31.6 24.4 27.6
4 33.3 23.6 27.2 20.1 26.5
5 33.3 19.8 12 20.3 22.9
6 32.9 23.4 14 14.3 24.6
7 16.2 19.5 14 16.5 17.8
8 10.6 20.6 12 18.9 18.7
9 13.9 16 14.8 20.9 19.7
10 21.7 30.1 14 34.6 20.6
11 19.2 34.6 29.7 36.1 29.4
12 29 37.2 34.1 42.4 22.4
Average 26.2 24.5 23 26.4 25.2
Max 36 37.2 38.4 42.4 37.9
Min 10.6 16 12 14.3 17.8

In ANN, data sets are divided into training and testing. The majority of data is
incorporated in training. In this study, MATLAB code is implemented to develop
forecasting model. Hidden layer neurons are calculated by equation suggested by
Chow et al. [8].

3 Results and Discussions

The six ANN models (ANN-1, ANN-2, ANN-3, ANN-4, ANN-5, and ANN-6) are
obtained with MATLAB software (R 2017a). ANN-1, ANN-2, and ANN-3 are used
to forecast PM10 , SO2 , and NO2 at 1 month ahead for Laxmie Talkie. ANN-4, ANN-
5, and ANN-6 are used to forecast PM10 , SO2 , and NO2 at 1 month ahead for Bharat
Yantra Nigam. For training ANN models 48 data points (from year 2013 to 2016) are
used and for testing 12 data (year 2017) points are utilized from Tables 2, 3, 4, 5, 6
and 7. The statistical analysis of ANN models is shown in Table 8. The comparison
between measured and forecasted PM10 , SO2 , and NO2 for Laxmi Talkie and Bharat
Yantra Nigam are shown in Tables 9 and 10.
Artificial Neural Networks Based Condition Monitoring … 429

Table 8 Statistical performance measures for ANN models

Models Forecast air pollutant RMSE training RMSE testing
ANN-1 PM10 0.158 0.2300
ANN-2 SO2 0.1792 0.2377
ANN-3 NO2 0.1692 0.2477
ANN-4 PM10 0.1000 0.1098
ANN-5 SO2 0.2000 0.2333
ANN-6 NO2 0.1700 0.1714

Table 9 Comparison between normalized value measured and forecasted PM10 , SO2 , and NO2 of
Laxmi Talkie for year 2017
Months Measured Forecasted Measured Forecasted Measured Forecasted
PM10 PM10 NO2 NO2 SO2 SO2
Jan 0.5123 0.3 0.8453 0.357 0.3219 0.3974
Feb 0.4837 0.6136 0.6906 0.3469 0.5048 0.3974
March 0.6236 0.7045 0.5599 0.6311 0.5048 1
Apr 0.592 0.5364 0.4532 0.5029 0.7084 0.641
May 0.5566 0.3773 0.3813 0.3046 0.3341 0.2179
June 0.2429 0.3136 0.3704 0.2769 0.5164 0.3974
July 0.4079 0 0.2745 0.2796 0.5048 0.0641
Aug 0.2066 0.2091 0.207 0.3178 0.2399 0.1026
Sep 0.0372 0.0409 0.2462 0.2098 0.3475 0.3205
Oct 0.3022 0.2273 0.305 0.2709 0.3413 0.2821
Nov 0.2084 0.5864 0.5359 0.3258 0.3219 0.4872
Dec 0.6143 1 0.4662 0.4936 0.4153 0.4231
Average 0.39897 0.40909 0.44462 0.35974 0.42175 0.39422
Max 0.6236 1 0.8453 0.6311 0.7084 1
Min 0.0372 0 0.207 0.2098 0.2399 0.0641

Based on the range of neurons, training and testing are performed. Neurons at
which models give least error are used for forecasting. Performance plots are shown
in Figs. 1, 2, 3, 4, 5, and 6. It presents decrease in mean square error with increase
in epochs (complete sweep of training, testing, and validation). The error histogram
(EH) plots are shown in Figs. 7, 8, 9, 10, 11, and 12 which shows difference between
target and predicted value.
430 V. Yadav and S. Nath

Table 10 Comparison between normalized value measured and forecasted PM10 , SO2 , and NO2
of Bharat Yantra Nigam for year 2017
Months Measured Forecasted Measured Forecasted Measured Forecasted
PM10 PM10 NO2 NO2 SO2 SO2
Jan 0.4291 0.5862 0.8585 0.8765 0.2306 0.4194
Feb 0.5818 0.4683 0.7547 0.6995 0.3667 0.4516
March 0.7091 0.6197 0.5346 0.6092 0.4167 1
Apr 0.6218 0.7027 0.5 0.2938 0.5003 0.5
May 0.4727 0.6374 0.3868 0.2627 0.4468 0.3871
June 0.4255 0.5308 0.4403 0.4202 0.3393 0.4355
July 0.2982 0.4583 0.2264 0.3591 0.3916 0.129
Aug 0.3927 0.4078 0.2547 0.1858 0.3539 0.1452
Sep 0.2727 0.3661 0.2862 0.2491 0.366 0.5806
Oct 0.4364 0.417 0.3145 0.2886 0.4118 0.4355
Nov 0.5818 0.4862 0.5912 0.3415 0.3916 0.5323
Dec 0.7055 0.6197 0.3711 0.604 0.4424 0.4194
Average 0.7564 0.5250 0.4599 0.4325 0.3881 0.4529
Max 0.51413 0.7027 0.8585 0.8765 0.5003 1
Min 0.7564 0.3661 0.2264 0.1858 0.2306 0.129

Fig. 1 Performance Plot (PP) of ANN-1

Artificial Neural Networks Based Condition Monitoring … 431

Fig. 2 PP of ANN-2

Fig. 3 PP of ANN-3

4 Conclusion

In this study, six ANN models are proposed to forecast air pollutant at 1 month ahead.
ANN-1, ANN-2, and ANN-3 are used to forecast PM10 , SO2 , and NO2 at 1 month
ahead for Laxmi Talkie. ANN-4, ANN-5, and ANN-6 are used to forecast PM10 , SO2,
and NO2 at 1 month ahead for Bharat Yantra Nigam. The root mean square error
(RMSE) for training ANN-1, ANN-2, and ANN-3 models are 0.158, 0.1792, and
432 V. Yadav and S. Nath

Fig. 4 PP of ANN-4

Fig. 5 PP of ANN-5

0.1692, respectively, and RMSE for testing ANN-1, ANN-2, and ANN-3 models are
0.2300, 0.2377, and 0.2477, respectively. RMSE for training ANN-4, ANN-5, and
ANN-6 models are 0.1000, 0.2000, and 0.1700, respectively, and RMSE for testing
ANN-4, ANN-5, and ANN-6 models are 0.1098, 0.2333, and 0.1714, respectively.
A good agreement was observed between the forecasted and measured air pollutant
concentrations at Laxmi Talkie and Bharat Yantra Nigam, Allahabad, U.P India. The
average value of PM10 , SO2 , and NO2 is found to vary from 227 to 269.8 µg/m3 , 3.7
to 5.1 µg/m3 , and 29.3 to 36.2 µg/m3 for Laxmi Talkie. For Bharat Yantra Nigam,
Artificial Neural Networks Based Condition Monitoring … 433

Fig. 6 PP of ANN-6

Fig. 7 EH plot of ANN-1

average value of PM10 , SO2 , and NO2 is found to vary from 174.8 to 236 µg/m3 , 2.9
to 4.2 µg/m3 , and 23 to 26.4 µg/m3 . As per air quality index, PM10 is moderately
polluted and SO2 and NO2 are in good range for Laxmi Talkie and Bharat Yantra
Nigam, Allahabad, U.P India. A good range of air pollutant causes minimal impact
whereas moderately polluted causes asthma, heart disease, and older adult.
434 V. Yadav and S. Nath

Fig. 8 EH plot of ANN-2

Fig. 9 EH plot of ANN-3

Artificial Neural Networks Based Condition Monitoring … 435

Fig. 10 EH plot of ANN-4

Fig. 11 EH plot of ANN-5

436 V. Yadav and S. Nath

Fig. 12 EH plot of ANN-6

Acknowledgements We would like to thank Uttar Pradesh Pollution Control Board (UPCB) Luc-
know and Department of Science and Technology (DST), New Delhi-110016 India for providing
inspire fellowship with Ref. No. DST/INSPIRE Fellowship/2016/IF160676.

References

1. V.S. Barai, K.A. Dikshit, S. Sharma, Neural network models for air quality prediction---a
comparative study. J. Environ. Sci. Eng. 1–18 (2013)
2. A.A. Zinatizadeh, M. Pirsahed, R.A. Kurdin, S. Zinadini, A. Dezfoulinijud, Dust level fore-
casting and its interaction with gaseous pollutants using artificial neural network: a case study
for Kermanshah, Iran. Iran. J. Energy Environ. 5(1), 51–58 (2014)
3. M. Alkasassbeh, Sheta F.A. Mouhammad, H. Faris, H. Turabieh, Prediction of PM10 and TSP
air pollution parameters using artificial neural network autoregressive, external input models---a
case study in salt. Jordan. J. Sci. Res. 14(7), 999–1007 (2013)
4. S.M. Baawain, A.S.A. Serihi, Systematic approach for the prediction of ground-level air pol-
lution (around an Industrial Port) using an artificial neural network. Aerosol Air Qual. Res. 14,
124–134 (2014)
5. A.G.N. Rahman, H.M. Lee, T.M.S. Latif, Forecasting of air pollution index with artificial neural
network. J. Environ. Nat. Resour. Sci. 63(2), 59–64 (2013)
6. A.S.F. Filho, M.F. Fernandes, Time-series forecasting of pollutant concentration levels using
particle swarm optimisation and artificial neural network. J. Quim Nova. 36(6), 783–789 (2013)
7. K.H. Elminir, A.H. Galil, Estimation of air pollutant concentration from meteorological
parameters using artificial neural network. J. Electr. Eng 57(2), 105–110 (2006)
8. L.N. Benjamin, S. Sharma, U. Pendharker, J.K. Shrivastava, Air quality prediction using
artificial neural network. Int. J. Chem. Stud. 2(4), 7–9 (2014)
9. H. Golizadeh, M. Darand, Forecasting of air pollution using artificial neural networks: the case
study; Tehran City. J. Appl. Sci. 9(21), 3882–3887 (2009)
10. S. Mahmoudzadeh, Z. Othma, M.A. Yazdani, A.A. Bakar, Carbon monoxide prediction using
artificial neural network and imperialist competitive algorithm. J. Basic Appl. Sci. 7(4), 735–
744 (2012)
Artificial Neural Networks Based Condition Monitoring … 437

11. J. Skyrzypski, J.E. Szakiel, Neural network prediction models as a tool for air quality
management in Cities. J. Environ. Prot. Eng. 34, 130–137 (2008)
12. A.U. Sahin, C. Bayat, N.O. Ucan, Application of cellular neural network (CNN) to the
prediction of missing air Pollutant data. J. Atmos. Res. 101, 314–326 (2011)
13. N.P. Sonaje, S.J. Mane, A.S. Kote, Modelling of respirable suspended particulate matter con-
centration using artificial neural networks in an urban area. Int. J. Eng. Sci. Technol. 3(4) (2013)
ISSN.2250-3498
Theoretical and Experimental
Investigation of Spur Gearbox
with Cracked Pinion Tooth Using
Dynamic Model for Condition
Monitoring

Yogesh Pandya and Anand Parey

Abstract Gearbox dynamic modeling is used to understand the effect of presence of

gear tooth fault and the role of change in mesh stiffness on vibration response using
computer simulation of a geared system. In the present study, a ten degree of free-
dom dynamic model is formulated to demonstrate the effect of tooth fault and mesh
stiffness variation on the dynamic of geared system. Further, for experimental study
the crack path is generated using finite element analysis based on fracture mechanics
in pinion tooth profile. Experiments have been performed on a drivetrain diagnos-
tic simulator using single-stage spur gear box and the vibration signals have been
recorded for healthy and cracked pinion tooth. The wavelet de-noising technique
is adopted to remove the unwanted noise from the vibration signals. A modulation
sideband approach in conjunction with conventional frequency spectrum visualiza-
tion has been introduced. A comparison of analytical and experimental results using
classical sideband analysis have been outlined, and it is concluded that the proposed
methodology helps in fault detection with increased deterioration level.

