12th Conference on Learning Factories, CLF2022
IoT and machine learning enable predictive maintenance for
manufacturing systems: a use-case of laser welding machine
implementation
Van Tung Tran∗, Chan Hian Leng, Parthasarathi Sagunthala, Lim Choon Pio, Chua Yong Quan
Singapore Institute of Manufacturing Technology (SIMTech), A*STAR, Singapore
Abstract
The concept of predictive maintenance (PdM) has been existing for many years but rarely implemented in industries due to their
high cost. With the advent of big data solutions, the Internet of Things (IoT), and machine learning, PdM has been affordable and
enabled to improve the maintenance capabilities in industries. In this paper, an IoT and machine learning based PdM framework
is presented. It consists of sequential modules starting from IoT sensing and edge devices for gathering information of critical
components of which dysfunctions can cause the machine shutdown. These data are then streamed to server via MQTT protocol
to store in database as historical data and to simultaneously perform fault detection and diagnosis in real-time by models built by
machine learning algorithms. A web-based dashboard is designed to assist users in training model and visualizing data analysis
results. The PdM framework is implemented for a laser plastic welding machine in Model Factory@SIMTech as a demonstration
for manufacturing industry applications.
© 2022 The Authors. This is an open access article.
Peer-review statement: Peer-review under responsibility of the scientific committee of the 12th Conference on Learning Factories
2022.
Keywords: IoT; Condition monitoring; Predictive maintenance; Fault diagnosis; Laser transmission welding machine
1. Introduction
In manufacturing operation, several factors e.g. material defects, fatigue, overloading, vibration, etc. can cause
degradation and failures in machinery equipment even it is designed and fabricated with intention to be highly robust
and long lifetime. These unexpected failures interrupt the production schedule, reduce the product quality, and lead
to costly downtime. Hence, maintenance is the necessity of rectifying the happened failures so that the machinery
equipment can be reinstated and the production facilities are sustained. Maintenance, on the other hand, is always
a high cost source in manufacturing industry as it is a large part contribution toward total production costs, from
15% to 70% depending on the industry as reported in studies [1, 2]. In response to globally intensified competition,
∗ Corresponding author. Tel.: +65-6319 4493.
E-mail address: Tung
[email protected] Electronic copy available at: https://ssrn.com/abstract=4073901
2 V.T. Tran et al., CLF 2022
manufacturing companies need to optimize production processes on one side and minimize the maintenance cost
with a efficient maintenance policy on the other side. Though, maintenance policies in the manufacturing industry
are still outdated and inefficient where corrective maintenance and preventive maintenance have been dominant [3].
Corrective and preventive maintenance policies have significant drawbacks, which are labor intensive and unavoidable
the unplanned downtime events (corrective maintenance) or high cost due to frequent replacements the expensive
components that are in their lifetime (preventive maintenance).
Condition-based maintenance (CBM) is an alternative and improved policy compared with the two previous ones
since it allows to continuously monitor equipment operation conditions to identify the need for maintenance in real-
time. This policy, however, still has issues. The main issue is that, even with continuous monitoring, the acquired data
only indicate the changes of machine condition before failure occurrence. It does not provide a tool to estimate the
future time period in which the abnormalities happen, and the failure might occur consequently. Hence, an efficient
schedule for maintenance activities ahead, for instance finding the optimal time for shutting down the manufacturing
process, is unable to be planned [3]. As a result, a new policy called predictive maintenance (PdM) wherein CBM
is a subcategory [4] is proposed to improve maintenance capability. According to studies [5–7], PdM enables to
lessen 25% – 35% the maintenance costs, eliminate 70% – 75% breakdowns, reduce 35% – 45% breakdown time,
and increase 25% – 35% production. A comprehensive PdM system is comprised multiple phases involving condi-
tion monitoring, fault detection and diagnosis, degradation assessment and prognosis, and maintenance activities and
planning. Some of these phases may be presented or not and could be customized depending on the equipment ar-
chitecture complexity, the system requirements, the available knowledge and data. The condition monitoring is one
of key phases used to monitor performance and operational condition of equipment. However, the cost of equipment
needed for the condition monitoring is often high leading to the limit of implementation over the past few decades.
