Intelligent Information Processing in the Problem of Propulsion

Intelligent Information Processing in the Problem of Propulsion System

System Condition Monitoring
Condition Monitoring*
＊

Oleksiy Bondarenko ＊＊
Tetsugo Fukuda ＊＊
Ken Miyachi＊＊＊
By Oleksiy Bondarenko**, Tetsugo Fukuda**, Ken Miyachi***

During the past years, in response to the gradually stringent regulations, the undertaken steps in the
shipping industry spurred the technological improvement in ship and engine designs. Thus, new engines
accept a wide variety of fuels, and electronic systems provide flexible control and online tuning. At the
same time, simultaneous assessment of performance and condition monitoring is becoming increasingly
important. In this respect, digitalization and the accompanying evolution of smart sensors and data
acquisition systems give the possibility of applying complex analytics and machine learning algorithms to
the compelling need of engine performance monitoring and failure identification. The present paper
proposes a solution to the problem of condition monitoring using a method of statistical analysis of
multidimensional information acquired from the sensors. The factor analysis method is used to derive a
performance index showing engine state deviation from the normal condition. At the same time, principal
factor loadings are used as features characterizing the contribution of every measured and analyzed
parameter to the variance of the performance index. The benchmark of the developed method is
illustrated using the simulation model of a diesel engine with incorporated models of failure states. The
latter was developed and validated from the data measured on the test engine.

1. Introduction the workload of monitoring and troubleshooting

With hardened competition and stringent tasks [1]. Thus, the condition monitoring and
regulations in the shipping industry, stronger unsupervised processing of the propulsion system
demand for greater efficiency, reduced emissions, operating parameters, creating value for the
and increased profitability becomes evident. In operators, is of great importance.
response to this, the ship and propulsion system While the condition monitoring and
designs are characterized by continuously rising information processing algorithms have been widely
complexity. As a result, engine manufacturers used in such sectors as aviation and nuclear power
introduce new engine designs that are compatible plants, there are not many applications in the
with a wide variety of fuels and equipped with marine field. Nevertheless, several applications
intelligent electronic systems providing flexible have been made in the past few years. For example,
control and online tuning. Simultaneously, the WinGD has proposed a WiDE (WinGD Integral
evolution of intelligent sensors and data acquisition Digital Expert) solution [2] that consists of data
systems and advanced telecommunications fostered acquisition, engine status visualization and
the equipment manufacturers and service providers diagnosis systems. The ClassNK introduced a set of
in the shipping industry to constantly implement consulting services called CMAXS for the
digitalization and computerization to encapsulate maintenance and management of onboard
and transfer an analog world to a digitalized world. equipment. Among them, e-GICSX is aimed at
Digitalization is one of the technological automatic diagnostic of propulsion system
improvements, which is also referred to as digital abnormal state [3].
shipping, is expected to increase ship efficient In condition monitoring, the development of a
utilization. On the other hand, digitalization suitable model of the target system is a primary
contributes to reducing the number of crew task allowing for data classification with respect to
members while bringing a significant increase in the reference (nominal) conditions. In recent times,
machine-learning (ML) methods are widely
* Received March 31, 2021 implemented for the purpose of monitoring and
** Member, National Maritime Research Institute
*** Mitsui E&S Machinery
Journal of the JIME Vol.00, No.00(2005) －1－ 日本マリンエンジニアリング学会誌 第 00 巻 第 00 号 (2005)

