A Hyper-heuristic Inspired Methodology for Failure Prediction in the Context of Industry 4.0

A Hype -heu is ic Inspi ed
Me hodology o Failu e
P edic ion in he Con ex o
Indus y 4.0
by
Ad iana Na ajas Gue e o
Doc o o Philosophy
May o 2023
Facul y o Enginee ing o Bilbao
Uni e si y o he Basque Coun y, UPV/EHU
(cc)2023 ADRIANA NAVAJAS GUERRERO (cc by-nc-nd 4.0)
A Hype -heu is ic Inspi ed
Me hodology o Failu e
P edic ion in he Con ex o
Indus y 4.0
by
Ad iana Na ajas Gue e o
Supe iso s: PhD. E a Po illo & PhD. Diana
Manja es
Depa men o Au oma ic Con ol and Sys ems
Enginee ing
Doc o o Philosophy
2023
Facul y o Enginee ing o Bilbao
Uni e si y o he Basque Coun y, UPV/EHU
Au ho ’s igh 2023 by
Na ajas Gue e o, Ad iana

Acknowledgmen s
No nega ´e que me es ´a cos ando sabe c´omo empeza con es as
l´ıneas. ´
Es as, que e´ıa an lejos po ene que esc ibi y ya han lle-
gado a es e lib o. Tampoco oy a nega que han sido a˜nos de es ue zo,
abajo, noches la gas, d´ıas e e nos, su imien o, oces, llo os, deses-
pe aci´on, dudas, isas, i encias, pe sonas y se endipias. Aun as´ı, so-
b e odo es a in e esan e y nu i i a e apa de mi ida ha sido cambio,
madu ez y i encias. Cambio en odos los sen idos, an o en el nue o
umbo que om´o mi ida iniendo a Bilbao y omando la decisi´on de
comenza es a a en u a, como en mi o ma de pensa , azona y mod-
i ica mis o mas de in e eac ua con odos los es ´ımulos que nos da la
ida. Pe o sob e odo, a ni el pe sonal he lib ado una de las mayo es
ba allas de mi ida con a mi peo enemiga, yo misma. Echando la
is a a ´as, di ´e que he madu ado eno memen e y que puedo deci ,
que me sien o o gullosa de lo que soy hoy en d´ıa. C eo que, si bien ha
sido una e apa du a, es a e apa la eco da ´e como la que me ayud´o
a pensa , azona , ene c i e io, que e i i , se ole an e, hones a,
emp´a ica, ca i˜nosa . . .
Sin duda, es a e apa ha sido el esul ado de un es ue zo pe sonal,
pe o sob e odo ha sido el u o del abajo en equipo y el apoyo
incondicional de muchas pe sonas que ya o maban pa e de mi ida,
que llega on pa a queda se o que se han ido. Es po ello que, sin odas
ellas, no es a ´ıa aqu´ı sen ada esc ibiendo es as l´ıneas hoy en d´ıa. Po
es a az´on, deseo que odas es as pe sonas sean pa e de mi lib o, de
mi his o ia y de es a esis.
En p ime luga , ag adece a mis di ec o as E a Po illo y Diana
Manja es, po se apoyo en buenos y malos momen os, po con ia
en mi incluso cuando yo dudaba de mi misma, po se c ´ı icas, po
no pe mi i me i a la oalla y po anima me cuando es aba c uzando
un eno me ´unel neg o. Jun o a ellas he ap endido muchas cosas en
es os a˜nos, an o en el plano acad´emico como en el pe sonal, que han
i
hecho que a d´ıa de hoy ambi´en sea como soy. Po odo, GRACIAS.
Tambi´en quie o ag adece a Tecnalia po da me la opo unidad
de empeza la que sin duda a a se la e apa m´as ele an e en mi
ida. Especialmen e da las g acias a odo el equipo de OPTIMA, a
mis AIdeanos y a oda la auna animal´ıs ica (cab as, acas, o i os,
koalas, pe ezosos...) que me ha acompa˜nado en los ´ul imos meses de
es a esis y me han sacado una son isa cada d´ıa. Debo hace una
menci´on especial a Se gio Gil, quien siemp e ha es ado a mi lado
apoy´andome, d´andome consejos y siendo pa e implicada del camino.
Me gus a ´ıa ex ende mi m´as since o ag adecimien o a I˜naki Olaba i-
e a, quien ha sido de g an ayuda du an e el p oceso y me ha ense˜nado
aliosas lecciones. Tambi´en quisie a ag adece a Isido o ci i´on po
sus cha las inspi ado as y su apoyo. No puedo deja de menciona
a I aide. Ella siemp e ha es ado p esen e con sus consejos sabios,
su apoyo incondicional, sus ab azos econ o an es, sus mensajes de
´animo, sus isas con agiosas y sus cha las inspi ado as. Su p esencia
ha sido undamen al en momen os de ince idumb e y di icul ades, y
su apoyo me ha dado la ue za y la mo i aci´on necesa ias pa a segui
adelan e.
Desde luego, no puedo ol ida me de mi gen e.
Ag adezco eno memen e a mis pad es odo lo que engo, soy y he
conseguido desde que engo uso de az´on. A ellos, po inculca me
buenos alo es, po apoya me incondicionalmen e en odo, po que -
e me, po se como son... sencillamen e po absolu amen e odo, os
quie o.
Es a esis no hubie a sido posible sin las buenas amis ades, las
isas omando unas ce ezas, los iajes, las in e minables llamadas de
el´e ono, las sesiones de e lexi´on pa a cambia el mundo, los a a de-
ce es, los ab azos, en de ini i a, sin las pe sonas de ´as de odos esos
momen os. . . po eso G acias a Igo , Idoia, Dani, Ma a, Elena, Paula,
I is, Nani, Ma ´ıa, Valle, Bego, Ped o, Ibai . . .
ii
Quie o hace dos menciones especiales, en p ime luga , a A i z.
A ´el le engo que ag adece se un amigo incondicional en es os a˜nos,
se apoyo y ayuda con inuos y po compa i m´usicas, en enamien-
os, Hea dles, a en u as, paseos y amis ad. En segundo luga , a mi
compa˜ne a y amiga desde el d´ıa en el que me emba qu´e en es a a en-
u a, a mi ho elana a o i a, a mi compa˜ne a de iaje, a I a xe. . .
GRACIAS.
Finalmen e, no que ´ıa acaba sin menciona a dos pe sonas, que
no son de oda la ida, pe o espe o que sean pa a oda la ida. Dos
se endipias que han llegado como un allo de luz en plena oscu idad,
dos lianas en las que en los peo es momen os me he podido colga y
han es ado ah´ı sin pedi lo, dos pe sonas inc e´ıbles que me han egalado
momen os ma a illosos, a ellos dos, I aia y Albe o GRACIAS po
odo.
iii
Table 5.6 Resul s o MRC, F1-sco e and TM in MV da aba-
ses by β, whe e; DB: Da abase, IQR: In e qua ile ange,
CS: Clus e ing Solu ion, TM: T us wo hiness me ic, D:
Di o ce, I: I is, Io: Ionosphe e, W: Wine, C: Ce ix, B:
Bupa, L: Lymphog aphy. . . . . . . . . . . . . . . . . . . 104
Table 5.7 Cha ac e is ics o UCR da abases. . . . . . . . . . . 106
Table 5.8 Resul s o MRC, F1-sco e and TM in MV da abases
by β, whe e; DB: Da abase, IQR: In e qua ile ange, CS:
Clus e ing Solu ion, TM: T us wo hiness me ic, S: Sony,
A: A ow, G: Gun, ECG: Elec oca diog am and M: Mo e. 107
Table 5.9 Ob ained da ase s in he case o s udy. . . . . . . . 112
Table 5.10 Pa ame e alues o PLAHS in he cold s amping
case s udy, whe e; Nd: To al numbe o elemen in da ase ,
NdL: numbe o labelled b eakage-s ops, NdnL: numbe
o unlabelled b eakage-s ops, Nk: numbe o know classes. 113
Table 5.11 Resul s o PLAHS in he cold s amping p ocess in
e ms o MRC, F1-sco e and TM and labelling p oposal,
whe e; IQR: In e qua ile ange, CS: Clus e ing Solu ion,
TM: T us wo hiness me ic. . . . . . . . . . . . . . . . . 114
Table 5.12 Time o execu ion pe p ocess signal, whe e is he
mean ime alue........................ 115
Table 6.1 Compa ison o he li e a u e and he p oposed hype -
heu is ic inspi ed app oach o AD. LER: Le el o Expe -
ise Requi emen , Wz: ime-window size, Th: Th eshold,
C: Calcula ed, A: Adap i e, Fx: Fixed, O: Op imized, R:
Random, CV: C oss Valida ion, GS: G id Sea ch, me a-
Op *: Compa ison be ween PSO, SHO, KH, SSA, me a-
Op **: Compa ison be ween PSO, DE, GA, GWO. . . . . 121
Table 6.2 Fea u es o s a is ics ex ac ed om TSM, whe e
Fq: F equency-domain and T: Time-domain, SNR: Signal
o Noise Ra io, S d: S anda d de ia ion. . . . . . . . . . . 127
Table 6.3 De ini ion and o mula ion o he ea u es used in
ime-domain, whe e Wzj: ime-window . . . . . . . . . . 134
x

Table 6.4 De ini ion and o mula ion o he ea u es used in
equency-domain, whe e Wzj: ime-window . . . . . . . 134
Table 6.5 Co esponding No e o each Fea u e, whe e F: F equency-
domain and T: Time-domain, SNR: Signal o Noise Ra io,
S d: S anda d de ia ion. . . . . . . . . . . . . . . . . . . . 146
Table 6.6 Feasible alues o no es whe e n is he ini ial an-
dom alue o no e, nnis a new alue o no e . . . . . . . 148
Table 6.7 Case o s udy: ypes o ailu e, numbe o TSM (M)
and numbe o SBD (N). . . . . . . . . . . . . . . . . . . . 157
Table 6.8 Op imal HS ope a o con igu a ion. . . . . . . . . . 161
Table 6.9 Mean AUC MOD alue pe HS ope a o se o Fail-
u eA.............................. 162
Table 6.10 Mean AUC MOD alue pe HS ope a o se o
Failu eB............................ 164
Table 6.11 Mean AUC MOD alue pe HS ope a o se o
Failu eC............................ 165
Table 6.12 Mean i e a ion con e gence o bes se ope a o s’s
alues.............................. 166
Table 6.13 Mean AUC MOD alue pe HS ope a o se o
Failu eD............................ 167
Table 6.14 Mean AUC MOD alue pe HS ope a o se in Fail-
u eE. ............................. 168
Table 6.15 Mean AUC MOD alue pe HS ope a o se o
Failu eF. ........................... 170
Table 6.16 Resul s o ailu e ype A: b eakage hole punch.
H: Ha mony, A-M: AUC MOD, A-R: AUC ROC, SNT:
Sensi i i y, SPC: Speci ici y, FPR: False Posi i e Ra e . . 173
Table 6.17 Resul s o ailu e ype B: b eakage hole punch.
H: Ha mony, A-M: AUC MOD, A-R: AUC ROC, SNT:
Sensi i i y, SPC: Speci ici y, FPR: False Posi i e Ra e . . 176
Table 6.18 Resul s o ailu e ype C: b eakage calib a ion punch.
H: Ha mony, A-M: AUC MOD, A-R: AUC ROC, SNT:
Sensi i i y, SPC: Speci ici y, FPR: False Posi i e Ra e . . 182
xi
Table 6.19 Resul s o ailu e ype D: b eakage punch. H: Ha -
mony, A-M: AUC MOD, A-R: AUC ROC, SNT: Sensi i -
i y, SPC: Speci ici y, FPR: False Posi i e Ra e. . . . . . . 185
Table 6.20 Resul s o ailu e ype E: wea Calib a ion Punch.
H: Ha mony, A-M: AUC MOD, A-R: AUC ROC, SNT:
Sensi i i y, SPC: Speci ici y, FPR: False Posi i e Ra e. . . 186
Table 6.21 Resul s o ailu e ype F: wea nugge e acua ion
ube. H: Ha mony, A-M: AUC MOD, A-R: AUC ROC,
SNT: Sensi i i y, SPC: Speci ici y, FPR: False Posi i e Ra e.192
Figu e 1.1 Le el o digi aliza ion o he e iewed manu ac u -
ing companies, sou ce: [2]. . . . . . . . . . . . . . . . . . . 3
Figu e 1.2 Deg ee o Indus y 4.0 echnologies implemen a ion
in SMS companies [3]. . . . . . . . . . . . . . . . . . . . 3
Figu e 1.3 Types o da a sou ces in he Manu ac u e Indus y
[2]. .............................. 4
Figu e 1.4 Type o analysis used in Manu ac u e Indus y [2]. 5
Figu e 2.1 S a e o he a scheme. . . . . . . . . . . . . . . . 13
Figu e 2.2 Indus ial e olu ions imeline. . . . . . . . . . . . 15
Figu e 2.3 Enabling echnologies o Indus y 4.0, sou ce: [4]
e-illus a ed and modi ied o his Thesis. . . . . . . . . . 16
Figu e 2.4 Imp o emen ac ions o he 8 main a eas o in e -
es and i s co esponding enabling echnologies in Indus y
4.0. Sou ce :[5] e-illus a ed and modi ied o his Thesis.
CPS: Cybe Physical Sys em, RTO: Real- ime op imiza-
ion, ML: ML, AR: Augmen ed Reali y, IoT: In e ne o
Things, AM: Addi i e manu ac u ing. . . . . . . . . . . . 22
Figu e 2.5 Ca ego iza ion and de ini ion o p ognosis me hods. 26
Figu e 2.6 Ou lie anomaly example [6]. . . . . . . . . . . . . 28
Figu e 2.7 Con ex ual anomalies example [6]. . . . . . . . . . 28
Figu e 2.8 Collec i e anomalies example [6]. . . . . . . . . . . 29
xii
Figu e 2.9 Semi-supe ised axonomy o AD. . . . . . . . . . 31
Figu e 2.10 Concep ual scheme o he p oposed axonomy o
SSLclus e ing. ........................ 34
Figu e 2.11 Taxonomy o hype -heu is ics. . . . . . . . . . . . 37
Figu e 2.12 Rele an pa ame e s in he AD con ex . . . . . . 44
Figu e 3.1 Hype -heu is ic inspi ed me hodology applied in
hisThesis. .......................... 57
Figu e 4.1 Shee me al cold-s amping p ocess. . . . . . . . . . 67
Figu e 4.2 Wi e od cold-s amping p ocess. . . . . . . . . . . 68
Figu e 4.3 Examples o shee me al cold-s amping pieces. . . 69
Figu e 4.4 Examples o wi e od cold-s amping pieces. . . . . 69
Figu e 4.5 Die holding and i e ing pin blocks in cold s amping
machine............................. 70
Figu e 4.6 Feed olle s and s aigh ene s. . . . . . . . . . . . . 71
Figu e 4.7 Bol cold o ming p ocess. . . . . . . . . . . . . . . 72
Figu e 4.8 Flow diag am o he cold o ming p ocess unde s udy. 73
Figu e 4.9 Signals cap u ed om he cold o ming p ocess in
wos amps........................... 75
Figu e 4.10 P ocess signal 1. . . . . . . . . . . . . . . . . . . . 75
Figu e 4.11 P ocess signal 2. . . . . . . . . . . . . . . . . . . . 76
Figu e 4.12 P ocess signal 3. . . . . . . . . . . . . . . . . . . . 76
Figu e 4.13 P ocess signal 4. . . . . . . . . . . . . . . . . . . . 77
Figu e 4.14 P ocess signal 5. . . . . . . . . . . . . . . . . . . . 77
Figu e 4.15 P ocess signal 6. . . . . . . . . . . . . . . . . . . . 78
Figu e 4.16 Flow diag am o da ase s c ea ion p ocess. . . . . 78
Figu e 5.1 Rela ion be ween amoun o labels and lea ning
me hods. ........................... 82
Figu e 5.2 Flow diag am o he p oposed PLAHS. . . . . . . . 86
Figu e 5.3 Example o he Bagging scheme p oposed in PLAHS. 88
Figu e 5.4 Example o clus e ing solu ions. . . . . . . . . . . . 96
Figu e 5.5 Flow diag am o he PLAHS labelling sys em. . . . 98
Figu e 5.6 TM p ocess e alua ion. . . . . . . . . . . . . . . . 100
xiii
Figu e 5.7 Co ela ion o TM and F1-sco e o unsupe ised
elemen s in UCI da abases. . . . . . . . . . . . . . . . . . 105
Figu e 5.8 Co ela ion o TM and F1-sco e o unknown ele-
men s in UCR da abases. . . . . . . . . . . . . . . . . . . 108
Figu e 5.9 Sepa abili y o da abases, whe e: S: Sony, ECG:
Elec oca diog am, A: A ow, G: Gun, M: Mo e, I: I is,
W: Wine, D: Di o ce, Io: Ionosphe e, B: Bupa, C: Ce ix
and L: Lymphog aphy. . . . . . . . . . . . . . . . . . . . . 110
Figu e 5.10 PLAHS implemen a ion in he eal use case. . . . 113
Figu e 5.11 Vo ing solu ions o BS8-BS9. . . . . . . . . . . . 115
Figu e 6.1 Pa ame e s ela ed wi h he ailu e p edic ion. . . 122
Figu e 6.2 Concep ual scheme o he p oposed HIMAFP. . . . 126
Figu e 6.3 Examples o wo possible se o heu is ic pa am-
e e s [Wz, Fe, ThH, ThL] o a speci ic ype o ailu e in
TSM1. The uppe igu e shows he expec ed beha iou
ha de ec s he anomaly in he p e ious ime-window o
he ailu e (Wz1). The lowe igu e shows a se o pa am-
e e s ha is no p edic o o he ype o ailu e o in e es . 136
Figu e 6.4 Flow diag am o he p oposed HIMAFP o collec-
i eAD............................. 138
Figu e 6.5 Reo ganiza ion and clipping p ocedu e. Fi s ow
shows he inpu da a o ma be o e segmen a ion, second
ow shows he eo ganiza ion pe TSM, las ow shows how
da a is clipped and inally adap ed o he HIMAFP. . . . 144
Figu e 6.6 Th eshold pa ame e explana ion. . . . . . . . . . 146
Figu e 6.7 Di e en ime-window simula ions o 
EV (E1 o
E6) and i s e alua ion wi h AUC MOD and AUC ROC. . 153
Figu e 6.8 E alua ion p ocess o AUC MOD s AUC ROC o
casesE1 oE6......................... 153
Figu e 6.9 Example o HIMAFP isualiza ion esul s. . . . . . 155
Figu e 6.10 Signal moni o ing o ailu e ype A. . . . . . . . 158
Figu e 6.11 Signal moni o ing o ailu e ype B. . . . . . . . 158
Figu e 6.12 Signal moni o ing o ailu e ype C. . . . . . . . 159
Figu e 6.13 Signal moni o ing o ailu e ype D. . . . . . . . 159
xi
Figu e 6.14 Signal moni o ing o ailu e ype E. . . . . . . . 160
Figu e 6.15 Signal moni o ing o ailu e ype F. . . . . . . . 160
Figu e 6.16 Me ic e olu ion AUC MOD in Failu e ype A o
all combina ion o HS ope a o alues o TSM6. . . . . . 163
Figu e 6.17 Me ic e olu ion AUC MOD in Failu e ype B o
all combina ion o HS ope a o alues o TSM1. . . . . . 164
Figu e 6.18 Me ic e olu ion AUC MOD o Failu e ype C o
all combina ion o HS ope a o alues o TSM2. . . . . . 166
Figu e 6.19 Me ic e olu ion AUC MOD o Failu e ype D o
all combina ion o HS ope a o alues o TSM2. . . . . . 168
Figu e 6.20 Me ic e olu ion AUC MOD o Failu e ype E o
all combina ion o HS ope a o alues o TSM5. . . . . . 169
Figu e 6.21 Me ic e olu ion AUC MOD o Failu e ype F o
all combina ion o HS ope a o alues in TSM1. . . . . . . 170
Figu e 6.22 HIMAFP SBDs solu ions o Ha mony H2.1 in
Failu eA(I) ......................... 175
Figu e 6.23 HIMAFP SBDs solu ions o Ha mony H2.1 in
Failu eA(II)......................... 176
Figu e 6.24 HIMAFP SBDs solu ions o Ha mony H6.1 in
Failu eA(I) ......................... 177
Figu e 6.25 HIMAFP SBDs solu ions o Ha mony H6.1 in
Failu eA(II)......................... 178
Figu e 6.26 HIMAFP SBDs solu ions o Ha mony H3.4 in
Failu eB ........................... 180
Figu e 6.27 HIMAFP SBDs solu ions o Ha mony H5.1 in
Failu eB ........................... 181
Figu e 6.28 HIMAFP SBDs solu ions o Ha mony H3.1 in
Failu eC............................ 183
Figu e 6.29 HIMAFP SBDs solu ions o Ha mony H3.3 in
Failu eC............................ 184
Figu e 6.30 HIMAFP SBDs solu ions o Ha mony H1.1 o
Failu eD ........................... 187
Figu e 6.31 HIMAFP SBDs solu ions o Ha mony H6.1 o
Failu eD ........................... 188
x

