Condition-Based Maintenance of Naval Propulsion

Systems with Supervised Data Analysis.

Francesca Cipollinia , Luca Onetoa , Andrea Coraddub,∗, Alan John
Murphyb , Davide Anguitaa
a
DIBRIS - University of Genova, Via Opera Pia 13, I-16145 Genova, Italy
b
Marine, Offshore and Subsea Technology Group, School of Engineering
Newcastle University, Newcastle upon Tyne, NE1 7RU, UK

Abstract
The behavior and interaction of the main components of Ship Propulsion
Systems cannot be easily modeled with a priori physical knowledge, consid-
ering the large amount of variables influencing them. Data-Driven Models
(DDMs), instead, exploit advanced statistical techniques to build models
directly on the large amount of historical data collected by on-board au-
tomation systems, without requiring any a priori knowledge. DDMs are
extremely useful when it comes to continuously monitoring the propulsion
equipment and take decisions based on the actual condition of the propul-
sion plant. In this paper, the authors investigate the problem of performing
Condition-Based Maintenance through the use of DDMs. In order to con-
ceive this purpose, several state-of-the-art supervised learning techniques are
adopted, which require labeled sensor data in order to be deployed. A naval
vessel, characterized by a combined diesel-electric and gas propulsion plant,
has been exploited to collect such data and show the effectiveness of the
proposed approaches. Because of confidentiality constraints with the Navy
the authors used a real-data validated simulator and the dataset has been
published for free use through the UCI repository.
Keywords: Data Analysis, Naval Propulsion Systems, Condition-Based

∗
Corresponding Author
Email addresses: [email protected] (Francesca Cipollini),
[email protected] (Luca Oneto), [email protected] (Andrea
Coraddu), [email protected] (Alan John Murphy),
[email protected] (Davide Anguita)

Preprint submitted to Ocean Engineering May 25, 2018

Maintenance, Supervised Learning.

1. Introduction
Data Analysis (DA) is improving our ability to understand complex phe-
nomena much more rapidly than a priori physical models have done in the
past (Anguita et al., 2010; Boucheron et al., 2005; Coraddu et al., 2017).
Real-world systems are usually very complex, and hard to model only rely-
ing on the a priori knowledge of the problem (Witten et al., 2016; Peng et al.,
2010). On the contrary, Data-Driven Models (DDMs) are built exploiting ad-
vanced statistical techniques together with the historical data produced and
stored by the logging and monitoring apparatus, without requiring any a pri-
ori knowledge of the underlining physical phenomena (Vapnik, 1998; Györfi
et al., 2006). These characteristics make DDMs a suitable solution in all those
contexts where large amount of historical data is available, such as manu-
facturing (Peng et al., 2010), communications (Nguyen & Armitage, 2008),
finance (Shin et al., 2005), healthcare (Mannini & Sabatini, 2010), social net-
works (Pang et al., 2002), commerce (Das & Chen, 2001), and transportation
(Petersen et al., 2012; Budai et al., 2006; Smith et al., 2013).
Many of these sectors traditionally based their profit on empirical ex-
perience of sector specialists, and on simplified models built upon a priori
knowledge of the specific problems (Waeyenbergh & Pintelon, 2002; Peng
et al., 2010). This approach obviously requires a significant amount of time
and experience. At the same time, during the last decades, production plants
and products have been equipped with many sensors for different purposes:
automation, quality check, monitoring, and logging. The results of this pro-
cess is the availability of a huge amount of historical and real-time data
(Linoff & Berry, 2011). Recently, industries have realized that these data,
despite their management costs, can be considered as an opportunity to im-
prove their business since historical information can be adopted to create
new services or improve the quality of their products (Linoff & Berry, 2011;
Oneto et al., 2016b). In particular, they can leverage this huge amount of
data thanks to the DMMs which can rapidly and effectively extract useful
and actionable information (Witten et al., 2016).
In the shipbuilding industry, one of the main objectives of shipwrights
companies is to improve the technological quality of their products. For
example, they design more efficient hull shapes and propeller geometries,

2
study innovative propulsion systems, and reduce the overall production costs
(Carlton, 2011). Recently, many of these companies are evaluating different
DA solutions for improving the quality of their products, for monitoring the
equipment, and for maintenance purposes as integrative activities to their
core business (Waeyenbergh & Pintelon, 2002; Wang et al., 2015). In fact,
ships are already equipped with a network of sensors that collect data for
security, diagnostic and monitoring purposes, which DA can directly exploit
by taking advantage of these technologies (Waeyenbergh & Pintelon, 2002;
Petersen et al., 2012). DA, for example, offers the possibility to extract, from
the raw sensor data, useful information about the efficiency of the ship (Smith
et al., 2013), to reduce the fuel consumption (Coraddu et al., 2015), and to
improve maintenance activities (Coraddu et al., 2016). These data represent
strategic information for shipyards, operators, ship owners, and crews, since
they can be used for advisory, control, and fault detection purposes (Oneto
et al., 2016a).
Among the different problems, maintenance is probably the most critical
one since it could require drydocking, and the cost of retrieving a stricken
vessel offshore is non-trivial (Widodo & Yang, 2007; Mobley, 2002). Correct
maintenance ensures that a ship works as it was designed, with the desired
level performances, without impacting the service (Management et al., 1984).
Maintenance policies can be divided into two main categories (Budai-Balke,
2009; Abu-Elanien & Salama, 2010): Corrective (CM), and Preventive (PM).
In the past, the most common approach to this problem relied on the
CM where maintenance is performed only after a breakdown of a compo-
nent (Kothamasu & Huang, 2007). However, replacing a malfunctioning
component after it has failed during service, results in exceptional costs and
inevitable lower incomes. In PM, instead, a component is replaced when it
reaches the end of its life cycle, which can be computed adopting many esti-
mations. One of the historical way to perform this operation is to determine
a conservative average of the component life cycle adopting the experience
gained with all the components belonging to a specific class. Similarly to CM,
this particular type of PM can bring unnecessary costs too, if the replaced
component could have been used more than originally forecast. Moreover,
this technique does not guarantee to limit the number of faults in a fleet, since
a breakdown could still happen before the replacement takes place. In this
case, there is a trade-off between the number of breakdowns and the lifetime
estimation of the components, which is not easy to reach since the actual
ship usage can be very different from ship to ship. Nevertheless, Condition-

3
Based Maintenance (CBM) can be considered as a specification of PM, which
aims at reducing both the costs of CM and non-predictive PM by relying on
the exact decay state of each component and then by efficiently planning
its maintenance (Mann et al., 1995; ISO BS, 2004). Since, in most cases,
the decay state of each component cannot be tracked with a sensor, CBM
requires a model able to predict it based on other sensors available. Consid-
ering the estimated state of decay, it is possible to schedule each component’s
replacement before failures occur, maximizing its life cycle, according to the
time required for each maintenance and to the geographical location of the
ship (Mobley, 2002). As a result, the additional costs of CM and PM can be
replaced with the lower ones of equipping the propulsion system with sensors
and by collecting, storing, and analyzing these data for the purpose of cre-
ating effective predictive DDMs (Widodo & Yang, 2007; Mobley, 2002). It is
worth noting that shipping companies already invested in maintenance to in-
crease the reliability of their fleet and to reduce the costs. Several examples of
Planned Maintenance System (PMS) are available and specifically designed
for maintenance monitoring and integrated maintenance with class surveys
(Danaos, 2017; DNV-GL, 2017; Spectec, 2017). Nevertheless, as reported in
(Jackson et al., 2005), the use of effective predictive models is necessary to
achieve reliability, availability and maintainability. For this reason, in this
paper, the authors address the problem of building effective predictive mod-
els of the main components decay state in a Naval Propulsion System (NPS)
for CBM purposes.Differently from the already available PMS systems, the
proposed CBM approach can be adopted for real-time applications, to fore-
cast and identify maintenance necessities. While physical models have been
exploited in the past, recently, the ones based on statistical analysis have
shown to outperform them (Peng et al., 2010). DDMs are adopted to corre-
late the available sensors measurements with the decay state of the different
components of an NPS. The authors build virtual sensors able to continu-
ously estimate the state of decay of the components based on other sensors
measurements which are indirectly influenced by this decay.
Many examples of DA techniques applied to different CBM problems can
be found in literature (Jardine et al., 2006). Among other, Support Vec-
tor Machines (Poyhonen et al., 2004), Hidden Markov Models (Bunks et al.,
2000) and Kalman filter (Simani et al., 2003) are the most frequently used.
Examples of DA approaches applied to the marine industry can be found in
Palmé et al. (2011), where a standard Neural Network approach is used to
improve monitoring of Gas Turbines, while Kernel based methods are ap-