Journal of the JIME Vol. 56, No. 6（2021）

（2021） ―1― 日本マリンエンジニアリング学会誌 第56巻 第 6号
号（2021）
（2021）
2 日本マリンエンジニアリング学会執筆要項
Intelligent Information Processing in the Problem of Propulsion System Condition Monitoring

diagnostic tasks. The ML model learns from the testbed engine. Thus, the propulsion system is
data provided to identify normal or abnormal referred to as the testbed engine loaded by a water
engine behavior and then characterizes the current brake.
state. Thus, in the literature [4], the performance For the dynamic engine model formulation, a
and potentialities of the regularized least square hybrid cycle-mean value (H-CMV) modeling
(RLS) and support vector machine (SVM) models approach is used [6, 7]. The H-CMV approach
were benchmarked by training them on data to consists in combining the continuous cycle-averaged
forecast the performance decay of propulsion model of air and exhaust gases flows through the
system on a naval vessel. Likewise, in the literature engine with the discrete consideration of the
[5], a robust framework for analyzing performance combustion cycle in the cylinder. Figure 1 illustrates
data and condition monitoring was constructed. the outline of the engine model. The model is
First, a one-class SVM classifier was trained to structured in a modular arrangement. Every entity
represent a nominal state, and then a healthiness represents an essential component of the engine,
index is represented by the probability of each new such as the compressor, air cooler, air receiver,
point belonging to the same performance profile exhaust gas receiver, turbine and cylinder unit,
with training data. The data-driven approaches, together with the shaft dynamics. The steady-state
however, provided limited solutions for specific performance maps, time-depended momentum,
machinery in specific operational conditions. What energy and mass equations, and non-linear
is more, a large amount of quality labeled data is algebraic equations are used to describe the
required to generalize the model for every case component's behavior.
being examined, including data from different
operational profiles of the ship and/or failure rates.
In the meantime, however, satisfactory results
can also be achieved using the methods of statistical
analysis of multidimensional information. In this
paper, the application of the method of factor
analysis to the problem of propulsion system
condition monitoring is scrutinized closely.
Following the method, the covariance matrix of
parameters deviation from the nominal state is
decomposed to a set of factor loadings components.
The latter provides the calculation of the relative
contribution to the explained variance, which is
used as a condition state indicator. The method is
benchmarked on data obtained from the propulsion
system simulator with the incorporated degradation
models. The following chapters detail the method
and the failure simulator.

2. Propulsion System Simulator

2.1 Introduction
A generic ship propulsion system may be considered
made up of three main components: the engine, the
propeller and the hull. However, in the present Fig. 1 Composition of a generic marine diesel
study, the application of the condition monitoring engine for H-CMV modeling

method is limited to the engine systems only. What The key differential equations describing both
is more, the data required to validate the simulator dynamic and thermodynamic engine behavior are
and degradation models were obtained from the based on the laws of conservation for mass and

Journal of the JIME Vol.00, No.00(2005) －2－ 日本マリンエンジニアリング学会誌 第 00 巻 第 00 号 (2005)

Journal of the JIME Vol. 56, No. 6（2021）
（2021） ―2― 日本マリンエンジニアリング学会誌 第56巻 第 6号 号
（2021）
（2021）
 日本マリンエンジニアリング学会執筆要項
Intelligent Information Processing in the Problem of Propulsion System Condition Monitoring 3

energy: 2.2.1 Air filter clogging

dP R T A turbocharger provides the air supply to the engine

dt V
Gin  Gout  ,
cylinder for the combustion of fuel. The dynamics of
(1)
dT dM the turbocharger (TCH) with respect to rotational
Cv M u  H in  H out  ,
dt dt speed is represented in the following form:
where Gin and Gout are the inlet and outlet mass dntc 1
flow rates, respectively; while Hin and Hout are inlet   QT  QC  , (3)
dt 2 Itc
and outlet energy rate, respectively; V is the
where QT is the torque developed by the turbine due
component volume; M and u are the mass and
to the expansion of exhaust gas, QC is the torque
internal energy, respectively; while P and T are the
required by the compressor to compress the air
pressure and temperature, respectively.
mass flow Gc from the pressure after filter Paf to
The dynamics of the rotating shaft is determined
pressure Ps in scavenging receiver, ntc is the
by the momentum conservation law:
turbocharger rotational speed, Itc is the inertia of
dn 1 Q Q ,

