Uni e si y o he Basque Coun y (UPV/EHU)
Depa men o compu a ion sciences and a i icial in elligence
Uni e si y o he Basque Coun y (UPV/EHU)
On he use o con ex in o ma ion o an
imp o ed applica ion o da a-based algo i hms
in condi ion moni o ing
Ke man López de Calle E xabe
1. Supe iso Basilio Sie a
Depa men o compu a ion sciences and a i icial in elli-
gence
Uni e si y o he Basque Coun y (UPV/EHU)
2. Supe iso Susana Fe ei o
In elligen In o ma ion Sys ems
Teknike
July 27, 2020
(cc) 2020 Ke man López de Calle E xabe (cc by 4.0)
Ke man López de Calle E xabe
On he use o con ex in o ma ion o an imp o ed applica ion o da a-based algo i hms in
condi ion moni o ing
July 27, 2020
Re iewe s: Basilio Sie a and Susana Fe ei o
Uni e si y o he Basque Coun y (UPV/EHU)
Uni e si y o he Basque Coun y (UPV/EHU)
Depa men o compu a ion sciences and a i icial in elligence
Manuel La dizabal Ibilbidea, 1
20018 Donos ia (Gipuzkoa)
Abs ac
The i up ion o senso ized, connec ed and au onomous machine y has al eady
ma ked a miles one in he manu ac u ing indus y. This pa adigm also known as
Indus y 4.0 consis s in he digi aliza ion o indus ial p ocesses and i is p o iding a
new and p e iously unde used sou ce o knowledge: da a.
The da a s eams ha a e massi ely being gene a ed and s o ed a e big sou ces o
in o ma ion. Wi h he p ope ools, he knowledge ex ac ed om hese da a can
enable moni o ing and con ol, boos he decision making and cus ome sa is ac-
ion as well as he e iciency, he p oduc i i y, and he op imal use o acili ies in
manu ac u ing.
The p ocess o knowledge ex ac ion o da a mining is ca ied ou using algo i hms
ha can handle la ge da a olumes and a e capable o ex ac ing pa e ns ha a e
hidden in he da a. Once pa e ns a e ound, hey can be cap u ed in Machine
Lea ning (ML) models o be la e applied in a ange o di e en ields wi h di e en
pu poses.
In he discipline o condi ion moni o ing, he discipline ha deals wi h he de ec ion
o he heal h s a us o he asse s, he so-called da a-based models ha e been a ound
o a long ime. S a is ical me hods as he quali y con ol cha s ha e been used
o de ec anomalies in ope a ing machines since he las cen u y. La e on, mo e
sophis ica ed classi ica ion and eg ession algo i hms such as a i icial neu al ne -
wo ks (ANNs) o suppo ec o machines (SVMs) ha e been used o de ec and
p edic ailu es wi h condi ion moni o ing pu poses. Howe e , mos o he wo ks in
he li e a u e o e look one o he mos impo an ac o s ha a e in ol ed in he
implemen a ion o algo i hms: he con ex . In sho , we can say con ex is composed
by he ac o s ha ha e in luence in he moni o ing o a machine. Hence, a de ailed
unde s anding o he oppo uni ies and limi a ions o he con ex o a pa icula
applica ion is needed o pu algo i hms in p oduc ion. As hey could limi he use
o ce ain algo i hms o could enable he use o some o he algo i hms ha a e no
sui ed o he p oblem a a i s sigh .
Mos o he algo i hms a e designed wi h a ce ain con ex in mind, ne e heless,
hey end o be applied wi hou conside ing ha he inal applica ion con ex migh
change. In condi ion moni o ing, his is he case o aul diagnosis algo i hms ha
iii
a e ained and alida ed in es igs, wi hou on-si e da a ha will enable e aining
he algo i hm in he eal-li e applica ion. O o he algo i hms de eloped in s eady
condi ions ha do no conside he eal ope a ion is a ying o e ime.
This wo k discusses he ole o he con ex in condi ion moni o ing algo i hms in
h ee di e en si ua ions dealing wi h he cons ain s and he chances ha a e
ela ed o each case o s udy. The applica ions discussed a e wind u bine gea boxes
ope a ing unde a ying ope a ing condi ions; o a ing machine y ope a ing unde
s eady condi ions wi h a lack o knowledge ega ding i s deg ada ion; and, an
elec omechanical ac ua o ha has been diagnosed wi h he help o a physical
model. The con ex s a e s udied and a e wa ds, solu ions a e p oposed o p o ide
ad hoc designed algo i hms ha compose condi ion moni o ing sys ems.
Al hough his hesis p ojec deals wi h h ee speci ic cases o s udy, he con ex s
ha a e add essed in his wo k can be ound in many o he condi ion moni o ing
p oblems, which makes he lessons ob ained in his wo k be ans e able o o he
eal condi ion moni o ing p oblems.
i
Labu pena
Sen so iza u ako, in e konek a u ako e a au onomoa den makine ia en age pena
dagoeneko muga i bilaka u da ab ikazioa en indus ian. P ozesu indus ialen
digi alizazioan da zan e a Indus y 4.0 izenaz ezagu ze a eman den pa adigma
honek o ain a e e abili gabea zen jakin za i u i ba ez ho ni zen gai u: da uak.
Masiboki so zen e a pila zen a i di en da uok e amin a egokiekin balia u a, da -
uek du en baliozko in o mazioa moni o izazio e a kon ola ahalbide zeko, e abaki
ha zea e a beze oa en gogobe e zea sus a u, zein ab ikazio ins alazioen e izien zia,
p oduk ibi a ea e a e abile a op imizazioan lagun zeko e abil dai eke. In o mazio
e auzke a p ozesua edo a da u mea za i za da u bolumen handiak maneia u e a
da ue an izku a u ik dauden pa oiak opa zeko gai di en algo i mo bi a ez egi en
da. Behin pa oiak opa u os ean, Makina Ikaske a (Machine Lea ning) modeloen
bi a ez pa oiok kap u a u e a ge oago hainba alo e an e a bes e ho enbes e
xede ekin aplika zen di a.
Osasun-egoe a en moni o izazioa en (Condi ion Moni o ing) diziplinan, hau da,
makinen osasuna en de ekzioaz a du a zen den diziplinan, da ue an oina i u ako
modeloak aski ezagunak di a. Kali a e kon ole ako g a ikoak, adibidez, ope a zen
dauden makine an anomaliak de ek a zeko e abiliak izan di a azken mendee a ik.
Be anduago, so is ika uagoak di en klasi ikazio zein e eg esio algo i moak, hala
nola sa e neu onalak edo a bek o e oina i makinak, osasun-egoe a moni o izazioa
helbu u zela ik e abili izan di a. Halabe , li e a u an age i di en lan gehien suenek
algo i moen inplemen azioan be ebiziko ga an zia duen ak o e ba ain za ha zea
ahaz en du e: kon es ua. Labu ki esanda, kon es ua makina ba en moni o izazioan
e agina du en ak o eek osa zen du ela esan genezake. Ho ega ik, aplikazio e -
eale an algo i moak p odukzioan ja i ahal iza eko kon es uak dauzkan muga e a
baliabideak ongi ule zeak be ebiziko ga an zia du. Izan e e, algo i mo jakin
ba zuen e abile a muga u ik egon bai ai eke kon es u zeha z ba zue an, aldi be ean,
lehen begi ada ba ean baz e u iko algo i mo ba e abil li eke baliabideei ondo
e epa a uz ge o.
Algo i mo gehienak kon es u ba helbu u dela ik diseina zen di a, halabaina, amaie -
ako aplikazioa en kon es ua ezbe dina izan li ekeela kon uan ha u gabe e abili ohi
di a. Osasuna en-egoe a en moni o izazioa en kasuan hau ge a zen da diagnosi -
ako en ena u iko algo i moekin en segu-bankue an en ena u e a es ea zen di e-
la ik aplikazio e ealean be an da uak so zeko dauden mugak kon uan ha u gabe,
ondo ioz, be iz en ena zeko inongo auke a ik gabe. Edo a egoe a egonko ean
ope a zen a i dela on za ha u a ga a u ako algo i moak, amaie ako aplikazioan
ope azioa aldako a dela kon uan ha u gabe.
Lan honek kon es uak osasun-egoe a en moni o izazio algo i moengan duen ola
az e zen du 3 egoe a ezbe dine an bakoi zak di uen baliabide e a mugei e epa a zen
zaiela ik. Hu engo aplikazio e emuak az e u di a: ope azio aldako eko haize
e o a ba en bide kagailua; deg adazio bilakae a ezezaguna du en e a ope azio
egonko ean diha du en makina bi aka iak; e a, modelo isiko ba en lagun zaz
diagnos ika u ako e agingailu elek omekanikoa. Kasuon kon es uak az e u e a
ondo en soluzioak p oposa zen di a esp esuki ga a u ako algo i moen bidez osasun-
egoe a en moni o izazioa hobe zeko bidean.
Lanak 3 kasu espezi iko az e zen badi u e e, be an az e u ako kon es uak bes e
hainba moni o izazio p oblema an opa dai ezke, ike kun za hone an ikasi ako
lezioak bizi za e ealeko bes e p oblema ba zue a a ans e i u dai ezkeela ik.
i
Acknowledgemen
Con on ing he obs acles o he pa h and lea ning om hem, xiki i- xiki i, his
ma ch has eached i s 35 h poin , lea ing behind ano he s age o his “ xuin” li e.
In he mean ime, in some way o ano he many o you ha e helped me ca y ou
his esea ch, and o ha easons, I would like o hank you.
S a ing om hose o you who ha e helped me echnically and ha e di ec ed my
wo k, my u o s, hank you Susana o you ha e been so close o me du ing all
his ime and you ha e been eady o help when I ha e needed i , and hanks o
you oo, Basi, o ending all he e-mail wi h encou aging wo ds. Thank you also o
bo h o you, Ai o and Ana, e en i you name is no o icial in he pape wo k, his
hesis p ojec has he oo p in o you guidance, as you ha e done i you s many
imes. Despi e being o a sho pe iod o ime, hank you e y much, C is obal, o
welcoming his o eigne and helping me as much as you ha e helped me. A he
same ime, I would like o specially hank hose o you ha ha e men o ed me, e en
by lea ing aside you own du ies o help me wi h my wo k, Ion, Ruben, Me i xell,
and he e a e many o he s I could men ion, as he unexpec ed help p o ided by
Alexand e when only a “ligh - e iew” o his wo k was equi ed. Thank you oo.
This phase, howe e , has equi ed o some elaxing o balance he o he side o
esea ching and lea ning, and ha would no be possible wi hou you, my a el
companions. So hank you membe s o he Teknike co ee b igade (San i, Ál a o,
Egoi z, Iñaki, C is ina, Edu and o he o he sho e m membe s), he Lunch ime
commi ee o C an ield (Pa an, Iñigo, Alex, Se gio, Law ence e a Dedy) and he
g oup o G eek philosophe s (Gio gios, Gio gios, Ye is and Alozie). Thank you o
he simple disconnec ion cha s. Thank you o showing me di e en coun ies and
cul u es. And hank you o helping me ind he sense o li e and he solu ion o he
uni e sal sal a ion.
A he same ime, I wan o specially hank hose o you ha ha e been on my side,
no jus in his phase, bu in he whole jou ney, I would like o hank you oo.
On he one hand, I wan o hank iends and kuad illa membe s, o making me
smile on he weekend a e he wo s week. And I am so y, bu , as you a e so many,
you will ha e o be sa is ied wi h a “chee s Julaspas” his ime.
ii
On he o he , I wan o hank you, he ones ha ha e isen and educa ed me and
ha e kep me on he igh ack du ing all his ime. F om G andpa-g andma, uncle-
aun s, and cousins o he un o ge able ime we ha e had, in pa icula , mo he and
a he , ou you encou agemen and you e o s o keep me ocused, and, Ishmael
and Pa xi, o making me pa o you li le mischie s. Finally, hanks o you, Jani e,
o sha ing my ea s and my joy and o being my suppo when I mos needed i .
Thank you all!
iii
Eske zak
Bideko oz opoei au e eginez e a haie a ik ikasiz, xiki i- xiki i, pa idu honi e e
ailega u zaio 35. an ua, bizi za “ xuin” honen bes e e apa ba a zean u ziz. Ta ean,
e a ba e a edo bes e a ike ke a hau au e a e ama en lagundu nauzuenak asko
za e e, e a ho ega ik, eske ak eman nahi dizkizue .
Teknikoki lagundu e a lana bide a u duzuen u o eengandik hasi a, eske ik asko
Susana denbo a guz i hone an ho en ge ukoa izan e a beha zene ako p es ego ea -
en; e a zeu i e e Basi, ida zi ako mezu guz ie an animoak ema ea en. Bai a zuei
e e, Ai o e a Ana, pape e an age u ez a en, esi honek baduelako zuen gida i za en
az a na, zeuena egin bai uzue hainba une an. Epe labu ez izan bada e e, eske ik
asko, C is obal, a o z honi ongie o ia eman e a lagundu nauzun bes e lagun zea-
ga ik. E a be ean be eziki eske u nahi nizueke men o e izan za e enei, eskua ean
zeneuka en lana u zi e a ni e dudak a gi zeko p es u age u za e enei, Ioni, Rubeni
e a Me i xelli, e a bes e asko aipa u li ezke, us ekabean, esia en e ebisio azka a
eska u a p es u asun e a lagun za osoa eskaini didazuenoi, Alexand e kasu, eske ak
zuei e e.
Fase honek o dea, ike ze e a ikas eaz gain, e laxa zea en o eka e e beha izan du e a
ho i ez li za eke posible izango bidelagun izan za e enenga ik izan ez bali z. Eske -
ak be az Teknike eko ka e b igadakoei (San i, Ál a o, Egoi z, Iñaki, C is ina, Edu
e a denbo a labu ez b igada-kide izan za e en bes eoi), C an ild-eko bazkalo duko
komi ekoei (Pa an, Iñigo, Alex, Se gio, Law ence e a Dedy) e a G ezia iloso oen
aldea i (Gio gios, Gio gios, Ye is e a Alozie). Mila eske egune oko deskonexio
hizke aldi sinpleenga ik. Mila eske he ialde e a kul u a ezbe dinak e akus eaga ik.
E a mila eske bizi za en zen zua e a mundua en salbamena opa zen lagun zea-
ga ik.
E a be ean, e apa hone az gain, bide osoan zeha ho egon za e enoi, ba ez-e e,
zuoi eman beha dizkizue eske ak.
Alde ba e ik, koad ilako kide e a bes elako lagunei eske u nahi nizueke, as e ik
oke ena en os eko as ebu uan nega egin a eko ba eak e agi ea en. E a, sen i zen
du baina, ho enbes e za e enez ge o, “Zu ega ik doa Julaspas” ba ekin kon o ma u
beha ko za e e o aingoan.
ix
me allu gy and conc e e indus ies eme ged, and he s anda ds o li ing o he
gene al popula ion consis en ly inc eased.
A ound 1850, he second indus ial e olu ion o he echnology e olu ion began,
hanks o he use o ail oads; la ge-scale i on and s eel p oduc ion; and he use o
elec ically powe ed p oduc ion, ha s a ed in an Ame ican slaugh e house in 1870
and soon sp ead o o he indus ies. Again, li ing s anda ds inc eased and, addi ion-
ally, he p ices o goods ell d ama ically due o he inc ease in p oduc i i y.
A he end o he 20 h cen u y, abou 1970, he digi al e olu ion o hi d indus ial
e olu ion began. This ime i was igge ed by he in en ion o he ansis o , which
allowed he u he de elopmen s o p og ammable logic con olle s (PLC) and he
consequen au oma ion o indus ial p ocesses.
Finally, in ou e a, he beginning o he 21s cen u y, he so called Indus y 4.0 o
he o h indus ial e olu ion has s a ed o ake shape hanks o he ini ial boos o
he Ge man go e nmen . This incoming e olu ion wan s o u n he indus y sma
by making he en i e supply chain accessible and con ollable h ough he in e ne .
In ha di ec ion, he e a e 5 componen s ha a e adop ing key oles in he indus y
o he u u e:
•
Cybe Physical Sys ems (CPS): CPSs a e sys ems ha allow he in e ac ion
be ween he physical machines and compu e sys em models while he e is a
da a exchange be ween hem.
•
In e ne o Things (IoT): This concep e e s o he end o connec ing e e y
single de ice o he in e ne , which gi es accessibili y ad an ages by p o iding
new da a exchange possibili y o de ices.
•
In e ne o Se ice (IoS): Th ough he in e ne o se ices, sys ems a e able o
ecei e da a om online esou ces ha a e ela ed o hei domain.
•
Sma ac o y: The goal behind he Sma Fac o y is o make p oduc ion
lexible by ha ing o al and cons an con ol o e he s a us o each p oduc .
Consequen ly, esou ces will be be e planned and he quali y o he inal
p oduc s will be con inuously assu ed.
•
Cybe secu i y: Due o he ull connec i i y o he sys ems, new isks will
appea , as ne wo ks would be di ec ly in con ol o he in eg i y o he as-
se s. The e o e, new e o s would be needed o a oid ulne abili ies in he
connec ions and he p o ec ion o sys ems/asse s.
4Chap e 1 Backg ound
This p og ession has pushed he humani y om using s one o wooden ools o
coding in he sma and connec ed indus y o he su i al o he species.
1.2 Main enance: When ou ools ail
In pa allel o he de elopmen o he i s ools by ances o s, he e ec o wea
and s ain appea ed and made he ools ail. Hence, ou ances o began he new
and "bo ing" a o epai ing and main aining he ools as soon as hose i s ools
b oke down. O e ime, machines and ools became mo e complex, bu he policy o
wai ing ill hey ailed o ix hem wi h spa e pa s las ed o cen u ies.
Howe e , ollowing he end o de eloping mo e complex and use ul ools, and
adding on he op o ha ou g owing dependence on he co ec unc ioning o
hese ools, new needs eme ged. As he shu downs o he machines could cause
a al ailu es o huge economical losses, indus ies s a ed o ocus on educing
down ime a e Wo ld Wa II. Fo ha pu pose, hey s a ed o app oxima e he li e
o he asse s and o eplace componen s ollowing ei he schedules o wea ela ed
pa ame e s such as cycles o kilome es o a oid b eakdowns. Ne e heless, he ixed
in e al eplacemen s lead o unneeded eplacemen s which p oduced addi ional
expenses. The e o e, since 1980 he e o s a e being made o de ec he heal h
s a us o he asse s, which is done by means o Condi ion Moni o ing, and o pe o m
main enance ac ions only when equi ed.
And his is how main enance is going om being co ec i e ( epai ing when ailu es
a e gi en) o condi ion based (de ec ing heal h s a us and planning main enance
acco dingly) going h ough p e en i e (pe o ming ixed ime epai s).
1.3 Condi ion Moni o ing: A b ain o he senses
As explained in he p e ious sec ion, Condi ion Moni o ing aims o de ec he cu en
heal h s a us o ou asse s. The main ad an ages o condi ion moni o ing a e ha ,
as he condi ion o he asse is known, eplacemen s a e only ca ied ou when
1.2 Main enance: When ou ools ail 5
equi ed; ca as ophic ailu es can be de ec ed; and, i is possible o plan eplace-
men s conside ing he p esen and he p edic ed wea o he asse s (also known as
P edic i e Main enance PdM).
Essen ially, condi ion moni o ing ope a es by aking physical measu emen s o he
asse s ha a e indica o s o de e io a ion. La e , he e olu ion o hese measu emen s
o e ime is analysed, and pa e ns o ends a e used o e lec he ac ual heal h
s a us o he asse s and o in e he e olu ion o he heal h s a us. Fo he measu e-
men o hese indica o s, a ious senso a e ypically used, such as: accele ome e s,
acous ic emission senso s, he mocouples, achome e s, oil deb is senso s, in a ed
he mog aphy, cu en senso s, e c.
Once measu emen s a e a ailable, he p ocedu e in Figu e 1.1 is ypically ollowed:
Fig. 1.1.: Elemen s o condi ion moni o ing.
Da a acquisi ion sys ems eco d he signals. Then, his signals a e i s p ocessed
(such as denoised and ans o med o equency domain) and desc ip o s (s a is ical
momen s such as mean and s anda d de ia ion) a e ex ac ed. Addi ionally, i he
ela ion be ween hese desc ip o s and he condi ion o he asse s is known, i is
possible o de elop a model ha e lec s he condi ion o he asse .
Depending o he unde s anding o he aul s o he machine and he his o ical da a
o he machine i sel o simila ones, hese models can be ca ego ized in o h ee
s ages, ega ding hei complexi y and le el o de ail o CM, as ep esen ed in Figu e
1.2.
Fig. 1.2.: Condi ion moni o ing s ages.
Each ca ego y ep esen s a highe deg ee o knowledge p o ided by he model and
he ca ego ies can be explained in sho as:
• De ec ion: The model dis inguishes no mal/abno mal condi ions.
• Diagnosis: The model knows which aul is causing he abno mal condi ion.
6Chap e 1 Backg ound
•
P ognosis o Remaining Use ul Li e (RUL): The model app oxima es how long
he machine will ope a e be o e aul causes ca as ophic damages.
These models, ha a e he co e o he condi ion moni o ing sys ems, can be de el-
oped in h ee di e en manne s. Ei he he model is based on physical p inciples
(physics based), o he model is in e ed om expe imen al da a (da a-based) o i
combines bo h (hyb id).
Pa icula ly, due o he ecen ad ances in compu ing, cloud da a s o age and sens-
ing, added o he complexi y o de eloping physics based models o ce ain asse s,
da a-d i en modelling is gaining pa icula a en ion. Fo ha eason, conside able
amoun o wo ks using la es s a e o he a da a-d i en models a e lou ishing in
he condi ion moni o ing li e a u e. F om simple k-nea es neighbou , going o he
mo e complex Suppo Vec o Machines, going h ough andom o es s and eaching
o he la es Deep Lea ning models.
Ne e heless, mos o he wo ks use da a ob ained in es igs, wi h aul s ha
ha e al eady been seeded; un o ailu e es s, whe e expe imen condi ions a e
s eady; o da a om physical models, which end o be qui e cleane han signals
om eal applica ions. The e o e, he e is a poo ep esen a ion o da a-based
models in p oduc ion en i onmen s, despi e he ac ha he p esence o Compu e
Main enance Managemen So wa e (CMMS) is in inc easing end.
1.4 Resea ch en i onmen : The e ilise ha g ows he
plan s
The de elopmen o his esea ch wo k is specially bonded o wo pa icula en i ies:
The Uni e si y o he Basque Coun y (EHU/UPV) and he esea ch cen e Teknike .
Bo h en i ies a e si ua ed in he Basque Coun y, a egion wi h a solid indus ial
base whe e indus y p oduces 23.5 % o he GDP whi a s ong go e nmen al bid o
ad anced manu ac u ing [@G u17]. In addi ion, he egion has a pa icula ly s ong
machine ool sec o , accoun ing o 78% o he whole machine ool p oduc ion o
Spain [@P e08].
A he same ime, he igu es p o ided by he Spanish Main enance Associa ion
(Asociación Española de Man enimien o - AEM) e lec ha , e en hough he awa e-
ness o adop ing imp o ed main enance echniques exis s in Spain (wi h CMMS
use going om 63% in 1995 o 98% in 2005), i s adop ion is ye in in an s ages.
1.4 Resea ch en i onmen : The e ilise ha g ows he plan s 7
The la es poll (2015) e eals ha he expendi u e in co ec i e main enance (%44)
is in dec easing end gi ing place o inc easing p e en i e main enance (%46),
ha has su passed co ec i e o i s ime. Addi ionally, p edic i e main enance
appea s o i s ime, accoun ing o he 10% o he emaining expendi u e. Fu -
he mo e, only 22% o he o al o he indus ies answe ing he poll used he CMMS
o ac ual moni o ing o he machines [Man15]. These igu e collide wi h he ones
ga he ed by Ad anced Technology Se ices (ATS), as, acco ding o he 2018 Plan
Enginee ing su ey (su eying wo ldwide bu mos ly Ame ican indus ies), he use
o p edic i e main enance wi h analy ics has inc eased om 47% in 2017 o %51 in
2018 [@Va 18], which sugges s Spanish ma ke is way behind i s mo e compe i i e
coun e pa s.
In his scena io, applied esea ch ha educes he gap om he heo y o main-
enance o i s inal applica ion and makes ou indus y compe i i e gains special
in e es .
1.4.1 TEKNIKER: G ow h make s
Teknike is a esea ch cen e loca ed in Eiba . The esea ch a eas o Teknike include
Ad anced manu ac u ing, su ace enginee ing, ICTs and p oduc enginee ing. I s
closeness o Eiba , c adle o he Basque machine ool indus y, has cen ed big pa
o i s esea ch in machine ools and manu ac u ing. Ne e heless, o he indus ial
sec o s such as ae onau ics, ag o ood, ene gy, in as uc u e and heal h a e also
co e ed by hei esea ch. Du ing he de elopmen o he hesis, he au ho has
g own bo h echnically and pe sonally hanks o he suppo o he su ounding
p o essional esea ch eam.
In elligen In o ma ion Sys ems - SII
Pa icula ly, hanks o he In elligen In o ma ion Sys ems uni ha has a as
expe ience on he ield o condi ion moni o ing as he a ious publica ions and
hei in ol emen in Regional, S a al and Eu opean p ojec s p o e. The ollowing
a e some p ojec s, in which I ha e pa icipa ed, ha ha e aken place du ing he
de elopmen o his hesis p ojec and a e ela ed o condi ion moni o ing:
8Chap e 1 Backg ound
•
Mainwind+: The goal o Mainwind+ p ojec was o de elop a non-des uc i e
oil-deb is sensing echnique ha would allow emo e moni o ing o Wind
Tu bine gea boxes. Teknike ook pa in he de elopmen o he op ical deb is
senso , he da a acquisi ion and s o age pla o m, and, inally, in he analy-
sis o he signals and hei ela ion wi h he ope a ion o he Wind Tu bine.
Mainwin+ was a HAZITEK p ojec (indus ial RD suppo p ojec unded by
he Basque Go e nmen ) ha las ed 2 and a hal yea s om July 2016 o
Decembe 2018. I was leaded by Inge eam Powe Technology and 9 indus ial
pa ne s and 7 basque esea ch agen s ook pa , being Teknike one o hem.
•
Vi ual: This ELKARTEK p ojec (basic collabo a i e esea ch unded by he
Basque Go e nmen ) deal wi h he de elopmen o hyb id models. Teknike
de eloped models used o he commissioning as well as models ha we e
la e used in combina ion wi h ope a ional da a o moni o ing pu poses. I
began on Ma ch 2018 and closed on Decembe 2019 being Ike lan he leading
esea ch pa ne and Teknike one among he emaining 7 esea ch pa ne s.
•
Man is: This Ho izon 2020 Eu opean p ojec wi h call ECSEL-2014-1 con-
sis ed in he de elopmen o a p oac i e main enance se ice pla o m and i s
associa ed a chi ec u e. Teknike de eloped a sma moni o ing de ice ha
moni o ed clu ch-b akes and uploaded he da a o a cloud based da a eposi-
o y, la e analysing he con en o he da a wi h machine lea ning echniques.
The p ojec began in Ap il 2015 and las ed o 36 mon hs. I was leaded
by Mond agon Goi Eskola Poli eknikoa S. Coop. and a o al o 54 Eu opean
pa ne s ook pa on i .
1.4.2 EHU/UPV: Gi e and sp ead
As s a ed by he mo o "Gi e and sp ead", he Uni e si y o he Basque Coun y has
p o ided he au ho wi h he knowledge ounda ions o he de elopmen o his
hesis. Fi s ly, hanks o he Renewable Ene gy Enginee ing bachelo , unde s anding
o basic enginee ing concep s was acqui ed, and, la e , he Mas e s and he PhD p o-
g ams ela ed o Compu a ional Enginee ing and In elligen Sys ems ha e p o ided
he skills o deal wi h he unde s anding and modelling o complex da a.
1.4 Resea ch en i onmen : The e ilise ha g ows he plan s 9
RSAIT
By means o he di ec o o he hesis, he hesis has been especially a ached o
he RSAIT esea ch g oup. This depa men has consis en knowledge o Machine
lea ning and s a is ics echniques, as well as in hei applica ion o he p e ious o
he ield o condi ion moni o ing.
1.4.3 C an ield Uni e si y: A e clouds ligh
In addi ion o he EHU/UPV and Teknike , pa o his hesis p ojec has aken place
in collabo a ion wi h C an ield Uni e si y h ough a b ie ye ui ul s ay ha las ed
om Sep embe o Decembe 2019. The suppo o he Th ough-li e Enginee ing
Se ices Ins i u e has been in aluable, pa icula ly hei knowledge ela ed o he
moni o ing o linea ac ua o s which de ini ely b ough he ligh o he long awai ed
success o he hyb id modelling.
10 Chap e 1 Backg ound
Mo i a ion 2
„I can no ain you he way I ained he Fi e. I
now see ha he way o ge h ough you is wi h
his. (Shows a bowl wi h bean buns)
—Mas e Shi u
(Kung u panda)
Shi u was he well-known aine o he Fu ious Fi e, he gua dians o he Valley o
Peace. Howe e , a new incoming h ea o ces Shi u o ind a new disciple, Po, who,
acco ding o Oogway ( he c ea o o Kung Fu), will be he nex D agon Wa io . Po
is a om being any hing ha Shi u could ha e expec ed. He is a lazy panda ha
d owns his so ows by ea ing. And Shi u’s a emp s o ain Po as he did wi h he
Fi e ob ain disas ous esul s. Howe e , Shi u’s aining ou come changes when he
unde s ands ha he con ex he is acing now is di e en . Tha he d i ing o ce
ha makes Po an o e -weigh ed lazy panda, his endless amine, can be used o u n
him in o he nex D agon Wa io .
2.1 Whe e o push he bounda ies: Con ex
Algo i hms a e powe ul ools ha ha e been p o en o be g ea solu ions o some
in ica e p oblems. They a e equi alen o a i eless g oup o wo ke s, ha can
execu e o de s in a minimal amoun o ime wi h minimum esou ces wi h su gical
p ecision. Howe e , one o he big p oblems hey ha e is he lack o in ui ion, he
lack o capabili y o p ocessing and adap ing o he con ex hey a e su ounded by.
