Anomaly Detections for Manufacturing Systems Based on Sensor Data—Insights into Two Challenging Real-World Production Settings

senso s A icle Anomaly De ec ions o Manu ac u ing Sys ems Based on Senso Da a—Insigh s in o Two Challenging Real-Wo ld P oduc ion Se ings Klaus Kamme e 1,* , Bu kha d Hoppens ed 1, Rüdige P yss 2, S e en S ökle 3, Johannes Allgaie 4and Man ed Reiche 1 1Ins i u e o Da abases and In o ma ion Sys em, Uni e si y o Ulm, 89081 Ulm, Ge many; [email p o ec ed] (B.H.); man [email p o ec ed] (M.R.) 2Ins i u e o Clinical Epidemiology and Biome y, Uni e si y o Wü zbu g, 97080 Wü zbu g, Ge many; uedige [email p o ec ed] 3Uhlmann Pac-Sys eme GmbH & Co. KG, 88471 Laupheim, Ge many; s oekle [email p o ec ed] 4ATR So wa e GmbH, 89231 Neu-Ulm, Ge many; [email p o ec ed] *Co espondence: klaus.kamme [email p o ec ed] Recei ed: 9 Oc obe 2019; Accep ed: 2 Decembe 2019; Published: 5 Decembe 2019   Abs ac : To build, un, and main ain eliable manu ac u ing machines, he condi ion o hei componen s has o be con inuously moni o ed. When ollowing a ine-g ained moni o ing o hese machines, challenges eme ge pe aining o he (1) eeding p ocedu e o la ge amoun s o senso da a o downs eam p ocessing componen s and he (2) meaning ul analysis o he p oduced da a. Rega ding he la e aspec , mani old pu poses a e add essed by p ac i ione s and esea che s. Two analyses o eal-wo ld da ase s ha we e gene a ed in p oduc ion se ings a e discussed in his pape . Mo e speci ically, he analyses had he goals (1) o de ec senso da a anomalies o u he analyses o a pha ma packaging scena io and (2) o p edic un a o able empe a u e alues o a 3D p in ing machine en i onmen . Based on he esul s o he analyses, i will be shown ha a p ope managemen o machines and hei componen s in indus ial manu ac u ing en i onmen s can be e icien ly suppo ed by he de ec ion o anomalies. The la e shall help o suppo he echnical e angelis s o he p oduc ion companies mo e p ope ly. Keywo ds: anomaly de ec ion; senso da a; machine lea ning; p oduc ion machines 1. In oduc ion Fo manu ac u ing companies, he managemen o machine ailu es is becoming inc easingly impo an . Due o he inc easing complexi y o he machines, down imes o any kind can a ec he o e all success o a company. Buying a spa e sys em is no an al e na i e solu ion as he acquisi ion cos s no mally su pass he bene i o he spa e sys em. In addi ion, mos eplacemen sys ems also equi e egula main enance, e en i hey a e no used. Tha is why companies a e looking o new ways o manage machine ailu es cos -e ec i ely. P edic i e Main enance is one opical subjec , among o he s, which is a p omising di ec ion o ackle machine ailu es be o e hey ac ually occu . Howe e , he selec ion o app op ia e echniques om he ield o P edic i e Main enance is challenging as nume ous aspec s ha e o be conside ed [ 1 ]. In addi ion o P edic i e Main enance, he e a e many o he app oaches in his con ex o coping wi h machine ailu es such as Condi ion Moni o ing o Con inuous Imp o emen s. Mo eo e , he end owa ds machine lea ning aises he ques ion o whe he machine ailu es can be easily p edic ed. Al hough echnical de elopmen s ha e imp o ed he possibili ies o manu ac u ing companies o cope wi h machine b eakdowns, hei p ac ical applica ion is s ill a challenging ask o many easons. On he basis o hese conside a ions, he Senso s 2019,19, 5370; doi:10.3390/s19245370 www.mdpi.com/jou nal/senso s Senso s 2019,19, 5370 2 o 18 wo k a hand p esen s wo eal-wo ld cases ha we e ca ied ou in coope a ion wi h manu ac u ing companies. Fo hese companies, he de ec ion o sys em e o s is o u mos impo ance. In his con ex , i was shown ha he de ec ion o anomalies [ 2 ] o a machine is c ucial o hese companies. Howe e , he meaning ul de ec ion o such anomalies is e y complex. In e es ingly, so a , many manu ac u ing companies o en employ a selec ed choice o echnical e angelis s ha a e only able o de ec anomalies based on hei p ac ical expe iences o e ime. Such expe s, in u n, a e e y expensi e by design. To elie e hem om manual decisions, his pape elabo a es on how machine ailu es o hese scena ios can be managed by analyzing senso da a o he p oduc ion machines. In pa icula , he wo examples will show ha di e en ypes o senso da a, as well as de ec ion echniques, should be conside ed. The i s p esen ed eal-wo ld se ing is ela ed o pha ma packing machines. The la e machines w ap able s in o indi idual packaging uni s (i.e., blis e s), and usually comp ise se e al o he componen s. Fo example, a p