Você está na página 1de 15

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/324835430

Fault Detection and Diagnosis of Two-Stroke Low-Speed Marine Engine with


Machine Learning Algorithms

Conference Paper · April 2018

CITATIONS READS

0 627

5 authors, including:

Dimitrios A. Papachristos Nikitas Nikitakos


TEI Piraeus University of the Aegean
30 PUBLICATIONS   87 CITATIONS    129 PUBLICATIONS   238 CITATIONS   

SEE PROFILE SEE PROFILE

Dalaklis Dimitrios A.I. Ölçer


World Maritime University World Maritime University
128 PUBLICATIONS   92 CITATIONS    66 PUBLICATIONS   461 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Additive manufacturing (3D printing) in the maritime industry View project

MARPART(2) MAN View project

All content following this page was uploaded by Dalaklis Dimitrios on 30 April 2018.

The user has requested enhancement of the downloaded file.


Fault Detection and Diagnosis of Two-Stroke Low-Speed Marine Engine
with Machine Learning Algorithms

G. Tsaganos *, D. Papachristos **, N. Nikitakos**, D. Dalaklis***, A. I. Ölcer***

* Merchant Marine Academy of Athens, Greece, tsaganos.aen@gmail.com


** University of the Aegean, Chios, Greece, dimpap2009@gmail.com, nnik@aegean.gr
*** World Maritime University, Malmo, Sweden, dd@wmu.se, aio@wmu.se

Abstract
Detection of marine engine faults is extremely important for the optimized operation of the ship and
ensuring profitability; improving the vessel’s overall environmental impact is also part of the same
equation. Early detection of these faults can significantly reduce incidents of engine breakdowns and
improve reliability, since timely resolving of these faults can ensure the non-interrupted ship
operation. The analysis in hand describes and evaluates the development/implementation of intelligent
diagnostic methods based on the use of machine learning algorithms, allowing the effective detection
and diagnosis of faults that are associated with a two-stroke slow-speed marine diesel engine. The
research was implemented with the free Weka data mining tool, which analyzes the data of the
operating parameters of the engine that are out of the prescribed bounds. The specific classification
technique predicts with accuracy the objective class of each case. This proposed method is based on
the construction of an ensemble classification model AdaBoost, which improves the performance of
the Simple Cart classifier. The resulted rank with the highest performance in the detection of faults
after experimental diagnostics equals to 96.5%. Consequently, the proposed method is feasible and
very appropriate for the detection and diagnosis of faults of two-stroke low-speed marine diesel
engines.

Keywords: Classification Algorithms, Weka, Cross-Validation, F-Measure, Accuracy, Confusion


Matrix, Ensemble Method

1. Introduction
For a rather extended period of time, the main objective of the developers of marine propulsion
engines was to increase their efficiency and reliability. The need for larger propellers, along with the
associated capacity limitations and low rotation rates to increase hydrodynamic efficiency, have all
together led to the prevalence of two-stroke low-speed engines for the propulsion of merchant ships.
Two-stroke engines have a very high degree of efficiency, as well as the ability to burn low-grade
heavy fuel. It is therefore not a coincidence that in recent years almost all new ships buildings over
2000dtw were equipped with diesel propulsion engines, the vast majority of which were two-stroke
low-speed ones (Kyrtatos, 1999) .
Engine control and reliability are both equally important for normal ship operations; special trained
engineers and researchers/naval architects are always involved in the respective engine development
efforts, as well as their repair and maintenance. The combination of the above mentioned facts has
provided a strong incentive for those involved with engine design and development to take advantage
of innovative technologies that improve the performance outcome. Modern information technology
(IT) applications can facilitate detection and fault diagnosis techniques; the terms artificial
intelligence, machine learning, fuzzy logic are just a few of those terms that were envisioned in the
recent past to offer optimal utilization of the engine for ship propulsion in all operating conditions
(Lan, Katagi, & Hashimoto, 1996) (Xiros & Kyrtatos, 2000). It is indicative that the authors of the
analysis at hand have very recently put forward a proposed IT application in order to facilitate
Monitoring-Reporting-Verification (MRV) anomaly detection with the aim to early identify and
correct the respective arising risks during ship’s operations (Nikitakos et al., 2018)
In any case, a diagnostic system is responsible for controlling engine operating parameters in order to
identify engine variations that are associated with specific failures; the use of an IT based system can
achieve both the reduction of human involvement (therefore reducing the probability of human error),
as well as to contribute into an increased safety level. It is indicative that timely fault detection can
help to avoid damage to the engine that can lead to economic loss or a poor environmental
performance. As a result, new IT techniques and electronic equipment are increasingly being sought to
create diagnostics systems suitable for engine fault detection and diagnosis, with reliability and
effectiveness in the focus of attention. In the current analysis, an optimal method of using a machine
learning algorithm is proposed, which provides a very high percentage of accuracy in the detection and
diagnosis of the faults. This is achieved through:
test of various machine learning algorithms and the comparison of the results obtained
(percentage of correct predictions, performance metrics and model construction time), and
implementation of ensemble methods to enhance and improve the results.