In recent years, thanks to IoT and advanced digital technologies in Industry 4.0, the condition monitoring is more
affordable. Together with the development of modern artificial intelligence (AI) where machine learning and deep
learning provide advancement, big data solution, cloud computing, and edge computing in the context of new industry
revolution, PdM has been further attracted a lot of interest from academia and industry.
As IoT enables presence of abundant sensors, therefore, huge real-time data collected from equipment in man-
ufacturing processes, it demands an efficient framework for handling the data streaming continuously and instantly
to provide valuable information and knowledge of system dynamics in maintenance context. With this aim, such
an architectural framework for PdM is presented in this paper, which could be ambitiously applied as a generic
scheme for various manufacturing machinery. The framework consists of IoT devices for data acquisition (DAQ),
timestamp synchronization with different data sources, data exchange via the Message Queuing Telemetry Transport
(MQTT) protocol and preprocessing, and machine learning algorithms for fault detection and diagnosis. A real-time
implementation equipped by a web-based user interface of the proposed framework for a laser welding machine in
ModelFactory@SIMTech is also introduced as a use-case of the industrial applications.
2. Related work
Generally, IoT-based PdM framework is variety depending on how edge devices are used, where the main com-
puting and data analysis are carried out, and what the type of AI model is developed. There are two main schemes
to employ the edge devices, that are edge DAQ and edge AI. In the former, edge devices are purely responsible for
gathering the sensor reading values and pass these through field gateways and network to feed the AI analytic models,
which is stored in either physical server or cloud server to carry out data processing, computing, and analysis. In the
latter, edge devices are used not only to acquire data but also perform the core tasks of predictive maintenance to
produce condition-related information for the analytic platform.
There are many frameworks proposed in the literature using the edge DAQ scheme. For example, Ayvaz et. al [8]
proposed a framework for a personal care goods plant. In their framework, data collected through different built-in
IoT sensors were converted into a single data type using MQTT protocol. Next, a sequential process including data
processing, imbalanced handling, correlation analysis, and machine learning algorithms were executed on-premise
servers to detect signals for potential failures. Ruiz-Sarmiento et. al. [9] proposed a similar framework and applied
for machinery of the hot rolling process and the heating coilers in stainless steel industry. Other frameworks can be
found in studies [10, 11]. For the edge AI approach, fewer studies have been published. For example, Katona and
Electronic copy available at: https://ssrn.com/abstract=4073901
V.T. Tran et al., CLF 2022 3
Fig. 1. The proposed PdM framework
Panfilov [12] proposed a framework that can distribute the computing capability to both edge devices and server
and applied it for HVAC system. Although the edge AI approach can reduce the delayed response times due to the
latency and avoid a high volume of data being transferred over the network, the edge DAQ has been widely used in
industry applications so far. The reason is that the edge computing capability is still limited for some AI models where
heavy training process is conducted. Furthermore, estimating the remaining lifetime using data driven methodologies
requires run-to-failure data, which need to be stored in physical or cloud server.
3. The proposed PdM framework
The proposed framework based on edge DAQ scheme in this paper is showed in Fig. 1. It starts with gathering
machinery data from different sources. These data sources include PLC data logger which is available in the machine
and/or a new sensorized system to monitor the conditions of components which are frequently failed during the
machinery usage. The add-in system could also be a redevelopment facilitated by brownfield connectivity to improve
the machinery capabilities. Due to the device’s clock difference, an IoT edge device is used not only as a DAQ
system to acquire the data from the existing system via MQTT protocol and new added sensors, but also as a time
synchronizing device to make sure that the machinery condition is collected as the same time. To do this, timestamp of
each value from a device in the system is used as the reference to finding timestamp of values from the other devices
to be the synchronized data. The synchronized data are then streamed to an on-premise server that contains modules
of data preprocessing, feature engineering, fault detection and root cause classification, and prognosis for further
processing via API to obtain condition-related information. Firstly, product cycle identification and pre-processing
procedures in the data preprocesing module are carried out to separate data belonging a product and transform their
values into a suitable format. Next, the feature engineering module involving feature extraction and feature selection
Electronic copy available at: https://ssrn.com/abstract=4073901
4 V.T. Tran et al., CLF 2022
are performed on the product cycle data. As a result, a feature sample x(i) = x1(i) , x2(i) , · · · , xn(i) , where i is the current
product cycle and n is the number of selected features, is obtained. x(i) is concurrently stacked to database as historical
data for prognosis and fed into a trained machine learning model for online diagnosing the current condition of
machine. Depending on the diagnostic result, if any fault occurs in the machine, the faulty type is described to provide
information to operators to carry out maintenance action in order to restate the machine functionality.
