Ocean Engineering 152 (2018) 404–415

Contents lists available at ScienceDirect

Ocean Engineering
journal homepage: [Link]/locate/oceaneng

Predicting ship machinery system condition through analytical reliability

tools and artificial neural networks
I. Lazakis a, Y. Raptodimos a, *, T. Varelas b
a
Department of Naval Architecture, Ocean and Marine Engineering, University of Strathclyde, Glasgow G4 0LZ, United Kingdom
b
Danaos Shipping Company Ltd., Piraeus 185 45, Greece

A R T I C L E I N F O A B S T R A C T

Keywords: Inadequate ship machinery maintenance can increase equipment failure posing a threat to the environment,
Predictive maintenance affecting performance, having a great impact in terms of business losses by reducing ship availability, increasing
Fault tree analysis downtime and moreover increasing the potential of major accidents occurring and endangering lives on-board.
FMEA This paper aims to provide a systematic approach for identifying critical ship machinery systems/components
Artificial neural networks and to analyse their physical parameters. Critical ship main engine systems/components are used as input in a
Time series
dynamic time series neural network, in order to monitor and predict future values of physical parameters related
to ship critical systems. The critical main engine systems/components and their relevant parameters to be
monitored are identified though the combination of Fault Tree Analysis (FTA) and Failure Mode and Effects
Analysis (FMEA). A case study of a Panamax size container ship is presented in which Artificial Neural Networks
(ANN) are used to predict the upcoming values of all main engine cylinders exhaust gas temperatures. The
forecasted results were validated through comparison with actual observations recorded on board the ship. The
proposed hybrid methodology successfully presents a systematic approach for initially identifying critical sys-
tems/components through reliability modelling and tools and subsequently monitoring their physical parameters
through neural networks.

1. Introduction structured Condition Based Maintenance (CBM) regime. This may not
replace all planned maintenance, but it can possibly reduce downtime,
Maintenance tasks affect the reliability and availability levels of the inspection and unnecessary servicing work. The advocates of change
shipping industry and are important factors in the lifecycle of a ship that argue that a move from scheduled, rule-based maintenance to a
can minimize down-time and reduce operating costs as is accounts for data-driven, risk-based approach can lead to more accurate and timely
20%–30% of a ship's operational expenses (Stopford, 2009). Also, due to maintenance, resulting in lower costs, greater availability of ship systems
the impact of shipping on the environment and the importance of the safe and increased safety (Tinsley, 2016).
operation of ships; ship owners and operators pursue to adopt a main- In this respect, according to British Standard (2012), CBM is defined
tenance plan and procedures that will reduce costs and promote the as the maintenance policy carried out in response to a significant dete-
lifecycle integrity of the ship. Although the maritime industry is rioration in a machine as indicated by a change in a monitored parameter
responsible for the massive transportation of goods worldwide, it is only of the machine condition. The heart of CBM is condition monitoring
recently that new approaches investigating the enhancement of ship's which aims in collecting data regarding equipment conditions. Condition
reliability, availability and profitability have been considered according monitoring technologies are applied through various tools by recording
€
to Lazakis and Olçer (2015). Though the industry is still predominantly and evaluating different measureable parameters. Data can include vi-
reliant on a time-based, prescriptive approach to maintenance, there are bration, acoustic, temperature, current signal, oil and lubricant mea-
a number of factors challenging the long-held norm. The increasing surements (Pascual, 2015). Compared to other industrial applications, in
complexity of shipboard systems, heightened expectation and competi- the shipping industry data pooling is not always possible as similar
tive needs as to ship and plant availability and efficiency and the influ- equipment in different conditions may have different failure patterns.
ence of the data revolution on vessel operations, favour a properly Another issue is the constant appearance of new equipment, which makes

* Corresponding author.
E-mail address: [Link]@[Link] (Y. Raptodimos).

[Link]
Received 23 February 2017; Received in revised form 14 September 2017; Accepted 8 November 2017
Available online 16 November 2017
0029-8018/© 2017 Elsevier Ltd. All rights reserved.
I. Lazakis et al. Ocean Engineering 152 (2018) 404–415

historical records obsolete. Moreover, data is not collected in a stand- continuous survey machinery cycle. Class NK introduced the concept of
ardised way so that it can lead to more informed and successful decision PrimeShip-Total Ship Care (2013) which has been designed to prevent
making (Dekker, 1996). Technological advances and high cost of pollution of the marine environment and ensure safety of ships at every
ownership have resulted in considerable interest in advanced mainte- stage of a ships life including maintenance. The product contributes to
nance techniques. Raza and Liyanage (2009) stated that there has been improved reliability and increased efficiency of hull structure analysis,
an increasing demand for testing and implementing intelligent tech- machinery shaft alignment and torsional vibrations and maintenance
niques as a subsidiary to existing condition monitoring programs and that management amongst others. DNV-GL also published a paper in 2014
ANNs have emerged as one of the most promising techniques in regarding condition monitoring in the shipping industry reviewing
this regard. existing condition monitoring technologies and methods for imple-
The question of how much data, which data, and how often this menting such technologies (Knutsen et al., 2014). The American Bureau
should be collected and how has also risen; as although companies adopt of Shipping (ABS) introduced the NS5 Entreprise software
CBM schemes, there seems to be an issue in processing, analysing and (2015b) aiming to handle the primary functions of operational
utilising the recorded operational data. This paper aims to identify crit- management-maintenance, supply chain, workforce, environmental and
ical systems of a ship's main engine through the combination of the FTA safety. Classification Societies encourage condition monitoring tech-
and FMEA tools and to further examine them through monitoring their niques on board ships, offer guidelines but do not oblige ship operators or
physical parameters. The physical parameters are used as input for time owners to implement such techniques in their operation and mainte-
series analysis and forecasting using artificial neural networks. €
nance. Furthermore, Lazakis and Olçer (2015) introduced a Reliability
The present research paper is organized as follows: Section 2 presents and Criticality Based Maintenance (RCBM) strategy by utilising a fuzzy
the maintenance overview and status in the maritime industry alongside multiple attributive group decision-making technique, which is further
the methodologies/tools implemented in detail. In Section 3 the sug- enhanced with the employment of Analytical Hierarchy Process (AHP).
gested methodology is demonstrated and explained. Section 4 presents The outcome of this study indicated that preventive maintenance is still
the case study application through which the methodology is applied the preferred maintenance approach by ship operators, closely followed
alongside the obtained results. Finally, the discussion and conclusion of by predictive maintenance; hence, avoiding the ship corrective mainte-
this research study is presented in Section 5. nance framework and increasing overall ship reliability and availability.
However, most shipping companies still follow the so called Planned
2. Research background/literature Maintenance System (PMS) based on ISM code (IMO, 1993) which is
centred on the preventive maintenance scheme, in which machinery
The evolution of maintenance was based not solely on technical but items are maintained based on operating hours or calendar intervals. A
rather on techno-economic considerations according to Pintelon and brief discussion of the methods applied in this paper is presented below,
Parodi-Herz (2008). Furthermore, according to Arunraj and Maiti while the specifics of how they were applied to the case study are given in
(2007), maintenance policies can be categorised into four generations as detail in Section 3.
seen in Fig. 1. The fourth generation is the most recent one, which focuses
on condition based maintenance, condition monitoring and failure 2.1. Failure modes and effects analysis (FMEA)
eliminations. It concentrates on reducing the proportion of equipment
failures and overall levels of failure probability through various tools and FMEA provides a systematic method for organizing the study of a
strategies, based on preventive and predictive maintenance approaches. particular system or process in terms of failure analysis. The aim of FMEA
Examining Fig. 1 and associating it with maintenance in the shipping is to review the system in order to provide details on how to identify
industry, it can be observed that maintenance was initially treated as a failures and their causes as well as determine the end results of the
necessary evil and has now started to be considered an important factor failures occurring. Thus, FMEA is a formalized method to consider all
in the operational management and lifecycle of ships. Lloyds Register components, their functions, failure modes and system failures (Iser-
(2013) published a report regarding machinery planned maintenance mann, 2006). It involves reviewing as many components, assemblies and
and condition monitoring which was revised in order to add machinery subsystems as possible to identify possible failures modes and causes and
condition based maintenance procedures and describe how a machinery effects of such failures. FMEA can be applied in a bottom-up approach
planned maintenance scheme can be accepted as an integral part of the which assists in mapping the overall failure potential of the system. This

