applied
sciences
A icle
A Decision-Making Tool Based on Explo a o y
Visualiza ion o he Au omo i e Indus y
Raquel Redondo 1,* , Ál a o He e o 1, Emilio Co chado 2and Ja ie Sedano 3
1G upo de In eligencia Compu acional Aplicada (GICAP), Depa amen o de Ingenie ía In o má ica,
Escuela Poli
é
cnica Supe io , Uni e sidad de Bu gos, A . Can ab ia s/n, 09006 Bu gos, Spain; [email p o ec ed]
2Depa amen o de In o má ica y Au omá ica, Uni e sidad de Salamanca, Plaza de la Me ced s/n,
37008 Salamanca, Spain; [email p o ec ed]
3Ins i u o Tecnológico de Cas illa y León, Pol. Ind. Villalonqueja , C/López B a o 70, 09001 Bu gos, Spain;
ja ie [email p o ec ed]
*Co espondence: [email p o ec ed]
Recei ed: 24 May 2020; Accep ed: 22 June 2020; Published: 25 June 2020


Abs ac :
In ecen yea s, he digi al ans o ma ion has been ad ancing in indus ial companies,
suppo ed by he Key Enabling Technologies (Big Da a, IoT, e c.) o Indus y 4.0. As a consequence,
companies ha e la ge olumes o da a and in o ma ion ha mus be analyzed o gi e hem compe i i e
ad an ages. This is o he u mos impo ance in ields such as Failu e De ec ion (FD) and P edic i e
Main enance (PdM). Finding pa e ns in such da a is no easy, bu cu ing-edge echnologies, such as
Machine Lea ning (ML), can make g ea con ibu ions. As a solu ion, his s udy ex ends Hyb id
Unsupe ised Explo a o y Plo s (HUEPs), as a isualiza ion echnique ha combines Explo a o y
P ojec ion Pu sui (EPP) and Clus e ing me hods. An ex ended o mula ion o HUEPs is p oposed,
adding o he i s ime he ollowing EPP me hods: Classical Mul idimensional Scaling, Sammon
Mapping and Fac o Analysis. Ex ended HUEPs a e alida ed in a case s udy associa ed wi h a
mul ina ional company in he au omo i e indus y sec o . Two eal-li e da ase s con aining da a
ga he ed om a Wa e je Cu ing ool a e isualized in an in ui i e and in o ma i e way. The ob ained
esul s show ha HUEPs is a echnique ha suppo s he con inuous moni o ing o machines in o de
o an icipa e ailu es. This con ibu ion o isual da a analy ics can help companies in decision-making,
ega ding FD and PdM p ojec s.
Keywo ds:
indus y 4.0; indus ial in e ne o hings; sma ac o ies; ad anced manu ac u ing;
indus ial big da a; p edic i e main enance; isualiza ion; machine lea ning; clus e ing; explo a o y
p ojec ion pu sui
1. In oduc ion
In he indus ial sec o , he e a e se e al issues ha mos co po a ions a e ying o add ess.
As a as echnology ad ances a e conce ned, some o hese p oblems could be sol ed, o a leas
hei nega i e impac could be educed. Recen ly, he concep o Indus y 4.0 [
1
] has been p oposed,
in ol ing se e al cu ing-edge echnologies such as obo ics, A i icial In elligence, Indus ial Big
Da a [
2
], Indus ial In e ne o Things (IIoT) [
3
], deep lea ning and deep analy ics, compu e ision,
isual da a analy ics, isual compu ing and digi al wins [
4
], among o he s. These esou ces a e g ea ly
con ibu ing o he sol ing o many o he p oblems in indus ial manu ac u ing, and a e imp o ing
manu ac u ing p ocesses.
Indus ial companies a e de eloping p ojec s in o de o adhe e o he “sma ac o y” [
5
,
6
]
concep . Wi h such p ojec s, ac o ies wan o be able o lea n and adap o changes in eal ime,
and in o de o
do ha , i is necessa y o ha e pe manen in o ma ion and da a ega ding he elemen s
Appl. Sci. 2020,10, 4355; doi:10.3390/app10124355 www.mdpi.com/jou nal/applsci
Appl. Sci. 2020,10, 4355 2 o 20
in ol ed in he plan . To be able o cap u e all his in o ma ion, senso s and IoT de ices need o be
ins alled. Thanks o IIoT, i is possible o ha e millions o da a i ems ela ed o ac o ies and hei
machines. Howe e , all hese da ase s a e useless and expensi e unless hey can be analyzed; he e is
whe e he p oposals o Big Da a and Visual Analy ics appea .
In gene al e ms, companies a e no able o accomplish he implemen a ion o he sma ac o y
pa adigm in all hei plan s. The cos s o becoming a sma ac o y a e po en ially una o dable;
he e o e, companies a e no senso izing all hei machines o p ocesses. The s o age esou ces (cloud)
equi ed o keep his huge quan i y o in o ma ion a e no ee. No mally, he da a o a machine in a
ac o y in ol es a g ea olume o in o ma ion wi h a high dimensionali y. Addi ionally, as he e a e
many di e en machines being pe manen ly moni o ed, he p oduced da a a e he e ogeneous and no
comple e (senso s and communica ions may ail).
When p ope ly used, all hese da ase s could make a ac o y mo e e icien in se e al
a eas: esou ce-sa ings, cos - educ ion, op imiza ion o p oduc ion imes, inc easing sus ainabili y,
minimizing ailu es and down ime, e c., bu he in es men in ga he ing, s o ing and analyzing hese
da ase s is usually e y high. Because o ha , i is e y impo an o be su e ha senso s, de ices and
ela ed da ase s a e ca e ully chosen, and ha e he p ope ea u es. These da a would ha e a high
dimensionali y, ha should be in balance wi h he cos s and equi emen s o he company. In o de o
o e come his well-known p oblem, named “ he cu se o dimensionali y”, Machine Lea ning (ML)
and in e ac i e isualiza ion o da a explo a ion could make g ea con ibu ions. As a esul , any
depa men o indus ial companies could bene i om he use o isual analy ics in decision-making.
In keeping wi h his idea, he p esen s udy p oposes he ex ension o Hyb id Unsupe ised
Explo a o y Plo s (HUEPs) [
7
], and he applying o hem o senso alida ion and condi ion moni o ing
as a decision-making ool based on isual analysis o subsequen p edic i e main enance (PdM).
HUEPs a e a ecen ly-p oposed echnique, whe ein Explo a o y P ojec ion Pu sui (EPP) [
8
] and
clus e ing [
9
] me hods could be combined o gene a e in o ma i e and in ui i e 3D isualiza ions o
high-dimensional da a (see Sec ion 2). When using di e en colo s and he glyph me apho , hey can
display mo e in o ma ion han a s anda d 2D o 3D p ojec ion, and in a mo e in ui i e way. HUEPs a e
alida ed and ex ended in he p esen pape by inco po a ing some addi ional EPP echniques, namely
Classical Mul idimensional Scaling (CMDS), Sammon Mapping (SM) and Fac o Analysis (FA).
This s udy helps o analyze he da ase s’ s uc u e and clea ly iden i y issues in machines,
an icipa ing po en ial ailu es. I he ini ial collec ed da ase , once analyzed, e eals a de ined s uc u e,
including clus e s “ epo ing” issues, his can be seen as a sign o a ep esen a i e da a se . The da a used
in his s udy ha e been p o ided by G upo An olin [
10
] (a mul ina ional company in he au omo i e
indus y sec o ), and e e o se e al wa e je indus ial ools (machines o pe o m indus ial cu ing
using ex emely-high p essu e wa e ) loca ed in an au omo i e plan . This p oposal ad ances p e ious
wo k, as he hyb idiza ion o unsupe ised lea ning o isualiza ion is applied o he i s ime in his
eal case s udy. Da ase s collec ed o PdM pu poses ha e ne e been analyzed be o e wi h HUEPs.
As was s a ed be o e, i is necessa y o balance cos s and bene i s. Da a isualiza ion con ibu es
o inding such balance, especially in cases o high olumes o in o ma ion wi h high dimensionali y.