In o he wo ds, i an algo i hm ained o classi y be ween dogs and ca s is shown a
cow, i will ha e no doub and will classi y i as canine o eline and sleep well ha
nigh . Due o his lack o in ui ion, enginee s a e needed o code he con ex in he
bes possible manne so ha b ainless IA has no choice o ake w ong decisions. And
ha is exac ly he opic o his hesis, which is no abou he use o he la es and
mos powe ul algo i hms, a he han how o p ope ly code he con ex and use i
11
in he bes possible manne o imp o e he pe o mance o ou decision aking a my,
ou u u e D agon wa io s.
As p esen ed in he p e ious chap e 1, huge ad ances in he heo e ical aspec s
o Condi ion Based Main enance ha e been done ecen ly, ne e heless, indus ial
con ex s ha e some complexi ies ha a e some imes o e looked in esea ch wo ks.
The use o “Con ex -awa eness” and “Con ex ” is widely adop ed by pe asi e and
mobile compu ing ield [SGW16] bu i s de ini ion is somehow uzzy. In an a emp
o clea ing ou he uzziness o he e m, a good de ini ion was p o ided by [DAS01]
which de ines con ex as
“Any in o ma ion ha can be used o cha ac e ise he
si ua ion o en i ies (i.e. whe he a pe son, place o objec ) ha a e conside ed
ele an o he in e ac ion be ween a use and an applica ion, including he use
and he applica ion hemsel es”
. As he concep o con ex has no a no ewo hy
p esence in he ield o main enance [SW15], we ha e adop ed he de ini ion o he
ield adap ing he p e ious de ini ion. Fo us, con ex would be:
“The in o ma ion
conside ed ele an o he cha ac e isa ion o he in e ac ion be ween an asse
and i s moni o ing”.
Following his de ini ion and, acco ding o ou expe ience, indus ial scena ios a e
go e ned by wo key ac o s ha in luence he applica ion o s a e-o - he-a moni-
o ing algo i hms in indus ial applica ions and, a he same ime, a e in e ela ed:
•Use case:
Use case is de ined by he asse ha is being moni o ed, he aul s i
is expec ed o ha e (i he e is any knowledge abou i ), he sub-componen s,
he numbe o uni s o he asse ha a e a ailable (mass-p oduce non-mass-
p oduced), also, and mos impo an ly he ope a ing condi ions: s eady, a y-
ing, commanded by demand o by ex e nal ac o s, epea abili y, e c.
•Da a a ailabili y:
Da a a ailabili y e e s o he kind and ype o da a ha is
a ailable o he de elopmen o he moni o ing p ocess. This includes, he
senso s used o he moni o ing wi h hei sampling equencies, he ex e nal
condi ions ha a e eco ded ( empe a u es, o he senso s e lec ing ope a ing
condi ion bu no ela ed o heal h), he his o ic da a o he machine a ailable,
he ypes o eco ds s o ed in he his o ic da a (a e he e any aul s? Is
e e y hing nominal da a?), e c.
The deg ee o complexi y and he ma u i y o he moni o ing sys ems will be di ec ly
a ec ed by hese ac o s ha compose he con ex . To ans e a moni o ing app oxi-
ma ion o a inal applica ion i mus ha e simila i no equal con ex . A he same
12 Chap e 2 Mo i a ion
ime, he knowledge ob ained in simila con ex is qui e gene alizable, and i migh
be ans e ed o a simila con ex .
This is he main poin o his hesis wo k ha iden i ies h ee indus ial con ex s
om h ee pa icula use cases. In each pa icula case an app oach ha imp o es
he cu en moni o ing o he asse is p esen ed, and, as he app oaches a e based in
he concep o con ex and a e use case agnos ic, we suppo ha he app oach can
be ans e ed o simila con ex s.
The concep o con ex is le e aged by he p oximi y Teknike has o eal indus ial
applica ions, and d i es his wo k o ocus he esea ch on a way ha allows he
ans e o he esea ch ou comes o he indus ial applica ions. In he cu en
scena io we a e in, his means wo king wi h non-mass-p oduced machines ha
a e oo complex o simula e, da a ha is di icul o ob ain, lack o knowledge
ega ding mos impo an indica o s and a ying ope a ing condi ions ha a e no
commanded.
2.2 Resea ch ques ion
Based on he p e ious expe ience acqui ed in Teknike and i s closeness o eal
indus ial applica ions a se o common scena ios is iden i ied. These scena ios
include:
•
Non mass-p oduced machines ope a ing unde s eady condi ions: This scena io
is gi en ypically in machine ool. Gene ally, he numbe o machines is
educed, and he knowledge and unde s anding o he possible aul s is sca ce.
•
Machines ope a ing unde a ying ope a ing condi ions which is no com-
manded: This scena io ep esen s machines whe e o cing ce ain ope a ion is
no possible and he e is need o ob ain a heal h indica o .
•
Diagnosis o machines ha ha e no aul ela ed eco ds: This scena io ap-
plies mos ly o cases whe e he numbe o machines is educed and, hence,
he his o ic da abase is also educed and has no eco ds o aul y machines.
The e o e, de eloping diagnosis algo i hms unde hese condi ions is complex.
These con ex s iden i ied by Teknike a e qui e equen . Tha is, clien s and ma-
chines can a y, bu he limi a ions and p oblema ic a e simila . Fo ha eason,
his p ojec o mula es and ies o answe he ollowing Resea ch Ques ions, ha
2.2 Resea ch ques ion 13
3.2 Case s udy: Ro a ing machine y
Ro a ing machines can be ound in nume ous indus ies, such as oil indus y, a ia-
ion, mining and anspo a ion, among o he s. Fu he mo e, hey end o ope a e
unde ad e se condi ions su e ing high empe a u es and high loads, hence, hey
su e om pe o mance deg ada ion and mechanical ailu e [Li+17]. Ro a ing
machines include pumps, mo o s, olling bea ings, gea s, gea boxes, sha s, ans...
and hey a e among he mos impo an equipmen in mode n indus ial applica ions
[Liu+18]. Fo example, bea ings, which a e mechanical componen s used equen ly
in mos o a ing de ices, can cons i u e as much as 44% o he o al numbe o aul s
in some de ices [Ce +18].
Due o he c i icali y o o a ing machines, hey ha e been widely add essed in
he CM li e a u e. E en hough di e en senso s can be used o hei moni o ing
such as acous ic emissions [Ce +18], empe a u e, p essu e, oil analysis, noise
[ASC11] o mo o cu en signa u e [B a+17], mos o he wo ks ely on he use
o ib a ions, as excess ib a ions e lec unbalance, misalignmen , wo n gea s o
bea ings, losseness among o he s as Vishwaka ma [Vis+17] s a es.
Rega ding he di e en signal p ocessing me hods used o wo k wi h ib a ion
signals, h ee ca ego ies can be dis inguished: ime domain, equency domain and
ime- equency domain echniques; he i s includes he use o s a is ical momen s
such as RMS, Ku osis, C es ac o o noise il e ing echniques such as Time Syn-
ch onous A e aging o he en elope; he second, includes he ans o ma ion o aw
signals in o equency domain by means o Fou ie T ans o m, in gene al, equency
domain ea u es a e be e indica o s o aul s, as he esonance equency compo-
nen (o aul componen ) can be be e de ec ed in equency domain when signals
a e s a iona y; las ly, o non-s a iona y applica ions, ime- equency echniques a e
used, such as he Sho Time Fou ie T ans o m (STFT) o he wa ele s, ha allow
he use o long ime in e als o mo e p ecise low- equency in o ma ion o sho
ime in e als o be e p ecision o high- equencies.
In his esea ch olling bea ings and gea s wo king unde s eady condi ions ha e
been conside ed. In addi ion o s opping he machine be o e ha ing ca as ophic
ailu e, ob aining as much knowledge as possible in ela ion o aul s/deg ada ion
du ing he moni o ing o he bea ings and gea s is a emp ed. Fo ha eason,
ocusing much in he complexi y o he ea u es ex ac ed om he ib a ion senso
20 Chap e 3 S eadily ope a ing machines
is ou o he scope o he esea ch, hence, simple s a is ical desc ip o s and basic
equency domain desc ip o s a e ex ac ed.
3.3 Inno a ion
As discussed in he he p e ious sec ion 3.1, he second ca ego y o moni o ing wo ks
sha e a common denomina o : cons an ope a ing condi ions. This ac is no i ial,
as his con ex can be exploi ed wi h algo i hms based on s a is ical dis ibu ions, as
any change in he dis ibu ion is gi en by he deg ada ion o incipien aul s in he
sys em due o he lack o in e e ence o a ying ope a ion. This simple assump ion
abou he s abili y o he con ex has been widely exploi ed o ailu e de ec ion,
ins ead, his wo k goes a s ep u he . In addi ion o answe ing he ques ion "When
should he machine s op be o e being oo la e?" he ques ion "I he machine is
s opping due o wea / aul , which indica o s a e e lec ing his damage he mos ?"
is also answe ed.
Two di e en da ase s a e employed in he esea ch: he i s one is gene a ed in FZG
es igs used o es lub ica ion as well as o he aspec s o spu -gea s; he second
da ase is aken om an open eposi o y, c ea ed o he PRONOSTIA challenge and
consis s o un o ailu e es s o bea ings [Nec+12].
In bo h cases, ea u es a e ex ac ed om he aw signals and he se o ea u es is
educed wi h a dimensionali y educ ion echnique. Fou di e en echniques a e
compa ed: LDA, Relie , Au oencode s and PCA. A e applying he echniques and
educing he dimensionali y o he ini ial da ase o a single dimension, a quali y
con ol cha is used o de e mine when he sys em is ou o con ol. Howe e , as he
eade migh ha e ealised, some o he p e ious algo i hms equi e o class alues
which is no some hing ha i is a ailable. And he e is whe e ad an age o he
s able con ex is aken o go one s ep u he and conside ha he la es da a aken
om a machine has o be always in he same o wo se condi ion han in p e ious
ins an s (conside ing no main enance ac ions). This way, i is possible o conside
he p oblem as a classi ica ion p oblem and use supe ised dimensionali y educ ion
echniques.
3.3 Inno a ion 21
In addi ion, he dimensionali y educ ion p ocess is epea ed pe iodically in o de
o de e mine which ea u es a e showing he g ea es di e ences in compa ison
o he ini ial da a-window in o de o ob ain ex a in o ma ion ela ed o he bes
indica o s o machine deg ada ion.
Las ly, some conside a ions abou he inal applica ion o he algo i hms a e included
by measu ing hei pe o mance and o he aspec s wi h a se o me ics designed
wi h he indus ial needs in mind: cos , in e p e abili y, and e ec i eness. This se o
me ics helps o be e unde s and o he pa ame e s ha some imes a e o e looked,
such as compu a ional cos s o he comp ehensi eness o he ea u e educ ion. The
me ics he e p esen ed p o ide a ai e amewo k o compa e ou algo i hms wi h
moni o ing pu poses.
3.4 Conclusion
Acco ding o he indings, i is possible o u ilise Dimensionali y Reduc ion echniques
o syn hesise he in o ma ion o la ge amoun s o a iables being aul s de ec ed
a he same ime. Rega ding he compa ison among algo i hms, wo algo i hms
ob ained be e esul s: Relie and LDA. The i s one, because ins ead o ea u e
p ojec ion makes anks o ea u es and, he e o e, hey a e s ill unde s andable. The
second, because i s accu acy was g ea e han he es o he algo i hms. I is no
coincidence ha bo h bes algo i hms we e supe ised, as including class alue ben-
e i s by gi ing ex a knowledge. None o he p e ious could ha e been used wi hou
using domain knowledge (compa ing nominal and la es obse a ions), hence, i is
clea ha he applica ion o ML algo i hms bene i s om domain knowledge. O , in
o he wo ds, be e unde s anding o he p oblems o sol e imp o es he chances o
using ools be e sui ed o i s solu ion.
Rega ding he me ics used in he wo k, i has o be conside ed ha o he wo ks
jus ocus on he accu acy, howe e , o he aspec s also a ec in he selec ion o he
"bes " algo i hm. The e o e, hese me ics p o ide a mo e comple e amewo k o
compa ison. Addi ionally, aking a look o he sub-dimensions he me ics ha we e
de eloped we e alida ed, as hey e lec ed co ec ly hei o iginal pu pose.
22 Chap e 3 S eadily ope a ing machines
Finally, i is p o en ha ob aining addi ional in o ma ion while he moni o ing
is ca ied ou when he con ex is kep s eady and all a iance is a ibu ed o
deg ada ion is possible. This has been achie ed by using supe ised dimensionali y
educ ion algo i hms ha dis inguish he ea u es ha change he mos om he
beginning o he s opping poin .
3.5 Publica ions
The s udy ca ied ou on his opic led o wo publica ions:
The Tescon con e ence ha ook place he 6 h and 7 h o No embe 2018, in
C an ield (UK), he wo k "Compa ison o Au oma ed Fea u e Selec ion and Reduc ion
me hods on he Condi ion Moni o ing issue" was p esen ed and la e published in
P ocedia Manu ac u ing.
López de Calle, K., Fe ei o, S., A naiz, A., Sie a, B., 2018. Compa ison o Au o-
ma ed Fea u e Selec ion and Reduc ion me hods on he Condi ion Moni o ing issue.
P ocedia Manu ac u ing 16, 2–9. h ps://doi.o g/10.1016/ j.p om g.2018.10.150
Addi ionally, he a icle i led "Dynamic condi ion moni o ing me hod based on
dimensionali y educ ion echniques o da a-limi ed indus ial en i onmen s", au-
ho ed by pa o he In elligen Sys em uni in Teknike (Ke man López de Calle,
Susana Fe ei o and Ai o A naiz) oge he wi h he Robo ics and au onomous sys-
ems g oup om he Uni e si y o he Basque Coun y (UPV/EHU) ep esen ed by
Basilio Sie a was published in he jou nal Compu e s in Indus y.
López de Calle, K., Fe ei o, S., A naiz, A., Sie a, B., 2019. Dynamic condi ion mon-
i o ing me hod based on dimensionali y educ ion echniques o da a-limi ed indus-
ial en i onmen s. Compu e s in Indus y 112, 103114. h ps://doi.o g/10.1016/
j.compind.2019.07.004
Bo h a icles a e a ached in PART II o he wo k.
3.5 Publica ions 23
Finding s able and epea able
con ex s in a ying ope a ing
condi ions
4
„Science is abou sailing (and aming)
unce ain y.
—Fe nando Blanco
@FBpsy on Twi e , Expe imen al Psychologis
S a ing in la e 2019 and du ing he i s and second qua e s o 2020 wi h a ye
unclea expi a ion da e, a p e iously unknown menace hangs o e he wo ld. This
menace akes he name o Co ona i us (COVID-19) and wi h a high mo ali y and
basic ep oduc ion a ios, his i us sp eads as ac oss he wo ld causing de as a ion
on i s way. Mos o he coun ies a e o ced o enac comple e lock-downs o he
ci izens in hei homes and only minimum se ices ( i s sec o ) a e kep wo king
in many coun ies. Despi e he measu es aken o a oid he sp ead o he i us,
i has caused nume ous a ali ies as well as incipien wo ldwide economical c isis.
In such a deplo able si ua ion, many ci izens and media do no unde s and why
go e nmen s did no ac as e . Meanwhile, go e nmen s hide behind he ac ha
hey ac ollowing scien i ic c i e ia. Once again, science/scien is a e o be blamed
o he bad decisions.
In ha con ex , Fe nando Blanco (@FBpsy on wi e ) wa ns wi h his wee media
and o he displeased ci izens abou how science wo ks: "I you expec science o
p o ide ce ain y and assu ance, i you hink ha (scien i ic) deba e is symp om o
igno ance, hen you ha e now clue o wha science is". The u h is ha science
is based on e idence, he e o e, as new e idence a ises p e ious assump ions a e
e ised and, in some cases, hey may be ejec ed and as new assump ions a e
alida ed.
The scope o his hesis is no ela ed o he COVID-19 (despi e g ea pa o i being
w i en du ing con inemen ). Howe e , du ing he co e age o ou ield (machine y
moni o ing algo i hms) we ha e had he chance o s udy new e idence ha has
25
allowed us o e ise and alida e hypo hesis o mula ed by p e ious wo ks, as well
as o d aw new hypo hesis.
4.1 Backg ound
Da a-based algo i hms used in condi ion moni o ing a e de eloped ypically ollow-
ing his p ocedu e:
1. The aul s ha need o be diagnosed a e iden i ied.
2. A es ig whe e hese aul s can be seeded is buil .
3.
Senso s/signals ha could be help ul o diagnose he aul s a e eco ded du ing
he es s.
4.
Meaning ul ea u es/ ans o ma ions a e used o diagnose o new ea u es/-
ans o ma ions a e es ed.
5. Da a-based algo i hms a e es ed o de e mine hei diagnos ic capabili ies.
This p ocedu e is as ly ex ended and has helped in he de elopmen o many signal
ans o ma ions and imp o ed bo h he ex ac ion o diagnos ic ea u es and he
use o diagnos ic algo i hms. Ne e heless, es igs a e buil ollowing ce ain
assump ions, and, e en i hey a e buil o esemble as much as possible o he inal
applica ion, he e a e some aspec s ha a e di icul o ep oduce and some ex e nal
in luences ha a e neglec ed in hei de elopmen . Consequen ly, he e idence and
conclusions d awn in a es ig (wi h ypically non a ying condi ions) migh no be
accu a e in a inal applica ion we e ope a ions a e no held cons an ( hey migh
no be conclusi e unde he new e idence). In o he wo ds, he ope a ion con ex
o es igs is no o ally accu a e ep oduc ion o he applica ion con ex , which is
ha dly s a iona y.
Being awa e o his ac , he applied esea ch in Teknike has la ely ocused on he
de elopmen o epe i i e es s (known as Finge p in s) in he applica ions ha allow
o cing machines o wo k unde ce ain epe i i e condi ions. These Finge p in s a e
aken pe iodically and allow ob aining s able da a om machines ha wo k unde
e y ansien condi ions. This way compa ing measu emen s om one poin in ime
o ano he is possible, as he measu emen s a e aken in simila s eady con ex s.
Finge p in has been sa is ac o ily used o he moni o ing o machine ool and
o he ypes o machine y, and i eases he knowledge ans e om es igs o inal
applica ions, as he Finge p in s can be aken in condi ions ha a e close o he
26 Chap e 4 Finding s able and epea able con ex s in a ying ope a ing con-
di ions
ones o he igs. A ep esen a i e example is he wo k p esen ed by Fe ei o e al. in
[Fe +16].
Ne e heless, no e e y machine can be o ced o ope a e in a speci ic way, o he
cos o doing i is oo expensi e o he sole pu pose o moni o ing. Fo example, he
case s udy he e p esen ed, wind u bines (WT), can no be o ced o ope a e in a
desi ed way as he d i ing o ce is he wind, which can no be con olled. The e o e,
al e na i es o Finge p in a e equi ed o hei moni o ing, which is wha he he
ollowing sec ions p esen .
4.2 Case s udy: Wind u bine gea boxes
The inc easing elec ical ene gy needs oge he wi h he abundance and a ailabil-
i y o wind ene gy a e igge ing he g ow h o wind u bine indus y [Tch+14;
Ham+09; NW13]. This su ge in wind ene gy demand comes oge he wi h an
inc ease o he size o he u bines ha has been ound co ela ed o highe ailu e
a es [Ech+08; Su+17; Ga +12].
Conside ing Ope a ion & Main enance (O & M) o wind u bines can comp ise om
10-20% o he o al cos o ene gy (COE) acco ding o Tchakoua e al. [Tch+14],
and i can each up o 30% in big o sho e wind a ms [CZH15], he use o Condi ion
Moni o ing Sys ems (CMS) is a mus . CMS a e p o en echniques ha success ully
de ec heal h s a us, de ec aul s and p o ide asse emaining li e es ima ions
[Tch+14; Ham+09; Ga +12]. The applica ion o CMS in wind u bines suppo s he
iden i ica ion o he s a e o heal h o he u bines emo ely and educes he need o
isual and on si e inspec ion o ha pu pose, which is pa icula ly cos sa ing in
o sho e wind a ms[NB07].
The a ious wo ks analysing ailu e s a is ics o gea boxes suppo he hesis o
gea box being one o he mos delica e componen s in WTs [CMM16]. These
wo ks show some con o e sial conclusions ega ding he endency o ailu e ha
gea boxes ha e as P a el, Fauls ich and Roh ig ecognise in [PFR17], as some o
hem ind high ailu e a es [HDR06], whe eas o he s ha e less gea box ela ed
ailu es epo ed [PFR17; Su+17]. In any case, mos o he s udies ela e he longes
WT down imes associa ed o gea box ailu es [NW13; PFR17] and ind i one o he
cos lies pa s o he u bines [CMM16].
Gea box moni o ing is a widely s udied opic in he li e a u e. The p ominen
app oaches in his ield a e based on ib a ion, oil deb is, acous ic emissions and
4.2 Case s udy: Wind u bine gea boxes 27
cu en signa u e analysis, among o he s. E en i he e is a p e alence o ib a ion
based moni o ing wo ks [Ham+09; NW13; Ga +12], oil deb is moni o ing (ODM)
echniques ha e been ound o in e es o gea box moni o ing, because o he highe
co ela ions hey show wi h wea c ea ion as demos a ed by Ka elus, Mie inen and
Leh o aa a in [KML17], and he addi ional capabili y o moni o ing he oil quali y
and he s a e o o he pa s o he gea box [Tch+14; Ham+09].
Ne e heless, he p o en ad an ages o condi ion moni o ing (CM) [NB07; Ham+09]
a e di icul o ans e o WT use cases, and hey ha e a smalle p esence in he li e -
a u e han es ig based ones [A +18]. This is because he a iabili y o ope a ion
condi ions o WTs a ec s he ex ac ion o indica o s while i specially damages he
sys ems o WTs [Ham+09]. Mos o he wo ks p esen ing eal in-se ice WT da a
a e based on he use o SCADA (Supe iso y Con ol and Da a Acquisi ion) da a,
which is eadily a ailable in gene al. Typically, i is used o compa e pe o mances
among WTs using powe cu es [Gon+19]. Addi ionally, empe a u es om he
SCADA ha e been modelled and compa ed o e ime o use di e ences as ala ms
as he di e en wo ks e iewed by [TW17] show. Howe e , he success o hese
echniques is limi ed [NW13]. Pa ly, because o ex e nal in luences (such as he
ou side empe a u e) ha equi e he ala ms o be manually supe ised by ope a o s
[TW17].
Consequen ly, he inclusion o addi ional CM senso s in ope a ing WTs is lou ishing
[Ga +12]. And an inc easing numbe o wo ks p esen indings om eal use cases,
including some ha show oil deb is senso s (ODS) o he moni o ing o in se ice
WTs as he wo ks o Ga cía Ma quez e al. [Ga +12] and he one by Nie and
Wang [NW13] s a e. As o he wo ks showing da a om in se ice u bines [Dup10;
Fen+13; KDT18] hey ise simila conclusions: cumula i e o a e aged alues a e
needed o a oid he noise caused by he a ying ope a ion, as he ope a ion a ec s
he sensed magni ude. These indings a e also suppo ed by he ex ensi e wo k
ca ied ou by Sheng [She16], we e a ull-scale WT gea box o 750 kW is es ed wi h
in-line and online senso s and samples aken du ing he es s. They men ion he
need o il e ing in luences caused by ope a ional condi ions; ecommend o ocus in
ends ins ead o in absolu e alues, and sugges conside ing big pa icle size (>14
µ
m) indica o s in pa icula . Also, hey iden i y ha damaged gea boxes ha e much
highe deb is gene a ion a es han heal hy ones.
Li e a u e ela ed o wind u bine moni o ing being conside ed, he di icul ies o
gea box diagnosis a e well assessed. Enginee ing a solu ion ha could cope wi h
he complexi ies o he a ying ope a ing con ex s was necessa y, which is, exac ly,
whe e he con ibu ion o his wo k esides.
28 Chap e 4 Finding s able and epea able con ex s in a ying ope a ing con-
di ions
4.3 Inno a ion
The wo k he e p esen ed adds new e idence o he li e a u e by publishing in se ice
s udies o 3 WTs ha a e moni o ed wi h ODSs and shows an app oxima ion used
o deal wi h he a ying condi ions o he u bines wi h he aim o de eloping
diagnos ic algo i hms. These a e he wo majo inno a ions p esen in his wo k:
Fi s ly, how op ical oil deb is senso s beha e in eal ope a ing condi ions is s udied.
In addi ion o ep oducing some moni o ing echniques ound in he li e a u e,
ce ain ac ions o he wind u bines (gene a o b eaking and boos ing) ha migh be
pa icula ly damaging he gea box (as hypo hesised by esea ch in es benches) is
analysed. Fo doing so, machine lea ning echniques a e employed o a oid manual
iden i ica ion o hese ac ions o e he whole da ase . La e , he ela ion be ween
hese ope a ions and he deb is senso s is s udied by analysing he dis ibu ion o
he co ela ion.
Addi ionally, in an a emp o p o ide a heal h indica o o he gea box, he concep
o Finge p in is ex ended o con ex s wi h a ying condi ions. Basically, he p emise
is as ollows: i i is no possible o o ce ce ain epe i i e ope a ion along he ime
o ob ain compa able measu emen s, le s ind which ope a ions a e be e sui ed o
aking measu emen s. In o de o gi e answe o ha p emise, a segmen a ion is
ca ied ou . Se e al ope a ing egions (OR) a e de ined and all ope a ion sequences
(OS) ha belong o each o he 5 compa ed ORs a e iden i ied. Once all his se-
quences (o ope a ion s a es) a e iden i ied, hei app op ia eness o be used as a
basis o a Heal h Indica o (HI) o he gea box is s udied. The aspec s conside ed
a e: he equency, how many OSs occu weekly; he du a ion, mean ime o OSs
and; he s abili y o he a ia ion o he di e en signals du ing he OS. Finally, once
he bes OR is iden i ied, i is possible o de elop a diagnos ic algo i hm ha il e s
undesi ed con ex and only akes alues o s able and epea able con ex s.
4.4 Conclusions
This wo k e eals in e es ing indings. Fi s ly, i is possible o alida e some o he
conclusions ha we e p e iously hypo hesised in es igs, such as he noise ound
in pa icle coun ing senso s [Fen+13; Dup10; She16] and he alidi y o using
cumula i e pa icle a es o educing ha noise [Fen+13; Dup10]. Also, g ea e
di e ences in he ODS han in he es o SCADA signals in damaged u bines we e
4.3 Inno a ion 29
o he ac ua o s. Recen wo ks in his a ea ha e ocused in he de ec ion o damages
and heal h s a us o he ac ua o s in o de o p o ide be e diagnos ic abili ies.
In he wo k ca ied ou by Eh mann, Isabey and Fleische [EIF16], he possible
senso s o moni o ack and pinion senso s a e discussed. They esea ch p e ious
wo ks ela ed o he moni o ing o ack and pinions, and, due o he lack o hem,
hey compa e ack and pinions o simila componen s. Finally, hey de e mine which
senso s could be be e sui ed o hei moni o ing, concluding ib a ion analysis,
posi ion e o and dynamic cha ac e is ics o he con ol loop and acous ic emission
could be well sui ed o moni o ing ack and pinions.
Fe ei o e al. [Sus+13] de elop a physical model o an ac ua o whe e ou
di e en aul s a e seeded: deg ada ion o he mo o , inc ease in ic ion, backlash
and ex e nal o ce. The ac ua ion p ocess is segmen ed and desc ip o s a e ex ac ed,
using hese desc ip o s di e en diangos ic Machinel Lea ning algo ih ms a e ained
and he bes ones (ID3 classi ica ion ee and Bayes Ne bayesian ne wo ks) a e used
as aul indica o s. They ace some di icul ies when diagnosing mild aul s bu he
o e all classi ica ion accu acy is high o se e e cases.
Vol age and cu en signals in ime and ime- equency domain a e used oge he
wi h PCA, op imal ans o ma ion and a suppo ec o machine algo i hm o moni o
an elec omechanical ac ua o by [Kn+15]. Thei algo i hm is capable o de ec ing
s oke ela ed ailu es and educed ol age.
Conside ing elec ical ailu es, in [Cai+16] a da a-d i en bayesian ne wo k based
algo i hm wi h he pu pose o de ec ing ailu es in he in e e s o elec omechanical
ac ua o s is de eloped. They ex ac some signal ea u es using as ou ie ans o m
and he dimensions o he samples a e educed using p incipal componen analysis
(PCA). A combina ion o bo h expe imen al and simula ed da a ha includes sho
ci cui and open ci cui damages wi h di e en combina ions o powe swi ches o
he SPWM is used o de ec ailu es.
An elec omechanical ac ua o used in an unmanned unde wa e ehicle is modelled
in [KM18]. They also model wo addi ional aul s: load aul s and coupling loss aul .
The h ee models (nominal and aul y ones) a e un in pa allel and paying a en ion
o he esiduals hey de e mine whe he he ac ua o is ope a ing co ec ly o i is
damaged. The esiduals hey compa e a e: Cu en , posi ion and angula speed.
An elec ohyd os a ic ac ua o is s udied in [GH14], hey apply a s a e es ima o
e e ed o as SVSF-VBL which is based on he smoo h a iable s uc u e il e (SVSF)
and sliding mode concep s. They conclude ha besides ge ing good es ima es,
SVSF-VBL app oach is alid o de ec ing ailu es.
36 Chap e 5 Ex ending con ex o machine lea ning algo i hms
In he wo k ca ied ou by [SI17] ib a ion senso s a e used o dis inguish aul y
o wo n ball sc ews om o he s in good shape. They design a de ice capable o
ha es ing ene gy om he ac ua o mo emen ha measu es ib a ions a he same
ime. A he end hey a e capable o ecognizing he wo n ac ua o s.
The e a e also some wo ks ha use he combina ion o models (hyb idisa ion) o
diagnosing and p ognosing ailu es in linea ac ua o s. This is he case o he esea ch
s a ed Na asimhan [Na +10] and con inued by Balaban [Bal15]. These wo ks
de elop a comple e condi ion moni o ing sys em o elec omechanical ac ua o s.
They ollow he TRASCEND diagnosis a chi ec u e, whe e a i s diagnos ic laye
de e mines i he ac ua o is de ia ing om i s common beha io , hen a quali a i e
model is igge ed in o de o diagnose he ailu e and is suppo ed by a da a-d i en
model ha imp o es diagnosis accu acy. A e ha , a Gaussian p ocess eg ession
(GPR) model (because he e is no explici aul deg ada ion model) is used o
es ima e he Remaining Use ul Li e (RUL). In hei in dep h s udy hey include
a ious aul modes (jam aul , spall, mo o ailu e, senso aul ), as well as di e en
load p o iles gene a ed simula ing eal ope a ing condi ions.