4
plied in Singer et al. (1995), and Coraddu et al. (2016). In Akinfiev et al.
(2008) and Bagavathiappan et al. (2013) image processing techniques are
adopted for hull condition assessment. In Basurko & Uriondo (2015) the en-
gine and propeller state is predicted adopting an Artificial Neural Network.
A complete overview can be found in Lazakis et al. (2016). In Ananthara-
man (2013) a model for NPS CBM is developed adopting Reliability Block
Diagrams and Fault Tree circuits, while in Michala et al. (2015) a CBM
is presented together with a Decision Support System, adopting Machinery
Risk Assessment techniques. In Mathur et al. (2001) and Byington et al.
(2002) a review of the different techniques applied in latest years for CBM
is given.
The novelty of this approach relies on the concurrent modeling of the four
main components decay status of an NPS: Propeller, Hull, Gas Turbine and
Gas Turbine Compressor. In fact, many works, present in literature, model
just one or maximum two of these components at the same time (Palmé
et al., 2011; Coraddu et al., 2016). The authors performed a regression
analysis where the target is to estimate the actual decay state of the compo-
nents described with an efficiency coefficient. The analysis has been carried
out comparing different state-of-the-art methodologies such as Kernel Meth-
ods (Shawe-Taylor & Cristianini, 2004), Neural Network (Rosenblatt, 1958),
Gaussian Processes (Rasmussen, 2006), Similarity Based Method (Duch,
2000), and Ensemble Methods (Breiman, 2001). Data coming from a Frigate
characterized by a COmbined Diesel ELectric And Gas (CODLAG) propul-
sion plant have been exploited to show the effectiveness of the proposed ap-
proaches and to benchmark them in a realistic maritime application. These
data will be available trough the UCI website, but they can already be down-
loaded at www.cbm.smartlab.ws. Confidentiality constraints with the Navy
require the authors to use a real-data validated complex numerical simulator
of a naval vessel (Coraddu et al., 2016). Nevertheless, even though the data
acquired do not belong to any active unit in service, the simulator which
was used to generate them, has been developed and tuned thanks to several
years experience on numerical modeling of real propulsion plants (Figari &
Altosole, 2007; Benvenuto & Campora, 2007).
The paper is organized as follows. Section 2 reports a general description
of the vessel, the numerical model, and the degradation phenomena. Section
3 presents a description of the dataset extracted from the numerical simulator
and published through UCI. Section 4 reports the proposed DDMs. Results
of the DDMs tested on these data are reported in Section 5 with conclusions

5
in Section 6.

2. Naval Propulsion System

2.1. Vessel Description
In this work authors focus on a Frigate, characterized by a CODLAG
NPS, widespread detailed in Coraddu et al. (2016). In particular, the gas
turbine (GT) mechanically drives the two controllable pitch propellers (CPP)
through a cross-connected gearbox. Besides, each shaft has its electric propul-
sion motor (EPM) mounted on the two shaft-lines. Two clutches between
the gearbox (GB) and the two EPM and another clutch between the GT and
the GB assure the possibility of using two different type of prime movers, i.e.
EPM and GT. Finally, the electric power is provided by four diesel genera-
tors (DG). In this work, only the GT operating conditions have been taken
into account.

2.2. Model Description

In this work, the authors consider an NPS numerical model developed
in the Matlab® Simulink® software environment within many years of re-
search (Altosole et al., 2009; Campora & Figari, 2003; Altosole et al., 2010).
The numerical model is composed of several modules each one representing
a single propulsion component such as the hull, the main engines, the pro-
pellers, the rudders, the gearbox, and the control system. In the previous
literature, the authors presented a model that considers the GT and GTC
decay performance (Coraddu et al., 2016; Benvenuto & Campora, 2005). The
model is now further improved to take into account the performance decay
of the HLL and PRP. This model is now readily to undertake an holistic
approach in addressing the performance decay by accounting the important
components as follows:
1. Gas Turbine (GT);
2. Gas Turbine Compressor (GTC);
3. Hull (HLL);
4. Propeller (PRP).

2.3. Degradation Model

The model described in Section 2.2 has been updated to take into account
the performance decay of the following components: GT, GTC, HLL, and
PRP. As reported in Altosole et al. (2014) and Coraddu et al. (2016), the

6
time domain performance decay has been modeled using suitable coefficients
embedded in the GTC low and high power modules of the model. This, allows
quantifying the variation of performance of each component and consequently
of the entire propulsion system.

GTC and GT Degradation Model

As reported in Tarabrin et al. (1998) and Meher-Homji et al. (1989), foul-
ing of the GTC increases the specific fuel consumption and the temperature
of the exhaust gas. In agreement with previous work, (Coraddu et al., 2016),
the effect of the fouling is simulated by reducing the numerical values of the
airflow rate Mc and of the isentropic efficiency ηc with a reduction factor kMc
and Kηc . The detailed description of the GTC and GT degradation model
is provided in Coraddu et al. (2016) and Altosole et al. (2014). The authors
applied a reduction factor, kMT , to the GT flow rate to represent the effect
of fouling.

PRP Degradation Model

An increase in the roughness of the blade surface is the primary cause of
marine PRP performance degradation (Khor & Xiao, 2011). This is caused
by accretions of marine organisms to the metal, alloy erosion and corrosion, or
combinations of these elements. The PRP decay status has been modeled by
increasing the torque coefficient (Kq ) and by reducing the thrust coefficient
(Kt ). The correction factors used for thrust reduction, kKt , and torque
increase, kKq , have been derived from Atlar et al. (2002); Wan et al. (2002).

HLL Degradation Model

The main factors that affect hull performance are the shape, the coating
used and the extent of fouling (Lindholdt et al., 2015). The resistance in-
crease due to fouling has been modeled utilizing a correction factor kH based
on the information reported in Hadler et al. (1962). Using the towing tank
data provided by the ship owner the resistance has been evaluated after one
and two years of vessel operations in agreement with the data available in
the literature (Altosole et al., 2014).

3. From Data to Condition-Based Maintenance

3.1. Dataset Creation
In this work, authors will use a real-data validated complex numerical
simulator of a Navy frigate described in Section 2 in order to build a realistic

7
set of data for designing and test purposes of DDMs. This dataset will be
released for the use to the research community on the widespread well-known
dataset repository of the University of California in Irvine (UCI) (Bache &
Lichman, 2013). Currently, it can be downloaded from www.cbm.smartlab.
ws.
The NPS model is characterised by the following input parameters:

• Speed: this parameter is controlled by the control lever. The latter can
only assume a finite number of positions lpi with i ∈ {0, · · · , 9}, which
in turn correspond to a finite set of possible configurations for fuel flow
and blade position. Each set point is designed to reach a desired speed
vi with i ∈ {0, · · · , 9}:

vi = 3 ∗ lpi [Knots] , ∀i ∈ {0, · · · , 9}. (1)

Note that, if the transients is not taken into account, lpi and vi are
deterministically related by a linear law. In the presented analysis the
transients between different speeds have been not considered.

• As reported in Section 2.3, the PRP thrust and torque decay limit over
two years of operations are:

kKt ∈ [0.9, 1.0], kKq ∈ [1.0, 1.1] (2)

kKt and kKq are respectively the components which define the decay
of the torque and the thrust provided by the propeller in time. They
are linearly correlated, since as the first decay of a certain quantity, the
latter decay of the same quantity (1 − kKt = kKq − 1). For this reason
only kKt will be analyzed, considering the linear dependency between
the two variables.