Fig. 1. Evolution of maintenance (Adapted by Arunraj and Maiti, 2007).

405
I. Lazakis et al. Ocean Engineering 152 (2018) 404–415

technique is most suited for the risk assessment of mechanical and events in each fault tree branch. A static gate indicates that the order of
electrical systems and the approach can be either quantitative or quali- the inputs of a gate do not matter, therefore are not sequence-dependent
tative. According to Ben-Daya and Knezevic (2009) FMEA performs three as in dynamic gates. On the other hand, in dynamic gates, the order of the
functions. These are initially the identification and recognition of po- occurrence of input events is vital for determining the output. If dynamic
tential failures including their causes and effects, the evaluation and gates are used, then the fault tree becomes a dynamic fault tree. A
prioritization of identified failure modes and the identification and thorough description of advanced fault tree symbols can be found in
suggestion of actions to either eliminate or reduce the chance of the NASA (2002) and NUREG-0492 (1981).
potential failures from occurring. A FTA is normally carried out by defining the FTA scope and identi-
The American Bureau of Shipping has released guidance notes (ABS, fying the top event, followed by defining the levels of the FT and con-
2015a) related to FMEA requiring the development and submission of nected the levels by using gates (Rausand and Arnljot, 2004). The FTA
FMEAs as part of Classification requirements for certain systems such as can be conducted in a qualitative or quantitative manner, depending on
Dynamic Positioning systems, drilling systems, dual fuel diesel engines the type of data available. If no data is available, a fault tree can be
etc. The International Association of Classification Societies (IACS) analysed qualitatively by using minimal cut sets. Finding minimum
(2014) published recommendations for the FMEA process for diesel en- cut-sets provides insight into weak points of complex systems. Qualita-
gines and reporting the FMEA process. To contribute and improve the tive analysis is used to identify what combinations of events cause the
ongoing efforts of Classification Societies and operators, Cicek and Celik top event to occur. Ultimately, failure rates are derived from
(2013) examined the application of FMEA in order to prevent and reduce well-substantiated historical data, including MTBF of components, units
the occurrence of crankcase explosion failure in order to improve ma- and subsystems (Pascual, 2015). FTA uses failure rates, mean time be-
chinery system reliability and enhance operational safety on board ships. tween failures and minimal cut sets to evaluate the reliability and
FMEA can favourably be combined with FTA because it yields the availability of the system.
possible system failures, which are inputs of FTA. Also, an FTA may use Specifically, Lazakis et al. (2010) presented a predictive maintenance
the basic failure mode FMEA records or an effect summary as one of its strategy utilising Failure Modes, Effects & Criticality Analysis (FMECA)
inputs (basic events). Therefore FMEA and FTA complement each other and FTA by considering the existing ship maintenance regime as an

(Isermann, 2006). Specifically, Souza and Alvares (2008) applied FMEA overall strategy including technological advances and decision support
in conjunction with Fault Tree Analysis (FTA) as a risk assessment tool for system by combining existing ship operational and maintenance tasks
the application of Reliability Centred Maintenance. The methodology with the FTA and FMECA tool. An innovative ship maintenance strategy
was used to study and analyse the failure mode of a hydraulic Kaplan is also presented by Turan et al. (2011) based on criticality and reliability
turbine and showed that the two tools can complement each other for the assessment while utilising the FTA tool with time-dependent dynamic
execution of an effective predictive maintenance plan on the basis that gates in order to accurately present the interrelation of the components
the FMEA analysis provided the information required for the FTA basic for a diving support vessel. However, the above papers utilise the Fault
event. Hidalgo et al. (2011) carried out the failure analysis of steering tree capabilities by using quantitative methods of analysis and data such
systems for LNG carriers. FTA was developed in order to identify the most as FR, MTBF. Laskowski (2015) applied the FTA as a tool for modelling
critical components for the steering gear system and then the application the reliability structure of a marine main engine by conducting qualita-
of FMEA was conducted for each critical component in order to identify tive means of analysis using the minimal cut sets method. Moreover,
the failure modes and provide appropriate maintenance policies based on Guan et al. (2016) presented a Fault Tree model considering fires and
reliability centred maintenance philosophy. Furthermore, Gao and Kang explosion in a dual fuel engine room as the top event. The primary factors
(2016) applied the FMEA method for the reliability analysis of the main that affect these kinds of accidents are determined through minimum cut
failure events and their interrelations of the offloading systems of an sets and based on the results; suggested measures are proposed to
FPSO. The main failure events were then demonstrated using a improve safety and reliability. Anantharaman et al. (2014) created a
Fault Tree. Fault Tree for a two stroke main engine lube oil system in order to
examine the reliability of the overall system and identify critical com-
2.2. Fault tree analysis (FTA) ponents and demonstrated that with the use of additional components in
the system, component reliability could be increased which contributed
FTA is a systematic technique used for acquiring information on a to the overall reliability of the main engine lubrication system.
system and finding out how the system or its components could
contribute to a failure and can assist decision-making process developed 2.3. Artificial neural network (ANN)
by safety and maintenance engineers who plan and organize mainte-
nance and monitoring activities (Manzini et al., 2009). It is a A neural network can be defined according to Haykin (1998) as a
failure-oriented, deductive, top-down approach which considers an un- massively parallel distributed processor made up of simple processing
desirable event associated with the system as the top event; the various units that has a natural propensity for storing experiential knowledge and
possible combinations of fault events leading to the top event are rep- making it available for use. It resembles the brain within two respects;
resented with logic gates. Therefore, the fault tree is a qualitative model knowledge is acquired by the network from its environment through a
which provides useful information on the various cases of undesired top learning process and interneuron connection strengths, known as syn-
events (Verma et al., 2010). The graphical representation is done through aptic weights, are used to store the acquired knowledge. An ANN consists
Fault Tree diagrams which are a graphical design technique following a of interconnection of neurons assembled in layers. Each layer has a
top-down approach. It uses a graphic model of the pathways within a number of simple, neuron processing elements called nodes or neurons
system that can lead to a projected, undesirable event or failure. The that interact with each other by using numerically weighted connections.
pathways interconnect contributory events and conditions, using stan- It consists of n layers of neurons of which two are input and output layers,
dard logic symbols and the basic constructs in a fault tree diagram are respectively. The former is the first and the only layer which receives and
gates and events. The fault tree analysis module is based on sets of rules transmits external signals while the latter is the last and the one that
and logic symbols from probability theory and Boolean algebra. Gates sends out the results of the computations. The n-2 inner ones are called
represent logic operators that link the various branches of the fault tree hidden layers which extract, in relays, relevant features or patterns from
together, can be either static or dynamic and determine whether the top received signals. Those features considered important are then directed
event can occur or not. The gates show the relationship of events needed to the output layer. Sophisticated neural networks may have several
for the occurrence of a higher event and serve to permit or inhibit the hidden layers, feedback loops, and time-delay elements, which are
fault logic up the tree. Basic events can be defined as the lower level designed to make the network as effective as possible in discriminating