4. A use case of the proposed framework
4.1. Problem statement
In this study, we present a use case of the proposed framework applied for a laser plastic welding machine at
Model Factory@SIMTech. Laser plastic welding or laser transmission welding is a well-known joining technique that
is widely utilized for plastic welding applications in medical, jewelry industry, automotive, and electronics. Unlike
others, energy in the transmission welding is concentrated in between two overlapping parts at their interface. The
laser beam is transmitted through the top transparent part and absorbed by the bottom absorbent part where the laser
energy is converted into heat. By conduction heat transfer from the bottom part to the top part, the polymers are melted
at the interface and a weld seam is formed through molecular diffusion. According to [13], there are several parameters
that affect to the laser welding process as well as product quality in which laser intensity and laser interaction time are
the main parameters.
The laser intensity I p and interaction time T are given as [14]:
(1 − Rt )P
Ip = (1)
πD2
L
T = (2)
S
where Rt , P, and D are the reflectivity of the transparent material (constant), laser power, and beam diameter, respec-
tively. Combining I p and T gives us the laser energy density IE in form:
(1 − Rt )P L
IE = (3)
πD2 S
As shown in equation (3) for a unit length of the weld, the energy input is independently governed by the laser
intensity, welding speed, and beam diameter. Because of that the welding speed and beam diameter are commonly
constant during the machine usage, the laser intensity (laser power) is the main parameter that its degradation over
time affects the welding process and welding seam quality. In order to monitor the laser intensity or laser power,
different types of sensors such as photodiode, thermopile, pyroelectrics, and beam profile can be used and are generally
expensive.
4.2. Sensorization and data acquisition
To diagnose the laser intensity, many machine parameter e.g. machine power, voltage, are acquired by PLC data
logger. In addition, joining pressure created by clamping force is also a key parameter that needs to be monitored.
This force is generated by a pneumatic cylinder, measured by two load cells installed on surface of the mechanical
clamp, and is read by an add-in PLC. The pneumatic cylinder is an another object of machine in maintenance, which
its leakage due to seal wear-out is the main failure mode and causes joining pressure reduction. A pressure sensor
is added to the sensing system and its measured values are collected by low cost card 16-bit resolution MCC 128
attached on top of a edge device. To synchronize the timestamp of the three data sources, the machine parameters
are published and subscribed by the edge device. The timestamp of these are then used as the reference to search the
matching timestamp of load cells and pressure values.
Electronic copy available at: https://ssrn.com/abstract=4073901
V.T. Tran et al., CLF 2022 5
Fig. 2. The training mode dashboard
4.3. Web-based dashboard
The analytic results are visualized on a dashboard designed as two modes: training mode is to train fault detection
and diagnosis models and live mode is to perform the trained model with the real-time data. In the training mode
as shown in Fig. 2, users can either select available data or conduct experiments to get new training sets. With the
sets selected, statistical feature extraction in the time domain and selection process are automatically carried out. As
a result, the important features that correlate to the laser intensity are obtained based on the threshold which can
be adjusted on the dashboard. In Fig. 2, the selected features are plotted as orange color and the discarded features
are depicted blue color. Various machine learning models are available for user choices, consisting of Decision Tree,
Random Forest, Neural Network, Support Vector Machine, K-Nearest Neighbor, and Gradient Booting.
The dashboard also allows users to input the hyper-parameters for each machine learning model, which is either
auto or manual. The auto mode uses Exhaustive Grid Search algorithm to help users tuning the hyper-parameters
automatically, while the manual mode requires users having basic knowledge of machine learning to provide the
parameters for each method accordingly. Finally, the training results are shown the performance of training and test
accuracy and F1 score on the gauge charts.
Fig. 3 shows the live mode dashboard where acquired data, features extracted, and diagnostic results are performed
in real-time. Data are acquired and product cycle detection procedure continuously check the input data to find the
started and ended points of a product cycles, so that the laser intensity of each product cycle is identified. The most
important features are also showed on the dashboard as condition monitoring process to indicate the degradation of
machine component.