• The HLL decay has been modeled according to the available literature
(Hadler et al., 1962) as described in Section 2.3. The decay limit over
two years of operations are:

kH ∈ [1, 1.2] (3)

• GTC decay:

kMc ∈ [0.95, 1.0] (4)

8
 • GT decay:

kMt ∈ [0.975, 1.0] (5)

It is worth underlying that the performance decay functions reported in

Section 2.3 are empirically derived and are only a function of time. The real
degradation behavior of the physical asset should be described through spe-
cific functions able to describe the time dependency and the real operational
profile. In order to overcome this issue, the authors considered each possible
combination of GTC, GT, HLL, and PRP decays. Under this assumption,
the authors sampled the range of decays with a uniform grid characterized by
a degree of precision sufficient to have a good granularity of representation.
Given the above premises, the evolution of the system can be exhaustively
explored by simulating all its possible states. The space of possible states is
described via the following tuple:

(lp, kKt , kKq , kH, kMc , kMt )i , i ∈ {1, · · · , 589223} (6)

since:

lp ∈ S lp = {0, 3, 6, · · · , 27} , (7)

kKt ∈ S kKt = {0.9, .9 + 0.1/14, .09 + 0.2/14, · · · , 1.0} , (8)
kKq = 2 − kKt , (9)
kH ∈ S kH = {1.0, 1.0 + 0.2/14, 1.0 + 0.4/14, · · · , 1.2} , (10)
kMc ∈ S kMc = {0.95, 0.95 + 0.05/14, 0.95 + 0.1/14, · · · , 1.0} , (11)
kMt ∈ S kMt = {0.975, 0.975 + 0.025/14, 0.975 + 0.05/14, · · · , 1.0} . (12)

Once these quantities are fixed, the numerical model is run until the
steady state is reached. Then, the model is able to provide all the quantities
reported in Table 1. These subsets of models outputs are the same quantities
that the automation system installed on-board can acquire and store.
The simulator was run on a server equipped with four Intel® Xeon®
CPU E5-4620 2.2 GHz, 128 GB of RAM, 120 GB SSD disk, and Matlab®
R2016a.

3.2. Condition-Based Maintenance

This study aims at estimating the four decay variables described in the
previous section, adopting different DA techniques. This section reports how

9
Table 1: Measured values available from the continuous monitoring system

# Variable name Unit

1 Lever (lp) [ ]
2 Vessel Speed [knots]
3 GT shaft torque (GTT) [kN m]
4 GT Speed (GT rpm) [rpm]
5 Controllable Pitch Propeller Thrust stbd (CPP T stbd) [N]
6 Controllable Pitch Propeller Thrust port (CPP T port) [N]
7 Shaft Torque port (Q port) [kN m]
8 Shaft rpm port (rpm port) [rpm]
9 Shaft Torque stbd (Q stdb) [kN m]
10 Shaft rpm stbd (rpm stbd) [rpm]
11 HP Turbine exit temperature (T48) [o C]
12 Generator of Gas speed (GG rpm) [rpm]
13 Fuel flow (mf) [kg/s]
14 ABB TIC control signal (ABB Tic) [ ]
15 GT Compressor outlet air pressure (P2) [bar]
16 CGT Compressor outlet air temperature (T2) [o C]
17 External Pressure (Pext) [bar]
18 HP Turbine exit pressure (P48) [bar]
19 TCS TIC control signal (TCS Tic) [ ]
20 Thrust coefficient stbd (Kt stbd) [ ]
21 Propeller rps stbd (rps prop stbd) [rps]
22 Thrust coefficient port (Kt port) [ ]
23 Propeller rps port (rps prop port) [rps]
24 Propeller Torque port (Q prop port) [kN m]
25 Propeller Torque stbd (Q prop stbd) [kN m]

10
the data generated can be used to create effective predictive DDMs for the
CBM of an NPS.
The data described in Section 3.1 contain two sets of information: one
regarding the quantities that the automation system installed on-board can
acquire and store and the other one regarding the associated state of decay
(efficiency coefficient) of the different NPS components (GT, GTC, HLL,
and PRP). This problem can be straightforwardly mapped into a classical
multi-output regression problem (Vapnik, 1998; Coraddu et al., 2016) where
the aim is to predict the actual decay coefficient based on the automation
data coming from the sensors installed on-board.
Since in authors’ analysis four different components of an NPS are taken
into account (GT, GTC, HLL, and PRP), they decided to perform an incre-
mental analysis by breaking down the problem into simpler ones. In partic-
ular, first, only one NPS decayed component at the time is considered, then
authors consider the possible combination of two NPS decayed components
and so on until finally all the four NPS components are contemporarily con-

sidered. Consequently, since four NPS components are considered, 4 = 41
problems have to be solved when
just one NPS decayed component at the
4
time is considered, then 6 = 2 problems have to be solved when two NPS
decayed components at the time are considered and so on. Finally
when all
the four NPS components are contemporarily considered 1 = 44 problem
has to be solved, for a total of 15 = 4i=1 4i problems. Therefore, from the
P

dataset described in Section 3.1, 15 sub-datasets corresponding to the cases

mentioned above have been extracted. For the sake of clarity in Table 2 the
15 problems with the corresponding decay values are reported, not that in
all the problems lp ∈ S lp .

4. Machine Learning Techniques

In this section, the authors will present the Machine Learning (ML) tech-
niques adopted in order to build the CBM DDMs for NPS described in Sec-
tion 2, based on the data outlined in Section 3.
Let authors consider an input space X ⊆ Rd and an output space Y. Note
that, for what concerns this paper, X takes into account the different sensors
measurements, also called features, reported in Table 1, while the output
space Y depends on the particular problem identified in Section 3.2. ML
techniques aim at estimating the unknown rule µ : X → Y which associates

11
Table 2: The 15 sub-problems corresponding to considering different decayed component
at the time

Decayed Non Decayed

Prob. Decayed Non Decayed
Component Component
# Component Component
Name Name

1 GT GTC,HLL,PRP kMt ∈S kMt kMc =1,kH=1,kKt =1

2 GTC GT,HLL,PRP kMc ∈S kMc kMt =1,kH=1,kKt =1
3 HLL GT,GTC,PRP kH∈S kH kMc =1,kMt =1,kKt =1
4 PRP GT,GTC,HLL kKt ∈S kKt kMc =1,kMt =1,kH=1
5 GT,GTC HLL,PRP kMt ∈S kMt ,kMc ∈S kMc kH=1,kKt =1
6 GT,HLL GTC,PRP kMt ∈S kMt ,kH∈S kH kMc =1,kKt =1
7 GT,PRP GTC,HLL kMt ∈S kMt ,kKt ∈S kKt kMc =1,kH=1
8 GTC,HLL GT,PRP kMc ∈S kMc ,kH∈S kH kMt =1,kKt =1
9 GTC,PRP GT,HLL kMc ∈S kMc ,kKt ∈S kKt kMt =1,kH=1
10 HLL,PRP GT,GTC kH∈S kH ,kKt ∈S kKt kMt =1,kMc =1
11 GT,GTC,HLL PRP kMt ∈S kMt ,kMc ∈S kMc ,kH∈S kH kKt =1
12 GT,GTC,PRP HLL kMt ∈S kMt ,kMc ∈S kMc ,kKt ∈S kKt kH=1
13 GT,HLL,PRP GTC kMt ∈S kMt ,kH∈S kH ,kKt ∈S kKt kMc =1
14 GTC,HLL,PRP GT kMc ∈S kMc ,kH∈S kH ,kKt ∈S kKt kMt =1
15 GT,GTC,HLL,PRP kMt ∈S kMt ,kMc ∈S kMc ,kH∈S kH ,kKt ∈S kKt

an element y ∈ Y to an element x ∈ X . Note that, in general, µ can be non-

deterministic. An ML technique estimates µ through a learning algorithm
AH : Dn ×F → h, characterized by its set of hyperparameters H, which maps
a series of examples of the input/output relation contained in a datasets of
n samples Dn : {(x1 , y1 ) , · · · , (xn , yn )} into a function f : X → Y chosen in
a set of possible ones F.
When both xi and yi with i ∈ {1, · · · , n} are available, the problems
is named supervised and consequently supervised ML technique are adopted
(Vapnik, 1998). Regression is one of the most popular examples of supervised
ML problems (Shawe-Taylor & Cristianini, 2004) which is characterized by
Y ∈ R.
The error that f commits in approximating µ is measured with reference
to a loss function ` : X ×Y ×F → [0, ∞). Obviously, the error that f commits
over Dn , is optimistically biased since Dn has been used, together with F,
for building f itself. For this reason, another set of fresh data, composed
of m samples and called test set Tm = {(xt1 , y1t ), · · · , (xtm , ymt
)}, needs to be
t t
exploited. Note that, xi ∈ X and yi ∈ Y with i ∈ {1, · · · , m}, and the
association of yit to xti is again made based on µ. It has to be clarified that
Tm must contain both xti ∈ X and yit ∈ Y with i ∈ {1, · · · , m} to estimate
the error of f .