Explo a o y analysis mus be one o he i s s eps in indus ial da a analy ics [
11
], and isualiza ions a e
ad isable wi h ega ds o knowing i he collec ed da a make sense [
12
]. As some s udies sugges [
13
],
isual compu ing plays a i al ole in bo h Indus y 4.0 and ad anced manu ac u ing. In [
14
], i is
shown how he applica ion o isual compu ing can empowe wo ke s in he amewo k o Indus y
4.0. A use case is p esen ed, showing how isual analy ics inco po a ed in a plan ’s Human–Machine
In e ace can imp o e he cogni i e p ocess o supe ising a p oduc ion line and applying co ec i e
main enance. In [
15
], he au ho s ocus on how he usion o g aphics, ision and media echnologies
can en ich he ole o he new ope a o 4.0. O he su eys p esen a me hodology o implemen ing a
da a-d i en PdM no only in he machine decision making, bu also in da a acquisi ion and p ocessing,
wi h a isual analysis o he Remaining Use ul Li e (RUL) o a machining ool [
16
]. As s a ed, companies
should no o ge he cos –imp o emen ade-o in ol ed in Indus y 4.0, explained in [
11
], whe e a
Appl. Sci. 2020,10, 4355 3 o 20
me hodology based on da a analy ics has been p oposed o he cos -e icien moni o ing o Indus y
4.0, wi h a eal use case wi hin he au omo i e indus y.
Some p e ious s udies, ha esea ched he applica ion o ML o PdM and aul s/anomalies
de ec ion, ha e p oposed a classi ica ion app oach, based on supe ised lea ning. Less e o has been
de o ed o in es iga ing he con ibu ions o unsupe ised lea ning, al hough some p e ious wo k
does exis . Fo ins ance, in [
17
], he au ho s p oposed he applica ion o he k-means me hod combined
wi h Fuzzy Logic o PdM pu poses. In [
18
], a combina ion o “cons ained k-means clus e ing, uzzy
modelling and LOF-Based sco e” has been applied in o de o de ec anomalies in an auxilia y ma ine
diesel engine. O he esea ch [
19
] applied i e clus e ing me hods (Hie a chical, k-medoids, k-means,
DBSCAN and OPTICS) in a heal h condi ion moni o ing model o a semiconduc o company’s
chemical apo deposi ion p ocess.
O he s udies on aul diagnosis o PdM analyze kinds o da ase s ha ha e been e y widely
used and es ed in he li e a u e: mo o bea ing da ase s [20–22], gea -box da ase s [20,23,24] and he
Tennessee Eas man P ocess (TEP) da ase [
25
]. In [
26
], ano he popula olling bea ing da ase has
been used o es ing a deep-dis ance me ic lea ning me hod.
Fo aul diagnosis in locomo i e olle bea ings, some au ho s p oposed [
27
] Fuzzy c-means
(FCM). Ano he pape p oposed a deep neu al ne wo k, named Ca AAE, o unsupe ised aul
diagnosis o olling bea ings [
28
]. The e a e also s udies ha p opose supe ised lea ning o bea ing
de ec classi ica ion [
29
]. In [
30
], he au ho s p oposed a nea es and a hes dis ance p ese ing
p ojec ion (NFDPP) algo i hm o explo e he ela ionships o a sample wi h i s a hes and nea es
neighbou s, which was alida ed in iden i ying compound aul s in locomo i e bea ings.
Some o he s udies ha e applied he well-known k-Nea es Neighbou (KNN) me hod; in [
31
],
a me hod was p oposed based on he E iden ial KNN, wi h applica ions in a powe plan . The e
a e also s udies ha apply KNN o a mo o gea s da ase [
32
]. Some o he au ho s ha e applied
PCA and A i icial Neu al Ne wo ks (ANN) o o a ion machines [
33
], bu no wi h a isualiza ion
pu pose. In o he s udies, ex ended PCA me hods ha e been also p oposed: WPD-PCA [
34
], FPCA [
35
],
FDKPCA [
36
] and DWRPCA [
37
], o aul de ec ion pu poses bu no om a isualiza ion pe spec i e.
Ex ensi e aul diagnosis s udies a e ound in he li e a u e conce ning bea ings da ase s; hese
kinds o da ase s a e equen ly used because o he impo ance o he ailu es in hese de ices. The e a e
also o he p ocesses, machines and de ices ha a e impo an in he indus y oo, and ha may be
equi ed o pe o m aul diagnos ics.
All in all, mos o he abo e-men ioned da ase s a e ou da ed, and a e no di ec ly ela ed o he
machines ound in an au omo i e ac o y. One o he main no el ies o he p esen esea ch is he
analysis o a no el and eal-li e case s udy, comp ising wo o he usual machines in he au omo i e
indus y. The ailu es a ec ing hese machines a e no ela ed wi h mo o s o hei bea ings, bu wi h a
di e en kind o p oblem. Fu he mo e, no el combina ions o HUEPs a e p oposed, inco po a ing
addi ional EPP echniques o he i s ime. This s udy p oposes he use o HUEPs as a isual ool,
in o de o de e mine i he collec ed da a om he ins alled senso s in di e en machines a e co ec
and well selec ed. HUEPs g ea ly con ibu e o he moni o ing o hese machines, and consequen ly o
he making o decisions in o de o an icipa e he associa ed ailu es.
The emaining sec ions o his s udy a e s uc u ed as ollows: Sec ion 2p esen s he echniques
and me hods ha a e applied o analyze he da a. Sec ion 3de ails he eal-li e case s udy and he
analyzed machines, while esul s a e p esen ed and discussed in Sec ion 4. Finally, Sec ion 5se s ou
he conclusions and p oposals o u u e wo k.
2. Hyb id Unsupe ised Explo a o y Plo s
In o de o ex ac knowledge om unlabelled da ase s, many di e en me hods could be applied.
Among all o he ML me hods, p ojec ion echniques a e conside ed a iable app oach o in o ma ion
seeking, as humans a e able o ecognize di e en ea u es and o de ec anomalies by inspec ing
g aphs. Such pa e ns may become pe cep ible i a ia ions a e made o he spa ial coo dina es o he
Appl. Sci. 2020,10, 4355 4 o 20
o iginal da ase s. Howe e , an a p io i choice ega ding which pa ame e s will e eal mos pa e ns
equi es p io knowledge o uniden i ied pa e ns, and his is no an easy ask. In o de o do ha ,
Explo a o y P ojec ion Pu sui (EPP) [
8
] is used o he pu pose o da a isualiza ion. Con as ed
wi h ea u e selec ion, EPP belongs o he ea u e ex ac ion pa adigm, as he esul ing dimensions
a e combina ions (could be linea o non-linea ) o he o iginal ea u es in he da ase . Con a ily,
clus e ing [
9
] is conce ned wi h g ouping oge he objec s, ha a e simila o each o he and dissimila
o objec s belonging o o he clus e s. The e o e, pa e ns wi hin he same clus e a e mo e simila o
each o he han hey a e o a pa e n belonging o a di e en clus e .
Recen wo k [
7
] has p oposed he independen applica ion o EPP echniques on he one hand,
and clus e ing hem on he o he . The comple e esul s o he wo o hem a e hen combined, oge he
wi h he glyph me apho , in a no el and di e en way, called he Hyb id Unsupe ised Explo a o y
Plo (HUEP). HUEPs a e a gene al-pu pose app oach, in which any EPP and clus e ing echnique could
be combined o gene a e 3D isualiza ions om high-dimensional da a. In Figu e 1, he p ocess o
gene a e a HUEP is shown. These depic ions a e in o ma i e and in ui i e, no only o da a scien is s,
bu o all he company s a (no equi ing p e ious knowledge ega ding ML). Once a p ope HUEP
con igu a ion has been selec ed by an expe , he o he company s a will only need o analyze he
ob ained isualiza ion. The HUEPs can be used by ‘non-expe ’ p o essionals due o hei simplici y,
as no u he decisions o pa ame e - unings a e equi ed. The p oposed HUEPs a e hyb id as hey
combine bo h explo a o y (dimensionali y educ ion) echniques wi h clus e ing ones. Besides, bo h
ypes o echnique a e unsupe ised as hey apply his kind o lea ning (no a ge class o alue is
p o ided o be ep oduced o new da a ins ances).
Appl.Sci.2020,10,xFORPEERREVIEW4o 21
Ino de  odo ha ,Explo a o yP ojec ionPu sui (EPP)[8]isused o  hepu poseo da a
isualiza ion.Con as edwi h ea u eselec ion,EPPbelongs o he ea u eex ac ionpa adigm,as
he esul ingdimensionsa ecombina ions(couldbelinea o non‐linea )o  heo iginal ea u esin
heda ase .Con a ily,clus e ing[9]isconce nedwi hg ouping oge he objec s, ha a esimila  o
eacho he anddissimila  oobjec sbelonging oo he clus e s.The e o e,pa e nswi hin hesame
clus e a emo esimila  oeacho he  han heya e oapa e nbelonging oadi e en clus e .