Keywords Vibration modeling · Dynamic modeling · Gearbox · Spur gear · Mesh

stiffness · Sidebands · Fault detection

Nomenclature

DDS Drivetrain Diagnostic Simulator

SB SideBand

Y. Pandya (B)
Department of Mechanical Engineering, Prestige Institute of Engineering
Management and Research, Indore, MP, India
e-mail: [email protected]
A. Parey
Discipline of Mechanical Engineering, IIT Indore, Indore, MP, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 439

H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_20
440 Y. Pandya and A. Parey

1 Introduction

1.1 General Overview

The technique of change in gearmesh stiffness is widely used to study the vibratory
property of the gearbox in gear dynamic modeling and computer simulation. The
vibration response of the gearbox arrangement can be obtained analytically under
various tooth fault conditions. It is also used to understanding the effect of the fault
and system parameters on mesh stiffness in vibration diagnostic signal.
To identify the condition of geared system a series of vibration signals can be
generated through computer simulation analytically. Experiments can be conducted
on single-stage and multistage gearboxes and vibration signals recorded for different
levels of faults in gear tooth. The vibration response obtained from analytical model
and experiments can be further analyzed by using different techniques available such
as statistical and signal processing in the spectrum and its diagnostic performance is
tested for comparative fault detection.

1.2 Literature Review and Research Gap Findings

Several studies have been conducted in the past to reproduce the vibratory properties
of the gearbox system analytically using dynamic models on the basis of single to
multi degree of freedom including the source of vibration. An extensive review on
dynamic models and their classification based on objective and different parameters
considered for the study of dynamics of gears was given by Ozguven and Houser
[1] in 1988. Parey and Tandon [2] in 2003 reviewed the dynamics of spur gear with
defects in detail. Ozguven and Houser [3, 4] studied single and multi degree of free-
dom nonlinear model for dynamic analysis of gear pair and the effect of variable
mesh stiffness and mesh damping, gear errors, profile modification, backlash, and
taking bearing and shaft dynamics into consideration. Bartelmus [5] suggested dif-
ferent factors for multi degree of freedom single and double-stage gearbox dynamics
modeling and simulation analytically. Kahraman and Singh [6–8] have considered
external and internal excitations for nonlinear frequency response characteristics of
a pair of spur gears in mesh, rotor bearing, and rotor bearing system having gear with
backlash, radial clearance and variation of mesh stiffness with time, respectively, and
solutions have been suggested using method of simulation and harmonic balance.
Munro et al. [9] have proposed an experimental method which reveals the change of
tooth pair mesh stiffness on the contact path and also into the tooth corner contact
region using properties of transmission error plots at different loads.
Howard et al. [10] presented a dynamic model using 16 degree of freedom for
gearbox vibration to include the effect of friction and change in torsional mesh
stiffness due to crack defect and analyzed the response from multi degree of freedom
dynamic gear model with defects. Fakhfakh et al. [11] modeled limited (two) degree
Theoretical and Experimental Investigation of Spur Gearbox … 441

of freedom system to reproduce the vibration response of a single-stage geared system

affected by localized crack and spalling tooth damage with unstable mesh stiffness.
In another attempt, Jia and Howard [12] presented a 26° of freedom double-stage
spur gearbox dynamic model for comparing the vibration response for local spalling
and crack fault. The variation of combined torsion mesh stiffness with tooth spalling
and crack has been included in the dynamic model after transforming it into linear
translational mesh stiffness. Chaari et al. [13] proposed a dynamic model using four
degrees of freedom to study the effect of fault in vibration response by using variation
in gearmesh stiffness measured analytically.
Wu [14] and Wu et al. [15] adopted the mathematical model given by Bartelmus
[5] which includes torsional and lateral vibration with six degrees of freedom for
single-stage spur gearbox. It is further used to study the vibration response of the
dynamic model and the effect of tooth crack of various sizes for fault detection. Parey
et al. [16] proposed a dynamic model for spur gearbox and its various components
using six degree of freedom including localized tooth defect. The model also includes
the effects of changes in mesh stiffness, change in damping, gear errors, and change
in gear profile.
Kim et al. [17] in their study, analyzed a dynamic model and its response of a spur
gear pair which includes translational motion resulting from deformation in bearing
and the center distance between pinion and gear, the contact ratio, and the pressure
angle.
Ma and Chen [18] presented a dynamic model with four degree of freedom with
local tooth fault such as crack and spalling. The difference and comparison between
defects were made using different plotting techniques for different characteristics to
identify the defect type. In a similar study, Ma et al. [19] investigated the effect tooth
fault on mesh stiffness and dynamics of a gear pair. The spalling characteristics were
also evaluated analytically and experimentally, and statistical techniques which detect
defect are used at low speed and small excitations. Osman and Velex [20] presented a
three-dimensional model in combination including contact fatigue modes to analyze
dynamic tooth load on gears using movement of crack and its dependency on speed
range. Liang et al. [21] reviewed the techniques of gearbox dynamics and role of
variational parameters in condition monitoring and fault diagnosis.
Mcfadden [22] has used the technique of time domain average for early stages
of failure. He has used regular signal by removing signals other than the mesh fre-
quency and its harmonics. Inalpolat and Kahraman [23] in their paper proposed an
analytical model to include sidebands activities of planetary gearsets and developed a
methodology to demonstrate modulation sidebands from gears. Further rich sideband
activities are also demonstrated at various speed and torque conditions.
Inalpolat and Kahraman [24] proposed a model which includes both amplitude
modulations with frequency modulation for vibration signal at a fixed location in the
planetary housing. The spectrum shows modulation sidebands at frequencies around
gear rotational frequency.
442 Y. Pandya and A. Parey

2 Methodology

2.1 Dynamic Modeling of a Spur GearBox with Crack

at Pinion Tooth Root

The analytical model of a single-stage spur gearbox with gear ratio 2.5:1 having
32 teeth on pinion and 80 teeth on gear is presented here. The specifications of
components are given in Table 1. The dynamic model of single-stage spur gear setup
is shown in Fig. 1a, b. The free body diagram (FBD) is shown in Fig. 2c.
A dynamic model based on assumptions that the mass moment of inertia of shaft
having flexible coupling is concentrated at gears and bearing. The gear profiles are
involute and the resonance of the gear case and tooth friction is neglected.
There are total 10 degrees of freedom in the model, which is represented as shown
in Fig. 1a, b. Six vertical motions (four at bearings (yb1 , yb2 , yb3 , and yb4 ) and two
(y1 and y2 ) at pinion and gear) and four rotational motions (at motor, load, pinion,
and gear), θ 1 , θ 2 , θ m , and θ L , have been considered for dynamic simulation.

Table 1 Main parameter of geared system

Type Pinion Gear
No of teeth (N) 32 80
Base circle radius R1 = 0.0246 m R2 = 0.0615 m
Mass m1 = 0.318 kg m2 = 1.45 kg
Moment of Inertia 1.026 × 10−4 kg m2 2.923 × 10−4 kg m2
Mass of bearings mbrng = 0.22328 kg
Moment of Inertia (motor) I m = 0.0021 kg m2
I L = 0.01582 kg m2
Moment of Inertia (load)
Motor torque T m = 11.9 N/m
Load torque T L = 44.2 N/m
Input shaft Frequency f r = 5 Hz
Gearmesh frequency f m = 160 Hz
I/O coupling stiffness (k c ) k ic and k oc = 4.4 × 104 N m/rad
I/O coupling damping coefficient qic/ qoc = 5 × 105 N m/rad
Radial stiffness of bearing k brng = 6.565 × 107 N/m
Damping coefficient of bearing qbrng = 1.8 × 105 N/m
Scale constant (μ) 3.99 × 10−6 s
Vertical radial stiffness of bearings at input and Constant
output Constant
Bearing damping coefficient Constant
Torsional stiffness of flexible coupling Constant
Damping coefficient of coupling
Theoretical and Experimental Investigation of Spur Gearbox … 443

(a) (b)

Fig. 1 a A 10-DOF dynamic model for single-stage spur gear box, b modeling of stiffness and
damping among meshing tooth

Fig. 2 Diagram of the 10-DOF nonlinear model of a gear pair

The transverse and torsional movement of gears and shafts are shown in Fig. 1a,
the damped vibratory model for mesh stiffness of the gears is shown in Fig. 1b [12].
The equations of motion of the dynamic model are given by
Rotational equation of motions for motor and load are

Im θ̈m = Tm − qIC (θ̇m − θ̇1 ) − kIC (θm − θ1 ) (1)

444 Y. Pandya and A. Parey

IL θ̈L = TL − qOC (θ̇2 − θ̇L ) − kOC (θ2 − θL ) (2)

The rotary equation of motion for pinion and gears are as follows:

I1 θ̈1 = kIC (θ1 − θm ) + qIC (θ̇1 − θ̇m ) − r1 kt (r1 θ1 − r2 θ2 − y1 + y2 ) − r1 qt (r1 θ̇1 − r2 θ̇2 − ẏ1 + ẏ2 )
(3)

I2 θ̈2 = −r2 kt (r2 θ2 − r1 θ1 − y2 + y1 ) − r2 qt (r2 θ̇2 − r1 θ̇1 − ẏ2 + ẏ1 ) − koc (θ2 − θL )
(4)

As the vibration in x-direction are ignored considering the natural damping,

translation motion equations in y-direction for pinion and gear are as follows:

m1 ÿ1 = −kt (r2 θ2 − r1 θ1 + y1 − y2 ) − q12 (r2 θ̇2 − r1 θ̇1 + ẏ1 − ẏ2 )

− kstrans (y1 − y13 ) − kstrans (y1 − y17 ) (5)

m2 ÿ2 = −kt (r1 θ1 − r2 θ2 − y1 + y2 ) − q12 (r1 θ̇1 − r2 θ̇2 − ẏ1 + ẏ2 )

− kstrans (y2 − y15 ) − kstrans (y2 − y19 ) (6)

Translation motion equations in y-direction for bearings are as follows:

mbrng ÿ13 = −qbrng ẏ13 − kstrans (y13 − y1 ) − kbrng y13 (7)

mbrng ÿ15 = −qbrng ẏ15 − kstrans (y15 − y2 ) − kbrng y15 (8)

mbrng ÿ17 = −qbrng ẏ17 − kstrans (y17 − y1 ) − kbrng y17 (9)

mbrng ÿ19 = −qbrng ẏ19 − kstrans (y19 − y2 ) − kbrng y19 (10)

where, I 1 , I m , I L , and I 2 , are the moment inertia’s of pinion, motor, load, and gear,
respectively. T m and T L , are the motor and output torque due to loading, respectively.
The θ 1 , θ L , θ 2 , θ m , represents the rotational displacements of pinion, load, gear, and
motor, respectively. m1 , m2 , mbrng , mass of pinion, gear and bearings, respectively.
qic , qoc , are the damping coefficients.
The torsional stiffness of flexible coupling attached to input and output shaft is
given by k ic , k oc , respectively. The damping coefficient is given by qbrng and the
bearing stiffness in vertical direction is given by k brng . The displacement of bearing
is named as yb1 , yb2 , yb3 , and yb4 . The y1 , y2 are the displacement of pinion and gear
in y-direction. k strans is the transverse stiffness of the shaft. k t is time-varying total
effective mesh stiffness of gear tooth in mesh and qt is mesh damping coefficient.
The radius of pitch circle of pinion and gear is given by R1 and R2 , respectively.
Theoretical and Experimental Investigation of Spur Gearbox … 445