12
4.1. Measuring the Error
In this work, many state-of-the-art ML techniques will be tested and
their performances will be compared to understand what is the most suited
solution for building CBM DDMs for NPS.
In order to perform this analysis, authors have to define different measures
of error, also called indexes of performance, able to well characterize the
quality of the different CBM DDMs for NPS. Once f has been chosen based
on Dn it is possible to use the fresh set of data Tm in order to compute its
error based on different losses. The choice of the loss strongly depends on
the problem under examination (Rosasco et al., 2004).
In a typical regression framework, there are two losses that are mainly
used for estimating the quality of a regressor: the absolute loss `1 (f (x), y) =
|f (x)−y| and the squared loss `2 (f (x), y) = (f (x)−y)2 . Based on these losses
it is possible to define different indexes of performance, which differently
weight the distance between yit and f (xti ) (Ghelardoni et al., 2013):
• the Mean Squared Error (MSE) is computed by taking the mean square
loss value of f over Tm : MSE = 1/m m (xti ), yit );
P
`
i=1 2 (f
• the Normalized Mean Square Error (NMSE) is similar to the MSE
but a normalization term composed of the mean value of the differ-
ence
Pm between the true values and their mean is applied: NMSE =
`2 (f (xti ), yit )/∆;
i=1
• the Relative Error Percentage (REP) is similar to NMSE but the nor-
malization term is composed of the sum of the squared true values.
Then pP the result is square rooted and reported in percentage: REP =
m t t
100 i=1 `2 (f (xi ), yi )/ i=1 (yi ) ;
Pm t 2

• the Mean Absolute Error (MAE) Pism similar to MSE but the absolute
loss is used instead: MAE = /m i=1 `1 (f (xi ), yit );
1 t

• the Mean Absolute Percentage Error (MAPE) P can be described as the

MAE expressed in percentage: MAPE = /m m 100
i=1
`1 (f (xti ), yit )/y t ;
i
• the Pearson Product-Moment Correlation Coefficient (PPMCC) is a
measure of the linearPdependency between f (xti ) and yit with i ∈ {1,
m
· · ·, m}: PPMCC = i=1 (yit − y)(f (xti ) − ŷ)/ m i=1 (f (xi ) − ŷ) ,
pP t
pPn t
2 2
i=1 (yi − y)
where authors have exploited the following notation
m m m
1 X t 1 X t 1 X
y= y, ∆= (y − y)2 , ŷ = f (xti ). (13)
m i=1 i m i=1 i m i=1

13
4.2. Machine Learning Techniques
In this section, authors will present the supervised learning algorithms
exploited in this paper for building CBM DDMs for NPS. Moreover, authors
will show how to tune their performances by tuning their hyperparameters
during the so-called Model Selection MS phase (Kohavi et al., 1995; Bartlett
et al., 2002; Anguita et al., 2012). Finally, authors will also check for possi-
ble spurious correlation in the data by performing the Feature Selection (FS)
phase (Guyon & Elisseeff, 2003; Friedman et al., 2001; Yoon et al., 2005).
In fact, once f is built based on the different learning algorithm and has
been confirmed to be a sufficiently accurate representation of µ, it can be
interesting to investigate how the model f is affected by the different fea-
tures that have been exploited to build f itself during the feature ranking
procedure (Guyon & Elisseeff, 2003). As authors will describe later, for some
algorithms, the feature ranking procedure is a by-product of the learning
process itself and allows to simply check the physical plausibility of f .

4.2.1. Supervised Regressive Learning Algorithms

Supervised ML techniques can be grouped into different families, accord-
ing to the space of function F from which the learning algorithm chooses
the particular f , the approximation of µ, based on the available data Dn .
In fact, techniques belonging to the same family, share an affine F. Among
the several possible ML families, authors choose the state-of-the-art ones
which are commonly adopted in real-world application, and, in each family,
the best performing techniques are selected. In particular, Neural Networks
(NNs), Kernel Methods (KMs), Ensemble Methods (EMs), Bayesian Meth-
ods (BMs), and Lazy Methods (LMs) are adopted.
NNs are ML techniques which combine together many simple models of a
human brain neuron, called perceptrons (Rosenblatt, 1958), in order to build
a complex network. The neurons are organized in stacked layers connected
together by weights that are learned based on the available data via back-
propagation (Rumelhart et al., 1988). The hyperparameters of an NN HNN
are the number of layers h1 and the number of neurons for each layer h2,i with
i ∈ {1, · · · , h1 }. Note that it is assumed that NN with only one hidden layer
has h1 = 1. If the architecture of the NN consists of only one hidden layer,
it is called shallow (SNN) (Cybenko, 1989; Bishop, 1995), while, if multiple
layers are staked together, the architecture is defined as deep (DNN) (Hinton
et al., 2006; Bengio, 2009). Extreme Learning Machines (ELMs) are a par-
ticular kind of SNN, where the weights of the first layer are randomly chosen

14
while the ones of the output layers are computed according to the Regular-
ized Least Squares (RLS) principle (Huang et al., 2015, 2006; Cambria &
Huang, 2013). The hyperparameters of the ELM HELM are the number of
neurons of the hidden layer, h1 , and the RLS regularization hyperparameter
h2 (Huang et al., 2004).
KMs are a family of ML techniques which exploits the “Kernel trick” for
distances in order to extend linear techniques to the solution of non-linear
problems (Cristianini & Shawe-Taylor, 2000; Scholkopf, 2001). In the case
of regression, KMs select f as the function which minimizes the tradeoff
between the sum of the accuracy over the data, namely the empirical error,
and the complexity of the solution, namely the regularization term (Rifkin
et al., 2003; Cristianini & Shawe-Taylor, 2000). Two of the most known and
effective KM techniques are Kernelized Regularized Least Squares (KRLS),
and Support Vector Regression (SVR). The hyperparameters of the KRLS
HKRLS are: the kernel, which is usually fixed and in this paper author chose
the Gaussian Kernel for the reasons described in Keerthi & Lin (2003); Oneto
et al. (2015), its hyperparameter h1 and the regularization hyperparameter
h2 . SVR, instead, is a method which roots in the Statistical Learning Theory
(Vapnik, 1998) and differs from the KRLS as it introduces an  insensitivity
value for the errors (Shawe-Taylor & Cristianini, 2004). The hyperparameters
of the SVR HSVR are the same as the ones of KRLS plus the insensitivity
value h3 .
EMs ML techniques relies on the simple fact that combining the output
of several classifiers results in a much better performance than using any
one of them alone (Germain et al., 2015; Breiman, 2001). Random Forest
(RF) (Breiman, 2001) and Random Rotation Ensembles (RRF) (Blaser &
Fryzlewicz, 2015), two popular state-of-the-art and widely adopted meth-
ods, combine many decision trees in order to obtain a effective predictors
which have limited hyperparameter sensitivity and high numerical robust-
ness (Fernández-Delgado et al., 2014; Wainberg et al., 2016). Both RF and
RRF have hidden hyperparameters which are considered fixed in this work
because of their limited effect (Orlandi et al., 2016).
BMs are ML techniques where, instead of choosing a particular f ∈ F a
distribution for choosing f ∈ F is defined (Gelman et al., 2014). Gaussian
Processes (GP) learning algorithm is a popular BM (Rasmussen, 2006) which
employs a collection of Gaussians in order to compute the posterior distribu-
tion of the f (x). In fact, this algorithm defines the probability distribution
of the output values as a sum of Gaussians whose variance is fixed according

15
to the training data. The hyperparameter of the GP HGP is the parameter
which governs the Gaussians width h1 .
LMs ML techniques are learning method in which the definition of f is
delayed until f (x) needs to be computed (Duch, 2000). LMs approximate
µ locally with respect to x. K-Nearest Neighbors (KNN) is one of the most
popular LM due to its implementation simplicity and effectiveness (Cover
& Hart, 1967). The hyperparameter of the KNN HKNN is the number of
neighbors of x to be considered h1 .