5.3 Inno a ion
This wo k p esen s an app oach ha would like o co e hose asse s in which he e
is no aul y da a a ailable bu diagnosing he asse s is needed by means o condi ion
moni o ing algo i hms. As he aining con ex o diagnos ic algo i hm is limi ed by
he lack o aul ela ed da a, i is necessa y o expand ha con ex o a wide one
ha includes aul s. This is only possible by ei he damaging he machine in pu pose
(which implies down ime and highs cos s); wai ing un il he machines su e s om
all he possible aul s ( ime consuming and undesi able) o by ec ea ing he physical
beha iou o he aul s wi h a physical model, which is he pa h we ha e ollowed.
A physical model o an elec o-mechanical ac ua o which is a win o he eal
ac ua o is buil . The mos common aul s (as indica ed by he FMECA analysis)
a e seeded in he model and da a eco ds a e c ea ed. Fea u es a e aken om
bo h he eal signals ( he ones om he es - ig) and he syn he ic signals ( he ones
c ea ed by he physical model). Because o he de e minis ic beha iou o he model,
addi ional syn he ic obse a ions a e gene a ed by adding he noise om he eal
desc ip o s.
Due he di icul ies aced du ing he combina ion o bo h da a sou ces, a me hod o
he dele ion o ea u es ha a e oo di e en among da a sou ces is de eloped. This
5.3 Inno a ion 37
way ea u es ha a e inconsis en om one da a sou ce o he o he a e emo ed
so ha he da ase is consis en while he de ec abili y o he aul s is main ained
wi h he emaining ea u es. Fo ha pu pose, he me hod pe o ms a classi ica ion
o aul s using only syn he ic da a and also classi ica ion o da a be ween eal o
syn he ic. This is epea ed i e a i ely emo ing he ea u e ha shows g ea es
di e ences among classes in he dis inc ion be ween eal/syn he ic da a. A he end,
he accu acy o bo h classi ica ions is analysed ( eal/syn he ic and aul s diagnosis)
and he ea u es le a he i e a ion wi h he wo s eal/syn he ic esul s and he
bes aul diagnosis esul s a e kep .
Once he da a a e combined, di e en scena ios a e es ed using Linea Disc iminan
Analysis (LDA) algo i hm. In hese scena ios LDA is ained using non- aul y da a
om he eal ac ua o and he aul y and non- aul y syn he ic da a gene a ed on
he physical model. Among hese scena ios single load cases, mul iple load cases,
class imbalance co ec ions, PCA p ojec ions, he e ec o se e i y and ch onological
p edic ions a e es ed. The alida ion o he model is done by using eal da a which
con ains aul s bu has no been shown o he ML algo i hms be o e.
Addi ionally, i needs special men ion ha he wo k ollows he open esea ch ap-
p oach. The aim o his app oach is o p o ide anspa en , open and ep oducible
esea ch ou comes ha can be c i ically e iewed and eused by o he esea che s.
Pa icula ly, he da a has been made open-access (i was ac ually open access hanks
o C an ield Uni e si y [@RS18a]) and, also, he sou ce code used o he analysis
has been uploaded o an open access eposi o y [@Lop+20].
5.4 Conclusions
In his wo k he possibili y o augmen he lea ning con ex o he algo i hms by
c ea ing syn he ic da ase s wi h physical models is p o ed. Fu he mo e, algo i hms
a e also used o de ec p e iously unseen ope a ing condi ions. Adding o ha
he ac ha be e esul s a e ob ained when c ea ing models ha diagnose in a
single load, he g ea po en ial hyb id modelling has o condi ion moni o ing is
highligh ed.
Consequen ly, he use o digi al- wins/hyb id models o a oid he limi a ions o
indus ial scena ios wi h lack o aul s is alida ed. This is done by showing o da a-
based models how machines beha e unde un egis e ed ope a ions o unde unseen
aul s, which enables he algo i hms o la e de ec hese ope a ions and aul s e en
38 Chap e 5 Ex ending con ex o machine lea ning algo i hms
i he e was no da a ob ained in hese si ua ions. This is he o he in e es ing aspec
o his wo k, ha co e s di e en load cases, which is mo e ep esen a i e o eal
ope a ion condi ions.
Addi ionally, simila ly o o he wo ks in he li e a u e, di icul ies ha e been en-
coun e ed du ing he de elopmen o he physical model. As he model could jus
pa ially esemble he eal beha iou and p oduce ea u es ha do no exac ly esem-
ble he eal ea u es. In ela ion o ha , a me hod ha allows emo ing he ea u es
ha do no ep esen consis en ly he eali y acco ding he measu ed ea u es ( he
ones om he ig) has been de eloped. The esul s sugges he me hod imp o es
signi ican ly he accu acy o he diagnosis algo i hm. Fu he mo e, he me hod is
gene ic enough o be ans e able o o he wo ks a emp ing o hyb idise da a om
physical models and eal ope a ion.
5.5 Publica ions
The esea ch o limi ed con ex s ha e been inco po a ed in he wo k "Hyb id mod-
elling o linea ac ua o diagnosis in absence o aul y da a eco ds". The pape
awai s he e iew by he Jou nal o In elligen Manu ac u ing, he p e-p in e -
sion o he wo k a ached o PART II o he disse a ion. In any case, eade s a e
encou aged o access he analysis which is a ailable in he eposi o y [@Lop+20].
López de Calle, K., Ruiz, C., Fe ei o, S., A naiz, A., Gómez, M., Sie a, B., S a ,
A., 2020 Hyb id modelling o linea ac ua o diagnosis in absence o aul y da a
eco ds 20.
5.5 Publica ions 39
Conclusions 6
He e a e he conclusions o he 3 yea long ad en u e:
RQ 1: Wha is he maximum in o ma ion ha can be ob ained om a s eady
ope a ion con ex ?
Acco ding o he indings, s eady con ex s ha ha e been b oadly add essed in
he li e a u e can be pushed a li le bi u he . Besides moni o ing and de ec ing
anomalies, dimensionali y educ ion algo i hms can be used o gain ex a knowledge
so ha he aul s can be ela ed wi h ce ain ea u es. Pa icula ly, supe ised
DR algo i hms such as LDA a e well sui ed o his ask, as hey can be used
o dis inguish among he ea ly heal hy measu emen s and he ones aken when
anomalies a e de ec ed. This app oxima ion se les he ounda ions o use p ocess
con ol algo i hms in indus ial en i onmen s o ill a da abase wi h in o ma ion
ela ed o aul s and aul indica o s, so ha u u e aul y obse a ions can be
compa ed o he his o ic da abase and hey can be diagnosed. I is concluded ha
DR can be used o unde s and which ea u es a e ela ed o he incoming aul while
he sys em is moni o ed.
S eady ope a ion con ex a e qui e in equen in eali y e en i hey a e one o
he scena ios mo e explo ed in he esea ch li e a u e. As men ioned in chap e 4,
Teknike has expe ience in he de elopmen o me hods o u n a ying ope a ion
in o s eady by meas o Finge p in , bu his app oach has limi a ions, such as he
cases whe e he ope a ion is no commanded by he ope a o . In hese cases s eady
ope a ion egions need o be iden i ied a pos e io which is exac ly wha is a emp ed
in he ollowing Resea ch Ques ion 2.
RQ 2: How can a ying ope a ing con ex s be analysed and s abilised o ease
machine moni o ing?
Chap e 4 has in oduced he eade o a con ex wi h a ying condi ions. In his case,
he a ying ope a ion is gi en by wind, which is s ochas ic and is used o command
he ope a ion o he u bine. In an a emp o ob ain a Heal h Indica o o he
u bine, he ull ope a ion o he u bine has been segmen ed in o ope a ion s a es
conside ing di e en ope a ion c i e ia. In o de o de e mine which c i e ia is bes ,
s eadiness, du a ion, and equency o occu ence o he ope a ion s a es has been
analysed. Once he bes c i e ia is de e mined, only he ope a ion s a es sugges ed
41
by he bes c i e ia a e used o collec da a om hese ope a ion s a es. Finally,
a e il e ing he da a, h esholds p o ided by labo a o y s udies a e es ablished o
de e mine maximum bea able wea alues.
Rega dless o he use case, his p ocedu e can be applied o simila machines whe e
he ope a ion is no s eady bu some ope a ion pa e ns a e epea ed. Va ying condi-
ions can be cancelled by iden i ying and isola ing da a om epea able ope a ing
condi ions ha occu du ing he a ying p ocess. A he same ime his enables o
use he app oach p oposed o s eady condi ions in a ying condi ions. This way
e en complex a ying scena ios can be deal as i hey we e s eady. Ne e heless,
a iabili y o ope a ion is no he only obs acle in he de elopmen o diagnos ic
algo i hms, he e is also he lack o aul y da a, which is exac ly wha RQ 3 ies o
answe .
RQ 3: How can he diagnos ic knowledge o a da a-based algo i hm be ex-
panded in a con ex wi h no aul y da a?
In chap e 5 a scena io wi hou aul y da a is p esen ed in which diagnosing aul s
is desi ed. In o de o expand he a ailable da a o a da a-d i en algo i hm, a
physical model ha ep oduces he ope a ion o he eal machine is buil and aul s
a e seeded in he model. This way i is possible o use he da a om he model
(syn he ic da a) o ain he da a-based algo i hm and, hence, diagnose aul s om
he machine be o e hey e e occu .
This p ocedu e can be used in o he machines i hey can be modelled and he mos
common aul s a e known. In conclusion, i is possible o expand he knowledge o
he da a-based algo i hms by adding new aul y da a eco ds om physical models
in con ex s wi hou eal aul y da a.
In addi ion o inding answe o he esea ch ques ions, o he indings ha need o
be men ioned a e:
The implemen a ion o algo i hms in indus ial scena ios aces addi ional ex e nal
di icul ies as he lack o mu ual unde s anding be ween main enance manage
and moni o ing algo i hm de elope . In his ega d, he wo ks ela ed o chap e
3 p esen some me ics as an a emp o se le a common g ound. These me ics
e lec h ee aspec s ha conce n manage s: cos , e ec i eness and in e p e abili y.
Each o hese me ics summa ises he in o ma ion o o he sub-dimensions, ha
p o ide de ailed in o ma ion o he beha iou o he DR algo i hms unde he hood:
how much ime hey equi e o be compu ed; how de e minis ic he esul s a e;
a ious ways o measu ing accu acy ( ue posi i e a e, posi i e p edic i e alue,
42 Chap e 6 Conclusions
noise de ec ion); he dis ibu ion o he weigh s o he ea u es (how spa se ea u es
a e); whe he hey a e linea combina ions o no ; and he change in ea u e weigh s
om he beginning o he end. This comple e se o sub-dimensions can be used o
be e unde s and how he DR algo i hms wo k. And also, o se a common language
by using concep s unde s ood by bo h he algo i hm de elope and main enance
manage ha can ease aking he decision o choosing " he bes " algo i hm, o e en
simply us ing and elaying on an algo i hm o ake decisions. Which p o es ha
unde s anding o he algo i hms and he needs o be sol ed can be ma ched by a
common g oup o me ics.
Ano he in e es ing inding is ela ed o he app op ia e use o machine lea ning
algo i hms and he bene i s domain knowledge p o ides o ML and ice e sa.
S a ing a chap e 3 whe e domain knowledge is used o enable classi ica ion
algo i hms (olde samples mus be in wo s heal h condi ion); going o chap e
5 whe e subsequen ins ances a e used o o ing and ob aining be e esul s
(consecu i e ins an s mus be in he same damage s age); going h ough chap e
4, whe e, in o de o unde s and whe he WTs migh be c ea ing mo e deb is
du ing b aking and boos ing o no , a ML algo i hm is ained o iden i y hese
ins an s in he da abase (imp o ing he knowledge o he domain); one conclusion
is clea : domain knowledge can be as bene icial o ML as ML can be o ob ain mo e
domain knowledge. The compa ison o ML algo i hms a e qui e equen in he
CM communi y, howe e , i seems like wha makes he eal di e ence is he p ope
in eg a ion o domain knowledge in he p oblem. Addi ionally, ML algo i hms a e
as ly used o aul diagnosis p o iding hem wi h di e en aul s (a ype o da a
which is di icul o ob ain). Ins ead, he e hey a e used o gain mo e insigh om
he da a by iden i ying ce ain pa e ns occu ing in a la ge da abase (a ype o
da a commonly ound in indus y) which could be done manually, bu would ake
insane amoun s o ime. Acco ding o he expe iences he e exposed, ML algo i hms
a e ex emely use ul ools wi h g ea po en ial. And, in ha sense, he challenge
is inding eal p oblems ha can be sol ed wi h his ools ins ead o making up
classi ica ion/ eg ession p oblems so ha he use o hese ools is jus i ied.
Also, du ing he de elopmen o he esea ch and pa icula ly in he las wo k
p esen ed in chap e 5, ep oducible esea ch p ac ices ha e been adop ed. This
includes publishing da a openly o using da a om open eposi o ies, as well as
uploading he code used o he analysis o open access eposi o ies. These p ac ices
a e gaining popula i y and in e es among o he scien i ic ields, as hey allow a
be e e iew o he wo k, and esea ch can be con inued by o he esea che s. I is
no common in condi ion moni o ing ield, howe e , some ecen wo ks ha ollow
hose p ac ices ha e been iden i ied. This is he case o he b illian wo k p esen ed
43
in [Zha+20]. Publishing code and da a is some imes di icul when alking abou
p ojec s ela ed o indus ial pa ne s. Ne e heless, i is ou since e belie ha
i conside ably aises esea ch quali y s anda ds and allows coope a ion among
esea che s.
All in all, he majo con ibu ion o his wo k esides in sol ing no he speci ic use
cases, bu he con ex s o he p oblem. As explained in chap e 2, inding solu ions
o con ex s ins ead han only o he use case implies ha simila con ex s could
use he same app oaches he e p esen ed. Tha is, he knowledge can be ans e ed.
Fo ins ance, dimensionali y educ ion algo i hms could be used o lea n om he
possible sou ces o ailu e o o he machines han o a ing machine y, ega dless o
no ha ing ib a ion senso s as long as hey ope a e in s eady condi ions. Simila ly,
he app oach used o deal wi h a ying condi ions in wind u bines could be expo ed
o o he machines ha also ope a e unde a ying condi ions ( he gea box o a ca
o example) and be used o ob ain a heal h indica o . And, he same happens
o he hyb id app oach used o elec omechanical ac ua o s, ha could be used
in any o he machine ha lacks aul y da a and can be modelled wi h a physical
model. The abili y o ans e he knowledge is pa icula ly in e es ing in he
moni o ing o indus ial asse s, as, many o he di icul ies in he applied scena ios
o his ield a e ela ed o he acquisi ion o ep esen a i e da a wi h good enough
quali y s anda ds and o deal wi h he a iabili y o he ope a ion. This wo k co e s
con ex s wi h a iabili y in he ope a ion, he ob ainmen o knowledge in s eady
moni o ing con ex s and goes one s ep u he by p oposing a way he de elop
diagnosis algo i hms o con ex s wi h no aul y da a. In sho , his hesis p ojec
p o ides a oad-map o he imp o emen o applied moni o ing s a egies om
a ying ope a ing condi ions o he diagnosis o machines.
6.1 Fu u e Wo ks
This hesis p ojec is de ined by he analysis o h ee di e en con ex s occu ing
in h ee di e en applica ions, ha , a he same ime, a e qui e ep esen a i e o
many o he si ua ions and cases ha a e gi en in machine condi ion moni o ing.
On he one hand, he ac ha a ious cases a e analysed lea es li le chance o go
in de ail in each o hem. On he o he , con on ing di e en scena ios p o ides a
be e o e iew o he cu en pa adigm and challenges o condi ion moni o ing.
In any case, some dead ends ha e been iden i ied h oughou he de elopmen o
he hesis ha could be add essed in ollowing wo ks o by o he esea che s.
44 Chap e 6 Conclusions
Rega ding he s eady ope a ing condi ions, s o ing a iables ha migh be o in e es
o dis inguish a ailu e is a good s a ing poin o diagnosis, bu i s ill equi es
o manual o human decision o in e p e incoming new aul s. Au oma ing his
piece o wo k would be o g ea in e es , and, o doing so, algo i hms ha use
ea u es ins ead o obse a ions would be needed. In addi ion, algo i hms ha could
lea n new classes on-line when no simila aul s a e ound in he da abase would be
equi ed.
Fo he case o a ying condi ions, he ope a ion egimes we e manually iden i ied
wi h he help o expe s. Howe e , au oma ed ime-se ies clus e ing me hods could
be used o segmen he dis inc egions in an au onomous way and la e , apply
he same p ocedu e he e p esen ed o measu e equency, du a ion and s abili y o
ope a ion s a es. This clus e ing au oma ion could ex emely bene icial in scena ios
wi h sca ce expe ience in ela ion o he op imal ope a ion egime.
Also, ega ding hyb id modelling, his wo k has s udied how o wo k wi hou ha ing
aul y da a. Ne e heless, his eal aul y da a migh appea in he long e m (e en
i only in he o m o mild o low se e i y aul s), and i could be in eg a ed in o he
models o imp o e diagnosis. These new algo i hms could p obably be based on he
use o inc emen al lea ning algo i hms ha allow pa ial e aining o he models
when new obse a ions a e a ailable.
Finally, a gene al lack o wo ks p esen ing da a om eal indus ial scena ios is
obse ed. Excep ionally, da a om wind u bines ends o be om eal applica ions,
bu in hese cases nei he da a no code a e disclosed. Fu he imp o emen s in
applying and p esen ing algo i hms in o ope a ional en i onmen s a e needed i
condi ion moni o ing esea ch is going o go any u he . Fo ha pu pose, s onge
coope a ion be ween esea che s and indus ies is needed, as well as he will o
disclose and p esen he indings o he esea ch communi y.
6.1 Fu u e Wo ks 45
[TW17]
Jannis Tau z-Weine and Simon J. Wa son. “Using SCADA da a o wind
u bine condi ion moni o ing – a e iew”. In: IET Renewable Powe Gene a ion
11.4 (2017), pp. 382–394 (ci . on p. 28).
[Tch+14]
Pie e Tchakoua, René Wamkeue, Mohand Ouh ouche, e al. “Wind u bine
condi ion moni o ing: S a e-o - he-a e iew, new ends, and u u e chal-
lenges”. In: Ene gies 7.4 (2014), pp. 2595–2630 (ci . on pp. 27, 28).
[Vis+17]
Manish Vishwaka ma, Rajesh Pu ohi , V. Ha shla a, and P. Rajpu . “Vib a ion
Analysis & Condi ion Moni o ing o Ro a ing Machines: A Re iew”. en. In:
Ma e ials Today: P oceedings 4.2 (2017), pp. 2659–2664 (ci . on p. 20).
[Zha+20]
Zhibin Zhao, Tian u Li, Jingyao Wu, e al. “Deep Lea ning Algo i hms o
Ro a ing Machine y In elligen Diagnosis: An Open Sou ce Benchma k S udy”.
en. In: a Xi :2003.03315 [cs, eess] (Ma . 2020). a Xi : 2003.03315 (ci . on
pp. 18, 44).
Webpages
[@G u17]
SPRI G upo. In es in Basque Coun y - S a egic sec o s. en-US. ex.ids:
S a egicSec o s2017a lib a yCa alog: www.sp i.eus. May 2017. URL:
h ps:
/ / www . sp i . eus / in es - in - basque - coun y / en / in es - basque -
coun y/sec o s/ ( isi ed on June 9, 2020) (ci . on p. 7).
[@Lop+20]
Ke man Lopez de Calle – E xabe, C is obal Ruiz – Ca cel, And ew S a , e
al. Hyb id Modelling o Linea Ac ua o Diagnosis in Absence o Faul y Da a
Reco ds. 2020. URL:
h ps://gi hub.com/klope x/Hyb id_EMA_Diagnosis
(ci . on pp. 38, 39).
[@P e08]
Gene al Sec e a ia o Communica ion P esidency o he Go e nmen . Maquina
He amien a. en. Lib a y Ca alog: www.basquecoun y.eus. Ap . 2008. URL:
h ps : / / www . basquecoun y . eus / 32 - sec o es / en / con enidos /
no icia/maquina_he amien a_08/en_ma_he/maquina_he amien a.
h ml ( isi ed on June 9, 2020) (ci . on p. 7).
[@RS18a]
C is obal Ruiz-Ca cel and And ew S a . Da a Se o "Da a-Based De ec ion
and Diagnosis o Faul s in Linea Ac ua o s". Ma . 7, 2018. URL:
h ps://co d.
c an ield.ac.uk/a icles/Da a_se _ o _Da a-based_De ec ion_and_
Diagnosis_o _Faul s_in_Linea _Ac ua o s_/5097649 (ci . on p. 38).
[@Tan19]
John Ola Tande. EERADeepWind’2019. en. 2019. URL:
h ps://www.sin e .
no/globalasse s/p ojec /ee a-deepwind-2019/ -2019_00247_ee a-
deepwind.pd / (ci . on p. 30).
[@Va 18]
Bob Va a. 2018 Main enance Su ey: Playing O ense and De ense. Ma . 15,
2018. URL:
h ps://www.plan enginee ing.com/a icles/2018-main enance-
su ey-playing-o ense-and-de ense/
( isi ed on July 28, 2020) (ci . on
p. 8).
52 Bibliog aphy
Pa II
Pa II: The esea ch
Compa ison o Au oma ed
Fea u e Selec ion and
Reduc ion me hods on he
Condi ion Moni o ing issue
7
•
Au ho s: López de Calle, Ke man; Fe ei o, Susana; A naiz, Ai o ; Sie a,
Basilio.
• Publishe : Else ie
• Jou nal: P ocedia Manu ac u ing
• Yea : 2018
• Qua ile (Scimago/WoS): Q2/-
• DOI: 10.1016/j.p om g.2018.10.150
55
A ailable online a www.sciencedi ec .com
ScienceDi ec
P ocedia Manu ac u ing 00 (2018) 000–000
www.else ie .com/loca e/p ocedia
2351-9789 © 2018 The Au ho s. Published by Else ie B.V.
This is an open access a icle unde he CC BY-NC-ND license (h ps://c ea i ecommons.o g/licenses/by-nc-nd/4.0/)
Pee - e iew unde esponsibili y o he scien i ic commi ee o he 7 h In e na ional Con e ence on Th ough-li e Enginee ing Se ices.
7 h In e na ional Con e ence on Th ough-li e Enginee ing Se ices
Compa ison o Au oma ed Fea u e Selec ion and Reduc ion
me hods on he Condi ion Moni o ing issue
Ke man López de Calle*a,b, Susana Fe ei oa, Ai o A naiza, Basilio Sie ab
aIK4-TEKNIKER, Iñaki Goenaga s ee , 5, 20600 Eiba , Gipuzkoa (SPAIN)
bUni e si y o he Basque Coun y (EHU-UPV), Facul y o In o ma ics, Manuel La dizabal Ibilbidea, 1, 20018 Donos ia, Gipuzkoa (SPAIN)
Abs ac
Condi ion Moni o ing is a key ask o condi ion-based main enance (CBM). The economic e iciency oge he
wi h he eliabili y o CBM is expanding i s use o a eas whe e i was no used be o e. Howe e , his expansion o
CBM sys ems is ballas ed by wo main obs acles: he main enance ela ed knowledge o ecen ly moni o ed asse s,
which is usually sca ce and no s uc u ed; and he lack o ailu e ela ed da a, which is a majo obs acle o da a-
d i en me hods. Wi h he aim o de eloping a me hod o au oma ically selec and/o educe ea u es, his wo k
compa es widely used dimensionali y educ ion and ea u e selec ion me hods, which a e capable o au oma ically
ob aining knowledge while he moni o ing is ca ied ou . Those me hods can la e be used o help ope a o s decide
which ea u es could be mo e ela ed o he deg ada ion o he machine. The pu pose o his wo k is o de e mine
which o hose compa ed me hods could be mo e app op ia e o de ec ing anomalies in o de o de elop moni o ing
sys ems.
© 2018 The Au ho s. Published by Else ie B.V.
This is an open access a icle unde he CC BY-NC-ND license (h ps://c ea i ecommons.o g/licenses/by-nc-nd/4.0/)
Pee - e iew unde esponsibili y o he scien i ic commi ee o he 7 h In e na ional Con e ence on Th ough-li e Enginee ing
Se ices.
Keywo ds: Condi ion Moni o ing; Da a Analy ics; Au onomous Main enance; Au oma ic Fea u e Selec ion; Dimensionali y Reduc ion;
* Co esponding au ho . Tel.: +34-943-206-744-9641;
E-mail add ess: Ke man.lopezdecalle@ eknike .es
57
2 Ke man López de Calle / P ocedia Manu ac u ing 00 (2018) 000–000
1. Backg ound
Condi ion Moni o ing is essen ially he design o a sys em which is able o moni o he condi ion o an asse h ough
senso s. Once he sys em de ec s he condi ion o he asse , i is possible o: i s , de ec anomalies; secondly, diagnose
hese anomalies; and, in some cases, e en p edic hem be o e hey occu .
Some au ho s ecognise ha , o a condi ion moni o ing p ocess o be eliable, i is equi ed o use mo e han a
single signal [1]. O he esea ches ha e shown ha appa en ly less in e es ing signals could be mo e app op ia e o
ce ain pu poses; ha is he case o [2], in which i is s a ed ha he di ec ion wi h lowes ib a ion esul ed mo e
in e es ing o de ec ing wea . A he same ime, he ange o di e en machines ope a ing in indus y and he equi ed
speci ic knowledge o moni o ing hose machines obs uc s he moni o ing ask. On he one hand, some ha e been
as ly s udied (such as mos o a ing machine y), whe eas, on he o he hand, he e a e o he machines which ha e
been less s udied (such as clu ch-b eaks and o he machine ools).
An ideal da a-d i en p ocedu e o condi ion moni o ing would be based on he ollowing schema Fig. 1 (a):
Fig. 1. (a) Ideal da a-d i en moni o ing app oach. (b) App oach in ecen ly moni o ed asse s.
Howe e , mos o he imes he e a e no enough labelled da a, o he quali y is no good enough, which is a e y
common phenomenon in asse s which ha e been ecen ly moni o ed. Fo ha eason, i is some imes necessa y o
lea n ‘on he ly’ and decide which ea u e is mo e ep esen a i e o he condi ion o he asse , as shown in Fig. 1 b).
This wo k ocuses on he cases in which da a becomes a ailable along he ime and he e is no way o ain a egula
classi ie due o he lack o labelled da a.
In o de o de e mine which ex ac ed ea u es a e mo e ela ed o he asse condi ion (Fea u e Selec ion) o how o
combine hem (Fea u e Ex ac ion), some au ho s ha e cons uc ed Heal h Indica o s om ea u es o make alida ions
[3]. Ne e heless, hose indica o s a e ex ac ed a e analysing he whole e olu ion o he ea u e om un- o- ailu e,
which is no applicable in he scope o his s udy.
In his wo k he eali y shown in Fig. 1 b) is simula ed. The simula ion con on s he di icul y o being an un-
supe ised p oblem, as he e is no label de e mining he eal condi ion o he asse . Fo ha eason, he e alua ion o
di e en me hods is no i ial and equi es p ope ly de eloped me ics.
The aim o ea u e selec ion and educ ion echniques is double in hese amewo ks. Fi s , hose me hods allow
de e mining which ea u es a e mo e ele an o he p ocess acco ding o some c i e ia; besides, he dimensionali y
58
Au ho name / P ocedia Manu ac u ing 00 (2018) 000–000 3
educ ion eases he moni o ing ask, as moni o ing a single dimension is less complica ed han acking se e al ea u es
a a ime.
2. Me hodology
As he pu pose o his wo k is o iden i y a echnique capable o moni o ing an asse and o de ec anomalies while
mo e knowledge is ex ac ed du ing he moni o ing p ocess, se e al me hods ha e been compa ed in o de o e alua e
which one pe o ms be e and unde which ci cums ances.
2.1. Simula ion:
Wi h his aim in mind, a simula ion was ca ied ou . The simula ion assumed a newly moni o ed de ice s a ing o
eed a da a eposi o y. A e an ini ial acking pe iod o ime, he s o ed da a was analysed pe iodically by di e en
algo i hms and a new acking dimension was ob ained acco ding o some c i e ia. The new dimension was hen
e alua ed by p ocess con ol limi s and i was decided whe he he sys em was wo king unde no mal o
abno mal/damaged condi ions.
As p e iously explained, he issue could be conside ed unsupe ised due o he lack o labels. Howe e , he addi ion
o a simple assump ion allowed he use o supe ised me hods. The assump ion consis ed o conside ing ha
deg ada ion le el canno be educed wi h he ime (unless main enance ac ions a e ca ied ou ) and, he e o e,
compa ing he mos ecen ins ances wi h he ini ial ones would be simila o ha ing “deg aded” and “non-deg aded”
classes.
2.2. Da a:
In o de o ob ain obus esul s, he o iginal idea was o compa e he algo i hms in di e en scena ios, his wo k
shows, howe e , how he algo i hms beha ed in he moni o ing o bea ings. The bea ing da ase p o ided by [4]
con ained un- o- ailu e es s o bea ings wi h wo ib a ion channels moni o ed h oughou he p ocess.
Th ee di e en s a egies we e analysed in ela ion o he da a eeding o he algo i hms. The i s one consis ed o
eeding all he da a cu en ly a ailable. The second one used he ini ial 20 ins ances and he las 20 ins ances e e y
ime he algo i hms we e e eshed. And he hi d one used 20 ins ances om he end and he emaining 20 ins ances
sampled om di e en ins an s along he p ocess.
The ollowing Table 1 con ains a de ailed explana ion o he da a-pa i ioning s a egies:
Table 1. De ails o da a eeding s a egies.
Da a spli ing
me hod
Takes all da a
De e minis ic
C ea es class
Explana ion
Fi s -las -n40
False
T ue
LDA and Relie
Takes i s n and las n alues in each e eshing.
Sampled-spli
False
False
LDA and Relie
The ‘ ecen ’ n/2 ins ances a e aken om he da a belonging o
he las 10% o he da a. The es is andomly sampled om he
emaining da a ollowing a p obabili y.