oduc loade componen pushes blis e s and lea le s in o ca ons. This p ocedu e, in u n, is p one o e o s. The e o e, he packaging p ocess needs o be con inuously moni o ed o educe cos ly down imes as well as o comply wi h ede al egula ions. The con inuous moni o ing p ocedu e, in u n, gene a es a la ge amoun o senso da a coming om senso s ha a e ela ed o he se e al componen s o he packaging machine. In addi ion, he pha ma packing machine can be indi idualized o each cus ome , which migh lead o many senso pa ame e se ings o he same machine ype, including all obscu ed componen s. The second p esen ed eal-wo ld se ing is ela ed o 3D p in ing machines in he ield o op ical p oduc s. In con as o he i s example, in he second se ing, no only senso alues om he machine i sel a e impo an , bu also senso s ha measu e he en i onmen o he machine. No e ha his ac dis inguishes his scena io om he i s one. He e, o example, an inc ease in he en i onmen al empe a u e o humidi y may ha e a signi ican in luence on he p oduc ion p ocess o he manu ac u ing machine. Anomalies, such as oom empe a u e ha has nega i e e ec s on he machines, should, he e o e, be a oided. Rega ding he anomaly de ec ion in gene al, plen y o algo i hms we e p oposed ha add ess di e en use cases [ 3 ]. The main objec i e o hese algo i hms is o analyze s eaming da a o de elop models ha can be used wi h an app op ia e numbe o pa ame e s and ac oss applica ions, i.e., o machines wi h di e en cha ac e is ics and pu poses. As in mos p ac ical cases, a e y la ge numbe o senso alues mus be analyzed simul aneously, de ailed knowledge o he indi idual pa ame e s is a leas mos ly o seconda y impo ance. Fo manu ac u e s, anomaly de ec ion algo i hms wi h a minimum o pa ame e se ings should be easy o apply and sui able o s eaming analysis [ 4 ]. Consequen ly, his ype o algo i hms should ea u e ma hema ical ope a ions wi h low compu a ional complexi y in o de o minimize la ency in long- unning s eaming analyses. Based on he wo eal-wo ld examples and hei di e en cha ac e is ics o p edic anomalies o p oduc ion machines, his wo k con ibu es o he ollowing majo insigh s (see also Figu e 1): • Fo he pha ma packaging machine scena io, senso s eaming da a was e alua ed wi h he goal o p edic anomalies o one pa icula machine ype. Mo e p ecisely, machine-in e nal senso s (see Figu e 1a) a e acqui ed and hei da a is ansmi ed o a senso s da a p ocessing se ice. The la e hen e alua es he senso da a wi h a dis ance p o iling me hod (see Figu e 1b). Fo his use case, i is shown ha he equi emen o e icien anomaly de ec ion me hods was ound, which is also able o cope wi h he huge amoun o senso da a o he analyzed packaging machine. Following his, he pha ma packaging company can add ess o he machine ypes as well. • Fo he 3D machine scena io, en i onmen al senso da a was e alua ed wi h he goal o p edic anomalies ha may a ec he p oduc ion machines. Mo e p ecisely, empe a u e senso s (see Figu e 1d) send da a o a message b oke ia he Message Queuing Teleme y T anspo p o ocol (MQTT) (see Figu e 1c). A machine lea ning se ice (see Figu e 1e) hen p edic s empe a u e alues ha may ha e nega i e e ec s. I a empe a u e anomaly is de ec ed, he senso da a p ocessing and he machine lea ning se ices no i y a machine ope a o ia a se ice message using he message b oke . In his use case, se e al in es iga ed machine lea ning app oaches a e Senso s 2019,19, 5370 3 o 18 p esen ed, which had he goal o e icien ly de ec anomalies o empe a u e alues. Fu he mo e, conside a ions on di e en sizes o aining and es da a a e discussed. • Fo p oduc ion companies, he de ec ion o anomalies becomes inc easingly impo . As mo e esea ch is equi ed o ge be e insigh s in o eal-wo ld scena ios and da ase s, his wo ks con ibu es wi h he esul s o wo complex use cases. The selec ion o app op ia e algo i hms and hei pa ame e iza ion is s ill challenging, which can be also seen om he ac ha less s anda d so wa e is o e ed in his con ex . B oke Machine Machine Lea ning Se ice s a /s op Machine Ope a o En i onmen al Senso s Senso Da a P ocessing Pipeline In e nal Senso s b e a c d Machine Da a Use Case En i onmen al Da a Use Case P oduc ion En i onmen ansmi en i onmen al senso da a subsc ibe publish ale s ansmi machine senso da a no i y publish machine e en s Figu e 1. Schema ic o e iew o he p esen ed use cases. The emainde o his pape is o ganized as ollows. Fi s , ela ed wo ks a e discussed in Sec ion 2. In Sec ion 3, he me hod o dis ance p o iling is applied o a