385
1 INTRODUCTION
Recently,relevantadvancesinpredictivemaintenance
havebeendeveloped,sinceithasbeendemonstrated
that it is the most suitable and efficient method
(Muszynska, 2005; Wang et al., 2017; White, 2010).
This technique is based on the basis that machinery
will show unusual behavior before of failing. In
the
maritimeindustryisespeciallyuseful(Gkerekosetal.,
2017;Goetal.,2013;Jakovlevetal.,2017)owingtothe
factthatanyunexpectedfailureduringajourneycan
makeadanger.
One of the most interesting predictive
maintenance is which based on vibration analysis
giventhatitis
thenon‐destructiveteststhatprovides
the greatest amount of information about internal
functioning of a machine (Martini and Troncossi,
2016; White, 2010). Therefore, knowing its normal
signature of vibration, it is possible to prevent a
breakdownbymonitoring.Inthisway,vibrationsare
indicatorofa potential problem.Therefore, to
count
on indirect monitoring systems would be very
advantageous to preventively maintain aboard and
hence to avoid damages. Separation systems for
heavy fuel oils are considered to be complex
mechanical systems and their reparation is usually
difficult.Owingtoitscomplexity,awidemonitoring
ofitsbehaviorisessentialin
ordertodetectincipient
failures. Thus, a sensor that can measure vibrations
should be incorporated. The maintenance of these
systems is based on preventive maintenance
scheduledfromtheirworkinghours.Inotherwords,
when a certain number has been achieved, a
preventive scheduled maintenance is carried out.
Nevertheless,withamonitoring,
itcouldbepossible
toextendthenextmaintenanceserviceifthesystemis
healthy or make it scheduled ahead of time if the
separatorsystemshowsfailureindications.Forthis,it
is essential to design a technique in order to
conveniently extract and process this amount of
information.
A New Intelligent Approach in Predictive Maintenance
of Separation System
G.N.Marichal,D.Ávila,A.Hernández&I.Padrón
UniversityofLaLaguna,SanCristóbaldeLaLaguna,Tenerife,Spain
ABSTRACT:Reducingcontaminantemissionsisanimportanttaskofanyindustry,includedthemaritimeone.
Infact,inApril2018,IMO(InternationalMaritimeOrganization)adoptedanInitialStrategyonreductionof
Greenhousegas(GHG)emissionsfromships.Anessentialpart
responsibleforproducingtheseemissionsisthe
dieselengine.Forthatreasonvesselsincludeseparationsystemsforheavyfueloils.Thepurposeofthisworkis
to improve the predictive maintenance techniques incorporating new intelligent approaches. An analysis of
vibrationsofthisseparationsystemwasmadeandtheircharacteristicswere
usedinaGeneticNeuro‐Fuzzy
Systeminordertodesignanintelligentmaintenancebasedonconditionmonitoring.Theachievedresultsshow
thattheproposedmethodprovidesanimprovementsinceitindicatesifamaintenanceoperationisnecessary
beforethescheduleoneorifitcouldbepossibleextendthenext
maintenanceservice.
http://www.transnav.eu
the International Journal
on Marine Navigation
and Safety of Sea Transportation
Volume 14
Number 2
June 2020
DOI:10.12716/1001.14.02.15
386
The application of artificial intelligence (AI) in
machine diagnosis has been widely investigated in
different fields (M. Samhouri , A. Al‐Ghandoor , S.
AlhajAli,I.Hinti,2009;Simanietal.,2003).Recently,
thereisa tendencyforthe useofgenetic algorithms
(Baojia et al., 2018; Cerrada
et al., 2015; Gou et al.,
2018;Heetal.,2017).Thepurposeofthisworkisto
combine predictive maintenance based on condition
monitoring, with a genetic neuro‐fuzzy system. The
proposedintelligentalgorithmintendstoprocessthe
measuredvibrationsandprovidesinformationabout
theinternalstateofthe
heavyfueloilseparators.For
that,severalactualmeasurementswerecarriedouton
boardseparatorsofRo‐Paxvessel.
In Section 2 a description of the heavy fuel oil
separation system used in the actual experimental
tests and an explanation of the measurement
procedure are explained. In Section 3 the signal
processing method based on artificial intelligence
techniquesisexposed.Section4showstheresultsand
finallyinSection5conclusionsarepresented.
2 EXPERIMENTALSTUDY
With the purpose of collecting vibration data from
separators,severalactualmeasurementswerecarried
outonboardcentrifugalheavyfueloilseparatorsof
Ro‐Pax
vessels to transport passengers and freight.