4.2.2. Model Selection

MS deals with the problem of tuning the hyperparameters of each learning
algorithm (Anguita et al., 2012). Several methods exist for MS purpose: re-
sampling methods, like the well-known k-Fold Cross Validation (KCV) (Ko-
havi et al., 1995) or the nonparametric Bootstrap (BTS) approach (Efron
& Tibshirani, 1994; Anguita et al., 2000) approaches, which represent the
state-of-the-art MS approaches when targeting real-world applications. Re-
sampling methods rely on a simple idea: the original dataset Dn is resampled
once or many (nr ) times, with or without replacement, to build two inde-
pendent datasets called training, and validation sets, respectively Lrl and Vvr ,
with r ∈ {1, · · · , nr }. Note that Lrl ∩ Vvr = , Lrl ∪ Vvr = Dn . Then, in order
to select the best combination the hyperparameters H in a set of possible
ones H = {H1 , H2 , · · · } for the algorithm AH or, in other words, to perform
the MS phase, the following procedure has to be applied:
nr
∗ 1 X 1 X
H : min `(AH,Lrl (xi ), yi ), (14)
H∈H nr v
r=1 (xi ,yi )∈Vvr

where AH,Lrl is a model built with the algorithm A with its set of hyper-
parameters H and with the data Lrl . Since the data in Lrl are independent
from the ones in Vvr , the idea is that H∗ should be the set of hyperparameters
which allows to achieve a small error on a data set that is independent from
the training set.
If r = 1, if l and v are aprioristically set such that n = l + v, and if the
resample procedure is performed without replacement, the hold out method is
obtained (Anguita et al., 2012). For implementing
the completen k-foldn cross
n n− n


validation, instead, it is needed to set r ≤ k k
k , l = (k − 2) , v =
k k
, and
n
t = k and the resampling must be done without replacement (Kohavi et al.,
1995; Arlot & Celisse, 2010; Anguita et al., 2012). Finally, for implementing

16
the bootstrap, l = n and Lrl must be sampled with replacement from Dn ,
while Vvr and Ttr are sampled without replacement from the sample of Dn
that have not been sampled in Lrl (Efron & Tibshirani, 1994; Anguita et al.,
2n−1


2012). Note that for the bootstrap procedure r ≤ n . In this paper the
BTS is exploited because it represents the state-of-the-art approach (Efron
& Tibshirani, 1994; Anguita et al., 2012).

4.3. Feature Selection

Once the CBM NPS models are built and have been confirmed to be
sufficiently accurate representation of the real decays of the components, it
can be interesting to investigate how these models are affected by the different
features used in the model identification phase (see Table 1).
In DA this procedure is called FS or Feature Ranking (Guyon & Elisseeff,
2003; Friedman et al., 2001; Yoon et al., 2005). This process allows detecting
if the importance of those features, that are known to be relevant from a
physical perspective, is appropriately described by the different CBM NPS
models. The failure of the statistical model to properly account for the
relevant features might indicate poor quality in the measurements or spurious
correlations. FS therefore represents an important step of model verification,
since it should generate consistent results with the available knowledge of the
physical system under exam.
In addition, the EMs can also be used to perform a very stable FS proce-
dure. The procedure is a combination of EMs, together with the permutation
test (Good, 2013), in order to perform the selection and the ranking of the
features. In details, for every tree, two quantities are computed: the first one
is the error on the out-of-bag samples as they are used during prediction,
while the second one is the error on the out-of-bag samples after a random
permutation of the values of variable j. These two values are then subtracted
and the average of the result over all the trees in the ensemble is the raw
importance score for variable j (mean decrease in accuracy). This procedure
was adopted since it can be easily carried out during the main prediction
process inexpensively.

5. Experimental results
In this section, the authors report the results obtained by the different
methods applied to the CBM of the main components of an NPS, in the

17
proposed regression framework, as described in Section 2, based on the data
described in Section 3.
As described in Section 4, the authors tackle a regression problem where
the actual value of the decay parameters needs to be estimated.
The different datasets considered in Section 3.2, corresponding to the
15 problems of Table 2 {P 1, · · · , P 15}, were divided into training and test
set, respectively Dn and Tm , as reported in Section 4. Moreover, different
dimensions of the training set n ∈ {10, 24, 55, 130, 307, 722, 1700, 4000} were
considered.
HLL
PRP 6
3.5
KRLS
KRLS
ELM
ELM
3 KNN 5 KNN
RF
RF
GP
GP
2.5 SVR
SVR 4
DNN DNN
RFR RFR
2 SNN
MAPE

SNN
MAPE

1.5

2
1

1
0.5

0 0
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
n n

GTC GT
1.8 0.8
KRLS KRLS
1.6 ELM 0.7 ELM
KNN KNN
1.4 RF RF
GP 0.6 GP
SVR SVR
1.2
DNN DNN
0.5
RFR RFR
1 SNN SNN
MAPE

MAPE

0.4
0.8
0.3
0.6

0.2
0.4

0.2 0.1

0 0
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
n n

Figure 1: MAPE of the models learned with the different algorithms (DNN, SNN, ELM,
SVR, KRLS, KNN, and GP) when varying n for P 15 and for the four main NPS compo-
nents.

For each supervised regression learning task and each ML technique, an

MS procedure was performed, as described in Section 4.2.2. Here-below, the

18
list of hyperparameters tested during the MS, with their respective intervals,
is reported:
1. DNN: the set of hyperparameters is HDN N = {h1 , h2,1 , · · · , h2,h1 } and
authors chose it in HDNN = {1, 3, 5, 7, 10} × {10, 101.2 · · · , 103 } × · · · ×
{10, 101.2 · · · , 103 };
2. SNN: the set of hyperparameters is HSN N = {h1 } and authors chose it
in HSNN = {1, 3, 5, 7, 10};
3. ELM: the set of hyperparameters is HELM = {h1 , h2 } and authors
chose it in HELM = {10, 101.2 , · · · , 103 } × {10−2 , 10−1.5 · · · , 102 };
4. SVR: the set of hyperparameters is HSV R = {h1 , h2 , h3 } and authors
chose it in HSVR = {10−6 , 10−5 , · · · , 103 } × {10−2 , 10−1.4 , · · · , 103 } ×
{10−2 , 10−1.4 , · · · , 103 };
5. KRLS: the set of hyperparameters is HKRLS = {h1 , h2 } and authors
chose it in HKRLS = {10−2 , 10−1.4 , · · · , 103 } × {10−2 , 10−1.4 , · · · , 103 };
6. KNN: the set of hyperparameters is HKN N = {h1 } and authors chose
it in HKNN = {1, 3, 7, 13, 27, 51};
7. GP: the set of hyperparameters is HGP = {h1 } and authors chose it in
HGP = {100 , 100.3 , · · · , 103 };
The performances of each model are measured according to the metrics
described in Section 4.1. Each experiment was performed 10 times in order to
obtain statistical relevant result, and the t-student 95% confidence interval is
reported when space in the table was available without compromising their
readability.
For SNN and DNN the Python Keras library (Chollet, 2015) has been
exploited. For ELM, SVR, KRLS, KNN, and GKNN a custom R implemen-
tation has been developed. For RF the R package of Liaw & Wiener (2002)
has been exploited. For RFE the implementation of Blaser & Fryzlewicz
(2015) has been exploited. For GP the R package of Zeileis et al. (2004) has
been exploited.

5.1. Regression Models Error Results

In this section, the results on the regression problem are reported. Note
that, based on the problem under exam, just a subset of the components may
decay. Moreover, since this dataset cardinality is limited, it is possible that a
particular combination of problem and cardinality of the training set cannot
be tested. For example, when just one component decays, nearly hundred
samples are available (see Section 3) and then a maximum of n = 55. Note
that, in a real-world scenario, the only useful problem is P 15, namely when

19
all the components decay contemporarily. In Figures 1 authors report the
MAPE of the models learned with the different algorithms when varying n for
P 15 and the four NPS components. In Figures 2 authors report the MAPE
of the DNN (the best performing model) when varying n for the different
problems under examination and the four NPS components.
HLL
PRP 6
3.5
P10:PRP.HLL
P4:PRP
P14:PRP.HLL.GTC
P10:PRP.HLL
3 P14:PRP.HLL.GTC 5 P15:PRP.HLL.GTC.GT
P13:PRP.HLL.GT
P15:PRP.HLL.GTC.GT
P3:HLL
P13:PRP.HLL.GT
2.5 P8:HLL.GTC
P9:PRP.GTC 4
P12:PRP.GTC.GT P11:HLL.GTC.GT
P7:PRP.GT P6:HLL.GT
2

MAPE
MAPE

1.5

2
1

1
0.5

0 0
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
n n

GTC GT
1.6 0.8
P14:PRP.HLL.GTC P15:PRP.HLL.GTC.GT
1.4 P15:PRP.HLL.GTC.GT 0.7 P13:PRP.HLL.GT
P9:PRP.GTC P12:PRP.GTC.GT
P12:PRP.GTC.GT P7:PRP.GT
1.2 P8:HLL.GTC 0.6 P11:HLL.GTC.GT
P11:HLL.GTC.GT P6:HLL.GT
P2:GTC P5:GTC.GT
1 0.5
P5:GTC.GT P1:GT
MAPE

MAPE

0.8 0.4

0.6 0.3

0.4 0.2

0.2 0.1

0 0
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
n n

Figure 2: MAPE of the models learned with DNN when varying n for the different problems
P 1, · · · , P 15 and for the four main NPS components.