The equations of motion have been solved by writing these equations in a series of
first-order differential equations compatible with the MATLAB ode15 s differential
equation solver. The mesh damping coefficient is proportional to the effective stiff-
ness of gear mesh and is given by, C t = μ k t , where μ is the constant scaling factor
in seconds, the value of which is given in Table 1 [15]. The representation of linear
and rotational motion parameters used in the solution of second-order differential
equation are as given below:

y(1) = y1 y(7) = dy(1) = ẏ1 dy(7) = ÿ1 , y(13) = y13 y(17) = y17
y(2) = y2 y(8) = dy(2) = ẏ2 dy(8) = ÿ2 , y(14) = dy(13) = ẏ13 y(18) = dy(17) = ẏ17
y(3) = θm y(9) = dy(3) = θ̇m dy(9) = θ̈m dy(14) = ÿ13 dy(18) = ÿ17
y(4) = θL y(10) = dy(4) = θ̇L dy(10) = θ̇L y(15) = y15 y(19) = y19
y(5) = θ1 y(11) = dy(5) = θ̇1 dy(11) = θ̈1 y(16) = dy(15) = ẏ15 y(20) = dy(19) = ẏ19
y(6) = θ2 y(12) = dy(6) = θ̇2 dy(12) = θ̈2 dy(16) = ÿ15 dy(20) = ÿ19

2.2 Analytical Dynamic Response of Spur Gear System

with Tooth Root Crack

With the established dynamic model and the equation of motion, computer simu-
lations have been performed for healthy to 50% cracked condition of pinion tooth.
The total time-varying mesh stiffness values are calculated for healthy and cracked
tooth (Table 1) in the previous section has been injected for individual tooth level
to simulate the crack size effect on vibration response of the geared system. It can
be observed from Figs. 3 and 4, that the influence of the fault is obvious for visual
observation.
The time domain vibration signal generated for pinion with healthy and cracked
tooth from 10, 20, 30, 40, and 50% crack length shows some periodic change in
vibration amplitude as the crack level increases.
The vibration impulse of cracked tooth repeats once in each shaft period
(T = 1/fr = 0.2 s). The gearbox was simulated over 0.42 s covering more than
two revolutions of shafts.
In general, the content of analytical vibration signal with fault shows a difference
from the signals from a normal gearbox. Figures 3 and 4 shows analytical simulation
covering more than two revolution of the pinion, where the vertical displacement
(yb1 ) of pinion at the bearing between input coupling and pinion is shown. The
waveform without crack is shown in Fig. 4a, and those with crack are shown in
Fig. 4b–f. Figures show that the presence of crack changes the resulting simulation.
The effect of crack is to reduce the magnitude of the bearing displacement, and that
is expected, as during engagement and disengagement of cracked tooth, work done
in crack opening and closing removes energy from the system which account for loss
of mesh stiffness.
446 Y. Pandya and A. Parey

Fig. 3 Pinion displacement in y-direction with crack level a healthy teeth, b 10%, c 20%, d 30%,
e 40%, f 50% cracked tooth

2.3 Sideband Analysis of Frequency Spectrum

The defects cause modulation sideband in vibration signal frequency content. The
modulation sideband around mesh frequency and its harmonics of gear can be used in
diagnosis of fault as the spectrum has higher number of sidebands with the increase
in number and size of defects in gear as shown in Fig. 5a–f. There is a difference in
stiffness and deformation characteristics of a cracked tooth in mesh and it has a visible
effect on the frequency content of the signal than the signal from an undamaged one.
The decreased contact between the meshing teeth is due to reduction in stiffness and
dimension of fault in teeth.
Theoretical and Experimental Investigation of Spur Gearbox … 447

Fig. 4 Depiction of vibration response for a healthy pinion teeth and b, c, d, e, f cracked pinion
tooth in two complete cycle of a 32 teeth pinion in mesh
448 Y. Pandya and A. Parey

Fig. 4 (continued)
Theoretical and Experimental Investigation of Spur Gearbox … 449

-6
(a) x 10
2 2fm

1.5
Amplitude

Sidebands
3fm
1 fm

0.5

0
0 50 100 150 200 250 300 350 400 450 500
Frequency (Hz)
-6
(b) 2 x 10
2fm

1.5
Amplitude

Sidebands
3fm
1
fm

0.5

0 50 100 150 200 250 300 350 400 450 500

Frequency (Hz)
-7
x 10
(c)
18
2fm
16
14
Amplitude

12 Sidebands
3fm
10
8 fm

6
4
2

0 50 100 150 200 250 300 350 400 450 500

Frequency (Hz)

Fig. 5 Sideband activities at gearmesh frequency and its harmonics in the frequency response,
a healthy teeth and b, c, d, e, f cracked tooth pinion
450 Y. Pandya and A. Parey

(d) -6
x 10
2fm

1.5
Amplitude

3fm
Sidebands
1
fm

0.5

0
0 50 100 150 200 250 300 350 400 450 500
Frequency (Hz)
-6
(e) x 10
2fm
1.5

3fm
Amplitude

Sidebands
1
fm

0.5

0
50 100 150 200 250 300 350 400 450 500
Frequency (Hz)
-6
x 10
(f)
1.6 2fm

1.4
1.2
Amplitude

Sidebands 3fm
1
fm
0.8
0.6
0.4
0.2
0
0 50 100 150 200 250 300 350 400 450 500
Frequency (Hz)

Fig. 5 (continued)
Theoretical and Experimental Investigation of Spur Gearbox … 451

Table 2 Values of sideband frequency around gearmesh frequencies

k (N × f r ) + (k × f r ) Hz (N × f r )−(k × f r ) Hz
N 1 = 32, f r = 5 Hz (N × f r = 160 Hz)
k =1 165(HSB1) 155(LSB3)
k =2 170(HSB2) 150(LSB2)
k =3 175(HSB3) 145(LSB1)

Analytically, the side band in the spectrum will approximately occur at the combi-
nation of the following fundamental rotational frequency, number of teeth, and their
harmonics (Table 2).
Multiple computer simulations are performed for normal gear and gear with crack.
The calculations were made for 5 Hz rotational speed of shaft and summarized as
follows: The different sidebands occur at (N × f r ) ± (k × f r ), a combination of
number of teeth (N), N 1 = 32, N 2 = 80, rotational frequency (f r = 5 Hz), and k =
1, 2, 3, etc. Table 3 shows the magnitude of sideband around mesh frequency and its

Table 3 Amplitude of sidebands around gearmesh frequency for pinion with healthy and cracked
tooth
S. no. Tooth Gearmesh frequency Lower sideband Higher sideband
condition (LSB) (HSB)
Freq (Hz) Amp Freq (Hz) Amp Freq (Hz) Amp
(m/s2 ) (m/s2 ) (m/s2 )
1 Healthy 160 8.06e-7 145 2.08e-8 165 6.62e-8
2 150 2.72e-8 170 3.88e-8
3 155 5.89e-8 175 2.41e-8
4 10% 160 7.30e-7 145 6.32e-8 165 8.35e-8
5 crack 150 5.41e-8 170 6.17e-8
6 155 6.72e-8 175 8.08e-8
7 20% 160 7.31e-7 145 6.34e-8 165 8.40e-8
8 crack 150 5.44e-8 170 6.25e-8
9 155 6.80e-8 175 8.20e-8
10 30% 160 7.66e-7 145 6.11e-8 165 8.07e-8
11 crack 150 5.44e-8 170 5.20e-8
12 155 7.33e-8 175 7.21e-8
13 40% 160 7.64e-7 145 5.72e-8 165 7.80e-8
14 crack 150 5.35e-8 170 4.92e-8
15 155 7.17e-8 175 6.84e-8
16 50% 160 8.00e-7 145 6.60e-8 165 7.88e-8
17 crack 150 6.76e-8 170 3.25e-8
18 155 8.18e-8 175 5.29e-8
452 Y. Pandya and A. Parey

multiple for the spur gear pair. The rotational frequency of shaft one and shaft two
is 5 Hz and 2 Hz, respectively, and the gearmesh frequency is 160 Hz.
From the spectrum as shown in Fig. 5a–f, it is observed that there is a visible
sideband activity around gearmesh frequency of 160 Hz at constant load and speed
condition for healthy and cracked pinion. As we increase the crack length from no
crack to 50% crack, the lower sidebands (LSB) and higher sidebands (HSB) start to
grow in comparison to the healthy condition at gearmesh frequency listed in Table and
its harmonics at f gm , 2f gm , 3f gm , and so on. However, in the present study sidebands
around gearmesh frequency have been considered only.
In this analysis, the sidebands around gearmesh frequency have been studied as
shown in Fig. 6. It has been observed that the magnitude of gearmesh frequency and
its sidebands have been increased from healthy tooth to cracked tooth with increasing
crack length.

Fig. 6 Modulation of sidebands around gearmesh frequency for different crack size for analytical
signal
Theoretical and Experimental Investigation of Spur Gearbox … 453

Fig. 7 a DDS gearbox, motor and break arrangement, b position of pinion, gears and location of
transducer in single-stage reduction

2.4 Experimental Setup and Method of Testing

A Drivetrain Dynamics Simulator (DDS) test setup is used to perform the Experi-
ments as shown in Fig. 7a, the DDS setup has two-stage parallel shaft arrangement
out of which only single-stage with a pinion and gear of having 14.5 degree pres-
sure angle is used. The gears are made of steel and supported by TGP steel shaft
of 1 diameter each supported by an inbuilt four deep groove ball bearings. The
splash lubrication is used at the bottom as shown in Fig. 7b. The gearbox contains
the followings:
1. One driver pinion gear of 32 teeth is mounted on the input shaft as per the
specifications given in Table 1, and one driven gear of 80 teeth is mounted on
the output shaft.
2. A 3HP, 3phase motor used for driving the input shaft is used for having variable
speed of the gearbox.
3. A The magnetic particle brake is connected to the output shaft having 4 to 220 lb.in
capacity.
The DDS schematic as shown in Fig. 8 provide flexibility in terms of mounting
more than one gear to save time to the change required for testing different gears.
The accelerometer is located on top of input shaft bearing housing as shown in
Fig. 7b. The magnetic particle brake is fixed to the output shaft for applying required
load during the test. The motor can be run on constant speed or variable speed by
controlling it using PC. The pinion on input shaft is meshed with gear on the output
shaft in single stage. A data acquisition system (DAQ) is used to connect PC with
transducers to recording and analyzing the vibration signals further in MATLAB
programming.
454 Y. Pandya and A. Parey

Fig. 8 Schematic of complete gearbox testing arrangement

2.4.1 Prediction of Crack (Fault) Propagation Path on Pinion Tooth

Root Using Finite Element Analysis

The crack propagation or fault path is generated using geometry of a 2-D involute
pinion tooth modeled as pert the specifications listed in Table 4. A FRacture Analysis
Code (FRANC) is used to generate the crack propagation path in the root of the 2-D
tooth model for simplicity. The analysis has been performed on the meshed model
for prediction of crack path.
The tooth is fixed at the edges of the rim and the load is applied normal to tooth
surface and at the highest point of single tooth contact (HPSTC) as shown in Fig. 9a.
The high-stress region in the tooth root or filler region has been marked after loading
and the location of crack is ascertained. The crack trajectories for different crack
length, viz., 1 mm 2 mm and 4 mm have been generated.
The automatic crack trajectory as identified by finite element analysis software is
further used to prepare the crack of different length on a 32 teeth pinion specimen
used, at the specified location on one of the tooth for the experimental study.

Table 4 Main specification

Name of specification Pinion Gear
of gear pair
No of teeth (N) 32 80
Diametral pitch (DP) 16/ inch
Module (m) 1.5875 mm
Pressure angle (α) 14.5°
Face width (B) 12.7 mm
Contact ratio 2.12
Young’s Modulus 2×105 MPa
Poison’s Ratio v = 0.3
Theoretical and Experimental Investigation of Spur Gearbox … 455

Fig. 9 a Meshed model of cracked pinion, b crack propagation path at tooth root

2.4.2 Creation of Predicted Crack Path on Pinion Tooth Using

CNC—WEDM

The CNC part program is written in ELCAM and wire electrode discharge machin-
ing (WEDM), wirecut ECOCUT is used to produce the predicted crack trajectory
(Fig. 9b) as shown in Fig. 10a–b, on the pinion tooth obtained via finite element
analysis. Different crack lengths are introduced on pinion, in the 10% increment and
upto 50% of crack length as shown in Fig. 11a–e.