Figu e 6.32 HIMAFP SBDs solu ions o Ha mony H2.2 in
Failu eE ........................... 190
Figu e 6.33 HIMAFP SBDs solu ions o Ha mon H2.4 in Fail-
u eE ............................. 191
Figu e 6.34 HIMAFP SBDs solu ions o Ha mon H3.1 in Fail-
u eF ............................. 193
Figu e 6.35 HIMAFP SBDs solu ions o Ha mon H5.1 in Fail-
u eF ............................. 194
Figu e 6.36 Boxplo g aphic o he dis ibu ion o TTF h ough
he di e en ha monies pe TSM in ailu e ype D. . . . . 196
Figu e 6.37 RTS o ha mony H2.1 o TSM2pe SBD whe e
RTS: Real Time Simula ion. . . . . . . . . . . . . . . . . . 197
Figu e 6.38 RTS o ha mony H4.1 o TSM4pe SBD whe e
RTS: Real Time Simula ion. . . . . . . . . . . . . . . . . 198
Figu e A.1 Me ic e olu ion AUC MOD in Failu e ype A in
TSM1 o all combina ion o HS ope a o alues. . . . . . 215
Figu e A.2 Me ic e olu ion AUC MOD in Failu e ype A in
TSM2 o all combina ion o HS ope a o alues. . . . . . 216
Figu e A.3 Me ic e olu ion AUC MOD in Failu e ype A in
TSM3 o all combina ion o HS ope a o alues. . . . . . 216
Figu e A.4 Me ic e olu ion AUC MOD in Failu e ype A in
TSM4 o all combina ion o HS ope a o alues. . . . . . 217
Figu e A.5 Me ic e olu ion AUC MOD in Failu e ype A in
TSM5 o all combina ion o HS ope a o alues. . . . . . 217
Figu e A.6 Me ic e olu ion AUC MOD in Failu e ype A in
TSM6 o all combina ion o HS ope a o alues. . . . . . 218
Figu e A.7 Me ic e olu ion AUC MOD in Failu e ype B in
TSM1 o all combina ion o HS ope a o alues. . . . . . 218
Figu e A.8 Me ic e olu ion AUC MOD in Failu e ype B in
in TSM2 o all combina ion o HS ope a o alues. . . . . 219
Figu e A.9 Me ic e olu ion AUC MOD in Failu e ype B in
in TSM3 o all combina ion o HS ope a o alues. . . . . 219
Figu e A.10 Me ic e olu ion AUC MOD in Failu e ype B
in in TSM4 o all combina ion o HS ope a o alues. . . 220
x i
Figu e A.11 Me ic e olu ion AUC MOD in Failu e ype B
in TSM5 o all combina ion o HS ope a o alues. . . . . 220
Figu e A.12 Me ic e olu ion AUC MOD in Failu e ype B
in TSM6 o all combina ion o HS ope a o alues. . . . . 221
Figu e A.13 Me ic e olu ion AUC MOD in Failu e ype C
in TSM1 o all combina ion o HS ope a o alues. . . . . 221
Figu e A.14 Me ic e olu ion AUC MOD in Failu e ype C
in TSM2 o all combina ion o HS ope a o alues. . . . . 222
Figu e A.15 Me ic e olu ion AUC MOD in Failu e ype C
in TSM3 o all combina ion o HS ope a o alues. . . . . 222
Figu e A.16 Me ic e olu ion AUC MOD in Failu e ype C
in TSM4 o all combina ion o HS ope a o alues. . . . . 223
Figu e A.17 Me ic e olu ion AUC MOD in Failu e ype C
in TSM5 o all combina ion o HS ope a o alues. . . . . 223
Figu e A.18 Me ic e olu ion AUC MOD in Failu e ype C
in TSM6 o all combina ion o HS ope a o alues. . . . . 224
Figu e A.19 Me ic e olu ion AUC MOD in Failu e ype D
in TSM1 o all combina ion o HS ope a o alues. . . . . 224
Figu e A.20 Me ic e olu ion AUC MOD in Failu e ype D
in TSM2 o all combina ion o HS ope a o alues. . . . . 225
Figu e A.21 Me ic e olu ion AUC MOD in Failu e ype D
in TSM3 o all combina ion o HS ope a o alues. . . . . 225
Figu e A.22 Me ic e olu ion AUC MOD in Failu e ype D
in TSM4 o all combina ion o HS ope a o alues. . . . . 226
Figu e A.23 Me ic e olu ion AUC MOD in Failu e ype D
in TSM5 o all combina ion o HS ope a o alues. . . . . 226
Figu e A.24 Me ic e olu ion AUC MOD in Failu e ype D
in TSM6 o all combina ion o HS ope a o alues. . . . . 227
Figu e A.25 Me ic e olu ion AUC MOD in Failu e ype E in
TSM1 o all combina ion o HS ope a o alues. . . . . . 227
Figu e A.26 Me ic e olu ion AUC MOD in Failu e ype E in
TSM2 o all combina ion o HS ope a o alues. . . . . . 228
Figu e A.27 Me ic e olu ion AUC MOD in Failu e ype E in
TSM3 o all combina ion o HS ope a o alues. . . . . . 228
x ii
Figu e A.28 Me ic e olu ion AUC MOD in Failu e ype E in
TSM4 o all combina ion o HS ope a o alues. . . . . . 229
Figu e A.29 Me ic e olu ion AUC MOD in Failu e ype E in
TSM5 o all combina ion o HS ope a o alues. . . . . . 229
Figu e A.30 Me ic e olu ion AUC MOD in Failu e ype E in
TSM6 o all combina ion o HS ope a o alues. . . . . . 230
Figu e A.31 Me ic e olu ion AUC MOD in Failu e ype F in
TSM1 o all combina ion o HS ope a o alues. . . . . . 230
Figu e A.32 Me ic e olu ion AUC MOD in Failu e ype F in
TSM2 o all combina ion o HS ope a o alues. . . . . . 231
Figu e A.33 Me ic e olu ion AUC MOD in Failu e ype F in
TSM3 o all combina ion o HS ope a o alues. . . . . . 231
Figu e A.34 Me ic e olu ion AUC MOD in Failu e ype F in
TSM4 o all combina ion o HS ope a o alues. . . . . . 232
Figu e A.35 Me ic e olu ion AUC MOD in Failu e ype F in
TSM5 o all combina ion o HS ope a o alues. . . . . . 232
Figu e A.36 Me ic e olu ion AUC MOD in Failu e ype F in
TSM6 o all combina ion o HS ope a o alues. . . . . . 233
x iii
NOMENCLATURE
αP opo ion o labelled da a c
βP opo ion o known elemen s pe class
λNumbe o boo s apped subda ase s
τNumbe o i e a ions
τLS Numbe o i e a ions in local sea ch
AD Anomaly De ec ion
AL Ac i e Lea ning
AMOSA Mul i-objec i e Simula ed Annealing Algo i hm
AUC MOD A ea Unde Cu e ROC Modi ied
AUC ROC A ea Unde Cu e ROC
C1 En opy o class p opo ions
CB Cons ain -Based
CG Cons ain s Guidance
CM Condi ion and Main enance
CN Cons ain s
Con Con ibu ion
CPPS Cybe -Physical P oduc ion Sys ems
CPS Cybe Physical Sys em
CS Clus e ing Solu ion
CV C oss Valida ion
xix
Chap e 1. In oduc ion
o whe e i is supposed o be. Wi h hese esul s, i is clea ha no
only Indus y 5.0 is a om being implemen ed, bu also in some
cases, Indus y 4.0 is s ill an eme ging challenge o many companies.
Rega ding he ypes o da a sou ces, Figu e 1.3 shows ha he
main sou ce o da a con inues o be he adi ional ones, such as
his o ical p oduc ion eco ds (54.8%), wo k o de s o p oduc s and
p ocesses (50.7%), da a eco ded by ope a o s (48.6%) and machine
pa ame e s (46.6%).
Figu e 1.3: Types o da a sou ces in he Manu ac u e Indus y [2].
Howe e , in indus ial en i onmen s, i is equen ha he a ail-
able da abases ha e ce ain sho comings, such as being pa ially la-
belled. To ha e a ully labelled da abase is specially impo an o
acili a e he p edic ion o ailu es and he implemen a ion o p edic-
i e main enance ope a ions. I is c ucial o iden i y he wo k o de
o label ha co esponds o he machine’s ope a ion a any pa icula
ime which enables a comp ehensi e analysis wi hou dis up ing he
machine’s ope a ing condi ions.
Finally, i can be no iced in Figu e 1.4 ha he mos commonly
applied echniques o he analysis o he da a a e 1) end analysis
and 2) s a is ical analysis, while he use o Machine Lea ning (ML)
echniques lags behind.
S a is ical me hods a e p e e ed o e ML p ima ily because ML
4

1.1. Mo i a ion
Figu e 1.4: Type o analysis used in Manu ac u e Indus y [2].
models based on da a end o be black boxes, making hem challenging
o in e p e . This lack o in e p e abili y is a signi ican d awback o
use s, who may ha e limi ed expe ience wi h bo h ML echniques and
he in e p e a ion o ou comes.
In gene al, bo h la ge and SMS companies a e willing o ake ad-
an age o he bene i s ha Indus y 4.0 o e s o 1) imp o e he
quali y o hei p oduc s and cus ome sa is ac ion, and 2) inc ease
he e iciency o hei p oduc ion sys ems [9, 2]. Howe e , looking a
he deg ee o Indus y 4.0 implemen a ion, i can be obse ed ha
he di icul ies o changing he cul u e o adi ional manu ac u ing
companies is one o he exis ing p oblems. This c ea es 1) a lack
o quali ied ope a o s and domain expe s in echnologies ela ed o
Indus y 4.0, 2) lack o knowledge in asks ha a e use ul o he im-
p o emen o hei pe o mance, 3) dis us in he ecommenda ions
p o ided by ce ain echnologies (such as ML), and 4) low in e es
in in es ing capi al o ans o ming he indus y owa ds he new
echnological e a [3].
In his sense, he gene al mo i a ion o his Thesis is wo old.
Fi s ly, o help manu ac u ing companies o ake a s ep u he in o
he echnological e a, losing hei ea and dis us o solu ions based
on new echnologies and echniques, and making his s ep easie o
hem. This s ep mus be use ul, easy o de elop, unde s andable, wi h
some kind o us alue, cos -e ec i e and he e ec o bene i should
be achie ed qui e ea ly. In his way, Indus y would gain con idence
5
Chap e 1. In oduc ion
and mo i a ion in he p ocess.
Secondly, a key Indus y 4.0 ac ion o help manu ac u ing com-
panies in pa icula is o de elop p edic i e main enance asks. In
his way, machine down imes could be educed and mal unc ions ha
lead o b eakdowns could be p edic ed, he eby imp o ing p oduc i -
i y and inc easing p o i s. In his sense, he design and de elopmen
o p edic i e main enance solu ions based on ML echniques ha en-
able o iden i y anomaly pa e ns in p ocess a iables ( ime se ies)
ela ed o ailu es is o g ea in e es . Fu he mo e, such solu ions
should ha e he ollowing cha ac e is ics:
•Capable o adap o se e al use cases, i.e., be gene ic.
•Easy o use and unde s and by an inexpe ienced use .
•The esul s should be p o ided in an unde s andable way. This
can be achie ed h ough he use o a iables o pa ame e s ha
a e di ec ly ela ed o he ope a ion o he machine y o he
p ocess.
•Capabili y o wo k wi h da abases ha con ain ew in o ma-
ion o li le amoun o labelled samples del ing in o a semi-
supe ised en i onmen .
•P o ide he use wi h a us me ic in he sys em i sel and
he esul s i deli e s. In his way he use can es ima e how
eliable he ou pu is, alle ia ing unce ain y and making he
model mo e unde s andable.
1.2 Objec i es
A e analysing he cu en s a e o Indus y 4.0, he posi ion o
companies in his espec , he eal le el o in eg a ion o his echno-
logical e a and he e o e de ec ing he needs in he con ex o Indus y
6
1.2. Objec i es
4.0, i has been obse ed ha he asks ela ed o p edic i e main e-
nance, i.e. he p edic ion o ailu es and b eakdowns, a e impo an
and necessa y in Indus y 4.0. Taking all his in o accoun , he main
objec i e o his Thesis ocuses on p oposing a me hodology capable
o p edic ing ailu es in indus ial machine y using ML echniques in
he con ex o Indus y 4.0.
In o de o de elop a solu ion, i is necessa y i s o all o explo e
and analyse he mos app op ia e echniques, wi hou igno ing un-
de s andabili y, explainabili y and use - iendliness o po en ial use s
who a e un amilia wi h ML echnologies.
In o de o de elop his main objec i e, a numbe o pa icula
objec i es mus be me :
Obj.1 S udy ad anced da a analysis echniques ocused on he de ec-
ion o anomalies in Indus y 4.0 in o de o p edic ailu es in
he indus ial p ocesses.
Obj.2 Iden i y ele an ea u es o p ocess a iable ime se ies in he
con ex o Indus y 4.0 ha help he de ec ion o anomalies o
p edic a ailu e in a machine y sys em. Among o he s, dis ance
and simila i y me ics and hei in luence on ML echniques a e
o be s udied.
Obj.3 Ob ain a me hodological basis, which p o ides a) a aluable
Know-How on he analysis o p ocess a iable ime se ies (gi ing
a be e insigh in o he p ocess, and hus imp o ing i ) and b)
explainabili y and a use - iendly sys em o he end use .
Obj.4 Design and de elop a gene al me hodology ha can wo k wi h
bo h labelled and pa ially labelled da abases, i.e. designing a
solu ion capable o in e ing knowledge om he labelled pa
o he da abases o label unknown samples.
Obj.5 S udy e alua ion me ics in semi-supe ised and supe ised en-
i onmen s o he iden i ica ion o pa e ns in p ocess a iable
7
Chap e 1. In oduc ion
ime se ies.
Obj.6 S udy exis ing us wo hiness me ics and implemen hem in
he p oposed solu ion. This way he use can be p o ided wi h
an es ima ion abou how eliable he solu ion is.
1.3 S uc u e
This sec ion ou lines he s uc u e o his Thesis, which is di ided
in o 6 addi ional chap e s.
In Chap e 2, a li e a u e e iew abou he di e en aspec s ad-
d essed in he Thesis is done. This s a e o he a ocuses on Indus-
y 4.0 and he enabling echnologies as well as he necessa y ac ion
poin s, such as a i icial in elligence and p edic i e main enance in o -
de o imp o e he pe o mance o indus ial p ocesses. Wi h his in
mind, he main objec i e is o analyse and compa e solu ions in he
li e a u e ha use ML echniques o de ec anomalies in indus ial
en i onmen s.
Chap e 3 p esen s he gene al s a egy p oposed o he au o-
ma ic and use - iendly me hodology o p edic i e main enance in
he con ex o Indus y 4.0. The p oposed s a egy is mainly based
on a hype heu is ic inspi ed app oach suppo ed by me a-heu is ics,
in pa icula Ha mony Sea ch (HS). In his sense, he wo main con-
ibu ions o his Thesis p esen ed in Chap e 5 and Chap e 6,
espec i ely, will exploi his gene al s a egy.
In o de o demons a e he Thesis p oposal based on he a o e-
men ioned hype -heu is ic s a egy, Chap e 4 p esen s a eal indus-
ial case consis ing in a cold s amping p ess o bol o ming.
Chap e s 5, and 6 p esen he di e en con ibu ions made as pa
o he hype -heu is ic inspi ed me hodology. Chap e 5 p esen s a
8
1.3. S uc u e
hype -heu is ic solu ion o labelling pa ially labelled da abases o-
ge he wi h a us wo hiness me ic. Chap e 6 p oposes a hype -
heu is ic s a egy ha aims o p o ide he use wi h a se o p edic i e
pa ame e s o u u e b eakages and ailu es in sys ems by de ec ing
anomaly pa e ns in he associa ed p ocess a iable ime se ies. Fi-
nally, Chap e 7 concludes he Thesis by analysing he esul s, de-
e mining he main con ibu ions o his Thesis and ou lining u u e
esea ch di ec ions.
9