5. Conclusions and future work
The framework for predictive maintenance based on IoT and machine learning has been presented in this paper. It
is equipped by multiple functions that can handle core tasks of predictive maintenance in real-time for manufacturing
systems. As deployed in architectural microservices, it is easier to build and maintain, flexibility, and scalability. The
proposed framework has been implemented to the laser transmission welding machine in Model Factory@SIMTech
as a use case. The framework is still in a developed progress; hence, it lacks of fault diagnosis for multiple critical
components and other key components such as remaining useful lifetime estimation. In future, fault diagnostic models
Electronic copy available at: https://ssrn.com/abstract=4073901
6 V.T. Tran et al., CLF 2022
Fig. 3. The live dashboard
for the pneumatic cylinder and rotary indexing table will be developed to enhance the framework to be at system
level. Moreover, predicting the lifetime of the critical components will be added to the framework to make it be
comprehensive and applicable in manufacturing industry. The real-time application of the framework for the laser
transmission welding machine can be also used for hands-on practices of training courses organized by SIMTech.
References
[1] C. Krupitzer, T. Wagenhals, M. Züfle, V. Lesch, D. Schäfer, A. Mozaffarin, J. Edinger, C. Becker, S. Kounev, A survey on predictive mainte-
nance for industry 4.0. arXiv preprint arXiv:2002.08224 (2020).
[2] P. Aivaliotis, Z. Arkouli, K. Georgoulias, S. Makris, Degradation curves integration in physics-based models: Towards the predictive mainte-
nance of industrial robots, Robotics and Computer-Integrated Manufacturing, 71 (2021) 102177
[3] A. Prajapati, J. Bechtel, S. Ganesan, Condition based maintenance: a survey, Journal of Quality in Maintenance Engineering, 18 (2012) 384-
400.
[4] A. Theissler, J. Pérez-Velázquez, M. Kettelgerdes, G. Elger, Predictive maintenance enabled by machine learning: use cases and challenges in
the automotive industry, Reliability Engineering & System Safety, 215 (2021) 107864
[5] G.P. Sullivan, R. Pugh, A.P. Melendez, W.D. Hunt, Operations & maintenance best practices: a guide to achieving operational efficiency, US
Department Energy, Fed Energy Management Program, 2010 (https://doi.org/10.2172/1034595).
[6] K. Matyas, T. Nemeth, K. Kovacs, R. Glawar, A procedural approach for realizing prescriptive maintenance planning in manufacturing indus-
tries, CIRP Annals - Manufacturing Technology, 66 (2017) 461-464
[7] R. Gao, L. Wang, R. Teti, D. Dornfeld, S. Kumara, M. Mori, M. Helu, Cloud-enabled prognosis for manufacturing, CIRP Annals-Manufacturing
Technology, 64 (2) (2015) 749-772.
[8] S. Ayvaz, K. Alpay, Predictive maintenance system for production lines in manufacturing: A machine learning approach using IoT data in
real-time. Expert Systems with Applications, 173, 114598.
[9] J.R. Ruiz-Sarmiento, J. Monroy, F. Moreno, F.A. Moreno, C. Galindo, J.M. Bonelo, J. Gonzalez-Jimenez, A predictive model for the mainte-
nance of industrial machinery in the context of industry 4.0, Engineering Applications of Artificial Intelligence, 87 (2020) 103289
[10] E. Traini, G. Bruno, G. D’antonio, F. Lombardi, Machine learning framework for predictive maintenance in milling, IFAC-PapersOnLine 52
(2019) 177-182.
[11] Z. Li, Y. Wang, K.S. Wang, Intelligent predictive maintenance for fault diagnosis and prognosis in machine centers: Industry 4.0 scenario,
Advances in Manufacturing, 5 (2017) 377-387.
[12] A. Katona, P. Panfilov, B. Katalinic, Building predictive maintenance framework for smart environment application systems, Proceedings of
the 29th DAAAM International Symposium, 2018, pp. 0460-0470.
[13] B. Acherjee, Laser transmission welding of polymers - A review on welding parameters, quality attributes, process monitoring, and applica-
tions, Journal of Manufacturing Processes, 64 (2021) 421-443.
[14] B. Acherjee, A.S. Kuar, S. Mitra, D. Misra, Modeling and analysis of simultaneous laser transmission welding of polycarbonates using an
FEM and RSM combined approach. Optics & Laser Technology, 44 (2012) 995-1006.
Electronic copy available at: https://ssrn.com/abstract=4073901