Recen wo k[7]hasp oposed heindependen applica iono EPP echniqueson heonehand,
andclus e ing hemon heo he .Thecomple e esul so  he woo  hema e hencombined, oge he 
wi h heglyphme apho ,inano elanddi e en way,called heHyb idUnsupe isedExplo a o y
Plo (HUEP).HUEPsa eagene al‐pu poseapp oach,inwhichanyEPPandclus e ing echniquecould
becombined ogene a e3D isualiza ions omhigh‐dimensionalda a.InFigu e1, hep ocess o
gene a eaHUEPisshown.Thesedepic ionsa ein o ma i eandin ui i e,no only o da ascien is s,
bu  o all hecompanys a (no  equi ingp e iousknowledge ega dingML).Onceap ope HUEP
con igu a ionhasbeenselec edbyanexpe , heo he companys a willonlyneed oanalyze he
ob ained isualiza ion.TheHUEPscanbeusedby‘non‐expe ’p o essionalsdue o hei simplici y,as
no u he decisionso pa ame e ‐ uningsa e equi ed.Thep oposedHUEPsa ehyb idas hey
combinebo hexplo a o y(dimensionali y educ ion) echniqueswi hclus e ingones.Besides,bo h
ypeso  echniquea eunsupe isedas heyapply hiskindo lea ning(no a ge classo  alueis
p o ided obe ep oduced o newda ains ances).

Figu e1.P ocess oob ainaHUEP.
In heo iginal o mula iono HUEPs[7],k‐means[38]andHie a chicalMe hodswe eappliedas
clus e ing echniques.Complemen a ily,PCA[39],MLHL[40]andCMLHL[41]we eappliedasEPP
echniques.Awide a ie yo EPP echniquesexis s,bu be o e hiss udy,only he h eep e iously‐
men ionedoneshadbeenappliedunde  he amewo ko HUEPs.EachEPP echniquep ojec s he
da ainadi e en way, hus o  hesameda ase  heob ained isualiza ionscouldbe e ydi e en .
Ashappens o mos o  heda aanalysisp oblems, he eisno a echnique ha alwaysge s hebes 
esul s o di e en da ase s.Thus,dependingon heanalyzedda a,oneHUEP ha usesace ainEPP
me hodmaybemo esui able hanano he one.Inkeepingwi h hisidea,wep oposedin hiss udy o
ex endHUEPswi ho he EPP echniques o ace ainda ase .Themainpu poseis o alida e ha 
HUEPscanbeex endedwi ho he EPPme hods,and op o e ha ce ainEPPme hodscouldbemo e
sui able hano he oneswhenanalyzing hep esen da ase .
Theo iginalHUEP o mula ionhasbeenex endedin hep esen s udybyinco po a ingand
alida ingsomeaddi ionalEPP echniques,namelyClassicalMul idimensionalScaling(CMDS),
SammonMapping(SM)andFac o Analysis(FA)—inyellowinFigu e1.Theya edesc ibedin he
ollowingsubsec ions.

Figu e 1. P ocess o ob ain a HUEP.
In he o iginal o mula ion o HUEPs [
7
], k-means [
38
] and Hie a chical Me hods we e applied
as clus e ing echniques. Complemen a ily, PCA [
39
], MLHL [
40
] and CMLHL [
41
] we e applied
as EPP echniques. A wide a ie y o EPP echniques exis s, bu be o e his s udy, only he h ee
p e iously-men ioned ones had been applied unde he amewo k o HUEPs. Each EPP echnique
p ojec s he da a in a di e en way, hus o he same da ase he ob ained isualiza ions could be e y
di e en . As happens o mos o he da a analysis p oblems, he e is no a echnique ha always ge s
he bes esul s o di e en da ase s. Thus, depending on he analyzed da a, one HUEP ha uses a
ce ain EPP me hod may be mo e sui able han ano he one. In keeping wi h his idea, we p oposed in
his s udy o ex end HUEPs wi h o he EPP echniques o a ce ain da ase . The main pu pose is o
alida e ha HUEPs can be ex ended wi h o he EPP me hods, and o p o e ha ce ain EPP me hods
could be mo e sui able han o he ones when analyzing he p esen da ase .
The o iginal HUEP o mula ion has been ex ended in he p esen s udy by inco po a ing and
alida ing some addi ional EPP echniques, namely Classical Mul idimensional Scaling (CMDS),
Sammon Mapping (SM) and Fac o Analysis (FA)—in yellow in Figu e 1. They a e desc ibed in he
ollowing subsec ions.
Appl. Sci. 2020,10, 4355 5 o 20
2.1. Classical Mul idimensional Scaling
Mul idimensional scaling (MDS) [
42
] is a se o me hods ha p ojec high-dimensional da a in o a
low-dimensional space using dis ances o dissimila i ies. Classical MDS (CMDS) is a membe MDS
ha shows he s uc u e o dis ance-like da a as a geome ical image [
43
]. Gene ally, MDS’s inpu
is no a da ase , bu he simila i ies o a se o i ems ins ead. CMDS uses a single dis ance ma ix o
Euclidian ype as inpu .
CMDS is also known as P incipal Coo dina es Analysis (PCoA) [
44
], To ge son Scaling o
To ge son–Gowe scaling. I is o en used o isualize da a when only hei dis ances o dissimila i ies
a e a ailable, bu in his esea ch, in o de o ex end and alida e HUEPs wi h his amily o me hods,
he o iginal da ase has been educed o a dis ance ma ix. This ma ix (pai wise dis ance be ween
pai s o obse a ions) c ea es a new con igu a ion o poin s using he ollowing me ics:
•Euclidean;
•Squa ed Euclidean;
•
S anda dized Euclidean (seuclidean): each coo dina e di e ence be ween obse a ions is scaled
by di iding by he co esponding elemen o he s anda d de ia ion;
•Ci yblock;
•Minkowski;
•Chebyshe : maximum coo dina e di e ence;
•Cosine: one minus he cosine o he included angle be ween poin s;
•Co ela ion: one minus he sample co ela ion be ween poin s;
•Hamming, which is he pe cen age o coo dina es ha di e ;
•
Jacca d: one minus he Jacca d coe icien , which is he pe cen age o non-ze o coo dina es
ha di e ;
•Spea man: one minus he sample Spea man’s ank co ela ion be ween obse a ions.
2.2. Sammon Mapping
Sammon Mapping (SM) [
45
,
46
] o Sammon P ojec ion is a p ojec ion me hod o analyzing
mul i a ia e da a. I can be seen as a ype o MDS me hod, using a non-linea me ic ha is equen ly
used o EPP. SM maps a high-dimensional da ase o a lowe dimensionali y one, conse ing he
in insic s uc u e o he da a when he pa e ns a e p ojec ed.
Unlike s anda d PCA and o he EPP echniques, SM is a non-linea app oach, as he esul canno
be ep esen ed as a linea combina ion o he o iginal a iables. Ne e heless, SM minimizes he
di e ences be ween co esponding pai wise poin dis ances in he wo spaces. PCA applies op imal
mapping o he da ase , while SM ies o ob ain a lowe -dimensional da ase ha keeps he o iginal
s uc u e as much as possible. Because o ha , he SM algo i hm has a high compu a ional load O(n
2
).
The sou ce da ase is ep esen ed as N ec o s in L-dimensional space, gi en by X
i
, i =1, ..., N.
I is sough o map hese in o d-dimensional space (wi h d<L), o ob ain ec o s Y
i
, i =1, ... N; d
ij
is he
pai wise dis ance be ween Yiand Yj, and dij*is he dis ance be ween Xiand Xj.
SM aims o minimize he ollowing e o unc ion, which is o en called Sammon’s s ess o
Sammon’s e o :
E=
1
Pi<jd∗
ij
n
X
i<j
d∗
ij −dij2
d∗
ij
(1)
As o iginally sugges ed [
45
], he minimiza ion could be achie ed by g adien sea ching echniques
o by o he me hods, bu hese equen ly in ol e i e a i e p ocedu es. Con e gen esul s a e no
always eached, and he numbe o i e a ions is decided expe imen ally [47].

Appl. Sci. 2020,10, 4355 6 o 20
2.3. Fac o Analysis (FA)
Fac o Analysis [
48
,
49
] is a da a analysis echnique ha could be used o dimensionali y educ ion.
I is used o educe a la ge numbe o a iables in o a smalle se o ac o s, whe e ac o s e e s o he
lowe numbe o unobse ed o unde lying a iables. FA’s objec i e is o disco e independen la en
a iables. I is equen ly used wi h da ase s wi h a la ge numbe o obse ed a iables ha could
e eal a smalle numbe o unde lying a iables.
FA is di e en ia ed om PCA as he o me en o ces a s ic s uc u e o a ixed numbe o common
(la en ) ac o s, while he la e de ines p ac o s in dec easing o de o impo ance. The main ac o
in FA is he one ha , a e o a ion, p o ides he maximal in e p e a ion, while he mos signi ican
ac o in PCA is he one ha maximizes he a iance. F equen ly, he main ac o in FA is di e en
om he di ec ion o he i s p incipal componen . PCA ex ac s ac o s based on he o al a iance o
he ac o s, bu FA ex ac s ac o s based on he a iance sha ed by he ac o s. PCA is used o ind he
smalles numbe o a iables ha explain he mos a iance, while FA is used o look o he la en
unde lying ac o s.