Figure1showsthemodelAlfaLavalSA861separator
usedinthiswork,whichtechnicalcharacteristicsare
exposedinTable1.
Figure1.MarineFuelSeparatorsAlfaLavalSA861.
As it was mentioned previously, vibration
measurementswereusedinordertodetectchangesin
thebehaviorofmachinery.Forthisreason,atriaxial
accelerometer sensor, Bruel & Kjaer 4504A (Bruel &
Kjaer, Naerum,Denmark) was attached to separator
system. In each separator ten consecutive vibration
measurements were made for 2
seconds with a
samplingfrequency(Fs)of2560Hz.Theusedsensor
hasthreeindependentoutputsforsimultaneoushigh‐
level measurements inthree mutually perpendicular
directions. The sensitivity of the accelerometer is 10
mV/g,anditmeasuresarangeofupto9.0KHz,and
its lower bound of the
sensing frequency range is 1
Hz.Theorientationoftheaccelerometerisasshown
in Figure 2: the X axis is vertical, the Y axis goes
throughfromrighttoleft,andtheZaxistraversesthe
device from front to back. Figure 2 shows the final
assemblywhereitcan
beseenthattheaccelerometer
wasconnectedtoaBrüel&KjaerPHOTON+dynamic
signal analyzer and this one to a laptop where the
measurements were recorded [30]. The PHOTON+
consistsofdataacquisitionhardwareandPCsoftware
able to measure, to record, to analyze and to post‐
process.Itallowsfor
real‐timesignalanalysis.Infact,
it could be used as a FFT analyzer with a
measurementdynamicrangeof115dBandan84kHz
real‐timerate.
Table1.TechnicalcharacteristicsofMarineFuelSeparators
AlfaLavalSA861.
_______________________________________________
CharacteristicsValues
_______________________________________________
BelttransmissionUFT‐21
Electricalcurrentfrequency50HZ
Motorpower(50Hz)18.5kW
Motorspeedsynchronous(50Hz) 3000rpm
Max.densityoffeed/sediment1100/2057
kg/m
3
Max.densityofoperatingliquid
1000kg/m
3
Feedtemperature,min./max.0°Cto100°C
Max.viscosityofoperatingliquid 700cStat50°C.
_______________________________________________
Figure2. Final assembly of accelerometer and PHOTON+
analyzer.
3 INTELLIGENTAPPROACH
Themaintenanceofmajorityofmachineriesonboard
is based on the number of hours. Particularly, the
separator systems are revised when they has been
working 12000h. This kind of maintenance is not
efficient,since itisoften possible to keepthe device
working as long as itdoes
notshow sign of failure.
387
On the other hand, it is also possible that a
maintenance operation should be done before a
scheduled one because of a failing piece. For this
reason, this work tries to find an efficiency‐based
maintenance method for the oil separators through
intelligent condition monitoring. The number of
working hours
has been considered as a key
parameter in this work. In this sense, it would be
convenient to associate this factor with the internal
state of the system, that in this paper it is studied
throughthevibrationsignature.
In order to collect data, real vibrations were
measured over on
board separators systems. These
recorded vibrations are analyzed by a signal
processingstage withthe purposeof obtaining their
internal characteristic parameters. In this research, a
FFTisappliedinordertogetthefrequencydomainof
the separator vibrations. The first five frequencies
with greater amplitude were collected, along with
their corresponding amplitudes as indirect
parameters.Withtheaimoflinkingthesepairofdata,
that is, the internal state of the separator, with the
number of working hours, an intelligent method
based on training was used. Specifically, a three
layers genetic neuro‐fuzzy system (Cordón et al.,
2004; Marichal
et al., 2016; Nobre, 1995), with an
analogous structure to the one proposed by Jang
(Jang,1993).Systeminputsareintroducedtothefirst
layer,whichrepresentthemembershipfunctions,and
theoutputsofthislayerareexpressedbyEquation1.
2
ij
1
ij
2
2
ij
i= 1,2, ..., N
φexp
j=1,2,...,N
i
Um
(1)
whereN
1=inputnumber;N2=numberofnodesofthe
intermediate layer; u
i= i‐th input, mij=center of the
membership function;σ
ij=the width of the
membershipfunction,φ
ij=outputneuronwiththei‐th
inputandtheoutputconnectedtothej‐thnodeofthe
intermediatelayer.
The second layer outputs correspond to the rule
system,anditisshowninEquation2.