Chap e 1. In oduc ion
10
CHAPTER 2
STATE OF THE ART
“E en hough no hing changes, i I change, e e y hing
changes.”
— Hono ´e de Balzac
This sec ion con ex ualises he scope o he Thesis, does ex ensi e
esea ch on he ela ed li e a u e,desc ibes he li e a u e gap in which
he p esen Thesis is de eloped, and inally p esen s he con ibu ion
o he Thesis. Figu e 2.1 shows an ou line o he e iewed opics in
he s a e o he a as well as he ela ionship be ween hem.
As Figu e 2.1 shows, in Sec ion 2.1, basic de ini ions and knowl-
edge abou Indus y, i s e olu ion o e he yea s and he di e en in-
dus ial e olu ions ha exis a e in oduced. Specially o Indus y
4.0, he key poin s, a eas and ac ion poin s o indus ial enhance-
men a e desc ibed oge he wi h he associa ed enabling echnolo-
gies. In Sec ion 2.2 one o he imp o emen ac ions, named p edic i e
main enance, is in oduced. I will explain why i is necessa y and
he ypes ha exis , as well as he echniques o me hods and ech-
11
Chap e 2. S a e o he A
nologies needed o make p edic i e main enance a eali y in Indus y.
Speci ically, his s a e o he a ocuses on ML echniques applied o
he de ec ion o anomalies. In sec ion 2.3, ce ain cha ac e is ics o
he da a associa ed wi h anomalies a e conside ed, such as he ype
o anomalies o be de ec ed, he ype o da a and he deg ee o pe -
cen age o labelling. In pa icula , an ex ensi e analysis o collec i e
Anomaly De ec ion (AD) echniques common o Time Se ies (TS) in
indus ial en i onmen s is de eloped. Depending on he pe cen age
o labelling, his sec ion is di ided in o 2 subsec ions, gi ing ise o
1) an explo a ion and analysis o he exis ing echniques o he au-
olabelling o Pa ially Labelled Da abases (PLD)(Subsec ion 2.3.1),
and 2) an ex ensi e analysis o he echniques and solu ions p oposed
in he li e a u e o de elop a collec i e ype o AD in TS wi h ully
labelled da abases (Subsec ion 2.3.2).
12
Indus y
Key poin s A eas and
ac ions o
imp o emen
P edic i e
Main enance
Model-based Da a d i en-
based Hyb id models
Anomaly
De ec ion
P oposed
solu ions
P oposed
solu ions
SSL Classi ica ion
SSL Clus e ing
SSL Reg ession
Heu is ic
Me aheu is ic
Hipe -heu is ic
Le el o
Expe ise
Requi ed (LER)
Low
Medium
High
Fea u e
Enginee ing
App oach
Heu is ic
Me aheu is ic
Indus y e olu ion
Indus y 4.0Indus y 1.0 Indus y 2.0 Indus y 3.0
Enabling
Technologies
Diagnosis P ognosis
Faul de ec ion
and isola ion
Faul
iden i ica ion
Remaining
Use ul Li e
Machine
Lea ning (ML)
Faul
p edic ion
Me hods
Indus y 5.0
Real- ime
op imiza ion
Machine Lea nign asks
Classi ica ion Reg ession Clus e ing
1
2
Ini ial da a
conside a ions
Type o
anomaly
Pe cen age o
Labels
Type o da a
Lea ning
app oach
Supe ised
Labelled
da abases
Unsupe ised
Unlabelled
da abases
Semi-supe ised
Pa ially labelled
da abases
Time Se ies
Mul i a ia e
Poin anomaly
Con ex ual
anomaly
Collec i e
anomaly
3
S a egy
Me hodology
App oach
Type o
knowledge Cons ain s
Labels
Fil e
W appe
Dis ance-based
Cons ain -based
Hyb id
S a egy
Me hod o
pa ame e
es ima ion
2.1
2.2
2.3.1
2.3.2
2.2
Hipe -heu is ic
25%
LABELS
1
100%
LABELS
2
2.3
Figu e 2.1: S a e o he a scheme.
13
Chap e 2. S a e o he A
in manu ac u ing p ocesses. One ask o imp o emen is he use
o obo s in collabo a ion wi h humans. In his way, asks ha
a e hea y o use s and can cause inju ies will be pe o med
by obo s. The e o e, he digi aliza ion o knowledge and asks
will make i easie o use s o manage and moni o p oduc ion.
The da a p o ided by companies ha ha e al eady aken hese
measu es is an inc ease o 45-55% in p oduc i i y [5].
4. Nowadays, in he in en o ies he e is an excess o bo h pu -
chased ma e ials and hose p oduced by he manu ac u ing com-
pany i sel . I is he e o e necessa y o ca y ou s ock moni-
o ing ac ions o p e en he company om ha ing a la ge cap-
i al s ock. These imp o emen ac ions a e no only ocused on
he exac accoun ing o s ock, bu also on he adequacy and
planning o he s ock necessa y o p oduc ion, hus elimina ing
excesses, he a angemen o s ocks in he wa ehouses o use as
li le space as possible o o e p oduc ion. Th ough asks such as
eal- ime supply chain op imisa ion, Indus y 4.0 can ypically
educe in en o y holding cos s by 20-50% [5].
5. I is indispu able ha quali y imp o emen is an impo an
esea ch a ea o he indus y, no only a he p oduc le el
bu also because o he addi ional cos s (in ime, ma e ials and
labou ) in ol ed in he ep ocessing o was e. These quali y
p oblems a e o en caused by uns able p ocesses, poo packag-
ing and mal unc ioning o b oken machine y. This can be sol ed
by eal- ime moni o ing and da a analysis ools ha de ec ab-
no mal beha iou and ale he wo ke o s op o modi y he
indus ial p ocess in ime. By applying such le e s, cos s e-
la ed o subop imal quali y can be educed by 10 o 20 % [5].
6. To achie e an op imal ma ch be ween supply and demand
i is necessa y o he indus y o unde s and he demand bo h
in e ms o quan i y and cha ac e is ics o he desi ed p oduc .
Ad anced and adequa e demand analysis based on da a models
makes i possible o inc ease he accu acy o demand by 85% pe
20

2.1. Indus y 4.0
week and o unde s and he cha ac e is ics mos sough a e by
cus ome s [5].
7. In e ms o ime o ma ke , pionee ing a p oduc b ings ex a
bene i s and also makes i easie o espond ea lie o po en ial
p oblems. Enabling ac ions o imp o e his a ea include con-
cu en enginee ing o apid expe imen a ion/p o o yping (e.g.
h ough 3D p in ing). This can educe ime o ma ke by 30-
50% [5].
8. Nowadays, o e ing a good a e -sales se ice and emo e main-
enance is a key elemen wi h g ea po en ial in he indus y.
He e, one o he mos demanded ac ions is emo e main enance.
These a e so wa e solu ions ha allow echnicians o es ablish
a secu e emo e connec ion o indus ial equipmen o ca y ou
a diagnosis wi hou he need o isi he si e. A educ ion in
main enance cos s o be ween 10-40% has been obse ed hanks
o emo e and p edic i e main enance using da a analy ics ech-
niques [5].
21
Chap e 2. S a e o he A
Se ice/
a e sales
Resou ce/
p ocess
Asse
u iliza ion
Labo
In en o ies
Quali y
Supply/
demand
ma ch
Vi ually guided
sel -se ice
Real- ime
yield op imiza ion
Sma ene gy
consump ion
Remo e moni o ing
and con ol
P edic i e
main enance
Human- obo
collabo a ion
Remo e moni o ing
and con ol
Digi al
managemen
S ock managemen
Real- ime supply
chain op imiza ion
Digi al quali y
managemen
Da a-d i en
demand p edic ion
Indus y 4.0
imp o emen
ac ions o he 8
main a eas
Time o
ma ke
Concu en
enginee ing
Rapid simula ion
and expe imen a ion
10 - 40%
educ ion o
main enance
cos s
P oduc i i y
inc ease
by 3 - 5%
20 - 50%
educ ion
in ime o
ma ke
Fo ecas ing
accu acy inc eased
o 85+%
Cos s o quali y
educed by 10 - 20% Cos s o
in en o y
holding
dec eased
by 20 - 50%
45 - 55% inc ease o
p oduc i i y in echnical
p o essions h ough
au oma ion o
knowledge wo k
30 - 50% educ ion
o o al machine
down ime
P edic i e
Main enance
Remo e
Main enance
Sma esou ces
consump ion
In elligen
IoTs
Real- ime
yield op imiza ion
Au oma ion o
knowledge wo k
Real- ime da a
analisys
Da a-d i en
demand p edic ion
Da a-d i en
demand p edic ion
Da a-d i en
p oduc design
Se ice/
a e sales
Time o
ma ke
Supply/
demand
ma ch
Quali y
Remo e
Main enance
P edic i e
Main enance
Vi ually guided
sel -se ice
Concu en
enginee ing
Rapid simula ion
and expe imen a ion
Da a-d i en
demand p edic ion
Op imiza ion o
wa ehouse space
Legend: Enabling Indus y 4.0 echnologies
IoT
ML
Cloud
CPS
Cobo sRTO
AR
Big
Da a
AM
Figu e 2.4: Imp o emen ac ions o he 8 main a eas o in e es and i s co esponding enabling echnologies in Indus y
4.0. Sou ce :[5] e-illus a ed and modi ied o his Thesis. CPS: Cybe Physical Sys em, RTO: Real- ime op imiza ion,
ML: ML, AR: Augmen ed Reali y, IoT: In e ne o Things, AM: Addi i e manu ac u ing.
22
2.2. P edic i e main enance
As can be seen, hese imp o emen ac ions in he Indus y 4.0
ha e a di ec e ec on 1) enhancing he alue o Indus y, 2) imp o -
ing p oduc ion and p oduc quali y and 3) inc easing he p o i s o
companies. As e e enced abo e, p edic i e main enance is one o he
key imp o emen ac ions ha is di ec ly co ela ed wi h d i ing he
co ec and op imal use o machine y. This is o c ucial impo ance
o companies, especially hose wi h hea y and expensi e machine y.
In o de o pe o m a good p edic i e main enance i is necessa y o
ha e a good his o y o da a cap u ed by di e en senso s and in elli-
gen de ices (IoT and CPS) placed in he di e en key poin s o he
machines. A e ob aining he da a and s o ing hem, i is usually
necessa y o apply in elligen analysis echniques (ML) ha c ea e
models capable o iden i ying pa e ns and de ec ing anomalies o ei-
he o esee undesi ed si ua ions, such as b eakages, u u e s oppages
o wea in any elemen o he machine. Wi h his analysis i is also
possible o pe o m op imisa ion in eal ime ( eal- ime op imisa ion
and cloud compu ing) and he e o e, imp o e p oduc i i y.
2.2 P edic i e main enance
Main enance echniques, oge he wi h Indus y 4.0, ha e de el-
oped new skills and imp o emen s o e he yea s. The ea lies main e-
nance echnique is basically co ec i e main enance, which akes place
only a e a ailu e has occu ed. A la e main enance echnique is p e-
en i e main enance (also called planned main enance), which se s a
pe iodic in e al o pe o m p e en i e main enance ega dless o he
heal h s a us o a physical asse . Wi h he apid de elopmen o mod-
e n echnology, p oduc s a e becoming mo e complex and highe qual-
i y and eliabili y a e demanded, which inc ease he cos o p e en i e
main enance. The e o e, mo e e icien main enance app oaches, such
as p edic i e main enance [19], a e needed. P edic i e main enance
is based on he con inuous moni o ing o a machine o a p ocess,
allowing main enance o be pe o med only when i is needed. I al-
23
Chap e 2. S a e o he A
lows he ea ly de ec ion and p edic ion o ailu es hanks o p edic i e
ools based on his o ical da a (e.g. ML), s a is ical in e ence me h-
ods and enginee ing app oaches [20]. As men ioned be o e, p edic i e
main enance b ings di e en bene i s o p oduc ion en i onmen s like
p oduc i i y imp o emen , educ ion o sys em ailu es, minimisa ion
o unscheduled machine y down imes, inc eased e iciency in he use o
inancial and human esou ces, and scheduling op imisa ion o main-
enance in e en ions [21].
2.2.1 P edic i e main enance ca ego ies
P edic i e main enance ocuses on wo main pa s: aul diagnosis
and aul p ognosis. Diagnosis in ol es all hose asks ha ocus on
he de ec ion, isola ion and iden i ica ion o aul s in a machine o
indus ial p ocess. Thus, he s eps o be ollowed in diagnosis [19] a e
1) de ec he aul s by de e mining when some hing in he moni o ed
sys em is w ong, 2) isola e he aul s by de e mining which elemen
is a aul and 3) iden i y he aul by de e mining he na u al cause
o ailu e.
A e he diagnos ic p ocess, he use usually has an idea o wha
aul s a e occu ing and why. Wi h his in o ma ion, p ognos ic asks
can be pe o med on he sys ems, which e e o he ask o p edic -
ing ha a ailu e will occu in he sys em. Two main asks can be
obse ed a his poin , ailu e p edic ion and Remaining Use ul Li e
(RUL) es ima ion. Faul o ailu e p edic ion de e mines ha a ail-
u e is imminen and es ima es when i may occu , i.e. he Time To
Failu e (TTF). The la e ask, RUL es ima ion, assesses how long a
machining componen has un il i canno unc ion anymo e in acco -
dance wi h i s in ended pu pose, gi en he cu en machine age and
condi ion, and he pas ope a ion p o ile [19]. P ognosis is said o be
much mo e e icien han diagnosis o ewe down imes in indus ial
p ocesses [19], howe e i can some imes be mo e unce ain. In his
espec , in he scope o he p esen Thesis i is conside ed o g ea
24
2.2. P edic i e main enance
in e es o p o ide he use wi h a con idence alue abou how much
he use can ely on such p edic ions.
2.2.2 Me hods o p edic i e main enance de elopmen
The way in which p edic i e main enance asks, bo h diagnosis
and p ognosis, a e pe o med can be done in di e en ways depend-
ing on 1) he ype o in o ma ion a ailable, i.e. whe he he e is
da a eco ded o e ime (senso da a), in o ma ion on p e ious ma-
chine condi ions and main enance (CM) [22] o he ope a ing condi-
ions/wo k o de (WO) ha a pa icula machine is unning, i.e., o
know he ope a ing pa ame e s o a machine [22, 23], and 2) he de-
g ee o expe knowledge equi ed o a ailable. Ha ing his in mind,
and as can be seen in Figu e 2.5, h ee di e en es ima ion me hods
can be chosen [19, 20, 22, 24, 25].
•Whi e-box o model-based model. This me hod equi es
a lo o expe knowledge as i is based on he physical laws
go e ning he analysed elemen s. Knowing he condi ions o he
machine is essen ial, bu ha ing expe imen al senso ised da a is
no necessa ily equi ed. In his case he esul is usually con-
cise and unde s andable o expe s bu i is ime-consuming,
equi es a lo o p io knowledge and aluable in o ma ion e-
la ed o he machine condi ions and main enance.
•Black box o da a-d i en model. This me hodology e-
qui es mainly a good his o ical senso da a [22] while da a om
condi ion main enance is no needed. This ype o model is
mainly cha ac e ised by he ac ha i does no equi e expe
knowledge and is based on ma hema ical models such as ML o
s a is ical algo i hms. They a e called black box models because
hey a e commonly di icul o unde s and o domain expe s,
as knowledge is ex ac ed om he da a and no om physical
laws.
25

Chap e 2. S a e o he A
•G ey box o hyb id models. These models a e a mix u e o
he wo p e ious ones, ying o use he bene i s o each one.
Wi h he ad en o Indus y 4.0 and he ease o massi e da a cap-
u e, da a-d i en models a e o g ea in e es o hei applica ion o
p edic i e main enance, especially when explici ela ions be ween he
ailu e and he ela ed ea u es a e unknown. I p ope ly managed,
his can esul in a as and eliable way o he p edic ion o ail-
u es and he op imized managemen o main enance asks, ul illing
all abo e men ioned ad an ages abou a good p edic i e main enance
in Indus y 4.0 and e en in he al eady imminen Indus y 5.0. Speci -
ically, his Thesis aims o design and de elop a new me hodology o
indus ial aul p edic ion, as de ailed la e .
WO
Low
Physics
modeling Model Type
AI
app oach
S a is ical
app oach
Da a d i en
me hods
Senso da a
Physical
app oach
Model Based
me hods
CM da a
AI, s a is ical &
physical app oaches
Hyb id
me hods
CM da a
Senso
da a WO
WO
High
Expe
knowledge
le el
Da a
modeling
Figu e 2.5: Ca ego iza ion and de ini ion o p ognosis me hods.
A e e alua ing 1) he impo ance o p edic i e main enance in
imp o ing p oduc i i y in companies wi hin he con ex o Indus y
4.0, 2) he ca ego ies o p edic i e main enance iden i ied in he li -
e a u e and 3) he me hodologies used o i s de elopmen , i can be
concluded ha pe o ming an adequa e and e icien p ognosis is o
i al impo ance o he Indus y. Fu he mo e, gi en he immense
26
2.3. Anomaly de ec ion o aul p edic ion
amoun o da a a ailable nowadays, i makes sense o use me hods
based on his o ical da a oge he wi h ML enabling echnologies o
de elop a co ec ailu e p edic ion sys em.
2.3 Anomaly de ec ion o aul p edic ion
Wi hin ML, AD can be applied o analyse signal pa e ns and
iden i y unexpec ed beha iou s [6] ha can p edic a subsequen ail-
u e. When de eloping ML algo i hms o AD, i is necessa y o pay
a en ion o se e al concep s, such as: he ype o da a, he ype o
anomaly and he a ailable labels.
The i s key aspec o he p ope selec ion o AD me hod is
he na u e o he da a. Each da a desc ibing an ins an in ime is
composed o one a ibu e o a iable (uni a ia e) o a se o a ibu es
(mul i a ia e). Rega ding mul i a ia e da ase s, hese may be o he
same da a ype o a mix u e o se e al ypes, e.g. ca ego ical o
con inuous. Finally, he e is he possibili y ha he a iables may
o may no be ela ed o each o he . Speci ically in Indus y, he
mos commonly used da ase s con ain a) indi idual mul i a ia e da a,
i.e., da a o di e en bu un ela ed a iables (he ea e e e ed o as
mul i a ia e da a) o b) TS, i.e. da a ela ed o one o mo e a iables
moni o ed o e a pe iod o ime [6, 19].
In e ms o he anomaly ype ha can occu and he e o e hose
o which he AD algo i hm is o be designed, h ee ypes o anomalies
a e dis inguished [6]:
•Abno mal poin s o ou lie s: hese a e andom poin s ha a e
ou o he pa e n o he es o he poin s. They usually appea
in a d as ic way and can be iden i ied wi h simple algo i hms,
such as he implemen a ion o ules o limi s based on he s a-
is ical dis ibu ion o he da a. Figu e 2.6 shows how he e a e
27
Chap e 2. S a e o he A
ce ain poin s such as O1and O2 ha a e ou side he es o he
majo i y g oups [6].
Figu e 2.6: Ou lie anomaly example [6].
•Con ex ual anomalies: hese a e poin s o ins ances ha may
a ise along a TS which, depending on he en i onmen o con ex
in which hey a e loca ed, a e called anomalies wi h espec o
hei su oundings. O e he las ew yea s, hese ha e been
he mos s udied by esea che s in he ield. As can be seen in
Figu e 2.7, whe e he e olu ion o he empe a u e o e a yea is
shown, he e a e wo speci ic momen s in ime whe e he alue
o he empe a u e is he same ( 1and 2). The g ea di e ence
be ween bo h is he con ex in which hey a e, since o he
momen 1, i is an expec ed o no mal alue. Fo 2i s alue
lea es he con ex o he en i onmen in which i is [6].
Figu e 2.7: Con ex ual anomalies example [6].
28
2.3. Anomaly de ec ion o aul p edic ion
•Abno mal pa e ns o collec i e anomalies: hese a e o med
when a se o small agmen s o he TS does no i he o he
pa e ns in he da a se . As can be seen in Figu e 2.8, he e
is an a ea (ma ked in ed) in he TS whe e he pa e n di e s
om he es o he se ies [6]. This ype o anomaly is he mos
commonly ound in he indus ial en i onmen s.
Figu e 2.8: Collec i e anomalies example [6].
Finally, he hi d key poin o selec ing an AD s a egy is ela ed
o he da a labels. I is especially impo an in he indus ial en-
i onmen o p edic i e main enance o know he labels o he da a
because, a bad o missing label can lead o an e oneous o inaccu-
a e solu ion. Howe e , i is wo h no ing ha his is usually a e y
cos ly p ocess o domain expe s o do, no only in e ms o quali y
bu also in e ms o ime. Addi ionally, in AD commonly he e is
a la ge imbalance in he da abases, wi h usually mo e no mal ags
han anomalous cases. Based on he ex en o which he labels a e
a ailable, AD echniques can ope a e in one o he ollowing h ee
modes:
•Supe ised lea ning. Techniques ained in supe ised mode
assume ha he da abase is comple ely labelled o no mal be-
ha iou as well as o he anomaly class. The ypical app oach
in his case is o build a p edic ion model o he no mal o
29
Chap e 2. S a e o he A
Finally, hype -heu is ics allow o he handling o la ge, compu-
a ionally complex sea ch p oblems ins ead o only sol ing single o
lowe complexi y p oblems. In ac , hype -heu is ics p o ide a supe-
io , au oma ed al e na i e o me a-heu is ics ha ypically equi e
ine- uning o ela ed pa ame e s o op imal pe o mance on a speci ic
p oblem [51]. Hype -heu is ics au oma e a combina ion o heu is ics
and/o me a-heu is ics as ano he sea ch p ocess o inding he bes
pa ame e s o con ol heu is ic o me a-heu is ic beha io , wi h he
o e all objec i e o balancing explo a i e and exploi a i e sea ch o
achie e a solu ion.
Acco dingly, hype -heu is ics based app oaches a e a e based on
hype -heu is ic algo i hms [52]and a e he less e e ed in he li -
e a u e. These a e also unde s ood as “a combined me a-heu is ic
sea ch me hod ha h ough he au oma iza ion o i s s ochas ic pa-
ame e s, such as: he selec ion, gene a ion, combina ion o adap a-
ion o se e al componen s o e icien ly sol e compu a ional NP-ha d
sea ch p oblems [53, 54]”. In he li e a u e di e en de ini ions o his
ype o algo i hms can be ound bu mainly a e de ined as “a high-
le el heu is ic app oach ha , gi en a pa icula p oblem and a se o
low-le el heu is ics, i s main objec i e is o ind he bes heu is ic o
sequence o heu is ics o sol e he p oblem a he han o p o ide he
solu ion o he p oblem di ec ly [55, 54]”.
Simila o o he domains, au oma iza ion is a key challenge o
p omo e he pene a ion o a gi en echnology. While heu is ics and
me a-heu is ics equi e human in e en ion o op imize hei ee pa-
ame e s o ob aining he bes balance in hei sea ch capabili ies,
hype -heu is ics p o ide an au oma ed solu ion o his challenge.
As hype -heu is ics a e gaining p ominence due o i s p ope ies,
di e en axonomies a e p oposed in he li e a u e which a e di e -
en ia ed in Figu e 2.11. The axonomy p oposed in [55] di e en ia es
be ween wo domains 1) he na u e o he heu is ics o use and 2)
he sou ce o knowledge o eedback in he p ocess. Rega ding he
36