3. A Real Case S udy: Wa e je Cu ing
As p e iously s a ed, in o de o alida e he p oposed HUEP ex ension, i has been applied o
a case s udy in ol ing a wa e je indus ial ool. As a esul , wo da ase s ha e been gene a ed by
collec ing da a om wo di e en machines. HUEPs a e gene a ed in o de o analyze hese da ase s,
which comp ise a g ea numbe o samples wi h a high numbe o ea u es.
The machines unde s udy a e loca ed in one ac o y om G upo An olin [
10
], a mul ina ional
company om he au omo i e indus ial sec o .
Wa e je cu ing is used in a ious manu ac u ing indus ies (such as ae ospace and au omo i e
ones) o cu ing, shaping, e c. [
50
]. This ool is used o cu se e al kinds o ma e ials by means o an
ex emely high p essu e je o wa e , o a mix u e o wa e and an ab asi e subs ance. In he p esen
wo k, he wa e je uses only wa e , and is applied du ing he ab ica ion o some pa s manu ac u ed in
he ac o y. Two o i s independen componen s ha e been selec ed o analysis due o hei c i ical ole:
•
In ensi ie : he wa e je pumps o in ensi ie s [
51
], which supply wa e a ex emely high p essu e
o wa e je machines;
•
Cyclone: he acuum cyclone uni loca ed in a wa e je machine, used o suc ioning he was e
gene a ed owa ds a chu e. I also holds he pieces du ing he cu .
To do his analysis, a ime in e al (Feb ua y 2020) has been selec ed among all he a ailable
da a as i includes samples o di e en anomalies/ ailu es. As a consequence o he main enance
ope a ions ha a e ca ied ou , he e a e no anomalies mos o he ime, and i is no easy o ind a
pe iod con aining examples o some di e en anomalies.
3.1. In ensi ie
The unc ion o a wa e pump o in ensi ie (Figu e 2) is o aise he wa e p essu e o he
le el needed in he wa e je machines o an op imal ope a ion. Explained in a e y simple manne ,
he in ensi ie wo ks like his: i begins when he low p essu e wa e en e s he in ensi ie (3–5 ba );
a e se e al il e s, he wa e eaches a pump ha aises he p essu e o 10 ba , and hen he wa e
lows o he cylinde s ha will be comp essed by a hyd aulic g oup, which, using a plunge /pis on
sys em, will be able o each an ex emely high p essu e (a ound 3000 ba ) The in ensi ie s unde s udy
ha e wo cylinde s, as can be seen in Figu e 2.
Appl. Sci. 2020,10, 4355 7 o 20
Appl.Sci.2020,10,xFORPEERREVIEW7o 21

(a)(b)
Figu e2.Exampleo in ensi ie andske cho senso s(blueand edci cles).(a)Pic u eo  he
in ensi ie .(b)Senso splacemen ske ch.
Acco ding o heexpe ienceo  hewo ke sin heplan ,i isknown ha be o ea ailu ehappens,
aninc easein empe a u eindi e en zoneso  hecylinde isobse ed.Ano he p oblemdi ec ly
ela edwi h heinc easeo  empe a u eis hewa e leaking,whichiscausedby hede e io a iono 
hecylinde ,andwhichcanbe isuallyobse ed.As hiscouldlead oac i ical ailu e,speci ic
senso sha ebeenins alledino de  omeasu e he empe a u ein hesealhead(SH)and he
hyd aulicpis on(HP)o  hein ensi ie .Addi ionally,leaksenso sha ebeenins alledineach
cylinde .Wi h heins alla iono  hesesenso s,asshowninFigu e2,adiagnosiso  hes a uso  he
machinecouldbepe o medino de  op e en  ailu es.
Theda ase analyzedin his esea chcomp ises7414samples( omFeb ua y2020)and36
pa ame e s( ea u es),desc ibedinTable1,whichwe ega he ede e y5min(excep  o  heinc ease
o wa e leak).
Table1.In ensi ie  ea u es.Va iablesga he ed omeachcylinde ,SHandHP.XXin he ea u e
name e e s o henumbe o eachsenso .
Fea u eNameDesc ip ionUni 
HPXXTemp_oC_a gHPa e age empe a u e°C
HPXXTemp_oC_maxHPmaximum empe a u e°C
HPXXTemp_oC_minHPminimum empe a u e°C
HPXXTemp_oC_s dHPs anda dde ia ion empe a u e°C
SHXXTemp_oC_a gSHa e age empe a u e°C
SHXXTemp_oC_maxSHmaximum empe a u e°C
SHXXTemp_oC_minSHminimum empe a u e°C
SHXXTemp_oC_s dSHs anda dde ia ion empe a u e°C
SHXXLeak_mLmSHinc easeleako wa e sincelas pe iod1.5mL/inc ease
As hein ensi ie has wocylinde sandeachcylinde has woSHand woHP, he ea e36 o al
ea u es:4 empe a u e ea u espe SH,4 empe a u e ea u espe HP,and1 ea u epe SH o  he
leakp oblem.The eisone empe a u esenso pe SHandHP,andoneleaksenso pe SH.As he
ea u esa e eco ded o agi en imepe iod, hemaximum,minimum,a e ageands anda d
de ia ions a is icsa ecalcula eddu ing hispe iod,ands o ed.In hecaseo  heleaksenso , he
inc emen whencompa ed o hep e iouspe iodiss o ed.
The ea e h eemainp oblemso  ailu es ha a ec  hein ensi ie :
Figu e 2.
Example o in ensi ie and ske ch o senso s (blue and ed ci cles). (
a
) Pic u e o he in ensi ie .
(b) Senso s placemen ske ch.
Acco ding o he expe ience o he wo ke s in he plan , i is known ha be o e a ailu e happens,
an inc ease in empe a u e in di e en zones o he cylinde is obse ed. Ano he p oblem di ec ly
ela ed wi h he inc ease o empe a u e is he wa e leaking, which is caused by he de e io a ion
o he cylinde , and which can be isually obse ed. As his could lead o a c i ical ailu e, speci ic
senso s ha e been ins alled in o de o measu e he empe a u e in he seal head (SH) and he hyd aulic
pis on (HP) o he in ensi ie . Addi ionally, leak senso s ha e been ins alled in each cylinde . Wi h he
ins alla ion o hese senso s, as shown in Figu e 2, a diagnosis o he s a us o he machine could be
pe o med in o de o p e en ailu es.
The da ase analyzed in his esea ch comp ises 7414 samples ( om Feb ua y 2020) and
36 pa ame e s ( ea u es), desc ibed in Table 1, which we e ga he ed e e y 5 min (excep o he
inc ease o wa e leak).
Table 1.
In ensi ie ea u es. Va iables ga he ed om each cylinde , SH and HP. XX in he ea u e name
e e s o he numbe o each senso .
Fea u e Name Desc ip ion Uni
HPXXTemp_oC_a g HP a e age empe a u e ◦C
HPXXTemp_oC_max HP maximum empe a u e ◦C
HPXXTemp_oC_min HP minimum empe a u e ◦C
HPXXTemp_oC_s d HP s anda d de ia ion empe a u e ◦C
SHXXTemp_oC_a g SH a e age empe a u e ◦C
SHXXTemp_oC_max SH maximum empe a u e ◦C
SHXXTemp_oC_min SH minimum empe a u e ◦C
SHXXTemp_oC_s d SH s anda d de ia ion empe a u e ◦C
SHXXLeak_mLm SH inc ease leak o wa e since las pe iod 1.5 mL/inc ease
As he in ensi ie has wo cylinde s and each cylinde has wo SH and wo HP, he e a e 36 o al
ea u es: 4 empe a u e ea u es pe SH, 4 empe a u e ea u es pe HP, and 1 ea u e pe SH o he
leak p oblem. The e is one empe a u e senso pe SH and HP, and one leak senso pe SH. As he
ea u es a e eco ded o a gi en ime pe iod, he maximum, minimum, a e age and s anda d de ia ion
s a is ics a e calcula ed du ing his pe iod, and s o ed. In he case o he leak senso , he inc emen
when compa ed o he p e ious pe iod is s o ed.
The e a e h ee main p oblems o ailu es ha a ec he in ensi ie :
Appl. Sci. 2020,10, 4355 8 o 20
•
Wa e leaks: he in ensi ie will s op wo king i he e is a se e e wa e leak. This is a c i ical ailu e
wi h high associa ed cos s, as i s ops p oduc ion;
•
High empe a u e in a cylinde : i high empe a u e las s a long ime, i could lead o a b eak in
he heade ;
•
De ec ed SH mal unc ion; his means ha i is necessa y o epai he SH, o o he wise i will c ash
and s op he p oduc ion. This mal unc ion/p oblem is pe cei ed by he main enance s a .