2. Background

2.1. Two-Stroke Low-Speed Marine Engine


The “Low-Speed Marine Engine” literally dominates in the field of propulsion for large seagoing
commercial vessels (tankers, container ships, bulk carriers etc.). These engines have prevailed in the
shipping market due to their high efficiency in high-power mode (> 50%), and tolerance to fuel
quality. They have a large stroke ratio - piston diameter and Rod layout with zygoma-thrust Rod-
crankshaft. The motors are directly connected to the propeller without a clutch, thanks to low rotation
speed (Lazarou, X., Ioannis, Κ., Kliani, IA., 2003)
Today, there are only three sluggish Diesel engine manufacturers that are involved with two-stroke
engines: BW MAN, Mitsubishi, and Wartsila NSD. These manufacturers offer different “families” of
a two-stroke low-speed diesel engine. No further technical details will be provided, but in figure 1 that
follows, the intersection of a two-stroke low-speed marine engine ΜΑΝ Β&W 7S60MC is presented.

Fig. 1. Intersection two-stroke low-speed marine engine ΜΑΝ Β&W 7S60MC (Source:Engine Selection Guide
Two-stroke MC/MC-C Engines, 2000)

2.1.1. Fault Diagnosis of Two-Stroke Marine Diesel Engines


The failure of an engine concerns the deviation of an operating parameter from an acceptable limit
value (prescribed). The main causes of damage to engines that equip sea-going vessels are mainly:
corrosion,
pollution,
failure due to fatigue or mechanical / thermal stresses,
damage causing the change in dimensions,
decomposition,
and permanent deformation of the surface of parts of the engine.
The above causes of damage can occur throughout the life of the engine. When the faults extend to a
significant level that is about to affect the operation of the engine, then the diagnosis should provide
both qualitative and quantitative identification (Margaronis, 1986). It is a rather self-explanatory fact
that “Fault Detection and Diagnosis” is an important field for the marine engine industry. A fault
diagnosis can be made with the use of sensors (and supervisory devices) to collect/assess the required
information about the status of an engine. Early detection and fault diagnosis can drastically contribute
to avoid the occurrence of abnormal events and prevent damage because of extensive wear-out, or
even a failure (Skountrianos, 2005). With the complexity of new engines, the diagnosis addresses the
issue: ‘If there is a problem, how will the (looming) damage be spotted’ (Margaronis, 1986).
Thus, the diagnostic process must include the ability to locate malfunctioning operating parameters
and at the same time provide the ability to recognize the existence of such malfunctions that occur in
the state of diesel engine parameters, which may be due to specific faults or general wear because of
extensive use (Lazarou, Ioannis, & Kliani, 2003). The fault diagnosis, can be divided into three
distinct phases (Margaronis, 1986):
measurement of structural and operating parameters,
collection of data for comparison with previous data from the same engine, and
identification of problems and causes that are causing them.
The main element of the diesel engine diagnostic test is that the previous stages/phases must be based
on information retrieved by the engine in normal operation, without for example permitting any
disassembly or another kind of direct access to the inside of the engine and by using only external
information/observations. The performance of a diagnostic system can be evaluated based on the
following parameters (Tselenti, 1998):
accuracy (number of correct answers),
diagnostic time,
number of successful diagnoses, and
a number of incorrect diagnoses.
Most diesel engine malfunctions are associated with the combustion process within the cylinders,
where a number of subsystems, such as camshaft valves, fuel injection system and pumps among
others work altogether. The gas pressure inside the cylinder is the most important information for
monitoring all the mechanisms and processes that occur within the combustion chamber of a Diesel
engine. Its measurement is the first point of concern for the majority of engine monitoring and fault
diagnosis methods. The processing of cylinder pressure values is particularly important as it can
provide the engineer with very useful information about engine power, heat output rate, ignition angle,
combustion duration and compression quality. Therefore, the pressure inside the cylinder and
estimation of derivative sizes that result from treatment (such as high-pressure combustion, power
indicative, Middle indicative pressure etc.), as the principal oversight of the operation of a size engine,
formulate the basis of fault diagnosis. Indicative examples are already available: the MAN BW
Diesel's Off-line PMI (Pressure Management Indicator) system is designed to provide engineers and
shipboard technicians with a portable electronic tool for measuring and analysing cylinder pressures in
MAN B&W two-stroke diesel engines (“User’s Guide PMI System,” 2005).