406
I. Lazakis et al. Ocean Engineering 152 (2018) 404–415

relevant features or patterns. Fig. 2 displays a simple structure of a typical could improve in accuracy by increasing the training parameters of the
ANN with one input, hidden and output layer respectively. back propagation algorithm. Furthermore, Guoping et al. (2013) com-
According to Nasr et al. (2012) ANNs provide an effective analysing bined a Fault Tree for an excavator hydraulic system with ANN for fault
and diagnosing tool to understand and simulate the nonlinear behaviour diagnosis. Their results demonstrated that the network was feasible and
of complex systems and can be used as a valuable performance assess- effective for fault diagnosis and further research should be conducted due
ment tool for operators and decision makers. ANN are trained to learn to the optimistic results of the methodology. Based on the literature
from past examples and capture subtle functional relationships among presented, a novel methodology is hence required to address the above
data provided even if the underlying relationships are hard to describe or identified gaps and shortcomings.
unknown. They can readily address modelling problems that are
analytically difficult and for which conventional approaches are not 3. Methodology
practical, including complex physical processes having nonlinear,
high-order, and time-varying dynamics and for which analytic models The overall methodology applied in this paper is illustrated in Fig. 3.
might not yet exist (Peng et al., 2010). Zhang et al. (2001) showed that Initially, a generic model is developed utilising the FMEA and FTA
neural networks are valuable tools for modelling and forecasting modelling tools in order to identify the most critical subsystems and
nonlinear time series while traditional linear methods are not as components of the main system under investigation. Subsequently, the
competent for this task. The lack of systematic approaches to neural ANN model is developed in order to transform the generic model into a
network model building is probably the primary cause of inconsistencies more specific one, by using physical parameters of the most critical items
in reported findings. of the FTA as input in the neural networks for time series analysis and
Although artificial neural networks have recently gained importance forecasting.
in time series applications (Aizenberg et al., 2016; Szoplik, 2015; Liu
et al., 2015; Laboissiere et al., 2015), some methodological shortcomings 3.1. FTA & FMEA model
still continue to exist, such as proper network selection, architecture and
learning algorithms. Aizenberg et al. (2016) performed time series System definition is a very important part of FTA. The basis of the
analysis using multilayer neural network for forecasting oil production in system definition is the Fault Tree diagram which defines all in-
the Gulf of Mexico. They concluded that the choice of embedding di- terconnections and components of the system. Also, reliability parame-
mensions from time series data is a challenging and ongoing task ters should be identified and the system definition must provide all
requiring additional research effort. Moreover, Noor et al. (2016) applied related assumptions regarding the system and the conditions that indi-
ANN modelling on a marine diesel engine in order to predict its perfor- cate that the components of the system have failed. The top gate should
mance in terms of output torque, brake power, brake specific fuel con- also be clearly defined as if a top event is not concisely defined then the
sumption and exhaust gas temperature using as input data various engine Fault Tree can possibly become too large and complex, resulting in an
speeds and loads. The network was based on a standard back-propagation unfocused system analysis. Logical relationship between top events and
Levenberg-Marquardt training algorithm and results were compared basic events of the Fault Tree can also be verified based on FMEA analysis
with those of a mathematical model. Results showed that the prediction and results. The two tools can complement each other on the basis that
error of the ANN model was lower than the mathematical model. Rap- the FMEA provides information required for the FTA basic event. Thus,
todimos and Lazakis (2016) created a dynamic autoregressive neural both can provide a complementary way of identifying errors and tracking
network for predicting the upcoming 5 hourly exhaust gas temperatures their possible influences. The connection and combination of these two
for a cylinder of a marine diesel engine. The prediction results were tools for this methodology is graphically demonstrated in Fig. 4.
compared with actual data and showed that the network was accurate in The FTA is a deductive, top-down approach in which the main system
predicting these values. (top event) is divided into its subsystems and they are consequently
In terms of combining reliability tools with ANN, Ruilin and Lowndes analysed to their components, which is the last level of the fault tree
(2010) proposed the use of a coupled FTA and ANN model in order to (basic events). As shown in Fig. 4, the basic events of the Fault Tree are
improve the prediction of the potential risk of coal and gas outburst used as input in the FMEA. The FMEA is an inductive, bottom-up
events during mining. Parameters identified from the FTA were subse- approach in which the components of the Fault Tree (basic events) are
quently used as input variables to a predictive ANN model in order to the failed items of the FMEA worksheet. The system tab in the worksheet
predict the potential risk occurrence of coal and gas outburst events. is considered the subsystems of the Fault Tree. For each failed item, all
Their findings presented that the FTA was successful in determining the possible failure events and failure causes are considered alongside their
critical events of the system and the ANN model although consistent, effect in a local and global scale for the subsystems and top event
respectively. Therefore, the combination of the FTA and FMEA provides
significant information regarding critical components, their possible
failures and causes and their effects not only in component, but also in
subsystem level, which affects the overall performance and reliability of
the main system (top event). Furthermore, the referred combination also
provides important information regarding which physical parameters
affect the performance of components as fluctuations in these physical
parameters can lead to possible failures.
The following steps are performed for the construction of a Fault Tree
(NASA, 2002): definition of the FTA scope, identification of the top
event, identification of the first level events, connection of the first level
events with the top event by means of gates, identification of the second
level events, connection of the second level events with first level by
using gates, repetition of the above steps for all subsequent event levels.
FTA can be performed qualitatively or quantitatively. If data is unavai-
lable, a Fault Tree can be analysed by using the minimal cut sets method.
By using qualitative analysis, the combinations of events that cause the
top event to occur can be identified.
Fig. 2. Artificial neural network structure (Zhang et al., 1998). A cut set is a set of basic events, which if they all occur, will result in