3.2. Cyclone
The ollowing is he no mal p ocess when manu ac u ing a piece in he wa e je machine (Figu e 3):
he wo ke places a pa inside he machine and gi es he s a o de , and he cu ope a ion is pe o med
in a con inuous way du ing he cycle. The wa e p essu e gene a ed by he in ensi ie eaches he le el
equi ed by he obo s in he wa e je o pe o ming he comple e cu o he piece. By means o a
acuum sys em, bo h he excess wa e and he cu ing le o e s om he wo k a e d ained h ough a
acuum cyclone uni .
Cyclone ailu es a e usually ela ed o he acuum sys em ha he cyclone mo o i sel uns.
This suc ion is esponsible o secu ing he piece o be cu , in addi ion o abso bing he emaining
wa e and cu ings le o e om he piece. The gene a ed was e om he wa e je is suc ioned owa ds
a ga bage chu e. The mos ep esen a i e ailu es in his a ea a e usually ela ed o blockages in he
suc ion ci cui and ai ou le s, and acuum mal unc ioning. The da a s o ed o de ec ing ailu es a e
ela ed o he cyclone engine, and a e ob ained om he PLC. Addi ionally, ib a ion and empe a u e
senso s ha e been ins alled in his mo o in o de o ob ain mo e in o ma ion abou i s s a us.
The da ase o his componen comp ises 623 samples (Feb ua y 2020) and 24 pa ame e s
( ea u es), desc ibed in Table 2, ga he ed o each manu ac u ing cycle.
Appl.Sci.2020,10,xFORPEERREVIEW8o 21
 Wa e leaks: hein ensi ie wills opwo kingi  he eisase e ewa e leak.Thisisac i ical ailu e
wi hhighassocia edcos s,asi s opsp oduc ion;
 High empe a u einacylinde :i high empe a u elas salong ime,i couldlead oab eakin
heheade ;
 De ec edSHmal unc ion; hismeans ha i isnecessa y o epai  heSH,o o he wisei willc ash
ands op hep oduc ion.Thismal unc ion/p oblemispe cei edby hemain enances a .
3.2.Cyclone
The ollowingis heno malp ocesswhenmanu ac u ingapiecein hewa e je machine(Figu e
3): hewo ke placesapa inside hemachineandgi es hes a o de ,and hecu ope a ionis
pe o medinacon inuouswaydu ing hecycle.Thewa e p essu egene a edby hein ensi ie 
eaches hele el equi edby he obo sin hewa e je  o pe o ming hecomple ecu o  hepiece.
Bymeanso a acuumsys em,bo h heexcesswa e and hecu ingle o e s om hewo ka e
d ained h ougha acuumcycloneuni .

(a)
Figu e 3. Con .
Appl. Sci. 2020,10, 4355 9 o 20
Appl.Sci.2020,10,xFORPEERREVIEW9o 21

(b)
Figu e3.Exampleo awa e je machinewi h hecycloneandske cho senso s( edandg eenci cles).
(a)Wa e je machine.(b)Senso splacemen ske ch.
Cyclone ailu esa eusually ela ed o he acuumsys em ha  hecyclonemo o i sel  uns.This
suc ionis esponsible o secu ing hepiece obecu ,inaddi ion oabso bing he emainingwa e 
andcu ingsle o e  om hepiece.Thegene a edwas e om hewa e je issuc ioned owa dsa
ga bagechu e.Themos  ep esen a i e ailu esin hisa eaa eusually ela ed oblockagesin he
suc ionci cui andai ou le s,and acuummal unc ioning.Theda as o ed o de ec ing ailu esa e
ela ed o hecycloneengine,anda eob ained om hePLC.Addi ionally, ib a ionand
empe a u esenso sha ebeenins alledin hismo o ino de  oob ainmo ein o ma ionabou i s
s a us.
Theda ase  o  hiscomponen comp ises623samples(Feb ua y2020)and24pa ame e s
( ea u es),desc ibedinTable2,ga he ed o eachmanu ac u ingcycle.
Table2.Cyclone ea u es.Va iablesga he edpe manu ac u ingcycle.
Fea u eNameDesc ip ionUni 
AccPeak_g_a gEngine ib a iona e ageG
AccPeak_g_maxEngine ib a ionmaximumG
AccPeak_g_minEngine ib a ionminimumG
AccPeak_g_s dEngine ib a ions anda dde ia ionG
CmdDu yEngineSpeed_HzFanRPMse poin Hz
CmdRes EngineSpeed_pe cen %RPMidlese poin %
CmdVacuumP essu e_mBa Vacuump essu ese poin mBa 
EngineTemp_oC_a gEngine empe a u ea e age°C
EngineTemp_oC_maxEngine empe a u emaximum°C
EngineTemp_oC_minEngine empe a u eminimum°C
EngineTemp_oC_s dEngine empe a u es anda dde ia ion°C
FanSpeed_Hz_a gFanspeeda e ageHz
Figu e 3.
Example o a wa e je machine wi h he cyclone and ske ch o senso s ( ed and g een ci cles).
(a) Wa e je machine. (b) Senso s placemen ske ch.
Table 2. Cyclone ea u es. Va iables ga he ed pe manu ac u ing cycle.
Fea u e Name Desc ip ion Uni
AccPeak_g_a g Engine ib a ion a e age G
AccPeak_g_max Engine ib a ion maximum G
AccPeak_g_min Engine ib a ion minimum G
AccPeak_g_s d Engine ib a ion s anda d de ia ion G
CmdDu yEngineSpeed_Hz Fan RPM se poin Hz
CmdRes EngineSpeed_pe cen % RPM idle se poin %
CmdVacuumP essu e_mBa Vacuum p essu e se poin mBa
EngineTemp_oC_a g Engine empe a u e a e age ◦C
EngineTemp_oC_max Engine empe a u e maximum ◦C
EngineTemp_oC_min Engine empe a u e minimum ◦C
EngineTemp_oC_s d Engine empe a u e s anda d de ia ion ◦C
FanSpeed_Hz_a g Fan speed a e age Hz
FanSpeed_Hz_max Fan speed maximum Hz
FanSpeed_Hz_min Fan speed minimum Hz
FanSpeed_Hz_s d Fan speed s anda d de ia ion Hz
VacuumP essu e1_mBa _a g Vacuum p essu e senso 1 a e age mBa
VacuumP essu e1_mBa _max Vacuum p essu e senso 1 maximum mBa
VacuumP essu e1_mBa _min Vacuum p essu e senso 1 minimum mBa
VacuumP essu e1_mBa _s d Vacuum p essu e senso 1 s anda d de ia ion mBa
VacuumP essu e2_mBa _a g Vacuum p essu e senso 2 a e age mBa
VacuumP essu e2_mBa _max Vacuum p essu e senso 2 maximum mBa
VacuumP essu e2_mBa _min Vacuum p essu e senso 2 minimum mBa
VacuumP essu e2_mBa _s d Vacuum p essu e senso 2 s anda d de ia ion mBa
Du a ion Cycle ime o pa p oduc ion ms
Appl. Sci. 2020,10, 4355 16 o 20
As has been shown o he p e ious da ase , i is wo h men ioning ha CMDS gene a es he bes
HUEPs o he cyclone da ase . The bes esul is ob ained wi h CMDS wi h Seuclidean +hie a chical
clus e ing (k=6, Ci yblock, and weigh ed), shown in Figu es 8d and 9c. In hese isualiza ions,
he c i ical ailu e (suc ion ci cui is blocked) samples ( ed x) a e isola ed in a unique g oup, as a e
hose samples associa ed o he o he anomaly (black +), which a e also depic ed in a sepa a ed g oup.
The HUEPs gene a ed o his da ase can clea ly isualize he c i ical ailu e in a sepa a e g oup.
This is done hanks o he combina ion wi h clus e ing esul s; he samples associa ed o he main
ailu e a e g ouped in he same clus e , clea ly isualizing hem in a sepa a e g oup. Al hough he
samples om he o he anomaly ( acuum mal unc ioning) a e clus e ed wi h many “no mal” samples,
he CMDS p ojec ion isualizes hem in a sepa a e way.