eal-wo ld da ase o a pha ma packaging machine, while Sec ion 4shows me hods o p edic anomalies o en i onmen al da a o a 3D machine’s manu ac u ing oom. Sec ion 5discusses he esul s and deals wi h he e ealed limi a ions. Sec ion 6 concludes he wo k wi h a summa y and ou look. 2. Rela ed Wo k Rega ding he acquisi ion o senso da a in manu ac u ing sys ems, which is an impo an p e equisi e o his wo k, di e en ela ed wo ks exis . The au ho s o [ 5 ] discuss equi emen s o da a acquisi ion o p oduc ion sys ems and in oduce an a chi ec u e based on he Open Pla o m Communica ions Uni ied A chi ec u e (OPC UA) o da a ansmission and he p ecision ime p o ocol (PTP) o ime synch oniza ion [ 6 ]. The au ho s o [ 7 ], in u n, p o ided an o e iew o me hods, echnologies, and exchange p o ocols o enable dynamic da a acquisi ion o senso da a in indus ial sys ems. Challenges ega ding he ep esen a ion and ans o ma ion o senso da a in cybe -physical sys ems a e p esen ed in [ 8 ]. In he ield o semiconduc o manu ac u ing, a ious case s udies exis o imp o e manu ac u ing p ocesses by analysing manu ac u ing senso da a [ 9 ]. Thus, Ad anced P ocess Con ol (APC) me hods u ilize con ol s a egies and analyses o iden i y machine aul s and hei causes [10]. The esul s a e hen used o, o example, op imize main enance schedules [11,12]. In gene al, di e en anomaly de ec ion echniques exis , which can be classi ied in o s a is ical me hods, such as S a is ical P o iling [ 13 ], Pa ame ic S a is ical Modeling [ 14 ], and Machine Lea ning App oaches [15]. Conce ning use cases o anomaly de ec ion in indus ial sys ems, he au ho s o [ 5 ] in oduce a model-based app oach o he p edic ion o ene gy consump ion in p oduc ion plan s in o de o de ec anomalies using he ANODA algo i hm [ 16 ]. [ 17 ] de ec ed anomalies by applying a Bayesian ne wo k and sco ing he esul ing ea u es acco ding o a sco ing model. The au ho s o [ 18 ], in u n, de eloped an assis ance sys em o da a acquisi ion, p ocess moni o ing, and anomaly de ec ion in indus ial and ag icul u al p ocesses and e alua ed h ee use cases. Anomaly de ec ion was de eloped Senso s 2019,19, 5370 4 o 18 indi idually o each use case and is based, o example, also on he dis ance p o iling app oach, wi h a local ou lie ac o and PCA-based anomaly de ec ion [19,20]. Rega ding s eam-based anomaly de ec ion in gene al, di e en app oaches exis . [ 21 ] conduc ed anomaly de ec ion based on an au o- eg essi e, da a-d i en model o he conside ed da a s eam. They analyzed da a s eams o eal-wo ld wind speed measu emen s. The au ho s o [ 22 ], in u n, p esen ed an app oach based on hal -space ees, which is a one-class anomaly de ec ion algo i hm. I s ad an ages wi h espec o he compu a ional complexi y a e cons an ime and memo y oo p in s. Finally, [ 23 ] in oduced an anomaly de ec ion me hod based on he Hie a chical Tempo al Memo y (HTM), which is a s eam-based sequence memo y algo i hm. Fu he mo e, hey in oduced da ase s con aining eal-wo ld da a s eams wi h labeled anomalies o enable benchma ks o s eam-based anomaly de ec ion. Fo ime se ies p edic ion , plen y o ela ed wo ks exis . In [ 24 ], an algo i hm called Ul a Fas Fo es T ee (UFFT) was in es iga ed ha examines he beha io o an ensemble o eg ession ees on s eaming da a. The wo k in oduces a hyb id adap i e sys em o he induc ion o andom o es s om s eaming da a. The UFFT sys em is an inc emen al algo i hm ha poses a cons an ime complexi y o p ocess each ins ance, wo ks online, and uses he Hoe ding bound o decide when a spli es is ins alled on a lea leading o a decision node [25]. The algo i hm uses a con inuous da a s eam and du ing he aining phase, sho - e m memo y is used. This me hod, in u n, is es ic ed o bina y classi ica ion. Howe e , i can be ex ended o a mul iple classi ie by inc easing he numbe o classi ie s and building a andom o es o bina y ees. Al hough UFFT is no be e in p edic ion, i is signi ican ly as e han he C4.5 algo i hm [ 26 ]. Howe e , UFFT is no able o de ec li le o ab up concep d i s like [ 27 ], who in es iga ed d i de ec ion. They p esen an algo i hm, which c ea es Reg ession T ees (RTs), ins ead o Random Fo es T ees, based on da a s eams in he p esence o concep d i s. RTs a e as e in lea ning, bu a e p one o ou lie s, as hey do no wo k as an ensemble. The Fas Inc emen al Reg ession T ee-D i De ec ion (FIRT-DD) algo i hm allows o model adop ion a any ime and is able o deal wi h local concep d i s and adap s locally. By doing so, global model adop ion is a oided o