Fi s -las -all
T ue
T ue
LDA and Relie
Spli s all a ailable da a in wo, conside ing he i s hal
‘old/undamaged’ and he second hal ‘ ecen ’ ins ances.
The ecalcula ions o he algo i hms ook place e e y ime 5 new ins ances en e ed he sys em, and he e eshing
p ocess began once 50 measu emen s we e s o ed in he simula ed da a- eposi o y.
59
4 Ke man López de Calle / P ocedia Manu ac u ing 00 (2018) 000–000
2.3. Desc ip o ex ac ion:
Raw da a usually consis s o as amoun s o measu emen s aken in e y sho ime in e als. Bu , o ou
moni o ing pu pose, ha da a needed o be somehow comp essed in desc ip o s, which a e alues syn he izing aw
da a by e aining ce ain in o ma ion.
This wo k is in ended o de elop a gene ic moni o ing echnique. Fo ha eason, he desc ip o s used assume no
complex unde s anding o he asse s, e en i i is no he case, as bea ings ha e been s udied in dep h and e y speci ic
desc ip o s ha e been iden i ied in he li e a u e. In o de o simula e ha nai e y, simple ime domain desc ip o s ha e
been used as p oposed by [5][6][7]. The ex ac ed desc ip o s we e: Mean, S anda d De ia ion, Clea ance, Impulse
Fac o , C es Fac o , Ku osis, Peak alue, Roo Mean Squa e, Roo , Shape ac o and Skewness.
2.4. Dimensionali y educ ion:
Despi e he possibili y o moni o ing he e olu ion o each o he desc ip o s independen ly, due o he complexi y
o his app oach in sys ems/asse s wi h a ious signal sou ces, i is usually decided i s o educe he dimensionali y
o a single dimension. This dimension should ep esen he s a e o he asse and i is o igina ed by combining o
p uning he p e iously ex ac ed desc ip o s in a ious ways by ollowing some c i e ia.
The dimensionali y educ ion/ ea u e selec ion me hods s udied in his wo k a e he ollowing:
• PCA: P incipal Componen Analysis is used o ep esen he inpu ea u es wi h new dimensions, wi h he aim
o inding new linea ly un-co ela ed dimensions explaining as much a iance as possible. The implemen a ion
used can be ound in [8].
• Au oencode s: Au oencode s a e Neu al Ne wo ks used o ex ac pa e ns om inpu alues. Thei mid laye s
a e cons ained wi h ewe neu ons han inpu and ou pu laye s so ha , when hey a e ained wi h he same
inpu s as class alues, hey need o ob ain a comp essed pa e n in hei hidden laye s. This comp essed pa e n
can be used o gene a e a new dimension. Implemen ed in R [8] wi h he au oencode package p o ided by
[9].
• LDA: Linea disc iminan analysis uses class alues. I measu es he mean and s anda d de ia ion alues o
each class in o de o la e calcula e he p obabili y o new ins ances o belong o one class o ano he . I c ea es
new p ojec ions wi h C-1 new di e en dimensions, whe e C is he numbe o classes. Implemen ed wi h R
[8] and package MASS [10].
• Relie : This supe ised ea u e selec ion algo i hm gi es weigh s o ea u es in ela ion o how well hey
dis inguish he class om simila (close in dis ance) ins ances belonging o o he classes, and, how di e en
he ea u e is in neighbou s o he same class. Implemen ed in R in he package FSelec o [11].
P io o being ed in o he algo i hms, he da a was i s p e-p ocessed by ex ac ing he mean alue and di iding i
by he s anda d de ia ion o compensa e he scale e ec .
I could be no iced ha mos o he a o emen ioned algo i hms wo k in an unsupe ised manne . Howe e , i is no
he case o he LDA and Relie , which equi e labels o wo k co ec ly. In o de o p o ide labels, some domain ela ed
bias was in oduced, which in his case mean adding a di e en label o he oldes and ano he label o mos ecen
samples. In his way, LDA and Relie would ely on hose desc ip o s showing g ea e di e ences om he beginning
o he end o he p ocess.
2.5. Moni o ing:
Once he da a ha e been p ocessed and he dimensionali y educed, i can be acked in o de o moni o ou sys em.
Moni o ing consis s o acking he e olu ion o p ocess- ela ed ea u es and de ec ing possible de ia ions om
no mali y. Typical moni o ing me hods calcula e he mean o he ea u e and es ablish uppe and lowe bounda ies.
Di e en s a egies a e used o de e mine when he p ocess is ou o con ol.
60
Au ho name / P ocedia Manu ac u ing 00 (2018) 000–000 5
In ou simula ion, uppe and lowe bounda ies we e ecalcula ed each ime he dimensions we e e eshed. In he
calcula ion o hose bounda ies, he ins ances conside ed ou -o -con ol by he las un we e no aken in o accoun .
The p ocess was de ined o be damaged o ou -o -con ol when wo consecu i e ou -o -con ol ins ances we e gi en.
2.6. E alua ion c i e ia:
E alua ing he pe o mance o dis inc algo i hms is di icul in CBM moni o ing du ies. E en i he e is a ela ion
be ween he measu ed signal and he heal h o he asse , ha ela ion is complex and no always empi ically de ined.
This means ha , he eal heal h s a e o he machine is unknown. The e o e, i is di icul o es ablish an objec i e
c i e ion ha de ines how well an algo i hm de ec s anomalies o /and damages due o he lack o eal labels.
Resea che s ag ee wi h he need o ea u e selec ion as a way o imp o ing model pe o mance by means o :
elimina ing noise; p o iding as e and mo e cos -e ec i e models; and, helping o be e unde s and he p ocess ha
gene a ed he da a [12][13]. Ne e heless, di e en esea che s ha e used di e en c i e ia o de ine he quali y o he
ea u es as discussed by [14], wi h some mo e gene ic me hods as p oposed by [13][15][16] and some o he e y case-
speci ic me hods such as [17][18]. Howe e , i was decided o u ilize o he c i e ia which would be e ma ch he
moni o ing pu pose o he algo i hms and he demands o he indus y.
A e some conside a ion, and, aking in o accoun ha he objec i e e alua ion o he dimensionali y
educ ion/ ea u e selec ion is complex, he ollowing c i e ia we e conside ed o e alua ing he ea u e selec ion and
dimensionali y educ ion algo i hms.
• Compu a ional cos : ime equi ed o compu e each ans o ma ion/ educ ion. No e ha i migh a y
using di e en implemen a ions o he algo i hms.
• Unde s andabili y and aceabili y: in e p e abili y o he ou pu ea u e by he ope a o , and capabili y
o ace he ea u es which cons uc ed he new dimension.
• E icacy: de ined as he dis ance o he limi gi en by he expe , o he limi gi en by he algo i hms. In
his case o bea ing moni o ing, he expe decision was o use he oo mean squa e (RMS) alue o he
ib a ion signal. The sco e is a di ision o he dis ance o class di ided by o al asse li e (in ins ances).
• De e mina eness: measu es he a e o change in he ank o weigh s in each e eshmen , and he change
compa ed o ini ial weigh se . Fo ha pu pose, he dis ance me ic o anks p oposed in [19] is used,
which anges om 0 o 1, being 1 absolu e equali y and 0 absolu e dispa i y.
Fig. 2. a) Compu a ional cos wi h andom sampling pa i ioning s a egy. b) Compu a ional cos wi h all-da a pa i ioning s a egy.
61
In
summa y,
al hough
he
inclusion
o
senso
de ices
is
a
g owing
end
in
he
indus y,
da a
ha
ep esen
anomaly
o
ailu e
si ua ions
a e
a ely
ob ained.
Fu he mo e,
he
analysis
and
knowledge
ex ac ed
om
he
use
o
expe imen al
es
benches
o
o he
app oaches
is
di ficul
o
ex apola e
o
he
ac ual
use
case
o
sys ems,
whe e
o he
aspec s
a e
in ol ed,
such
as
he
complexi y
and
he
dynamic
ope a ing
condi ions
among
o he s.
The e o e,
i
is
necessa y
o
implemen
o he
moni o ing
s a egies
ha
allow,
wi hou
p io
knowledge
in
he
domain
o
ope a ion
o
he
equipmen ,
o
s a
moni o ing
and
ex ac ing
ele an
in o ma ion
du ing
i s
li e.
This
documen
p oposes
an
al e na i e
p ocedu e
o
he
cases
in
which
he
analysis
and
he
implemen a ion
o
he
algo i hms
collide
wi h
he
p e iously
men ioned
limi a ions.
Fig.1
ep esen s
he
wo
poin s
o
iew:
Fig.
1(a)
ep esen s
he
app oach
used
when
he e
is
a
ep esen a i e
da ase
o
po en ial
ailu e
modes
o
anomalies,
and
Fig.
1(b)
ep esen s
he
app oach
de ailed
in
his
wo k,
o
be
used
when
he e
is
no
a
p io i
knowledge
abou
he
mos
ep esen a i e
ea u es
o
heal h,
no
a
ep esen a i e
se
o
da a
defining
i s
no mali y
and
no-no mali y
(anomalies
o
ailu es).
In
bo h
o
he
app oaches
p esen ed
in
Fig.
1,
he
moni o ed
asse
is
wo king
while
he
senso s
a e
gene a ing
da a
along
i s
li e ime.
This
da a
is
p ocessed,
some
ea u es
a e
ex ac ed
and,
la e ,
dimensionali y
educ ion
(DR)
echniques
a e
used.
Las ly,
da a
models
a e
used
in
o de
o
de ec ,
diagnose
and
p edic
possible
ailu es.
The
main
di e ence
o
he
app oaches
is
he
ac
ha
ou
app oach
does
no
assume
ha ing
aul y
da a
o
DR
no
o
he
building
o
da a-based
models,
and,
he e o e,
i
is
limi ed
o
de ec ing
anomalies.
Fu he mo e,
he
DR
is
ca ied
ou
conside ing
o he
c i e ia
ha
a e
no
ela ed
o
he
op imiza ion
o
ailu e
de ec ion.
The
es
o
he
a icle
is
o ganized
as
ollows.
Sec ion
2
p esen s
a
e iew
o
p e ious
wo ks
ela ed
o
Condi ion
moni o ing
(CM)
and
he
posi ion
o
he
p esen
wo k
in
ha
con ex .
Sec ion
3
desc ibes
he
p oposed
app oach:
he
simula ion,
ca ied
ou
using
da a
om
eal
use-cases
and
emula ing
an
ope a ing
machine
ha
mus
be
s opped
be o e
he
ailu e;
he
Fea u e
Ex ac ion
echniques,
used
o
desc ibe
he
signals;
he
basis
o
he
DR
algo i hms,
u ilized
o
educe
he
dimension
and
gain
insigh
o
he
p ocess;
he
moni o ing
echniques
ha
define
when
he
p ocess
is
ou
o
con ol;
and
las ly,
he
defini ion
o
he
e alua ion
c i e ia.
Sec ion
4
b iefly
desc ibes
he
expe imen al
se up
used
o
ga he ing
he
da ase s.
Nex ,
Sec ion
5
explains
he
esul s
o
he
analysis
and
he
e alua ion.
And
finally,
Sec ion
6
closes
he
a icle
wi h
he
mos
impo an
conclusions
and
u u e
wo ks.
2.
Li e a u e
e iew
2.1.
Condi ion
moni o ing
Condi ion-based
main enance
(CBM)
is
widely
ex ended
due
o
i s
cos
e ec i eness
and
i s
inc easing
capabili ies
o
imp o ing
p oduc i i y
and
main enance
planning.
The
g ow h
in
he
use
o
CM
is
so
widesp ead
ha
p o ide s
also
seek
o
mode nize
exis ing
equipmen
wi h
CM
capabili y.
The e
is
a
g ea e
numbe
o
machines
whe e
he
da a
is
al eady
a ailable
and,
he e o e,
in elligen
algo i hms
a e
needed
o
ca y
ou
he
analysis
while
conside ing
he
ope a ion
mode
o
he
equipmen .
In
e e ence
o
he
de elopmen
o
algo i hms,
he e
a e
mainly
wo
me hodologies:
da a-d i en
app oaches
and
ma hema ical
o
physical
modeling.
Da a-d i en
app oaches
collec
in o ma ion
om
he
senso s
in
o de
o
ha e
eal
li e
da a
du ing
long
pe iods
o
ime
while
wai ing
o
he
ailu e
o
occu
and
using
his
da a
o
iden i y
ends
o
ea u es
which
co espond
and
cha ac e ize
aul s.
In
some
cases,
such
as
some
mechanical
componen s
o
sys ems
(i.e.,
gea s,
bea ings,
specific
ypes
o
gea boxes,
e c.),
eal
expe imen a ion
can
be
eplaced
by
expe imen a ion
in
he
es
benches
(which
can
be
ela i ely
simila
o
he
machine
and
can
mimic
i s
beha io
o
pe o mance
in
eal
ope a ion).
These
es
benches
allow
he
simula ion
o
specific
ypes
o
ailu es
and
algo i hms
o
hei
de ec ion
o
be
de eloped.
Ma hema ical
o
physical
modeling
allows
he
algo i hm
o
be
de eloped
om
he
fi s
p inciples
and
can
be
used
o
p edic
ailu e
modes
and
hei
e ec
on
measu ed
pa ame e s.
Va ious
wo ks
in
he
li e a u e
ha e
ollowed
one
o
he
o he
me hodology
aiming
he
moni o ing
o
he
condi ion
o
mechani-
cal
componen s.
Rega ding
he
da a-d i en
ones,
in
[1],
a
sys ema ic
me hodology
o
gea box
moni o ing
and
aul
classifica ion
is
de eloped
and
e alua ed
o
a
da ase
o
gea box
ib a ion
da a.
Mo eo e ,
Bechhoe e
and
He
[2]
desc ibe
a
p ocess
o
da a
d i en
p ognos ics
and,
as
an
example,
uses
a
gea
aul
un
o
ailu e
es .
Thei
app oach
uses
he
se
o
ea u es
in o
a
heal h
indica o
h ough
a
s a is ical
p ocess.
Sa a anan
and
Ramachan-
d an
[3,4]
s udied
gea
box
aul
diagnosis
using
decision
ee
classifica ion
and
a ificial
neu al
ne wo ks.
The
ib a ion
signals
o
a
spu
be el
gea
box
in
di e en
condi ions
a e
used
o
demons a e
he
applica ion
o
a ious
wa ele s
in
ea u e
Fig.
1.
(a)
Typical
app oach
o
condi ion
moni o ing.
(b)
App oach
o
his
a icle.
2
K.
López
de
Calle
e
al.
/
Compu e s
in
Indus y
112
(2019)
103114
68
ex ac ion.
Del
Río
e
al.
[5]
p esen s
expe imen al
es ing
and
da a
collec ion
o
heal hy
and
aul y
gea s
o
de ec
and
p edic
ea ly
s ages
o
ailu es
in
a
gea box,
conside ing
ib a ion
signals,
which
can
be
ob ained
om
EMA
(elec omechanical
ac ua o s)
by
means
o
iaxial
accele ome e
in
on-g ound
es ing.
[6]
p oposes
bea ing
moni o ing
( he
mos
ulne able
pa
o
induc ion
mo o s)
based
on
s a o
cu en
signals
p ocessed
wi h
a
deep
lea ning
a chi ec u e
om
an
expe imen al
campaign
ha
a ificially
simula es
bea ing
damage.
Fu he mo e,
B a o-
imaz
e
al.
[7]
ocused
on
he
analysis
o
mo o
cu en
signa u e
o
aul
diagnosis
o
gea boxes
ope a ing
unde
ansien
speed
egimes.
A
specifically
designed
and
ho oughly
moni o ed
es
bench
was
used
o
ully
cha ac e ize
he
expe imen s,
in
which
gea s
in
di e en
heal h
s a us
we e
es ed.
The e
a e
o he
wo ks
ha
adop
physics
based
models,
as
in
[8],
whe e
a
simple
dynamic
model
o
a
single
s age
gea box
is
used
o
show
accu a ely
he
e ec
o
he
o a ing
inpu ,
ou pu
and
mesh
equency
componen s
in
he
s a o
cu en
signa u e
o
induc ion
machine-based
elec omechanical
sys ems.
In
o de
o
alida e
he
p oposed
model,
a
es -bed
based
on
a
wound-
o o
4
kW
h ee-phase
induc ion
machine
connec ed
o
a
gea box
was
used.
Fu he mo e,
o he
au ho s
ha e
used
bo h
echniques,
Li
e
al.
[9]
pe o ms
a
hyb id
aul
diagnosis
o
a
gea box
using
ou
classifie s.
Eigh
aul
s a es,
including
gea
de ec s,
bea ing
de ec s
and
combina ion
o
gea
and
bea ing
de ec s,
a e
simula ed
on
a
single-s age
gea box
o
e alua e
he
p oposed
ea u e
ex ac ion
and
selec ion
scheme.
Ne e heless,
he
use
o
condi ion
moni o ing
is
no
limi ed
o
mechanical
componen s.
Due
o
i s
p o en
benefi s,
i
has
been
ex ended
o
a
wide
a ie y
o
sys ems
and
final
applica ions
as
de ailed
below.
Fo
example,
Dju djano ic
e
al.
[10]
demons a es
he
abili y
o
use
in o ma ion
om
mul iple
senso s
( om
a
CNC
la he
machine)
o
assess
he
d i
away
om
he
sha p
ool
machine
ope a ion
and
o
dis inguish
be ween
a
sha p
and
a
wo n
ool
h ough
expe imen s.
Nex ,
in
[11],
he
algo i hm
based
on
he
usion
o
mul iple
senso
inpu s
and
da a
collec ed
om
expe imen a ion
on
he
welding
machine
is
p esen ed.
I
ma ches
obse ed
sys em
signa u es
agains
hose
obse ed
du ing
i s
no mal
beha iou .
La e ,
Dju djano ic
e
al.
[12]
showed
he
de elopmen
o
an
Agen
o
enable
mul i-senso
assessmen
and
p edic ion
o
pe o mance
o
p oduc s
and
machines
conside ing
some
signal
p ocessing
echniques
and
ea u e
ex ac ion
(such
as
equency-bands
ene gies,
wa ele s,
p incipal
componen s,
e c.),
combined
wi h
a
heal h
assessmen
(based
on
Gaussian
unc ions,
Hidden
Ma ko
Models,
Pa icle
fil e ,
e c.)
om
a
web-enabled
E-
manu ac u ing
es -bed
ealized
by
a
ca
manu ac u e .
Mo e
ecen ly,
Bleakie
and
Dju djano ic
[13]
applied
condi ion
moni-
o ing
and
aul
modeling
o
a
se
o
s anda d
buil -in
senso s
on
a
mode n
300-mm
echnology
indus ial
plasma
enhanced
chemical
apo
deposi ion
(PECVD)
ool.
In
ano he
field
o
applica ion,
Fe ei o
e
al.
[14]
p esen s
a
de elopmen
o
heal h
moni o ing
sys em
o
an
elec o-mechanical
nose-landing
gea
doo
ac ua o
o
an
unmanned
ae ial
ehicle,
based
on
a
combina ion
o
simula ion
modeling
and
da a-d i en
echniques.
The
aim
o
he
wo k
is
o
de ec
some
ailu es
a
ea ly
s ages
o
a oid
a
ca as ophic
aul
ha
may
cause
se ious
damage
o
he
unmanned
ae ial
ehicle
(UAV).
Also,
la e ,
Ruiz-Ca cel
and
S a
[15]
p esen s
a
da a-based
condi ion
moni o ing
ool
o
linea
ac ua o s
es ed
success ully
using
compu a ional
simula ions
and
hen
hey
p opose
in
[16]
a
se
o
algo i hms
ha
makes
use
o
ea u es
ex ac ed
om
he
con olle
h ough
he
analysis
and
expe i-
men a ion
wi h
a
specially
designed
es
ig
o
ec ea e
aul
scena ios
unde
di e en
ope a ion
condi ions.
Howe e ,
o
se e al
easons,
bo h
me hodologies
ha e
disad an ages
ha
complica e
he
ans e
o
he
algo i hms
o
he
final
applica ions.
On
he
one
hand,
he
ypical
da a-based
me hodology
is
ime
consuming
and
he
da a
cap u e
o
all
ailu e
and
ope a ion
modes
is
needed
bu
no
gua an eed.
On
he
o he
hand,
ma hema ical
and
physical
modeling
mus
be
alida ed
h ough
he
combina ion
o
eal-li e
da a
and/o
he
delibe a e
in oduc ion
o
aul s
in o
he
sys em;
howe e ,
his
ask
becomes
expensi e
and
cumbe some
o
la ge
machines
wi h
complex
dynamics.
Fu he mo e,
i
equi es
a
lo
o
expe
knowledge
abou
he
machine
and
p ocesses.
In
addi ion,
he
eal
p oblems
o
he
indus y
go
u he
and
a e
di ficul
o
ackle,
especially
in
small
and
medium-sized
companies
in
manu ac u ing.
These
companies
ha e
new
equipmen
ha
needs
o
be
moni o ed.
The
pe o mance
o
hese
machines
is
unknown
e en
by
he
manu ac u e
and
he e
is
nei he
ime
no
esou ces
o
expe imen a ion.
Each
machine
is
indi idual
wi h
i s
pa icula i ies
and
he e
is
no
gua an ee
ha
he
knowledge
ex ac ed
om
one
machine
can
be
ans e ed
o
ano he
due
o
con ex ual
ac o s.
2.2.
Dimensionali y
educ ion
Rega ding
he
use
o
DR
echniques,
hei
use
is
well
es ablished
in
CM.
DR
echniques
a e
ecognized
ools
o
imp o ing
model
pe o mance
by
means
o :
he
elimina ion
o
educ ion
o
noise,
p o iding
as e
and
mo e
cos -e ec i e
models,
and
helping
o
be e
unde s and
he
p ocess
ha
gene a ed
he
da a
as
explained
in
[17,18].
Also,
applied
o
CM,
he
usion
o
ea u es
has
p o ided
be e
esul s
as
desc ibed
in
[19];
and
some
appa en ly
less
in e es ing
ea u es
ha e
been
demons a ed
mo e
app op ia e
o
ce ain
pu poses
as
e ealed
in
[20].
They
ha e
also
been
used
in
CM
o
cons uc
Heal h
Indica o s
as
p esen ed
in
[21,22].
Fu he mo e,
i
is
ecognized
ha
he
educ ion
o
ea u es
used
o
moni o ing
pu poses
can
be
ansla ed
in
cos
educ ion,
as
i
educes
compu a ional
equi emen s
as
well
as
he
numbe
o
senso s
equi ed
o
ob aining
aw
measu emen s
as
shown
in
[23].
Fo
hese
easons,
he
use
o
DR
echniques
is
widely
ex ended
in
CM,
as
de ailed
in
[16,23].
2.3.
E alua ion
c i e ia
The
e alua ion
o
he
pe o mance
o
dis inc
algo i hms
is
a
di ficul
ma e
in
CM.
E en
i
he e
is
a
ela ion
be ween
he
measu ed
signal
and
he
heal h
o
he
asse ,
ha
ela ion
is
complex
and
no
always
empi ically
defined.
Consequen ly,
he
eal
heal h
s a e
o
he
machine
migh
be
unknown.
The e o e,
i
is
di ficul
o
es ablish
an
objec i e
c i e ion
ha
defines
how
well
an
algo i hm
de ec s
anomalies
o /and
damages
due
o
he
lack
o
eal
labels.
DR
algo i hm
pe o mances
a e
ypically
e alua ed
and
compa ed
by
aining
a
classifie
and
measu ing
i s
accu acy
o
ela ed
me ics
( alse
posi i e
a e,
e o
a e,
e c.)
as
explained
in
[18].
Some
wo ks
go
one
s ep
u he
and
include
addi ional
me ics
such
as
in
[24],
whe e
un ime,
p opo ion
o
selec ed
ea u es
and
a
sensi i i y,
and
confidence
h esholds
a e
included.
In
CM
applica ions
ha
include
DR,
accu acy
and
un ime
ela ed
me ics
p e ail
[23,25–27].
Howe e ,
some
wo ks
men ion
he
need
o
conside
o he
addi ional
indica o s
such
as
powe
consump ion
in
[23].
In
any
case,
accu acy
and
un ime
a e
no
he
only
me ics
ha
eflec
how
well
algo i hms
a e
sui ed
o
he
final
implemen a ion
in
CM
sys ems.
The
e idence
has
shown
ha
in e p e able
models
a e
mo e
accep ed
by
he
use s
acco ding
o
[28].
Fu he mo e,
he e
is
an
inc easing
gene al
eluc ance
o
implemen
non-in e p e able
algo i hms
(also
called
black
box
algo i hms)
eflec ed
in
he
new
egula ions
(GDPR)[29].
Tha
is
why
he
e alua ion
o
CM
algo i hms
mus
include
addi ional
and
specific
me ics
ailo ed
o
i s
final
applica ion.
K.
López
de
Calle
e
al.
/
Compu e s
in
Indus y
112
(2019)
103114
3
69
Focusing
on
he
de elopmen
o
ea ly
s age
moni o ing
algo i hms
o
he
cases
o
new
o
ecen ly
moni o ed
and
less
known
asse s,
his
wo k
es s
a
simula ion
based
on
he
p ocedu e
showed
in
Fig.
1(b)
and
e alua es
he
dimensionali y
educ ion
algo i hms
unde
a
specific
e alua ion
c i e ia
based
on
indus ial
needs.
3.
P oposed
app oach
The
app oach
p esen ed
below
consis s
o
wo
main
pa s
de ailed
in
dep h
in
he
ollowing
sec ions:
Simula ion
and
e alua ion
c i e ia.
Fi s ,
he
simula ion
based
on
he
diag am
o
Fig.
1(b)
is
explained,
om
which
he
pe o mance
o
he
DR
algo i hms
is
measu ed
unde
he
ep oduc ion
o
a
eal
scena io.
Nex ,
he
e alua ion
c i e ia
a e
shown,
i.e.,
he
me ics
based
on
he
mos
ele an
implemen a ion
ac o s.
3.1.
Simula ion
Essen ially,
he
simula ion
consis s
o
aking
un- o- ailu e
da a
om
he
di e en
da a
sou ces
(desc ibed
in
Sec ion
4)
and
emula ing
a
ecen ly
moni o ed
de ice
wi h
his
da a.
The
app oach
his
wo k
ollows
is
p esen ed
in
Fig.
2.
Fi s ly,
an
ini ial
acking
pe iod
o
ime
wi h
no
moni o ing
bu
jus
acking
is
assumed.
A e
ha
pe iod
o
ime,
he
da a
s o ed
un il
cu en
is
analyzed
h ough
Loop
1
and
Loop
2
and,
i
he e
is
no
anomaly
de ec ed,
cu en
is
mo ed
10
ins ances
(simula ing
new
da a
gene a ed
by
he
machine).
This
p ocess
is
epea ed
un il
an
anomaly
is
de ec ed
in
Loop
2
o
he
whole
da ase
has
been
moni o ed
wi h
he
sys em
wi hou
finding
anomalies.
In
mo e
de ail,
he
asse
li e
simula ion
ollows
hese
s eps:
S ep
(A)
P ocess
simula ion:
The
da a
un il
cu en
is
spli
in o
wo
windows
(fi s
and
las
window)
ha
keep
equal
size
and
include
he
fi s
25
and
he
las
(un il
cu en
)
25
ins ances
o
he
p ocess.
S ep
(B)
Label
cons uc ion:
Labels
a e
gi en
o
he
windows
coming
om
p ocess
simula ion.
This
is
done
by,
assuming,
om
he
domain
knowledge,
ha
he
measu emen s
aken
in
pos e io
ins an s
mus
be
equally
o
mo e
deg aded,
bu
ne e
less
han
he
ones
aken
a
he
beginning.
The e o e,
non-
deg aded
class
is
gi en
o
ini ial
eco ds
(fi s
window)
and
deg aded
label
o
mos
ecen
eco ds
(las
window).
This
app oxima ion
allows
he
use
o
supe ised
me hods
o
dimensionali y
educ ion.
S ep
(C)
Dimensionali y
educ ion:
DRalgo i hmsa eusedwi h he
da a
om
he
p e ious
s ep.
As
a
esul ,
a
ea u e
ans o ma ion
and
weigh s
ela ed
o
ea u e
alues
a e
gene a ed.
Weigh s
a e
s o ed
(see
S ep
D)
and
he
pa ame e s
used
in
he
DR
a e
ans e ed
o
he
nex
s ep.
This
phase
is
explained
in
mo e
de ail
in
Dimensionali y
educ ion
algo i hms
(DR)
sec ion.
S ep
(D)
Weigh
s o ing:
The
weigh s
ob ained
in
he
aining
phase
o
he
algo i hms
a e
s o ed
in
o de
o
s udy
hem
wi h
he
me ics
explained
in
Sec ion
3.2.
S ep
(E)
Signal
econs uc ion:
Using
he
pa ame e s
om
he
DR
s ep
and
he
whole
signal
da a
un il
cu en
,
he
signals
a e
p ojec ed
in
he
new
and
single
dimension
(in
he
case
o
Fea u e
P ojec ion
algo i hms)
o
jus
he
bes
ea u e
is
le
(in
he
case
o
Fea u e
Selec ion
algo i hms).
S eps
(F)–(I)
Moni o ing(Loop
2):
The
p ojec ed
ea u e
is
acked
and
i
is
decided
whe he
he
sys em
is
unde
con ol
o
i
is
ou
o
con ol.
Mo e
de ails
o
hese
s eps
a e
gi en
in
he
sec ion
Moni o ing.
Fig.
2.
Schema
o
he
simula ion.
4
K.
López
de
Calle
e
al.
/
Compu e s
in
Indus y
112
(2019)
103114
70
S ep
(J)
P ocess
con inua ion:
I
he
p ocess
is
no
ou
o
con ol,
he
simula ed
machine
should
keep
wo king,
cu en
is
inc eased
by
10
da a
ins ances
and
Loop
1
is
igge ed
again.
Each
algo i hm
has
been
es ed
sepa a ely,
s o ing
all
he
weigh s
and
p ojec ions
c ea ed
along
he
asse
li e
simula ion.
No e
ha
as
di e en
algo i hms
a e
es ed,
he
p ojec ions
sugges ed
by
each
algo i hm
can
lead
o
dispa a e
esul s
ega ding
when
he
machine
should
s op
wo king.
3.1.1.