Appl.Sci.2020,10,xFORPEERREVIEW16o 21

(e)( )
Figu e8.HUEP isualiza ionso  hecycloneda ase .EPP+hie a chical.(a)HUEP:PCA+hie a chical
clus e ing(k=6,Ci yblock,andweigh ed).(b)HUEP:MLHL+hie a chicalclus e ing(k=6,Ci yblock,
andweigh ed).(c)HUEP:CMLHL+hie a chicalclus e ing(k=6,Ci yblock,andweigh ed).(d)HUEP:
CMDS–Seuclidean+hie a chicalclus e ing(k=6,Ci yblock,andweigh ed).(e)HUEP:
SM+hie a chicalclus e ing(k=6,Ci yblock,andweigh ed).( )HUEP:FA+hie a chicalclus e ing(k=
6,Ci yblock,andweigh ed).
Ashasbeenshown o  hep e iousda ase ,i iswo hmen ioning ha CMDSgene a es he
bes HUEPs o  hecycloneda ase .Thebes  esul isob ainedwi hCMDSwi hSeuclidean+
hie a chicalclus e ing(k=6,Ci yblock,andweigh ed),showninFigu e8dandFigu e9c.In hese
isualiza ions, hec i ical ailu e(suc ionci cui isblocked)samples( edx)a eisola edinaunique
g oup,asa e hosesamplesassocia ed o heo he anomaly(black+),whicha ealsodepic edina
sepa a edg oup.TheHUEPsgene a ed o  hisda ase canclea ly isualize hec i ical ailu eina
sepa a eg oup.Thisisdone hanks o hecombina ionwi hclus e ing esul s; hesamplesassocia ed
o hemain ailu ea eg oupedin hesameclus e ,clea ly isualizing heminasepa a eg oup.
Al hough hesamples om heo he anomaly( acuummal unc ioning)a eclus e edwi hmany
“no mal”samples, heCMDSp ojec ion isualizes heminasepa a eway.
Onceagain,CMDSis heEPPme hodgene a ing hebes HUEP,andaddi ional esul sob ained
wi h hisme hoda eshowninFigu e9.Themainpu poseo  his igu eis ocompa e he
isualiza ionsgene a edwhenclus e ingpa ame e sa e a ied.

(a)(b)
Appl.Sci.2020,10,xFORPEERREVIEW17o 21

(c)
Figu e9.HUEP isualiza ionsgene a edbyCMDS–Seuclideananddi e en clus e ingpa ame e s
o  hecycloneda ase .(a)HUEP:CMDS–Seuclidean+k‐means(k=3andCo ela ion).(b)HUEP:
CMDS–Seuclidean+k‐means(k=6andCo ela ion).(c)HUEP:CMDS–Seuclidean+hie a chical
clus e ing(k=6,Ci yblock,andweigh ed).
Figu e9shows he esul sob ainedbycombiningCMDS–Seuclideanwi hk‐meansand
hie a chicalclus e ing.Thehie a chicalclus e ingme hodcansepa a ein oclus e #1 hec i ical‐
ailu esamples( edx),bu k‐meansg oups heseda a oge he wi h“no mal”samples.I can hus
besaid ha ,asin hecaseo  hep e iousda ase ,hie a chicalclus e ingleads obe e  isualiza ion
esul s o  hecycloneda ase .
Ingene al e ms, heHUEPsgene a edbyagglome a i e(hie a chical)clus e inga emo e
in o ma i e han hosegene a edbyk‐means o  he woda ase sunde analysis.Thisisconsis en 
wi hgene alheu is ics[52],asagglome a i eclus e ingismo eapp op ia ei g oupsa eexpec ed
obedi e en sizes.Opposingly,k‐meansis hebes op ionwhen heexpec edg oupsha e
app oxima elysimila sizes.In he woanalyzedda ase s, he ea emanymo eda asamples
associa edwi h he“no mal” unc ioningo  hecomponen s han hoseassocia edwi h ailu es.
Consequen ly, heHUEPs’wouldha ebe e  isualiza ionsbyusingagglome a i eclus e ing a he 
hank‐means.The alida ionsca iedou wi hbo hda ase sha eshown ha HUEPsgene a edby
CMDSa e hemos use ulones,as heyp o ide isualiza ions ha sepa a e ailu es om“no mal”
samplesin heclea es way.
5.ConclusionsandFu u eWo k
Thiss udyhasshown ha HUEPsa ea echnique ha suppo s hemoni o ingo senso sand
machinesino de  oan icipa e ailu es.Fi s ly,i canbeused oeasilysee ha  heda a ha hasbeen
ga he edha eas uc u e,and ha  heycouldbe ep esen a i eandin o ma i e.Then,HUEPscan
showsepa a eg oups,di e en ia ing ailu es omno mal unc ioning.Thep oposedHUEP
ex ensionsuppo sdecision‐makingasi depic sda aina isualway ha assis swi h his ask.This
hasbeen es edand alida edinacomplexindus ialscena io,wi hassocia edda ase scomp ising
Figu e 9.
HUEP isualiza ions gene a ed by CMDS–Seuclidean and di e en clus e ing pa ame e s
o he cyclone da ase . (
a
) HUEP: CMDS–Seuclidean+k-means (k=3 and Co ela ion). (
b
) HUEP:
CMDS–Seuclidean+k-means (k=6 and Co ela ion). (
c
) HUEP: CMDS–Seuclidean+hie a chical
clus e ing (k=6, Ci yblock, and weigh ed).
Once again, CMDS is he EPP me hod gene a ing he bes HUEP, and addi ional esul s ob ained
wi h his me hod a e shown in Figu e 9. The main pu pose o his igu e is o compa e he isualiza ions
gene a ed when clus e ing pa ame e s a e a ied.
Figu e 9shows he esul s ob ained by combining CMDS–Seuclidean wi h k-means and hie a chical
clus e ing. The hie a chical clus e ing me hod can sepa a e in o clus e #1 he c i ical- ailu e samples
( ed x), bu k-means g oups hese da a oge he wi h “no mal” samples. I can hus be said ha , as in

Appl. Sci. 2020,10, 4355 17 o 20
he case o he p e ious da ase , hie a chical clus e ing leads o be e isualiza ion esul s o he
cyclone da ase .
In gene al e ms, he HUEPs gene a ed by agglome a i e (hie a chical) clus e ing a e mo e
in o ma i e han hose gene a ed by k-means o he wo da ase s unde analysis. This is consis en
wi h gene al heu is ics [
52
], as agglome a i e clus e ing is mo e app op ia e i g oups a e expec ed o be
di e en sizes. Opposingly, k-means is he bes op ion when he expec ed g oups ha e app oxima ely
simila sizes. In he wo analyzed da ase s, he e a e many mo e da a samples associa ed wi h he
“no mal” unc ioning o he componen s han hose associa ed wi h ailu es. Consequen ly, he
HUEPs’ would ha e be e isualiza ions by using agglome a i e clus e ing a he han k-means.
The alida ions ca ied ou wi h bo h da ase s ha e shown ha HUEPs gene a ed by CMDS a e he
mos use ul ones, as hey p o ide isualiza ions ha sepa a e ailu es om “no mal” samples in he
clea es way.
5. Conclusions and Fu u e Wo k
This s udy has shown ha HUEPs a e a echnique ha suppo s he moni o ing o senso s and
machines in o de o an icipa e ailu es. Fi s ly, i can be used o easily see ha he da a ha has been
ga he ed ha e a s uc u e, and ha hey could be ep esen a i e and in o ma i e. Then, HUEPs can
show sepa a e g oups, di e en ia ing ailu es om no mal unc ioning. The p oposed HUEP ex ension
suppo s decision-making as i depic s da a in a isual way ha assis s wi h his ask. This has been
es ed and alida ed in a complex indus ial scena io, wi h associa ed da ase s comp ising a g ea
numbe o samples and a high numbe o ea u es. As a esul , PdM can be ca ied ou in a manne
complemen a y o o he ools, an icipa ing de ia ions in ope a ing condi ions.
I has been p o en ha he p oposed ex ension o HUEPs ou pe o ms he o iginal o mula ion in
isualizing he da ase s om he p esen case s udy. I is wo h men ioning ha HUEPs gene a ed by
CMDS a e mo e use ul han he o he ones, as he di e en anomalies a e g ouped and isualized in a
clea e way. Addi ionally, CMDS is an in e es ing EPP echnique because i can be applied wi h se e al
dis ance me ics ha adjus o he da ase unde s udy. On he o he hand, o he analyzed da ase s
(whe e he “no mal” samples a e mo e nume ous han he ailu e ones), agglome a i e clus e ing
g oups da a in a mo e consis en way han k-means.
I has been p o en ha , depending on he da ase , i could be be e o use a combina ion o
me hods, o gene a e one HUEP o ano he . Tha is, he bes isualiza ions o he in ensi ie a e no
gene a ed wi h he same pa ame e combina ion as ha used o he cyclone da ase . Hence, i is
impo an o ca y ou an exhaus i e expe imen a ion wi h e e y di e en combina ion, in o de o
iden i y he bes isualiza ion. Once his is pe o med, any pe son amilia wi h he manu ac u ing
p ocess (such as skilled ope a o s and main enance s a , among o he s) would be quali ied o analyze
he isualiza ions ob ained, and ake decisions based on hem.