gain e iciency. The change de ec ion algo i hm is based on change de ec ion uni s (CDUs) ha moni o he g owing p ocess. CDUs equi e ew memo y spaces pe node and a small, cons an amoun o ime complexi y o each sample. Re . [ 27 ] de ined concep d i s as a change o he unde lying join p obabili y, i.e., a change o P(Y|X) , and dis inguished be ween h ee main app oaches o concep d i s: (1) me hods ha explici ly de ec concep d i s, (2) me hods ha use ensembles o decision models, and, (3) me hods ha a e based on da a managemen using a sliding window (simila o ou p esen ed app oach). Fas Inc emen al Model T ees wi h D i De ec ions (FIMT-DD), he successo o FIRT-DD, is based on andomized model ees and combines hem o ensembles, which leads o a andom o es eg ession [ 28 ]. Each lea and node o a ee is buil on andomly and independen ly chosen a ibu es. The au ho s addi ionally c ea ed an Online Reg ession Fo es (ORF), i.e., ou o en FIMT-DD, wi h a ee dep h o i e. Howe e , he e is no s iking ou pe o mance o one o he p esen ed algo i hms when compa ing a single online ee, including op ional spli s wi h a andom o es consis ing o en ees [ 29 ]. While pe o mance, in gene al, depends on he da ase , empi ical analyses showed ha a single online op ion ees wi h a e aging i s bes o mos compa ed da ase s. Al oge he , he combina ion o (1) how anomaly de ec ion me hods ha e been in eg a ed in o echnical se ings o eal-wo ld examples o p oduc ion machines and (2) applica ion examples o how anomaly de ec ion o senso da a can be pe o med in a da a-d i en manne , wi hou majo pa ame e op imiza ions, has no been p esen ed by he discussed o he wo ks so a as done in his pape . 3. Pha ma Packing Use Case—Machine Senso Da a The ollowing use case p esen s he applica ion o dis ance p o iling o a eal-li e da ase o a pha ma packaging machine wi h he goal o de ec anomalies o a machine componen . In he i s s ep, ele an senso da a was collec ed. Then, he ob ained senso da a we e p ocessed by a da a p ocessing pipeline. Senso s 2019,19, 5370 5 o 18 P ac ical esul s show ha dis ance p o iling can be a aluable me hod o de ec he anomalies o he pha ma packaging machine componen s. Use Case Desc ip ion: Uhlmann Pac-Sys eme GmbH & Co. KG is a mechanical enginee ing company headqua e ed in Laupheim, Ge many. Uhlmann is a supplie o pha maceu ical w apping packaging machines. The blis e machines o Uhlmann pu able s in o indi idual packaging uni s. Fo example, he o e ed machines o m blis e s wi h indi idual cou s o able s om a plas ic o aluminum oil s and (see Figu e 2). Table s, in u n, a e ed and so ed in o blis e cou s. The la e a e comple ed wi h a co e shee , while inished blis e s a e inally punched ou by he Uhlmann machines. Figu e 2. Uhlmann blis e machine B1440i. As all Uhlmann packaging machines a e used in he pha maceu ical indus y, coun y-speci ic laws and egula ions mus be conside ed and ul illed. These legal equi emen s a e ela ed o he alida ion o machines, including he p o ision o de ailed documen a ion o all p ocess s eps ha a e pe o med du ing d ug packaging. No ably, e e y packaging machine deli e s senso da a, which can be con inuously moni o ed in o de o de ec anomalies. As an impo an p elimina y echnical s ep o he de ec ion o anomalies, he acquisi ion o senso da a and he gene a ion o ac ua o signals mus be p o ided by p og ammable logic con olle s (PLC) [ 30 ]. Along wi h he example o a p oduc loade , his p elimina y s ep will be sho ly delinea ed, as i shows he cha ac e iza ion o he esul ing senso da a. The p oduc loade s a ion o an Uhlmann blis e machine loads blis e s and lea le s in o a ca on. The p oduc loade consis s o se e al senso s and se omo o s. Speci ically, wo unc ional assemblies wi h se omo o s a e unning in pa allel, while he senso s o each single se omo o gene a e ou signals ha ep esen he 1. mechanical posi ion (MP) o he assembly (P oduc Loade .Posi ion), 2. he di e ence o a e e ence alue and he ac ual alue o he MP (P oduc Loade .D e ), 3. he powe consump ion o he assembly (P oduc Loade .Cu en ), 4. and he e e ence signal o he assembly (P oduc Loade .Re e ence). O e all, eigh signals pe p oduc loade can be acqui ed. The physical p ocess o p oduc loading and p oduc eleasing is execu ed in con inuous cycles. No e ha he packaging pe o mance o a machine is he e o e exp essed in cycles pe minu e. Du ing hese cycles, di e en anomalies may occu . A machine componen can lock, which leads o a p oduc ion hal . This can be caused, in u n, by a aul y eeding o he