dt 2 I sh e l  (2) the turbocharger shaft.
The isentropic work required for an air pressure
where n is the shaft rotational speed, Ish is the
increase in a compressor, appropriately modified by
inertia moment of the whole shaft line; Qe and Ql
the efficiency ηic, yields the following expression for
are the engine torque and load torque, either
compressor torque:
propeller or water-brake.
 ka 1 
Since the attention is focused on the C pa Ta Gc   Ps  ka 
degradation models of particular components,
 a QC    1 , (4)
2 ntciC   Paf  
detailed discussion on the complete engine model is
 
omitted for this paper. However, full description and
where Cpa is the specific heat constant of air, ka is
model validation for a low-speed two-stroke engine
the ratio of the specific heat for air, Paf is the
can be found in [7].
pressure after the filter.
The pressure drop, ΔPaf, through the air filter is
2.2 Components Degradation Models
proportional to the square of the air mass flow in
Two types of faults can be recognized in the
the following form:
propulsion system: the abrupt faults, which can
ξaf Gc2
only be discovered at the time of occurrence and Paf  , (5)
a
consequently are not identifiable by the condition
where ξaf is the flow coefficient, ρa is the air density.
monitoring; the incipient faults are characterized by
The air filter clogging is introduced as a
the time-dependent degradation of components
degradation function [8] of the flow coefficient in the
performance. The latter is highly suitable for
following form:
identification by condition monitoring. In order to
develop and study the characteristics of the  t 
 d 
ξaf t  k 1  e  , (6)
condition monitoring system, the model of the  
propulsion system described in the preceding
where k, τd are the decay coefficient and time
chapter has been updated to take into account the
constant, respectively.
performance degradation of several components.
2.2.2 Air cooler fouling
One cannot deny that the compressor's
In order to keep the density of air after the
performance, air cooler (AC) and turbine, which
compressor as high as possible, an air cooler is
forms the engine air loop, affects the entire
installed before the engine cylinder's input. The
propulsion system's performance. Thus, faults
temperature of the air after AC is modeled as
related to the air loop are considered in this paper.
follows:
These are clogging of compressor air filter, fouling of 1
air cooler, and fouling of the exhaust duct after the  G 3
turbine. Ts  Tcw  ac Tc  Tcw   c  , (7)
 Gc 
 mcr 

Journal of the JIME Vo00， No． 00(2005) －3－ 日本マリンエンジニアリング学会誌 第 00 巻 第 00 号 (2005)

Journal of the JIME Vol. 56, No. 6（2021）
（2021） ―3― 日本マリンエンジニアリング学会誌 第56巻 第 6号 号
（2021）
（2021）
 4 日本マリンエンジニアリング学会執筆要項
Intelligent Information Processing in the Problem of Propulsion System Condition Monitoring

where Tcw is the water temperature at the cooler from the available historical data. However, such
inlet, ηac is the cooler effectiveness, Tc is the air data is not usually available, and thus the
temperature after the compressor. particular experiment on the test engine was set up
The fouling of AC can be caused by air to identify the degradation model parameters.
impurities that, together with oil vapors, produce a
layer covering the surface on the airside. From the 2.3 Identification and Validation of Degradation
waterside, the scale can form a layer of insulation. Models
The main consequence is the degradation of AC The actual degradation behavior of the engine
effectiveness and the rise of outlet temperature. components is both time and operation
Thus, the AC fouling is introduced as a degradation profile-dependent, and thus a vast combination of
function of the cooler effectiveness in the following conditions is required. However, for the purpose of
form: the present research, not the degradation behavior
 t  but the validity of the simulation model response to
 