407
I. Lazakis et al. Ocean Engineering 152 (2018) 404–415

Fig. 3. Overall hybrid Predictive Methodology.

the top event of the fault tree occurring. A minimal cut set is a combi- Gates are logic operators within the Fault Tree that determine how
nation (intersection) of primary events sufficient for the top event. The events are generated and can be either static or dynamic. A basic event
combination is a minimal combination in that all the failures are required represents the lowest level of a fault tree. A static gate indicates that the
for the top event to occur; if one of the failures in the cut set does not order of the inputs of a gate do not matter, therefore are not sequence-
occur, then the top event will not occur. To determine the minimal cut dependent as in dynamic gates. On the other hand, in dynamic gates,
sets of a fault tree, the tree is first translated to its equivalent Boolean the order of the occurrence of input events is vital for determining the
equations. These equations can be used to determine the associated output. If dynamic gates are used, then the fault tree becomes a dynamic
minimal cut sets. The minimal cut set expression for the top event can be fault tree. The most common static gates include the AND, OR and Voting
written in the general form according to (NASA, 2002): gates while dynamic gates include the Sequence Enforcing-gate, Priority
AND-gate, Spare-gate and Functional Dependency-gate amongst other
T ¼ M1 þ M2 þ ::: þ Mk (Relex, 2009).

where T is the top event and Mi are the minimal cut sets, each of them
consisting of a combination of specific component failures. The general n- 3.2. ANN architecture
component minimal cut can be expressed as:
Once the critical components and subsystems of the top gate have
Mi ¼ X1  X2  :::  Xn been identified from the FTA and the FMEA, the neural network is con-
structed for time series analysis and prediction using as input physical
where X1, X2, …, Xn are basic component failures.
parameters of the identified systems. The methodology followed for the

Fig. 4. Combination of FTA-FMEA tools.

408
I. Lazakis et al. Ocean Engineering 152 (2018) 404–415

Fig. 5. Artificial neural network methodology.

time series artificial neural network is show in Fig. 5. P

ðxi xmean Þðdi dmean Þ
The data consisting of physical parameter measurements is pre-
processed in the ANN by mapping data to a matrix row with minimum R ¼ P N
P 0:5
ðdi dmean Þ2 ðxi xmean Þ2
and maximum values from 1 to 1 in order to conduct proper analysis N N
and improve efficiency of the network training. Moreover, the neural
network architecture has to be established to design a network capable of where P ¼ number of output processing elements; N ¼ number of ex-
analysing and forecasting time series data. A time series is a sequential set emplars in the data set; yij ¼ network output for exemplars i at processing
of data points, measured typically over successive times. It is mathe- element j; and dij ¼ desired output for exemplars i at processing element j.
matically defined as a set of vectors y(t), t ¼ 0,1, 2, …, d where t represents The ANN is then trained based on the above modelling parameters.
the time elapsed (Hipel and McLeod, 1994). The variable y(t) is treated as The trained ANN can then be converted to close-loop mode and the data
a random variable. The measurements taken during an event in a time is reformatted to simulate the network's closed loop response in order to
series are arranged in proper chronological order. The future values of a carry out multi-step ahead predictions. The trained ANN is converted to
time series y(t) are predicted only from the past values of that series. This close-loop by replacing the feedback input and creating a feedback
form of prediction is called nonlinear autoregressive and can be writ- connection from the network output to the network input, thus making
ten as: the network a Recurrent Neural Network (RNN). The first two time steps
of the input are used as input delay states in order to model the dynamic
yðtÞ ¼ f ðyðt  1Þ; :::; yðt  dÞÞ
system. RNNs can store sequential information in the form of historical
where y(t) is the observation at time t and d is the dimension of the input data and can be used in forecasting. In an RNN, the input nodes of the
vector or number of past observations used to predict the future; and f is a network are used as the value of the current condition Xt and values of the
non-linear function. The time series data is prepared by shifting time by previous time-series condition (Xt-1, Xt-2, Xt-3, …, Xt-d and Xn). The value of
the minimum amount to fill input states and layer states. This allows the the output Xtþ1 can provide a one-step-ahead prediction of a time-series
original time series data to remain unchanged, easily customizing it for condition, which is a function of the current value Xt and time-lagged
networks with different numbers of delays. Finally, data is divided into values of the previous condition (Xt-1, Xt-2, Xt-3, …, Xt-d and Xn). The
two subsets in the network for training and testing. The training set is predicted value Xtþ1 of a time series, one-step ahead in the future is given
used for computing the gradient and updating the network weights and by the following equation:
biases and the test data is used to measure the network generalisation
Xtþ1 ¼ FðXt ; Xt1 ; Xt2 ; …; Xtl ; …; Xn Þ
capabilities.
Successful ANN modelling is based upon the number of neurons, where, l is the time lag, Xtþ1 is the predicted value, Xt is the current value
number of hidden layers, values of the weights and biases, type of the or condition and Xt-d is the values of previous condition lagged by time d.
activation function, structure of the network, training styles and algo-
rithms as well as data structure (Raza and Liyanage, 2009). However, the
4. Case study & results
best structure is the one which can predict behaviour of the system as
accurately as possible. A crucial step in the building of a neural network
The methodology presented in the previous sections is applied for a
model is the determination of the number of processing elements and
Panamax container ship case study. The FMEA and FTA are performed for
hidden layers in the network. Hidden nodes are used to capture the
the main engine of the ship, a MAN B&W 8K90MC-C type engine. The
nonlinear structures in a time-series. Since no theoretical basis exists to
physical parameters (e.g. temperature, pressure) of the most critical
guide the selection, in practice the number of hidden nodes is often
components identified from the Fault Tree are then used as input in ANNs
chosen through experimentation. A nonlinear autoregressive neural
in order to perform time series analysis and predict their upcoming future
network is trained in order to proceed with the prediction results. A
values. The data used in the neural networks was collected during an on
hyperbolic tangent transfer function in the hidden layer and linear
board measurement campaign while the vessel was operating in the
transfer function in the output layer are employed, capable of approxi-
Mediterranean region. More information regarding the measurement
mating any function with a finite number of discontinuities.
campaign can be found in Raptodimos et al. (2016). Part of the data
The neural network is a feed-forward back propagation network.
collected is also used as seen in Section 4.3 for validating the neural
During training, the network weights and biases are updated after all of
network results.
the inputs and target values have been presented to the network. The
network is autoregressive as the only inputs are lagged target values. The
4.1. Fault tree & FMEA case study results
neural network is trained using the Bayesian regularization back-
propagation algorithm (Kayri, 2016). The term backpropagation refers to
Fig. 6 defines the boundaries of the main engine FTA in order to
the process by which derivatives of network error, with respect to the
define what is included in the analysis carried out.
network weights and biases, can be computed. The performance of the
As observed in Fig. 6, the main engine system is divided into six sub-
network is evaluated using the Mean Square Error (MSE) average sum of
systems. These include the cooling, lubrication oil, fuel, air, cylinder
square errors and Correlation Coefficient (R) given by the following
block assembly and engine block and components subsystem. Regarding
equations respectively (Oladokin et al., 2006):
the cooling system, it is divided into the jacket water cooling and central
PP PN  2 cooling system. The jacket water cooling system consists of the jacket
j¼0 i¼0 dij  yij fresh water cooling pump and jacket water cooling and the central
MSE ¼
NP cooling system of the sea chest strainer, sea water pipes and central