2.3. Anomaly de ec ion o aul p edic ion
Na u e o he
space sea ch
Hype -heu is ics
Feedback Techniques
O line lea ning
Online Lea ning
No-lea ning
Gene a ion
Selec ion
Cons uc ion
Pe u ba ion
Cons uc ion
Pe u ba ion
G eedy and
peckish
Random
selec ion
Me a-heu is ic
based
By lea ning
mechanisms
Figu e 2.11: Taxonomy o hype -heu is ics.
na u e o he heu is ics and how hey a e implemen ed in he hype -
heu is ic wo ca ego ies can be dis inguished [56]: a) selec ion which
a e me hodologies ha sea ch he bes con igu a ion o gi en heu is-
ics o b) gene a ion which, in con as , a e me hodologies ha gen-
e a e hype -heu is ics om a se o gi en heu is ics. Wi hin his clas-
si ica ion, wo ypes o low-le el me hodologies can be dis inguished:
1) cons uc i e heu is ics, which g adually c ea es a sui able inal so-
lu ion om s a o inish. The goal is o in elligen ly choose he mos
app op ia e heu is ic o he cu en p oblem s a e. The p ocess con-
inues un il a comple e solu ion is achie ed, and since he p oblem has
a ixed size, he e is a na u al endpoin o he cons uc ion p ocess,
o 2) pe u ba i e o local sea ch heu is ics, which de elop a pa ial
modi ica ion o a heu is ic solu ion p e iously c ea ed. The aim is
o i e a i ely choose and selec he heu is ics based on he cu en
solu ion ha is al eady comple e.
Del ing in o he knowledge used as eedback om he sea ch p o-
cess, i can be ind h ee main ca ego ies o hype -heu is ics 1) online
lea ning, whe e he lea ning p ocess occu s du ing he esolu ion o a
gi en sea ch p oblem, while he hype -heu is ic in e ac s wi h a de-
37
Chap e 2. S a e o he A
ined en i onmen , such as u ilizing ein o cemen lea ning o heu is-
ic selec ion and employing me a-heu is ics as high-le el sea ch s a e-
gies o e a heu is ic sea ch space. As a esul , he high-le el s a egy
can le e age ask-speci ic local p ope ies o iden i y he mos sui able
low-le el heu is ic o apply, 2) o line lea ning, in his case he idea is
o accumula e knowledge, o en in he o m o ules o p og ams. The
main idea is o use a gi en se o aining ins ances o lea n and hen,
be gene alized o sol e unseen ins ances. O line lea ning echniques
allow he algo i hm o lea n om p e ious p oblem-sol ing expe i-
ences, esul ing in a mo e e icien and e ec i e me hodology. Some
examples o o line lea ning app oaches using hype -heu is ics include
case-based easoning, and gene ic p og amming. By le e aging hese
echniques, he algo i hm can iden i y pa e ns and apply p e iously
lea ned knowledge o imp o e i s p oblem-sol ing abili ies, whe e one
lea ns om one aining se o apply i o ano he , and 3) no-lea ning
e e s o hose me hodologies whe e he e is no eedback du ing he
sea ch p ocess.
Rega ding he echniques used o de elop hype -heu is ics, ou
me hods can be dis inguished [57]: 1) Hype -heu is ic based on an-
dom selec ion, which is he mos basic and s aigh o wa d o he
hype -heu is ic amily, as i andomly selec s a low-le el heu is ic
om a gi en se a each decision poin wi hou conside ing pas pe -
o mance, 2) G eedy and peckish hype -heu is ic, which selec s and
applies he low-le el heu is ic ha p oduces he la ges imp o emen
o he cu en objec i e alue a each decision poin . I no imp o ing
low-le el heu is ic is a ailable, he heu is ic chooses he one ha leads
o he smalles de e io a ion. G eedy hype -heu is ic equi es a p e-
limina y e alua ion o each low-le el heu is ic in he se o selec he
bes one, which makes i a mo e compu a ional complex app oach,
3) Me a-heu is ic based hype -heu is ic, whe ein a me a-heu is ic is
he me hod o global sea ch ha ope a es in he solu ion space o a
p oblem and employs s a egies o escape local op ima and 4) Hype -
heu is ic by lea ning mechanisms, ha uses a ious echniques o lea n
he pe o mance his o y o low-le el heu is ics. A each decision poin ,
38
2.3. Anomaly de ec ion o aul p edic ion
a hype -heu is ic chooses a p omising low-le el heu is ic based on he
e ec i eness o each one ga he ed om ea lie s ages o p e ious uns.
Once he p oposed axonomy in Figu e 2.11 is p esen ed, he e-
la ed s a e o he a can be analysed acco dingly. Table 2.1 sum-
ma ises he classi ica ion and compa ison done o he li e a u e e e -
ences.
Table 2.1: Compa ison o he li e a u e. TS: Time Se ies, MV: Mul-
i a ia e, L: Labels, CN: Cons ain s, F: Fil e , W: W appe , CB:
Cons ain -Based, DB: Dis ance-me ic Based, CG: Cons ain s Guid-
ance, S: Seeding, OF: Op imiza ion Func ion, H: Heu is ics, MH:
Me a-heu is ics, HH: Hype -Heu is ics.
Re e ence Da a Knowledge S a egy Me hodology App oach
[41, 58, 59]
MV
L
W CB (S) H
[42, 60] W CB (CG) H
[38] W DB MH
[61] W CB (OF) H
[62] F CB (OF) H
[50, 63, 64, 65] F CB (OF) MH
[39]
CN
W Hyb (DB + OF) H
[66, 67] W CB (CG) H
[68] W CB (CG) H ensemble
[69, 43, 70, 71] W DB H
[72, 73] F CB (OF) H
[37] F CB (OF) H ensemble
[74, 75]
TS
LW CB(CG) H
[48] F CB (OF) MH
[76, 77]
CN
W DB H
[35] W CB (CG) H
[78] F CB (OF) H
Mul i a ia e da a
In he con ex o MV da a, i can be obse ed ha he e a e
solu ions ha use labels and cons ain s. In bo h ypes o knowledge,
w appe s a egies a e mo e common and ha e mo e di e si y in he
me hodologies han il e ones. Fil e s a egies a e mainly cons ain -
39
Chap e 2. S a e o he A
based by means o an objec i e unc ion.
In ega ds o he solu ions ha use labels and w appe s a egies,
bo h heu is ic and me a-heu is ic algo i hms a e used. Conce ning
heu is ic algo i hms, labels a e in oduced as a) ini ial seeds o a K-
means [41, 58, 59] wi h p ede ined pa ame e s, b) a guiding me hod
in he o ma ion o clus e s in DBSCAN [42, 60] o c) a e alua ing
me hod o membe ship ag eemen in Fuzzy-C-Means [61]. In con-
as , he use o me a-heu is ics is de eloped by using a dis ance-based
me hodology and SA as he main algo i hm [38].
Rega ding he solu ions ha use labels and il e s a egies, i is
obse ed ha me a-heu is ics a e commonly employed. These op i-
mise an objec i e unc ion ei he using GA [64, 63] o a mul i-objec i e
op imisa ion wi h AMOSA [65, 50] o di e en e alua ion indexes wi h
p ede ined pa ame e s. Finally, wi h ega ds o il e s a egy and
heu is ics ew e e ences use an objec i e unc ion as in [62] which
employs a uzzy algo i hm.
Del ing in o solu ions using cons ain s, i can be obse ed ha no
e e ence employing me a-heu is ics is encoun e ed. Howe e , he use
o heu is ic ensemble me hods has p oli e a ed in his ield [68, 37].
Speci ically, o he solu ions using cons ain s oge he wi h a
w appe s a egy, hyb id [39] and dis ance-based [69, 43, 70, 71] me hod-
ologies can be ound. These solu ions a e complex and p oblem-
dependen whe ein he lea ning me ic o he clus e ing i sel can be
modi ied easily as in hie a chical [69, 43], EM [39, 70] o K-means
[71] clus e ing algo i hms. Addi ionally, he e a e solu ions ha use
cons ain s o guide clus e o ma ion and imp o e he quali y o he
solu ion by means o bo h heu is ics (K-means [67], DBSCAN [66])
and ensemble me hods [68].
Addi ionally, using his ype o knowledge (cons ain s) and il-
e s a egies, he solu ions a e based on cons ain -based me hods
40
2.3. Anomaly de ec ion o aul p edic ion
employing objec i e unc ions using heu is ic algo i hms (K-means)
and AL laye [72, 73] o an ensemble s uc u e [37] o hie a chical
clus e ing me hods.
Time se ies da a
Th oughou he li e a u e e iew, i can be obse ed ha he so-
lu ions p oposed o TS da a a e sca ce due o hei complexi y bo h
in e ms o size and noise [76, 35]. The e a e solu ions ha join ly
use semi-supe ised clus e ing p io o de elop a classi ica ion sys em
[79, 26].
In [74, 75], i can be ound a solu ion based on w appe s a egy
oge he wi h heu is ics (G aph based clus e ing) guided by he use
o labels. By con as , au ho s in [48] combine wo me a-heu is ics
(GA and SA) using a il e s a egy, hus le e aging he s eng hs o
bo h and ob aining p omising esul s in he con ex o aul diagnosis
o bea ings.
I knowledge is in oduced in he o m o cons ain s, au ho s usu-
ally de elop no an holis ic solu ion bu a pa ial one adap able o
di e en solu ions. In his espec , i is common o use il e s a e-
gies applied oge he wi h an AL echnique [78] o w appe s a egies
using dis ance-based me hods [76]. Finally, some comple e solu ions
based on w appe s a egies and heu is ics can be ound. In [35] au-
ho s use he use -supplied cons ains as guidance (AL) o wo ypes
o clus e ing, while [77] uses a complex me ic lea ning in eg a ing he
dis ance equi emen s acco ding o he cons ain s o he elemen s.
Discussion
As his li e a u e e iew has shown, he e a e a wide a ie y o e -
e ences using SSL clus e ing based solu ions o labelling PLD. In gen-
41

Chap e 2. S a e o he A
e al, i can be obse ed ha cons ain -based me hodologies a e he
mos commonly used, as well as solu ions ha use a speci ic heu is ic.
Many o he p oposed solu ions a e ich in knowledge and echniques,
anging om he simples heu is ic algo i hms o he u ilisa ion o
a) complex in eg a ed lea ning me ics o e alua e cons ain s, b) en-
semble algo i hms o c) me a-heu is ic sys ems wi h in e nal pa am-
e e iza ion and op imiza ion. Howe e , i has no been ound ye a
p oposal capable o : a) in eg a ing a me hod applicable o bo h ypes
o da a (MV, TS), b) being comple ely au onomous in e ms o op i-
miza ion o in e nal pa ame e s o he algo i hms and o he me hod
i sel o he co ec labelling o semi-supe ised da abases, c) being
use - iendly and d) p o iding a us wo hiness me ic o imp o ing
he use con idence on he sys em. The la e is e y impo an since
in semi-supe ised sys ems he e is no ce ain y abou he accu acy
o he solu ion, specially due o he lack o labelled samples.
2.3.2 Collec i e anomaly de ec ion solu ions in ime se-
ies da a in he con ex o Indus y 4.0
Th oughou his subsec ion a desc ip ion and analysis o he cha -
ac e is ics o he anomalies ha can be ound in indus ial en i on-
men s and he pa ame e s o he signals ha can be p edic o s o
ailu es in indus ial p ocesses o machine y sys ems a e p esen ed.
In his way and oge he wi h AD echniques based on ML, p edic i e
main enance sys ems can be de eloped.
As men ioned be o e, many indus ial p ocesses su e om a deg a-
da ion o he no mal beha iou ha concludes on a ailu e o ma-
chine y, ool o p ocess. This deg ada ion o he no mal beha iou
is unde s ood as an anomaly in he p ocess, ha is eco ded in he
p ocess signals (TS) as collec i e anomaly. Ac ually, he common
AD p ocedu e s a s om gene a ing an expe imen al da abase wi h
p eselec ed a iables associa ed o N ailu e cases. A e wa ds, he
42
2.3. Anomaly de ec ion o aul p edic ion
expe imen al da abase is analysed and he abno mal pa e ns ha
enable o p edic he ailu es a e usually iden i ied based on heu is ic
ules. Finally, an algo i hm is designed and de eloped o implemen
a eal- ime de ec ion o such anomaly pa e ns [80, 81, 82, 83, 84].
Howe e , hese echniques a e cos ly and edious as hey a e cen ed
on he speci ic use case.
The de ec ion o his deg ada ion is key o p edic he ailu e and
in his Thesis i is unde s ood as he pe iod o ime o ime-window
immedia ely p io o he ailu e whe e he beha iou s ops ollowing
he no mal end. This ime-window depends on he dynamics and
cha ac e is ics o he ype o ailu e in a pa icula p ocess and may
be unknown be o ehand. Likewise, i is also common o igno e o he
impo an pa ame e s as:
(a) The pa icula p ocess a iable o a iables ha enable he p e-
dic ion o he ailu e among all he Mp e-selec ed p ocess a i-
ables, he eina e e e ed o as a Time Se ies Measu emen
(TSM ).
(b) The gene al s a is ic o ex ac ed ea u e (Fe) o he p ocess
a iables wi hin he deg ada ion ime-window size ha enable
he p edic ion o he ailu e.
(c) The size o he ime-window (Wz) o pe o m he ea u e ex-
ac ion.
(d) The h eshold alue (Th) ha he s a is ic ea u e alue should
exceed o no o be ca ego ized as a ailu e.
An example o he abo e-men ioned pa ame e s can be seen in
Figu e 2.12. Fi s ly, he e is a p ocess a iable o TSM eco ded un-
il a b eakage o an elemen occu s. In his case, he “maximum”
Fe is ex ac ed h ough he ime-windows o size Wz gi ing ise o
he ex ac ed FeW z(TSM1) alues. I is obse ed ha wi h he es i-
43
Chap e 2. S a e o he A
Wz
FeWz (TSM1)
Th
TSM1
Failu e
Figu e 2.12: Rele an pa ame e s in he AD con ex .
ma ed Th, based in some ules, in he deg ada ion window Wz, he
anomalous poin s ha p edic he u u e b eakage can be de ec ed.
Rega ding he de ec ion and diagnosis o aul s in indus y, a
mul i ude o algo i hms and me hodologies ela ed o collec i e da a-
d i en AD ha e been de eloped o e he yea s.
Among he solu ions ha use da a-d i en me hods, wo main ap-
p oaches can be dis inguished: a) he ule based heu is ic, ha p o-
ides he use wi h a single possible solu ion [85] wi h a low com-
pu a ional cos a he expense o a high le el pa ame e isa ion and
expe knowledge [86], and b) he me a-heu is ic [87, 86, 88], which
ollows a p oblem-independen op imiza ion s a egy o adjus ing he
pa ame e s o he me a-heu is ic algo i hm, ha ob ains an op imal
solu ion [89]. Howe e , hese la e end o be oo p oblem-speci ic
o knowledge-in ensi e o be implemen ed in cheap, easy- o-use com-
pu e sys ems.
Following his idea, hype -heu is ic builds sys ems which can han-
dle classes o p oblems a he han sol ing jus one p oblem del ing
wi h he issues o me a-heu is ics. Using hype -heu is ic can be in e -
es ing in AD as i is able o op imise bo h pa ame e s and algo i hms
a he same ime.
44
2.3. Anomaly de ec ion o aul p edic ion
In he scope o his Thesis, no e e ence has been ound in he
li e a u e ha pe o ms a classi ica ion o p oposals e alua ing hese
concep s, i.e., heu is ics, me a-heu is ics and hype -heu is ics based
app oaches, in e ms o ailu e p edic ion. Fi s o all, i is wo h
no ing ha he as majo i y o e e ences in he li e a u e uses a
heu is ic me hod. In hese cases he p io knowledge equi ed abou
he p oblem a hand is high. This can be demanded ei he 1) in he
ini ial s eps, whe e da a is conscien iously p ep ocessed and selec ed
acco ding o he p oblem [90], 2) in he pa ame e se ing p ocess o
a gi en p oblem, such as Wz and h esholds, o adjus he in e nal
hype -pa ame e s o he algo i hms, o o make c ucial decisions nec-
essa y o he AD sys em o accu a ely de ec anomalies [91, 92, 93].
In his li e a u e e iew, a classi ica ion is p oposed based on he
amoun o knowledge equi ed by each p oposal. The ca ego iza ion
is decided depending on 3 ac ions equi ed by he use : a) he use
has o ix bo h ypes o pa ame e s (p oblem pa ame e s and hype -
pa ame e s o he algo i hm); b) he use has o pe o m a deep da a
p ep ocessing; c) he use has o in e ac wi h he ailu e p edic ion
sys em. In his con ex , h ee ca ego iza ions a e de ined:
•High Expe Knowledge, when he h ee ac ions abo e a e e-
qui ed.
•Medium Expe Knowledge, when wo o he h ee ac ions abo e
a e equi ed.
•Low Expe Knowledge, when one o he h ee ac ions abo e is
equi ed.
Table 2.2 summa ises he classi ica ion and compa ison done be-
ween he ela ed li e a u e in he con ex o collec i e AD. The a-
ble classi ies he solu ions by i s expe knowledge equi emen s and
indica es he me hods used o pa ame e es ima ion di e en ia ing
be ween ixed, calcula ed and op imized pa ame e s.
45
Chap e 2. S a e o he A
Once a da abase is a ailable in which he wo k o de s o ope a ing
condi ions a e ully iden i ied, he p edic i e main enance asks can
be pe o med.
As has been obse ed in he indus ial pa adigm, i is common
o ind he si ua ion whe e he explici ela ionships be ween p ocess
a iables and ailu es a e unknown. The e is a long his o y o ML
echniques, by which i can be s a ed ha hey a e a good solu ion
o de e mine he implici ela ionships be ween p ocess a iables and
pa ame e s and ailu es enabling o p edic ailu es and b eakdowns,
as hey a e based on his o ical da a and expe ience.
A sys em applied o label da abases in he con ex o Indus y
4.0, o p edic i e main enance asks, should ul il he ollowing
equi emen s:
•In eg a e a solu ion applicable o he common ypes o
da a in indus ial en i onmen s (MV, TS).
•Be comple ely au onomous in e ms o op imiza ion o
in e nal pa ame e s o he algo i hms and o he solu ion
i sel o he co ec labelling o PLD.
•Be use - iendly and unde s andable.
•P o ide a us wo hiness me ic o imp o ing he use
con idence on he sys em.
In his case, based on da a models and applying ML echniques,
he AD ask can p o ide a p ope solu ion. Sec ion 2.3.2 shows ha
he de ec ion o collec i e ype anomalies in he indus ial en i on-
men is gaining ele ance. In he li e a u e e iew, i has been ob-
se ed, i s ly, ha heu is ic solu ions ha equi e a medium-high
le el o expe knowledge a e he main employed. These solu ions
become p oblem-dependen , wi h a la ge numbe o pa ame e s o be
known and p ede ined and, consequen ly, wi h poo gene alisa ion ca-
pabili ies. Secondly, he use o me a-heu is ics o alle ia e he bu den
52