ac 
t  k 1  e d
. (8) the degradation of the components is of high
 
relevance.
The experiments were performed on the
2.2.3 Exhaust duct fouling two-stroke marine diesel test engine kindly
From the propulsion system efficiency point of view, provided by Mitsui E&S Machinery. Table 1
the exhaust gas energy recovery after the turbine is summarizes the specification of the test engine. The
commonly applied by installing a steam boiler. Also, engine was run at a constant speed and load
more and more ships are being equipped with corresponding to 75% of the Maximum Continuous
selective catalytic reduction (SCR) systems. Rating (MCR). The previously described
Therefore, fouling can be caused by soot in the degradation models were introduced as follows:
exhaust gases and oil vapors that produce layers exhaust duct fouling was imitated by gradually
covering the surfaces of boilers and SCR. The main closing a valve on the exhaust stack of the test
consequence is the degradation of TCH performance engine; the AC fouling was imitated by gradually
due to increased backpressure after the turbine. reducing the flow of cooling water through the cooler.
The isentropic work of gas expansion in the The air filter clogging imitation was considered the
turbine appropriately modified by the efficiency ηit most complicated due to the risk of the compressor
yield the expression for turbine torque: surging. Thus, the degradation was introduced in
 ke 1 
the reverse order. Engine load was reduced to 25%
C pe TeGeiT   Po  ke 
 QT 1   , (9) MCR, the air filter was partially closed with a cover,
2 ntc   Pe   engine load was returned to 75% MCR, and then
 
where Cpe is the specific heat constant of gas, ke is the cover was gradually sliding.
the ratio of the specific heat for gas, Ge is the The degradation functions, introduced by Eqs.
exhaust mass flow, Po is the pressure after the (6), (8) and (10), are essentially continuous in time;
turbine. nonetheless, the degradations were introduced in a
Likewise, the air filter clogging, the exhaust stepwise discrete fashion, as illustrated in Fig.2.
duct fouling is introduced as a degradation function The first step towards the validation of the
of the flow coefficient in the following form: engine simulation model with introduced failures is
the identification of degradation models introduced
ξed Ge2  t 
 d  on the test engine. Figure 3 shows the time
Po  , ξed  t   k 1  e  , (10)
e   evolution of parameter degradation measured on
where ΔPo is the pressure drop through the duct, ξed the test engine and the model identification results.
is the flow coefficient, ρe is the gas density. After successfully identifying the degradation, the
The parameters of the presented above second step compares the engine simulation model
degradation functions can be empirically derived responses with those measured on the test engine
under similar conditions. The results of engine

Journal of the JIME Vol.00, No.00(2005) －4－ 日本マリンエンジニアリング学会誌 第 00 巻 第 00 号 (2005)

Journal of the JIME Vol. 56, No. 6（2021）
（2021） ―4― 日本マリンエンジニアリング学会誌 第56巻 第 6号 号
（2021）
（2021）
 日本マリンエンジニアリング学会執筆要項
Intelligent Information Processing in the Problem of Propulsion System Condition Monitoring 5

model responses are depicted in Fig. 4. As can be

seen, although there are some biases between
measured and simulated state variables, the engine
model correctly reproduces the responses of the test
engine to all the failure models introduced.
Therefore, the engine simulation model can be used
as an equivalent substitute for the actual engine
with failures.

1.0
Change of parameter

0.8

0.6

0.4 Continuous Parameter

Degradation
0.2 Stepwise Discrete
Approximation
0.0
0 20 40 60 80 100 120 140 160
Time, min

Fig. 2 Stepwise approximation of a continuous function

Fig. 3 Degradation models identification results

Table 1. Mitsui-MAN engine specification at MCR
Engine Type 4S50ME-T9
No of Cylinders 4
Bore/Stroke, [mm/mm] 500/2214
Power, [kW] 7120
Speed, [rpm] 117
IMEP, [bar] 22

Fig. 4 Validation of the engine simulation model responses to the introduced failure events

Journal of the JIME Vo00， No． 00(2005) －5－ 日本マリンエンジニアリング学会誌 第 00 巻 第 00 号 (2005)

Journal of the JIME Vol. 56, No. 6（2021）
（2021） ―5― 日本マリンエンジニアリング学会誌 第56巻 第 6号 号
（2021）
（2021）
6 Intelligent Information Processing in the Problem of Propulsion System Condition Monitoring