packaging box. Fu he mo e, wea and ea o he ball bea ings can lead o inc eased ic ional esis ance o a comple e ailu e o he ball bea ing lub ica ion. Bo h anomalies a e epo ed by he echnical e angelis s o be de ec able by analyzing he powe consump ion o he p oduc loade . Fundamen als: To de ec anomalies in ime se ies da a, dis ance p o iling can be applied [ 31 ]. A dis ance p o ile, in u n, is a ec o D o he Euclidean dis ances be ween a gi en ime se ies pa e n pand e e y possible subsequence siin he espec i e ime se ies s(see Figu e 3). Senso s 2019,19, 5370 6 o 18 Pa e n p Subsequence nm Dis ance P o ile D= Time Se ies s dp,1 dp,2 dp,3 dp,n-m+1 sn-m+1 ... Figu e 3. Dis ance p o ile o a signal. In his pape , Mueen’s Algo i hm o Simila i y Sea ch (MASS) was used o calcula e he dis ance p o ile [ 32 ]. When calcula ing he dis ance p o ile, he esul ing ime complexi y is O(nm) . MASS uses a con olu ion-based me hod o calcula e he dis ance p o ile o a pa e n in O(n×log n) . Fu he mo e, a z-no maliza ion is applied o he gene a ed dis ance p o ile ec o du ing he calcula ion as well. I he Euclidean dis ance be ween a pa e n p and a subsequence si is smalle han he alue o a h eshold, hen p and si a e conside ed o be simila . Senso Da a Acquisi ion: When p ocessing senso da a o he pha ma packing machines, di e en s eps ha e o be pe o med [ 33 ]. Fi s , senso da a has o be collec ed om a PLC and ans e ed o a collec ion componen . We de eloped a bina y ansmission p o ocol o ans e senso da a di ec ly om he PLC o a collec ion componen ia TCP socke s. The p o ocol o e s di e en ame ypes o ime synch oniza ion, signal me ada a desc ip ion, and senso da a poin ansmission. Senso Da a P ocessing: The ans e ed aw da a s eam has o be spli in o windows and p e-p ocessed i i ea u es dis u bing noise. A e he p e-p ocessing s ep, he da a is p ocessed h ough as -Fou ie ans o ma ions and inally s o ed i no u he p ocessing s eps a e equi ed. In his s ep, we adjus ed he al eady exis ing echnical de elopmen o enable he in eg a ion o anomaly de ec ion app oaches a e he p e-p ocessing s ep. To enable his, he u ilized p ocessing pipeline mus p o ide a high a iabili y o e e y senso signal, which means ha he pipeline is gene ally di icul o manage and con igu e. To add ess hese issues, we ha e de eloped a Senso Da a P ocessing (SDP) amewo k in C# o Mic oso . NET o collec ing, p ocessing, s o ing, and isualizing aw senso da a in a con inuous p ocessing chain, acco ding o he da a s eam p ocessing model [ 34 ]. The SDP de ines a g aph-based p ocessing model, in which p ocessing nodes a e connec ed o each o he o handle all a o emen ioned and equi ed pipeline s eps in a con olled way (see Figu e 4). Figu e 4. Schema o he senso da a p ocessing pipeline. Senso Da a Analysis: The da ase used in his wo k was gene a ed by an Uhlmann endu ance es a angemen in he Uhlmann echnical cen e in Laupheim, Ge many. To be mo e speci ic, an isola ed p oduc loade s a ion was unning in a con inuous ope a ion mode. Fo ou p ac ical e alua ion, senso da a we e collec ed o e 10 days by he SDP and s o ed in a MongoDB da abase. In o al, 1,127,790 imes amps pe signal (i.e., ime synch oniza ion ames) and 18,736,022,320 da a poin s (i.e., loa ing-poin numbe s ep esen ing he measu ed senso da a) we e eco ded du ing he men ioned ime pe iod, including NULL alues o he signal gaps. In e ms o disk space, he eco ded alues co espond o oughly 22 GB. Fu he no e ha he da ase comp ises a ew signal gaps due o connec i i y losses be ween he PLC and he s o age componen . Impo an ly, he da ase con ained eco ds o bea ing damage, which e en ually led o he ailu e o he machine. Apa om ha , he da ase can be conside ed as being heal hy, as he endu ance es and an e alua ion o he log Senso s 2019,19, 5370 7 o 18 da a gene a ed by he es execu ion sys em showed no o he abno mali ies. Figu e 5shows he mechanical posi ion and powe consump ion alues o i e a bi a y p oduc loade cycles ( x1 o x5 ) o he conside ed da ase . Du ing a cycle, he mechanical posi ion alues inc ease mono onously, while a p oduc is loaded in o a ca on (x-axis). The powe consump ion (y-axis) shows posi i e alues i he p oduc loade is inc easing in i s speed, while nega i e alues show b eaking ac ions. When a new cycle s a s, he posi ion is se back o ze o. Figu e 5. Powe consump ion (blue) and e e ence signal (o ange) o an assembly. Dis ance P o iling: Fo his p ac ical use case, he SDP p ocessing pipeline was adjus ed by changing he ecei e node o ecei e BMTT da