2.2. Machine Learning


In Machine Learning (ML), the computer includes learning algorithms, aiming through automatic
learning, to learn a system, make accurate predictions, and intelligently make decisions with little or
no human intervention or help. Learning is always based on some kinds of observations in the data,
and an additional important role is fulfilled by the registered examples, experience and system training
(Schapire, 2013). ML is an effective empirical approach for both regression and/or classification
(supervised or unsupervised) of nonlinear systems. The ML-based methods have been widely applied
to science and engineering problems for near two decades (Lary et al. 2016).
ML involves five (5) steps of performing a task: data collection, data preparation, model training,
model evaluation and performance improvement. Learning is divided into two main categories:
Supervised Learning and Unsupervised Learning (Fig.2). Supervised learning algorithms learn from
“labelled” training data to produce a model that can make predictions on previously unseen data.
Supervised learning includes two categories of algorithms: Classification (SVM, neural networks,
Naïve Bayes classifiers, decision trees, discriminant analysis, and nearest neighbor-kNN), and
Regression (Linear/ nonlinear regression, generalized linear models, decision trees, and neural
networks). Furthermore, unsupervised learning includes: Cluster analysis (k-means), Hierarchical
clustering, and othe algorithms (Gaussian mixture models, hidden Markov models, and self-organizing
neural network maps) (Amozegar & Amozegar, 2015), (Ayubi Rad & Yazdanpanah, 2015), (Hu
Jinhai,Xie Shousheng,Cai Kailong,He Xiuran, 2007), (Li et al., 2012), (Sahin, Yavuz, Arnavut, &
Uluyol, 2007), (Sharkey, Chandroth, & Sharkey, 2000), (Twiddle & Jones, 2002), (Wong, Zhong,
Yang, & Vong, 2016).

Figure 2. The Basic Types of ML.


3. Research Objectives
The scopes of the research were based primarily:
In developing methods of data mining and the development of capacities of machine learning
algorithms to solve problems.
The findings of a series of researches into engine failure diagnoses, which conclude that
resolving engine fault diagnosis is one of the major problems faced by naval engineers, and
even more so in one of the world's most important engine categories, namely the two-stroke
low-speed marine diesel engine.
The purpose of the investigation is to create a timely (early) and reliable system fault diagnosis
utilizing the open source tool Weka. This diagnostic system could be a very useful tool for marine
engineers to conduct troubleshooting within the engine room.

4. Methodology
The research field involves mainly the differentiation of pressures and temperatures. Maintaining them
within the appropriate boundaries plays a key role in engine performance for manufacturers. This is
why when significant parameter deviations are detected, the reasons of their cause should then be
controlled. The estimated deviation for a functional parameter is calculated from the mean value of the
deviations resulting from the cylinder counts, and is the arithmetic average of the pressure
measurements of a particular cylinder. The calculation results in the estimated values of the
operational parameters of all the cases.