3. Condition Monitoring Framework case of the weighted principal components model,

and thus principal components are used to estimate
3.1 Method of Factor Analysis
the factors.
The statistical method of factor analysis (FA) plays
The correlation matrix for a set of p measured
a central role in extracting non-trivial information
parameters during a time of length N is given as:
from a set of raw data collected from sensors of the
monitored component. FA is the method designed to Y TY
R ,
analyze interrelationships within a set of variables. N
The critical assumption of the method is that the (12)
j
measured variables can be correlated in such a way X i, j  X
Yi, j  , i 1,
, N , j 1,, p
that their correlation may be reconstructed by a few X j
(latent) common parameters, called factors. These where Y is a matrix (N x p) of scaled parameter
factors, in turn, could represent the underlying deviations from the base level, σx is the parameter
structure and inherent interdependencies in a standard deviation.
concise and interpretable form [9, 10]. In turn, the (p x p) correlation matrix R can be
The general factor model can be written as expressed as the sum of p-eigenvalues multiplied by
Y A F Ψ,
 (11) their eigenvectors and their transpose that is a
where Y = (y1, y2,…,yp)T is a vector of measured singular value decomposition (SVD) of matrix R:
variables, F = (f1, f2,…,fr)T is a vector of r < p latent
R  USV T , (13)
variables or factors, A is a (p x r) matrix of fixed
Eventually, the estimator of the factor loadings
coefficients (factor loadings) and Ψ = diag(u1,
matrix is defined as:
u2,…,up) is a diagonal matrix of random error terms.
A U S, (14)
The latter consists of errors of measurements,
The principal components analysis gives a set
together with unique individual effects associated
of axes to which variables could be referred, and
with each variable yi. Figure 5 demonstrates the
thus the factor loadings are the projections of the
relationships among variables in the factor model.
variables onto the factor axes. Although
mathematically unique, a set of factors is the only
one of an infinite number of sets that describes data
configuration just as well. Still, it does not
guarantee maximum statistical significance.
Therefore, as the second step of the analysis, the
maximum likelihood principle can be used to specify
the elements of matrix A that best model the
correlation structure of data Y.
In the beginning, the factors F and data Y are
assumed as a joint Gaussian distribution:
Fig. 5 Representation of factor relationships
Y   0  AAT  Ψ A 
3.2 Algorithm of Factor Analysis  F  �  0 ,   , (15)
The FA algorithm analyses the correlations      
 AT
I  
among the parameter deviations from the base To maximize the parameters of the joint
(nominal) level of monitored parameters. The latter distribution above, the log-likelihood function of the
can be provided from the accompanying nominal parameters yields:
state simulation model (digital twin), or some initial
set of parameters can be considered as a base level.  1 N 
  2 log Ψ  2 log 2 ... 
For the sake of brevity, only brief details of the   A,Ψ    E  , (16)
algorithm are given below, and full theoretical   1 Y  AF T Ψ 1 Y  AF 
 2    

background can be found in [9, 10].
Factor analysis can be viewed as a particular

Journal of the JIME Vol.00, No.00(2005) －6－ 日本マリンエンジニアリング学会誌 第 00 巻 第 00 号 (2005)

Journal of the JIME Vol. 56, No. 6（2021）
（2021） ―6― 日本マリンエンジニアリング学会誌 第56巻 第 6号 号
（2021）
（2021）
 日本マリンエンジニアリング学会執筆要項
Intelligent Information Processing in the Problem of Propulsion System Condition Monitoring 7