Fea u e
ex ac ion
(FE)
Ins ead
o
using
aw
da a
measu emen s
o
he
moni o ing,
some
s a is ical
desc ip o s
a e
ypically
used.
These
desc ip o s
a e
gene a ed
as
a
esul
o
a
ea u e
ex ac ion
p ocess
which
ends
o
be
specific
acco ding
o
he
senso
p oducing
he
da a
s eam.
In
ou
pa icula
da ase s,
consis ing
o
ib a ion,
mic ophone,
cu en
in ensi y
and
encode
measu emen s;
bo h
simple
s a is ical
desc ip o s
as
well
as
mo e
specific
ones
we e
ex ac ed.
Rega ding
he
simple
ime
domain
s a is ical
desc ip o s,
he
ones
p oposed
in
[21,30,31]
we e
ex ac ed
in
he
ime
domain
aw
da a.
The
ex ac ed
desc ip o s
we e:
Mean,
S anda d
De ia ion,
Clea ance,
Impulse
Fac o ,
C es
Fac o ,
Ku osis,
Peak
alue,
Roo
Mean
Squa e,
Roo ,
Shape
ac o
and
Skewness.
Besides,
some
mo e
specific
desc ip o s
we e
ex ac ed
o
he
ib a ions
and
he
mic ophones.
In
he
case
o
ib a ions,
equency
domain
s a is ics
we e
used.
F om
he
fi s
fi e
ha monics
o
he
gea
mesh
equency
10
Hz
and
30
Hz
windows
we e
aken
and
he
maximum
and
he
RMS
alues
we e
ex ac ed
simila ly
o
[32].
Also,
o
he
mic ophone,
he
ene gy
index
(EI)
p oposed
in
[33]
was
calcula ed
using
ou
di e en
powe
alues
(1,
3,
5,
10).
De ailed
explana ion
o
he
ea u es
wi h
mo e
complexi y
is
gi en
in
[34].
No
all
desc ip o s
we e
ex ac ed
in
bo h
da a
sou ces,
mo e
de ails
ega ding
he
da a
sou ces
and
desc ip o s
can
be
ound
in
Sec ion
4.
Las ly,
be o e
using
he
desc ip o s
wi h
he
DR
algo i hms,
he
desc ip o s
we e
p e-p ocessed
by
ex ac ing
he
mean
alue
and
di iding
hem
by
he
s anda d
de ia ion
o
compensa e
he
scale
e ec .
3.1.2.
Dimensionali y
educ ion
algo i hms
(DR)
Assuming
ha
basing
a
CM
sys em
on
a
single
ea u e
in
un easible
[19],
he
ole
o
DR
echniques
is
wo old
in
his
wo k.
Fi s ,
hese
me hods
de e mine
which
ea u es
a e
mo e
ele an
o
he
p ocess
acco ding
o
some
c i e ia
in
a
on-line
way;
besides,
hey
ease
he
moni o ing
ask,
as
moni o ing
a
single
dimension
is
less
complica ed
han
acking
se e al
ea u es
a
a
ime.
Two
amilies
o
DR
algo i hms
a e
used
in
his
wo k:
ea u e
selec ion
and
ea u e
p ojec ion.
Fea u e
selec ion
(FS)
algo i hms
iden i y
he
ea u es
ha
maximize
he
ela ion
among
he
inpu
a iables
and
he
a ge
ea u e
(a.k.a
class)
o
jus
educe
he
numbe
o
ea u es
by
educing
he
amoun
o
edundancy
among
he
a iables.
Fea u e
p ojec ion
(FP)
algo i hms,
ins ead
o
choosing
among
he
inpu
a iables,
c ea e
new
mappings
o
he
ea u es.
Each
algo i hm
has
i s
own
c i e ion
o
op imize
he
a iable
selec ion/ educ ion.
In
addi ion
o
he
p e ious
classifica ion
o
DR
echniques,
hese
echniques
can
also
be
classified
ega ding
he
use
hey
make
o
he
class
a iable.
When
hey
need
a
a iable
o
in e es ,
hey
a e
called
supe ised.
Con a ily,
i
hey
do
no
equi e
he
class
a iable
o
he
mapping
o
selec ion,
hey
a e
known
as
unsupe ised.
This
wo k
has
u ilized
he
ollowing
algo i hms,
all
o
hem
implemen ed
in
he
R
so wa e
[35]:
PCA:
P incipal
Componen
Analysis
is
used
o
ep esen
he
inpu
ea u es
wi h
new
dimensions.
These
new
dimensions
a e
linea ly
un-co ela ed
and
explain
as
much
a iance
as
possible.
The
implemen a ion
used
can
be
ound
in
[35]
based
on
he
desc ip ion
o
[36].
Au oencode s
(AE):
Au oencode s
a e
A ificial
Neu al
Ne wo ks
(ANN)
used
o
ex ac
pa e ns
om
inpu
alues.
Thei
mid-
laye s
a e
cons ained
wi h
ewe
neu ons
han
inpu
and
ou pu
laye s
so
ha ,
when
hey
a e
ained
wi h
he
same
inpu s
as
class
alues,
hey
need
o
ob ain
a
comp essed
pa e n
in
hei
hidden
laye s.
This
comp essed
pa e n
can
be
used
o
gene a e
a
new
dimension.
The
echnique
is
well
explained
in
[37],
i
is
implemen ed
wi h
he
au oencode
package
p o ided
by
Dubossa sky
and
Tyshe skiy
[38].
LDA:
Linea
disc iminan
analysis
uses
class
alues.
I
measu es
he
mean
and
s anda d
de ia ion
alues
o
each
class
in
o de
o
la e
calcula e
he
p obabili y
o
new
ins ances
belonging
o
one
class
o
ano he .
I
c ea es
new
p ojec ions
wi h
C-1
new
di e en
dimensions,
whe e
C
is
he
numbe
o
classes.
The
p oposed
new
dimensions
y
o
maximize
he
di e ence
be ween
classes.
I
is
implemen ed
using
he
MASS
package
[36]
wi h
he
desc ip ion
o
[36].
Relie :
This
supe ised
ea u e
selec ion
algo i hm
gi es
weigh s
o
ea u es
in
ela ion
o
how
well
hey
dis inguish
he
class
om
simila
(close
in
dis ance)
ins ances
belonging
o
o he
classes,
and
how
di e en
he
ea u e
is
in
neighbou s
o
he
same
class.
I
is
implemen ed
using
he
FSelec o
package
[39]
based
on
he
desc ip ion
o
Kononenko
[40].
I
should
be
no ed
ha
some
o
he
a o emen ioned
algo i hms
wo k
in
an
unsupe ised
manne .
Howe e ,
i
is
no
he
case
o
he
LDA
and
Relie ,
which
equi e
labels
o
wo k
co ec ly.
In
o de
o
p o ide
labels,
some
domain
ela ed
bias
was
in oduced,
as
explained
in
Simula ion
sec ion.
In
addi ion,
i
mus
be
men ioned
ha
nei he
PCA
no
Relie
educe
dimensions
by
hemsel es,
wha
hey
do
is
gi e
a
anking
o
ea u e
impo ance
om
which
only
he
bes
ea u e
was
la e
used.
Table
1
classifies
he
algo i hms
used
in
his
a icle
acco ding
o
he
classifica ion
explained
p e iously
in
his
sec ion.
No e
ha
each
o
he
p e ious
algo i hms
uses
weigh s
in
o de
o
assess
he
impo ance
o
he
ea u e
o
he
co esponding
model.
The
weigh s
do
no
ha e
equal
meaning
om
one
algo i hm
o
o he ,
howe e ,
hey
do
ep esen
he
impo ance
he
algo i hm
gi es
o
each
pa icula
ea u e
in
compa ison
o
he
o he s.
In
he
pa icula
case
o
au oencode s,
ha
do
no
assign
impo ance
o
he
a iables,
he
weigh s
a e
equi alen
o
he
weigh s
o
he
fi s
laye ,
as
i
is
he
only
connec ion
o
he
a iables
o
he
laye
wi h
he
single
neu on.
Fo
he
aining
o
he
algo i hms
he
de aul
pa ame e s
o
he
implemen a ion
a e
used
unless
o
he
o
he
case
o
au oencode ,
in
which
he
mid
laye
has
single
neu on
wi h
anh
ac i a ion
unc ion,
lambda
(weigh
decay)
is
se
o
0.002,
be a
(spa si y
penal y)
is
equal
o
6,
ho
(spa si y
pa ame e )
0.01
and
epsilon
(weigh
ini ializa ion)
0.001
a e
used.
3.1.3.
Moni o ing
Once
he
da a
is
p ocessed
and
he
dimensionali y
is
educed,
he
new
dimensions
a e
used
in
o de
o
moni o
he
sys em
as
Table
1
Classifica ion
o
he
algo i hms.
Fea u e
p ojec ion
Fea u e
selec ion
Supe ised
LDA
Relie
Unsupe ised
PCA,
AE
K.
López
de
Calle
e
al.
/
Compu e s
in
Indus y
112
(2019)
103114
5
71
displayed
in
Fig.
2
Loop
2.
Moni o ing
consis s
o
acking
he
e olu ion
o
p ocess- ela ed
ea u es
and
de ec ing
possible
de ia ions
om
no mali y.
As
his
s ep
is
ca ied
ou
a e
he
DR
s ep,
and
only
a
single
ea u e
is
kep ,
adi ional
S a is ical
P ocess
Con ol
sys ems
can
be
used.
Moni o ing
is
ca ied
ou
by
using
a
Shewha
cha
[41].
In
s ep
F,
a
window
equal
o
he
ini ial
acking
pe iod
o
ime
(as
in
s ep
A)
is
aken.
The
mean
(m)
and
h eshold
alues
(3
s)
o
he
Shewha
cha
a e
calcula ed
(s ep
G).
I
any
fi e
consecu i e
ou -o -con ol
(OOC)
poin s
a e
ound,
an
anomaly
is
diagnosed,
and
he
sys ems
s ops
he
p ocess
(s ep
H).
O he wise,
i
he
sys em
is
unde
con ol,
Loop
2
con inues
by
adding
1
ins ance
o
he
ini ial
window
(s ep
I)
un il
ei he
he
sys em
is
s opped
o
cu en
is
eached.
I
cu en
is
eached
and
no
anomaly
is
de ec ed,
he
whole
p ocess
con inues
in
s ep
J
by
adding
mo e
da a
poin
o
he
p ocess
simula ion.
3.2.
E alua ion
c i e ia
The
simula ion
con on s
he
di ficul y
o
being
an
unsupe -
ised
p oblem,
as
he e
is
no
label
de e mining
he
eal
condi ion
o
he
asse .
Fo
ha
eason,
he
e alua ion
o
di e en
me hods
is
no
i ial
and
equi es
p ope ly
de eloped
me ics.
Wi h
he
in en ion
o
designing
a
gene ic
compa ison
amewo k
o
indus ial
applica ion
cases,
he
p oposed
solu ion
has
aken
in o
accoun
he
ollowing
h ee
ac o s:
E ficacy:
Sa isfies
he
desi ed
esul s
(de ec s
he
anomalies).
In e p e abili y:
The
ou pu s
a e
unde s andable.
E ficiency:
Wi h
he
ocus
on
op imizing
esou ce
needs.
The
es
o
he
sec ion
is
dedica ed
o
explaining
in
de ail
he
exac
me ics
used
o
each
dimension.
3.2.1.
E ficiency
T ansla ing
cos
o
indus ial
needs
is
synonymous
wi h
o
money
and
ime.
In
he
pa icula
case
o
so wa e
p oduc s,
he
compu a ion
ime
i sel
is
ela ed
o
bo h,
as
i
de e mines
he
ha dwa e
equi emen s
and
he
highes
equency
a e
o
decision
aking.
This
dimension
has
been
ep esen ed
by
wo
sub-
dimensions:
Compu a ion
ime:
Time
equi ed
by
he
algo i hm
o
ob ain
he
new
dimensions.
In
o de
o
o ce
0–1
ange,
he
in e se
alues
o
he
minimum
and
maximum
imes
equi ed
by
he
compa ed
algo i hms
a e
used.
De e mina eness
om
p e ious:
De e mina eness
is
defined
in
his
wo k
as
he
di e ence
o
he
weigh s
om
one
ins an
o
ano he .
In
o de
o
measu e
ha
change,
a
dis ance
me ic
o
anks
p oposed
in
he
wo k
[42]
and
implemen ed
in
he
gespeR
package
[43]
has
been
used.
This
me ic
also
anges
om
0
o
1,
whe e
1
means
absolu e
equali y
and
0
absolu e
dispa i y.
De e mina eness
has
been
used
in
wo
di e en
sub-dimensions,
in
he
case
o
de e mina eness
om
p e ious,
i
s ands
as
he
a e age
dis ance
o
he
weigh s
in
espec
o
he
weigh s
o
he
p e ious
ins an s
o
each
o
he
ecalcula ions.
The
idea
o
including
his
me ic
in
he
cos
esides
in
he
ac
ha
he
slowe
(highe
de e mina eness
om
p e ious)
he
weigh s
change
om
he
p e ious
i e a ion,
he
easie
i
is
o
educe
he
calcula ion
equency
wi hou
losing
in o ma ion.
3.2.2.
In e p e abili y
This
second
dimension
akes
in o
accoun
he
amoun
o
a iables
used
by
he
models,
whe he
o
no
he
models
equi e
a
ma hema ical
mapping
and
how
much
he
models
change
he
alues
o
he
weigh s:
Spa si y:
This
s ands
o
he
ac ion
o
spa se
weigh s
o
an
algo i hm.
As
he
whole
ea u e
space
has
been
educed
o
he
fi s
selec ed
ea u e
o
componen ,
a
spa se
weigh
(and
he e o e
ea u e
no
con ibu ing
o
ha
ea u e/componen )
is
conside ed
a
weigh
which
is
10
imes
smalle
han
he
maximum
weigh
alue
o
ha
algo i hm.
Comp ehensi eness:
Spa si y
conside s
he
whole
ou pu
weigh s
o
he
dimensionali y
educ ion
weigh s.
Howe e ,
he
dimen-
sion
used
o
acking
is
a
single
ea u e
in
he
case
o
FS
algo i hms,
in
con as
o
dimensionali y
educ ion
echniques
ha
use
a ious
signals
in o
a
single
one.
Fo
ha
eason,
a
bina y
me ic
has
been
c ea ed
gi ing
1
alues
o
comple e
comp ehensi eness
and
0
o
hose
me hods
using
ma hema ical
mappings
(Fea u e
P ojec ion
me hods),
as
hei
aceabili y
is
mo e
complex.
A e age
de e mina eness:
This
me ic
a e ages
he
de e mina e-
ness
o
weigh s
in
he
fi s
ins an
o
he
weigh s
in
he
ou -o -
con ol
ins an ( e e ed
o
de e mina eness
om
ini ial)
wi h
he
de e mina eness
om
p e ious
( he
same
me ic
as
he
p e ious
dimension).
This
a e aged
me ic
shows,
on
he
one
hand,
how
much
he
ans o ma ion
has
e ol ed
om
he
beginning
(0
meaning
o ally
di e en
ank
and
1
meaning
no
e olu ion)
and,
on
he
o he
hand,
how
e a ic
he
ans-
o ma ions
a e
(i
de e mina eness
om
p e ious
is
high,
i
means
he e
is
a
small
change,
whe eas
i
i
is
big,
i
s ands
o
e a ic
beha iou ).
3.2.3.
E ficacy
Conside ing
a
condi ion
moni o ing
amewo k,
e ficacy
has
been
conside ed
as
he
capabili y
o
an
algo i hm
o:
de ec
anomalies;
gi ing
he
smalles
numbe
o
alse
posi i es;
and,
a
he
same
ime,
ex ac
as
much
aluable
in o ma ion
as
possible.
This
dimension
has
been
dissemina ed
in
h ee
di e en
sub-
dimensions
ha
measu e
how
accu a e
p edic ions
a e
and
how
well
noise
is
de ec ed
and
elimina ed.
Due
o
he
di ficul ies
o
assess
he
eal
heal h
s a e
o
an
asse ,
expe
knowledge
was
conside ed
o
defining
he
g ound
u h
ea u e
o
each
da ase
(see
Sec ion
4)
and
compa ed
o
he
esul s
p o ided
by
he
algo i hms.
T ue
posi i e
a e
(TPR):
This
me ic
ep esen s
he
ac ion
o
cases
ha
a e
abno mal
(posi i e)
acco ding
o
he
expe
ea u e
and
a e
also
conside ed
abno mal
by
he
algo i hm,
di ided
by
he
o al
numbe
o
eal
posi i e
cases.
Posi i e
p edic i e
alue
(PPV):
This
me ic
assesses
which
ac ion
o
he
posi i e
p edic ions
is
eally
posi i e.
Tha
is,
om
all
he
da a
ins an s
ha
a e
ound
ou
o
con ol
by
he
es ed
algo i hms,
how
many
a e
eally
ou
o
con ol
acco ding
o
he
expe
ea u e.
Noise
de ec ion
(ND):
In
o de
o
es ima e
he
denoising
capabili y
o
he
dimensionali y
educ ion
algo i hms,
ou
syn he ic
a iables
ha e
been
c ea ed
and
added
o
he
da ase .
Fo
he
c ea ion
o
each
syn he ic
a iable,
he
ange
o
a
andomly
chosen
a iable
is
measu ed,
and
hen
his
ange
is
used
o
c ea e
andom
uni o m
noise.
Noise
de ec ion
me ic
akes
0–1
alues,
0
being
none
o
he
syn he ic
a iables
a e
iden ified
as
noise
(meaning
weigh
was
no
spa se)
and
1
being
he
ou
syn he ic
a iables
a e
iden ified
as
noise
(had
spa se
weigh s).
No e
ha
all
me ics
ha e
0
o
1
ange
and
he e o e
can
be
easily
a e aged
o /and
compa ed.
Each
dimension
is
compu ed
by
a e aging
he
co esponding
sub-dimensions.
6
K.
López
de
Calle
e
al.
/
Compu e s
in
Indus y
112
(2019)
103114
72
4.
Expe imen al
se up
Fo
he
sake
o
obus ness,
wo
di e en
da a
sou ces
ha e
been
used
o
he
simula ions
and
e alua ions.
Bo h
sou ces
we e
o igina ed
in
es
igs,
bu
wi h
he
pu pose
o
es ing
di e en
componen s.
As
he
simula ions
a e
based
on
hese
da ase s
ha
a e
aken
unde
s eady
ope a ing
condi ions,
he
load
and
speed
a ia ions
(o
any
o he
ac o s
influencing
signals
du ing
ope a ion)
a e
no
conside ed.
Da ase
1:
Bea ings
This
da ase
p o ided
by
Nec oux
e
al.
[44]
is
open
an
con ains
un- o- ailu e
es s
o
bea ings
wi h
wo
ib a ion
channels
moni o ed
h oughou
he
p ocess.
Fo
he
moni o ing
pu pose,
some
desc ip o s
a e
ex ac ed
om
he
aw
da a.
The
se
o
ime
domain
s a is ical
desc ip o s
explained
in
Fea u e
ex ac ion
(FE)
sec ion
has
been
used.
In
his
da ase ,
oo
mean
squa e
(RMS)
alues
o
accele ome e s
ha e
been
used
as
he
eal
heal h
indica o s
o
e alua ing
he
e ficacy.
Da ase
2:
Gea s
This
da ase
is
p i a e
and
has
been
gene a ed
in
IK4-TEKNIKER
[34].
I
was
gene a ed
in
a
FZG
es
ig,
which
is
a
s anda dized
wo k
bench
well
known
o
spu
gea
es ing.
The
o iginal
pu pose
o
he
es
was
o
assess
he
e olu ion
o
mic o-pi ing
and
pi ing,
which
a e
kinds
o
wea
gene a ed
on
he
su ace
o
he
gea s.
To
his
e ec ,
he
es
had
o
be
s opped
pe iodically
in
o de
o
measu e
he
wea .
The e o e,
in
addi ion
o
he
signals
moni o ed
by
he
senso s,
some
su ace
deg ada ion
measu es
we e
made
a ailable.
The
es
ig
had
di e en
da a
sou ces
in eg a ed:
3
axial
accele ome e s,
a
mic ophone,
cu en
in ensi y
me e s
and
an
encode .
Fo
each
one
o
he
senso s,
he
same
ime
domain
s a is ical
desc ip o s
ha
we e
aken
o
he
bea ings
da ase
we e
ex ac ed.
Addi ionally,
he
mo e
specific
desc ip o s
desc ibed
in
Sec ion
3.1.1
(FE)
we e
also
aken
o
ib a ions
(ha monics)
and
he
mic ophone
(ene gy
index).
In
he
case
o
gea s
da ase ,
as
he e
a e
some
mic o-pi ing
measu emen s
in
addi ion
o
he
eco ded
signals,
an
indica o
was
c ea ed
using
he
measu emen s
o
mic o-pi ed
su aces
o
bo h
wheel
and
pinion
gea s.
The
measu emen s
o
bo h
su aces
was
a e aged
wi h
a
RMS
alue
and,
as
he e
a e
less
deg ada ion
eco ds
han
senso
measu emen s,
linea i y
has
been
assumed
( om
each
measu ed
mic o-pi ed
ins an
o
he
nex
one)
o
impu e
deg ada ion
alues
o
he
es
o
he
senso
measu emen s.
Wi h
he
help
o
a
co ela ion
analysis
and
an
expe ,
he
mos
app op ia e
ea u es
we e
chosen.
The
ea u es
shown
in
Table
2
ha e
been
used
in
each
case
as
heal h
indica o s.
Fig.
3
shows
an
example
o
bes
ea u e
used
as
heal h
indica o
oge he
wi h
he
in e pola ed
deg ada ion
a iable.
Table
2
Heal h
indica o s
by
gea .
Gea
Class
ea u e
Tes
5
Accele ome e
X
H2
Max
Tes
14
Accele ome e
Y
H1
RMS
10
Hz
Tes
18
Accele ome e
Y
C es
Fig.
3.
Le
side:
Example
o
bes
ea u e
acco ding
o
he
expe
in
Gea /Tes
5.
Righ
side:
RMS
o
wheel
and
pinion
deg aded
su ace
alues
wi h
linea
in e pola ion
in
he
gaps.
Table
3
Cha ac e is ics
o
he
di e en
da ase s.
Da ase
Fea u es
Leng h
Gea
5
141
263
Gea
14
141
558
Gea
18
141
543
Bea ing
1.1
22
466
Bea ing
1.2
22
144
Bea ing
2.1
22
151
Bea ing
2.2
22
797
Bea ing
3.1
22
85
Bea ing
3.2
22
1637
Table
4
A e age
o
sub-dimension
esul s
by
da a
sou ce.
Da a
sou ce
Model
E ficiency
In e p e abili y
E ficacy
Comp.
E .
De .
p e .
Spa si y
Comp .
A g
De .
TPR
PPV
ND
Bea ings
Au oencode
0.937
0.716
0.583
0
0.638
0.233
0.243
0.667
LDA
0.995
0.815
0.885
0
0.780
0.615
0.648
1.000
PCA
1.000
0.877
0.359
0
0.727
0.558
0.508
0.667
Relie
0.000
0.667
0.295
1
0.559
0.792
0.778
0.625
Gea s
Au oencode
0.987
0.695
0.870
0
0.580
0.000
0.000
0.833
LDA
0.999
0.646
0.956
0
0.598
0.333
0.333
0.917
PCA
1.000
0.839
0.475
0
0.549
0.333
0.333
0.833
Relie
0.000
0.721
0.275
1
0.495
0.333
0.333
0.833
K.
López
de
Calle
e
al.
/
Compu e s
in
Indus y
112
(2019)
103114
7
73
As
he
du a ion
o
he
es s
a ied
in
each
pa icula
case
because
o
he
dependence
on
how
long
he
bea ings/gea s
needed
o
b ake/ge
pi ing,
he
da ase s
ha e
di e en
leng hs.
The
final
dimensions
o
he
da ase s
a e
displayed
in
Table
3.
5.
Resul s
and
discussion
A e
ca ying
ou
he
simula ions,
he
esul s
ha e
been
e alua ed
acco ding
o
he
sub-dimensions
and
dimensions
explained
in
Sec ion
3.2.
Comple e
esul s
and
agg ega ed
ables
o
each
case
a e
displayed
in
he
appendix
(Tables
A.5–D.8).
Table
4
shows
he
mean
alues
o
he
di e en
algo i hms
in
each
da a
sou ce.
O e all,
he
algo i hms
end
o
beha e
simila ly
om
one
case
o
use
case
o
ano he ,
showing
g ea e
di e ences
in
E ficacy
me ics.
TPR
and
PPV
a e
he
mos
changing
sco es,
and
in
bo h
cases
hey
ob ain
low
alues,
penalizing
specifically
Au oencode s.
O he
in e es ing
poin s
a e
he
huge
di e ences
ha
Relie
has
wi h
he
es
o
he
algo i hms
ega ding
Compu a ional
E ficiency.
This
happens
because
o
he
ob en-
ion
o
he
me ic,
as
he
me ics
has
0–1
ange
which
is
achie ed
by
scaling
elapsed
ime
o
ha
ange using
heminimum
Fig.
4.
A e aged
main
dimension
sco es
by
algo i hm
and
case
o
use.
Fig.
5.
Rada
cha s
ep esen ing
a e aged
sub-dimension
sco es
o
all
cases
o
use
o
each
algo i hm.
(a)
Au oencode ,
(b)
LDA,
(c)
PCA,
(d)
Relie .
8
K.
López
de
Calle
e
al.
/
Compu e s
in
Indus y
112
(2019)
103114
74
and
maximum
alues.
Relie
algo i hm
has
been
slow
(possibly
due
o
easons
ela ed
o
implemen a ion)
and
i
has
aken
conside ably
longe
imes
han
he
es
o
he
algo i hms.
Adding
he
scaling
o
ha ,
i
always
ob ains
he
lowes
alues
(0)
achie ing
e y
simila
sco es
o
he
es
o
he
algo i hms.
This
nuance
o
he
scale
should
be
aken
in o
accoun ,
also
he
eal
elapsed
ime
in
addi ion
o
he
sub-
dimension
should
be
conside ed.
I
is
also
in e es ing
o
no e
he
capabili y
o
de ec ing
noise
ha
algo i hms
ha e
demons a ed,
in
which
LDA
clea ly
excels.
As
he
noise
gene a ed
we e
uni o mly
dis ibu ed
a iables,
one
could
expec
o
ob ain
be e
noise
de ec ion
sco es
by
supe ised
algo i hms.
Howe e ,
possibly
because
i
does
no
end
o
gene a e
spa se
weigh s
(see
Spa si y),
i
is
no
he
case
o
Relie .
Visualiza ion
o
he
h ee
main
dimensions
also
p esen s
in e es ing
addi ional
in o ma ion.
Fig.
4
sugges s
ha
LDA
should
be
selec ed
o
hese
cases
o
use
ollowed
by
PCA,
hen
Relie
and
las ly
Au oencode .
Howe e ,
his
isualiza ion
also
shows
ha ,
o
applica ions
aiming
o
ha e
in e p e able
esul s,
Relie
would
be
a
be e
op ion
due
o
i s
high
sco es
in
In e p e abili y
and
E ficacy.
The
ada -cha s
displayed
in
Fig.
5
ep esen
he
sco es
ob ained
in
each
o
he
sub-dimensions
by
each
algo i hm
and
show
wo
di e en
pa e ns:
one
o
FP
algo i hms
and
ano he
o
he
FS
ones.
This
di e ence
is
mos ly
caused
by
bo h
he
comp ehensi eness
and
compu a ion
e ficiency
sco es.
The
fi s
because
i
only
belongs
o
FS
algo i hms;
and,
he
la e ,
because
o
he
a o emen ioned
easons
ela ed
o
scale.
Taking
some
o
he
weigh s
o
he
10
mos
aluable
a iables
o
each
algo i hm
along
he
simula ion
p ocess
(S ep
D
Weigh
s o ing
in
Fig.
2),
Fig.
6
is
c ea ed.
The
e olu ion
o
he
weigh s
p esen ed
in
he
g aph
shows
how
ecalcula ing
he
models
has
de ec ed
changes
in
he
impo ance
o
he
ea u es
as
a
consequence
o
changes
in
he
inpu
da a,
as
he
e olu ion
in
ac ions
along
he
ime
show.
No e
ha
each
algo i hm
ends
o
beha e
in
a
de e minis ic
way
h oughou
he
simula ion
p ocess,
ei he
hey
assign
big
alues
o
a
small
numbe
o
a iables
o
hey
end
o
gi e
homogeneous
alues
o
each
a iable.
This
e ec
has
been
co ec ly
eflec ed
by
he
spa si y
dimension,
as
i
ob ains
small
alues
o
he
case
o
Relie
(see
Table
C.7
Gea
Tes
14)
whe eas
LDA
ob ains
a
qui e
high
spa si y
alue
(meaning
i
ends
o
ocus
on
a
small
g oup
o
a iables).
In
ela ion
o
he
e olu ion
o
he
weigh s
o e
ime,
he
di e en
de e mina eness
alues
ha e
demons a ed
hei
use
o
de ec ing
bo h
he
change
om
ini ial
alues
and
he
change
in
each
DR
ecalcula ion
s ep.
As
an
example,
in
he
e olu ion
o
he
weigh s
shown
in
Fig.
6,
i
is
possible
o
see
ha
he
algo i hm
ha
has
he
lowes
a e age
de e mina eness
alue,
Au oencode ,
is
clea ly
he
algo i hm
wi h
he
g ea es
change
in
weigh s
om
he
beginning
o
he
end.
Meanwhile,
PCA,
which
has
he
highes
de e mina eness
om
p e ious
sco e,
shows
he
smoo hes
changes
in
each
i e a ion.
No ice
ha ,
e en
i
he
algo i hms
should
selec
he
mos
in e es ing
a iables
o
he
p ocess,
in
his
pa icula
case
o
gea
es
14,
he e
is
no
clea
ag eemen
be ween
Fig.
6.
E olu ion
o
he
weigh s
o
Gea
Tes
14.
In
all
cases,
he
colo ed
p opo ions
ep esen
op
10
a iables
acco ding
o
each
algo i hm
( hey
a e
di e en )
while
black
colo
ep esen s
he
sum
o
he
es
o
he
a iables.
K.
López
de
Calle
e
al.
/
Compu e s
in
Indus y
112
(2019)
103114
9
75
algo i hms.