Table 1. Functional Parameters Groups & ΜΑΝ Β&W 7S60MC Engine Limits.

Group Engine parameters Construction Limits Classes/Categories


Characteristics per cylinder
All All Deviations Within Limits ΟΚ
power_c1 to Power (KW) Greater than 2000 KW Power
power_c7
rpm_c1 to rpm_c7 Speed Less than average cylinder RPM_low
Rotation per minute value -5%
(rpm)
rpm_c1 to rpm_c7 Speed Less than average cylinder RPM_very_low
Rounds per minute value -10%
(rpm)
pmi_c1 to pmi_c7 Average indicative Greater than average Pmi_high
pressure (bar) cylinder value + 1 bar
Greater than average Pmi_low
cylinder value - 1 bar
pmi_c1 to pmi_c7 Average indicative Greater than average Pmi_high
exhaust_gass_temp pressure (bar) & cylinder value Exhaust_gass_temp
_c1 to Exhaust gas + 1 bar & > 410 οC _high
exhaust_gass_temp temperature Less than average cylinder Pmi_low
ο
_c7 C diameter - 1 and Less than Exhaust_gass_temp
330 °C _low
pcomp_c1 to Pressure Greater than average Pcomp_high
pcomp_c7 compression (bar) cylinder value + 3 bar
Less than average cylinder Pcomp_low
value - 3 bar
pmax_pos_c1 to Maximum burning Greater than 20 degrees Pmax_pos
pmax_pos_c7 pressure location
(deg)
pmax_c_c1 to Max Combustion Greater than average Pmax_c_high
pmax_c_c7 Pressure (bar) cylinder value +2 bar
Greater than average Pmax_c_low
cylinder value -2 bar
ignition_c1 to Fuel Injection Greater than average Ignition_ angle _high
ignition_c7 Angle (deg) cylinder value +0,7 degrees
Less than average roller Ignition_ angle _low
-0,7 degrees
exhaust_gass_temp Exhaust gas Greater than 410 οC Exhaust_gass_temp
_c1 to temperature _high
ο
exhaust_gass_temp C Less than 330 οC Exhaust_gass_temp
_c7 _low

The methodology of the analysis in hand followed these consecutive steps:


Setting a sample set (MAN BW 7S60MC engine failure scenarios by using the engine
simulator of the Engineering School of AEN Aspropyrgos).
Collection and processing of data (use of MAN_7S60MC.arff fault file, including 1000 valid
fault records, through rules-based software, developed specifically for validation of data,
according to engine operating details as defined by the manufacturer, the standard database
was formatted and created).
Study of the characteristics of the chosen engineering algorithms to be tested.
Parameterization and execution of algorithms.
Shaping the relevant model.
A comparative study of the results of the algorithms in order to find the most suitable
algorithm per case and the corresponding settings of these parameters.
Optimization of algorithms through combinational methods in order to increase their accuracy.
Defining a final fault assignment model.
Each algorithm was trained using 1000 snapshots. Verification of the model for estimating the
performance of the methodology was done by applying the 10-fold cross-validation procedure. That is,
out of the total of 1000 snapshots, 100 screenshots were used for testing and the rest of training.
The choice of cross-validation was based on the fact that it minimizes the problem of over-fitting,
improves the prediction and reduces the variance of the estimate between the data. Extensive
experiments have shown that this is the best choice for an accurate estimate (Ross et al., 2009).
Finally, a series of algorithms were tested in the WEKA environment to select the best algorithms that
matched the problem. The following algorithms were used for each classification class:
The Naive Bayes as representative of Bayesian Networks
The J48 (C 4.5) and Simple Cart as representatives of Decision Trees
The LWL as representative of Lazy algorithms
The MultilayerPerceptron as representative of the Neural Network
The SMO as representative of Support Vector Machines
The MODLEN as representative of classification rules
The AdaBoost, MultiBoost and Decorate as representatives of meta-algorithms.
Below are listed the criteria that measure the performance of the classification system, where they
evaluate the quality and accuracy of the classification.
The use performance metrics are “Accuracy” and “F-Measure”.
Accuracy is used as a statistical measure of how well a test in a classification correctly determines or
excludes a situation.
In the current analysis, multi-class outputs and metrics (such as accuracy) lose their credibility because
sometimes they can be misleading; for example a model with relatively ‘high’ accuracy might be
available, but the specific model can also perform all kinds of mistakes for classes that are really
critical to the problem.
For this reason, the function F (F-Measure) (Singh Sabharwal, n.d.), which is the weighted average of
the Precision and Recall metrics, was used to evaluate the system.
2 precision recall
F_Measure
precision recall
For each fault category, calculated the metrics and then the average using weights (Sokolova &
Lapalme, 2009).