The maximization of the likelihood function with analysis applied to the data obtained from the
respect to the matrix A, is performed by applying developed simulator of the engine with fault models.
the iterative Expectation-Maximization (EM) In order to demonstrate the qualitative difference of
algorithm. For the sake of brevity, the details of the the results and due to limitations of paper length,
algorithm are omitted and can be found in [11]. only the cases of air filter clogging and AC fouling
Although, as was mentioned earlier, the factors are demonstrated and discussed.
contain the essential information related to the The engine simulation model was set to run at
variance of measured parameters, the factors a load corresponding to 75% MCR and speed 106
loadings are considered more informative for the rpm. The steady-state simulation is followed by the
condition monitoring purpose. This is because the response of engine state variables to the introduced
components of factor loadings matrix A reflect the incipient failure of air filter clogging. The time
degree of contribution of every parameter deviation constant in the degradation model is set to τd = 3600
onto the explained variance of every factor. Thus, a sec. The fault is introduced at T = 1800 sec. Figure 6
performance index of the component condition is illustrates the engine state variables time evolution.
defined as the relative share of the explained The sensors noise imitation is also added to the
variance formed by the factors loadings of the results. The parameters listed in Table 2 were
measured parameters: recorded every 1 sec.
p
Table 2. List of recorded parameters
 ai2,n
D
n
  2  F 
n  i 1
, 
n 1,, N (17) Engine Speed, [rpm] Ne
  N p
 ai2,n TCH Speed, [rps] Ntc
n 1 i 1 Scav. Air Pressure, [kPa] Ps
As one may note, the number of performance Scav. Air Temperature, [K] Ts
indices is equal to the number of processed
Exh. Gas Pressure, [kPa] Pe
parameters p, and this fact does not facilitate
Exh. Gas Temperature, [K] Te
efficient condition monitoring. Simultaneously, the
Maximum Compression Pressure, [bar] Pc
eigenvalues of the correlation matrix R usually
Maximum Combustion Pressure, [bar] Pz
decay rapidly, which suggests that only a few first
components of R and thus a few first factors can be
The FA algorithm is applied in a discrete fashion as
considered as principal containing a significant
follows:
share of the explained variance. Therefore, N in Eq.
- Select the array of recorded parameters (N =
(17) is limited to N ≤ 2.
600-time steps)
As mentioned above, the factor loadings
- Calculate correlation matrix R (Eq. (12))
indicate the relationship strength between
- Estimate matrix of factor loadings A (Eq. (14))
parameters and factors variation. Therefore, this
- Refining elements of A with EM algorithm (Eq.
property can be considered a feature that
(16))
characterizes every measured parameter's
- Evaluate performance index D1 for the
contribution to the variance of the performance
principal factor
index. Thus, the relative contribution of loadings to
- Evaluate the relative contribution of factor
the variance of factors is given as:
loadings λi,1 (Eq. (18))
ai2,n
i,n  , n 1, 2 - Slide the analysis window to get a new set of
p (18)
2
 ai,n recorded parameters (N = N + Δt, Δt = 150)
i 1 Figure 7 illustrates the relative contribution of
parameter variation to the variation of the
3.3 Application of FA to Condition Monitoring performance index. Regarding the considered
This section presents a case study failures, as one may note, there is a qualitative
demonstrating the FA methodology implementation difference in the factor loadings patterns, implying
to the condition monitoring framework. The that every failure may have a unique set of features.

Journal of the JIME Vo00， No． 00(2005) －7－ 日本マリンエンジニアリング学会誌 第 00 巻 第 00 号 (2005)

Journal of the JIME Vol. 56, No. 6（2021）
（2021） ―7― 日本マリンエンジニアリング学会誌 第56巻 第 6号 号
（2021）
（2021）
8 日本マリンエンジニアリング学会執筆要項
Intelligent Information Processing in the Problem of Propulsion System Condition Monitoring

At the same time, Figure 8 shows the evolution of state which corresponds to a very early moment of
performance indices D1. As can be seen, both indices degradation development.
clearly indicate the onset of an abnormal engine

Fig.6 Engine simulation model responses to the air filter clogging and AC fouling faults

Journal of the JIME Vol.00, No.00(2005) －8－ 日本マリンエンジニアリング学会誌 第 00 巻 第 00 号 (2005)

Journal of the JIME Vol. 56, No. 6（2021）
（2021） ―8― 日本マリンエンジニアリング学会誌 第56巻 第 6号 号
（2021）
（2021）
 日本マリンエンジニアリング学会執筆要項
Intelligent Information Processing in the Problem of Propulsion System Condition Monitoring 9