1
12 2
ε min φ ,φ ,..., φ 1,...,
ijjNj
jN (2)
FinallyEquation3representstheglobaloutput,
2
2
jk
1
3
1
1,...,
N
j
j
k
N
j
j
s
YkN
(3)
where N
3= genetic neuro‐fuzzy output number;
sv
jk=estimatedvalueofthek‐thoutputgivenbythej‐
thnode.
AsEquations(1)–(3)show,thegeneticneuro‐fuzzy
system depends on the center and width of the
membership function, the estimated system outputs
and the number of nodes of the intermediate layer.
These parameters are fixed through a
three‐phase
learningalgorithm.Initialvaluesandanoptimization
of the number of nodes of the hidden layer are
establishedinthefirsttwophases.Thenatthethird
phasetheseparametersarereset.
3.1 FistLevel:UnsupervisedLearningPhase
Inthisfirstphase,theinitialvaluestom
ijandsvjkare
provided by a Kohonen’s self‐organizing [37] map,
wheretheirinputsare:
13
12 12NN
VUUUYYY (4)
Thevector(U
1U2…UN1)istheinputvectortothe
genetic neuro‐fuzzy system, and (Y
1 Y2…YN3) is the
desired output vector. Equation (5) shows the
necessary initial weight vector of the self‐organizing
map and it is acquired by the mean between the
maximumandminimumoftheinputsetbytheuser.
13
123 2
1, 2, ,
jjjjNNj
Wwwww j N
(5)
An update of the weights is achieved after a
monodimensional Kohonen self‐organizing map is
applied,sinceitprovidethewinnernode.
2
; 1, 2,..., ; 1, 2
ij ij
mwj Ni
(6)
This algorithm is an unsupervised learning
algorithm; therefore, once the process has been
completed and the winner node has been obtained,
thecenterofthemembershipfunctions(m
ij)isfixed,
and then the estimated system outputs (sv
jk) will be
carriedout.
This phase is crucial, since the initial assignment
establishes the starting point. In the following
learning phases, these parameters will be modified
fromtheinitialones.
3.2 SecondLevel:TheGeneticAlgorithmPhase
In this phase a genetic neuro‐fuzzy system is built,
since in the
previous one N2, mij, and sνjk were
obtained, but then there are still values for the
parametersσ
ij missing. Moreover, an optimization
process is necessary in order to obtain a minimum
number of rules. Then, in this phase will be
accomplish a more reduced number of nodes in the
hiddenlayer.
Thegeneticalgorithm(Cordónetal.,2004;Nobre,
1995;RajasekaranandPai,2003)usedinthisphase
is
based on the biological model of genetic evolution.
On the one hand, there is an individual with basic
information,particularlyavector;ontheotherhand,
there are genes, in this work they are the vector
components. Therefore, the components of each
vector represent the hidden nodes by a Boolean
parameter and the width of the membership
functions. Following, a fitness function is defined,
taking into account the difference between the real
outputsandtheindividualoutputs.Oncethegenetic
neuro‐fuzzy algorithm is applied, individual
388
satisfactoryvalues forσ
ij, andan adequateset of N2
rules(nodesonthehiddenlayer)canbereached.
3.3 ThirdLevel:SupervisedLearningPhase
Thelastphaseattemptstoimprovetheinitialvalues
forthem
ij,σij,andsvjkparameters.Owingtothefact
thatthesystemusedinthisworkissimilartoathree
layer neural network, the same mathematical
expression as the neurons in a radial basis neural
network(Chenetal.,1991)hasbeenusedtoexpress
the nodes on the input layer of the
genetic neuro‐
fuzzy system. Furthermore, the least mean squared
learningalgorithmhasbeenalsoapplied.Finally,the
criterion function (Equation 8), defined as the error
function between the outputs of the genetic neuro‐
fuzzy system (ψ
k) and the real outputs (Yk), is
intendedtominimize.
3
N
2
kk
k1
1
EYψ
2
(8)
4 RESULTS
As it was mentioned in previous sections, vibration
signals were collected in the vessel and a signal
processing was carried out in order to extract the
information about the state of the system. In this
work, a traditional FFT was applied and the five
dominant frequencies were
obtained in each
measurement. An example of a vibration
measurement is shown in Figure 3. This graph
displaysavibrationmeasuredontheYaxis.
Figure3.AnexampleofvibrationsignalontheYaxis.
Figure 4 displays the corresponding frequency
spectrum of previous vibration. At this point is
necessary to remark that every measurement
includinginthisworkwascarriedoutinarealvessel
inthemiddleofactualjourneys.