2.4. Conclusions
o pa ame e isa ion has no been widely exploi ed in he li e a u e.
In ac , mos p oposals use hem in hyb id o complex models o es-
ima e hype -pa ame e s bu wi h he limi a ion o using p ede ined
p ese pa ame e s. In ac , ew e e ences a e ound based on hype -
heu is ic inspi ed app oaches so as, some o hem y o be ei he
1) an in eg a ed and op imised ad-hoc solu ion o aul p edic ion
by means o inno a i e Deep-lea ning s uc u es, o 2) a bes model
selec ion amewo k ha gi es o he inal use he bes s a is ical
model and some ex ac ed ea u es o he TS ha de ec he anomaly.
Howe e , hese app oaches a e no use iendly, unde s andable and
ha e a educed numbe o p e-designed algo i hms, i.e., some Deep-
lea ning and s a is ical algo i hms espec i ely, and ex ac ed ea u es
o choose om, limi ing he bes possible solu ions and making hem
less gene alized.
In conclusion, in he scope o AD he p oposed solu ions a e lim-
i ed by se e al condi ions: 1) he need o inpu da a p ep ocessing,
2) he need o expe domain knowledge o classi y he aul s, 3) he
subop imal de ini ion o he ele an pa ame e s o he p oblem, 4)
he e alua ion in benchma k o syn he ic da ase s [110] which does
no gua an ee i s alidi y o speci ic eal-wo ld use cases, and 5) he
lack o unde s andable and in e p e able in o ma ion o he inal use
abou he TTF, impo an ea u es [111] o ele an pa ame e s ha
help o de ec he anomalies, making he p oposals a black-box model
o hem.
A e he analysis o he li e a u e in he con ex o ailu e p e-
dic ion in indus ial en i onmen s h ough he de ec ion o collec i e
anomalies, which is he common ype o anomaly iden i ied in he in-
dus ial en i onmen s, i can be de e mined ha he e is no p oposed
solu ion ha join ly add esses he ollowing poin s: 1) use ML ech-
niques, 2) be gene ic and no p oblem-dependen , 3) be easy o use
and unde s andable o he use , 4) p o ide he use wi h he op imal
se o pa ame e s ha a e in insically ela ed o he aul s in o de o
p edic hem, 5) be able o se he necessa y pa ame e s ensu ing he
53
Chap e 2. S a e o he A
bes esul s h ough op imisa ion echniques, 6) sea ch o he bes
se o heu is ics o ules ha p edic ailu es a he han he solu ion
di ec ly, and 7) be ained and es ed o e a eal indus ial use case.
All hese poin s p o ide he use wi h an op imal solu ion ha
iden i ies he po en ial pa ame e s and heu is ics which a e p edic o s
o he ela ed ailu es. This sys em is unde s andable s eng hening
he use ’s knowledge and easy o use.
A holis ic sys em based on collec i e AD echniques applied
o p edic i e main enance asks, should ul il he ollowing e-
qui emen s:
•Use ML and op imiza ion echniques.
•Be gene ic and no p oblem-dependan .
•Be easy o use and unde s andable o he use , which
should only in oduce he da a and ead he esul s, un-
de s anding hem and aking he ele an ac ions.
•P o ide he use wi h he op imal se o aul p edic o
pa ame e s o de ec ing collec i e anomalies.
•Be able o se he pa ame e s i needs. Look o he
bes se o heu is ics ha p edic he aul s ins ead o he
solu ion di ec ly.
•Be ained and es ed o e an indus ial use case.
54
CHAPTER 3
A HYPER-HEURISTIC INSPIRED
METHODOLOGY FOR FAILURE PREDICTION
This chap e desc ibes he in eg al hype -heu is ic inspi ed me hod-
ology p oposed in his Thesis o ailu e p edic ion. Speci ically, his
chap e illus a es how he ull me hodology is de eloped and uni ied
o ackle wo main and ele an issues in he con ex o Indus y 4.0.
Figu e 3.1 shows he concep ual scheme o he p oposed hype -
heu is ic inspi ed me hodology. Speci ically, his me hodology has
wo main objec i es o ackle he gaps ound in he li e a u e speci-
ied in Chap e 2: 1) o au oma ically label PLD and 2) o ob ain he
bes se o pa ame e s ha a e p edic o s o a ailu e om a FLD in
he con ex o Indus y 4.0. In he o me his me hodology, named
PLASH, has he ollowing objec i es: 1) be a use - iendly and au-
onomous sys em o labelling PLD, 2) e alua e and op imize he di -
e en clus e ing solu ions pe o med by a new semi-supe ised me ic
PSOM and 3) p o ide he use wi h he esul s in an unde s andable
way oge he wi h a us wo hiness me ic. Rega ding he ask o
collec i e anomaly de ec ion o ailu e p edic ion, named HIMAFP,
55
Chap e 3. A hype -heu is ic inspi ed me hodology o ailu e
p edic ion
he aim is o ind he op imal se o pa ame e s ha a e p edic o s o
a speci ic ype o ailu e ha allows o eliably and obus ly p edic
he ailu e pa ame e s p o iding he inal use wi h in o ma ion o
de ec he anomaly and an icipa e he ailu e.
As shown in Figu e 3.1 he hype -heu is ic me hodology can be
di ided in o h ee main s ages. The ini ial componen , i.e., Da a
p ocessing ocuses on de eloping he equi ed da a ea men o
di e en ypes o da a inpu s o be p ope ly p ocessed in he ol-
lowing componen . The cen al pa o he me hodology, i.e. HS
p ocedu e, ocuses on an op imiza ion sea ch p ocess ha employs
a well-known and es ablished me aheu is ic algo i hm called Ha mony
Sea ch. Finally, he me hodology includes a s age o esul s isual-
iza ion and in e p e a ion o he use in a clea and unde s and-
able manne .
56
Hype -heu is ic Me hodology
Ini ializa ion
nI e = 1 nI e = ζ ?
STOP
HMCR
PAR
RSR
Me ic
e alua ion
So ing and
selec ion o
he bes
ha monies
Yes
Imp o isa ion
nI e = nI e +1
No
Bes Ha monies
Selec ed
Is he da abase
comple ely
labelled?
Ini ial
Da abase
NO YES Fully labelled
da abase (FLD)
Pa ially labelled
da abase (PLD)
X%<100%
Da a p ocessing
100%
HS p ocedu e
A Hype -heu is ic Inspi ed
App oach o Au oma ic
Failu e P edic ion in he
Con ex o Indus y 4.0
Q1
PLAHS: a Pa ial Labelling
Au onomous Hype -heu is ic
Sys em o Indus y 4.0 wi h
applica ion on classi ica ion o
cold s amping p ocess
Q1
Resul s isualiza ion and in e p e a ion
Final solu ion
wi h he mos
ep esen a i e
pa ame e s ha
p edic he
ailu e
Use app o es
labelling
p oposal
Real ime ailu e
de ec ion sys em
Decision Making
owa ds
P edic i e
Main enance
Me hodology
PLD FLD
Figu e 3.1: Hype -heu is ic inspi ed me hodology applied in his The-
sis.
By ollowing his app oach, he use simply needs o inpu he
da abase o pe o ming he analysis. The e a e wo po en ial scena -
ios based on he da abase a he beginning.
•I he da abase is PLD, i s ly he sys em will au oma ically la-
bel he unlabelled samples o he da abase. As s a ed be o e,
his pa o he me hodology (in o ange colo in he Figu e 3.1) is
57

Chap e 3. A hype -heu is ic inspi ed me hodology o ailu e
p edic ion
called PLAHS. A e he da abase has been labelled and he e-
o e i is a FLD, an AD ask is de eloped o ob ain a se o
ailu e p edic o pa ame e s.
The jou nal pape associa ed wi h his me hodology is ound in:
Applied So Compu ing (Q1) “Mino e ision”. Na ajas
Gue e o, A., Po illo, E. & Manja es, D., PLAHS: A Pa ial
Labelling Au onomous Hype -Heu is ic Sys em o Indus y 4.0
wi h Applica ion on Classi ica ion o Cold S amping P ocess.
A ailable a SSRN 4288726.
•I he da abase is FLD, he sys em can di ec ly de elop he AD
ask o ob ain a se o pa ame e s ha a e po en ial p edic o s
o he ailu e o in e es . As men ioned be o e, his pa o he
me hodology (in pu ple in he Figu e 3.1) is called HIMAPF.
The jou nal pape associa ed wi h his me hodology is ound in:
Compu e s and Indus ial Enginee ing (Q1). Na ajas-
Gue e o, A., Manja es, D., Po illo, E., & Landa-To es, I.
(2022). A hype -heu is ic inspi ed app oach o au oma ic ail-
u e p edic ion in he con ex o indus y 4.0. Compu e s &
Indus ial Enginee ing, 171, 108381.
3.1 Da a p ocessing
In his s age an ini ial ea men and p ocessing o he da a is
conduc ed, i.e., i is adjus ed o he necessa y cha ac e is ics o each
o he main objec i es add essed in his Thesis h ough he p oposed
hype -heu is ic inspi ed app oach.
In bo h cases, i.e., he au o-labelling (PLAHS) and he au oma ic
ailu e p edic ion (HIMAPF), a unique da a p ocessing me hodology
is p oposed. In gene al i is necessa y o o ganize he da a in a cohe -
en way wi h he p oposed hype -heu is ic inspi ed me hodology, and
p ep ocess he da a acco ding o he equi emen s o he HS-based
58
3.2. Ha mony sea ch p ocedu e
solu ion. Speci ically, o PLAHS, he labeled and unlabeled da a
should be o ganized p ope ly o ex ac eliable in o ma ion, while
in HIMAFP, da a p ocessing is necessa y o isola e ailu e cases o
independen conside a ion by he HS.
In Chap e 5, Subsec ion 5.2.1, an ex ensi e explana ion o he
da a p ocessing echniques used o he au o-labelling ask is desc ibed
in dep h. I de ails how he da a is u ilized and in eg a ed o analy-
sis pu poses. Simila ly, Chap e 6, Sec ion 6.3 also o e s an in-dep h
analysis o how he da a is p ocessed o he au oma ic ailu e p edic-
ion ask.
3.2 Ha mony sea ch p ocedu e
In he second s age, an op imisa ion p ocess akes place o bo h
scena ios, which will p o ide he basis and cohe ence o he hype -
heu is ic s a egy. Op imiza ion me hods a e widely employed o a -
ious pu poses, such as indus ial planning, econome ics, scheduling,
decision making, enginee ing, and compu e science applica ions. The
op imiza ion ield is an ac i e a ea o esea ch, wi h new echniques
con inuously being de eloped [112]. Op imisa ion in ol es choosing
he mos sui able op ion om a gi en se o al e na i es, subjec o
ele an cons ain s. The p ocess en ails minimizing o maximizing
he objec i e o cos unc ion o he p oblem. The p ocess i e a i ely
selec s alues om a pe missible se un il he op imal ou come is a -
ained, o he s opping c i e ion is me .
Me a-heu is ic algo i hms a e well-es ablished and e ec i e me h-
ods o sol ing op imiza ion p oblems wi h sa is ac o y ou comes [113].
These algo i hms possess se e al ea u es, such as simplici y, obus -
ness, and lexibili y, which make hem an a ac i e a ea o esea ch
o e icien ly sol e compu a ional NP-ha d sea ch p oblems [114, 53].
Many me a-heu is ic algo i hms d aw inspi a ion om na u al phe-
59
Chap e 3. A hype -heu is ic inspi ed me hodology o ailu e
p edic ion
nomena, such as Pa icle Swa m Op imiza ion (PSO), SA, GA, and
HS. These algo i hms a e in elligen ly designed and can p oduce e -
ec i e solu ions o a wide ange o op imiza ion p oblems [112].
Chap e 2, Sec ion 2.3.1 s a es ha using me a-heu is ic algo-
i hms is a common me hod o de elop hype -heu is ic algo i hms.
Thus, he p oposed hype -heu is ic inspi ed me hodology o sol e
bo h asks is de eloped by means o a me a-heu is ic algo i hm.
In pa icula , he p oposal o hype -heu is ic inspi ed me hodol-
ogy is implemen ed by a me a-heu is ic algo i hm, speci ically he HS
algo i hm, o explo e and exploi he esolu ion space o an op imal
solu ion o a p oblem. The HS algo i hm is a popula ion-based al-
go i hm ha uses a se o solu ions, called ha monies, s o ed in he
Ha mony Memo y (HM). I is achie ed h ough an i e a i e p ocess,
applying se e al imp o isa ion ope a o s o ind he bes i ness alue
o a solu ion ec o , called ha mony. HS algo i hm emula es he col-
labo a i e beha io o musicians who adjus hei ins umen s’ pi ches
o achie e a ha monious sound. HS is a highly e ec i e me a-heu is ic
algo i hm ha gene a es a solu ion ec o in elligen ly h ough he
explo a ion and exploi a ion o he sea ch space.
Since i s eme gence in 2001 in [115] by Geem e al., he HS algo-
i hm has been ecognized as a highly e icien popula ion-based me a-
heu is ic algo i hm o combina o ial op imiza ion. I has eached
signi ican in e es om esea che s ac oss di e se ields, who ha e
enhanced i s pe o mance by ine- uning i s pa ame e s and in eg a -
ing i s componen s wi h o he me a-heu is ic algo i hms [112]. This
algo i hm has shown in se e al pieces o esea ch [116, 117, 118] ha
hanks o i s in e nal ope a o s (HMCR memo y conside a ion, PAR
adjus men a e and RSR andomness) an op imal in e ac ion be-
ween explo a ion and exploi a ion in he sea ch o he bes solu ion
is achie ed.
As shown in Figu e 3.1, his p ocess is mainly composed o i e
60
3.2. Ha mony sea ch p ocedu e
main pa s i.e., 1) ini ializa ion, 2) imp o isa ion, 3) he p ope
me hodology depending on he scena io, 4) me ic e alua ion and
5) so ing and selec ion o he bes ha monies.
3.2.1 Ha mony sea ch algo i hm
Ini ializa ion. This s ep is only pe o med du ing he i s i e a-
ion and he ini ial ha mony is p oposed. To c ea e he HM, wo
ac o s mus be conside ed. Fi s ly, he Size o he Ha mony Memo y
(HMS) ha de e mines he numbe o ha monies ha will cons i u e
he o e all HM. Secondly, he choice o encoding sys em is c ucial.
The encoding consis s o a ce ain numbe o no es (nx), which se e
as pa ame e s o be op imized and o m he basis o he ha mony.
The ha monies o he HM a e gene a ed andomly among he easible
alues o each no e. An example o he s uc u e o he HM is shown
abo e.
In he HM 3.1 i is shown a ma ix o di e en nxno es pe ha -
mony and he espec i e i ness unc ion (nx) o a leng h o HSM
ha monies.
HM =









n11 n12 ··· n1x (n1)
n21 n22 ··· n2x (n2)
n31 n32 ··· n3x (n3)
.
.
..
.
.....
.
..
.
.
nHSM1nHSM2··· nHSMx (nHSM )









(3.1)
Imp o isa ion. This is a p ocess whe ein a new HM is gene a ed
by applying consecu i ely h ee di e en p obabilis ic pa ame e s ha
modi ies he ini ial HM. These p obabilis ic pa ame e s a e: HMCR
(Ha mony Memo y Conside ing Ra e), PAR (Pi ch Adjus ing Ra e)
61
Chap e 4. Case s udy: A cold-s amping p ess o bol
manu ac u ing
Figu e 4.2: Wi e od cold-s amping p ocess.
and malleable, such as low alloy s eel, aluminium alloys (p e e ably
magnesium alloys wi hou coppe ), b ass, sil e and gold. Commonly,
di e en elemen s a e added o he s eel con ibu ing o he inal cha -
ac e is ics o he ma e ial o be used. Some o he elemen s mos
commonly employed in alloys a e [121]:
•Coppe . I imp o es he co osion esis ance o he ma e ial
[122].
•Nickel. I educes ha dening and dis o ion empe a u e when
ha dened. The nickel alloy ex ends he c i ical empe a u e
le el, wi hou ca bides o oxides. This inc eases s eng h wi h-
ou dec easing duc ili y [122].
•Ch omium. I inc eases ha denabili y and imp o es wea and
co osion esis ance. I esul s in he c ea ion o ch omium ca -
bides ha a e e y ha d. Howe e , hey a e mo e duc ile han
a s eel o he same ha dness p oduced simply by inc easing i s
ca bon con en . The addi ion o ch omium ex ends he c i ical
empe a u e ange o he s eel [122].
68