Fig.7 Relative contribution of measured parameters factor loadings to the variance of principal factor F1

Journal of the JIME Vo00， No． 00(2005) －9－ 日本マリンエンジニアリング学会誌 第 00 巻 第 00 号 (2005)

Journal of the JIME Vol. 56, No. 6（2021）
（2021） ―9― 日本マリンエンジニアリング学会誌 第56巻 第 6号 号
（2021）
（2021）
10 Intelligent Information Processing in the Problem of Propulsion System Condition Monitoring

Reference
1) Man, Y., Lundh, M. & MacKinnon, S.N. Managing
unruly technologies in the engine control room: from
problem patching to an architectural thinking and
standardization. WMU J Marit. Affairs Vol.:17,
pp.497–519 (2018).
2) WinGD: WiDE WinGD Integrated Digital Expert.
https://www.wingd.com/getattachment/Digital-solutio
ns/WiDE/WiDE_Brochure_8pp_FINAL.pdf/?lang=en-
GB
3) Introduction of ship maintenance management
system ClassNK CMAXS (in Japanese),
https://classnkcs.co.jp/archive/pdf/seminar01.pdf
4) Coraddu A, Oneto L, Ghio A, Savio S, Anguita D,
Figari M. Machine learning approaches for improving
condition-based maintenance of naval propulsion
plants. In Proc.: the Institution of Mechanical
Engineers, Part M: J. of Eng. for the Mar. Env. 2016;
Fig.8 The performance indices D1 of the engine condition
230(1): 136-153.
5) Gkerekos C., Lazakis I., and Theotokatos G. Ship

4. Conclusion Machinery Condition Monitoring Using Performance

Data Through Supervised Learning. In Proc.: Smart
The proposed framework for propulsion system Ship Technology Conference. Royal Institute of Naval
condition monitoring is based on the statistical Architects, London, (2017), pp.105-111
method of factor analysis and proved its 6) Baldi F, Theotokatos G, Andersson K. Development of
effectiveness. The performance index provides for a combined mean value–zero dimensional model and
efficient recognition of failure event which occurs in application for a large marine four-stroke diesel
the monitored component. Moreover, the engine simulation. Appl Energy 2015;154:402–15.
performance index clearly shows the onset of the 7) Bondarenko O., Fukuda T. (2020) Development of a
failure event, providing early detection of diesel engine's digital twin for predicting propulsion
degradation even in case of a minor change of the system dynamics. Energy, Vol.196.
recorded parameters. 8) Ken D.Q., Seaton R.E. A theoretical analysis of
Furthermore, the incipient failures cause thermal surface fouling. Br Chem. Eng. (1959); 4(5),
deviation of a group of the engine parameters, such pp.258-262.
as TCH speed, scavenging air pressure, exhaust gas 9) Harman H.H. Modern Factor Analysis. Second
temperature, etc. However, these deviations are Edition. The University Chicago Press. (1968)
usually of minor significance far from the alarm 10) Cureton E.E., D'Agostino R.B. Factor Analysis, An
monitor system thresholds. In this respect, the Applied Approach. Psychology Press. (1983)
relative contributions of parameter factor loadings 11) Basilevsky A. Statistical Factor Analysis and Related
to the principal factor variance clearly show the Methods, Theory and Application. A
anomaly pattern and are unique for every failure. Wiley-Interscience Publication. (1993)
Thus the latter can be used as a set of features
characterizing the failure, which are highly suitable
for the ML algorithms such as clustering and
classification. In turn, the ML algorithms can
identify the cause of engine performance
degradation, and such algorithms are currently
under development.

Journal of the JIME Vol.00, No.00(2005) －10－ 日本マリンエンジニアリング学会誌 第 00 巻 第 00 号 (2005)

Journal of the JIME Vol. 56, No. 6（2021）
（2021） ― 10 ― 日本マリンエンジニアリング学会誌 第56巻 第 6号 号
（2021）
（2021）