5. Presentation of Result
In the current analysis, seven (7) basic algorithms were used from all classes of classification, such as
NaiveBayes of Bayesian Networks, Multilayer Perceptron of Neural Networks, SMO of Vector
Support Machines, LWL of Lazy Algorithms, MODLEM of Algorithms rule creation, and J48, and
Simple Cart of decision trees. In addition, an attempt has been made to improve the performance of
the basic algorithms by the AdaBoost, MultiBoost and Decorate ensemble methods. All algorithms
were implemented using the Weka data mining tool to analyze their accuracy and performance. Also,
the performance metrics, F-Measure, Accuracy and the time required to build the model were used to
evaluate the efficiency of each algorithm. The algorithms were then compared and evaluated
according to the above performance metrics. In addition, a comparison was made between ensemble
methods and basic algorithms to examine the usefulness of ensemble methods to improve the
performance of basic algorithms in engine fault diagnosis.
The following table (Table 2) provides in an adequate level of detail the comparisons of the basic
algorithms with the metric performance F-Measure, the Accuracy predictive precision statistical
measure, and the model's construction time. It was identified that the most efficient and accurate in its
predictions is the Simple Cart algorithm with relatively little time to construct its model. The worst
performance of the algorithms selected in the survey is MultiLayer Perceptron. Also, the SMO
algorithm did not show satisfactory prediction performance and accuracy. On the other hand, the
Naïve Bayes, LWL, MODLEM and J48 algorithms showed almost the same performance and
precision, and a short time to build their models.
By comparing the ensemble methods it was discovered that the most accurate predictions of damage
and a higher performance measure have AdaBoost with a basic classifier Simple Cart, improving
Simple Cart algorithm performance by 1.1%. The improvement, however, results in much larger
construction time of the model. Second most effective in forecasting accuracy rate and in F-Measure
performance measure is the MultiBoost with the Simple Cart base classifier, improving the
performance of simple Simple Cart by 0.5% with also the big-time construction of the model. The J48
basic classifier was enhanced and yielded satisfactorily with the AdaBoost, MultiBoost and Decorate
ensemble methods, significantly increase its F-Measure performance and Accuracy rate by 7%. The
least effective is MultiBoost, with a basic MODLEM classifier, achieving little improvement.
Table 2. Comparison and Ranking of Research Algorithms.
Classification Algorithms F-Measure Accuracy % Time of Model
Construction (sec)

Naïve Bayes 0.841 86,1 0,05


Multilayer Perceptron 0.530 54,4 333,53

SMO 0.747 75,5 1,34


LWL 0.890 90,1 0
MODLEM 0.905 90,4 0,56
J48 0.881 89,1 0,58
Simple Cart 0.955 95,5 5,73
AdaBoost J48 0.950 95,1 24,41
AdaBoost SimpleCart 0.965 96,6 238,99
MultiBoost MODLEM 0.919 92,5 13,73
MultiBoost J48 0.951 95,2 16,3
MultiBoost SimpleCart 0.959 96 224,86
Decorate J48 0.950 95,2 84,94

Finally, the comparative performance improvement results for the basic classification algorithms
MODLEM, J48 and Simple Cart with the combination methods AdaBoost, MultiBoost and Decorate
are shown in the following graphs, where in two separate axes on the first axis a blue column is
displayed, Figure 3, 4, 5, the performance metric F-Measure, the second axis shows the percentage of
the correct Accuracy snapshots in green.