4.1. Cold-s amping p ocess
Figu e 4.3: Examples o shee me al cold-s amping pieces.
Figu e 4.4: Examples o wi e od cold-s amping pieces.
4.1.2 Pa s o a cold o ming machine o bol manu-
ac u ing.
As men ioned be o e, he case unde s udy is a s amping p ocess
ha wo ks wi h a wi e od placed on a coil as aw ma e ial and is
commonly used o he manu ac u e o bol s and nu s. The machine y
ha de elops his wo k is composed mainly o 3 blocks as Figu es 4.5
and 4.6 show.
69
Chap e 4. Case s udy: A cold-s amping p ess o bol
manu ac u ing
•Die-holding block o T ans e : This is he pa in which he dies
a e housed, as well as he en i e se o ools o he cold p ess.
I has a cylind ical shape, so ha he dies a e adjus ed o he
diame e o he wo kpiece. This block is mainly composed o
wo pa s: a module (commonly known as a ans e ) which can
be mo ed and olded down, and a lowe pa whe e he dies a e
housed. By mo ing he ans e , he dies can be accessed o
change o epai hem. This pa is also esponsible o mo ing
he ma e ial du ing s amping by means o cams and di e en
mechanisms.
•Ri e ing pin block: This block, unlike he die se , consis s o a
single piece. The clamping sys em is simila o ha o he inne
block. A he back o he piece he e a e wedges, which allow he
use o modi y he posi ion o each i e ing pin independen ly
o he o he s.
Ri e ing pin block Die holding block
Ri e ing pin
Die
Figu e 4.5: Die holding and i e ing pin blocks in cold s amping ma-
chine.
•Feed olle s and s aigh ene s: These olle s ha e he unc ion
o d agging he ma e ial (wi e od). They a e moun ed in such
a way ha he ma e ial is held unde p essu e be ween he wo
olle s and when he olle s o a e, he ma e ial ad ances. P io
o passing h ough he olle s, he ma e ial is passed h ough a
70
4.1. Cold-s amping p ocess
s aigh ene so as no o encoun e any p oblems when pulling
he ma e ial. The olle s ha e a g oo e o he diame e o he
ma e ial in which he wi e i s. In o de o a oid ma ks on
he ma e ial, he edges be ween he g oo e and he ou side a e
ounded.
Feed olle s and s aigh ene s
aw ma e ial
Figu e 4.6: Feed olle s and s aigh ene s.
4.1.3 S eps in he cold- o ming p ocess o bol s man-
u ac u ing.
As Figu e 4.7 shows, he bol manu ac u ing p ocess consis s o
h ee main s ages. Fi s , he ma e ial, which is in he o m o wi e on
a coil, is ed in o he machine y h ough he olle s and s aigh ene s.
Once he wi e en e s, i i s goes h ough he cu ing p ocess, whe e
he pa ame e s a e se o cu i o he desi ed size o leng h.
A e cu ing, he piece o ma e ial is anspo ed un il he i s die
and is hen placed in he ans e sys em. I is hen mo ed h ough
con inuous die and ool openings o displace and o m he wo king
me al in o he desi ed p oduc . The i e ing pin (mobile ool in cha ge
o s iking he piece), pushes he ma e ial in o he die, hus gi ing he
shape o he die. Finally, a e passing h ough all he s a ions o he
p ocess, he bol is eady o ul il i s unc ion.
71
Chap e 4. Case s udy: A cold-s amping p ess o bol
manu ac u ing
The cold s amping p ocess is a high-speed manu ac u ing p ocess,
so he empe a u e and p essu e mus be jus igh a each s ep. I
can be used o educe o inc ease he diame e s and leng hs o he
aw ma e ial and can emo e small amoun s o ma e ial by punching
and imming. All o hese cold s amping p ocesses and ope a ions
a e ca ied ou in a con inuous planned p ocess [51].
6 s ages o ming p ocessCu ing s age
4 in e media e
s ages
Wi e od in oduc ion
Figu e 4.7: Bol cold o ming p ocess.
4.2 Cold o ming da abase
The use case employed in his Thesis pe o ms he s eps desc ibed
abo e. In his pa icula case, as Figu e 4.8 shows, he p ocess con-
sis s o six s a ions whe e each one pe o ms a di e en ope a ion on
he ma e ial. In o de o collec da a om he p ocess, a se ies o
piezoelec ic load cell senso s moni o ing he o ce o s amping ha e
been ins alled a each o he six p ocess s a ions.
The p ocess is moni o ed o e he cou se o a yea and he da a
is s o ed in ex iles, each o which con ains eco ds o he s eng h
o he six signals in 5-minu e pe iods (acquisi ion equency 1 kHz).
Finally, hese iles a e s o ed in a da a ile.
72
4.2. Cold o ming da abase
Da a
S o age
File sys em
Raw Ma e ial 6 S age cold s amping machine Piezoelec ic load
cell senso s
Ma e ial in oduc ion
Cold s amping bol
o ming
Moni o iza ion and Da a
adquisi ion
Figu e 4.8: Flow diag am o he cold o ming p ocess unde s udy.
4.2.1 Da a con ained in he da abase
Th oughou he moni o ing pe iod, p oduc ion s oppages ha e oc-
cu ed o h ee di e en easons: 1) machine s oppage, which can be
caused by unexpec ed b eakage o wea o di e en componen s, 2)
olun a y s oppage, ha a e due o a ious pa s eplacemen o sched-
uled main enance o 3) s oppage due o b ankamp, which is ela ed o
a condi ion moni o ing sys em ha exis s in he p oduc ion plan in
which some limi s in he e o cu es a e se so as unde some speci ic
ules he machine is s opped. Fo his case s udy, machine s oppage
eco ds will be used and analysed, which a e hose ha appea un-
expec edly due o b eakage o wea and ea and he e o e necessa y
o be de ec ed in ad ance. Thus, i is in ended o de elop a p edic-
i e main enance on hese ools and hus an icipa e he occu ence o
b eakage o wea ha lead o a machine s oppage..
Fu he mo e, a documen con aining manually eco ded da a by
he machine ope a o , he ea e named S ops Regis e s ile (SR- ile),
is a ailable o documen and classi y any p oduc ion s oppages. The
eco d includes he cause o he s oppage and he associa ed WO in
which he machine is cu en ly engaged. Fo he domain expe is
pa icula ly impo an o know in which WO each b eakage-s op oc-
cu s o a p ope analysis o he p ocess. This means ha he machine
mus wo k wi h ce ain ope a ing pa ame e s ha cha ac e ise each
WO. Speci ically, he concep s ha a e egis e ed in his ile a e he
73

Chap e 4. Case s udy: A cold-s amping p ess o bol
manu ac u ing
ollowings:
•Type o s oppage ( olun a y, machine o bank amp).
•Da e and ime o he eco ded b eakdown.
•Reason o he s oppage depending on he ype o he s oppage.
•Tools in ol ed.
•O he obse a ions.
Howe e , as discussed in he in oduc ion, he eali y in eal da-
abases is ha hey con ain noise, ou lie s, and a e pa ially labelled.
This case s udy p esen s he ollowing d awbacks:
•In he manual eco ding o s oppages, da a abou WOs, he ea-
son o elemen a ec ed by he s op among o he pa ame e s a e
missing in some cases. This could be caused by di e en ac o s
such as ope a o shi changes o human e o s. The e o e i is
a pa ially labelled da abase.
•The e is a ime lag be ween he end o he ime se ies ep esen -
ing a s op and hei manual eco ding in he SR- ile. Despi e
ha ing access o a as his o ical da abase, one o he main chal-
lenges aced is he ma ching be ween he moni o ed signal and
i s co esponding manual egis a ion. As a esul , he num-
be o accu a ely eco ded s oppages ha can be co ec ly a -
ibu ed o he speci ic WO and he co esponding b oken o
wo n elemen has been educed.
As p e iously men ioned, he e a e six di e en ypes o signals
belonging o he o ce eco ds o he di e en s a ions. Figu e 4.9
illus a es he shape o he ime se ies esul ing om moni o ing wo
s amps. Conside ing each o he signals o wo s amps, he Figu es
4.10 o 4.15 a e ob ained.
74
4.2. Cold o ming da abase
Figu e 4.9: Signals cap u ed om he cold o ming p ocess in wo
s amps.
Figu e 4.10: P ocess signal 1.
75
Chap e 4. Case s udy: A cold-s amping p ess o bol
manu ac u ing
Figu e 4.11: P ocess signal 2.
Figu e 4.12: P ocess signal 3.
76
4.2. Cold o ming da abase
Figu e 4.13: P ocess signal 4.
Figu e 4.14: P ocess signal 5.
77
Chap e 7. Conclusions and Fu u e Wo k
iendly and comp ehensible solu ions ha inco po a e a ious
algo i hms and ea u es.
The e o e, a comp ehensi e ML solu ion, ha is gene ic and no
p oblem-dependen , is easy o use and unde s and, p o ides op-
imal pa ame e s ela ed o aul s, and can be sel -pa ame e ized
wi h op imiza ion echniques is needed. The solu ion should also
sea ch o he bes se o heu is ics o ules ha p edic ailu es
a he han he solu ion di ec ly, and be ained and es ed on
eal da ase s. This in eg a ed solu ion would o e he op imal
ailu e p edic ion pa ame e s while being use - iendly and p o-
mo ing use knowledge.
7.2 Con ibu ions
The key con ibu ion o his Thesis is he c ea ion o a hype -
heu is ic inspi ed me hodology ha le e ages op imiza ion algo i hms
and ML o p edic ailu es in indus ial sys ems wi hin he con ex o
Indus y 4.0. This me hodology is hype -heu is ic inspi ed in e ms
o : a) educing he expe knowledge equi ed, b) p o iding an easy-
o-use compu e sys em, c) ope a ing on a ange o ela ed p oblems
a he han on one na ow class o p oblems. Chap e 3 ho oughly
elucida es his me hodology, which consis s o wo dis inc main s ages
and success ully add esses he p ima y goal o he hesis, which is he
de elopmen o a ML-based solu ion o p edic ing indus ial sys em
ailu es.
As men ioned and desc ibed in Chap e 3 his me hodology has
2 main objec i es: 1) o au oma ically label PLD and 2) o ob ain
he bes se o pa ame e s ha a e p edic o s o a ailu e om a
FLD in he con ex o Indus y 4.0. In he ask o labelling PLD his
me hodology has he ollowing objec i es: 1) being a use - iendly and
au onomous sys em o labelling PLD, 2) e alua e and op imize he
di e en clus e ing solu ions by a new semi-supe ised me ic PSOM
204

7.2. Con ibu ions
and 3) p o ide he use wi h bo h he esul s in an unde s andable way
and a us wo hiness me ic. The me hodology applied o e his ield,
in his Thesis is called PLAHS. Rega ding, he ask o collec i e AD
o ailu e p edic ion, he aim is o ind he op imal se o pa ame e s
ha a e p edic o s o a speci ic ype o ailu e ha allows o eliably
and obus ly p edic he ailu e pa ame e s p o iding he inal use
wi h in o ma ion o de ec he anomaly and an icipa e he ailu e. In
his espec , he me hodology name o his ield is HIMAFP.
In pa icula , he subsequen poin s desc ibe he con ibu ions o
his Thesis, which a e di ec ly linked o he accomplished and p o-
posed objec i es.
Con .1 The hesis p esen s a comp ehensi e and de ailed s udy o wo
speci ic a eas. Fi s ly, he a ailable echniques o labeling unla-
beled samples o da abases a e in es iga ed. This is pa icula ly
signi ican in he con ex o p edic i e main enance, as ha ing
mo e in o ma ion leads o mo e accu a e ML-based models. To
his end, a ho ough su ey o semi-supe ised clus e ing-based
echniques in he li e a u e is conduc ed and a new axonomy
is p oposed. No ably, he e alua ion o he solu ions based on
whe he hey a e heu is ic, me a-heu is ic, o hype -heu is ic,
which has no been done in p e ious esea ch, is pa o he
con ibu ions o his hesis.
Secondly, ML solu ions o de ec ing collec i e anomalies in he
con ex o Indus y 4.0 a e explo ed. The con ibu ion is wo-
old: 1) p oposing a classi ica ion o hese solu ions based on he
LER o implemen hem, and 2) e alua ing and classi ying he
solu ions based on whe he hey a e heu is ic, me a-heu is ic,
o hype -heu is ic app oaches. I should be no ed ha no p io
esea ch has conside ed his classi ica ion.
This con ibu ions accomplish he objec i e 1 (Obj. 1) p oposed
in Chap e 1.
205
Chap e 7. Conclusions and Fu u e Wo k
Con .2 Th ough he design and de elopmen o he in eg al hype -
heu is ic inspi ed me hodology, i has been possible o de elop
an ex ensi e s udy no only abou he ex ac ed Fe om he
TSM ha may be p edic o s o ailu e, bu also o he p edic o
pa ame e s, such as he Wz p io o he b eak whe e anoma-
lous beha iou is ound and he Th ha de ec s i . To de elop
his s udy, he pa o he in eg al me hodology ha has been
designed is HIMAPF.
In summa y, he esul s ob ained om he p oposed hype -
heu is ic inspi ed me hodology HIMAFP o collec i e AD and
ela ed aul p edic ion applied o e a eal da abase shows in-
e es ing esul s. The HIMAFP has demons a ed i s use ulness
in e ms o : a) p o iding domain expe s wi h aluable knowl-
edge abou he beha iou o signi ican p ocess a iables in eal
cases wi h lack o in o ma ion, b) de eloping an easy- o-use use -
iendly me hodology, c) adap ing o di e en ypes o ailu es
and i s co esponding anomalies.
The s udy has in ol ed es ing six di e en ypes o b eakages
in a cold s amping p ocess used o bol manu ac u ing. The
indings demons a e ha HIMAFP is e ec i e in iden i ying
he op imal combina ion o ele an ea u es om he TSM and
hei associa ed h esholds, which can indica e he imminen
occu ence o a ailu e. HIMAFP can also handle he dynamic
beha io s o abno mal TSM Fe and Wz o ob ain TTF in o ma-
ion. This in o ma ion is aluable o de eloping eal- ime AD
sys ems and making impo an co ec i e, p e en i e, o p e-
sc ip i e decisions in indus ial p ocesses. The s udy highligh s
he impo ance o ob aining in o ma ion abou ea u es in bo h
he ime and equency domains, as e idenced by he es ing o
19 di e en ea u es. The op imal esul s ha e been ob ained
using equency domain ea u es in 30% o he o al ha monies,
emphasizing hei signi icance and alida ing hei inclusion in
he s udy.
This con ibu ion accomplish objec i es 2 (Obj. 2) and 3 (Obj.
206
7.2. Con ibu ions
3) p oposed in Chap e 1.
Con .3 In he con ex o designing and de eloping a gene al solu ion
ha can wo k wi h bo h labelled and pa ially labelled da a-
bases, a solu ion capable o in e ing knowledge om he la-
belled pa o he da abases o label unknown samples has been
de eloped. Speci ically, he design and de elopmen owa ds
an in elligen hype -heu is ic inspi ed sys em (PLAHS) o au-
onomously label PLD, ha enable o classi y signals om in-
dus ial p ocesses in he con ex o Indus y 4.0, has been p o-
posed and is pa o he con ibu ions o his Thesis.
PLAHS is a comple e sys em ha combines a boo s apping
da a p ocessing echnique, he HS me a-heu is ic algo i hm, and
an ensemble o ing me hod o pa ially labeled da abases. This
sys em is inno a i e o se e al easons. Fi s ly, i p o ides a la-
beling solu ion o ime se ies and mul i a ia e da a. Secondly, i
in eg a es he HS algo i hm o op imize he in e nal pa ame e s
o h ee commonly used clus e ing algo i hms (K-means, DB-
SCAN, and HAC), while also deli e ing he bes clus e ing and
op imized solu ion o each case s udy. Thi dly, i in oduces a
new e alua ion me ic (PSOM) ha conside s bo h supe ised
and unsupe ised pa s, weigh ed by he pe cen age o labels.
Fou hly, i is use - iendly, allowing use s o inpu hei da ase
and easily unde s and he esul s. Finally, he sys em p o ides a
no el us wo hiness me ic (TM) o measu e i s beha io and
he use ’s le el o con idence in he esul s ob ained.
PLAHS has demons a ed i s abili y o adap , e alua e, and
p o ide he bes labelling solu ion o a ious ypes o da a, in-
cluding mul i a ia e and ime se ies da a, and scena ios wi h
di e en pe cen ages o labelled da a anging om 15 o 90%.
The sys em’s e ec i eness has been shown h ough se e al ex-
pe imen s: 1) se en da abases om he UCI eposi o y ha e
been used o es di e en pe cen ages o knowledge and de-
g ees o sepa abili y, 2) six da abases om he UCR eposi o y
207
Chap e 7. Conclusions and Fu u e Wo k
ha e been used o es di e en deg ees o labelled elemen s and
in insic sepa abili y and 3) a eal case s udy in ol ing a cold
s amping p ess o bol manu ac u ing has been conduc ed.
In addi ion, his app oach has demons a ed i s hype -heu is ic
inspi a ion in a ious ways: a) a low le el o expe knowledge
equi emen , b) an easy o use sys em, c) i s abili y o ope a e
on di e en ypes o p oblem.
This con ibu ion accomplish objec i e 4 (Obj. 4) p oposed in
Chap e 1.
Con .4 Th oughou he de elopmen o his Thesis, di e en s udies and
analyses ha e been de eloped ela ed o he e alua ion me ics
in he di e en ypes o lea ning wi hin he ML echniques. In
his way, he ollowing expe imen s ha e been de eloped.
In he i s ials and s udies, a sys em o op imiza ion o clus e -
ing pa ame e s and selec ion o he bes clus e ing algo i hms
has been p oposed in HSOCC, whe e he in e nal alida ion
me ics a e in es iga ed and analyzed, speci ically he well-known
Silhoue e me ic.
In he con ibu ion and labeling p oposal, PLAHS, an analy-
sis and bibliog aphic sea ch is pe o med on di e en e alua ion
me ics in semi-supe ised en i onmen s. The gene al ule when
de eloping labelling sys ems is o use ex e nal alida ion indices
whe e a con usion ma ix is cons uc ed wi h he known samples
and ex apola ed o he global solu ion. O he me ics used in
he li e a u e measu e he deg ee o cohesion o he g oups ha
a e o med. In he p oposal and con ibu ion o his Thesis, a
dual me ic called PSOM has been de eloped based on a supe -
ised pa e alua ed by a p oposed and no el me ic called MRC
and an unsupe ised pa e alua ed by Silhoue e. In PSOM
bo h pa s a e weigh ed by he pe cen age o labels. Wi h e-
ga d o he esul s in he expe imen a ion de eloped in PLAHS,
he p oposed me ics o e alua e semi-supe ised (PSOM) and
supe ised (MRC) en i onmen s ha e shown o be able o gi e
208
7.2. Con ibu ions
a co ec and accu a e assessmen . In his way, by using he
MRC me ic and he F1-sco e o he supe ised pa , i has been
possible o de elop a us wo hiness me ic (TM) ha is no
only able o gi e he end-use an e alua ion o how eliable he
p oposed labeling solu ion is, bu also, i is able o en a i ely
calcula e he alue o he F1-sco e o he unknown pa . This
co ela ion has an R2 i alue o mul i a ia e bases o 0.813
and o ime se ies o 0.755 as he esul s in Chap e 5 e eal.
No exis ing e e ence has been ound p oposing such a me ic
o assessing eliabili y in semi-supe ised en i onmen s and i s
applica ion o labeling, as de e mined h ough a comp ehensi e
e iew o he s a e o he a li e a u e on his opic.
Finally, in ully supe ised en i onmen s and in he con ex o
he de ec ion o collec i e anomalies in ime se ies, he e is a
wide ange o ex e nal alida ion me ics ha epo he deg ee
o sa is ac ion o he solu ions. Some o he mos widely imple-
men ed a e AUC ROC , Sensi i i y, Speci ici y, Recall, P eci-
sion, o Accu acy among o he s. The analysis o hese me ics
has been pe o med in he HIMAPF me hodology. As pa o
he con ibu ion o his hesis, an e alua ion me ic based on
AUC ROC , called AUC MOD , has been de eloped and ali-
da ed. Speci ically, o cases whe e i is impo an o know when
he anomalies appea , since he AUC ROC does no ake i in o
conside a ion. By con as , as shown and men ioned, he p o-
posed AUC MOD me ic enhances he alue o he HIMAFP
o i s applica ion o he indus y whe e he ime o occu ence
o an anomaly is i al o p edic he ailu es.
As a summa y, his hesis has pu o wa d ou con ibu ions
wi h ega ds o e alua ion and us wo hiness me ics. These
include: 1) a semi-supe ised e alua ion me ic called PSOM, 2)
a supe ised e alua ion me ic named MRC speci ically designed
o pa ially supe ised en i onmen s, 3) supe ised e alua ion
me ics in ended o collec i e AD in he con ex o Indus y
4.0 called AUC MOD, and 4) a us wo hiness me ic (TM)
209