This
can
be
caused
by
di e en
easons:
due
o
s ong
co ela ions
in
he
a iables
(as
hey
a e
o igina ed
om
same
o
simila
senso s);
o ,
because
o
he
di e en
c i e ia
used
by
DR
algo i hms
o
educe
dimensionali y
( a iables
maximizing
he
sepa a ion
be ween
deg aded
and
non-deg aded
ins ances
o
combina ions
o
a iables
e aining
as
much
in o ma ion
as
possible).
Fig.
7
depic s
he
new
dimension
c ea ed
by
each
DR
algo i hm
in
S ep
(E)
Signal
econs uc ion.
This
figu e
aises
some
con a-
dic ions.
The
s opping
poin s
ound
by
mos
o
he
algo i hms
( e ical
line
in
he
g aph)
a e
eally
close
o
he
s opping
poin
ob ained
by
he
expe
based
dimension.
Howe e ,
his
is
no
eflec ed
by
E ficacy
me ics
Table
C.7
in
Gea
14,
jus
because
he
poin s
ound
ou
o
con ol
by
he
algo i hms
(ma ked
wi h
a
s a
in
he
g aph)
do
no
coincide
wi h
he
poin s
ound
ou
o
con ol
by
he
expe
ea u e,
he e o e,
he
E ficacy
sco e
(wi hou
conside ing
ND)
is
null
o
all
he
algo i hms,
when
his
should
no
be
he
case.
E alua ing
he
me ics
globally,
he
ollowing
aspec s
a e
no ewo hy:
E ficacy
shows
some
limi a ions.
On
he
one
hand,
i
displays
accu a ely
he
capabili y
o
de ec ing
noise.
On
he
o he
hand,
conside ing
DR
me hods
can
ha e
mo e
in o ma ion
han
he
bes
ea u e,
a
100%
ag eemen
be ween
expe
and
algo i hms
would
be
complica ed
o
each.
Fu he mo e,
he
a o emen ioned
di e -
ences
be ween
s opping
poin s
ha e
no
been
conside ed,
hei
inclusion
as
sub-dimension
could
a ec
he
E ficacy
sco e
g ea ly.
Addi ionally,
i
has
o
be
aken
in o
accoun
ha
he
expe
c i e ia
is
used
as
base-line.
Due
o
he
la ge
amoun
o
ea u es
a ailable
in
he
Gea s
da ase ,
he
expe
was
helped
by
a
co ela ion
analysis,
which
helped
o
find
a iables
highly
co ela ed
wi h
he
deg ada ion.
Consequen ly,
he
expe
c i e ia
(p obably
ying
o
find
mono onic
and
linea ly
co ela ed
ea u es)
could
ha e
seen
i s
p io
judgmen s
ein o ced
e en
i
he
ea u es
sugges ed
by
he
co ela ion
we e
no
he
bes .
The e o e,
he
base-line
should
be
ques ioned
in
all
he
compa isons
o
ea u es
o
moni o ing
pu poses,
whe e
he
eal
class
alue
is
ha dly
known.
As
a
esul ,
basing
wo
hi ds
o
he
E ficacy
only
on
TPR
and
PPV
is
no
accu a e,
and
his
me ic
should
be
eadjus ed
o
eflec
eali y
in
a
mo e
de ailed
way.
Fig.
7.
Dimensions
c ea ed
by
algo i hms
in
Gea
Tes
14.
The
isible
dimensions
co espond
o
he
Signal
econs uc ion
s ep
E
in
he
simula ion.
Solid
blue
line
co esponds
o
he
mean
alue
o
he
Shewha
cha
and
he
ho izon al
discon inued
lines
symbolize
he
h esholds.
The
s opping
poin
is
ep esen ed
as
he
e ical
discon inued
line.
The
poin s
wi h
ed
s a s
a e
ou
o
con ol.
10
K.
López
de
Calle
e
al.
/
Compu e s
in
Indus y
112
(2019)
103114
76
In e p e abili y
me ic
sco es
li e
up
o
he
expec a ions.
Spa si y
is
a
good
measu e
o
he
endency
o
ocus
on
small
subse s
o
a iables,
ne e heless,
i
does
no
conside
ha
FS
algo i hms
use
a
single
dimension
o
ep esen ing
he
p ocess
ega dless
o
how
he
es
o
he
weigh s
a e
dis ibu ed.
Howe e ,
his
ac
is
co ec ed
by
he
comp ehensi eness
sub-
dimension,
which
penalizes
DR
echniques
using
ma hema ical
mappings.
Rega ding
a e age
de e mina eness,
i
is
no
e y
clea
whe he
i
p oduces
sa is ac o y
esul s,
as
i
migh
ha e
been
penalizing
models
ha
ha e
encoun e ed
big
changes
in
ea u es
om
he
ini ial
calcula ions,
which
is
no
in insically
bad.
In
ela ion
o
he
E ficiency,
igno ing
he
limi a ions
o
he
scaling,
i
ep esen s
he
cos s
equi ed
o
implemen a ion.
I
conside s
he
ime
needed
o
each
ecalcula ion,
and
he
possibili y
o
educing
he
amoun
o
ecalcula ions
wi hou
losing
in o ma ion.
All
in
all,
ep esen ing
he
e alua ion
wi h
h ee
main
dimensions
seems
a
good
in e media y
in
o de
o
educe
he
gap
be ween
he
use
and
he
implemen e ,
as
hese
me ics
could
be
used
as
a
common
language
be ween
bo h.
Rega ding
he
bes
DR
algo i hm
o
he
moni o ing
pu poses,
LDA
shows
he
bes
po en ial
in
bo h
da a
sou ces,
wi h
good
E ficacy
and
E ficiency
bu
wi h
wo se
esul s
han
Relie
o
In e p e abili y.
The
ope a o
should
conside
including
ei he
one
algo i hm
o
he
o he .
Howe e ,
hese
esul s
do
no
imply
hese
algo i hms
a e
always
he
bes
o
he
cases
o
use
(bea ings/gea s),
and
simula ions
and
e alua ions
should
be
done
in
simila
o
di e en
applica ions.
Finally,
i
is
impo an
o
no e
ha
domain
knowledge
is
essen ial
o
he
de elopmen
o
CM
echniques.
Wi hou
domain
knowledge
i
would
no
be
possible
o
define
app op ia e
ea u es
no
o
use
he
small
nuance
o
he
deg ada ion
in
o de
o
use
su pe ised
algo i hms.
In
his
pa icula
case,
ha
would
imply
no
being
capable
o
using
supe ised
algo i hms
(LDA,
Relie ),
which
a e
among
he
bes
candida es.
The e o e,
i
is
impo an
o
no e
ha
he
inclusion
o
domain
knowledge
is
decisi e
o
he
implemen a ion
o
any
da a-based
model.
This
is
because,
i
he
knowledge
is
co ec ly
ansla ed
in o
he
da a
d i en
model,
he
pe o mance
will
inc ease
sha ply.
6.
Conclusions
This
wo k
p esen s
an
inno a i e
app oach
based
on
he
use
o
dimensionali y
educ ion
algo i hms
o
he
ea ly
de ec ion
o
anomalies
and
he
ob en ion
o
p ocess
ela ed
knowledge
in
scena ios
wi h
limi ed
da a.
Fou
di e en
DR
algo i hms
ha e
been
compa ed
o
moni o ing
pu poses
in
wo
di e en
case
scena ios:
bea ing
moni o ing
and
gea s
moni o ing.
Finally,
he
esul s
o
he
simula ion
ha e
been
e alua ed
wi h
a
se
o
me ics
c ea ed
ad
hoc
conside ing
he
equi emen s
ound
in
he
indus y.
DR
algo i hms
ha e
demons a ed
he
capabili y
o
de ec ing
changing
a iables
in
a
sys em,
as
he
e olu ion
o
he
weigh s
has
shown.
Addi ionally,
ha
change
has
been
moni o ed
in
o de
o
de e mine
accep able/unaccep able
h esholds.
Fu -
he mo e,
as
his
app oach
can
easily
wo k
wi h
la ge
se s
o
a iables,
i
becomes
a
good
s a ing
poin
o
ecommending
a iables
in
highly
mul i a ia e
da ase s,
o
be e
unde s and
and
de ec
changing
a iables
and
o
ini ia e
CM
sys ems
wi h
li le
da a.
Some
specifically
designed
me ics
ha e
been
used
o
e alua e
he
algo i hms.
These
me ics
a e
based
on
he
indus ial
implemen a ion
needs,
mo e
p ecisely
on
he
e alua ion
o
he
In e p e abili y,
E ficacy
and
E ficiency.
This
inclusion
o
addi-
ional
dimensions
o
he
E ficacy
akes
o he
impo an
ac o s
in o
conside a ion.
These
o he
ac o s
a e
neglec ed
in
o he
wo ks,
bu
di ec ly
a ec
he
final
chance
o
implemen ing
he
algo i hms.
The
se
o
p oposed
sub-dimensions
co ec ly
eflec s
i s
undamen al
pu pose,
wi h
some
need
o
adjus men
o
E ficacy,
ha
should
be
imp o ed
by
adding
mo e
in o ma i e
sub-dimensions
in
he
u u e.
The e
is
no
single
and
bes
algo i hm
o
CM.
Ne e heless,
in
he
case
scena ios
p esen ed,
wo
algo i hms
s and
ou :
LDA,
which
sco es
he
bes
a e age
e ficiency
and
e ficacy;
and
Relie ,
which
p o ides
he
mos
unde s andable
ea u es,
and
sco es
well
in
he
o he
dimensions.
The
final
decision
should
be
le
o
domain
expe s
who
can
make
he
bes
decision
conside ing
he
p oposed
me ics
and
he
equi emen s
o
he
final
applica ion.
This
wo k
ocuses
on
wo,
o en
una ended,
pa adigms:
he
applica ion
o
CM
in
sys ems
wi h
da a
limi a ions;
and
he
need
o
include
addi ional
me ics
besides
pe o mance
o
he
analysis
o
he
implemen a ion
o
algo i hms.
Howe e ,
some
issues
a e
no
su ficien ly
add essed
he e
and
i
would
be
in e es ing
o
explo e
and
analyze
hem
in
u u e
wo ks.
Fo
ins ance,
he
inaccu acies
o
E ficacy
me ics,
which
do
no
conside
he
dis ance
be ween
s opping
poin s
should
be
added
as
a
sub-dimension
o
E ficacy;
he
s udy
o
he
ag eemen
be ween
he
impo an
ea u es
acco ding
o
di e en
DR
algo i hms;
he
possibili y
o
ex ending
he
algo i hm,
by
inc emen ally
lea ning
om
ailu es
om
one
es
o
ano he ;
o ,
las ly,
he
ex ension
o
he
app oach
o
o he
indus ial
cases
o
use
no
ela ed
o
o a ing
machine y
and/o
in ol ing
a ying
ope a ions.
Decla a ion
o
in e es
The
au ho s
decla e
no
compe ing
in e es s.
Appendix
A.
Model
agg ega ed
sco es
in
sub-dimensions
Table
A.5
Mean
o
algo i hm
sco es
by
model.
Model
Comp.
E .
De .
p e .
Spa si y
Comp e.
A g
De .
TPR
PPV
ND
Au oencode
0.954
0.709
0.679
0
0.618
0.156
0.162
0.722
LDA
0.997
0.759
0.908
0
0.720
0.521
0.543
0.972
PCA
1.000
0.864
0.397
0
0.667
0.483
0.449
0.722
Relie
0.000
0.685
0.288
1
0.537
0.639
0.630
0.694
K.
López
de
Calle
e
al.
/
Compu e s
in
Indus y
112
(2019)
103114
11
77
Ene gies 2019,12, 3373 2 o 19
p o en success o heal h s a us de ec ion, aul iden i ica ion and p edic ion [
1
,
2
,
9
]. Such echnology
allows iden i ying he s a e o he asse s emo ely, educing conside ably he need o isual inspec ions
on si e, which is a cos ly ma e mos ly o o sho e a ms [11].
Pa icula ly, gea boxes ep esen a delica e componen o WTs [
10
]. The a ious ailu e s a is ic
analysis ca ied ou lead o some con o e sy on i s endency o ailu e [
12
], as some o hem ha e less
ailu es epo ed [
8
,
12
], whe eas o he s ind high ailu e a es ela ed o gea boxes [
13
]. Anyway, mos
o he s udies associa e he longes WT down imes o his componen [
3
,
12
] and emphasise he need o
p ope moni o ing echniques o a oid hem. Addi ionally, i is conside ed one o he cos lies pa s
o he WTs [
10
] and al hough he e a e a emp s o eplace hem wi h di ec d i es o educe cos s,
s udies ques ion his assump ion, and s ill he majo i y o o sho e wind u bines ely on gea boxes [
14
].
Consequen ly, hei moni o ing is o i al impo ance.
Gea box and gea box ela ed subsys em condi ion moni o ing (CM) has been b oadly add essed
on he li e a u e. Fo ha pu pose a wide a ie y o app oaches ha e been es ed, such as ib a ion, oil
deb is moni o ing (ODM), acous ic emissions and cu en signa u e among o he s. Vib a ion-based
CMS p e ail o e o he kinds o senso s [
2
,
9
], wi h much esea ch ca ied ou on he ield o signal
p ocessing o ib a ion signals in ime, equency, ime/ equency and o de domains [
3
]. Howe e ,
some wo ks ind ODM echniques also in e es ing, o hei highe co ela ion wi h wea c ea ion [
15
],
o o he added alue hey ha e o moni o ing bo h he oil quali y and he s a e o he gea box
pa s [1,2].
Ne e heless, e en i he ad an ages o CM a e p o en [
2
,
11
], i s ans e om expe imen al es s
o eal WT use cases is less known in he li e a u e [
16
]. This is because he a iabili y o ope a ion
condi ions o WTs a ec s he ex ac ion o indica o s while i specially damages he sys ems o WTs [
2
].
Mos o he wo ks p esen ing eal in-se ice WT da a a e based on he use o SCADA (Supe iso y
Con ol and Da a Acquisi ion) da a, which is eadily a ailable in gene al. Typically, i is used o
compa e pe o mances among WTs using powe cu es [
17
]. These benchma king p ocedu es a e
usual o o he O&M issues such as pi ch misalignmen co ec ion [
18
,
19
] and he iden i ica ion o
de ec i e anemome e s [
20
]. Addi ionally, empe a u es om he SCADA ha e been modelled and
compa ed o e ime o use di e ences as ala ms as he di e en wo ks e iewed by he au ho s o [
21
]
show. Howe e , he success o hese echniques is limi ed [
3
]. Pa ly, because o ex e nal in luences
(such as he ou side empe a u e) ha equi e he ala ms o be manually supe ised by ope a o s [
21
].
Consequen ly, he inclusion o addi ional CM senso s in ope a ing WTs is lou ishing [
9
], and an
inc easing numbe o wo ks p esen indings om eal cases o use:
•
In he wo k by he au ho s o [
22
], wo case s udies a e analysed: in he i s one, physical
p inciples ha ela e he di e ence in empe a u es wi h he e iciency, o a ional speed and powe
ou pu a e used, and he app oach is alida ed by using he de ia ion o he empe a u e wi h
espec o powe in o de o o esee a ailu e; in he second one, ib a ion and pa icle coun e
senso s a e used and he e olu ion o he signals is s udied be o e and a e he eplacemen o a
bea ing. They sugges using cumula i e pa icle coun s o be e de ec ailu es ins ead o di ec
pa icle c ea ion measu emen s and o combine a ious senso s in o de o imp o e con idence in
he diagnosis.
•
In he wo k by he au ho s o [
23
], hey c ea e a heal h indica o based on he cen oids p oposed
by a Sel O ganising Map in o de o g oup WTs acco ding o heal h s a us using SCADA da a.
This way ope a o s a e gi en addi ional in o ma ion ega ding he heal h s a e o he WTs, and
can plan consequen ly.
•In he wo k by he au ho s o [16], cu en and ib a ion analyses a e used o diagnose a u bine
d i e ain, hey emphasise on he di icul y o using signal p ocessing echniques ha a e
only p o en a labo a o y scale, and ecognise he complexi y o calcula ing he emaining
use ul li e (RUL) and es ablishing damage/heal hy h esholds, especially wi h a lack o a ailable
his o ical da a.
84
Ene gies 2019,12, 3373 3 o 19
•
In he wo k by he au ho s o [
24
], he ib a ion signa u e o a sample o heal hy wind u bines is
shown, mos ele an indica o s a e iden i ied on a e aged powe spec a and he dependence o
ampli udes on he ope a ion is s udied. They conclude ha he high impac o wind speed on
ib a ion ampli udes has o be aken in o accoun o de elop CMS.
•
In he wo k by he au ho s o [
14
], he da a ga he ed be o e and a e a plane a y gea was changed
due o spalling is examined. They in eg a e empe a u e, ib a ion and pa icle coun e signals in
o de o educe alse ala ms, and p o e he abili y o dis inguish heal hy and wa ning s a es.
•
In he wo k by he au ho s o [
25
], hey sugges he use o mo ing a e ages (o bo h sho and
long e m ends) o ODM o gene a e a coun a e p opaga ion model. Then, hey es ablish an
accep ance h eshold based on he equi alen maximum angle o spall which is ela ed o bea ing
geome y; and, las ly, hey es ima e emaining use ul li e (RUL).
Mos o he wo ks ela ed o on se ice WT u ilise ib a ion and/o oil deb is senso s [
3
,
9
].
The wo ks based on ODM om he p e iously men ioned ones [
14
,
22
,
25
] ag ee on he same di icul ies
o he de elopmen o ODM sys ems: he need o a e aging o using cumula i e alues ins ead
o using di ec ly pa icle gene a ion a es; and he endency o pa icle c ea ion a e o a y wi h
ope a ion. These indings a e suppo ed by he ex ensi e wo k o he au ho s o [
26
], in which a
ull-scale WT gea box o 750 kW is es ed wi h in-line and online senso s and samples aken along he
ime. In hei indings, he need o il e ing in luences caused by ope a ional condi ions is ema ked;
hey ecommend o ocus in ends ins ead o in absolu e alues, and sugges conside ing big pa icle
size (>14
µ
m) indica o s in pa icula ; also, hey iden i y ha damaged gea boxes ha e much highe
deb is gene a ion a es han heal hy ones.
Taking in o conside a ion he in e es o ha ing eal on-se ice WT ope a ion da a analysed, and
ha some o he limi a ions o ODM o WT a e al eady iden i ied on he li e a u e, his wo k aims o
p o ide a be e insigh o he de elopmen o ODMs. Fo ha pu pose, he da a ob ained in h ee
WTs moni o ed wi h oil deb is senso s a e s udied o a pe iod o six mon hs; he eadings o he
senso a e compa ed o o he adi ional SCADA based moni o ing echniques; and, las ly, a s udy o
he di e en ope a ion s a es is ca ied ou o de e mine which il e ing c i e ia is be e o de elop an
heal h index ha conside s ope a ing condi ions.
2. Da a and Me hodology
2.1. Wind Fa m and Tu bines
This s udy analyses he da a p oduced by 3 WTs which a e loca ed in he wind a ms a Bayo and
Mon e os, in Za agoza (Spain). Bo h wind a ms a e close one ano he and unde go simila in luences
o he wind. The na u al ba ie s o he Ibe ian Sys em moun ain anges in he sou h and he Py enees
moun ain anges in he no h cons i u e a unnel e ec ha c ea es he me eo ological occu ence
known as cie zo; a d y, usually cold and accele a ed lux o ai in ensi ied by he na u al unnel going
h ough he Eb o alley. Cie zo is mo e equen du ing win e and he beginning o sp ing, and is
compensa ed by he an agonis ic phenomenon known as bocho no, ha goes in he opposi e di ec ion
o cie zo and ends o be so e . Addi ionally, hese oposing phenomena p o ide he wind wi h copious
kine ic ene gy and make he egion an in e es ing loca ion o he exploi a ion o wind ene gy [27].
The WTs ha e a 58 m diame e o o and h ee blades. Thei a ed powe is 850 kW and cu -in and
cu -ou wind speeds a e 3 m/s and 20 m/s, espec i ely. They ha e plane a y gea boxes wi h 1/62
ansmission a ios coupled wi h asynch onous gene a o s. The mine al lub ican is cleaned by o line
oil il e s and he online oil deb is op ical senso s is ins alled in a bypass o he lub ica ion sys em.
Rega ding he heal h s a us o he gea boxes, isual and endoscopic inspec ions ca ied ou on-si e
e eal di e en le els o damage. Two o he gea boxes show medium wea le els (WT 1 and WT 2)
wi h mic opi ing p esen in mos o he gea s, whe eas he las one is diagnosed wi h medium–high
wea le el showing g ea e su aces damaged by mic opi ing in some gea s and pi ing in he sun
gea . Howe e , no co ec i e ac ions ha e been ecommended ye .
85
Ene gies 2019,12, 3373 4 o 19
2.2. Op ical Oil Deb is Senso
Oil samples can be aken and analysed o line in labo a o ies, howe e , his p ocedu e delays
he decision making p ocess and equi es o access he WTs. The e o e, online oil deb is senso s
a e an a ac i e way o de e mining he quali y o he lub ican and sa egua d he componen s o
he gea box.
In pa icula , his wo k uses a op ical oil deb is senso . This kind o senso s moni o he
luid condi ion and con amina ion using op ical echnology by cap u ing high- esolu ion images
o he mo ing luid, and la e applying ad anced p ocesses o image digi isa ion and spec al
analysis. They de ec , quan i y and classi y he pa icles bigge han 4 mic ons by size and/o shape,
in addi ion o dis inguishing hese pa icles om ai bubbles [
28
]. Besides wind u bine lub ica ion
sys em moni o ing, his kind o echnology is well-sui ed o o he indus ial applica ions such as
au omo i e, s eel sec o , was ewa e ea men o cemen indus ies [
29
] as all o he p e ious use
lub ica ion sys ems.
2.3. Da ase
The s udy is based on a da ase consis ing o six mon hs long eco ds o 3 WTs. The da a eco ds
a e aken wi h one minu e equency om he SCADA. A he same ime, addi ional measu emen s
p o ided by online op ical oil deb is senso s a e aken. Va iables om he SCADA ep esen he
ope a ion o he WTs, whe eas he ones p o ided by he senso indica e he amoun o pa icles o size
g ea e han 4, 6 and 14 mic ome e s (ISO.4, ISO.6 and ISO.14, espec i ely) p esen on he lub ica ing
oil acco ding o he ISO 4406 s anda d [
30
]. These alues o he oil senso ep esen he pa icle
gene a ion a e, as he oil is being con inuously il e ed. De ails o he a iables o he SCADA and he
oil deb is senso wi h he uni s o measu emen a e p esen ed in he Table 1.
Table 1. Va iables a ailable in he da ase and measu emen uni s by da a sou ce.
Sou ce Sys em Uni s o Measu emen
SCADA
Pi ch angle ◦
Gea box empe a u e ◦C
Wind speed m/s
Gene a o speed pm
Ac i e powe kW
Oil deb is senso
ISO.4 scale
ISO.6 scale
ISO.14 scale
Fo p i acy easons he da a is shown in a no malised way along his wo k wi hin a 0 o 1 ange
co esponding o minimum and maximum alues o each o he a iables in he da ase .
2.4. Me hodology
In o de o gain be e insigh on he use o oil deb is senso s o ob ain heal h indica o s, he
s udy has wo pa s: an explo a ion and co ela ion analysis s age, in which an o e iew o he da a is
p esen ed and some me hods o he li e a u e con as ed; and he compa ison o ope a ion egions
and heal h index (HI) de elopmen , whe e di e en ope a ing egimes a e compa ed and he mos
app op ia e one is chosen as he basis o de elop a HI. The me hods used in each o he pa s a e
p esen ed below.
2.4.1. Explo a ion and Co ela ion Analysis
In an ini ial s age, a ious isualisa ion and co ela ion echniques a e used:
86
Ene gies 2019,12, 3373 5 o 19
•
Pea son co ela ion: Coe icien used o measu e he deg ee o linea associa ion be ween wo
a iables, p esen ed in he wo k by he au ho s o [31].
•
Spea man’s co ela ion: Nonpa ame ic coe icien ha e lec s he deg ee o mono onici y
be ween wo a iables explained in he wo k by he au ho s o [32].
•
P incipal componen analysis (PCA): PCA is a o hogonal ans o ma ion ha u ns a se
o a iables in o a se o linea ly unco ela ed a iables. This echnique is widely used
o isualisa ion pu poses in o de o educe mul idimensional spaces o lowe dimensional
ep esen a ions wi h he minimum in o ma ion loss [33].
•
Exponen ial mo ing a e age: In con as o egula mo ing a e age whe e he a e age o a
window o alues in aken, his ype o mo ing a e age is used when la es alues a e need o
ha e mo e impo ance. The implemen a ion used in his wo k can be ound on he wo k by he
au ho s o [34].
•
Local eg ession (LOESS): This nonpa ame ic me hod is used on local subse s o da a.
The implemen a ion is based on he wo k by he au ho s o [35].
•
Decision ees: Decision ees a e machine lea ning (ML) algo i hms used o classi ica ion.
Thei goal is o p edic alues o a a ge a iable based on he inpu s. T ees a e buil by spli ing
inpu a iables wi h c i e ia ha maximise he p obabili y o ha ing ins ances o ce ain g oup
in each pa i ion. The ees used in his wo k used Gini impu i y index o pa i ioning and a e
implemen ed in he wo k by he au ho s o [
36
], which is based on he wo k by he au ho s o [
37
].
2.4.2. Compa ison o Ope a ion Regions and Heal h Index (HI) De elopmen
Du ing he ini ial explo a ion, he in luence o he ope a ion in he pa icle c ea ion a es is
de ec ed; ne e heless, as he e is no clea co ela ion iden i ied be ween ope a ion a iables and
pa icle c ea ion, i is decided o conside only he measu emen s ha a e aken unde he same
ope a ion condi ions. Fu he mo e, a me hodology is used o de ine which ope a ing condi ions a e
he mos app op ia e o moni o ing pu poses. The ollowing echniques we e used.
•
Ope a ion egion (OR) de ini ion: In o de o ind he op imal ins an s o aking measu emen s,
se e al ope a ing egions (OR) a e explo ed. Each ope a ing egion is de ined by a se o
ules/c i e ia, such as wind speed in ange (x m/s, y m/s), ac i e powe equals nominal powe ,
e c. The ORs analysed in his wo k a e sugges ed by expe s in he domain, and a e depic ed in
he ollowing Figu e 1wi h a sho explana ion added in Table 2.
•
Ope a ion s a es (OS): Ins ead o conside ing each ime ins an indi idually as a da a poin wi h
ce ain associa ed a iable alues (pi ch angle, ac i e powe , ISO.4. . . e c.) g oupings o da a
poin s ha e been s udied. These g oups o ope a ion s a es a e gene a ed conside ing he di e en
ope a ion egions p e iously p esen ed, and co espond o a se o ch onologically con inuous
poin s o e ime ul illing he ules p oposed by he OR. E e y ime he machine wo ks unde he
c i e ia o an OR we say i has en e ed in a new OS ela ed o ha OR, ha las s as long as he WT
keeps wo king unde he cons ain s o he OR.
Table 2. Desc ip ions o he di e en ope a ing egions.
Ope a ing Region Cha ac e is ics
Nominal S able powe gene a ion. Va ying pi ch
N. & low pi ch Simila o nominal, bu mo e es ic i e and no including high wind speeds,
delimi ed using pi ch alues.
Ramp o nominal Ranges om abou he middle o he powe cu e o beginning o nominal ope a ion.
Ramp Values aken only du ing he powe amp.
P e- amp Values aken be o e he gene a o speed amp s a s.
87
Ene gies 2019,12, 3373 6 o 19
Figu e 1. Ope a ion egions o e ac i e powe agains wind speed plo .
•
Ope a ion clus e ing: Wi h he pu pose o analysing he s eadiness o he di e en OR he
ollowing p ocedu e was used in o de o gene a e da a clus e ep esen ing he a iabili y o he
ope a ion in each OS acco ding o he di e en ORs.
1.
Scale all he a iables be ween 0 and 1 co esponding o he maximum and minimum alues
o each a iable.
2.
Taking an OR (Example:Nominal) ind he espec i e numbe o OS occu ences in he
da ase {OS}m
i=1, whe e mis he numbe o occu ences.
3. C ea e a ma ix o each OSiwhe e i=1, 2, . . . , m:
OSi=
ai
11 ai
12 . . . ai
1p
ai
21 ai
21 . . . ai
2p
.
.
.....
.
.
ai
ni1. . . . . . ai
nip
(1)
whe e
p
is equal o he numbe o senso s conside ed and
ni
is he leng h o he
i
- h OS;
he e o e, hese ma ices con ain he alues o he pope a ion a iables along he OS.
4.
Then, he di e ence ec o o each a iable is calcula ed by OS. This ec o s ep esen he
a iabili y o he ope a ion du ing he OS and gi e as a esul he new ma ix D:
Di=
di
11 di
12 . . . di
1p
di
21 di
22 . . . di
2p
.
.
..
.
.....
.
.
di
ni−1,1 di
ni−1,2 . . . di
n−1,p
,i=1, 2, . . . , m(2)
whe e
djk =ai
j+1,k−ai
jk
, ha is, each elemen o he di e ence ec o is he di e ence
be ween he measu emen in ha ins an
(j)
and he ollowing measu emen
(j+
1
)
, o
each j=1, 2, . . . , (n−1),k=1, 2, . . . , p.
88
Ene gies 2019,12, 3373 7 o 19
5. Then ma ix Ris compu ed.
R=
11 12 . . . 1p
21 22 . . . 2p
.
.
..
.
.....
.
.
m1 m2. . . mp
(3)
R
is he esul o compu ing he columnwise quad a ic mean o he
Di
ma ices, and
ep esen s he a e age alues o he a iabili y conside ing bo h nega i e and posi i e
alues. They a e compu ed in he ollowing way.
ik =
u
u
∑ni−1
j=1djk
ni−1(4)
6. F om he Rma ix, wo me ics a e ob ained:
(a) Cen oid: The a e age posi ion o he poin s con ained in W. Compu ed as ollows.
Cen oid = (µ1, . . . , µp)(5)
whe e o all k=1, . . . , p he a e age µkis calcula ed as ollows.
µk=1
m
m
∑
i=1
ik (6)
(b)
Clus e dispe sion: Mean o he a iable a iance alue ha ep esen s how dispe se
he clus e is; i is calcula ed as ollows.