a, a windowing node (see below), and a p ocessing node implemen ing he dis ance p o iling based on he MASS, and an ou pu node o publish MASS esul s o a b oke (see Figu es 1c and 4). A p e-p ocessing node is no necessa y as he MASS p ocessing node calcula es z-no maliza ion and dis ance p o iles simul aneously. As he SDP pipeline p ocesses da a s eams, windowing had o be applied [ 34 ]. The e o e, co ela ion windows a e used. The la e a e a specializa ion o session windows [ 35 ]. In con as o session windows, co ela ion windows a e igge ed by one e en a he beginning o a new window. In he SDP, co ela ion windows may be also gene a ed based on o he signals. He e, a window igge is execu ed based on he mas e encode , ep esen ed by a saw- oo h signal (see Figu e 5). Tha means, he egula clocking o he mas e encode (abou 720° ) was used o di ide he ela ed signals in o uni o m windows by de ec ing i s alling slope (da aF ame.posi ion[i+1] - da aF ame.posi ion[i]) <− 500). The MASS algo i hm was hen applied o he da a o he powe consump ion signal o he p oduc loade , as i o e s he mos exp essi eness o he mechanical p ocess and ea u es a e y high esolu ion, i.e., one da a poin pe 2 milliseconds. In con as , he mechanical posi ion o e s only abou 1000 di e en , mono onously inc easing alues (see Figu e 5). Mo eo e , a compa ison be ween he indi idual pa e ns becomes possible. The e o e, a new coun e alue is c ea ed o each ecognized new pa e n o a window. In o de o classi y a new pa e n, he Euclidean Dis ance is used o de e mine he signals o exis ing pa e ns wi h he signals o new pa e ns. In p ac ice, he coun e alue inc eases apidly a he beginning o an analysis se ies, as well as in he ange o any anomaly and d ops below a ce ain limi he ea e . No e ha he numbe o pa e ns caused by de ec ed anomalies a e g adually conside ed “no mal” by MASS, and a e he e o e smoo hed. Con e sely, his means ha a e a la ge numbe o de ec ed pa e ns, i is no assumed ha he anomaly has been disappea ed. Resul s: On he gi en 10-days da ase , he applica ion o he MASS wi h he de eloped SDP p ocessing node ook a o al un ime o 60 min. As a esul , he MASS de ec ed ewe han 5 pa e ns pe Senso s 2019,19, 5370 8 o 18 da a poin analysis a e 22 h in he da ase , i.e., he baseline was eached. In e es ingly, abou 14 h p io o bea ing damage, he numbe o de ec ed pa e ns inc eased apidly (see Figu e 6). A 10:00 a.m., he numbe o de ec ed pa e ns e u ned o 1 pa e n pe sample. A 01:30 p.m., he numbe o de ec ed pa e ns inc eased again o 7, whe eas a 01:48 p.m., he machine inally s opped wo king due o bea ing damage. Figu e 7a shows window-laye ed plo s o he powe consump ion in he no mal condi ion ( ew de ec ed pa e ns), whe eas Figu e 7b in bad condi ion (many de ec ed pa e ns). The x-axis ep esen s he e e ence signal o he p oduc loade in deg ee, while he y-axis ep esen s he ac ual powe consump ion in milliampe e (mA) o he wo assemblies. In Figu e 7b, powe consump ion shows a highe a iance. In e es ingly, he idle phase shows he highes a iance o powe consump ion. One eason o his may be he con ol loop used as du ing slow mo emen s, he ac ual posi ion is app oached o he e e ence posi ion wi h smalle speed changes,. A highe a iance also means a lowe mo o p ecision o he mo emen s in he no mal s a e. Figu e 6. Numbe o de ec ed pa e ns (y-Axis) and ime (x-Axis) o a p oduc loade o he las 14 h collec ed. Figu e 7. Phase olded plo o ( a ) no mal si ua ion wi h less de ec ed pa e ns and ( b ) a de ec ed anomaly wi h many pa e ns. Summa y: The da ase o he p oduc loade componen o a pha ma packaging machine was analyzed o de ec anomalies in he collec ed da ase . The selec ion o a p ope anomaly de ec ion algo i hm is challenging as many a ian s o a machine exis and he selec ion o pa ame e s o Senso s 2019,19, 5370 9 o 18 each machine equi es high e o s. The e o e, he de elopmen o a pa ame e less algo i hm ha i s in o he se ing o he pha ma packaging machine should be a majo goal. In his pape , a dis ance p o ile anomaly de ec ion algo i hm was p esen ed and applied. Speci ically, he dis ance p o ile was calcula ed wi h he MASS based on he z-Euclidean Dis ance. The la e o e s a ious ad an ages, e.g., i equi es ew pa ame e s. Fo he senso da a acquisi ion, he de eloped senso da a p ocessing amewo k (SDP) was used. The la e is based on a p ocessing pipeline model ha is con igu able and ex ensible, i.e., o he anomaly de ec ion algo i hms can be lexibly in eg a ed. Al hough i is a a he simple me hod, MASS showed p omising esul s on he p esen ed eal-wo ld da ase . I was possible o de ec anomalies in a meaning ul manne , i.e., based on he numbe o de ec ed pa e ns, 13 h as well as 18 min be o e he damage. P ac ically, his means ha a echnical e angelis o he pha ma packaging company can be e analyze he da a o decide whe he bea ing damage can be a oided by s opping he machine o eplacing componen s be o e damage ac ually occu s. Howe e , i has also been shown ha he in eg a ion o he anomaly de ec ion me hod equi es conside able echnical e o s. 4. 