Fig. 3. Comparative improvement of performance metrics of the basic algorithm J48 with the ensemble methods
AdaBoost, MultiBoost and Decorate.
Fig. 4. Comparative improvement of performance metrics of the basic algorithm Simple Cart with the ensemble
methods AdaBoost and MultiBoost.

Fig. 5. Comparative improvement of performance metrics of the basic algorithm MODLEM with the ensemble
method MultiBoost.

The classification system performance is given through a table called the “confusion matrix”. In the
confusion matrix Table 3, the results of the AdaBoostM1 algorithm are presented with a basic Simple
Cart classifier, analysing the number of correct and incorrect predictions of the algorithm by fault
class, the main diagonal of the table shows the number of correct predictions in the fault detection for
each fault class, and the number of incorrect predictions of fault outside the main diagonal.
The total sum in the column of each fault category (except for the main diagonal) is the Failure
Positive Predictions (FP), i.e. there appears to be an indication of failure to that category of failure
while it is not actually present. While the total sum in the line of each fault class (excluding the main
diagonal) is the Failure Negative Predictions (FN), i.e. there appears to be no indication of damage to
that particular category of failure while in fact there is.
Table 3. Confusion Matrix of the AdaBoostM1 algorithm with a Simple Cart classifier.

6. Conclusions
Detection of marine engine faults is extremely important for the optimized operation of the ship and
ensuring profitability. The diagnostic methods presented in this research effort have shown that they
can provide reliable diagnostic tools. Summarizing the findings of the research, the following stand
out:
Simple Cart, which belongs to the regression trees, is used to the basic algorithms, which are
used when the dependent values such as the data of the present work continuous, therefore it
holds an advantage over with respect to the J48 which belongs to the classification trees and is
more suitable when the dependent values are categorical.
The model manufactured by MultiLayer Perceptron had poor performance and this is because
of the relatively small amount of data it had available for its training, and the SMO did not
perform satisfactorily, since the classification categories were 17 much more than two, and
this differs from the results of relevant research that bibliographic research has highlighted,
since the type and range of data has an important role in the results.
In Combinatorial methods, thanks to the advantages of the method 'Boosting', the improved
prediction performance of basic algorithms, detects much more efficiently the engine fault
categories, from what can be achieves each algorithm separately, extracting more precise and
more reliable diagnostic conclusions.
Therefore, the proposed algorithm showed clear superiority in the experimental process of fault
detection and diagnosis based on the use of ensemble method AdaBoost, with the Simple Cart decision
tree, which was superior to the other algorithms, giving the best predictive performance F- Measure
0.965, but at the cost of increasing the time needed to build the model.
The result of this research, indicates originally how useful are the techniques of machine learning
algorithms, and secondarily the possibilities provided by the data mining tool Weka, in the analysis of
operating parameters for the application of the diagnostic system for detection and fault diagnosis on
two-stroke slow speed marine diesel engines. Early detection of these faults can significantly reduce
incidents of engine breakdowns and improve reliability, since timely resolving of these faults can
ensure the non-interrupted ship operation.

References
Amozegar, M., & Amozegar, M. (2015). Aircraft Jet Engine Health Monitoring Through System Identification
Using Ensemble Neural Networks. Retrieved from http://spectrum.library.concordia.ca/980252/

Ayubi Rad, M. A., & Yazdanpanah, M. J. (2015). Designing supervised local neural network classifiers based
on EM clustering for fault diagnosis of Tennessee Eastman process. Chemometrics and Intelligent Laboratory
Systems, 146, 149–157. https://doi.org/10.1016/j.chemolab.2015.05.013

Engine Selection Guide Two-stroke MC/MC-C Engines. (2000) (5th ed.). Retrieved from
https://www.fsb.unizg.hr/ship-design/esg.pdf

Hu Jinhai, Xie Shousheng, Cai Kailong, He Xiuran, P. J. (2007). Classification Method of DivClassification
Method of Diverse AdaBoost-SVM and Its Application to Fault Diagnosis of Aeroengine. Retrieved from
http://en.cnki.com.cn/Article_en/CJFDTOTAL-HKXB200705013.htm

Kyrtatos, Ν. (1999). Important Research and Development for mariners, Diesel Engines, (Ifo 380), 1–12. (in
Greek).