From the different figures it is possible to observe that:

• as expected the larger is n the better performances are achieved by the
learned models (see Figures 2 and 1);
• the models learned with ELM, SNN, and especially DNN generally
show the best performances (see Figure 1);
• the larger is the number of decaying components authors consider the
lower performances are achieved by the learned models (see Figure 2);

20
 • the most complicate decay to predict is the one of the HLL, in fact
the problems where the HLL decays are the ones which show lower
accuracies (see Figure 2);
• unfortunately, as it can be seen from 2 and 1, this approach requires
a significant amount of labelled data, which could not be retrieved in
many operational scenarios. In fact, while the sensors data coming from
the automation system are easy to collect, the information regarding
the associated state of decay is not so easy to retrieve. Collecting the
state of decay of the different NPS components requires the intervention
of an experienced operator and, in some cases, to stop the vessel or even
to put the ship in a dry dock.
10-4 P1:PRP 10-3 P2:HLL
4.5 1.8

4 1.6

3.5 1.4
Mean Decrease in Accuracy

Mean Decrease in Accuracy

3 1.2

2.5 1

2 0.8

1.5 0.6

1 0.4

0.5 0.2

0 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
feature feature

10-6 P4:GT
10-5 P3:GTC 9
5

4.5 8

4 7
Mean Decrease in Accuracy
Mean Decrease in Accuracy

3.5 6

3
5
2.5
4
2
3
1.5
2
1

0.5 1

0 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
feature feature

Figure 3: FS performed with RF for the four main NPS components for problem P 1, P 2,
P 3, and P 4

Finally in Figure 3 and 4 the FS phase is reported, performed with RF, for

21
P 1, P 2, P 3, and P 4 (see Figure 3) and for P 15 (see Figure 4). In particular,
for each problem and each feature, the mean decrease in accuracy is reported
as described in Section 4.3.
10-4 P15:PRP
8 10-3 P15:HLL
3

7
2.5
6
Mean Decrease in Accuracy

Mean Decrease in Accuracy

2
5

4 1.5

3
1
2

0.5
1

0 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
feature feature

10-5 P15:GT
10-4 P15:GTC 6
3.5

3 5
Mean Decrease in Accuracy
Mean Decrease in Accuracy

2.5
4

2
3

1.5

2
1

1
0.5

0 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
feature feature

Figure 4: FS performed with RF for the four main NPS components for problem P 15.

From Figure 3 and 4 it is possible to observe that:

• as expected, when just one component decays (see Figure 3), few pre-
dictors have strong predictive power while when all the components
decay (see Figure 4) many predictors need to be considered in order to
achieve satisfying accuracies;
• from Figure 3, it is possible to note that the RF model can adequately
accounts for the relevant features as the outcome is consistent with
the available knowledge (note that even if DNN has higher predictive
capabilities with respect to RF the latter is still competitive, see Figure
1). In fact, for the P 1 and P 2 (PRP and HLL decays) the features

22
 22 and 24 (Thrust coefficient stbd and port) have strong predictive
power. Moreover, for the P 3 (GTC decay) the features describing
the thermodynamic process, 17, 19, and 20 (GT Compressor outlet air
pressure, External Pressure ,and HP Turbine exit pressure) have the
highest predictive power. Finally, several features are necessary for the
GT decay prediction in P 4.

6. Conclusions
The behavior and interaction of the main components of Ship Propulsion
Systems cannot be easily modeled with a priori physical knowledge, consid-
ering the large amount of variables influencing them. In fact, in this work,
authors showed that by exploiting the most recent statistical techniques and
the large amount of historical data collected by the current on-board automa-
tion systems it is possible to build effective Data-Driven Models which do
not require any a priori knowledge. In particular, the developed models are
able to continuously monitor the propulsion equipment to avoid Preventive
or Corrective Maintenance and take decisions based on the actual condition
of the propulsion plant. A naval ship, characterized by a COmbined Diesel
ELectric And Gas propulsion plant, has been exploited to show the effective-
ness of the proposed approaches. Because of confidentiality constraints with
the Navy the authors used a real-data validated simulator and the dataset
has been published for free use.
In conclusion, this study proved that it is possible to treat a Condition-
Based Maintenance problem in a supervised fashion adopting regression tech-
niques. These models can be adopted for real-time applications directly on-
board, to easily and quickly identify maintenance necessities. Nevertheless,
the dataset adopted for this study can difficulty be collected in a real-world
scenario. In this paper, the authors propose an example of DMMS for CBM,
providing evidence of the possible advantages in the experimental section
with an indirect measurement of the potential savings. The proposed models
are quite effective and reliable, and, for this reason, the authors’ approach
can correctly identify the state of decay of the system, bringing to an opti-
mal predictive maintenance strategy with a reduction in the operational and
maintenance costs. On the other hand, the authors would like to address
some findings which need to be solved. First, supervised DMMS require a
significant amount of labelled data to efficiently monitor the NPS compo-
nents state, which would require the operational stop or even the drydocking

23
of the vessel. Vessels are rarely available in the harbour for maintenance,
and the necessary parameters cannot be easily acquired when the ship is op-
erative. Second, the experiments carried out are relative to a specific vessel,
and this work cannot be straightforwardly mapped to any new ship without a
further investigation on transfer learning to understand how to use available
data from other vessels. To solve the first issue, the authors suggest the pos-
sibility of adopting a relaxation on the decay value prediction requirement,
by transforming the regression problem into a simpler one by predicting only
the necessity of each component to be replaced or not, instead of the precise
component decay. Moreover, in the future, authors will focus their efforts on
the development of supervised and unsupervised methods which could solve
this problem and allow a simplification of the data collection procedure.

References
Abu-Elanien, A. E. B., & Salama, M. M. A. (2010). Asset management
techniques for transformers. Electric power systems research, 80 , 456–464.

Akinfiev, T. S., Armada, M. A., & Fernandez, R. (2008). Nondestructive

testing of the state of a ship’s hull with an underwater robot. Russian
Journal of Nondestructive Testing, 44 , 626–633.

Altosole, M., Benvenuto, G., & Campora, U. (2010). Numerical modelling

of the engines governors of a codlag propulsion plant. In International
Conference on Marine Sciences and Technologies.

Altosole, M., Benvenuto, G., Figari, M., & Campora, U. (2009). Real-time
simulation of a cogag naval ship propulsion system. Journal of Engineering
for the Maritime Environment, 223 , 47–62.

Altosole, M., Campora, U., Martelli, M., & Figari, M. (2014). Performance
decay analysis of a marine gas turbine propulsion system. Journal of Ship
Research, 58 , 117–129.

Anantharaman, M. (2013). Using reliability block diagrams and fault tree

circuits, to develop a condition based maintenance model for a vessel’s
main propulsion system and related subsystems. Transnav: International
Journal On Marine Navigation And Safety Of Sea Transportation, 7 .

24
Anguita, D., Boni, A., & Ridella, S. (2000). Evaluating the generalization
ability of support vector machines through the bootstrap. Neural Process-
ing Letters, 11 , 51–58.

Anguita, D., Ghio, A., Greco, N., Oneto, L., & Ridella, S. (2010). Model
selection for support vector machines: Advantages and disadvantages of
the machine learning theory. In International Joint Conference on Neural
Networks.

Anguita, D., Ghio, A., Oneto, L., & Ridella, S. (2012). In-sample and out-of-
sample model selection and error estimation for support vector machines.
IEEE Transactions on Neural Networks and Learning Systems, 23 , 1390–
1406.