409
I. Lazakis et al. Ocean Engineering 152 (2018) 404–415

Fig. 6. Boundary conditions of main engine.

cooler. The lubrication oil system includes the lube oil filter, pump, assumption is made on the basis that a failure in the lube oil filter will
valves and lube oil cooler. Furthermore the fuel system resembles the have a knock back effect on the lube oil pump, valves then cooler. Thus
lube oil system with the addition of the fuel injectors. The air system is the presented configuration allows measuring the impact of a component
further separated into the main air system and scavenging air system. In failure on another system component, the so-called domino effect. The
the cylinder block assembly system, the system has been separated into fuel system has been modelled using a Voting gate of 3 out of 5 systems
the cylinder system which includes the cylinder head and liner and the assuming that any three of the five components of the fuel system have to
piston assembly including the piston crown, rings, stuffing box and fail in order for the fuel system to fail. The air system has been modelled
connecting rod. Finally, the engine block and components group contains as an AND gate assuming that both the main air system and scavenge air
components of the main engine such as the crankshaft, crankcase, system have to fail in order for the air system to fail. The cylinder block
camshaft and various bearings. In total, 39 basic events were modelled in assembly gate has been modelled with an AND gate consisting of the
the Fault Tree representing the components of the various main engine cylinder system containing the cylinder head and liner as basic events
sub-systems as demonstrated in Table 1. and the piston system modelled with a Voting gate of 2 out of 4 com-
A four level Fault Tree for the main engine is constructed including 38 ponents as seen in Fig. 7. Finally, the engine block and components gate
basic events and 14 gates as described in Table 1. Qualitative analysis is is composed of a Voting gate assuming that any three of the bearings,
performed in order to obtain the minimal cut sets, which provide insight crankshaft, crankcase, camshaft or exhaust valves have to occur as seen
into weak points of complex systems. Fig. 7, displays the overall Fault in Fig. 8.
Tree for the main engine. The Fault Tree is modelled with time- As previously discussed, cut sets are employed in order to obtain the
dependent dynamic gates in order to represent the interrelation of the critical components and subsystems of the top gate of the Fault Tree.
main engine system and components in an accurate and comprehensive Table 2 displays the top five minimal cut sets obtained from the main
manner. Dynamic logic gate is applied for improving veracity of engine FTA. All the minimal cut sets are of third order. As observed, the
fault tree. first three cut sets involve systems/components that were modelled in the
The cooling system is modelled using an AND gate assuming both the cylinder block assembly gate related to the cylinder and piston. The
jacket water cooling and central cooling system have to fail in order for fourth cut set is related to systems in the engine block and components
the cooling system to fail. The lubrication oil system has been modelled subsystem while the fifth is related to components of the fuel system of
using a Priority-AND gate, assuming that the lube oil filter failure has to the main engine.
occur before the pump, valves and then lube oil cooler occur. This

4.2. Physical parameters of main engine systems

Table 1
Basic events used in Fault Tree. For each system and associated component, physical parameters such
Lube Oil Filter Air Cooler, Fuel Pumps Piston Connecting as temperature and pressure are measured in order to monitor their
Piping Rod condition. Due to availability of data during the on board measurement
Main Lube Oil Pump Air Cooler Fuel Valves Camshaft Bearing campaign and time constraints, these parameters were selected as a first
Lube System Valves Scavenge Fuel Injector Thrust Bearing step for the analysis of the novel methodology approach mentioned in
Air Port this paper. Moreover, key parameters in performance observations of
Lube Oil Cooler Scavenge Main Air Main Bearings
Air Receiver Compressor
marine engines include amongst others: engine speed, barometric pres-
JFW Cooling Pump Scavenge Air Distributor Crankshaft sure, compression pressure, fuel pump index, exhaust gas temperatures
Air Manifold and pressures, scavenge air temperature and pressure, air and cooling
Jacket Water Cooling Air Receiver Air Starting Crankcase water temperatures prior and after scavenge air cooler.
Valves
The first three minimal cut sets of the FTA related to the cylinders and
Sea Chest Strainer Cylinder Head Air Filter Camshaft
Sea Water Pipes Cylinder Liner Auxiliary Exhaust Valves pistons of the main engine and by combining information retained from
Blower the FMEA worksheet, the physical parameters related to these systems
Central Cooler Piston Crown Piston Ring Piston Rod and components can be identified. Table 3 displays a fraction of the
Stuffing Box FMEA for the marine cylinder block assembly system. The FMEA for the
Fuel Piping System Fuel Oil Filter
case study was set up by using the components of the Fault Tree as input

410
I. Lazakis et al. Ocean Engineering 152 (2018) 404–415

Fig. 7. Main engine fault tree diagram.

Fig. 8. Engine block & components subsystem fault tree diagram.

in the Failed Item column of the FMEA spreadsheet. Afterwards, for each The FMEA indicated that for a possible failure event of the cylinder
failed item, a list of possible failure events and failure causes is analysed head overheating or being cracked, failure causes could be faulty exhaust
alongside their local effect on the system and their global effect of the valves and also the effects would be high temperature alarms, compres-
Fault Tree top gate which is the main engine. Information collected for sion loss and cylinder damage. In the case of the cylinder liner wear or
the formation of the FMEA is based on Cicek and Celik (2013), Turan piston ring scuffing, insufficient lubrication could be a possible failure
et al. (2011), Mokashi et al. (2002), Emovon (2016), INCASS EU project cause. Moreover, as observed from the possible failure events and causes
deliverables (INCASS, 2014b; INCASS, 2014a) and experts opinion. In- from the FMEA for the identified systems from the FTA; and by exam-
formation was collected from three Class societies, two ship operators ining the local and global effects, the possible physical parameters that
and two consultancy companies. Specifically, experts included IACS can be monitored and examined are the main engine lubrication inlet
Classification Societies surveyors and researchers with many years of pressure and temperature, piston cooling oil pressure and cylinder
onsite experience. Moreover, experts from ship operators and marine exhaust gas temperatures amongst others. For the purpose of this paper,
consultancy companies included technical managers, superintended en- the exhaust gas temperatures for the main engine cylinders are used as
gineers, and Chief engineers sailing onboard the mentioned vessel. input for the ANN as presented in the following section due to availability
of data collected in the on board measurements campaign.
Table 2
Main Engine top five minimal cut sets.