Fig. 10 a Cracked Pinion tooth, b use of WEDM for crack generation

456 Y. Pandya and A. Parey

Fig. 11 WEDM generated crack path at the pinion tooth root, a 10%, b 20%, c 30%, d 40% and
e 50% crack length

2.4.3 Experimental Vibration Signal of a Spur GearBox with Faulty

Pinion

The vibration signals were measured using accelerometer location which is shown
in Fig. 7b. The sampling frequency was 3200 Hz. The experimental results were
obtained at constant shaft speed of 5 Hz (300 rpm). The characteristic frequencies of
healthy and cracked pinion are shown in Table 2. The vibration response of the spur
gear with different tooth condition, namely, healthy, 10, 20, 30, 40, and 50% crack
length of fully developed crack are shown in Fig. 12a–f.
For the further processing of the experimental vibration data, the unwanted noise
is removed using simple signal de-noising using wavelet.
In the present study, wavelet tool box of MATLAB is used for de-noising.
Daubechies wavelet (db2) at level 5 which is widely used in solving such problems
has been used here for decomposition. The resultant de-noised signals corresponding
to original signals are shown in Fig. 13.
After preprocessing the experimental signal, it is observed that in the time domain
signal it is hard to observe the effect of fault such as crack as shown in Fig. 13a–f.
Viewing the signal in frequency domain as show in Fig. 14 is an effective approach
in comparison to the time domain signals.
Theoretical and Experimental Investigation of Spur Gearbox … 457

Fig. 12 Raw experimental signals for a healthy, b 10%, c 20%, d 30%, e 40% and f 50% cracked
pinion tooth

It is observed that all considerable amplitudes are around the mesh harmonic order
f gm . The sideband appears around gearmesh frequency and its harmonics, when a
crack is present in frequency spectrum as shown in Fig. 15. However, amplitude
of sidebands around gearmesh frequency has been considered only in the present
study. The amplitude of lower (LSB) and higher sideband (HSB) are calculated and
summarized in Table 5 where it can be seen from Fig. 15 that the magnitude of the
fundamental mesh frequency and their side bands are affected by the presence and
dimension of the crack and their magnitude is modulated with increase in crack size
in the pinion tooth.
458 Y. Pandya and A. Parey

Fig. 13 De-noised experimental signals, a healthy, b 10%, c 20%, d 30%, e 40% and f 50% cracked
pinion tooth

3 Results and Discussions

The dynamic model of a ten degree of freedom spur gear system with crack is estab-
lished and the failure characteristics that a single tooth crack has upon the resultant
dynamic motion has been demonstrated using time history and spectrum analysis.
The model focuses on the variation of the tooth mesh stiffness in the presence of a
fault such as tooth crack. The analysis results show that the failure will cause unstable
region in the time waveform and existence of LSB and HSB band in the frequency
spectrum for both analytical and experimental signals. A number of experiments of
the gearbox with crack defect were carried out to analyze the failure characteristics
Theoretical and Experimental Investigation of Spur Gearbox … 459

Fig. 14 Frequency spectrum of de-noised experimental signal for a healthy, b 10%, c 20%, d 30%,
e 40% and f 50% cracked pinion tooth

of the measured vibration signals, which agrees qualitatively with the theoretical
results. This model can thus be used for predicting the mesh stiffness variation due to
localized tooth defect, and the principles could be extended to all cases where tooth
deflection is dominant. The results presented in this study, provides a theoretical
foundation and experimental evidence of the fault diagnosis of gearbox.
460 Y. Pandya and A. Parey

Fig. 15 Modulation of sidebands around gearmesh frequency for different crack size

4 Conclusions

It has been observed that variation in mesh stiffness in the presence of a fault such as
tooth root crack as shown in Fig. 16 causes the unstable region in the vibration signa-
ture of the geared system and same phenomenon is observed in analytical results as
well as in vibration signature of gearbox with increasing deterioration level (Table 6).
There is a difference in stiffness and deformation characteristics of a tooth with a
crack in the root than an undamaged one in mesh and have influence on the frequency
spectrum. The modulation may be either due to frequency or amplitude. The locations
of sidebands are determined using the carrier frequency, the number of teeth and
‘k’ around gearmesh frequency to show the increase in sideband activity through
numerical and experimental results.
The effect of including varying mesh stiffness with and without fault in dynamic
model is demonstrated. The analytical and experimental results show that the spec-
trum contains sideband activities in the presence due to faults such as crack due
to amplitude modulations around the fundamental gearmesh frequency. At the end,
it is therefore significant to study sideband analysis and can be used for condition
monitoring of dynamic system such as gearbox and for fault indication.
Theoretical and Experimental Investigation of Spur Gearbox … 461

Table 5 Amplitude of sidebands around gearmesh frequency for pinion with healthy and cracked
tooth
S. Tooth Gearmesh frequency Lower sideband Higher sideband
no. condition (LSB) (HSB)
Freq Amp Freq Amp Freq Amp
(Hz) (m/s2 ) (Hz) (m/s2 ) (Hz) (m/s2 )
1 Healthy 159.8 0.00115 155.7 0.0008411 165 0.0001113
2 150 0.0007231 170.1 0.000536
3 144.3 0.001028 174.2 0.0001628
4 10% crack 159.8 0.001228 155.7 0.001139 165 0.0001165
5 150 0.0006674 170.1 0.001445
6 144.3 0.001369 174.2 0.0001977
7 20% crack 159.8 0.001213 155.7 0.001155 165 0.000131
8 150 0.0006735 170.1 0.001462
9 144.3 0.001357 174.2 0.000203
10 30% crack 159.8 0.00162 155 0.0002947 165 0.0000792
11 150 0.0006321 170.1 0.001165
12 145.1 0.0003293 175.1 0.000067
13 40% crack 159.8 0.003035 155 0.0009125 165 0.0001069
14 150 0.001036 170.1 0.0005988
15 145.1 0.001036 175.1 0.0001057
16 50% crack 159.8 0.002185 155 0.001201 165 0.0002856
17 150 0.0006779 170.1 0.0007596
18 145.1 0.001344 175.1 0.0001604

Fig. 16 Variation of gearmesh stiffness in spur gear pair in mesh for pinion rotation including
healthy tooth and tooth with 10, 20, 30, 40, and 50% cracked tooth
462 Y. Pandya and A. Parey

Table 6 Variation of mesh stiffness for cracked pinion

Crack length Crack intersection angle Percentage change in mesh stiffness
L v k t Max. (×109 ) k t Min. (×109 )
Healthy – 1.3 1.008
10% 67° 1.29(0.8%) 1.002(0.6%)
20% 69° 1.27(2%) 0.994(1.4%)
30% 72° 1.23(5.4%) 0.985(2.3%)
40% 75° 1.19(8.5%) 0.978(3%)
50% 77° 1.16(10.8%) 0.972(3.6%)

References

1. H.N. Ozguven, D.R. Houser, Mathematical models used in gear dynamics—a review. J. Sound
Vib. 121(3), 383–411 (1988)
2. A. Parey, N. Tandon, Spur gear dynamic models including defects—a review. Shock. Vib. Dig.
35(6), 465–478 (2003)
3. H.N. Ozguven, D.R. Houser, Dynamic analysis of high speed gears by using loaded static
transmission error. J. Sound Vib. 125(1), 71–83 (1988)
4. H.N. Ozguven, A non-linear mathematical model for dynamic analysis of spur gears including
shaft and bearing dynamics. J. Sound Vib. 145(2), 239–260 (1991)
5. W. Bartelmus, Mathematical modeling and computer simulation as an aid to gearbox
diagnostics. Mech. Syst. Signal Process. 15(5), 855–871 (2001)
6. A. Kahraman, R. Singh, Non-Linear dynamics of spur gear pair. J. Sound Vib. 142, 49–75
(1990)
7. A. Kahraman, R. Singh, Non-linear dynamics of a geared rotor bearing system with multiple
clearance. J. Sound Vib. 144(3), 409–506 (1991)
8. A. Kahraman, R. Singh, Interaction between time varying mesh stiffness and clearance non-
linearities in a geared system. J. Sound Vib. 146(1), 135–156 (1991)
9. R.G. Munro, D. Palmer, L. Morrish, An experimental method to measure gear tooth stiffness
throughout and beyond the path of contact, in Proceedings of the Institution of Mechanical
Engineers, 2001, 215(c), 793–803 (2001)
10. I. Howard, S. Jia, J. Wang, The dynamic modeling of a spur gear in mesh including friction
and a crack. Mech. Syst. Signal Process. 15(5), 831–853 (2001)
11. T. Fakhfakh, F. Chaari, M. Haddar, Numerical and experimental analysis of a gear system with
teeth defects. Int. J. Adv. Manuf. Technol. 25, 542–550 (2005)
12. S. Jia, I. Howard, Comparison of localized spalling and crack damage from dynamic modeling
of spur gear vibrations. Mech. Syst. Signal Process. 20(2), 332–349 (2006)
13. F. Chaari, T. Fakhfakh, M. Haddar, Analytical modeling of spur gear tooth crack and influence
on gear mesh stiffness. Eur. J. Mech. A/Solid 28, 461–468 (2009)
14. S. Wu, Gearbox dynamic simulation and estimation of fault growth, Master’s Dissertation,
University of Alberta, 2007
15. S. Wu, M.J. Zuo, A. Parey, Simulation of spur gear dynamics and estimation of fault growth.
J. Sound Vib. 317(3–5), 608–624 (2008)
16. A. Parey, M.E. Badaoui, F. Guillet, N. Tandon, Dynamic modelling of spur gear pair and
application of empirical mode decomposition-based statistical analysis for early detection of
localized tooth defect. J. Sound Vib. 294, 547–561 (2006)
17. W. Kim, H.H. Yoo, J. Chung, Dynamic analysis for a pair of spur gears with translational
motion due to bearing deformation. J. Sound Vib. 329, 4409–4421 (2010)
Theoretical and Experimental Investigation of Spur Gearbox … 463

18. R. Ma, Y. Chen, Research on the dynamic mechanism of the gear system withlocal crack and
spalling failure. Eng. Fail. Anal. 26, 12–20 (2012)
19. R. Ma, Y. Chen, Q. Cao, Research on the dynamic and fault mechanism of spur gear pair with
spalling defect. J. Sound Vib. 331, 2097–2109 (2012)
20. T. Osman, Ph Velex, A model for the simulation of the interactions between dynamic tooth
loads and contact fatigue in spur gears. Tribol. Int. 46, 84–96 (2012)
21. X. Lianga, M.J. Zuoa, Z. Fen, Dynamic modeling of gearbox faults: A review. Mech. Syst.
Signal Process. 98(1), 852–876 (2018)
22. P.D. Mcfadden, Examination of a technique for the early detection of failure in gears by signal
processing of the time domain average of the meshing vibration. Mech. Syst. Signal Process.
l(2), 173–183 (1987)
23. M. Inalpolat, A. Kahraman, A theoretical and experimental investigation of modulation
sidebands of planetary gear sets. J. Sound Vib. 323(3), 677–696 (2009)
24. M. Inalpolat, A. Kahraman, A dynamic model to predict modulation sidebands of a planetary
gear set having manufacturing errors. J. Sound Vib. 329(4), 371–393 (2010)
Experimental Study of Sideband
Harmonics in Vibration Spectrum
of Two-Stage Planetary Gear Box
for Condition Monitoring

Yogesh Pandya and Anand Parey

Abstract The fault diagnosis via condition monitoring of a two-stage planetary

gearbox using the vibration signal processing techniques is the aim of the present
work. The online vibration data were recorded for a set of normal gear and gear with
fault, at a particular speed and various loading condition using Drivetrain dynamics
simulator setup. A spectral analysis is carried out in order to demonstrate and examine
the presence of sidebands activity around the gearmesh frequency resulting from the
localized tooth fault, viz, broken tooth. The chapter is to investigate and develop an
understanding of the sideband activities around gearmesh frequencies of a double-
stage planetary gearbox.