Arlot, S., & Celisse, A. (2010). A survey of cross-validation procedures for

model selection. Statistics surveys, 4 , 40–79.

Atlar, M., Glover, E. J., Candries, M., Mutton, R. J., & Anderson, C. D.
(2002). The effect of a foul release coating on propeller performance. In
Marine Science and Technology for Environmental Sustainability.

Bache, K., & Lichman, M. (2013). UCI machine learning repository.

Bagavathiappan, S., Lahiri, B. B., Saravanan, T., Philip, J., & Jayakumar, T.
(2013). Infrared thermography for condition monitoring-a review. Infrared
Physics & Technology, 60 , 35–55.

Bartlett, P. L., Boucheron, S., & Lugosi, G. (2002). Model selection and
error estimation. Machine Learning, 48 , 85–113.

Basurko, O. C., & Uriondo, Z. (2015). Condition-based maintenance for

medium speed diesel engines used in vessels in operation. Applied Thermal
Engineering, 80 , 404–412.

Bengio, Y. (2009). Learning deep architectures for ai. Foundations and

trends® in Machine Learning, 2 , 1–127.

Benvenuto, G., & Campora, U. (2005). A gas turbine modular model for
ship propulsion studies. In Symposium on High Speed Marine Vehicles.

25
Benvenuto, G., & Campora, U. (2007). Performance prediction of a faulty
marine diesel engine under different governor settings. In International
Conference on Marine Research And Transportation.

Bishop, C. M. (1995). Neural networks for pattern recognition. Oxford uni-

versity press.

Blaser, R., & Fryzlewicz, P. (2015). Random rotation ensembles. The Journal
of Machine Learning Research, 2 , 1–15.

Boucheron, S., Bousquet, O., & Lugosi, G. (2005). Theory of classification:

A survey of some recent advances. ESAIM: probability and statistics, 9 ,
323–375.

Breiman, L. (2001). Random forests. Machine learning, 45 , 5–32.

Budai, G., Huisman, D., & Dekker, R. (2006). Scheduling preventive railway
maintenance activities. The International Journal of Advanced Manufac-
turing Technology, 57 , 1035–1044.

Budai-Balke, G. (2009). Operations research models for scheduling railway

infrastructure maintenance. Rozenberg Publishers.

Bunks, C., McCarthy, D., & Al-Ani, T. (2000). Condition-based maintenance

of machines using hidden markov models. Mechanical Systems and Signal
Processing, 14 , 597–612.

Byington, C. S., Roemer, M. J., & Galie, T. (2002). Prognostic enhance-

ments to diagnostic systems for improved condition-based maintenance.
In Aerospace Conference.

Cambria, E., & Huang, G.-B. (2013). Extreme learning machines. IEEE
Intelligent Systems, 28 , 30–59.

Campora, U., & Figari, M. (2003). Numerical simulation of ship propul-

sion transients and full-scale validation. Journal of Engineering for the
Maritime Environment, 217 , 41–52.

Carlton, J. (2011). Marine propellers and propulsion. Butterworth-

Heinemann.

Chollet, F. (2015). Keras. https://keras.io.

26
Coraddu, A., Oneto, L., Baldi, F., & Anguita, D. (2015). Ship efficiency
forecast based on sensors data collection: Improving numerical models
through data analytics. In OCEANS 2015-Genova.

Coraddu, A., Oneto, L., Baldi, F., & Anguita, D. (2017). Vessels fuel con-
sumption forecast and trim optimisation: A data analytics perspective.
Ocean Engineering, 130 , 351–370.

Coraddu, A., Oneto, L., Ghio, A., Savio, S., Anguita, D., & Figari, M. (2016).
Machine learning approaches for improving condition-based maintenance
of naval propulsion plants. Proceedings of the Institution of Mechanical
Engineers Part M: Journal of Engineering for the Maritime Environment,
230 , 136–153.

Cover, T., & Hart, P. (1967). Nearest neighbor pattern classification. IEEE
transactions on information theory, 13 , 21–27.

Cristianini, N., & Shawe-Taylor, J. (2000). An introduction to support vector

machines and other kernel-based learning methods. Cambridge university
press.

Cybenko, G. (1989). Approximation by superpositions of a sigmoidal func-

tion. Mathematics of Control, Signals, and Systems (MCSS), 2 , 303–314.

Danaos (2017). Planned maintenance system. https://danaosmc.

wordpress.com/2016/03/30/planned-maintenance-system.

Das, S., & Chen, M. (2001). Yahoo! for amazon: Extracting market sen-
timent from stock message boards. In Asia Pacific finance association
annual conference (APFA).

DNV-GL (2017). Shipmanager technical. https://www.dnvgl.com/

services/planned-maintenance-system-for-technical-ship-
management-shipmanager-technical-1509.

Duch, W. (2000). Similarity-based methods: a general framework for classi-

fication, approximation and association. Control and Cybernetics, 29 .

Efron, B., & Tibshirani, R. J. (1994). An introduction to the bootstrap. CRC

press.

27
Fernández-Delgado, M., Cernadas, E., Barro, S., & Amorim, D. (2014). Do
we need hundreds of classifiers to solve real world classification problems.
The Journal of Machine Learning Research, 15 , 3133–3181.

Figari, M., & Altosole, M. (2007). Dynamic behaviour and stability of marine
propulsion systems. Journal of Engineering for the Maritime Environment,
221 , 187–205.

Friedman, J., Hastie, T., & Tibshirani, R. (2001). The elements of statistical
learning. Springer series in statistics Springer, Berlin.

Gelman, A., Carlin, J. B., Stern, H. S., & Rubin, D. B. (2014). Bayesian
data analysis, vol. 2 . Chapman and Hall/CRC.

Germain, P., Lacasse, A., Laviolette, F., Marchand, M., & Roy, J. F. (2015).
Risk bounds for the majority vote: From a pac-bayesian analysis to a
learning algorithm. The Journal of Machine Learning Research, 16 , 787–
860.

Ghelardoni, L., Ghio, A., & Anguita, D. (2013). Energy load forecasting
using empirical mode decomposition and support vector regression. IEEE
Transactions on Smart Grid , 4 , 549–556.

Good, P. (2013). Permutation tests: a practical guide to resampling methods

for testing hypotheses. Springer Science & Business Media.

Guyon, I., & Elisseeff, A. (2003). An introduction to variable and feature

selection. The Journal of Machine Learning Research, 3 , 1157–1182.

Györfi, L., Kohler, M., Krzyzak, A., & Walk, H. (2006). A distribution-free
theory of nonparametric regression. Springer Science & Business Media.

Hadler, J. B., Wilson, C. J., & Beal, A. L. (1962). Ship standardisation trial
performance and correlation with model predictions. In The Society of
Naval Architects and Marine Engineers.

Hinton, G. E., Osindero, S., & Teh, Y. W. (2006). A fast learning algorithm
for deep belief nets. Neural computation, 18 , 1527–1554.

Huang, G., Huang, G.-B., Song, S., & You, K. (2015). Trends in extreme
learning machines: A review. Neural Networks, 61 , 32–48.

28
Huang, G.-B., Zhu, Q.-Y., & Siew, C.-K. (2004). Extreme learning machine:
a new learning scheme of feedforward neural networks. In IEEE Interna-
tional Joint Conference on Neural Networks.

Huang, G. B., Zhu, Q. Y., & Siew, C. K. (2006). Extreme learning machine:
theory and applications. Neurocomputing, 70 , 489–501.

ISO BS (2004). 13372: 2012, condition monitoring and diagnostics of

machines-vocabulary. In British Standards.

Jackson, Y., Tabbagh, P., Gibson, P., & Seglie, E. (2005). The new depart-
ment of defense (dod) guide for achieving and assessing ram. In Reliability
and Maintainability Symposium.

Jardine, A. K. S., Lin, D., & Banjevic, D. (2006). A review on machinery

diagnostics and prognostics implementing condition-based maintenance.
Mechanical systems and signal processing, 20 , 1483–1510.

Keerthi, S. S., & Lin, C. J. (2003). Asymptotic behaviors of support vector

machines with gaussian kernel. Neural computation, 15 , 1667–1689.

Khor, Y. S., & Xiao, Q. (2011). Cfd simulations of the effects of fouling and
antifouling. Ocean Engineering, 38 , 1065–1079.