3D P in ing Machine—Tempe a u e En i onmen Da a The ollowing use case p esen s he applica ion o p edic ion models o a eal-li e da ase o a measu ing oom o 3D p in ing machines wi h he goal o p edic anomalies o empe a u e alues. In he i s s ep, ele an senso da a we e collec ed. Then, he ob ained senso da a a e p ocessed and p omising ea u es a e selec ed. Finally, machine lea ning models p edic upcoming alues and wa n echnical machine ope a o s abou possible anomalies. Use Case Desc ip ion: An indus ial company in he ield o 3D p oduc ion machines ope a es a measu ing oom equipped wi h nine empe a u e senso s and addi ional senso s o ai humidi y (%), ai p essu e (mBa ), and ai low (m/s). The oom con ains an a bi a y numbe o machines and hei ope a o s en e and lea e he oom a a bi a y poin s in ime. Machines wi hin he oom a e allowed o ope a e wi hin a empe a u e h eshold, which is de ined indi idually o each machine. A empe a u e con ol uni ies o keep he empe a u e in he oom be ween he uppe and lowe h eshold. The challenge o his p ojec is o igu e ou empe a u e anomalies o he measu ing oom ha is solely based on he p o ided en i onmen al da a and does no con ain any con ex ual in o ma ion o he ope a ed machines, such as he numbe o ope a o s o he ac ual numbe o he machines in he oom. Fu he mo e, no s anda ds o accep able e o s we e se o he p ojec . The models p esen ed below e u n a p edic ion e o on a basis o which he echnical e angelis s can decide a e a no i ica ion whe he he e o a e is accep able o no . Senso Da a Acquisi ion: The measu ing oom is sending he cu en senso in o ma ion wi h a ansmission a e o one alue pe 3 min. Figu e 8illus a es he used a chi ec u e. He eby, all alues a e combined using he Ja aSc ip Objec No a ion (JSON). Nex , all sen messages a e ans e ed ia he Message Queuing Teleme y T anspo p o ocol (MQTT), which is a publish/subsc ibe based message p o ocol o machine- o-machine (M2M) communica ion. The machine lea ning code is implemen ed as a sepa a e se ice using he sciki -lea n py hon lib a y [ 36 ]. The sciki se ice subsc ibes o he en i onmen al alues and uses hem o ain a machine lea ning model. The model is cons an ly e alua ed and, whene e he p edic ed empe a u e alue exceeds a de ined h eshold, he esponsible machine ope a o s (i.e., he echnical e angelis s) a e ala med. In con as o he i s use case, he e, he in eg a ion o he anomaly de ec ion me hod is echnically easie . Howe e , he selec ion o p ope algo i hms is mo e challenging han o he i s use case. Senso s 2019,19, 5370 16 o 18 min_samples_spli =2, min_weigh _ ac ion_lea =0.0 , p eso =False , andom_s a e=0, s p l i e = ’ bes ’ ) MLPReg esso ( a c i a ion = ’ elu ’ , alpha =0.0001 , ba ch_size= ’ au o ’ , be a_1 =0.9 , be a_2 =0.999 , ea ly_s opping=False , ep si lon =1e −08, hidden_laye _sizes =(100 ,) , lea ning_ a e= ’ cons an ’ , l e a n i n g _ a e _ i n i =0. 001 , max_i e =200 , momen um=0.9 , n_i e _no_change =10 , nes e o s_momen um=T ue , powe _ =0.5 , andom_s a e=None , sh u l e=T ue , sol e = ’ l b gs ’ , o l = 0.0001 , a li d a i o n _ ac i o n =0.1 , e bose=T ue , wa m_s a =False ) RandomFo es Reg esso ( boo s ap=T ue , c i e i o n = ’mse ’ , max_dep h=10 , max_ ea u es= ’ au o ’ , max_lea _nodes=None , min_impu i y_dec ease =0.0 , min_impu i y_spli =None, min_samples_lea =1 , min_samples_spli =2, min_weigh _ ac ion_lea =0.0 , n_es ima o s =1000 , n_jobs=−1, oob_sco e=False , andom_s a e=1, e bose =0 , wa m_s a =False ) Re e ences 1. Hoppens ed , B.; P yss, R.; S elze , B.; Meye -B ö z, F.; Kamme e , K.; T eß, A.; Reiche , M. Techniques and Eme ging T ends o S a e o he A Equipmen Main enance Sys ems—A Bibliome ic Analysis. Appl. Sci. 2018,8, 916. [C ossRe ] 2. Hawkins, D. Iden i ica ion o Ou lie s; Chapman and Hall: London, UK, 1980. 3. Hodge, V.; Aus in, J. A su ey o Ou lie De ec ion Me hodologies. A i . In ell. Re . 