Keywords Vibration · Experimental analysis · Planetary gearbox · Condition

monitoring

1 Introduction

Planetary gearboxes are used in automobile gearboxes, processing machines and

heavy earth moving machines for their outstanding power transmission efficiencies,
large speed reduction and compactness.
The sideband analysis is a technique to study the amplitude of the frequency
content of a signal and can be used for planetary gearbox vibration for condition
monitoring. A lot of work has been carried out in the area of planetary gearbox fault
diagnosis but limited published work is available. In the past, a number of researchers
have studied the vibration signal and concluded that there is presence of modulation
sidebands. In one such study, McFadden and Smith [1] presented the effect of sensor

Y. Pandya (B)
Department of Mechanical Engineering, Prestige Institute of Engineering Management and
Research, Indore, MP, India
e-mail: [email protected]
A. Parey
Discipline of Mechanical Engineering, IIT Indore, Indore, MP, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 465

H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_21
466 Y. Pandya and A. Parey

transducer location on vibration signal of planetary gearbox. McFadden [2] used

the time domain averages technique for tooth meshing vibration in an Epicyclic for
gear failure detection. McNames [3] studied the amplitudes of the main peaks and
concluded that the complete vibration spectrum is the combination of individual
planet gear vibration passing from a fixed transducer location and having major
frequencies and number of planets.
Inalpolat and Kahraman [4] suggested that having inphase and uniformly spaced
N planet gears, the symmetric sidebands will occur with fundamental gearmesh
frequency Zr × ωc at (Zr ± N )ωc, if arrangement satisfies the condition, ψi =
2π(i − 1)/N , Zr × ψi/2π = n, where n is an integer. They further mentioned
that the sidebands are not symmetric around fundamental gearmesh and exhibits
almost zero amplitude. All significant harmonics with sizable amplitude appear at
integer multiple of N at H = n × N, around the mesh harmonic, where n is an
integer, for uniformly spaced planets and consecutively phased gearmeshes if it
 N the mathematical condition, ψi = 2π(i − 1)/N , Zr × ψi/2π = n, and
satisfy
i=1 Zr × ψi = mπ , where m and n are integers.
Mosher [5] defined the characteristics of vibration signals of a planetary gearbox
using kinematics and through experimental measurement for condition monitoring.
Ken [6] outlined the results of spin test of the twostage planetary gearbox vibration
for failure causes and remedies and various equation for calculating the gearbox
parameters.
Feng and Zao [7] presented a comprehensive study of different models of con-
dition monitoring of planetary gearbox for fault detection using theoretical and
experimental analysis.
Lie et al. [8] conducted a review of the published work on methodologies of
condition monitoring and fault detection of planetary gearbox.
Zhang et al. [9] have explored the effectiveness of Sideband Energy Ratio (SER)
method for planetary gearbox fault detection for analyzing limited number of side
bands.
Besides knowing the behaviour of vibration data, previous studies fall short of
providing a general formulation for planetary vibration condition monitoring for
diagnostics purpose, which provides a reasonable foundation for amplitude modu-
lated sidebands of planetary gearsets with localized tooth fault that can be used for
fault diagnosis purpose in planetary gearsets based on their sideband behaviour.
For the actual test data of the tests of planetary transmissions, a detailed exper-
imental study of modulation sidebands of planetary gearsets was carried out for
two-stage Epicyclic Gear Train (EGT) and the results are presented.

2 Experimental Setup

For the purpose of experimental investigation we have to collect the vibration signal
for different external load values, as the gearbox with faulty gear is more vulnerable
to external load than the gearbox in healthy condition.
Experimental Study of Sideband Harmonics … 467

Experiments were performed on Epicyclic Gear Train (EGT) inbuilt in Drivetrain

dynamics simulator (DDS) setup as shown in Fig. 1, consists of two-stage Epicyclic
Gear Train. The first stage of planetary arrangement at the input shaft is having N s
= 20, N P = 40 and N R = 100 number of teeth on sun gear, planet gears (three in
number) and ring gear, respectively. The second stage of transmission is having N s
= 28, N P = 36 and N R = 100 number of teeth on sun gear, planet gears (four in
number) and ring gear, respectively of 14.5° pressure angle each made of steel.
The axis of shaft is supported by 2 deep groove ball bearings. The gearing system
is lubricated by splash lubrication provided at the bottom of gear housing. The input
shaft of the gearbox is powered by a 3 phase, 3 HP, 0–5000 rpm variable speed
motor. The output shaft is connected with 4–220 lb. in capacity magnetic particle
brake (Fig. 2).

Fig. 1 (a) and (b) Schematic of planetary arrangement of two-stage EGT

Fig. 2 DDS planetary gearbox attached with sensors and magnetic particle brake
468 Y. Pandya and A. Parey

3 Test Procedure and Results

The test has been conducted for healthy gear and gear with broken tooth and the
vibration signals were recorded under varying load condition. The gear load was
varied from no load to 90% load, to generalize the effect of load on vibration spectrum
for a rotational speed of 6.7 Hz of input shaft. The sampling frequency should be
double the maximum frequency to be acquired hence it is set to a value of 12,800
samples per second as our maximum frequency to be acquired is 5 kHz, which is
well above the minimum sampling frequency.
Four channels were used in the multi analyser to capture the vibration from the
gearbox in real time with varying load condition mentioned above after initial running
the gearbox for some time. At first, the healthy pinion 1 has been fixed on the carrier
shaft of stage 1 and meshed with healthy planet gears of carrier of output shaft of
stage 2. The vibration activities were captured for the loading condition mentioned
above for the pair of healthy sun gear and planets. In the second stage a defected sun
gear with broken tooth as shown in Fig. 3 is fixed and the same procedure has been
repeated to acquire the vibration signals.
In the present study eleven gears whose details are mentioned in Table 1 are used.
The three gears with number of teeth, viz. 80, 40 and 72 are free from localized
defects. During the experiment, the load is applied for short duration of 15 s in steps
from 0 to 90% for each of healthy gear pair and gear pair with defects. Sufficient time

Fig. 3 Arrangement of gears in planetary gearbox in first and second stage

Experimental Study of Sideband Harmonics … 469

Table 1 Detailed specification of planetary gear arrangement

Parameters Details First stage Second stage
S RPM Sun gear RPM (input speed) 402 67.2
TR Ring gear teeth 100 100
TP Planet gear teeth 40 36
TS Sun gear teeth 20 28
GTvalue Gear train value 0.2 0.28
C RPM Carrier RPM 67.2 14.7
PRPM Planet RPM 168 40.8
PRPMabs Planet RPM absolute 100.8 26.1
RRPM Ring gear RPM 0 0
PGMF Planet gearmeshing frequency CPM 6720/112 1468.8/24.5
S GMF Sun gearmeshing freq CPM 8040/134 1881.6/31.37
R Stage ratio 5.9821 4.5714

has been given to stabilize the system in particular speed condition. After acquiring
the vibration signals in time domain, it is processed further in multichannel FFT
analyzer in real time to observe, analyze and record the side band activities around
gearmesh frequency.

4 Results and Discussion

The size of sideband around gearmesh frequency and its harmonics are shown in
Fig. 4a–h for varying load. The gear load is varied from, viz., no load, 10, 20, 40,
60, 70 and 90% load for a constant rotation of input shaft.
The sidebands will occur at combination of the fundamental rotational frequency
(Fr), number of teeth (N) on ring gear and their harmonics (N × Fr) ± (k × Fr) for
k = 1, 2, 3, …, etc. at 870, 900, 930, 990, 1020, 1050 Hz for shaft one in which the
first gearmesh frequency occurs at 960 Hz.

5 Conclusion

The sideband activities around the gearmesh frequency and its harmonics is a indica-
tor of gearbox health in general or gear condition in particular. It is therefore studied
here that how these activities can be captured for condition monitoring experimen-
tally with the support of available literature. In the present study only amplitude of
sidebands around gearmesh frequency has been considered. Where it can be seen
470 Y. Pandya and A. Parey

(a) 30hz 90 for normal gear and gear with fault (b) 30hz 80 for normal and gear with fault

(c) 30hz 70 for normal and faulted gear (d) 30hz 60 for normal and faulted gear

(e) 30hz 50 for normal and gear with fault (f) 30hz 40 for normal and gear with fault

(g) 30hz 20 for normal and gear with fault (h) 30hz 0 for normal and gear with fault

Fig. 4 a–h The sideband activities around gearmesh frequency and its harmonics with varying
load
Experimental Study of Sideband Harmonics … 471

from results that the magnitude of the sidebands around mesh frequency and harmon-
ics are affected by the presence and dimension of the tooth fault and their magnitude
is modulated.

References

1. P.D. McFadden, J.D. Smith, An explanation for the asymmetry of the modulation sidebands
about the tooth meshing frequency in epicyclic gear vibration, in Proceedings of the Institution
of Mechanical Engineers, vol. 199, no. C1 (1985), pp. 65–70
2. P.D. McFadden, A technique for calculating the time domain averages of the vibration of the
individual planet gears and the sun gear in an epicyclic gearbox. J. Sound Vib. 144(1), 163–172
(1991)
3. J. McNames, Fourier series analysis of epicyclic gearbox vibration. ASME J. Vib. Acoust. 124,
150–152 (2002)
4. M. Mosher, Understanding vibration spectra of planetary gear systems for fault detection, in
Proceedings of ASME Design Engineering Technical Conferences, Paper No. PTG-48082 (IL,
USA, 2003)
5. K. Singleton, A Case Study: Analysis of a Two Stage Planetary Gearbox Vibration (KSC
Consulting LLC, 2006)
6. M. Inalpolat, A. Kahraman, A theoretical and experimental investigation of modulation
sidebands of planetary gear sets. J. Sound Vib. 323, 677–696 (2009)
7. Z. Feng, M.J. Zuo, Vibration signal models for fault diagnosis of planetary gearboxes. J. Sound
Vib. 331, 4919–4939 (2012)
8. Y. Lei, J. Lin, Ming J. Zuo, Z. He, Condition monitoring and fault diagnosis of planetary
gearboxes: a review. Measurement 48, 292–305 (2014)
9. M. Zhang, K. Wang, D. Wei, Ming J. Zuo, Amplitudes of characteristic frequencies for fault
diagnosis of planetary gearbox. J. Sound Vib. 432, 119–132 (2018)
Vibration-Based Fault Diagnosis
of a Bevel and Spur Gearbox Using
Continuous Wavelet Transform
and Adaptive Neuro-Fuzzy Inference
System

Amandeep Singh Ahuja, Dada Saheb Ramteke and Anand Parey

Abstract This Chapter is based on a research study aimed at identifying faults in

two different types of gearboxes under nonstationary conditions. The method pro-
posed for fault diagnosis employs the independent angular resampling technique
for processing the raw gearbox vibration signatures along with a hybrid intelligent
classifier, namely, ANFIS, for fault diagnosis. With wavelet coefficients being input
directly to an intelligent classifier for diagnosing gear faults, promising results have
been reported in the recent past. However, considering that the computation bur-
den associated with the implementation of an ANFIS classifier is dependent to a
large extent on the network inputs, the angular domain averaged wavelet amplitude
maps are segmented, each segment representing 60° of pinion rotation. Coefficients
extracted from fragmented scalograms are then fed to a hybrid intelligent classifier,
namely, ANFIS, for fault identification. The research outcome indicates reasonably
good gear fault diagnostic potential in the case of the bevel gearbox. The method-
ology proposed is then extended to spur gearbox fault diagnosis with promising
results.