Dispe sion =
p
∑
k=1
σk(7)
Each σk o k=1, 2, . . . , pis he s anda d de ia ion compu ed in he ollowing way.
σk=1
m
m
∑
i=1
( ik −µk)2
This p ocedu e is epea ed sepa a ely o each WT and OR. The e o e, he e a e i e ORs by h ee
WTs, a o al o 15 da a clus e s.
•
Ope a ion s a e and clus e me ics: F om he clus e s o da a and he OS some me ics a e
calcula ed ha help iden i ying he mos in e es ing OR. These me ics a e as ollows.
–
Weekly occu ence a io: A e age numbe o imes pe week he WT en e s in an OS as
de ined in he OR.
–
S eadiness: The euclidean dis ance om he cen oid (o mean poin ) o a clus e o he o al
s eadiness (no a ia ion) poin .
–
Dispe sion: Indica es how sp ead he da a poin s wi hin a clus e a e. De ined p e iously in
Clus e dispe sion.
89
Ene gies 2019,12, 3373 8 o 19
3. Resul s
As in Sec ion 2.4, esul s chap e is di ided in wo pa s. The i s pa , Sec ion 3.1, explains he
explo a o y analysis ha is ca ied ou o e he da ase and he ela ions ound be ween he a iables.
The second pa , Sec ion 3.2, shows he s eps ha we e aken in o de o iden i y he bes condi ions
o ob aining measu emen s along he ime in o de o ob ain a heal h index o he gea boxes.
3.1. Explo a ion and Co ela ion Analysis
Taken a sample o he whole da ase , Pea sons and Spea mans co ela ions a e s udied. In o de o
iden i y any possible di e ence be ween powe gene a ion and du ing no gene a ion, co ela ion is also
measu ed in sepa a e samples. Howe e , no signi ican co ela ions (nei he Pea sons no Spea mans)
a e ound be ween ope a ion a iables and pa icle gene a ion da a. Rega ding he ope a ional
a iables, some show high deg ee o associa ion because o he con ol sys em. Fu he mo e, he
associa ion o he same a iables among WTs o e o e lapped ime spans yields high co ela ion
which means hey ace simila en i onmen al condi ions (wind). Howe e , his is no he case o ODSs,
ha do no co ela e om one WT o ano he . Ne e heless, in ligh o he s ong co ela ions be ween
pa icle indica o s (ISO.4, ISO.6 and ISO.14), and ollowing he ad ice p o ided by he au ho s o [
26
],
i is decided o ollow he s udy using ISO.14 indica o as only indica o o pa icles in o de o
simpli y he s udy.
A e he b u e co ela ion s udy, he a iables a e isually s udied agains he wind speed in
Figu e 2. The di e en a iables o he SCADA da a plo ed agains he wind speed show he ypical
pa e ns ha can be ound in wind u bines, and a e ex emely simila one o ano he . The g ea es (bu
ye small) di e ences a e ound in gea box empe a u e, sugges ing he e could be some di e ences in
he cooling sys em o on he e iciency o he gea boxes.
Figu e 2.
SCADA a iables agains wind speed by WT. (
a
) Ac i e powe agains wind speed, (
b
) pi ch
angle agains wind speed, (
c
) gene a o speed agains wind speed and (
d
) oil empe a u e agains
wind speed.
As he signals o he ODSs a e disc e e, much noisie and is almos impossible o isualise
any hing in he aw measu emen s, he measu emen s a e gi en some p e ea men s by a e aging he
alues in 0.33 m/s wind speed bins, which c ea es he pa e n isible in Figu e 3c.
90
Ene gies 2019,12, 3373 9 o 19
Figu e 3.
A e aged alues o gea box empe a u e, ac i e powe and ISO.14 in 0.33 m/s wind speed
bins. (a) Gea box empe a u e. (b) Ac i e powe . (c) ISO.14.
A e aged alues show g ea di e ences be ween he pa icle c ea ion a es among WTs. A he
same ime, he in luence o ope a ion o e pa icle gene a ion is isible. In e es ingly, he beha iou s
do no coincide exac ly be ween WTs: WT 1 and WT 3 show big simila i ies, wi h high wea c ea ion
wi h low wind speeds and lowe wea c ea ion a medium speed o nominal ope a ion; meanwhile,
WT 2 shows a di e en pa e n, as i s wea c ea ion inc eases p opo ionally wi h wind speed.
These di e ences in he beha iou s o he WTs ega ding pa icle c ea ion and ope a ion a e also
p esen in he a e aged alues o wea gene a ion du ing powe p oduc ion and no powe p oduc ion
(including: idling, gene a o u ning wi hou ac i e powe gene a ion and idling because o o e load)
as Figu e 4demons a es.
Figu e 4. Boxplo s o ISO.14 pa icle c ea ion a es du ing di e en ope a ions.
Again, WT 2 does no ac as he o he WTs. In any case, he p e ious igu es sugges he pa icle
c ea ion is g ea e du ing no powe gene a ion, meaning b aking and accele a ion could be causing
highe pa icle c ea ions. Fu he mo e, he e is a clea dis inc ion in he mean le el o pa icle c ea ion
a es ha ma ch he isual diagnos ics o he gea boxes, showing highe alues in WT 3, and lowe
alues o WT 1 and WT 2. This a ia ion o pa icle le els ha indica es dispa a e damage se e i y,
is complex o de ec by jus paying a en ion o he he SCADA a iables. As Figu e 3a,b shows,
he same binned in a iables ypically used o condi ion moni o ing by benchma king (Ac i e powe
and gea box empe a u e) a e no su icien ly di e en in o de o make compa isons be ween u bines
and de e mine whe he WTs could be damaged. Whe eas hese di e ences a e clea ly isible in he
binned ISO.14 alues (Figu e 3c).
91
Ene gies 2019,12, 3373 10 o 19
This ac is clea e when cumula i e pa icle c ea ions a e used. Ins ead o using aw signals,
using cumula i e pa icle a es p o ides a be e insigh o he deg ada ion p ocess, as i allows us
dis inguishing changes in he slopes. In Figu e 5, we see clea and inc easing di e ences be ween
WTs in he ends gene a ed when plo ing cumula i e pa icle c ea ions agains cumula i e powe
gene a ion. I cumula i e empe a u e is obse ed, he di e ences among WTs, e en i exis en ,
a e qui e small which educes he possibili y o co ec ly diagnose ailu es using only SCADA da a.
Addi ionally, he p esence o simila shapes o he h ee u bines in bo h empe a u e and pa icle
c ea ion and conside ing he same pe iods o ime a e s udied indica e some common ac o could be
causing he sha p inc ease in he middle o he cu es, which is isible in bo h a iables.
Figu e 5. Cumula i e alues agains cumula i e ac i e powe . (a) Tempe a u e. (b) ISO.14.
In o de o ep oduce p e ious indings in he li e a u e, i is decided o analyse b aking and
accele a ion egis e ed in he SCADA da a. Fo doing so, he Gene a o speed o i e days o ope a ion
is aken and i is manually labelled adding “b aking”, “boos ing” o “o he ” labels. Then, i e lagged
a iables o he gene a o speed, a exponen ially smoo hed gene a o speed (using a bin size o 15) and
a di e ence ec o a e c ea ed. Wi h his da a a decision ee is ained using he de aul pa ame e s o
classi ica ion cases and i is used o segmen he emaining gene a o speed da a in b ake/boos /o he .
Wi h he da a spli in hese g oups, i is possible o s udy he sequences occu ing in he da a. Boos ing
and b aking sequences a e s udied by measu ing he spea mans co ela ions o he ISO.14 a iables
wi h he smoo hed gene a o speed. In Figu e 6 he dis ibu ion o he co ela ions ob ained by WT
is p esen ed.
Figu e 6. Densi y plo o he spea man co ela ion o ISO.14 in espec o he exponen ially smoo hed
gene a o speed by wind u bine, e ical lines ep esen qua iles. (a) WT 1, (b) WT 2 and (c) WT 3.
92
Ene gies 2019,12, 3373 11 o 19
The dis ibu ion o he co ela ion shows he e di e en beha iou s. In he b aking sequences, WT 1
and WT 3 (Figu e 7a,c) ha e bimodal dis ibu ions wi h a mino mode in s ong posi i e co ela ion
alues and he majo mode in e y s ong nega i e co ela ion alues. This implies ha he e is a
p edominan endency o c ea e mo e pa icles du ing s ops (speed dec eases and pa icle gene a ion
inc eases), bu is no always he case, as in some cases he co ela ion is posi i e (speed dec ease wi h
pa icle dec ease). In WT 2 he opposi e beha iou is iden i ied, e en i he co ela ion dis ibu ion is
also bimodal, he majo mode is on posi i ely co ela ed alues, meaning in his WT he e is a endency
o dec ease pa icle gene a ion when he gene a o is s opping. Rega ding he boos ing sequences, he
o e all co ela ion alues a e qui e low, which implies he e is no clea ela ion be ween he inc easing
speed and he pa icle c ea ion. The p edominance o he majo mode in e y nega i e co ela ion
alues oge he wi h he qua ile lines so a om he 0 alue indica es b aking gene a e an inc ease in
pa icle c ea ion, a leas o u bines WT 1 and WT 3.
Taking WT 3, boos ing and b aking sequences we e analysed in dep h. The ollowing Figu e 7
p esen s wo examples o sequences wi h s ong spea man co ela ion alues wi h nega i e and
posi i e co ela ion.
Figu e 7.
Examples o OSs showing high spea man co ela ion be ween ISO.14 and smoo hed gene a o
speed. (a) Posi i e co ela ion. (b) Nega i e co ela ion.
In e es ingly, despi e he he e is a clea unbalance in he numbe occu ences, s opping can lead
o bo h an inc ease o a dec ease in pa icle gene a ion. No e ha he gene a o speed dec eases as e
han he ISO.14 le el, and he na u e o he exponen ially smoo hed speed is mo e simila o he one o
he ISO.14 a iable ha is mo e in luenced by he ine ia o he sys em han gene a o speed.
The same p ocedu e is ollowed o boos ing, in his case, conside ing p edominan co ela ion is
nea 0 (meaning he e is no mono onici y) examples wi h low co ela ion a e also s udied. Figu e 8
displays occu ences wi h high posi i e co ela ions (a), highly nega i e co ela ions (b) and no
co ela ion (c).
Wi h he uni o m dis ibu ion o co ela ion o boos ing cases and he di e en cases shown
in Figu e 8 he e is no way o iden i ying an expec ed beha iou o he pa icle gene a ion du ing
boos ing sequences. Fu he mo e, Figu e 8c e eals an unexpec ed beha iou du ing idling. As he
senso is gi ing high ISO.14 pa icle le els. This ac occu s mos ly in WT 3 bu is also epo ed in
WT 1, bu wi h a lowe equency. O -line oil il e s should ope a e con inuously ega dless o he
ope a ion o he machine, bu his inding sugges he il e could be s opping in ce ain si ua ions,
which explains also he big di e ence o pa icle gene a ion ound in Figu e 4.
93
Ene gies 2019,12, 3373 18 o 19
4. Ma i-Puig, P.; Blanco, A.M.; Cá denas, J.J.; Cusidó, J.; Solé-Casals, J. Fea u e selec ion algo i hms o wind
u bine ailu e p edic ion. Ene gies 2019,12, 453. [C ossRe ]
5.
Manwell, J.F.; McGowan, J.G.; Roge s, A.L. Wind Ene gy Explained: Theo y, Design and Applica ion; John Wiley
& Sons, L d.: Hoboken, NJ, USA, 2010. [C ossRe ]
6.
C ab ee, C.J.; Zappalá, D.; Hogg, S.I. Wind ene gy: UK expe iences and o sho e ope a ional challenges.
P oc. Ins . Mech. Eng. Pa A J. Powe Ene gy 2015,229, 727–746. [C ossRe ]
7.
Echa a ia, E.; Hahn, B.; an Bussel, G.J.W.; Tomiyama, T. Reliabili y o Wind Tu bine Technology Th ough
Time. J. Sol. Ene gy Eng. 2008,130, 031005. [C ossRe ]
8.
Su, C.; Yang, Y.; Wang, X.; Hu, Z. Failu es analysis o wind u bines: Case s udy o a Chinese wind a m.
In P oceedings o he 2016 P ognos ics and Sys em Heal h Managemen Con e ence (PHM-Chengdu 2016),
Chengdu, China, 19–21 Oc obe 2016. [C ossRe ]
9.
Ga cía Má quez, F.P.; Tobias, A.M.; Pina Pé ez, J.M.; Papaelias, M. Condi ion moni o ing o wind u bines:
Techniques and me hods. Renew. Ene gy 2012,46, 169–178. [C ossRe ]
10.
Ca oll, J.; McDonald, A.; McMillan, D. Failu e a e, epai ime and unscheduled O&M cos analysis o
o sho e wind u bines. Wind Ene gy 2016. [C ossRe ]
11.
Nilsson, J.; Be ling, L. Main enance managemen o wind powe sys ems using condi ion moni o ing
sys ems—Li e cycle cos analysis o wo case s udies. IEEE T ans. Ene gy Con e s.
2007
,22, 223–229.
[C ossRe ]
12.
P a el, S.; Fauls ich, S.; Roh ig, K. Pe o mance and eliabili y o wind u bines: A e iew. Ene gies
2017
,10,
1904. [C ossRe ]
13.
Hahn, B.; Du s ewi z, M.; Roh ig, K. Reliabili y o Wind Tu bine–Expe iences o 15 yea s wi h 1500 WTs.
In Wind Ene gy; Peinke, J., Schaumann, P., Ba h, S., Eds.; Sp inge : Be lin/Heidelbe g, Ge many, 2006.
[C ossRe ]
14.
Kol sidopoulos Papa zimos, A.; Dawood, T.; Thies, P.R. Da a Insigh s om an O sho e Wind Tu bine
Gea box Replacemen . J. Phys. Con . Se . 2018,1104. [C ossRe ]
15.
Ka elus, J.; Mie inen, J.; Leh o aa a, A. De ec ion o gea pi ing ailu e p og ession wi h on-line pa icle
moni o ing. T ibiol. In . 2018,118, 458–464. [C ossRe ]
16.
A igao, E.; Koukou a, S.; Hon ubia-Esc ibano, A.; Ca oll, J.; McDonald, A.; Gómez-Láza o, E. Cu en
signa u e and ib a ion analyses o diagnose an in-se ice wind u bine d i e ain. Ene gies
2018
,11, 960.
[C ossRe ]
17.
Gonzalez, E.; S ephen, B.; In ield, D.; Mele o, J.J. Using high- equency SCADA da a o wind u bine pe o mance
moni o ing: A sensi i i y s udy. Renew. Ene gy 2019,131, 841–853. [C ossRe ]
18.
Elosegui, U.; Egana, I.; Ulazia, A.; Iba a-Be as egi, G. Pi ch angle misalignmen co ec ion based on
benchma king and lase scanne measu emen in wind a ms. Ene gies 2018,11, 3357. [C ossRe ]
19.
As ol i, D. A S udy o he Impac o Pi ch Misalignmen on Wind Tu bine Pe o mance. Machines
2019
,7, 8.
[C ossRe ]
20.
Rabanal, A.; Ulazia, A.; Iba a-Be as egi, G.; Sáenz, J.; Elosegui, U. MIDAS: A Benchma king Mul i-C i e ia
Me hod o he Iden i ica ion o De ec i e Anemome e s in Wind Fa ms. Ene gies 2019,12, 28. [C ossRe ]
21.
Tau z-Weine , J.; Wa son, S.J. Using SCADA da a o wind u bine condi ion moni o ing—A e iew.
IET Renew. Powe Gene . 2017,11, 382–394. [C ossRe ]
22.
Feng, Y.; Qiu, Y.; C ab ee, C.J.; Long, H.; Ta ne , P.J. Moni o ing wind u bine gea boxes. Wind Ene gy
2013
,
16, 728–740. [C ossRe ]
23.
Blanco, M.A.; Gibe , K.; Ma i-Puig, P.; Cusidó, J.; Solé-Casals, J. Iden i ying heal h s a us o wind u bines
by using sel o ganizing maps and in e p e a ion-o ien ed pos -p ocessing ools. Ene gies
2018
,11, 723.
[C ossRe ]
24.
Escale , X.; Meba ki, T. Full-Scale Wind Tu bine Vib a ion Signa u e Analysis. Machines
2018
,6, 63.
[C ossRe ]
25.
Dupuis, R. Applica ion o Oil Deb is Moni o ing Fo Wind Tu bine Gea box P ognos ics and Heal h
Managemen . In P oceedings o he Annual Con e ence o he P ognos ics and Heal h Managemen Socie y,
Po land, OR, USA, 10–16 Oc obe 2010.
26.
Sheng, S. Moni o ing o Wind Tu bine Gea box Condi ion h ough Oil and Wea Deb is Analysis: A Full-Scale
Tes ing Pe spec i e. T ibol. T ans. 2016. [C ossRe ]
100
Ene gies 2019,12, 3373 19 o 19
27.
Cuad a P a s, J. El clima de A agón. Geog a ía Física de A agón. Aspec os Gene ales y Temá icos; Asociación de
Geóg a os Español: Mu cia, Spain, 2004; pp. 15–26.
28.
Mabe, J.; Zubia, J.; Go i xa egi, E. Pho onic low cos mic o-senso o in-line wea pa icle de ec ion in
lowing lube oils. Senso s 2017,17, 586. [C ossRe ] [PubMed]
29.
Lopez, P.; Mabe, J.; Mi ó, G.; E xebe ia, L. Low cos pho onic senso o in-line oil quali y moni o ing:
Me hodological de elopmen p ocess owa ds unce ain y mi iga ion. Senso s
2018
,18, 2015. [C ossRe ]
[PubMed]
30.
B i ish S anda d. BS ISO 4406:1999 Hyd aulic Fluid Powe —Fluids—Me hod o Coding he Le el o Con amina ion
by Solid Pa icles; In e na ional O ganiza ion o S anda diza ion: Gene a, Swi ze land, 1999. [C ossRe ]
31.
Pea son, K. VII. No e on eg ession and inhe i ance in he case o wo pa en s. P oc. R. Soc. Lond.
1895
,
58, 240–242.
32.
Spea man, C. The p oo and measu emen o associa ion be ween wo hings. Am. J. Psychol.
1904
,15,
72–101. [C ossRe ]
33.
Ho elling, H. Analysis o a complex o s a is ical a iables in o p incipal componen s. J. Educ. Psychol.
1933
,
24, 417–441. [C ossRe ]
34.
Ul ich, J. TTR: Technical T ading Rules. R Package Ve sion 0.23-4. 2018. A ailable online: h ps://c an. -
p ojec .o g/web/packages/TTR/TTR.pd (accessed on 15 July 2019).
35.
Cle eland, W.; G osse, E.; Shyu, W. Local eg ession models. In S a is ical Models in S; Chambe s, J.M.,
Has ie, T.J., Eds.; So wa e Paci ic G o e: Wadswo h, OH, USA, 1992; pp. 309–376.
36.
The neau, T.; A kinson, B. pa : Recu si e Pa i ioning and Reg ession T ees. R Package Ve sion 4.1-13.
2018. A ailable online: h ps://c an. -p ojec .o g/web/packages/ pa / igne es/longin o.pd (accessed
on 15 July 2019).
37.
B eiman, L.; F iedman, J.; Olshen, R.; S one, C. Classi ica ion and Reg ession T ees; Wadswo h In . G oup:
Wadswo h, OH, USA, 1984; Volume 37, pp. 237–251.
c
2019 by he au ho s. Licensee MDPI, Basel, Swi ze land. This a icle is an open access
a icle dis ibu ed unde he e ms and condi ions o he C ea i e Commons A ibu ion
(CC BY) license (h p://c ea i ecommons.o g/licenses/by/4.0/).
101
Hyb id modelling o linea
ac ua o diagnosis in absence
o aul y da a eco ds
10
•
Au ho s: Lopez de Calle – E xabe, Ke man; Ruiz – Ca cel, C is obal; S a ,
And ew; Fe ei o, Susana; A naiz, Ai o ; Gomez, Me i xell; Sie a, Basilio;
• Publishe : Submi ed o Sp inge
• Jou nal: Submi ed o Jou nal o In elligen Manu ac u ing
• Yea : 2020
• Qua ile (Scimago/WoS): -
• DOI: -
103
1
Hyb id modelling o linea ac ua o diagnosis in
absence o aul y da a eco ds
Au ho s: Ke man López de Calle – E xabe12* , C is obal Ruiz – Ca cel 3, And ew
S a 3, Susana Fe ei o1, Ai o A naiz1, Me i xell Gómez12, Basilio Sie a2
1: TEKNIKER. Iñaki Goenaga s ee , 5 - 20600 Eiba , Gipuzkoa, Spain
2: UNIVERSITY OF THE BASQUE COUNTRY (UPV/EHU). Manuel La dizabal
Pasealekua, 1 - 20018 Donos ia-San Sebas ián, Gipuzkoa, Spain
3: CRANFIELD UNIVERSITY. College Rd, MK43 0AL, C an ield, Bed o d, UK
*: Co esponding au ho .
Keywo ds: Condi ion based main enance, aul diagnosis, linea ac ua o s,
hyb id modelling.
Abs ac :
The ad an ages o Condi ion Based Main enance o e al e na i e main enance s a egies ha e been widely
p o en. The use o de ec ion, diagnosis and p ognos ic algo i hms allows ope a o s o adjus schedules o op imise
he cos o main aining he machine y while a oiding b eakdowns and down imes. Howe e , he p e ious
echniques a e no always applicable o he whole ange o machines we ind in indus y. Pa icula ly in cases
whe e machines a e unique o no -mass-p oduced, i is di icul o ob ain aul y da a eco ds, and his limi a ion
complica es he de elopmen o diagnos ic moni o ing algo i hms.
This wo k p oposes an app oach ha combines he da a om physical models wi h da a-based models (also known
as hyb id modelling) o so ou he lack o aul y da a eco ds. Such app oach acili a es he de elopmen o
diagnosis algo i hms ha a e gi en in some indus ial scena ios. Fo doing so, aul y da a eco ds a e gene a ed
in a physical model and used o augmen he eal non- aul y da a eco ds so ha diagnos ic algo i hms can be
ained. This p ocedu e allows using da a-based algo i hms o de ec eal aul y da a and o analyse hei diagnos ic
po en ial.
Ou app oach is es ed in a condi ion moni o ing case s udy applied o linea ac ua o s. A speci ic es ig was
buil and used o collec da a om heal hy and aul y cases ( he la e only used o alida ion pu poses).
Addi ionally, a physical model which simula es nominal (heal hy) and aul y condi ions was used o gene a e
syn he ic da a. Syn he ic and eal measu ed da a we e combined o de elop a diagnos ic model and he algo i hm
was alida ed in he de ec ion o eal aul y cases. Besides he de ec ion o ce ain aul s, his app oach has been
p o ed o be aluable o de ec also unseen ope a ing condi ions.
The esul s ob ained in his wo k p o e he alidi y o hyb id models o hose cases in he indus y whe e he e
a e physical o economical limi a ions o ob ain ce ain da a eco ds ha , he e o e, di icul he implemen a ion
o diagnos ic algo i hms.
105
2
In oduc ion:
The p e alence o condi ion-based main enance (CBM) o e p e en i e and co ec i e main enance is a p o en
ac . Fu he mo e, good CBM s a egies should be based on he analysis and p edic ions o his o ical and eal-
ime da a, and should be a ailable o upda ing ecommenda ions as soon as new da a exis s (Bousdekis e al.,
2018). Howe e , he e a e some cases whe e ce ain da a ypes ( aul y eco ds) migh no be a ailable in he
his o ical da a. Consequen ly, he de elopmen o diagnos ic algo i hms is obs uc ed, as i is necessa y o possess
eco ds ob ained du ing aul y condi ion ope a ions o de elop hem. A he same ime, ha ing aul y machines is,
ob iously, undesi ed and a oided by he indus y. And only when many machines a e ope a ing, he e migh be
he chance o ob ain some aul ela ed da a. This scena io o lack o aul y da a is pa icula ly common in non-
mass-p oduced machines. The e o e, designing diagnos ic algo i hms unde hese ci cums ances is complex.
The al e na i e app oach he e p oposed consis s in he combina ion o physics-based and da a-based models. This
me hod also known as hyb id modeling, has been p e iously sugges ed by o he wo ks in he li e a u e and ies
o bene i om he ad an ages o bo h modelling echniques p ecision and applicabili y (Medjahe and Ze houni,
2013; Mish a e al., 2015). Some o he easons behind he use o hyb id models a e: he di icul y o ob aining
comple ely accu a e physical models (Ma ei e al., 2015); he cos o unning pa allel models (Balaban e al., 2015;
Na asimhan e al., 2010); he complexi y and lack o unde s anding o all he physics go e ning he sys em (An
e al., 2013; Benkedjouh e al., 2015); imp o ed pa e n ecogni ion o da a-d i en models (Balaban e al., 2015);
and no ha ing enough da a eco ds (mos ly ela ed o ailu es) (Mish a e al., 2015) o lack o aul g ow h models
(Balaban e al., 2015).
Rega ding he ield o applica ion, hyb idiza ion has been u ilized in a ange o di e en applica ions: o a ing
machine y (An e al., 2013; Le u iondo e al., 2017; Li e al., 2019; Qian e al., 2017), ailway swi ches (Ma ei e
al., 2015), ba e y li e p edic ion (Liao and Kö ig, 2014), mecha onic sys ems (Medjahe and Ze houni, 2013)
and, also, in elec o-mechanical ac ua o s (Balaban e al., 2015; Na asimhan e al., 2010).
Pa icula ly, his wo k ocuses on he moni o ing o linea ac ua o s, a widely ex ended mechanism o he
gene a ion o linea mo ion. In he pas , his linea mo ion p o iles we e p oduced wi h cam- ollowe and c ank-
od mechanisms. La e on, hese echniques e ol ed o hyd aulic and pneuma ic ac ua ion sys ems, which ha e
been ex ensi ely used o applica ions equi ing mo ion con ol o high o ces. In ecen imes, linea ac ua o s
a e ound in a a ie y o sys ems, such as al es, doo opene s, ai c a sys ems, machine ools, obo a ms, e c.
Fu he mo e, elec o-mechanical ac ua o s (EMA) in pa icula a e gaining popula i y in ae onau ics, as hey ha e
some ad an ages o e he ypically used hyd aulic ac ua o s, such as inc eased sa e y and eliabili y, easie and
educed main enance, educed weigh , olume and complexi y o ansmission pa hs, and a highe e iciency
(Qiao e al., 2018). Howe e , his echnology is no ma u e ye , and needs u he esea ch.
Cu en ly, he e a e wo main ends in he la es esea ch ega ding linea ac ua o s.
Fi s , some au ho s ha e ocused on he design and de elopmen o aul ole an ac ua o s, a.k.a. high edundancy
ac ua o s. The unde laying idea behind aul ole an ac ua o s is designing ac ua o s ha a e composed by se e al
small ac ua ion elemen s coupled, so ha hey will be able o wo k e en unde aul y condi ions, his duplici y in
ac ua o s gi es he name high edundancy ac ua o s (HRA). HRA was inspi ed by human muscula u e. They
mimic human muscles, ha a e composed by many indi idual muscle cells each o hem con ibu ing o he o ce
and a el o he muscle (Da ies e al., 2008). Cu en esea ch ocuses on he di e en ypes o a chi ec u es hese
ac ua o s ha e, and in he e ec hese a chi ec u es ha e in he esilience o he ac ua o . Fo example, in he wo k
by (An ong e al., 2014) a 3 se ial elemen in 4 pa allel s uc u e is modelled, la e he model is alida ed wi h a
eal ig in (An ong e al., 2016) conside ing open-loop and close-loop con ols. O he wo ks such as he one by
(Manoha e al., 2018) model a 3×3 se ies-in-pa allel a chi ec u e HRA and seed aul s, hey s udy he deg ada ion
and he possibili y o keep ac ua ing e en unde 3 aul y ac ua o s.
In ela ion o he la es wo ks in he ield o condi ion moni o ing o linea ac ua o s, ecen esea ch ocuses on
he de ec ion and diagnosis o aul s and he emaining use ul li e es ima ion o he ac ua o s. Fo ha pu pose
moni o ing sys ems a e ypically based on posi ion e o and cu en (Ruiz-Ca cel and S a , 2015, 2018a). Also,
in lesse ex en ib a ion (Balaban e al., 2015; Sudhawiyangkul and Isa ako n, 2017) when o a ing elemen s a e
in ol ed (balls c ew/ ack and pinion), empe a u e (Balaban e al., 2015) and also acous ic emissions could be
used as Eh man colla es (Eh mann e al., 2016). The ypical aul s ha a e conside ed a e: lack o lub ica ion,
106
3
spalling, backlash (Ruiz-Ca cel and S a , 2018a), ex e nal o ces (jamming)(Susana Fe ei o e al., 2013),
including some o he s mo e ela ed o he cu en supply o EMAs, such as open ci cui , close ci cui (Cai e al.,
2017); acco ding o Na asimhan (Na asimhan e al., 2010) and Balaban (Balaban e al., 2015) in addi ion o he
p e ious aul s, winding sho s, and common senso aul s like bias, d i and scaling should be conside ed. Linea
ac ua o moni o ing sys ems ha e also conside ed a ie y o ope a ing condi ions, which includes, a ious loads
and mo ion p o iles as sinusoidal, apezoidal (Ruiz-Ca cel and S a , 2018a), sine sweep and iangula
(Na asimhan e al., 2010). I is also common o ind some p e- ea men o he signals be o e plugin hem in o he
diagnosis o p ognosis models. In ha sense, hey can be as simple as s a is ical desc ip o s aken om ime
domain (Ruiz-Ca cel and S a , 2018a) going h ough he use o as ou ie ans o m (FFT) o ib a ions
(Sudhawiyangkul and Isa ako n, 2017) o cu en s (Cai e al., 2017) o equency domain going o he mo e
complex ime- equency domain echniques such as Disc e e Wa ele T ans o m (DWT), Wigne -Ville-
Dis ibu ion (WVD) o he Mel-F equency-Ceps al-Coe icien (MFCC) ha Knöbel employs (Knöbel e al.,
2015). And, o he educ ion o he dimensionali y o he da a, p incipal componen analysis (PCA) mus be
no ed, o i s use is widely ex ended as he amoun o wo ks using i sugges s (Cai e al., 2017; Knöbel e al., 2015;
Mazzoleni e al., 2019; Ruiz-Ca cel and S a , 2018a). Rega ding he inal diagnosis models, he wo ks ocusing
in aul diagnosis ha e used di e en algo i hms o ha pu pose, o example, (Cai e al., 2017) employs Bayesian
ne wo ks, (Knöbel e al., 2015) uses suppo ec o machines and Neu al Ne wo ks a e used by (Balaban e al.,
2015).