1 Cylinder Head Piston Crown Piston Ring 4.3. Artificial neural network case study results
2 Cylinder Head Piston Crown Piston Rod Stuffing Box
3 Cylinder Liner Piston Crown Piston Ring The neural network uses a univariate time series data set. The data
4 Crankcase Crankshaft Camshaft collected on board the container ship case study represents 30 continuous
5 Fuel Oil Filter Fuel Pumps Fuel Injector
per hour recordings of the exhaust gas temperatures for every one of the

411
I. Lazakis et al. Ocean Engineering 152 (2018) 404–415

Table 3
FMEA sample for FTA critical components.

System Failed Item Failure Event Failure Cause Local Effect Global Effect

Cylinder Cylinder Head Cracked Overheating, fatigue Compession loss, cylinder damage, engine misfire Possible engine stop
Cylinder Head Overheating Cracks, faulty exhaust valves High temperature alarm, smoke, cylinder damage Possible engine stop, engine damage
Cylinder Liner Wear Fatigue, lubrication oil quality Compession loss, increased lubrication consumption Engine performance reduction
Piston Piston Ring Scuffing Insufficient lubrication Scuffing mark on liner surface, oil smoke from exhaust Engine performance reduction
Piston Rod Malfunction Faulty oil scraper rings Combustion gas in crankcase Engine stop, possibility of explosion
Stuffing Box

eight cylinder engine. The ANN constructed consists of one hidden layer of the specific state and condition of each individual cylinder. Specif-
with 8 hidden nodes as illustrated in Fig. 9. A set of eight parallel neural ically, an increase of around 10  C for the exhaust gas temperatures for all
networks with the above configuration were created in order to forecast cylinders after 17–18 h of operation are observed and are mainly caused
the next 10 hourly values of each cylinder exhaust gas temperature by by an increase in the engine and fuel load. This is due to the engine
using in each network the 30 hourly temperature readings corresponding governor regulating the engine speed, as the vessel was also sailing at a
to each individual cylinder. constant speed of 10 knots during transient operation in the Mediterra-
The set of neural networks are trained and the network performance nean Sea. Moreover, localised weather conditions such as current, waves
for cylinder no.8 are demonstrated by presenting the regression plots of and wind direction could all be possible factors.
the correlation coefficient R for the training and testing data set as shown Moreover, the data used represents cylinder exhaust gas temperatures
in Fig. 10 and the error autocorrelation as shown in Fig. 11. For the other while the vessel was in transient operation. Since no failures occurred
set of networks for the other 7 cylinders, equivalent results during the on board measurement and no significant fluctuations were
were obtained. observed in the vessel's speed and engine's rpm, the obtained data does
The regression plot showing the correlation coefficient R in Fig. 10, is not cover the whole operational range of the system and further research
a good measure of how well the network has fitted the data. The is required to investigate the capability of the developed neural network
regression plot shows the actual network outputs plotted in terms of the to forecast parameters during various engine loads and speeds. However,
associated target values. Regression values measure the correlation be- the neural network developed was capable of accurate time series pre-
tween outputs and targets. A correlation coefficient R value of 1 implies a dictions by using appropriate training algorithms and network structure.
perfect fit of outputs exactly equal to targets. Fig. 10 displays the network Besides the network performance, the first 5 hourly forecasted values
outputs with respect to targets for training and test sets. 70% of the data from the network are compared in Table 4 with the actual values
was used for training purposes and the remaining was used for the test recorded on board the vessel for validation purposes.
data set. Bayesian regularization does not use a validation set but in- As observed from Table 4, the error between the actual and predicted
cludes this in the training set. The training data indicate a good fit as does values is under 1%. As such, the performance and accuracy of the trained
the test data results showing values of R equal to 90% in both sets. neural networks are verified, indicating its satisfying predictive time
Besides the correlation coefficient, the error autocorrelation is used to series capabilities.
validate the network performance. The error autocorrelation function
defines how the forecast errors are interrelated in time. For a faultless 5. Discussion & conclusions
prediction model, there should be one non-zero value that should occur
at zero lag implying that the forecast errors are entirely uncorrelated with The combination of the FTA with the FMEA tool provided a good
each other. Thus, if the network has been trained well then besides the insight for identifying subsystems and their components affecting the
centre line which shows the mean squared error, all other lines should main engine through qualitative analysis using minimal cut sets. A
fall within the confidence limits as successfully shown in Fig. 11. minimal cut set is the smallest set of basic events which result in the
The results obtained from the set of neural networks for predicting the occurrence of the top event. FMEA also provided insight into possible
future upcoming ten hourly values in time of the cylinder exhaust tem- failure events and causes and contributed in identifying which physical
peratures are illustrated in Fig. 12. parameters could be further examined in order to investigate the per-
Fig. 12 illustrates the graph of the recorded values and predicted ones formance and operation of the systems of interest.
against time for the exhaust gas temperatures of every cylinder of the In the case in which physical parameters are not applicable to a
ship main engine. The first 30 hourly measurements are the recorded specific system by examining both the FTA and FMEA, such as the 5th
values from the on board measurement campaign, while the last 10 minimal cut sets in Table 2 referring to the fuel oil filter, then counter-
hourly data points, from 30 h until 40 h are the ones forecasted from each measures and maintenance actions have to be implemented in order to
neural network. successfully maintain this component or system. Overall, the combina-
The recorded temperatures for all cylinders are within the range of tion of these tools can contribute to an initial reliability assessment of a
220 –270  C. Variations in the cylinder temperatures could be the result complex system. The FTA can assist in identifying critical items that
affect the overall performance and reliability of a system. In terms of the
FTA, if historical data is available such as failure rates or mean time
between failures, then quantitative analysis can also be performed in
order to examine the criticality of various event contributing to the
overall failure or occurrence of the top undesired event.
In an era in which big data and the internet of things is becoming a
reality and the shipping industry is endeavouring to advance, there is no
clear definition of big data and it is still challenging to quantify the
volume of data required for successful machine learning, data training
and analysis. Developments in sensor technology progressively permit
shipping companies and operators to engage in monitoring activities in a
more reliable and cost-effective manner, with a higher Return on In-
vestment (ROI). This will permit future applications of data-driven
Fig. 9. Closed-loop dynamic neural network. models in the maritime industry more feasible and possible. However,

412
I. Lazakis et al. Ocean Engineering 152 (2018) 404–415

Fig. 10. Sample regression results for training and test data of neural network (Cylinder no.8).

privacy issues, data protection and privacy-related risks pose challenges functional relationships among the data even if the underlying re-
for the development of big data methods and tools. Moreover, for lationships are hard to describe or unknown. They do not rely on priori
effective analysis of collected data, availability of skilled workers and principles or statistics models and can significantly simplify the model
software tools is necessary in addition to the required resources for synthesized process. Moreover, the ANN methodology described was
investigating and developing this type of tools. Non-standardised applied for all 8 cylinders of the main engine by creating a set of parallel
collection of machinery data also limits the infrastructure and general networks. Because neural networks are a data-based method, they are
applicability of such methods in the maritime industry, as incompatible universally applicable to systems from different industrial application
data standards may cause issues. Therefore, suitable uncertainty quan- fields. Modelling an effective and accurate neural network structure is
tification is crucial in order to derive correct data analysis results and based upon the number of neurons, number of hidden layers, values of
avoid false-positive results. the weights and biases, type of the activation function, structure of the
Neural networks learn from past examples and capture subtle network, training algorithms as well as data structure. Usually, the ideal

Fig. 11. Sample of autocorrelation of error (cylinder no.8). Fig. 12. Cylinders no.1–8 exhaust gas temperature predictions.