Keywords Gearbox fault diagnosis · Energy–Shannon’s entropy ratio ·

Continuous wavelet transform · ANFIS

1 Introduction

In order to transmit power between intersecting shafts, bevel gears are extensively
employed. The bevel gear finds application in diverse areas such as in automobiles,
power plants, marine industry, and aircraft. Spur and planetary gearbox fault diag-
nosis has attracted enormous attention in the recent past. However, fault diagnosis
of bevel gearboxes merits further attention considering their widespread application

A. S. Ahuja · D. S. Ramteke · A. Parey (B)

Discipline of Mechanical Engineering, Indian Institute of Technology Indore, Indore,
Madhya Pradesh, India
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 473

H. Malik et al. (eds.), Soft Computing in Condition Monitoring and Diagnostics
of Electrical and Mechanical Systems, Advances in Intelligent Systems
and Computing 1096, https://doi.org/10.1007/978-981-15-1532-3_22
474 A. S. Ahuja et al.

and significance in power transmission. The following paragraph highlights some of

the research endeavor aimed at diagnosing faults in bevel gearboxes.
Saravanan et al. [1] reported promising results in fault diagnosis of a spur bevel
gearbox fault diagnosis based on an analysis of gearbox vibration signatures coupled
with fuzzy technique. Further, Saravanan et al. [2] studied the efficacy of wavelet
transformation-derived features for bevel gearbox fault diagnosis employing ANN
and proximal support vector machine (PSVM). Predominant statistical features for
fault identification were recognized using the J48 algorithm. Feature extraction
employing discrete wavelets and fault classification using ANN was investigated
by Saravanan and Ramachandran [3]. Statistical parameters obtained from wavelet
coefficients and shortlisted for superior fault identification potential through the J48
algorithm were employed as inputs to train and test the SVM and PSVM for identify-
ing faults in a bevel gearbox in [4]. Saravanan and Ramachandran [5] proposed a fast
single-shot PSVM to identify 24 classes of faults in a bevel gearbox and compared
the efficacy and time interval required for classifying all classes of faults together.
In [6], they utilized discrete wavelet transformation-derived features along with the
decision tree algorithm for diagnosing defects in a spur bevel gearbox.
A considerable amount of the above literature on bevel gearbox fault diagnosis
is focused on the identification of statistical parameters with superior fault discrim-
ination capability. Moreover, in such works, the experimental rig comprises a belt-
driven bevel gearbox. Gear speeds fluctuations owing to uncertainties of load and
drive mechanisms, and as a result of possible slippage between the belts and pulleys,
have largely been neglected. Further, research endeavor aimed at diagnosing faults
in gearboxes under nonstationary conditions owing to load and speed fluctuations
is limited [7–11]. Several research studies on gear fault identification under nonsta-
tionary scenarios are based on gearbox vibration signature analysis under run-up of
the drive mechanism [12–16]. A constant angular acceleration is assumed during the
linearly varying speed profile.
ANNs have been extensively implemented for gearbox fault diagnosis. Moreover,
fuzzy inference systems, too, have been utilized for diagnosing gearbox defects [1].
However, neuro-fuzzy hybrid classifiers possess the advantage of utilizing the neural
network properties to tune fuzzy systems that are largely rule-based and fundamen-
tally based on the manner in which information is processed by humans. An example
of a successfully employed neuro-fuzzy classifier for modeling functions that are
nonlinear in nature is the ANFIS or Adaptive Neuro-Fuzzy Inference System. The
hybrid architecture of the ANFIS enables it to learn the data sample characteristics
and adapt its parameters in order to minimize the network error. The succeeding
paragraph enumerates some of the research efforts wherein ANFIS has successfully
been implemented to diagnose gearbox faults.
Discrete wavelet transform (DWT) and ANFIS were employed by Wu et al. [17]
to identify locations of gear faults in an experimental setup as well as to classify
various conditions of gear faults. Discrete wavelet packet transform along with prin-
cipal component analysis (PCA) was employed by Attoui et al. [18] for extracting
suitable features with ANFIS being implemented for gearbox fault classification. A
neuro-fuzzy ensemble was implemented by Zhang et al. [19] for identifying four
Vibration-Based Fault Diagnosis of a Bevel and Spur Gearbox … 475

types of faults in a motorcycle gearbox. Utilizing wavelet packet transformation-

derived statistical parameters as input feature vectors, Soleimani [20] compared two
intelligent classifiers, the ANN and ANFIS, for their potential in diagnosing gear
faults.
This Chapter is based on a study that is aimed at fault diagnosis of a belt-driven
bevel gearbox and takes into consideration the nonstationary factors emerging from
the uncertainties of load and drive mechanisms and also the slippage possible between
pulleys and belts. No attempt is made to identify statistical parameters with superior
fault discrimination capability, though a reduction in the size of the feature set has
been undertaken on the basis of energy–entropy ratio results. A simple and effective
technique called the independent angular resampling (IAR) technique [10, 11] is
implemented for translating raw nonstationary gearbox vibration signatures into the
rotational domain. Thirty independent signals in the rotational domain are then aver-
aged to arrive at angular domain averaged (ADA) signatures. An independent signal
in the angular domain that represents one revolution of the pinion is fragmented
into six sections, each representing 60° rotation of the pinion. Decomposition with
CWT results in wavelet amplitude maps (WAMs) or scalograms with only six data
points. CWCs from the six columns are then employed directly as input to an ANFIS
for diagnosing gearbox faults. The present Chapter considers two cases for fault
diagnosis: in the first case, wavelet coefficients derived from the complete range of
scales are utilized for fault identification and in the second case, coefficients derived
from specific scales are employed. Vibration signatures acquired from another type
of gearbox are finally processed in a similar manner to ascertain if the proposed
methodology can be applied to diagnose faults in other gearbox types.
The subsequent sections of the Chapter are structured in the following manner:
Sect. 2 describes the machinery fault simulator used to acquire bevel gearbox vibra-
tion signatures. In Sect. 3, the IAR technique is introduced from the perspective of
the present work, while Sect. 4 gives an overview of CWT and its application in this
Chapter. An introduction to the ANFIS is given in Sect. 5. Sections 6 and 7 report the
gear fault diagnosis results obtained in case of the bevel gearbox and spur gearbox,
respectively. The Chapter ends with conclusions drawn from the present study in
Sect. 8.

2 Test Rig and Data Acquisition

The machinery fault simulator (MFS), comprises a belt-driven bevel gearbox driven
by a variable speed electrical motor. The magnetic load at the gearbox output is
capable of generating a torque up to 5 N-m. A speed reduction of the ratio 2.5:1 is
achieved owing to a smaller pulley driving a larger one. For example, if the motor
speed is set to 30 Hz, the pinion must ideally rotate at 12 Hz. In reality, however, owing
to various factors including belt slippage, the pinion speed fluctuates rendering the
task of diagnosing gearbox faults a challenging one. Figure 1 shows the experimental
setup employed in the present work for acquiring the gearbox vibration signatures.
476 A. S. Ahuja et al.

Motor

Data acquisition system Bearings Pulley

(small)

Belts
Motor speed controller
Magnetic load

Bevel gearbox

Pulley
(large)

Fig. 1 Bevel gearbox experimental setup (machinery fault simulator)

Figure 2 shows the gearbox case-mounted accelerometer used to capture vibration

signals in the z-axis. The larger of the two pulleys is mounted with two reflective strips
spaced 146° apart on the inner surface. Such a multiple pulse tachometer setup enables
those specific instants of time corresponding to three known rotational positions of
the drive shaft during any given revolution to be known. This is a prerequisite of the
IAR technique devised for translation of raw gearbox signals from time to rotational
domain. A laptop is interfaced to the data acquisition system that receives inputs
from the accelerometer and tachometer.
The three gearbox conditions under which vibration signals are collected include
a pinion in healthy condition, a pinion with one tooth chipped, and a pinion with one
tooth missing/ broken. In each experiment, the gear wheel continues to remain the
same. Figure 3 shows the three pinions employed while conducting experiments.
The motor rotational speed is set to 30 Hz and gearbox signatures are captured
under sampling frequency equaling 12,800 Hz. Figure 4 shows time domain vibration
signals acquired over 30 complete revolutions of the pinion at 0% load. Two readings
were acquired under each identical load–speed condition of which only reading 1
(R1) is being analyzed in the present work.
Ascertaining whether the gearbox has a healthy pinion or a pinion with one of the
teeth chipped, solely through a visual examination of the acquired vibration signals,
could be misleading, though the missing tooth pinion signal clearly exhibits impulses
of large amplitude. Time to rotational domain conversion of the gearbox vibration
Vibration-Based Fault Diagnosis of a Bevel and Spur Gearbox … 477

(a) (b)

z-axis

(c)

(d)

Fig. 2 a Accelerometer, b and c multiple pulse tachometer arrangement, and d pulse generation

(a) (b) (c)

One tooth
broken
One tooth
chipped

Fig. 3 a Pinion in a normal/healthy condition, b having one tooth chipped, and c having one broken
tooth

signatures is accomplished using the IAR technique. Initially, this Chapter deals
with bevel gearbox vibration signature analysis to ascertain gear faults. Section 7 of
this Chapter considers a similar analysis of fault identification in a spur gearbox to
confirm the applicability of the technique to another type of gearbox.
478 A. S. Ahuja et al.

Fig. 4 Raw gearbox signals TDSHP/ 0L/ 30Hz/ R1/ z-axis

25
incorporating each of the
three pinion types 20

15

10

acceleration (m/s )
2
5

-5

-10

-15

-20

-25
0 0.5 1 1.5 2 2.5 3
sample number 4
x 10
TDSCHT/ 0L/ 30Hz/ R1/ z-axis
25

20

15

10
acceleration (m/s )
2

-5

-10

-15

-20

-25
0 0.5 1 1.5 2 2.5 3
sample number 4
x 10

TDSMT/ 0L/ 30Hz/ R1/ z-axis

50

40

30

20
acceleration (m/s )
2

10

-10

-20

-30

-40

-50
0 0.5 1 1.5 2 2.5 3
sample number 4
x 10
Vibration-Based Fault Diagnosis of a Bevel and Spur Gearbox … 479

3 Implementation of IAR Technique

IAR technique aims at converting time domain gearbox signals into angular domain.
More precisely, the technique aims at conversion of raw gearbox signals to as many
angular domain signals as the number of complete pinion revolutions considered.
These independent angular domain signals can then be combined to arrive at the
angular domain signal corresponding to the originally acquired time domain sig-
nal. This combined angular domain signal can be decomposed using CWT and the
resulting WAM is utilized to compute the energy–entropy ratio of CWCs with the
objective of identifying the optimal scales. Alternatively, an averaging of indepen-
dent signals in the rotational domain can be undertaken to augment the quality of
the signal by diminishing its noise content. Figure 5 shows merged angular domain
signals corresponding to the three gearbox health conditions considered.
There is a striking similarity between the time domain signals acquired using
the accelerometer and the IAR technique-generated angular domain signals. The
change in axes from sample number in Fig. 4 to angle (in degrees) in Fig. 5 is to be
noted. While the number of samples in the three cases in Fig. 4 is unequal (owing to
fluctuations in pinion speed), the number of samples in Fig. 5 is equal representing
30 complete revolutions of the pinion.
Under fluctuating speed conditions (such as in the case under study), conventional
time synchronous averaging may not directly be suitable owing to unequal number of
samples between successive tachometer pulses. Another effective method of enhanc-
ing the signal quality is to carry out averaging of signals obtained in the rotational
domain through the IAR technique. These signals can be averaged for arriving at
ADA signatures corresponding to each of the pinion conditions considered. Some of
the phenomena not clearly evident in the time domain signal may reveal itself in the
ADA signals. Figure 6 shows the ADA signals for the three pinion health conditions.
A higher amplitude of the chipped tooth ADA signal compared to that of the
healthy pinion can be noticed in Fig. 6. Moreover, the missing tooth pinion ADA
signal exhibits high-amplitude impulses at around 50 degrees of rotation. This indi-
cates the missing tooth pinion location which was not so evident in both time domain
as well as merged angular domain vibration signatures.
The above gear fault diagnostic methods, however, to a large extent, rely on
the judgment of operator and are, therefore, susceptible to error. Development of a
method wherein the condition of the gearbox can be ascertained automatically by
feeding in parameters derived from gearbox signatures is essential. This is undertaken
in the present work by employing an ANFIS classifier which is fed with wavelet
coefficients obtained from gearbox vibration signatures.
480 A. S. Ahuja et al.