Kohavi, R. et al. (1995). A study of cross-validation and bootstrap for ac-

curacy estimation and model selection. In International Joint Conference
on Artificial Intelligence.

Kothamasu, R., & Huang, S. H. (2007). Adaptive mamdani fuzzy model for
condition-based maintenance. Fuzzy Sets and Systems, 158 , 2715–2733.

Lazakis, I., Dikis, K., Michala, A. L., & Theotokatos, G. (2016). Advanced
ship systems condition monitoring for enhanced inspection, maintenance
and decision making in ship operations. Transportation Research Procedia,
14 , 1679–1688.

Liaw, A., & Wiener, M. (2002). Classification and regression by randomfor-

est. R News, 2 , 18–22.

Lindholdt, A., Dam-Johansen, K., Olsen, S. M., Yebra, D. M., & Kiil, S.
(2015). Effects of biofouling development on drag forces of hull coatings for

29
 ocean-going ships: a review. Journal of Coatings Technology and Research,
12 , 315–444.

Linoff, G. S., & Berry, M. J. A. (2011). Data mining techniques: for market-
ing, sales, and customer relationship management. John Wiley & Sons.

Management, G. B. Q., Committee, S. S., & Institution, B. S. (1984). British

Standard Glossary of Maintenance Management Terms in Terotechnology.
British Standards Institution.

Mann, L., Saxena, A., & Knapp, G. M. (1995). Statistical-based or condition-

based preventive maintenance? Journal of Quality in Maintenance Engi-
neering, 1 , 46–59.

Mannini, A., & Sabatini, A. M. (2010). Machine learning methods for classi-
fying human physical activity from on-body accelerometers. Sensors, 10 ,
1154–1175.

Mathur, A., Cavanaugh, K. F., Pattipati, K. R., Willett, P. K., & Galie,
T. R. (2001). Reasoning and modeling systems in diagnosis and prognosis.
In SPIE Aerosense Conference.

Meher-Homji, C. B., Focke, A. B., & Wooldridge, M. B. (1989). Fouling of

axial flow compressors - causes, effects, detection, and control. In Eigh-
teenth Turbomachinery Symposium.

Michala, A. L., Lazakis, I., & Theotokatos, G. (2015). Predictive maintenance

decision support system for enhanced energy efficiency of ship machinery.
In International Conference on Shipping in Changing Climates.

Mobley, R. K. (2002). An introduction to predictive maintenance.

Butterworth-Heinemann.

Nguyen, T. T. T., & Armitage, G. (2008). A survey of techniques for in-

ternet traffic classification using machine learning. IEEE Communications
Surveys & Tutorials, 10 , 56–76.

Oneto, L., Anguita, D., Coraddu, A., Cleophas, T., & Xepapa, K. (2016a).
Vessel monitoring and design in industry 4.0: A data driven perspective.
In IEEE International Forum on Research and Technologies for Society
and Industry Leveraging a better tomorrow .

30
Oneto, L., Fumeo, E., Clerico, G., Canepa, R., Papa, F., Dambra, C., Mazz-
ino, N., & Anguita, D. (2016b). Advanced analytics for train delay pre-
diction systems by including exogenous weather data. In International
Conference on Data Science and Advanced Analytics.

Oneto, L., Ghio, A., Ridella, S., & Anguita, D. (2015). Support vector
machines and strictly positive definite kernel: The regularization hyper-
parameter is more important than the kernel hyperparameters. In IEEE
International Joint Conference on Neural Networks.

Orlandi, I., Oneto, L., & Anguita, D. (2016). Random forests model selection.
In European Symposium on Artificial Neural Networks, Computational In-
telligence and Machine Learning.

Palmé, T., Breuhaus, P., Assadi, M., Klein, A., & Kim, M. (2011). New
alstom monitoring tools leveraging artificial neural network technologies.
In Turbo Expo: Turbine Technical Conference and Exposition.

Pang, B., Lee, L., & Vaithyanathan, S. (2002). Thumbs up? sentiment
classification using machine learning techniques. In ACL-02 conference on
Empirical methods in natural language processing-Volume 10 .

Peng, Y., Dong, M., & Zuo, M. J. (2010). Current status of machine prognos-
tics in condition-based maintenance: a review. The International Journal
of Advanced Manufacturing Technology, 50 , 297–313.

Petersen, J. P., Winther, O., & Jacobsen, D. J. (2012). A machine-learning

approach to predict main energy consumption under realistic operational
conditions. Ship Technology Research, 59 , 64–72.

Poyhonen, S., Jover, P., & Hyotyniemi, H. (2004). Signal processing of vi-
brations for condition monitoring of an induction motor. In Symposium
on Control, Communications and Signal Processing.

Rasmussen, C. E. (2006). Gaussian processes for machine learning. In Gaus-

sian Processes for Machine Learning.

Rifkin, R., Yeo, G., & Poggio, T. (2003). Regularized least-squares classifica-
tion. Nato Science Series Sub Series III Computer and Systems Sciences,
190 , 131–154.

31
Rosasco, L., De Vito, E., Caponnetto, A., Piana, M., & Verri, A. (2004). Are
loss functions all the same? Neural Computation, 16 , 1063–1076.

Rosenblatt, F. (1958). The perceptron: a probabilistic model for information

storage and organization in the brain. Psychological review , 65 , 386.

Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1988). Learning repre-

sentations by back-propagating errors. Cognitive modeling, 5 , 1.

Scholkopf, B. (2001). The kernel trick for distances. In Advances in neural

information processing systems.

Shawe-Taylor, J., & Cristianini, N. (2004). Kernel methods for pattern anal-
ysis. Cambridge university press.

Shin, K., Lee, T. S., & Kim, H. (2005). An application of support vector ma-
chines in bankruptcy prediction model. Expert Systems with Applications,
28 , 127–135.

Simani, S., Fantuzzi, C., & Patton, R. J. (2003). Model-based Fault Diagno-
sis in Dynamic Systems Using Identification Techniques. Springer-Verlag
London.

Singer, R. M., Gross, K. C., & King, R. W. (1995). A pattern-recognition-

based, fault-tolerant monitoring and diagnostic technique. In Argonne
National Lab., Idaho Falls, ID (United States).

Smith, T. W. P., O’Keeffe, E., Aldous, L., & Agnolucci, P. (2013). As-
sessment of shipping’s efficiency using satellite ais data. In International
Council on Clean Transportation.

Spectec (2017). Amos maintenance-materials-management. http://www.

spectec.net/maintenance-materials-management.

Tarabrin, A. P., Schurovsky, V. A., Bodrov, A. I., & Stalder, J. P. (1998). An

analysis of axial compressor fouling and a blade cleaning method. Journal
of Turbomachinery, 120 , 256–261.

Vapnik, V. N. (1998). Statistical learning theory. Wiley New York.

32
Waeyenbergh, G., & Pintelon, L. (2002). A framework for maintenance con-
cept development. International journal of production economics, 77 , 299–
313.

Wainberg, M., Alipanahi, B., & Frey, B. J. (2016). Are random forests truly
the best classifiers? The Journal of Machine Learning Research, 17 , 1–5.

Wan, B., Nishikswa, E., & Uchida, M. (2002). The experiment and numerical
calculation of propeller performance with surface roughness effects. Journal
of the Kansai Society of Naval Architects, Japan, 2002 , 49–54.

Wang, H., Osen, O. L., Li, G., Li, W., Dai, H., & Zeng, W. (2015). Big data
and industrial internet of things for the maritime industry in northwestern
norway. In IEEE Region 10 Conference TENCON .

Widodo, A., & Yang, B. (2007). Support vector machine in machine condition
monitoring and fault diagnosis. Mechanical systems and signal processing,
21 , 2560–2574.

Witten, I. H., Frank, E., Hall, M. A., & Pal, C. J. (2016). Data Mining:
Practical machine learning tools and techniques. Morgan Kaufmann.

Yoon, H., Yang, K., & Shahabi, C. (2005). Feature subset selection and fea-
ture ranking for multivariate time series. IEEE transactions on knowledge
and data engineering, 17 , 1186–1198.

Zeileis, A., Hornik, K., Smola, A., & Karatzoglou, A. (2004). kernlab-an s4
package for kernel methods in r. Journal of statistical software, 11 , 1–20.

33