As a u u e line o wo k, au ho s p opose he combina ion o HUEPs wi h he ou pu s o supe ised
models, in o de o implemen a holis ic ool. Fu he mo e, a ool is p oposed ha gene a es speci ic
isualiza ions o machine ope a o s and main enance s a . This would help hem in supe ising he
manu ac u ing and in comple ing he machine’s main enance. Addi ionally, HUEPs a e being applied
o some o he componen s/machines, in o de o alida e hei abili y o ope a ion moni o ing and
ailu e de ec ion. The use o HUEPs o quali y pu poses will also be explo ed, in o de o imp o e he
de ec ion o quali y de ec s.
Au ho Con ibu ions:
Concep ualiza ion,
Á
.H., E.C. and J.S.; me hodology,
Á
.H.; so wa e, R.R.; o mal analysis,
R.R. and
Á
.H.; da a cu a ion R.R.; w i ing—o iginal d a p epa a ion, R.R. and
Á
.H.; w i ing— e iew and edi ing,
R.R.,
Á
.H., E.C. and J.S.; supe ision,
Á
.H., E.C. and J.S. All au ho s ha e ead and ag eed o he published e sion
o he manusc ip .
Funding: This esea ch ecei ed no ex e nal unding.
Appl. Sci. 2020,10, 4355 18 o 20
Acknowledgmen s:
The au ho s would like o hank he ehicle in e io s manu ac u e , G upo An olin, o i s
collabo a ion in his esea ch.
Con lic s o In e es : The au ho s decla e no con lic o in e es .
Abb e ia ions
The ollowing abb e ia ions a e used in his manusc ip , in alphabe ic o de :
ANN A i icial Neu al Ne wo ks
CMDS Classical Mul idimensional Scaling
CMLHL
Coope a i e Maximum-Likelihood Hebbian Lea ning
EPP Explo a o y P ojec ion Pu sui
FA Fac o Analysis
FD Failu e De ec ion
HP Hyd aulic Pis on
HUEP Hyb id Unsupe ised Explo a o y Plo
IoT In e ne o Things
KNN k-Nea es Neighbou
MDS Mul idimensional scaling
ML Machine Lea ning
MLHL Maximum-Likelihood Hebbian Lea ning
PCA P incipal Componen Analysis
PdM P edic i e Main enance
SH Seal Head
SM Sammon Mapping
Re e ences
1.
Zhou, K.; Liu, T.; Zhou, L. Indus y 4.0: Towa ds Fu u e Indus ial Oppo uni ies and Challenges.
In P oceedings o he 2015 12 h In e na ional Con e ence on Fuzzy Sys ems and Knowledge Disco e y
(FSKD), Zhangjiajie, China, 15–17 Augus 2015; pp. 2147–2152. [C ossRe ]
2.
Khan, M.; Xiao ong, W.; Xiaolong, X.; Wanchun, D. Big Da a Challenges and Oppo uni ies in he Hype o
Indus y 4.0. In P oceedings o he 2017 IEEE In e na ional Con e ence on Communica ions (ICC), Pa is,
F ance, 21–25 May 2017. [C ossRe ]
3.
Del Campo, G.; Cala a a, S.; Canada, G.; Olloqui, J.; Ma inez, R.; San ama ia, A. IoT Solu ion o Ene gy
Op imiza ion in Indus y 4.0: Issues o a Real-Li e Implemen a ion. In P oceedings o he 2018 Global
In e ne o Things Summi (GIoTS), Bilbao, Spain, 4–7 June 2018. [C ossRe ]
4.
Va hoopan, M.; Johny, M.; Zoi l, A.; Knoll, A. Modula Faul Asc ip ion and Co ec i e Main enance
Using a Digi al Twin. In P oceedings o he 16 h IFAC Symposium on In o ma ion Con ol P oblems in
Manu ac u ing INCOM 2018, Be gamo, I aly, 11–13 June 2018; pp. 1041–1046. [C ossRe ]
5.
Sha iq, S.I.; Szcze bicki, E.; Sanin, C. Manu ac u ing Da a Analysis in In e ne o Things/In e ne o Da a
(IoT/IoD) Scena io. Cybe n. Sys . 2018,49, 280–295. [C ossRe ]
6.
Qu, Y.J.; Ming, X.G.; Liu, Z.W.; Zhang, X.Y.; Hou, Z.T. Sma Manu ac u ing Sys ems: S a e o he A and
Fu u e T ends. In . J. Ad . Manu . Technol. 2019,103, 3751–3768. [C ossRe ]
7.
He e o,
Á
.; Jim
é
nez, A.; Bay ak a , S. Hyb id Unsupe ised Explo a o y Plo s: A Case S udy o Analysing
Fo eign Di ec In es men . Complexi y 2019,2019. [C ossRe ]
8. F iedman, J.H. Explo a o y P ojec ion Pu sui . J. Am. S a . Assoc. 1987,82, 249–266. [C ossRe ]
9.
Jain, A.K.; Mu y, M.N.; Flynn, P.J. Da a Clus e ing: A Re iew. ACM Compu . Su .
1999
,31, 264–323.
[C ossRe ]
10. G upo An olin. A ailable online: h ps://www.g upoan olin.com/(accessed on 14 May 2020).
11.
Pa a, J.; Del Se , J.; Neb o, A.J.; Zu u uza, U.; He e a, F. Analyze, Sense, P ep ocess, P edic , Implemen , and
Deploy (ASPPID): An Inc emen al Me hodology Based on Da a Analy ics o Cos -E icien ly Moni o ing he
Indus y 4.0. Eng. Appl. A i . In ell. 2019,82, 30–43. [C ossRe ]
12.
Skiena, S.S. Visualizing Da a. In The Da a Science Design Manual; Tex s in Compu e Science; Sp inge : Cham,
Swi ze land, 2017; pp. 155–200. [C ossRe ]
Appl. Sci. 2020,10, 4355 19 o 20
13.
Posada, J.; To o, C.; Ba andia an, I.; Oya zun, D.; S icke , D.; De Amicis, R.; Pin o, E.B.; Eise , P.; Döllne , J.;
Valla ino, I. Visual Compu ing as a Key Enabling Technology o Indus ie 4.0 and Indus ial In e ne .
IEEE Compu . G aph. Appl. 2015,35, 26–40. [C ossRe ]
14.
Segu a,
Á
.; Diez, H.V.; Ba andia an, I.; A belaiz, A.;
Á
l a ez, H.; Sim
õ
es, B.; Posada, J.; Ga c
í
a-Alonso, A.;
Uga e, R. Visual Compu ing Technologies o Suppo he Ope a o 4.0. Compu . Ind. Eng.
2020
,139, 105550.
[C ossRe ]
15.
Posada, J.; Zo illa, M.; Dominguez, A.; Sim
õ
es, B.; Eise , P.; S icke , D.; Rambach, J.; Dollne , J.; Gue a a, M.
G aphics and Media Technologies o Ope a o s in Indus y 4.0. IEEE Compu . G aph. Appl.
2018
,38, 119–132.
[C ossRe ]
16.
Jimenez-Co adi, A.; I igoien, I.; Bo o, F.; Sie a, B.; Rod iguez, G. P edic i e Main enance on he Machining
P ocess and Machine Tool. Appl. Sci. 2020,10, 224. [C ossRe ]
17.
Uhlmann, E.; Pon es, R.P.; Geise , C.; Hohwiele , E. Clus e Iden i ica ion o Senso Da a o P edic i e
Main enance in a Selec i e Lase Mel ing Machine Tool. In P oceedings o he 4 h In e na ional Con e ence
on Sys em-In eg a ed In elligence: In elligen , Flexible and Connec ed Sys ems in P oduc s and P oduc ion,
Hanno e , Ge many, 19–20 June 2018; pp. 60–65. [C ossRe ]
18.
Diez-Oli an, A.; Pag
á
n, J.A.; Sanz, R.; Sie a, B. Da a-D i en P ognos ics Using a Combina ion o Cons ained
K-Means Clus e ing, Fuzzy Modeling and LOF-Based Sco e. Neu ocompu ing 2017,241, 97–107. [C ossRe ]
19.
Yoo, Y.; Pa k, S.H.; Baek, J.-G. A Clus e ing-Based Equipmen Condi ion Model o Chemical Vapo Deposi ion
P ocess. In . J. P ecis. Eng. Manu . 2019, 1677–1689. [C ossRe ]
20.