Fig. 5 IAR ADSHP/ 0L/ 30Hz/ R1/ z-axis

25
technique-generated angular
domain signals pertaining to 20

the three pinion types 15

10

acceleration (m/s )
2
5

-5

-10

-15

-20

-25
0 2000 4000 6000 8000 10000
angle (degrees)

ADSCHT/ 0L/ 30Hz/ R1/ z-axis

25

20

15

10
acceleration (m/s )
2

-5

-10

-15

-20

-25
0 2000 4000 6000 8000 10000
angle (degrees)

ADSMT/ 0L/ 30Hz/ R1/ z-axis

50

40

30

20
acceleration (m/s )
2

10

-10

-20

-30

-40

-50
0 2000 4000 6000 8000 10000
angle (degrees)
Vibration-Based Fault Diagnosis of a Bevel and Spur Gearbox … 481

Fig. 6 ADA signals ADASHP/ 0L/ 30Hz/ R1/ z-axis

10
corresponding to the three
pinion types 8

acceleration (m/s )
2
2

-2

-4

-6

-8

-10
0 50 100 150 200 250 300 350
angle (degrees)

ADASCHT/ 0L/ 30Hz/ R1/ z-axis

10

4
acceleration (m/s )
2

-2

-4

-6

-8

-10
0 50 100 150 200 250 300 350
angle (degrees)

ADASMT/ 0L/ 30Hz/ R1/ z-axis

20

15

10
acceleration (m/s )
2

-5

-10

-15

-20
0 50 100 150 200 250 300 350
angle (degrees)
482 A. S. Ahuja et al.

4 Transformation to Time–Frequency Domain

Mathematical transformation of a signal is carried out to extract meaningful fea-

tures in another domain. In the present Chapter, continuous wavelet transformation
of gearbox signatures is undertaken for extracting features in the time–frequency
domain. CWT of a signal x(t) may be defined by the following expression [21]:

+∞
1 t −τ
Wx (s, τ ) = √ x(t).Ψ ∗ ( )dt (1)
s s
−∞

where Ψ ∗ (t) is the complex conjugate of the mother wavelet function Ψ (t), τ and s
represent the translation and scale parameters, respectively.
Utilizing continuous wavelet coefficients (CWCs) in the form of direct inputs to
ANNs and support vector machines (instead of employing vibration signal-derived
statistical parameters) for gear fault identification, promising results are reported
in [22]. In this Chapter, the IAR technique-generated independent angular domain
signals are fragmented into six sections, each representing 60 degrees of pinion
rotation. These signals with six data samples each are averaged to generate averaged
signals in the rotational domain. Decomposition of these angular domain averaged
(ADA) signals with CWT is then carried out and wavelet coefficients at various scales
were employed as direct inputs to the ANFIS.
Bafroui and Ohadi [15] considered energy–Shannon’s entropy ratio to arrive at an
optimal range of wavelet scales and reported improved test accuracy in an ANN with
wavelet coefficient-derived parameters being fed to the ANN. The WAM obtained
through continuous wavelet transformation of the merged rotational domain signal
of Fig. 5a is shown in Fig. 7. Considering the impulsive characteristics associated

Fig. 7 Scalogram associated WAM/ ADSHP/ 0L/ 30Hz/ R1/ z-axis

with the healthy (normal)
61
pinion rotational signal
57
53
49
45
41
37
scale

33
29
25
21
17
13
9
5
1
1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
angle (degrees)
Vibration-Based Fault Diagnosis of a Bevel and Spur Gearbox … 483

Fig. 8 Energy–Shannon’s x 10
4

entropy ratio plot

corresponding to the WAM
of Fig. 8 6

energy entropy ratio

4
X: 25 X: 49
Y: 3.127e+04 Y: 3.146e+04

0
10 20 30 40 50 60
scale

with geared system faults, the Morlet wavelet is utilized for the transformation of
signals into the time–frequency domain.
The energy–entropy ratio of CWCs at the 64 scales of the healthy pinion WAM
is represented graphically in Fig. 8. Clearly, a range of scales with relatively higher
values of energy–entropy ratio can be seen in the plot.

5 ANFIS as a Hybrid Intelligent Classifier

ANFIS is a hybrid intelligent classifier that possesses the advantages of systematic

modeling and reduced reliance on expert knowledge owing to its adaptive nature.
Figure 9 demonstrates the various layers comprising the ANFIS architecture. The
fixed nodes in the architecture are represented by circles and adaptive nodes by
squares. The inputs to the system are represented by x and y.
All nodes of Layer I are adaptive in nature and output the fuzzy membership
grade of the inputs. This layer may espouse any membership function (MF), e.g.,
Gaussian, triangular and bell-shaped MF,. with adaptable parameters that govern
the MF. In Layer II, the nodes  are all fixed nodes. They perform multiplier action
and are labeled by the symbol . Layer II outputs are also referred to as the firing
strengths of rules. Layer III of the architecture also incorporates only fixed nodes.
They perform the function of normalizing the rule firing strengths (as obtained from
the preceding layer) and are represented by the symbol N . Layer III, therefore, outputs
the normalized firing strengths. Layer IV of the architecture incorporates adaptive
nodes that simply output the product of the normalized firing strengths from Layer
III and a polynomial that incorporates adaptable parameters. Layer V includes only 
one fixed node that sums all input signals and is represented by the symbol .
484 A. S. Ahuja et al.

Fig. 9 ANFIS architecture

Thus, the ANFIS architecture incorporates two layers with adaptive parameters, i.e.,
Layers I and IV. Layer I includes parameters that can modify the input MFs and
are referred to as premise parameters. Layer IV also includes adaptable parameters
associated with the polynomial which are known as consequent parameters. The
learning algorithm performs the function of tuning the adaptable parameters such
that the system generates an output consistent with the training set. Details of the
hybrid algorithm prove effective in training the classifier can be found in [23] and
[24].

6 Results and Discussions

The WAMs corresponding to the ADA signals for each pinion health condition
comprise six segments representing 60, 120, 180, 240, 300, and 360° of pinion
rotation. Figure 10 shows the angular domain averaged WAMs corresponding to the
three pinion health conditions.
The three WAMs in Fig. 10 appear different from each other and wavelet coeffi-
cients from these maps, when fed straight to the ANFIS, can help determine gearbox
faults. Two cases are considered in this Chapter:
(i) Wavelet coefficients extracted from the complete range of scales are input to
the classifier, and
(ii) Coefficients derived only from those scales identified as optimal are employed
in the form of inputs.
Gaussian MFs are selected with each input being divided into two regions, i.e., small
and large. The number of data samples from each class is 64 when data is fed in
from all scales and 25 when data is fed in from only the optimal scales. In each case,
half the data samples from each class are utilized to train and half to test the ANFIS.
Thus, two cases, as enumerated in Table 1, are considered in the present work.
Vibration-Based Fault Diagnosis of a Bevel and Spur Gearbox … 485

Fig. 10 Segmented WAMs Healthy pinion

for healthy, chipped tooth, 61
and missing tooth pinions 57
53
49
45
41
37

scale
33
29
25
21
17
13
9
5
1
1 2 3 4 5 6
angle (degrees) [/60]

Chipped tooth pinion

61
57
53
49
45
41
37
scale

33
29
25
21
17
13
9
5
1
1 2 3 4 5 6
angle (degrees) [/60]

Missing tooth pinion

61
57
53
49
45
41
37
scale

33
29
25
21
17
13
9
5
1
1 2 3 4 5 6
angle (degrees) [/60]
486 A. S. Ahuja et al.

Table 1 Training and test data set breakdown

Signal type Case No. of data samples Training samples Test samples from
from each class from each class each class
Vibration (z-axis) 1 64 32 32
2 25 13 12

Case 1: Data is fed in from all 64 scales, i.e., there are six columns consisting of
CWCs from all 64 scales pertaining to the three classes.
When the ANFIS is trained with 32 randomly selected data samples from each
class, the final error convergence value is 0.00020132 over 400 training epochs as
shown in Fig. 11.
The trained ANFIS classification accuracy, when training data samples are input
to the system, is shown in Fig. 12. It can be observed that all the 32 training data
samples from each class are correctly classified.
The ANFIS, having been trained with the training data set (comprising 32 ran-
domly data samples from each class), its classification accuracy with the test data set
is tested. It can be observed from Fig. 13 that some of the test data samples to the
trained classifier are incorrectly classified.

Fig. 11 Network error

convergence curve (Case 1)

Fig. 12 Trained ANFIS

classification accuracy when
fed with training samples
Vibration-Based Fault Diagnosis of a Bevel and Spur Gearbox … 487

Fig. 13 Trained ANFIS

classification accuracy when
fed with test samples

Case 2: In this case, only those data samples extracted from Scales 25 to 49 are used
for training and testing the hybrid classifier. There are still six columns consisting of
CWCs. Here, only the number of data samples has been reduced from 64 to 25 from
each class. Among the 25 data samples from each class, 13 samples and 12 samples,
respectively, are utilized to train and test the classifier. Figure 14 shows the network
error convergence curve wherein the final training error over 200 training epochs is
0.00026494.
Figure 15 represents the trained ANFIS classification accuracy utilizing the
training set. All the 13 data samples from each class are correctly classified.
Figure 16 represents the trained ANFIS test accuracy utilizing the test data sam-
ples. All the 12 data samples from each class are correctly classified. The test results
are unlike Case 1 where some of the test instances were incorrectly classified.

Fig. 14 Network error

convergence curve (Case 2)

Fig. 15 Trained ANFIS

classification accuracy
utilizing the training set
488 A. S. Ahuja et al.

Fig. 16 Trained ANFIS

classification accuracy
utilizing the test set

The results demonstrate that data samples from the optimal scales of the vibra-
tion signals’ WAMs have superior fault discrimination potential. In order to test the
proposed method for its generalization to other types of gearboxes, the procedure
is repeated for diagnosing faults in a spur gearbox. However, the procedure is now
applied to a four-class problem instead of the three-class problem as in the case of
the bevel gearbox.

7 The Procedure Applied to Spur Gearbox Fault Diagnosis

The experimental test rig called the drive train diagnostic simulator (DDS),
incorporating a single-stage spur gearbox, is shown in Fig. 17.
Vibration signatures from the single-stage spur gearbox are acquired along the
z-axis [indicated by a red arrow in Fig. 17b] using an accelerometer. The output shaft
of the electrical motor driving the gearbox is mounted with a tacho-wheel having
twin optical reflective tapes. A data acquisition system receiving inputs from the
accelerometer and tachometer is interfaced with a laptop.
In the case of the spur gearbox, four pinion health scenarios (pinion in healthy
condition, pinion with one tooth cracked, pinion with one tooth chipped, and pinion
with one tooth broken) are considered. Vibration signatures are collected from the
gearbox under each gear condition at 20 Hz and no load condition. Figure 18 shows
the four pinion types installed while conducting experiments.
The gear wheel continues to remain unchanged in each experiment and a sam-
pling frequency of 8192 Hz is selected during signal acquisition. Signal segments
representing 50 complete gearbox drive shaft revolutions are considered for analysis.
Figure 19 shows the gearbox vibration signals captured with each of the four pinion
types installed. The number of samples corresponding to 50 revolutions is not the
same in each of the four cases owing to fluctuations in speed of the gearbox drive
shaft.
Vibration-Based Fault Diagnosis of a Bevel and Spur Gearbox … 489

(a)

(b)
Accelerometer

Gearbox
housing
Tachometer

Coupling
Input shaft

Motor

Tacho-reflector wheel Bearing housing

(c)

Fig. 17 a DDS, b sensor placement, and c DDS schematic diagram

490 A. S. Ahuja et al.

Fig. 18 a Pinion in a normal/healthy condition, b having one tooth cracked, c having one tooth
chipped, and d with one tooth broken