Lan, W. C., Katagi, T., & Hashimoto, T. (1996). Quasi Steady State Simulation of Diesel Engine Transient
Performance and Design of Mechatronic Governor*. Retrieved from
http://www.jime.jp/e/publication/bulletin/english/pdf/mv24n011996p01.pdf

Lazarou, X., Ioannis, Κ., Kliani, I.A. (2003). Internal Combustion Engines Volume Two. (in Greek).

Li, Z., Yan, X., Guo, Z., Zhang, Y., Yuan, C., & Peng, Z. (2012). Condition Monitoring and Fault Diagnosis
for Marine Diesel Engines using Information Fusion Techniques. Electronics and Electrical Engineering, 123(7),
109–112. https://doi.org/10.5755/j01.eee.123.7.2387

Margaronis, Ι. E. (1986). Function Diagnostic Systems in Marine Diesel Engines. Doctoral thesis, National
Technical University of Athens. (in Greek).

Nikitakos, N., Dalaklis, D., & Siousiouras, P., (2018), Real Time Awareness for MRV Data, in Ölçer, A.I.,
Kitada, M., Dalaklis, D., Ballini, F. (Eds.), Trends and Challenges in Maritime Energy Management, Springer-
Cham (ISBN 978-3-319-74576-3).
Ross, K. A., Jensen, C. S., Snodgrass, R., Dyreson, C. E., Jensen, C. S., Snodgrass, R., … Chen, L. (2009).
Cross-Validation. In Encyclopedia of Database Systems (pp. 532–538). Boston, MA: Springer US.
https://doi.org/10.1007/978-0-387-39940-9_565

Sahin, F., Yavuz, M. Ç., Arnavut, Z., & Uluyol, Ö. (2007). Fault diagnosis for airplane engines using
Bayesian networks and distributed particle swarm optimization. Parallel Computing, 33(2), 124–143.
https://doi.org/10.1016/j.parco.2006.11.005

Schapire, R. (2013). Theoretical Machine Learning, 1–7.

Sharkey, A. J. C., Chandroth, G. O., & Sharkey, N. E. (2000). A Multi-Net System for the Fault Diagnosis of
a Diesel Engine. Neural Computing & Applications, 9(2), 152–160. https://doi.org/10.1007/s005210070026
Skountrianos, H. (2005). Modeling and Diagnosis-Recognition of Non-Linear Dynamic Systems Faults with
Neural Networks of Local Models. Doctoral thesis, National Technical University of Athens. (in Greek).

Sokolova, M., & Lapalme, G. (2009). A systematic analysis of performance measures for classification tasks.
Information Processing and Management, 45, 427–437. https://doi.org/10.1016/j.ipm.2009.03.002

Tselenti, G. N. (1998). Fault diagnosis through vibration control in industrial line Laundry production.
Technical University of Crete. (in Greek).

Twiddle, J. A., & Jones, N. B. (2002). A high-level technique for diesel engine combustion system condition
monitoring and fault diagnosis. Proceedings of the Institution of Mechanical Engineers, Part I: Journal of
Systems and Control Engineering, 216(2), 125–134. https://doi.org/10.1243/0959651021541499

User’s Guide PMI System. (2005). In MAN B&W Diesel A/S (2.3, p. 82). Retrieved from
http://marengine.com/ufiles/MAN-PMI_off.pdf

Wong, P. K., Zhong, J., Yang, Z., & Vong, C. M. (2016). Sparse Bayesian extreme learning committee
machine for engine simultaneous fault diagnosis. Neurocomputing, 174, 331–343.
https://doi.org/10.1016/j.neucom.2015.02.097

Xiros, N. I., & Kyrtatos, N. P. (2000). A neural predictor of propeller load demand for improved control of
diesel ship propulsion. In Proceedings of the 2000 IEEE International Symposium on Intelligent Control. Held
jointly with the 8th IEEE Mediterranean Conference on Control and Automation (Cat. No.00CH37147) (pp.
321–326). IEEE. https://doi.org/10.1109/ISIC.2000.882944

View publication stats

Você também pode gostar