Themainpurposeofthisworkistoreachthatthe
exposed intelligent method
allows relating the
vibration signature with the corresponding working
hours.Forthatreason,thefivedominantfrequencies
with theircorresponding amplitude FFTs were used
as input vectors to the genetic neuro‐fuzzy system.
Theoutputintelligentsystemwouldbethenumberof
hours of the oil separator had been running
at the
moment that vibration signal was measured. With
this input‐output dataset, oncethe systemhasbeen
trained,itwillbeabletoprovidethenumberofhours
thatithadbeenworking.
Figure4.FrequencyspectrumofthesignalshowninFigure
3.
Several trials were carried out with the input–
output data set in each training phase previously
explained, in order to obtain the parameters that
provide an satisfactory error value. The training
process has been developed with 70% of the data,
since the other 30% was reserved to check the
generalization capability
of the algorithm. If the
genetic neuro‐fuzzy system provides adequate
outputstounknowninputvalues(datathat had not
been used in the training process), then a suitable
levelofgeneralizationhasbeenreached.
Afteralltrainingphaseshavebeenconcluded,one
with a minimum error function is chosen.
Figure 5
shows the results of the genetic algorithm phase.
After12generations,thebestfitnessvaluewasquite
similar to the mean, and after 48 generations, the
average value between individuals was zero; this
means that the genetic neuro‐fuzzy system has
achievedgoodtraining.
389
Figure5. Evolution of the fitness value and the average
distancesbetweenindividuals.
Moreover, if more particular results were
evaluated, satisfactory parameters were obtained as
well. In this case, the genetic neuro‐fuzzy system
reached a training error of 2277 h, and a
generalization error of 1872.8 h. Analyzing these
error,itcouldbesaidthatareacceptable,sincetheon
boardheavyfuel
oilseparatorsusedforthisworkhad
beenworkingbetween3196and12236hours.
Asafinalpoint,a comparativetable is presented
inordertoensurethatthegeneticalgorithmreacheda
suitable generalization level. Table 2 shows a
comparison between the real value of the working
hours corresponding
to a separator system and the
value provided by the trained genetic neuro fuzzy
system. It is essential to highlight that the input
valuesusedforgeneratingthistableareunknownfor
the system, that is, they were not included to the
training.
Table2. Comparison between outputs provided by the
genetic neuro‐fuzzy system, and the real values of the
workinghours.
_______________________________________________
RealWorkingHoursWorkingHoursProvided
byGeneticNeuroFuzzy
_______________________________________________
3196 2776
3090 4338
3078 4231
6300 6411
12236 10145
5621 4725
6476
4996
_______________________________________________
5 CONCLUSIONS
The exposed research presents an intelligent
condition‐based maintenance for heavy fuel oil
separators based on genetic neuro‐fuzzy system.
Vibration measurements were carried out on board
real vessels with the purpose of to relating these
vibrations and the internal state of the separation
systems.Thecollecteddata
wereprocessedinorderto
obtaintheircharacteristicparameters.Inthiswork,a
Fast Fourier Transform allowed for extracting the
frequency domains of the separator vibrations and
their corresponding amplitudes. This packet of data
wasusedasinputsetforthetrainingalgorithm.Each
input vector was fixed with the
number of working
hours that each fuel oil separator had been running
foruntilthemeasurementmoment.
Once the training process has been finished, it is
possibletoconcludethatthereisavibrationsignature
capable of providing useful information in order to
preventing damages. This is because of there is
a
relationshipbetweenvibrationandtheinternalstate,
and, therefore, a trained system can indicate the
numberofworkinghoursthatthesystemhavebeen
runningfor.Thefactthatamonitoringandatrained
system are included presents an advance over the
traditional preventive scheduled maintenance.
Whereas the preventive
maintenance is carried out
when a certain number of hours has been achieved,
the proposed method can indicate whether it is
possibletoextendthenextmaintenanceserviceifthe
separator is healthy, or if it is required to execute
maintenance ahead of timeif any failure indications
are shown. This
potential is an advantage to
shipowners, since it can prevent breaks or delay a
revision, and consequently, it would involve an
economicimprovement.
FUNDING
ThisworkhasbeensupportedbytheSpanishGovernment
[MAQ‐STATUS DPI2015‐69325‐C2] and [DPI2015‐69
1808271602] of Ministerio de Economía y Competitividad
and with European Funds of Regional Development
(FEDER).