Jia, F.; Lei, Y.; Guo, L.; Lin, J.; Xing, S. A Neu al Ne wo k Cons uc ed by Deep Lea ning Technique and I s
Applica ion o In elligen Faul Diagnosis o Machines. Neu ocompu ing 2018,272, 619–628. [C ossRe ]
21.
Lei, Y.; Jia, F.; Lin, J.; Xing, S.; Ding, S.X. An In elligen Faul Diagnosis Me hod Using Unsupe ised Fea u e
Lea ning Towa ds Mechanical Big Da a. IEEE T ans. Ind. Elec on. 2016,63, 3137–3147. [C ossRe ]
22.
Delgado-P ie o, M.; Ci incione, G.; Espinosa, A.G.; O ega, J.A.; Henao, H. Bea ing Faul De ec ion by a
No el Condi ion-Moni o ing Scheme Based on S a is ical-Time Fea u es and Neu al Ne wo ks. IEEE T ans.
Ind. Elec on. 2013,60, 3398–3407. [C ossRe ]
23.
Pacheco, F.; De Oli ei a, J.V.; S
á
nchez, R.-V.; Ce ada, M.; Cab e a, D.; Li, C.; Zu i a, G.; A
é
s, M. A S a is ical
Compa ison o Neu oclassi ie s and Fea u e Selec ion Me hods o Gea box Faul Diagnosis unde Realis ic
Condi ions. Neu ocompu ing 2016,194, 192–206. [C ossRe ]
24.
Jia, F.; Lei, Y.; Lin, J.; Zhou, X.; Lu, N. Deep Neu al Ne wo ks: A P omising Tool o Faul Cha ac e is ic
Mining and In elligen Diagnosis o Ro a ing Machine y wi h Massi e Da a. Mech. Sys . Signal P ocess.
2016
,
72–73, 303–315. [C ossRe ]
25.
Yang, Y.-H.; Pan, Y.-K.; Zhang, L.-P.; Liu, X.-Z. Incipien Faul De ec ion Me hod Based on S eam Da a
P ojec ion T ans o ma ion Analysis. IEEE Access 2019,7, 93062–93075. [C ossRe ]
26.
Li, X.; Zhang, W.; Ding, Q. A Robus In elligen Faul Diagnosis Me hod o Rolling Elemen Bea ings Based
on Deep Dis ance Me ic Lea ning. Neu ocompu ing 2018,310, 77–95. [C ossRe ]
27.
Lei, Y.; He, Z.; Zi, Y.; Chen, X. New Clus e ing Algo i hm-Based Faul Diagnosis Using Compensa ion
Dis ance E alua ion Technique. Mech. Sys . Signal P ocess. 2008,22, 419–435. [C ossRe ]
28.
Liu, H.; Zhou, J.; Xu, Y.; Zheng, Y.; Peng, X.; Jiang, W. Unsupe ised Faul Diagnosis o Rolling Bea ings Using
a Deep Neu al Ne wo k Based on Gene a i e Ad e sa ial Ne wo ks. Neu ocompu ing
2018
,315, 412–424.
[C ossRe ]
29.
Jianbo, Y. Local and Nonlocal P ese ing P ojec ion o Bea ing De ec Classi ica ion and Pe o mance
Assessmen . IEEE T ans. Ind. Elec on. 2012,59, 2363–2376. [C ossRe ]
30.
Li, W.; Zhang, S.; Rakheja, S. Fea u e Denoising and Nea es –Fa hes Dis ance P ese ing P ojec ion o
Machine Faul Diagnosis. IEEE T ans. Ind. In o m. 2016,12, 393–404. [C ossRe ]
31.
Chen, X.-L.; Wang, P.-H.; Hao, Y.-S.; Zhao, M. E iden ial KNN-Based Condi ion Moni o ing and Ea ly
Wa ning Me hod wi h Applica ions in Powe Plan . Neu ocompu ing 2018,315, 18–32. [C ossRe ]
32.
Wang, D. K-Nea es Neighbo s Based Me hods o Iden i ica ion o Di e en Gea C ack Le els unde
Di e en Mo o Speeds and Loads: Re isi ed. Mech. Sys . Signal P ocess. 2016,70–71, 201–208. [C ossRe ]
33.
Luwei, K.C.; Yunusa-Kal ungo, A.; Sha’Aban, Y. In eg a ed Faul De ec ion F amewo k o Classi ying
Ro a ing Machine Faul s Using F equency Domain Da a Fusion and A i icial Neu al Ne wo ks. Machines
2018,6, 59. [C ossRe ]
Appl. Sci. 2020,10, 4355 20 o 20
34.
You, D.; Gao, X.; Ka ayama, S. WPD-PCA-Based Lase Welding P ocess Moni o ing and De ec s Diagnosis
by Using FNN and SVM. IEEE T ans. Ind. Elec on. 2015,62, 628–636. [C ossRe ]
35.
Zhao, C.; Gao, F. Faul -Rele an P incipal Componen Analysis (FPCA) Me hod o Mul i a ia e S a is ical
Modeling and P ocess Moni o ing. Chemom. In ell. Lab. Sys . 2014,133, 1–16. [C ossRe ]
36.
Deng, X.; Tian, X.; Chen, S.; Ha is, C.J. Faul Disc iminan Enhanced Ke nel P incipal Componen Analysis
Inco po a ing P io Faul In o ma ion o Moni o ing Nonlinea P ocesses. Chemom. In ell. Lab. Sys .
2017
,
162, 21–34. [C ossRe ]
37.
Yang, Y.-H.; Chen, X.; Zhang, Y.; Liu, X. A No el Decen alized Weigh ed Relie F-PCA Me hod o Faul
De ec ion. IEEE Access 2019,7, 140478–140487. [C ossRe ]
38.
MacQueen, J.B. Some Me hods o Classi ica ion and Analysis o Mul i a ia e Obse a ions; Wes e n Managemen
Science In e na ional Uni e si y o Cali o nia: Los Angeles, CA, USA, 1966.
39.
Jolli e, I.T. P incipal Componen Analysis; (Sp inge Se ies in S a is ics); Sp inge -Ve lag: New Yo k, NY, USA,
2002. [C ossRe ]
40.
Co chado, E.; Macdonald, D.; Fy e, C. Maximum and Minimum Likelihood Hebbian Lea ning o Explo a o y
P ojec ion Pu sui . Da a Min. Knowl. Disco . 2004,8, 203–225. [C ossRe ]
41.
Co chado, E.; Fy e, C. Connec ionis Techniques o he Iden i ica ion and Supp ession o In e e ing
Unde lying Fac o s. In . J. Pa e n Recogni . A i . In ell. 2003,17, 1447–1466. [C ossRe ]
42.
To ge son, W.S. Mul idimensional Scaling: I. Theo y and Me hod. Psychome ika
1952
,17, 401–419. [C ossRe ]
43.
Wang, J. Classical Mul idimensional Scaling. Geom. S uc . High-Dimens. Da a Dimens. Reduc .
2012
, 115–129.
[C ossRe ]
44.
Gowe , J.C. P incipal Coo dina es Analysis. In Wiley S a sRe : S a is ics Re e ence Online; Ame ican Cance
Socie y: A lan a, GA, USA, 2015; pp. 1–7. [C ossRe ]
45.
Sammon, J.W. A Nonlinea Mapping o Da a S uc u e Analysis. IEEE T ans. Compu .
1969
,C–18, 401–409.
[C ossRe ]
46. Hende son, P. Sammon mapping. Pa e n Recogni . Le . 1997,18, 1307–1316.
47.
Le ne , B.; Gu e man, H.; Aladjem, M.; Dins ein, I. On he Ini ialisa ion o Sammon’s Nonlinea Mapping.
Pa e n Anal. Appl. 2000,3, 61–68. [C ossRe ]
48. Hä dle, W.K.; Sima , L. Fac o Analysis. Appl. Mul i a . S a . Anal. 2015, 359–384. [C ossRe ]
49. Cle , T. Fac o Analysis. Appl. S a . Mul i a . Da a Anal. Bus. Econ. 2019, 433–446. [C ossRe ]
50. Kong, C. Wa e -Je Cu ing. CIRP Encycl. P od. Eng. 2019, 1803–1807. [C ossRe ]
51.
KMT S eamline SL-V Pumps Ca alog.Pd . A ailable online: h ps://www.km wa e je .com/KMT%
20S eamline%20SL-V%20Pumps%20Ca alog.pd (accessed on 14 May 2020).
52.
Cleophas, T.J.; Zwinde man, A.H. Densi y-Based Clus e ing o Iden i y Ou lie G oups in O he wise
Homogeneous Da a (50 Pa ien s). Mach. Lea n. Med. Cookb. 2014, 9–11. [C ossRe ]
©
2020 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/).