Ac ua o moni o ing sys ems ha e also bene i ed om he use o physical models. In ha ega d, some di e en
ends ha en been iden i ied. The wo ks by (Balaban e al., 2015) o (Kemp and Ma in, 2018) un physical models
and compa e he model da a o he one gene a ed in eali y, u he mo e, (Kemp and Ma in, 2018) uns addi ional
aul y models o de ec which aul shows smalle esiduals and is he e o e he one he happening in he ig. Ruiz-
Ca cel (Ruiz-Ca cel and S a , 2015) uses he model o alida e a s a is ical p ocess con ol (SPC) based
moni o ing echnique ha is la e alida ed in a eal ig in (Ruiz-Ca cel and S a , 2018a). Simila ly, (Susana
Fe ei o e al., 2013) and (Cai e al., 2017) and (Knöbel e al., 2015) use da a om he physical model o ain
machine lea ning diagnos ic algo i hms. Fe ei o uses a da a se ully gene a ed wi h physical model da a om
he design s age o he ac ua o , Knöbel uses nominal da a om he ig and en iches he da ase wi h aul s om
he physical model and, las ly, Cai combines he da a gene a ed in he ig oge he wi h he one om he physical
model. No e ha he diagnos ic models used in he p e ious wo ks a e ained wi h ully syn he ic (gene a ed in
he physical model) o da a combined om he model and he eal use case. In con as , he scena io he e p esen ed
is one in which ob aining eal aul y da a is no possible.
In his scena io, he wo k he e p esen ed con inues wi h he esea ch ini ia ed in (Ruiz-Ca cel and S a , 2015)
whe e a da a-based anomaly de ec ion algo i hm was de eloped o an EMA and es ed in a physical model. La e ,
in (Ruiz-Ca cel and S a , 2018a), his app oach was imp o ed and alida ed in a eal ac ua o . This second wo k
showed he capabili y o de ec ing ailu es and o gi e addi ional indica ions o which ea u es we e mo e dis an
om no mali y. The esea ch wo k p esen ed in his pape deals wi h he au oma ed diagnosis o he aul s and
copes wi h a scena io ha is qui e equen in he indus y: No ha ing aul y eco ds. The inal aim o he wo k
is o p o e ha , wi hou ha ing eal aul y eco ds, i is possible o lea n o dis inguish and iden i y ailu es wi hou
seeing hem in he eal li e in ad ance. Tha is, we a emp o an icipa e o eal ailu es by using syn he ic aul y
da a p oduced by he physical model. Fo ha eason, eal aul y da a is only used o alida ion pu poses along
he wo k.
Me hodology:
Essen ially, o en ich he eal da a wi h unseen aul s, i is necessa y o ha e a eliable physical model ha , besides
con empla ing he same ope a ing condi ions as he eal use case, mus ep esen also aul y condi ions. Fo doing
so he schema displayed in he ollowing Figu e 1 has been ollowed:
107
4
Figu e 1: Schema o da a pipeline.
The es ig is used o gene a e da a in a ious ope a ing condi ions. Fu he mo e, some mechanical aul s can be
seeded in he es ig. In pa allel, a i s app oxima ion o he physical model is ca ied ou , and, once he eal
nominal da a is a ailable, i is possible o adjus he model o beha e as he eal use case. Then, he expec ed aul s
a e seeded in he physical model. La e , ea u es a e ex ac ed om bo h eal nominal and syn he ic da a and,
using he noise o iginally ound in nominal da a, mo e syn he ic da a samples a e c ea ed and he ea u es ha do
no esemble simila i y be ween eal and syn he ic da a a e emo ed. Finally, he syn he ic da a augmen a ion can
be alida ed by aining a Machine Lea ning algo i hm ha is es ed a e wa ds on he de ec ion o eal aul s.
The ollowing subsec ions desc ibe o p ocess in mo e de ail:
1. Tes ig:
The main elemen o he es ig consis s o a ball-sc ew mechanism wi h a h eaded sha con aining a helical
aceway o he displacemen o he bea ing balls housed inside he nu . Va ying loads a e gene a ed by a aching
a seconda y ac ua o . The ac ua o s a e connec ed h ough a load cell, so ha he cell p o ides eed back o he
con olle o he seconda y ac ua o . Wi h his con olle , di e en ope a ing condi ions can be ep esen ed by
changing he load se poin . The commanded load se poin s o he second ac ua o a e: 20 Kg , 40 Kg and -40
Kg .
The signal p o ided by a linea ansduce is used by he posi ion con olle o command he main ac ua o o
p oduce wo mo ion p o iles: apezoidal (cons an speed) and sinusoidal (smoo h accele a ion and decele a ion).
Howe e , only apezoidal p o iles a e conside ed in his wo k. These apezoidal p o iles a e de ined o comple e
a 120mm ex ension in 5 seconds, s op o 3 seconds and e ac in ano he 5 seconds. Mo o cu en in his ac ua o
is also measu ed o moni o ing pu poses.
In addi ion, 3 ypes o aul s a e seeded wi h inc easing se e i ies: he bol in he holding he seal o he is igh ened
o inc ease ic ion simula ing lack o lub ica ion; su aces o he sc ew a e damaged s a ing wi h 1 mm diame e
de ec s and g adually inc easing up o 4 mm diame e de ec s including he emo al o he sidewall o neighbo ing
channels; las ly, backlash de ec is ob ained by subs i u ing he balls (o iginally 3.15 mm o diame e ) by smalle
ones (3 mm and 2.5 mm diame e ).
Fu he de ails ega ding he es ig and he seeded aul s can be ound in (Ruiz-Ca cel and S a , 2018a). Each
es in he inal da ase con ains 5 mo ion cycles, and each es is epea ed 10 imes. The ull da a se con ains es s
o bo h mo ion p o iles (sinusoidal and apezoidal), he a ying load condi ions (20 Kg , 40 Kg and -40 Kg ),
and he no mal and aul y condi ions. The whole da ase is accessible a (Ruiz-Ca cel and S a , 2018b).
108
5
2. Physical model:
The model o he ac ua o ollows he model de eloped in (Ruiz-Ca cel and S a , 2018a) and was de eloped in
Ma lab using Simscape oolbox. The model con ains an elec ical mo o which is con olled by a PID using he
posi ion measu emen and desi ed se poin , a ack and pinion block ha ans e s he o a ional mo ion o linea
one and, o simplici y sake, an equi alen o ce in subs i u ion o he seconda y ac ua o . In addi ion, he model
also includes an equi alen o a ional ine ia, a mass, and linea ic ion block. The physical model allows a ying
he in ensi y o he o ce o simula e he same ope a ing condi ions ha a e es ed in he es ig.
2.1. Model alida ion
Fo he combina ion o da a om he physical model and he ig i is necessa y ha hey coincide. The e o e, he
physical model mus be alida ed and adjus ed i necessa y. The Ma lab Op imiza ion oolbox was used o he
adjus men o he physical model using nominal displacemen and cu en measu emen s om he es ig o he
alida ion and adjus men .
The inal signals om he es ig and he signals om he physical model a e displayed in he ollowing Figu e
2:
Figu e 2: Real and syn he ic signals. a) Real cu en b) Real posi ion e o c) Syn he ic cu en d) Syn he ic posi ion e o
2.2. Faul seeding
Acco ding o ou app oach, o he pu pose o o eseeing aul s i is necessa y o seed hem i s in he physical
model. This wo k ocuses on he diagnos ic o lack o lub ica ion and spalling aul s. Lack o lub ica ion is
in oduced in he physical model by inc easing he ansla ional ic ion coe icien du ing he whole ansla ion
o he ac ua o , which is done wi h inc easing ic ion alues o s udy g adually inc easing se e i ies. Rega ding
he spalling aul , i is simula ed by a se e e inc ease o ic ion in he pa icula poin o he mo ion pa h whe e
he aul is loca ed on he sc ew. This es is also done wi h a ied ic ion alues co esponding o mo e se e e
spalls.
3. Fea u e ex ac ion:
Raw signals con ain in o ma ion ha can be exploi ed o de elop diagnos ic algo i hms. Howe e , edundan da a
is emo ed and key indica o s a e ypically ex ac ed om he aw signal. This p ocess is known as ea u e
ex ac ion. The ea u e ex ac ion schema adop ed in his wo k ollows he one in (Ruiz-Ca cel and S a , 2018a).
Howe e , due o he added di icul y o manipula ing wo dis inc da a sou ces (syn he ic and eal) only he
simples ea u es a e kep educing he di e ences be ween da a sou ces.
109
12
Figu e 9: Con ex de ec ion wi h and wi hou PCA p e- ea men and wi h and wi hou using eal da a.
Bes -case scena io
Las ly, in o de o ha e a baseline o compa ison, he bes cases we e compa ed o he hypo he ical o ha ing he
eal aul y da a ep esen ing he bes -case scena io, in which aul y eco ds a e a ailable and he e o e, he e is no
limi a ion no need o using a physical model o gene a e syn he ic da a o aul s.
Due o i s low pe o mance, Load condi ion -40 was le aside du ing his compa ison, consequen ly, load
condi ions 40 and 20 we e combined. As in p e ious es s, syn he ic da a om aul y and nominal cases was
en iched wi h nominal eal da a, in his es , howe e , i is compa ed o a model ained o e da a ully o igina ed
in he es ig. The se o ea u es was educed o he inal ea u es p esen ed abo e and he e ec o ch onology
was compa ed agains indi idual p edic ion. Addi ionally, conside ing he pa icula i y o he da ase ha , only
Kappa s a is ic was measu ed, and simple Kappa baselines we e s ablished o bo h mul iclass (conside ing
nominal, spalling aul and lub ica ion aul ) and bina y (spalling aul and lub ica ion aul ) cases. Fo he
mul iclass case, he baseline p edic ion conside ed ha e e y obse a ion belonged o Spalling aul (class wi h
g ea es numbe o obse a ions), o bina y, se e al lub ica ion aul misclassi ica ion a es we e included.
Figu e 10: Compa ison o esul s agains he bes -case scena io (ha ing eal da a) e ical ed lines ep esen addi ional
baseline scena ios. a) Mul iclass p oblem (nominal/spalling/lub ica ion). b) Bina y class p oblem (lub ica ion/spalling).
Discussion:
Di icul ies ha e a isen du ing he de elopmen o he es s, such as he di icul y o ob ain a physical model wi h
enough accu acy o ep oduce aul y da a wi h ealis ic cu en and posi ion e o signals. Al hough his has been
pa ially accomplished (see he small di e ences in ig 2), he me hod p oposed o dimensionali y educ ion has
educed he misma ch be ween he obse a ions p oduced in he model and he eal ones, which is p o en by he
huge pe o mance disag eemen be ween he diagnosis esul s be ween “All F.” and “Final F.” ha Figu e 6 shows.
Fu he mo e, when PCA is used nai ely o e all ea u es, he esul s is s ill much wo se han ou p oposed me hod
(see “ALL F. PCA” in he same igu e).
Rega ding he abili y o diagnose aul s, besides he al eady men ioned p e alence o he educed se o ea u es
he i s diagnosis i e a ion (Figu e 6, Figu e 7 and Figu e 8) e eals some in e es ing indings. On he one hand,
116
13
using ch onology is he bes s a egy in compa ison o he o he s (using se e e cases o using augmen a ion), in
any case, none o he p e ious ha e an ex eme con as in compa ison o di ec ly using he inal se o ea u es
(using augmen a ion is e en wo se han jus using he ea u es). On he o he , same majo conce ns a ise: i s ly,
when modelling each load sepa a ely load -40 gene a es ca as ophic esul s (which is possible educing he
o e all accu acy o he p e ious models) in compa ison o he es o loads (Figu e 8); and, he e a e qui e big
di e ences om accu acy o kappa me ics, leading o hink he e is a class which is no being de ec ed p ope ly,
which is ob ious when looking o he Figu e 7 as he class wise accu acy o nominal cases is close o null. This
lack o sensi i i y wi h nominal label (which is al eady e lec ed by he con as be ween accu acy and kappa
me ics) migh be caused by he combina ion o he ollowing ac o s:
- Se e i y: Ini ial deg ada ion s ages o he spalling aul do no ha e much impac no in he ope a ion no
in he signal, he e o e, dis inguishing hem om nominal cases is non- i ial (o e en no possible).
- In e mi ence: Spalling aul does no mani es in all he cycles. Some imes, e en i he spall is in i s
bigges size, he balls managed o un smoo hly h ough he bol , which leads o an inco ec labelling o
spalling aul .
- Imbalance: As he e is a ixed numbe o obse a ions pe se e i y case, he numbe o obse a ions o
he da ase ela ed o spalling aul is g ea e han he one o nominal cases (single se e i y case
conside ed, see Table 1). This class imbalance combined o he p e ious ac o s g ea ly a ec s he
gene aliza ion capabili y o he algo i hms. No e ha in igu e Figu e 7 he only es wi h balanced class
wise accu acies is he augmen ed case which imp o es he nominal case de ec ion a he expense o
educing o e all accu acy.
This di icul y o disce n nominal cases and spalling aul s is p esen in e en mo e a o able scena ios as igu e
Figu e 10 sugges . Rega dless o no conside ing load case -40, he o e all diagnosis o mul iclass case (Figu e 10
a) ) is lowe han he simple baseline o conside ing e e y hing as spalling, due o he a o emen ioned easons.
Ne e heless, he bina y classi ica ion p oblem (Figu e 10 b)) shows a be e esul s. Fi s ly, i demons a es ha
he use o ch onology imp o es kappa, as in bo h cases baseline and ou app oach, esul s using ch onology ha e
highe accu acies. Secondly, i p o es ha he aul s can be disce ned unde no eal da a condi ions wi h a hyb id
model wi h li le chance o e o . As he baselines sugges only abou 5% o he lub ica ion aul s would be
misma ched as spalling, possibly co esponding o cases wi h e y mild lub ica ion aul .
Howe e , o making he p e ious scena io possible, i is necessa y o de ec he loads unde which he algo i hm
is wo king, as some load condi ions o e e y low accu acies (L -40). This ac is p obably ela ed o he physical
model being be e adap ed o some load cases han o he s. In any case, de ec ing he ope a ing con ex is no
e y challenging as he Figu e 9 shows ha ex emely high accu acies a e eached in load de ec ion e en when
no eal da a is used. This has e y in e es ing implica ions, as i shows ha a physical model can ep oduce unseen
scena ios ha can be la e lea ned by machine lea ning models.
Conclusions:
Hyb id models and digi al wins ha e ecen ly gained a en ion in he li e a u e. This wo k demons a es he alue
o ha ing a physical model, win o he moni o ed asse , o he imp o emen o he diagnosis and, also, he
de ec ion o new ope a ion con ex s. In o he wo ds, his wo k shows how da a limi a ions (one o he majo
d awbacks o da a-based models) can be pa ially o e come wi h he da a c ea ed om he physical model.
This s udy has ied o gain insigh in he expansion o he bounda ies diagnos ic algo i hms ha e when dealing
wi h scena ios wi h no aul y eco ds. Fo ha pu pose, da a om a es ig has been used o ep esen a eal use
case and a physical model has been used o c ea e da a o aul y and nominal/heal hy cases. A eal scena io wi h
no aul y da a has been simula ed by using only syn he ic da a o ain he diagnos ic algo i hms, la e , he
pe o mance o ou app oach has been e alua ed by es ing he p edic ions wi h he eal aul y da a.
Acco ding o ou esul s, using physical models o augmen he inpu da a pool o diagnos ic algo i hms seems a
iable way o p o iding a good s a ing poin o diagnos ic algo i hms. The esul s we ha e ob ained show a
s ong capabili y o di e en ia e among aul s, wi h a wo se pe o mance in he dis inc ion among nominal and
spalling aul s. Howe e , we ha e p o ed ha , e en i he eal da a was a ailable o he modelling, he same
p oblem would ha e occu ed due o he in e mi en beha io , he imbalance among classes and he so ailu e
se e i ies. Hence, aul s a e no i ial o de ec e en wi h he eal da a.
117
14
Addi ionally, besides he use in aul de ec ion, he possibili y o de ec di e en con ex in an ex emely accu a e
way is p o en, e en i hey we e p e iously unseen. This implies ha algo i hms’ ho izons can be expanded o
de elop no mali y models in ope a ion egions p e iously unobse ed o ha diagnos ic algo i hms ha only make
p edic ions in he egions hey pe o m op imally can be de eloped.
As sugges ed by (Cai e al., 2017), he ob ious di e ences among eal and syn he ic da a ha e led o di icul ies
du ing he in eg a ion o he da a. The model was alida ed wi h eal nominal da a, ye he sligh di e ences in
he na u e o he signals has caused he da a o no ma ch exac ly. As coun e measu e, a ea u e emo al me hod
has been p oposed. The me hod is based in he ecu si e elimina ion o ea u es by moni o ing he diagnos ic
capabili y and hei capabili y o be used o de ec he o igin o he da a sou ce and inding a poin wi h an
accep able ade-o . Ou esul s show ha he me hod has made a signi ican di e ence in he inal diagnosis,
and we belie e his me hod is gene ic enough o be ans e ed o o he wo ks ha y o combine physical model
da a and eal da a. Fu he mo e, ou esul s show signi ican imp o emen in compa ison o he PCA, which is he
p e e ed ool acco ding o ou e iew.
Also, he added alue o domain knowledge has been mani es ed. The be e unde s anding o he p oblem ( aul s
should no ix by hemsel es in subsequen cycles) has allowed o make use o ch onological in o ma ion by adding
a majo i y o ing laye o he diagnosis algo i hm, which has shown he bes accu acy compa ed o he es o
diagnosis algo i hms.
Las ly, o he sake o ep oducibili y and scien i ic igo , we ha e gi en open access o he p ocedu e and esul s
we ha e ob ained by publishing he code used o he p e-p ocessing and he analysis, as well as o he da a se s
used. We ha e he s ong belie ha his p ac ice will ease he ep oduc ion o ou esul s and gi es o he
esea che s he oppo uni y euse pa o ou wo k o o imp o e i .
In espec o he loose ends le in he esea ch, wo in e es ing esea ch ields ha e been iden i ied: he poo esul s
in he dis inc ion o heal hy and spalling aul s sugges he need o me hods o imp o e he de ec ion o in e mi en
aul s and deal wi h imbalance; and, secondly, he possibili y o s udy how he diagnosis abili y could be imp o ed
as soon as aul y da a poin s appea . We belie e he o me could be add essed by using al e na i e policies o
majo i y o ing, policies ha could explo e di e en ways o combing diagnosis p edic ions o inc ease he
de ec ion o hose in e mi en kind o aul s. Rega ding he la e , a possible di ec ion could be he use o online
machine lea ning algo i hms, a b anch o machine lea ning ha deals wi h algo i hms ha a e upda ed as soon as
new da a becomes a ailable
Bibliog aphy:
An, D., Choi, J.H., Kim, N.H., 2013. Op ions o P ognos ics Me hods: A e iew o da a-d i en and physics-based
p ognos ics, in: 54 h AIAA/ASME/ASCE/AHS/ASC S uc u es, S uc u al Dynamics, and Ma e ials
Con e ence. p. 1940.
An ong, H., Dixon, R., Wa d, C., 2014. Modelling and building o expe imen al ig o high edundancy ac ua o ,
in: 2014 UKACC In e na ional Con e ence on Con ol (CONTROL). pp. 384–388.
An ong, H., Dixon, R., Wa d, C.P., Wa d, P., An ong, H., Dixon, R., Ch is ophe , P., Wa d, P., 2016.
ScienceDi ec Elemen s : Closed-loop Elemen s : Closed-loop Elemen s : Open- and Closed-loop Model.
IFAC-Pape sOnLine 49, 254–259. h ps://doi.o g/10.1016/j.i acol.2016.10.563
Balaban, E., Saxena, A., Na asimhan, S., Roychoudhu y, I., Koopmans, M., O , C., Goebel, K., 2015. P ognos ic
heal h-managemen sys em de elopmen o elec omechanical ac ua o s. Jou nal o Ae ospace
In o ma ion Sys ems 12, 329–344.
Benkedjouh, T., Medjahe , K., Ze houni, N., Rechak, S., 2015. Heal h assessmen and li e p edic ion o cu ing
ools based on suppo ec o eg ession. Jou nal o In elligen Manu ac u ing 26, 213–223.
118
15
Bousdekis, A., Magou as, B., Apos olou, D., Men zas, G., 2018. Re iew, analysis and syn hesis o p ognos ic-
based decision suppo me hods o condi ion based main enance. Jou nal o In elligen Manu ac u ing
29, 1303–1316.
Cai, B., Zhao, Y., Liu, H., Xie, M., 2017. A Da a-D i en Faul Diagnosis Me hodology in Th ee-Phase In e e s
o PMSM D i e Sys ems. IEEE T ansac ions on Powe Elec onics 32, 5590–5600.
h ps://doi.o g/10.1109/TPEL.2016.2608842
Da ies, J., S e en, T., Dixon, R., Goodall, R.M., Zolo as, A.C., Pea son, J., 2008. Modelling o high edundancy
ac ua ion u ilising mul iple mo ing coil ac ua o s. IFAC P oceedings Volumes 41, 3228–3233.
Eh mann, C., Isabey, P., Fleische , J., 2016. Condi ion moni o ing o ack and pinion d i e sys ems: necessi y and
challenges in p oduc ion en i onmen s. P ocedia CIRP 40, 197–201.
Kemp, M., Ma in, E.J., 2018. Faul isola ion o an elec o-mechanical linea ac ua o , in: P oceedings o
P ognos ics and Heal h Managemen Socie y Con e ence, 1. P esen ed a he Annual Con e ence o he
PHM Socie y. h ps://doi.o g/10.36001/phmcon .2018. 10i1.539
Knöbel, C., Ma sil, Z., Rekla, M., Reu e , J., Gühmann, C., 2015. Faul de ec ion in linea elec omagne ic
ac ua o s using ime and ime- equency-domain ea u es based on cu en and ol age measu emen s,
in: 2015 20 h In e na ional Con e ence on Me hods and Models in Au oma ion and Robo ics (MMAR).
pp. 547–552.
Le u iondo, U., Salgado, O., Gala , D., 2017. Valida ion o a physics-based model o a o a ing machine o
syn he ic da a gene a ion in hyb id diagnosis. S uc u al Heal h Moni o ing 16, 458–470.
h ps://doi.o g/10.1177/1475921716676053
Li, Z., Wu, D., Hu, C., Te penny, J., 2019. An ensemble lea ning-based p ognos ic app oach wi h deg ada ion-
dependen weigh s o emaining use ul li e p edic ion. Reliabili y Enginee ing & Sys em Sa e y 184,
110–122.
Liao, L., Kö ig, F., 2014. Re iew o hyb id p ognos ics app oaches o emaining use ul li e p edic ion o
enginee ed sys ems, and an applica ion o ba e y li e p edic ion. IEEE T ansac ions on Reliabili y 63,
191–207.
Lopez de Calle – E xabe, K., Ruiz – Ca cel, C., S a , A., Fe ei o, S., A naiz, A., Gomez, M., 2020. Hyb id
modelling o linea ac ua o diagnosis in absence o aul y da a eco ds. Gi Hub eposi o y.
h ps://doi.o g/10.5281/zenodo.3961536
Manoha , G.A., Vasu, V., S ikan h, K., 2018. ScienceDi ec Modeling and simula ion o high edundancy ac ua o
o aul - ole ance. Ma e ials Today: P oceedings 5, 18867–18873.
h ps://doi.o g/10.1016/j.ma p .2018.06.234
Ma ei, I., Ganguli, A., Honda, T., de Klee , J., 2015. The case o a hyb id app oach o diagnosis: A ailway
swi ch., in: DX@ Sa ep ocess. pp. 225–234.
Mazzoleni, M., P e idi, F., Scandella, M., Pispola, G., 2019. Expe imen al De elopmen o a Heal h Moni o ing
Me hod o Elec o-Mechanical Ac ua o s o Fligh Con ol P ima y Su aces in Mo e Elec ic Ai c a s.
IEEE Access 7, 153618–153634. h ps://doi.o g/10.1109/ACCESS.2019.2948781
Medjahe , K., Ze houni, N., 2013. Hyb id p ognos ic me hod applied o mecha onic sys ems. The In e na ional
Jou nal o Ad anced Manu ac u ing Technology 69, 823–834.
Mish a, M., Le u iondo, U., Salgado, O., Gala , D., 2015. Hyb id modelling o ailu e diagnosis and p ognosis
in he anspo sec o : Acqui ed da a and syn he ic da a. Dyna 90, 139–145.
119
16
Na asimhan, S., Roychoudhu y, I., Balaban, E., Saxena, A., 2010. Combining model-based and ea u e-d i en
diagnosis app oaches-a case s udy on elec omechanical ac ua o s.
Qian, P., Ma, X., C oss, P., 2017. In eg a ed da a-d i en model-based app oach o condi ion moni o ing o he
wind u bine gea box. IET Renewable Powe Gene a ion 11, 1177–1185.
Qiao, G., Liu, G., Shi, Z., Wang, Y., Ma, S., Lim, T.C., 2018. A e iew o elec omechanical ac ua o s o
Mo e/All Elec ic ai c a sys ems. P oceedings o he Ins i u ion o Mechanical Enginee s, Pa C:
Jou nal o Mechanical Enginee ing Science 232, 4128–4151.
Ruiz-Ca cel, C., S a , A., 2018a. Da a-based de ec ion and diagnosis o aul s in linea ac ua o s. IEEE
T ansac ions on Ins umen a ion and Measu emen 67, 2035–2047.
Ruiz-Ca cel, C., S a , A., 2018b. Da a se o “Da a-based De ec ion and Diagnosis o Faul s in Linea Ac ua o s.”
h ps://doi.o g/10.17862/c an ield. d.5097649
Ruiz-Ca cel, C., S a , A., 2015. De elopmen o a no el condi ion moni o ing ool o linea ac ua o s, in: The
Twel h In e na ional Con e ence on Condi ion Moni o ing and Machine y Failu e P e en ion
Technologies. pp. 1–12.
Sudhawiyangkul, T., Isa ako n, D., 2017. Design and ealiza ion o an ene gy au onomous wi eless senso sys em
o ball sc ew aul diagnosis. Senso s and Ac ua o s A: Physical 258, 49–58.
Susana Fe ei o, And es Jimenez, Jon Mada iaga, E a No illo, San iago Fe nandez, And es Alia, Es eban
Mo an e, 2013. Heal h moni o ing o elec o-mechanical nose landing gea doo ac ua o o a ua , based
on simula ion modelling and da a-d i en echniques. Chemical Enginee ing T ansac ions 33, 655–660.
h ps://doi.o g/10.3303/CET1333110
Wilson, G., B yan, J., C ans on, K., Ki zes, J., Nede b ag , L., Teal, T.K., 2017. Good enough p ac ices in
scien i ic compu ing. PLoS Compu Biol 13, e1005510. h ps://doi.o g/10.1371/jou nal.pcbi.1005510
Decla a ion:
All manusc ip s mus con ain he ollowing sec ions unde he heading 'Decla a ions'.
I any o he sec ions a e no ele an o you manusc ip , please include he heading and
w i e 'No applicable' o ha sec ion.
To be used o non-li e science jou nals
Funding (in o ma ion ha explains whe he and by whom he esea ch was suppo ed)
Con lic s o in e es /Compe ing in e es s (include app op ia e
disclosu es)
Au ho s decla e no con lic s o in e es .
A ailabili y o da a and ma e ial (da a anspa ency)
120
17
Code a ailabili y (so wa e applica ion o cus om code)
Au ho s' con ibu ions (op ional: please e iew he submission guidelines om
he jou nal whe he s a emen s a e manda o y)
No applicable
To be used o li e science jou nals + a icles wi h biological applica ions
E hics app o al (include app op ia e app o als o wai e s)
No applicable
Consen o pa icipa e (include app op ia e s a emen s)
No applicable
Consen o publica ion (include app op ia e s a emen s)
No applicable
121
Some inal hough s 11
„Asi beha da a ja ai u, au e a uko bada.
Asie an ez di a gauzac oso uac ecus en. Gueldica
egui en di a au e apenac
—Ipui oncak (1804) by Vicen a Mogel
(Fi s Basque emale w i e )
An onia Vicen a Moguel li ed du ing a ime when Basque language was conside ed
o be a language o uneduca ed illage s, u he mo e, w i ing was he kind o du y
a lady could no do. The b a e An onia p o ed hose belie s w ong when she w o e
a book in Basque adap ing some ables om Aesop. The book she w o e, Ipui onac,
included he quo e ha begins his chap e , which ansla es o some hing like "You
ha e o s a and con inue i you gonna e e mo e o wa d. A i s you won
´
see
ma e s inished. P og ess is made ’s ep by s ep’ ".
Two cen u ies la e , my g andmo he Agus ina Solozabal, who li es in he illage
besides An onia’s (and sha es An onia’s b a e y as well), ound he way o syn hesise
he whole quo e in o a sho ye p ecise exp ession, and managed o ans e his
pill o knowledge o he descendan s. He philosophy, "Txiki i- xiki i", ep esen s
he i eless e o ha keeps d i ing humani y one s ep beyond. The kind o e o
ha d aws a smile in he a elle ’s ace when he/she looks back o he now dis an
s a ing poin .
En olling on a PhD is saying yes o a olle -coas e jou ney whe e emo ions ha e
cons an ups and downs (as li e i sel ) due o pape ejec ions, pape publica ion
accep ances, aul y expe imen s (do expe imen s e e go as expec ed?) and o he
icissi udes ha one aces du ing his complex jou ney. Howe e , I would like o
gi e a piece o ad ice o hose who doub whe he o s a he jou ney o no , o
hose o he whose ene gies ha e been d ained by he slope o he jou ney: Txiki i-
xiki i.
I would like o hank you, o you who ha e ead so a (e en i you migh ha e
jumped some chap e s o pages) and desi e good luck in you own jou ney.
123
124 Chap e 11 Some inal hough s