413
I. Lazakis et al. Ocean Engineering 152 (2018) 404–415

Table 4 authors would like to express their sincere gratitude to Dr Victoria Cat-
Comparison of on board values with ANN predictions for the Exhaust Gas Temperatures for terson (University of Strathclyde) for her valuable input.
all cylinders.

Cylinder 1 On Board 262.2 262.4 262.4 262.4 262.4 References

ANN Prediction 263.0 260.0 262.0 262.0 263.0
Error 0.31% 0.92% 0.16% 0.16% 0.23%
ABS, 2015a. Guidance notes on failure mode and effects analysis (FMEA) for
Cylinder 2 On Board 232.1 231.9 231.8 231.8 231.7 classification. In: ABS. TX USA, Houston.
ANN Prediction 232.0 232.0 232.0 233.0 232.0 ABS, 2015b. NS5 Enterprise-Asset Management Solutions. TX USA, Houston.
Error 0.02% 0.03% 0.07% 0.53% 0.13% Aizenberg, I., Sheremetov, L., Villa-Vargas, L., Martinez-Mu~ noz, J., 2016. Multilayer
Neural Network with Multi-Valued Neurons in time series forecasting of oil
Cylinder 3 On Board 264.4 264.7 264.5 264.5 264.5 production. Neurocomputing 175 (Part B), 980–989.
ANN Prediction 265.0 264.0 264.0 265.0 262.0 Anantharaman, M., Khan, F., Garaniya, V., Lewarn, B., 2014. A step by step approach for
Error 0.22% 0.27% 0.21% 0.17% 0.96% evaluating the reliability of the main engine lube oil system for a Ship's propulsion
system. TransNav Int. J. Mar. Navig. Saf. Sea Transp. 8, 367–371.
Cylinder 4 On Board 245.6 245.3 245.1 244.9 244.8 Arunraj, N.S., Maiti, J., 2007. Risk-based maintenance-Techniques and applications.
ANN Prediction 246.0 245.0 246.0 246.0 247.0 J. Hazard. Mater. 142, 653–661.
Error 0.16% 0.13% 0.36% 0.44% 0.90% Ben-Daya, M., Knezevic, J., 2009. Handbook of Maintenance Management and
Engineering. Springer, New York, USA.
Cylinder 5 On Board 238.3 238.0 237.8 237.7 237.7
BS, 2012. Condition Monitoring and Diagnostics of Machines. BS ISO 13372:2012. BSI,
ANN Prediction 237.0 237.0 235.0 236.0 238.0 London, UK.
Error 0.56% 0.40% 1.19% 0.74% 0.12% Cicek, K., Celik, M., 2013. Application of failure modes and effects analysis to main
Cylinder 6 On Board 251.2 250.7 250.5 250.4 250.4 engine crankcase explosion failure on-board ship. Saf. Sci. 51, 6–10.
ANN Prediction 250.0 249.0 251.0 250.0 251.0 CLASSNK, 2013. PrimeShip-intergrated System for Ship Performance Capability. Japan,
Tokyo.
Error 0.47% 0.69% 0.19% 0.17% 0.24%
Dekker, R., 1996. Applications of maintenance optimization models: a review and
Cylinder 7 On Board 264.8 263.5 263.4 263.9 264.0 analysis. Reliab. Eng. Syst. Saf. 51, 229–240.
ANN Prediction 259.0 261.0 261.0 261.0 262.0 Emovon, I., 2016. Multi-criteria Decision Making Support Tools for Maintenance of
Error 2.25% 0.97% 0.94% 1.10% 0.77% Marine Machinery Systems. Doctor of Philosophy. Newcastle University.
Gao, S., Kang, J., 2016. Group maintenance strategy for FPSO offloading system based on
Cylinder 8 On Board 255.7 253.9 254.7 256.1 256.5 reliability analysis. In: 35th International Conference on Ocean. Offshore and Arctic
ANN Prediction 256.0 255.0 255.0 254.0 256.0 Engineering, Busan, Korea.
Error 0.12% 0.45% 0.12% 0.83% 0.20% Guan, Y., Zhao, J., Shi, T., Zhu, P., 2016. Fault tree analysis of fire and explosion accidents
for dual fuel (diesel/natural gas) ship engine rooms. J. Mar. Sci. Appl. 15, 331–335.
Guoping, L., Qingwei, Z., Xiao, M., 2013. Combination of fault tree and neural networks
in excavator diagnosis. Indonesian J. Electr. Eng. Comput. Sci. 11, 1787–1796.
Haykin, S., 1998. Neural Networks: a Comprehensive Foundation. Prentice Hall PTR, NJ,
network topology is found through testing different network configura-
USA.
tions and choosing the one whose performance criteria are greater. Hidalgo, E.M.P., Silva, D.W.R., Souza, G., 2011. FMEA and FTA analysis applied to the
Another closely related issue in ANN model building is related to steering system of LNG carriers for the selection of maintenance policies. In:
discovering the best way to split up the data, and how large the training Proceeding of the 21st Brazilian Congress of Mechanical Engineering. Natal, Brazil.
Hipel, K.W., McLeod, A.I., 1994. Time Series Modelling of Water Resources and
and/or test data sets should be. Environmental Systems. Elsevier, Amsterdam, The Netherlands.
In conclusion, the FMEA and FTA tools can complement each other, IACS, 2014. Recommendation for the FMEA Process for Diesel Engine Control Systems.
providing a good general model for obtaining critical systems of interest IACS Rec. 2014. London, UK.
IMO, 1993. International Safety Management (ISM) Code. Resolution A741(18). London,
alongside their possible failure causes and effects and relevant physical UK.
parameters, which if monitored can provide useful information regarding INCASS, 2014a. Deliverable D4.1 Specification of Requirements for Machinery and
the performance of the systems and components of interest. The use of Equipment. INCASS-Inspection Capabilities for Enhanced Ship Safety, UK. EC FP7
Project.
the ANN offers a more specific direction in terms of examining and INCASS, 2014b. Deliverable D4.2 Stakeholders' Data Requirements. INCASS-Inspection
monitoring the condition of the identified Fault Tree components though Capabilities for Enhanced Ship Safety, UK. EC FP7 Project.
the time series analysis of their physical parameters. The set of parallel Isermann, R., 2006. Fault-diagnosis Systems: an Introduction from Fault Detection to
Fault Tolerance. Springer Science & Business Media, Berlin, Germany.
networks contribute to the development of an overall prediction model Kayri, M., 2016. Predictive abilities of bayesian regularization and levenberg–marquardt
which will be utilised within a maintenance and condition monitoring algorithms in artificial neural networks: a comparative empirical study on social data.
framework in order to assess the performance of ship machinery based on Math. Comput. Appl. 21, 20.
Knutsen, K.E., Manno, G., Vartdal, B.J., 2014. Beyond Condition Monitoring in the
past observations and current information and can be used for prognostic
Maritime Industry. DNV-GL, Norway.
and diagnostic purposes and applications. In order to ensure that the Laboissiere, L.A., Fernandes, R.A.S., Lage, G.G., 2015. Maximum and minimum stock
prediction model is trustworthy and reliable, uncertainty and sensitivity price forecasting of Brazilian power distribution companies based on artificial neural
analysis methods should also be investigated. Future steps in this direc- networks. Appl. Soft Comput. 35, 66–74.
Laskowski, R., 2015. Fault Tree Analysis as a tool for modelling the marine main engine
tion include the examination of more machinery systems by also reliability structure. Sci. J. Marit. Univ. Szczecin 71–77.
including alongside their physical parameters various engine loads. €
Lazakis, I., Olçer, A., 2015. Selection of the best maintenance approach in the maritime
Furthermore, the development of alarm levels and thresholds should be industry under fuzzy multiple attributive group decision-making environment. Proc.
Inst. Mech. Eng. Part M J. Eng. Marit. Environ. 1–13.
developed for the prediction model. The alarms and thresholds can be Lazakis, I., Turan, O., Aksu, S., 2010. Increasing ship operational reliability through the
based on either user defined limits, original equipment manufacturer implementation of a holistic maintenance management strategy. Ships Offshore
recommendations or parameter deviations from an established baseline. Struct. 5, 337–357.
Liu, H., Tian, H.-Q., Liang, X.-F., Li, Y.-F., 2015. Wind speed forecasting approach using
Finally, training and testing of the neural network with big data sets will secondary decomposition algorithm and Elman neural networks. Appl. Energy 157,
assist in investigating the accuracy of the network in its predictive ca- 183–194.
pabilities and selecting the optimum neural network algorithms and Lloyds Register, 2013. Machinery Planned Maintenance & Condition Monitoring. London,
UK.
configurations. Manzini, R., Regattieri, A., Pham, H., Ferrari, E., 2009. Maintenance for Industrial
Systems. Springer Science & Business Media, London, UK.
Acknowledgments Mokashi, A.J., Wang, J., Vermar, A.K., 2002. A study of reliability-centred maintenance in
maritime operations. Mar. Policy 26, 325–335.
NASA, 2002. In: Stamatelatos, M., Vesely, W. (Eds.), Fault Tree Handbook with Aerospace
The work in this paper is partially funded by INCASS project. INCASS Applications (Washington, DC).
has received research funding from the European Union's Seventh Nasr, M.S., Moustafa, M.A.E., Seif, H.A.E., El Kobrosy, G., 2012. Application of Artificial
Framework Programme under grant agreement No. 605200. This publi- Neural Network (ANN) for the prediction of EL-AGAMY wastewater treatment plant
performance-Egypt. Alexandria Eng. J. 51, 37–43.
cation reflects only the authors' views and European Union is not liable Noor, M., MH, M.Y., MM, N., 2016. Prediction of marine diesel engine performance by
for any use that may be made of the information contained herein. The using artificial neural network model. J. Mech. Eng. Sci. (JMES) 10, 1917–1930.