In Fig. 19, relatively lower amplitude of vibration signal compared to the other
three cases can be observed in the case of the healthy pinion. Moreover, distinguishing
between the healthy and cracked tooth pinions based solely on visual observation
of the time domain vibration signals is challenging. Further, impulses at regular
intervals can be observed in case of the chipped tooth pinion while impacts of large
amplitude are evident in case of the missing tooth pinion. Gear fault diagnosis purely
on the basis of visual assessment of vibration signatures could be misleading and it is
preferable to have an automated mechanism for diagnosing gear faults to minimize
human error.
As in the case of the bevel gearbox, raw gearbox vibration signals are translated
into the rotational domain. The combined angular domain signal for the healthy
pinion and the corresponding WAM is shown in Fig. 20.
The energy–Shannon’s entropy ratios of wavelet coefficients pertaining to the 64
scales of the healthy pinion scalogram are plotted in Fig. 21. The 30 scales ranging
from 32 to 61 are identified with relatively higher energy–entropy ratios.
The 360° of rotation associated with each independent revolution is split into six
equal segments. Figure 22 shows the segmented angular domain averaged WAMs
for the four pinion conditions. CWCs from Segments 1 to 6 are employed directly
as six inputs to the ANFIS.
First, CWCs from all scales are employed for training and testing the ANFIS.
There are six columns of CWCs and 64 data samples from each class. An equal
number of randomly selected samples are utilized for training and testing the classi-
fier. Figure 23 shows that over 200 training epochs, the final error convergence value
is 0.021458.
The training and test performances of the trained ANFIS utilizing CWCs from
all scales are shown in Fig. 24 and Fig. 25, respectively. With CWCs belonging to
the complete range of scales being utilized to train the ANFIS, the training accuracy
obtained is 100%. However, Fig. 25 indicates that a number of test samples to the
ANFIS are misclassified.
With CWCs derived from optimal scales being utilized to train and test the ANFIS,
the final error convergence value over 150 training epochs is 0.00083938 as shown
in Fig. 26.
Vibration-Based Fault Diagnosis of a Bevel and Spur Gearbox … 491

Healthy Pinion Cracked Tooth Pinion

30 30

20 20

10 10
acceleration (m/s )

acceleration (m/s )
2

2
0 0

-10 -10

-20 -20

-30 -30
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Sample number 4 Sample number 4
x 10 x 10
Chipped Tooth Pinion Missing Tooth Pinion
30 30

20 20

10 10
acceleration (m/s )

acceleration (m/s )
2

0 0

-10 -10

-20 -20

-30 -30
0 0.5 1 1.5 2 0 0.5 1 1.5 2
sample number 4 sample number 4
x 10 x 10

Healthy Pinion Cracked Tooth Pinion

30 30

20 20

10 10
acceleration (m/s )

acceleration (m/s )
2

0 0

-10 -10

-20 -20

-30 -30
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Sample number 4 Sample number 4
x 10 x 10
Chipped Tooth Pinion Missing Tooth Pinion
30 30

20 20

10 10
acceleration (m/s )

acceleration (m/s )
2

0 0

-10 -10

-20 -20

-30 -30
0 0.5 1 1.5 2 0 0.5 1 1.5 2
sample number 4 sample number 4
x 10 x 10

Fig. 19 Vibration signatures acquired over 50 complete pinion revolutions under the four pinion
conditions
492 A. S. Ahuja et al.

Healthy pinion Healthy pinion

8
61
57
6
53
49
4
45
acceleration (m/s )
2

41
2
37

scale
33
0
29
25
-2
21
17
-4
13
9
-6
5
1
-8 2000 4000 6000 8000 10000 12000 14000 16000 18000
0 2000 4000 6000 8000 10000 12000 14000 16000 18000
angle (degrees) angle (degrees)

Fig. 20 Angular domain signal for the healthy pinion and the corresponding WAM

Fig. 21 Energy–Shannon’s 4
x 10
entropy ratio at the 64 scales
of the WAM in case of the 5
healthy pinion
4.5
energy-Shannon entropy ratio

3.5

2.5

2 X: 32
Y: 1.38e+04
1.5
X: 61
1 Y: 1.368e+04

0.5

0
10 20 30 40 50 60
scale

Figures 27 and 28 represent the training and test accuracies of the trained ANFIS
feeding CWCs from only the optimal scales. Fifteen randomly selected data samples
from the optimal range of scales each are utilized to train and test the trained system.
Feeding CWCs from the identified optimal range of scales directly to the ANFIS
leads to reasonably good training and test accuracy. As seen in Fig. 28, only one of
the test samples belonging to Class 3 (pinion with chipped tooth) is misclassified as
Class 4 (broken tooth pinion). The other test samples are nearly correctly classified.
Vibration-Based Fault Diagnosis of a Bevel and Spur Gearbox … 493

Healthy pinion Cracked tooth pinion

61 61
57 57
53 53
49 49
45 45
41 41
37 37
scale

scale
33 33
29 29
25 25
21 21
17 17
13 13
9 9
5 5
1 1
1 2 3 4 5 6 1 2 3 4 5 6
angle (degrees) [/60] angle (degrees) [/60]

Chipped tooth pinion Missing tooth pinion

61 61
57 57
53 53
49 49
45 45
41 41
37 37
scale

scale

33 33
29 29
25 25
21 21
17 17
13 13
9 9
5 5
1 1
1 2 3 4 5 6 1 2 3 4 5 6
angle (degrees) [/60] angle (degrees) [/60]

Fig. 22 Angular domain averaged WAMs representing 360° of rotation in segments of 60° each

Fig. 23 ANFIS error

convergence curve

Fig. 24 Trained ANFIS

classification accuracy
utilizing training samples
494 A. S. Ahuja et al.

Fig. 25 Trained ANFIS

classification accuracy
utilizing test samples

Fig. 26 Error convergence

curve utilizing wavelet
coefficients from only
optimal scales

Fig. 27 Trained ANFIS

classification accuracy
utilizing training data set
from optimal scales

Fig. 28 Trained ANFIS

classification accuracy
utilizing test set from
optimal scales
Vibration-Based Fault Diagnosis of a Bevel and Spur Gearbox … 495

8 Conclusion

The study presented in this Chapter aims at diagnosing faults in a bevel gearbox under
nonstationary conditions. Implementing the IAR technique through slight modifica-
tions in hardware, the inherently nonstationary gearbox vibration signatures are trans-
lated into the rotational domain. Continuous wavelet transformation of the merged
angular domain signals is undertaken and a range of scales with a relatively higher
energy–entropy ratios is identified. Independent rotational domain signals are frag-
mented into six equal sections, each representing 60° rotation of the pinion. Thus,
each angular domain averaged WAM consists of six data samples, each representing
60° of pinion rotation. CWCs from each segment are then input to an ANFIS classi-
fier directly. Reasonably good training and test accuracy is attained with continuous
wavelet coefficients derived from optimal scales being employed as direct inputs to
the ANFIS. The results of the study demonstrate the efficacy of the method proposed
in diagnosing defects in both the bevel as well as the spur gearbox.

References

1. N. Saravanan, S. Cholairajan, K.I. Ramachandran, Vibration-based fault diagnosis of spur bevel

gear box using fuzzy technique. Expert Syst. Appl. 36(2), 3119–3135 (2009)
2. N. Saravanan, V.N.S.K. Siddabattuni, K.I. Ramachandran, Fault diagnosis of spur bevel gear
box using artificial neural network (ANN), and proximal support vector machine (PSVM).
Appl. Soft Comput. 10(1), 344–360 (2010)
3. N. Saravanan, K.I. Ramachandran, Incipient gear box fault diagnosis using discrete wavelet
transform (DWT) for feature extraction and classification using artificial neural network (ANN).
Expert Syst. Appl. 37(6), 4168–4181 (2010)
4. N. Saravanan, V.N.S.K. Siddabattuni, K.I. Ramachandran, A comparative study on classifica-
tion of features by SVM and PSVM extracted using Morlet wavelet for fault diagnosis of spur
bevel gear box. Expert Syst. Appl. 35(3), 1351–1366 (2008)
5. N. Saravanan, K.I. Ramachandran, A case study on classification of features by fast single-shot
multiclass PSVM using Morlet wavelet for fault diagnosis of spur bevel gear box. Expert Syst.
Appl. 36(8), 10854–10862 (2009)
6. N. Saravanan, K.I. Ramachandran, Fault diagnosis of spur bevel gearbox using discrete wavelet
features and decision tree classification. Expert Syst. Appl. 36(5), 9564–9573 (2009)
7. M.A. Jafarizadeh, R. Hasannejad, M.M. Ettefagh, S. Chitsaz, Asynchronous input gear damage
diagnosis using time averaging and wavelet filtering. Mech. Syst. Signal Process. 22, 172–201
(2008)
8. N. Ahamed, Y. Pandya, A. Parey, Spur gear tooth root crack detection using time synchronous
averaging under fluctuating speed. Measurement 52, 1–11 (2014)
9. V. Sharma, A. Parey, Gear crack detection using modified TSA and proposed fault indicators
for fluctuating speed conditions. Measurement 90, 560–575 (2016)
10. A. Singh, A. Parey, Gearbox fault diagnosis under fluctuating load conditions with independent
angular re-sampling technique, continuous wavelet transform and multilayer perceptron neural
network. IET Sci. Meas. Technol. 11(2), 220–225 (2017)
11. A. Singh, A. Parey, Gearbox fault diagnosis under non-stationary conditions with independent
angular re-sampling technique applied to vibration and sound emission signals. Appl. Acoust.
144, 11–22 (2019)
496 A. S. Ahuja et al.

12. H. Li, Y. Zhang, Angle domain average and CWT for fault detection of gear crack, in Proceed-
ings of the IEEE Transaction, Fifth International Conference on Fuzzy Systems and Knowledge
Discovery, vol. 3 (Jinan, Shandong, China, 2008), pp. 137–141
13. G. Meltzer, Y. Ivanov, Fault detection in gear drives with non-stationary rotational speed—Part
I: the time-frequency approach. Mech. Syst. Signal Process. 17, 1033–1047 (2003)
14. G. Meltzer, Y. Ivanov, Fault detection in gear drives with non-stationary rotational speed—Part
II: the time-quefrency approach. Mech. Syst. Signal Process. 17, 273–283 (2003)
15. H.H. Bafroui, A. Ohadi, Application of wavelet energy and Shannon entropy for feature extrac-
tion in gearbox fault detection under varying speed conditions. Neurocomputing 133, 437–445
(2014)
16. H. Li, Y. Zhang, H. Zheng, Gear fault detection and diagnosis under speed-up condition based on
order cepstrum and radial basis function neural network. J. Mech. Sci. Technol. 23, 2780–2789
(2009)
17. J.D. Wu, C.C. Hsu, G.Z. Wu, Fault gear identification and classification using discrete wavelet
transform and adaptive neuro-fuzzy inference. Expert Syst. Appl. 36, 6244–6255 (2009)
18. I. Attoui, A. Boudiaf, N. Fergani et al., Vibration-based gearbox fault diagnosis by DWPT
and PCA approaches and an adaptive neuro-fuzzy inference system, in 16th International
Conference on Sciences and Techniques of Automatic Control and Computer Engineering
(STA) (Tunisia, 2015)
19. L. Zhang, G. Xiong, L. Liu, Q. Cao, Gearbox health identification by neuro-fuzzy ensemble.
J. Mech. Sci. Technol. 3, 603–608 (2018)
20. A. Soleimani, Gear fault identification using artificial neural network and adaptive neuro-fuzzy
inference system. Appl. Mech. 110–116, 2562–2569 (2012)
21. R. Polikar, The Wavelet Tutorial, 2nd edn., Part 1, Fundamental concepts and an overview of
the wavelet theory (2006). https://cseweb.ucsd.edu. Accessed Jun 2018
22. L. Jedlinski, J. Jonak, Early fault detection in gearboxes based on support vector machines and
multilayer perceptron with a continuous wavelet transform. Appl. Soft Comput. 30, 636–641
(2015)
23. J.S.R. Jang, Self-learning fuzzy controllers based on temporal back propagation. IEEE Trans.
Neural Netw. 3(5), 714–723 (1992)
24. J.S.R. Jang, ANFIS: adaptive-network-based fuzzy inference system. IEEE Trans. Syst. Man
Cybern. 23(3), 665–685 (1993)