Chap e 7. Conclusions and Fu u e Wo k
ailo ed o semi-supe ised en i onmen s.
This con ibu ions accomplish objec i es 5 (Obj. 5) and 6 (Obj.
6) p oposed in Chap e 1.
7.3 Resul s
F om he de elopmen s and con ibu ions done in his Thesis, he
ollowing jou nal a icles ha e been published:
•Compu e s and Indus ial Enginee ing (Q1). Na ajas-
Gue e o, A., Manja es, D., Po illo, E., & Landa-To es, I.
(2022). A hype -heu is ic inspi ed app oach o au oma ic ail-
u e p edic ion in he con ex o indus y 4.0. Compu e s &
Indus ial Enginee ing, 171, 108381.
•Applied So Compu ing (Q1) “Mino e ision”. Na ajas
Gue e o, A., Po illo, E., & Manja es, D. PLAHS: A Pa ial
Labelling Au onomous Hype -Heu is ic Sys em o Indus y 4.0
wi h Applica ion on Classi ica ion o Cold S amping P ocess.
A ailable a SSRN 4288726.
Addi ionally, he ollowing con e ence pape and p esen a ion ha e
been made:
•In 14 h In e na ional Con e ence on So Compu ing
Models in Indus ial and En i onmen al Applica ions
(SOCO 2019). Na ajas-Gue e o, A., Manja es, D., Po illo,
E., & Landa-To es, I. (2020). A no el heu is ic app oach o
he simul aneous selec ion o he op imal clus e ing me hod and
i s in e nal pa ame e s o ime se ies da a. In 14 h In e na-
ional Con e ence on So Compu ing Models in Indus ial and
En i onmen al Applica ions (SOCO 2019) Se ille, Spain, May
210
7.4. Fu u e wo k
13–15, 2019, P oceedings 14 (pp. 179-189). Sp inge In e na-
ional Publishing.
•I Cong eso Anual de Es udian es de Doc o ado (CAED).
In his case a p esen a ion en i led “Au oma ic sys em o ail-
u e p edic ion in Indus y ” was p esen ed.
7.4 Fu u e wo k
This Thesis has made i possible o iden i y se e al po en ially
in e es ing a eas o u u e esea ch. The mos signi ican ones a e:
•In Chap e 5 he selec ed clus e ing algo i hms a e among he
mos equen ly u ilized ones in he li e a u e,i.e. K-Means, DB-
SCAN and HAC. As a po en ial a enue o u u e in es iga ion,
i would be aluable o expand he numbe o algo i hms consid-
e ed and co e a b oade ange o po en ial solu ions. Fu he -
mo e, his sys em boas s he ad an age o being easily ex end-
able and gene alizable in e ms o i s con igu a ions. I could
be wo hwhile o inco po a e a g ea e a ie y o p econ igu ed
dis ance measu es o clus e o ma ion, such as cosine o Ma-
halanobis dis ances, among o he s.
•Chap e 6 desc ibes he pa o he hype -heu is ic me hodol-
ogy ha is dedica ed o collec i e AD o p edic ing ailu es
in indus ial en i onmen s (HIMAPF), i would be in e es ing
ex end he p oposed me hodology by in eg a ing in he op imi-
sa ion p ocess a sliding window ha , wi h he ini ial pa ame e s
ob ained in HIMAPF, can ine- une hese pa ame e s. In his
way, he objec i e is o ind he new se o ine- uned p edic o
pa ame e s such as: he size o he sliding window (Ws) used
o ex ac he ea u e (Fe) sequen ially o ob ain a ime se ies
o he cha ac e is ic (Fews), he size o he deg ada ion window
(Wz) whe e he expec ed anomaly p io o ailu e is loca ed, and
211
Chap e 7. Conclusions and Fu u e Wo k
he h esholds (ThHand ThL) ha de ec anomalous beha io
in FeW s wi hin he deg ada ion window (Wz). These ine- uned
pa ame e s could make i possible o de ec collec i e anomalies
ha p edic ailu e in a mo e p ecise, adjus ed, and accu a e
manne .
•In he con ex o Chap e 6, i would be wo h explo ing a po-
en ial u u e line o esea ch which ocuses on iden i ying which
se o se s o ea u es ex ac ed om he signals ha e he po en-
ial o p edic aul s. This could be achie ed by conduc ing a
s udy wi hin he op imiza ion me hodology, aimed a de e min-
ing which combina ions o ea u e pa ame e s wo k well oge he
o p o ide an accu a e p edic ion. Ra he han simply p o iding
a lis o indi idual ea u e pa ame e s, he s udy would seek o
iden i y combina ions o ea u es ha , when used oge he , ha e
he po en ial o p o ide a mo e comp ehensi e and eliable aul
p edic ion. This app oach could be mo e e ec i e in iden i y-
ing aul s ea ly, be o e hey de elop in o mo e se ious p oblems,
ul ima ely imp o ing he o e all eliabili y and e iciency o he
sys em unde conside a ion.
212
Appendices
213
Appendix A. Resul s o Chap e 6
0 5 10 15 20 25 30
I e a ions
0.66
0.68
0.70
0.72
0.74
0.76
0.78
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.10: Me ic e olu ion AUC MOD in Failu e ype B in in
TSM4 o all combina ion o HS ope a o alues.
0 5 10 15 20 25 30
I e a ions
0.70
0.75
0.80
0.85
0.90
0.95
1.00
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.11: Me ic e olu ion AUC MOD in Failu e ype B in TSM5
o all combina ion o HS ope a o alues.
220

A.1. Ha mony Sea ch ope a o s analysis
0 5 10 15 20 25 30
I e a ions
0.70
0.75
0.80
0.85
0.90
0.95
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.12: Me ic e olu ion AUC MOD in Failu e ype B in TSM6
o all combina ion o HS ope a o alues.
A.1.3 Type C
0 5 10 15 20 25 30
I e a ions
0.825
0.850
0.875
0.900
0.925
0.950
0.975
1.000
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.13: Me ic e olu ion AUC MOD in Failu e ype C in TSM1
o all combina ion o HS ope a o alues.
221
Appendix A. Resul s o Chap e 6
0 5 10 15 20 25 30
I e a ions
0.825
0.850
0.875
0.900
0.925
0.950
0.975
1.000
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.14: Me ic e olu ion AUC MOD in Failu e ype C in TSM2
o all combina ion o HS ope a o alues.
0 5 10 15 20 25 30
I e a ions
0.84
0.86
0.88
0.90
0.92
0.94
0.96
0.98
1.00
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.15: Me ic e olu ion AUC MOD in Failu e ype C in TSM3
o all combina ion o HS ope a o alues.
222
A.1. Ha mony Sea ch ope a o s analysis
0 5 10 15 20 25 30
I e a ions
0.84
0.86
0.88
0.90
0.92
0.94
0.96
0.98
1.00
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.16: Me ic e olu ion AUC MOD in Failu e ype C in TSM4
o all combina ion o HS ope a o alues.
0 5 10 15 20 25 30
I e a ions
0.800
0.825
0.850
0.875
0.900
0.925
0.950
0.975
1.000
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.17: Me ic e olu ion AUC MOD in Failu e ype C in TSM5
o all combina ion o HS ope a o alues.
223
Appendix A. Resul s o Chap e 6
0 5 10 15 20 25 30
I e a ions
0.70
0.75
0.80
0.85
0.90
0.95
1.00
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.18: Me ic e olu ion AUC MOD in Failu e ype C in TSM6
o all combina ion o HS ope a o alues.
A.1.4 Type D
0 5 10 15 20 25 30
I e a ions
0.80
0.85
0.90
0.95
1.00
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.19: Me ic e olu ion AUC MOD in Failu e ype D in TSM1
o all combina ion o HS ope a o alues.
224
A.1. Ha mony Sea ch ope a o s analysis
0 5 10 15 20 25 30
I e a ions
0.70
0.75
0.80
0.85
0.90
0.95
1.00
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.20: Me ic e olu ion AUC MOD in Failu e ype D in TSM2
o all combina ion o HS ope a o alues.
0 5 10 15 20 25 30
I e a ions
0.70
0.75
0.80
0.85
0.90
0.95
1.00
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.21: Me ic e olu ion AUC MOD in Failu e ype D in TSM3
o all combina ion o HS ope a o alues.
225

Appendix A. Resul s o Chap e 6
0 5 10 15 20 25 30
I e a ions
0.86
0.88
0.90
0.92
0.94
0.96
0.98
1.00
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.22: Me ic e olu ion AUC MOD in Failu e ype D in TSM4
o all combina ion o HS ope a o alues.
0 5 10 15 20 25 30
I e a ions
0.64
0.65
0.66
0.67
0.68
0.69
0.70
0.71
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.23: Me ic e olu ion AUC MOD in Failu e ype D in TSM5
o all combina ion o HS ope a o alues.
226
A.1. Ha mony Sea ch ope a o s analysis
0 5 10 15 20 25 30
I e a ions
0.80
0.85
0.90
0.95
1.00
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.24: Me ic e olu ion AUC MOD in Failu e ype D in TSM6
o all combina ion o HS ope a o alues.
A.1.5 Type E
0 5 10 15 20 25 30
I e a ions
0.75
0.80
0.85
0.90
0.95
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.25: Me ic e olu ion AUC MOD in Failu e ype E in TSM1
o all combina ion o HS ope a o alues.
227
Appendix A. Resul s o Chap e 6
0 5 10 15 20 25 30
I e a ions
0.88
0.90
0.92
0.94
0.96
0.98
1.00
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.26: Me ic e olu ion AUC MOD in Failu e ype E in TSM2
o all combina ion o HS ope a o alues.
0 5 10 15 20 25 30
I e a ions
0.75
0.80
0.85
0.90
0.95
1.00
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.27: Me ic e olu ion AUC MOD in Failu e ype E in TSM3
o all combina ion o HS ope a o alues.
228
A.1. Ha mony Sea ch ope a o s analysis
0 5 10 15 20 25 30
I e a ions
0.825
0.850
0.875
0.900
0.925
0.950
0.975
1.000
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.28: Me ic e olu ion AUC MOD in Failu e ype E in TSM4
o all combina ion o HS ope a o alues.
0 5 10 15 20 25 30
I e a ions
0.88
0.90
0.92
0.94
0.96
0.98
1.00
Fi ness a e age o 10 mon eca lo
['0.5', '0.1', '0.1']
['0.5', '0.1', '0.3']
['0.5', '0.3', '0.1']
['0.5', '0.3', '0.3']
['0.5', '0.5', '0.1']
['0.5', '0.5', '0.3']
['0.7', '0.1', '0.1']
['0.7', '0.1', '0.3']
['0.7', '0.3', '0.1']
['0.7', '0.3', '0.3']
['0.7', '0.5', '0.1']
['0.7', '0.5', '0.3']
['0.9', '0.1', '0.1']
['0.9', '0.1', '0.3']
['0.9', '0.3', '0.1']
['0.9', '0.3', '0.3']
['0.9', '0.5', '0.1']
['0.9', '0.5', '0.3']
Figu e A.29: Me ic e olu ion AUC MOD in Failu e ype E in TSM5
o all combina ion o HS ope a o alues.
229
Bibliog aphy
[6] V. Chandola, A. Bane jee, and V. Kuma , “Anomaly de ec ion:
A su ey,” ACM compu ing su eys (CSUR), ol. 41, no. 3, pp.
1–58, 2009.
[7] H. Lasi, P. Fe ke, H.-G. Kempe , T. Feld, and M. Ho mann,
“Indus y 4.0,” Business & in o ma ion sys ems enginee ing,
ol. 6, no. 4, pp. 239–242, 2014.
[8] M. B eque, L. De Nul, A. Pe idis e al., “Indus y 5.0: Towa ds
a sus ainable, human-cen ic and esilien eu opean indus y,”
2021.
[9] A. G. Gonzalez, D. R. Quinone o, and S. F. Vega, “Assessmen
o he deg ee o implemen a ion o indus y 4.0 echnologies:
Case s udy o mu cia egion in sou heas spain,” Enginee ing
Economics, ol. 32, no. 5, pp. 422–432, 2021.
[10] Y. Ko en, The global manu ac u ing e olu ion: p oduc -p ocess-
business in eg a ion and econ igu able sys ems. John Wiley &
Sons, 2010.
[11] R. N. Langlois and P. L. Robe son, “Explaining e ical in e-
g a ion: Lessons om he ame ican au omobile indus y,” The
Jou nal o Economic His o y, ol. 49, no. 2, pp. 361–375, 1989.
[12] M. T. Okano, “Io and indus y 4.0: he indus ial new e olu-
ion,” in In e na ional Con e ence on Managemen and In o -
ma ion Sys ems, ol. 25, 2017, p. 26.
[13] X. Xu, Y. Lu, B. Vogel-Heuse , and L. Wang, “Indus y 4.0 and
indus y 5.0—incep ion, concep ion and pe cep ion,” Jou nal o
Manu ac u ing Sys ems, ol. 61, pp. 530–535, 2021.
[14] P. K. R. Maddikun a, Q.-V. Pham, B. P abade i, N. Deepa,
K. De , T. R. Gadekallu, R. Ruby, and M. Liyanage, “Indus y
5.0: A su ey on enabling echnologies and po en ial applica-
ions,” Jou nal o Indus ial In o ma ion In eg a ion, ol. 26, p.
100257, 2022.
[15] J. Gan z, D. Reinsel e al., “Ex ac ing alue om chaos,” IDC
i iew, ol. 1142, no. 2011, pp. 1–12, 2011.
[16] J. Lee, B. Baghe i, and H.-A. Kao, “A cybe -physical sys ems
a chi ec u e o indus y 4.0-based manu ac u ing sys ems,”
Manu ac u ing le e s, ol. 3, pp. 18–23, 2015.
236