414
I. Lazakis et al. Ocean Engineering 152 (2018) 404–415

NUREG-0492, 1981. Fault Tree Handbook U.S. Nuclear Regulatory Commission. Relex, 2009. Relex Reliability Studio Reference Manual. Relex Software Corporation,
Washington, DC. USA.
Oladokin, V.O., Charles-Owaba, O.E., Nwaouzru, C.S., 2006. An application of artificial Ruilin, Z., Lowndes, I.S., 2010. The application of a coupled artificial neural network and
neural network to maintenance management. J. Ind. Eng. Int. 2, 19–26. fault tree analysis model to predict coal and gas outbursts. Int. J. Coal Geol. 84,
Pascual, D.G., 2015. Artificial Intelligence Tools: Decision Support Systems in Condition 141–152.
Monitoring and Diagnosis. Crc Press, USA. 
Souza, R., Alvares, A.J., 2008. FMEA and FTA analysis for application of the reliability-
Peng, Y., Dong, M., Zuo, M.J., 2010. Current status of machine prognostics in condition- centered maintenance methodology: case study on hydraulic turbines. In: AVCM
based maintenance: a review. Int. J. Adv. Manuf. Technol. 50, 297–313. Symposium Series in Mechatronics, pp. 803–812.
Pintelon, L., Parodi-Herz, A., 2008. Maintenance: an Evolutionary Perspective. Complex Stopford, M., 2009. Maritime economics 3e. Routledge, New York, USA.
System Maintenance Handbook. Springer, London, UK. Szoplik, J., 2015. Forecasting of natural gas consumption with artificial neural networks.
Raptodimos, Y., Lazakis, I., 2016. An artificial neural network approach for predicting the Energy 85, 208–220.
performance of ship machinery equipment. In: Maritime Safety and Operations 2016 Tinsley, D., 2016. Dawning of New Era in Asset Maintenance. Marine Power & Propulsion
Conference Proceedings. Supplement 2016. The Royal Institution of Naval Architects, London.
Raptodimos, Y., Lazakis, I., Theotokatos, G., Varelas, T., Drikos, L., 2016. Ship sensors Turan, O., Lazakis, I., Judah, S., Incecik, A., 2011. Investigating the reliability and
data collection and analysis for condition monitoring of ship structures and criticality of the maintenance characteristics of a diving support vessel. Qual. Reliab.
machinery systems, 1/27 2016. In: Smart Ships 2016 Conference. Royal Institution of Eng. Int. 27, 931–946.
Naval Architects, London. Verma, A.K., Srividya, A., Karanki, D.R., 2010. Reliability and Safety Engineering.
Rausand, M., Arnljot, H., 2004. System Reliability Theory: Models, Statistical Methods, Springer, London, UK.
and Applications. John Wiley & Sons. Zhang, G., Eddy Patuwo, B., Hu, M.Y., 1998. Forecasting with artificial neural networks:
Raza, J., Liyanage, J.P., 2009. Application of intelligent technique to identify hidden the state of the art. Int. J. Forecast. 14, 35–62.
abnormalities in a system: a case study from oil export pumps from an offshore oil Zhang, G.P., Patuwo, B.E., Hu, M.Y., 2001. A simulation study of artificial neural
production facility. J. Qual. Maintenance Eng. 15, 221–235. networks for nonlinear time-series forecasting. Comput. Oper. Res. 28, 381–396.

415