REFERENCES
Baojia, C., Baojia, S.,Fafa, C., Hongliang, T., Wenrong, X.,
Zhang,F.,Zhao,C.,2018.Faultdiagnosismethodbased
onintegrationofRSSDandwavelettransformtorolling
bearing.Measurement.
Cerrada, M., Sánchez, R., Cabrera, D., Zurita, G., Li, C.,
Cerrada, M., Sánchez, R.V., Cabrera, D., Zurita, G., Li,
C.,2015.
Multi‐StageFeatureSelectionbyUsingGenetic
Algorithms for Fault Diagnosis in Gearboxes Based on
VibrationSignal.Sensors15,23903–23926.
Chen,S.,Cowan,C.F.N.,Grant,P.M.,1991.Orthogonalleast
squares learning algorithm for radial basis function
networks.IEEETrans.NeuralNetworks2,302–309.
Cordón, O., Gomide, F., Herrera, F., Hoffmann,
F.,
Magdalena,L.,2004.Tenyearsofgeneticfuzzysystems:
currentframeworkandnewtrends.FuzzySetsSyst.141,
5–31.
Gkerekos, C., Lazakis, I., Theotokatos, G., 2017. Ship
Machinery condition monitoring using performance
data through supervised learning. ISBN Smart Sh.
Technol.9781909024632,105–111.
Go, H., Kim, J.‐S., Lee, D.‐H.,
2013. Operation and
preventive maintenance scheduling for containerships:
Mathematical model and solution algorithm. Eur. J.
Oper.Res.229,626–636.
Gou,X.,Bian,C.,Zeng,F.,Xu,Q.,Wang,W.,Yang,S.,2018.
A Data‐Driven Smart Fault Diagnosis Method for
Electric Motor, in: 2018 IEEE International Conference
on Software Quality,
Reliability and Security
Companion(QRS‐C).IEEE,pp.250–257.
He, J., Yang, S., Gan, C., He, J., Yang, S., Gan, C., 2017.
Unsupervised Fault Diagnosis of a Gear Transmission
ChainUsingaDeepBeliefNetwork.Sensors17,1564.
390
Jakovlev, S., Andziulis, A., Daranda, A., Voznak, M.,
Eglynas,T.,2017.Researchonshipautonomoussteering
control for short‐sea shipping problems. Transport 32,
198–208.
Jang, J.‐S.R., 1993. ANFIS: adaptive‐network‐based fuzzy
inference system. IEEE Trans. Syst. Man. Cybern. 23,
665–685.
M.Samhouri,A.Al‐Ghandoor,S.
AlhajAli,I.Hinti,W.M.
a, 2009. An Intelligent Machine Condition Monitoring
System Using Time‐Based Analysis: Neuro‐Fuzzy
Versus Neural Network. Jordan J. Mech. Ind. Eng. 3,
294–305.
Marichal, G.N., Hernández, A., Rojas, J.A., Melón, E.,
Rodríguez,J.A.,Padrón,I.,2016.SistemaInteligentede
apoyoamaniobrasde
grandesbuquesenpuertos.RIAI‐
Rev.Iberoam.Autom.eInform.Ind.
Martini,A.,Troncossi,M.,2016.Upgrade ofanautomated
lineforplastic capmanufacturebasedonexperimental
vibrationanalysis.CaseStud.Mech.Syst.SignalProcess.
3,28–33.
Muszynska,A.,2005.Rotordynamics,CRCTaylor&Francis
Group.NewYork.
Nobre, F.S.M.,
1995. Genetic‐neuro‐fuzzy systems: a
promising fusion, in: Proceedings of 1995 IEEE
International Conference on Fuzzy Systems. The
International Joint Conference of the Fourth IEEE
International Conference on Fuzzy Systems and The
Second International Fuzzy Engineering Symposium.
IEEE,pp.259–266.
Rajasekaran,S.,Pai,G.A.V.,2003.NeuralNetworks,Fuzzy
Logic
and Genetic Algorithm: Synthesis and
applications.
Simani, S., Fantuzzi, C., Patton, R.J., 2003. Model‐based
Fault Diagnosis in Dynamic Systems Using
IdentificationTechniques.
Wang, J., Zhang, L., Duan, L., Gao, R.X., 2017. A new
paradigm of cloud‐based predictive maintenance for
intelligent manufacturing. J. Intell. Manuf. 28, 1125–
1137.
White, G.,
2010. Introducción al Análisis de Vibraciones.
AZIMA,16‐98.