Bibliog aphy
[17] P. Flach, Machine lea ning: he a and science o algo i hms
ha make sense o da a. Camb idge uni e si y p ess, 2012.
[18] R. Azuma, Y. Baillo , R. Beh inge , S. Feine , S. Julie , and
M. Blai , “Recen ad ances in augmen ed eali y. ieee compu e
g aphics and applica ions,” IEEE Compu e G aphics and Ap-
plica ions, ol. 21, no. 6, pp. 34–47, 2001.
[19] A. K. Ja dine, D. Lin, and D. Banje ic, “A e iew on ma-
chine y diagnos ics and p ognos ics implemen ing condi ion-
based main enance,” Mechanical sys ems and signal p ocessing,
ol. 20, no. 7, pp. 1483–1510, 2006.
[20] T. P. Ca alho, F. A. Soa es, R. Vi a, R. d. P. F ancisco, J. P.
Bas o, and S. G. Alcal´a, “A sys ema ic li e a u e e iew o
machine lea ning me hods applied o p edic i e main enance,”
Compu e s & Indus ial Enginee ing, ol. 137, p. 106024, 2019.
[21] J. Dalzochio, R. Kuns , E. Pigna on, A. Bino o, S. Sanyal,
J. Fa illa, and J. Ba bosa, “Machine lea ning and easoning
o p edic i e main enance in indus y 4.0: Cu en s a us and
challenges,” Compu e s in Indus y, ol. 123, p. 103298, 2020.
[22] M. Paolan i, L. Romeo, A. Felice i, A. Mancini, E. F on oni,
and J. Lonca ski, “Machine lea ning app oach o p edic i e
main enance in indus y 4.0,” in 2018 14 h IEEE/ASME In-
e na ional Con e ence on Mecha onic and Embedded Sys ems
and Applica ions (MESA). IEEE, 2018, pp. 1–6.
[23] F. A ellano-Espi ia, M. Delgado-P ie o, A.-D. Gonzalez-Ab eu,
J. J. Saucedo-Do an es, and R. A. Oso nio-Rios, “Deep-
compac -clus e ing based anomaly de ec ion applied o elec-
omechanical indus ial sys ems,” Senso s, ol. 21, no. 17, p.
5830, 2021.
[24] P. Czop, G. Kos , D. S lawik, and G. Wszo lek, “Fo mula ion
and iden i ica ion o i s -p inciple da a-d i en models,” Jou nal
o Achie emen s in ma e ials and manu ac u ing Enginee ing,
ol. 44, no. 2, pp. 179–186, 2011.
[25] D. An, N. H. Kim, and J.-H. Choi, “P ac ical op ions o selec -
ing da a-d i en o physics-based p ognos ics algo i hms wi h
e iews,” Reliabili y Enginee ing & Sys em Sa e y, ol. 133, pp.
223–236, 2015.
237
Bibliog aphy
[26] G. Fo es ie and C. Wemme , “Semi-supe ised lea ning us-
ing mul iple clus e ings wi h limi ed labeled da a,” In o ma ion
Sciences, ol. 361, pp. 48–65, 2016.
[27] X. J. Zhu, “Semi-supe ised lea ning li e a u e su ey,” 2005.
[28] E. Swana and W. Doo samy, “An unsupe ised lea ning ap-
p oach o condi ion assessmen on a wound- o o induc ion gen-
e a o ,” Ene gies, ol. 14, no. 3, p. 602, 2021.
[29] J. Chen, C. Lu, and H. Yuan, “Bea ing aul diagnosis based on
ac i e lea ning and andom o es ,” Vib oenginee ing PROCE-
DIA, ol. 5, pp. 321–326, 2015.
[30] N. Pi oonsup and S. Sin hupinyo, “Analysis o aining da a
using clus e ing o imp o e semi-supe ised sel - aining,”
Knowledge-Based Sys ems, ol. 143, pp. 65–80, 2018.
[31] J. E. Van Engelen and H. H. Hoos, “A su ey on semi-supe ised
lea ning,” Machine Lea ning, ol. 109, no. 2, pp. 373–440, 2020.
[32] G. Kos opoulos, S. Ka los, S. Ko sian is, and O. Ragos, “Semi-
supe ised eg ession: A ecen e iew,” Jou nal o In elligen
& Fuzzy Sys ems, ol. 35, no. 2, pp. 1483–1500, 2018.
[33] E. Bai , “Semi-supe ised clus e ing me hods,” Wiley In e dis-
ciplina y Re iews: Compu a ional S a is ics, ol. 5, no. 5, pp.
349–361, 2013.
[34] Y. Qin, S. Ding, L. Wang, and Y. Wang, “Resea ch p og ess
on semi-supe ised clus e ing,” Cogni i e Compu a ion, ol. 11,
no. 5, pp. 599–612, 2019.
[35] T. Van C aenendonck, W. Mee , S. Dumanˇci´c, and H. Blockeel,
“Cob as s: A new app oach o semi-supe ised clus e ing o
ime se ies,” in In e na ional Con e ence on Disco e y Science.
Sp inge , 2018, pp. 179–193.
[36] N. G i a, M. C ucianu, and N. Boujemaa, “Unsupe ised and
semi-supe ised clus e ing: a b ie su ey,” A e iew o machine
lea ning echniques o p ocessing mul imedia con en , ol. 1, pp.
9–16, 2004.
[37] T. Yang, N. Pasquie , and F. P ecioso, “Semi-supe ised con-
sensus clus e ing based on closed pa e ns,” Knowledge-Based
Sys ems, ol. 235, p. 107599, 2022.
238
Bibliog aphy
[38] P. P lo´nski and K. Za emba, “Full and semi-supe ised k-means
clus e ing op imised by class membe ship hesi a ion,” in In e -
na ional Con e ence on Adap i e and Na u al Compu ing Algo-
i hms. Sp inge , 2013, pp. 218–225.
[39] S. Basu, M. Bilenko, and R. J. Mooney, “A p obabilis ic ame-
wo k o semi-supe ised clus e ing,” in P oceedings o he en h
ACM SIGKDD in e na ional con e ence on Knowledge disco -
e y and da a mining, 2004, pp. 59–68.
[40] I. S ojano i´c, I. B aje i´c, P. S. S animi o i´c, L. A. Kazako -
se , and Z. Zd a e , “Applica ion o heu is ic and me aheu is ic
algo i hms in sol ing cons ained webe p oblem wi h easible
egion bounded by a cs,” Ma hema ical P oblems in Enginee -
ing, ol. 2017, 2017.
[41] S. Basu, A. Bane jee, and R. Mooney, “Semi-supe ised clus-
e ing by seeding,” in In P oceedings o 19 h In e na ional Con-
e ence on Machine Lea ning (ICML-2002. Ci esee , 2002.
[42] L. Lelis and J. Sande , “Semi-supe ised densi y-based clus e -
ing,” in 2009 Nin h IEEE In e na ional Con e ence on Da a
Mining. IEEE, 2009, pp. 842–847.
[43] L. Zheng and T. Li, “Semi-supe ised hie a chical clus e ing,”
in 2011 IEEE 11 h In e na ional Con e ence on Da a Mining.
IEEE, 2011, pp. 982–991.
[44] M. Blocho, “Heu is ics, me aheu is ics, and hype heu is ics o
ich ehicle ou ing p oblems,” in Sma Deli e y Sys ems. El-
se ie , 2020, pp. 101–156.
[45] M. Abdel-Basse , L. Abdel-Fa ah, and A. K. Sangaiah, “Me a-
heu is ic algo i hms: A comp ehensi e e iew,” Compu a ional
in elligence o mul imedia big da a on he cloud wi h enginee -
ing applica ions, pp. 185–231, 2018.
[46] T. Doke oglu, E. Se inc, T. Kucukyilmaz, and A. Cosa , “A
su ey on new gene a ion me aheu is ic algo i hms,” Compu e s
& Indus ial Enginee ing, ol. 137, p. 106040, 2019.
[47] A. A. Ba a’a, A. D. Abbood, A. A. Hasan, C. Pizzu i, M. Al-
Ani, S. ¨
Ozdemi , and R. D. Al-Dabbagh, “A e iew o heu is ics
239
Bibliog aphy
and me aheu is ics o communi y de ec ion in complex ne -
wo ks: Cu en usage, eme ging de elopmen and u u e di-
ec ions,” Swa m and E olu iona y Compu a ion, ol. 63, p.
100885, 2021.
[48] J. Xiong, X. Liu, X. Zhu, H. Zhu, H. Li, and Q. Zhang, “Semi-
supe ised uzzy c-means clus e ing op imized by simula ed an-
nealing and gene ic algo i hm o aul diagnosis o bea ings,”
IEEE Access, ol. 8, pp. 181 976–181 987, 2020.
[49] R. Ko ha i and V. Jain, “Lea ning om labeled and unlabeled
da a using a minimal numbe o que ies,” IEEE T ansac ions
on Neu al Ne wo ks, ol. 14, no. 6, pp. 1496–1505, 2003.
[50] A. K. Alok, S. Saha, and A. Ekbal, “A new semi-supe ised
clus e ing echnique using mul i-objec i e op imiza ion,” Ap-
plied In elligence, ol. 43, no. 3, pp. 633–661, 2015.
[51] A. Na ajas-Gue e o, D. Manja es, E. Po illo, and I. Landa-
To es, “A hype -heu is ic inspi ed app oach o au oma ic ail-
u e p edic ion in he con ex o indus y 4.0,” Compu e s &
Indus ial Enginee ing, p. 108381, 2022.
[52] E. Bu ke, G. Kendall, J. Newall, E. Ha , P. Ross, and S. Schu-
lenbu g, “Hype -heu is ics: An eme ging di ec ion in mode n
sea ch echnology,” in Handbook o me aheu is ics. Sp inge ,
2003, pp. 457–474.
[53] A. Swie cz, “Hype -heu is ics and me aheu is ics o selec ed
bio-inspi ed combina o ial op imiza ion p oblems,” Heu is ics
and Hype -Heu is ics-P inciples and Applica ions, 2017.
[54] E. K. Bu ke, M. Gend eau, M. Hyde, G. Kendall, G. Ochoa,
E. ¨
Ozcan, and R. Qu, “Hype -heu is ics: A su ey o he s a e
o he a ,” Jou nal o he Ope a ional Resea ch Socie y, ol. 64,
no. 12, pp. 1695–1724, 2013.
[55] E. K. Bu ke, M. Hyde, G. Kendall, G. Ochoa, E. ¨
Ozcan,
and J. R. Woodwa d, “A classi ica ion o hype -heu is ic ap-
p oaches,” in Handbook o me aheu is ics. Sp inge , 2010, pp.
449–468.
[56] J. H. D ake, A. Khei i, E. ¨
Ozcan, and E. K. Bu ke, “Recen ad-
ances in selec ion hype -heu is ics,” Eu opean Jou nal o Op-
e a ional Resea ch, ol. 285, no. 2, pp. 405–428, 2020.
240
Bibliog aphy
[57] K. Chakhle i ch and P. Cowling, Hype heu is ics: Recen De-
elopmen s, 06 2008, ol. 136, pp. 3–29.
[58] M. Leng, X. Chen, L. Li, M. Leng e al., “K-means clus e ing
algo i hm based on semi-supe ised lea ning,” 2008.
[59] X. Wang, C. Wang, and J. Shen, “Semi–supe ised k-means
clus e ing by op imizing ini ial clus e cen e s,” in In e na ional
con e ence on web in o ma ion sys ems and mining. Sp inge ,
2011, pp. 178–187.
[60] J. Li, J. Sande , R. Campello, and A. Zimek, “Ac i e lea ning
s a egies o semi-supe ised dbscan,” in Canadian Con e ence
on A i icial In elligence. Sp inge , 2014, pp. 179–190.
[61] V. Maca io and F. d. A. de Ca alho, “An adap i e semi-
supe ised uzzy clus e ing algo i hm based on objec i e unc-
ion op imiza ion,” in 2012 IEEE In e na ional Con e ence on
Fuzzy Sys ems. IEEE, 2012, pp. 1–8.
[62] H. Gan, Y. Fan, Z. Luo, R. Huang, and Z. Yang, “Con idence-
weigh ed sa e semi-supe ised clus e ing,” Enginee ing Applica-
ions o A i icial In elligence, ol. 81, pp. 107–116, 2019.
[63] C. Ch ysouli and A. Te as, “Spec al clus e ing and semi-
supe ised lea ning using e ol ing simila i y g aphs,” Applied
So Compu ing, ol. 34, pp. 625–637, 2015.
[64] A. Demi iz, K. Benne , and M. Emb ech s, “Semi-supe ised
clus e ing using gene ic algo i hms,” A i . Neu al Ne w. Eng,
09 1999.
[65] S. Saha, A. Ekbal, and A. K. Alok, “Semi-supe ised clus e ing
using mul iobjec i e op imiza ion,” in 2012 12 h In e na ional
Con e ence on Hyb id In elligen Sys ems (HIS). IEEE, 2012,
pp. 360–365.
[66] C. Ruiz, M. Spiliopoulou, and E. Menasal as, “C-dbscan:
Densi y-based clus e ing wi h cons ain s,” in In e na ional
wo kshop on ough se s, uzzy se s, da a mining, and g anula -
so compu ing. Sp inge , 2007, pp. 216–223.
[67] K. Wags a , C. Ca die, S. Roge s, S. Sch ¨odl e al., “Con-
s ained k-means clus e ing wi h backg ound knowledge,” in
Icml, ol. 1, 2001, pp. 577–584.
241

Bibliog aphy
[68] M. Okabe and S. Yamada, “Clus e ing using boos ed con-
s ained k-means algo i hm,” F on ie s in Robo ics and AI,
ol. 5, p. 18, 2018.
[69] P. Ra ho e, J. C. Bezdek, P. San i, and C. Ra i, “Coni a :
Clus e endency assessmen and clus e ing wi h pa ial back-
g ound knowledge,” a Xi p ep in a Xi :2008.09570, 2020.
[70] V. Melnyko , I. Melnyko , and S. Michael, “Semi-supe ised
model-based clus e ing wi h posi i e and nega i e cons ain s,”
Ad ances in da a analysis and classi ica ion, ol. 10, no. 3, pp.
327–349, 2016.
[71] A. Vou os and E. Vasilaki, “A semi-supe ised spa se k-means
algo i hm,” Pa e n Recogni ion Le e s, ol. 142, pp. 65–71,
2021.
[72] S. Basu, A. Bane jee, and R. J. Mooney, “Ac i e semi-
supe ision o pai wise cons ained clus e ing,” in P oceedings
o he 2004 SIAM in e na ional con e ence on da a mining.
SIAM, 2004, pp. 333–344.
[73] H. Guo, J. Ma, and Z. Li, “Ac i e semi-supe ised k-means clus-
e ing based on silhoue e coe icien ,” in In e na ional Con e -
ence on In elligen and In e ac i e Sys ems and Applica ions.
Sp inge , 2018, pp. 202–209.
[74] D. Tiano, A. Boni a i, and R. Ng, “Fea s: Fea u e-based ime
se ies clus e ing,” in P oceedings o he 2021 In e na ional Con-
e ence on Managemen o Da a, 2021, pp. 2784–2788.
[75] ——, “Fea u e-d i en ime se ies clus e ing.” in 24 h In e na-
ional Con e ence on Ex ending Da abase Technology, EDBT
2021, 2021, pp. 349–354.
[76] G. He, Y. Pan, X. Xia, J. He, R. Peng, and N. N. Xiong, “A as
semi-supe ised clus e ing amewo k o la ge-scale ime se ies
da a,” IEEE T ansac ions on Sys ems, Man, and Cybe ne ics:
Sys ems, ol. 51, no. 7, pp. 4201–4216, 2019.
[77] J. Zhou, S.-F. Zhu, X. Huang, and Y. Zhang, “Enhancing ime
se ies clus e ing by inco po a ing mul iple dis ance measu es
wi h semi-supe ised lea ning,” Jou nal o Compu e Science
and Technology, ol. 30, no. 4, pp. 859–873, 2015.
242
Bibliog aphy
[78] H. A. Dau, N. Begum, and E. Keogh, “Semi-supe ision d a-
ma ically imp o es ime se ies clus e ing unde dynamic ime
wa ping,” in P oceedings o he 25 h ACM In e na ional on Con-
e ence on In o ma ion and Knowledge Managemen , 2016, pp.
999–1008.
[79] J. Tan, W. Fu, K. Wang, X. Xue, W. Hu, and Y. Shan, “Faul
diagnosis o olling bea ing based on semi-supe ised clus e ing
and suppo ec o da a desc ip ion wi h adap i e pa ame e
op imiza ion and imp o ed decision s a egy,” Applied Sciences,
ol. 9, no. 8, p. 1676, 2019.
[80] A. A iandiaga, E. Po illo, J. A. S´anchez, I. Cabanes, and
I. Pombo, “Vi ual senso s o on-line wheel wea and pa
oughness measu emen in he g inding p ocess,” Senso s,
ol. 14, no. 5, pp. 8756–8778, 2014.
[81] R.-M. Hage, I. Hage, C. Ghna ios, I. Jawahi , and R. Hamade,
“Op imized abu sea ch es ima ion o wea cha ac e is ics and
cu ing o ces in compac co e d illing o basal ock using pcd
ool inse s,” Compu e s & Indus ial Enginee ing, ol. 136, pp.
477–493, 2019.
[82] A. Ha ami-Ma bini, S. M. Sajadi, and H. Malekpou , “Op i-
mal con ol and simula ion o p oduc ion planning o ne wo k
ailu e-p one manu ac u ing sys ems wi h pe ishable goods,”
Compu e s & Indus ial Enginee ing, ol. 146, p. 106614, 2020.
[83] W. J. Lee, G. P. Mendis, M. J. T iebe, and J. W. Su he land,
“Moni o ing o a machining p ocess using ke nel p incipal com-
ponen analysis and ke nel densi y es ima ion,” Jou nal o In-
elligen Manu ac u ing, ol. 31, no. 5, pp. 1175–1189, 2020.
[84] E. Po illo, M. Ma cos, I. Cabanes, and D. O i e, “Real- ime
moni o ing and diagnosing in wi e-elec o discha ge machining,”
The In e na ional Jou nal o Ad anced Manu ac u ing Technol-
ogy, ol. 44, no. 3-4, pp. 273–282, 2009.
[85] M. Hu, Z. Ji, K. Yan, Y. Guo, X. Feng, J. Gong, X. Zhao, and
L. Dong, “De ec ing anomalies in ime se ies da a ia a me a-
ea u e based app oach,” IEEE Access, ol. 6, pp. 27 760–27 776,
2018.
243
Bibliog aphy
[86] A. H. Gandomi, X.-S. Yang, S. Tala aha i, and A. H. Ala i,
Me aheu is ic applica ions in s uc u es and in as uc u es.
Newnes, 2013.
[87] M. A. D´ıaz-Co ´es, E. Cue as, J. G´al ez, and O. Cama ena,
“A new me aheu is ic op imiza ion me hodology based on uzzy
logic,” Applied So Compu ing, ol. 61, pp. 549–569, 2017.
[88] E. ¨
Ozcan, B. Bilgin, and E. E. Ko kmaz, “A comp ehensi e
analysis o hype -heu is ics,” In elligen da a analysis, ol. 12,
no. 1, pp. 3–23, 2008.
[89] N. Lap e , S. Amizadeh, and I. Flin , “Gene ic and scalable
amewo k o au oma ed ime-se ies anomaly de ec ion,” in
P oceedings o he 21 h ACM SIGKDD In e na ional Con e -
ence on Knowledge Disco e y and Da a Mining, 2015, pp. 1939–
1947.
[90] W. Lu, Y. Li, Y. Cheng, D. Meng, B. Liang, and P. Zhou, “Ea ly
aul de ec ion app oach wi h deep a chi ec u es,” IEEE T ans-
ac ions on Ins umen a ion and Measu emen , ol. 67, no. 7, pp.
1679–1689, 2018.
[91] H. Cai, J. Feng, J. Moyne, J. Iskanda , M. A macos , F. Li, and
J. Lee, “A amewo k o semi-au oma ed aul de ec ion con ig-
u a ion wi h au oma ed ea u e ex ac ion and limi s se ing,”
in 2020 31s Annual SEMI Ad anced Semiconduc o Manu ac-
u ing Con e ence (ASMC). IEEE, 2020, pp. 1–6.
[92] F. Calab ese, A. Rega ie i, L. Bo i, C. Mo a, and F. G. Gal-
izia, “Unsupe ised aul de ec ion and p edic ion o emaining
use ul li e o online p ognos ic heal h managemen o mechan-
ical sys ems,” Applied Sciences, ol. 10, no. 12, p. 4120, 2020.
[93] J. A. Ca ino, M. Delgado-P ie o, J. A. Iglesias, A. Sanchis,
D. Zu i a, M. Millan, J. A. O. Redondo, and R. Rome o-
T oncoso, “Faul de ec ion and iden i ica ion me hodology un-
de an inc emen al lea ning amewo k applied o indus ial ma-
chine y,” IEEE access, ol. 6, pp. 49 755–49 766, 2018.
[94] H. Izakian and W. Ped ycz, “Anomaly de ec ion in ime
se ies da a using a uzzy c-means clus e ing,” in 2013
Join IFSA Wo ld Cong ess and NAFIPS Annual Mee ing
(IFSA/NAFIPS). IEEE, 2013, pp. 1513–1518.
244
Bibliog aphy
[95] L. Xie, S. H˚ab ekke, Y. Liu, and M. A. Lund eigen, “Ope a-
ional da a-d i en p edic ion o ailu e a es o equipmen in
sa e y ins umen ed sys ems: A case s udy om he oil and gas
indus y,” Jou nal o Loss P e en ion in he P ocess Indus ies,
ol. 60, pp. 96–105, 2019.
[96] J. Yu, “Faul de ec ion using p incipal componen s-based gaus-
sian mix u e model o semiconduc o manu ac u ing p o-
cesses,” IEEE T ansac ions on Semiconduc o Manu ac u ing,
ol. 24, no. 3, pp. 432–444, 2011.
[97] M. A. A oui, S. Ve on, and A. Kobi, “Faul de ec ion wi h
condi ional gaussian ne wo k,” Enginee ing Applica ions o A -
i icial In elligence, ol. 45, pp. 473–481, 2015.
[98] H. Chen, B. Jiang, W. Chen, and Z. Li, “Edge compu ing-
aided amewo k o aul de ec ion o ac ion con ol sys ems
in high-speed ains,” IEEE T ansac ions on Vehicula Tech-
nology, ol. 69, no. 2, pp. 1309–1318, 2019.
[99] X. Li, X. Yang, Y. Yang, I. Benne , and D. Mba, “A no el
diagnos ic and p ognos ic amewo k o incipien aul de ec-
ion and emaining se ice li e p edic ion wi h applica ion o
indus ial o a ing machines,” Applied So Compu ing, ol. 82,
p. 105564, 2019.
[100] J. Liu, Y.-F. Li, and E. Zio, “A s m amewo k o aul de ec-
ion o he b aking sys em in a high speed ain,” Mechanical
Sys ems and Signal P ocessing, ol. 87, pp. 401–409, 2017.
[101] B. Luo, H. Wang, H. Liu, B. Li, and F. Peng, “Ea ly aul de ec-
ion o machine ools based on deep lea ning and dynamic iden i-
ica ion,” IEEE T ansac ions on Indus ial Elec onics, ol. 66,
no. 1, pp. 509–518, 2018.
[102] D. Zang, J. Liu, and H. Wang, “Ma ko chain-based ea u e
ex ac ion o anomaly de ec ion in ime se ies and i s indus ial
applica ion,” in 2018 Chinese Con ol And Decision Con e ence
(CCDC). IEEE, 2018, pp. 1059–1063.
[103] J. Chen, W. Hu, D. Cao, B. Zhang, Q. Huang, Z. Chen,
and F. Blaabje g, “An imbalance aul de ec ion algo i hm o
a iable-speed wind u bines: A deep lea ning app oach,” En-
e gies, ol. 12, no. 14, p. 2764, 2019.
245