2004 ,22, 85–126. [C ossRe ] 4. Kno , E.; Ng, R. Algo i hms o Mining Dis ance-Based Ou lie s in La ge Da ase s. In P oceedings o he 24 h In e na ional Con e ence on Ve y La ge Da a Bases (VLDB ’98), New Yo k, NY, USA, 24–27 Augus 1998 ; Mo gan Kau mann Publishe s Inc.: San F ancisco, CA, USA, 1998; pp. 392–403. 5. Fal inski, S.; Fla , H.; Pe hig, F.; K oll, B.; Vodenˇca e i´c, A.; Maie , A.; Niggemann, O. De ec ing Anomalous Ene gy Consump ions in Dis ibu ed Manu ac u ing Sys ems. In P oceedings o he 10 h In e na ional Con e ence on Indus ial In o ma ics, Beijing, China, 25–27 July 2012; pp. 358–363. Senso s 2019,19, 5370 17 o 18 6. Lei ne , S.; Mahnke, W. OPC UA–Se ice-o ien ed A chi ec u e o Indus ial Applica ions. ABB Co p. Res. Cen . 2006,48, 61–66. 7. Schlech endahl, J.; Keine , M.; K e schme , F.; Lechle , A.; Ve l, A. Making Exis ing P oduc ion Sys ems Indus y 4.0-Ready. P od. Eng. 2015,9, 143–148. [C ossRe ] 8. Ji ko sk ` y, V.; Obi ko, M.; Maˇ ík, V. Unde s anding Da a He e ogenei y in he Con ex o Cybe -Physical Sys ems In eg a ion. IEEE T ans. Ind. In o m. 2016,13, 660–667. [C ossRe ] 9. Moyne, J.; Iskanda , J. Big Da a Analy ics o Sma Manu ac u ing: Case S udies in Semiconduc o Manu ac u ing. P ocesses 2017,5, 39. [C ossRe ] 10. Feng, J.; Jia, X.; Zhu, F.; Moyne, J.; Iskanda , J.; Lee, J. An Online Vi ual Me ology Model Wi h Sample Selec ion o he T acking o Dynamic Manu ac u ing P ocesses Wi h Slow D i . IEEE T ans. Semicond. Manu . 2019,32, 574–582. [C ossRe ] 11. Rami ez-He nandez, J.A.; C ab ee, J.; Yao, X.; Fe nandez, E.; Fu, M.C.; Janaki am, M.; Ma cus, S.I.; O’Conno , M.; Pa el, N. Op imal P e en i e Main enance Scheduling in Semiconduc o Manu ac u ing Sys ems: So wa e Tool and Simula ion Case S udies. IEEE T ans. Semicond. Manu . 2010 ,23, 477–489. [C ossRe ] 12. Ramí ez-He nández, J.A.; Fe nandez, E. Op imiza ion o P e en i e Main enance scheduling in semiconduc o manu ac u ing models using a simula ion-based App oxima e Dynamic P og amming app oach. In P oceedings o he 49 h IEEE Con e ence on Decision and Con ol (CDC), A lan a, GA, USA, 15–17 Decembe 2010; pp. 3944–3949. 13. Hill, D.J.; Minske , B.S.; Ami , E. Real- ime Bayesian Anomaly De ec ion in S eaming En i onmen al Da a. Wa e Resou . Res. 2009,45. [C ossRe ] 14. Candès, E.J.; Li, X.; Ma, Y.; W igh , J. Robus P incipal Componen Analysis? J. ACM 2011 ,58, 11. [C ossRe ] 15. Nai ac, A.; Townsend, N.; Ca , R.; King, S.; Cowley, P.; Ta assenko, L. A Sys em o he Analysis o Je Engine Vib a ion Da a. In eg . Compu . Aided Eng. 1999,6, 53–66. [C ossRe ] 16. Vodenˇca e i´c, A.; Büning, H.K.; Niggemann, O.; Maie , A. Using Beha io Models o Anomaly De ec ion in Hyb id Sys ems. In P oceedings o he 2011 XXIII In e na ional Symposium on In o ma ion, Communica ion and Au oma ion Technologies, Sa aje o, Bosnia and He zego ina, 27–29 Oc obe 2011; pp. 1–8. 17. Ebe ha d , J.S., III; Radano, T.A.; Pe e son, B.E. Applica ion o Machine Lea ned Bayesian Ne wo ks o De ec ion o Anomalies in Complex Sys ems. U.S. Pa en 9,349,103, 24 May 2016. 18. Windmann, S.; Maie , A.; Niggemann, O.; F ey, C.; Be na di, A.; Gu, Y.; P omme , H.; S eckel, T.; K üge , M.; K aus, R. Big Da a Analysis o Manu ac u ing P ocesses. J. Phys. 2015,659, 012055. [C ossRe ] 19. Phua, C.; Alahakoon, D.; Lee, V.C.S. Mino i y Repo in F aud De ec ion: Classi ica ion o Skewed Da a. SIGKDD Explo . Newsl. 2004,6, 50–59. [C ossRe ] 20. Eickmeye , J.; Li, P.; Gi ehchi, O.; Pe hig, F.; Niggemann, O. Da a D i en Modeling o Sys em-Le el Condi ion Moni o ing on Wind Powe Plan s. In P oceedings o he 26 h In e na ional Wo kshop on P inciples o Diagnosis, Pa is, F ance, 31 Augus –3 Sep embe 2015. 21. Hill, D.; Minske , B. Anomaly De ec ion in S eaming En i onmen al Senso Da a: A Da a-d i en Modeling App oach. En i on. Model. So w. 2010,25, 1014–1022. [C ossRe ] 22. Tan, S.C.; Ting, K.M.; Liu, T.F. Fas Anomaly De ec ion o S eaming Da a. In P oceedings o he 22nd In e na ional Join Con e ence on A i icial In elligence, Ba celona, Ca alonia, Spain, 16–22 July 2011. 23. Ahmad, S.; La in, A.; Pu dy, S.; Agha, Z. Unsupe ised Real- ime Anomaly De ec ion o S eaming Da a. Neu ocompu ing 2017,262, 134–147. [C ossRe ] 24. Gama, J.; Medas, P.; Rocha, R. Fo es T ees o On-line Da a. In P oceedings o he ACM Symposium on Applied Compu ing, Nicosia, Cyp us, 14–17 Ma ch 2004; pp. 632–636. 25. Hoe ding, W. P obabili y Inequali ies o Sums o Bounded Random Va iables. J. Am. S a . Assoc. 1963 ,58, 13. [C ossRe ] 26. Quinlan, J. C4.5: P og ams o Machine Lea ning; Mo gan Kau mann Publishe s Inc.: San F ancisco, CA, USA, 1993. 27. Ikonomo ska, E.; Gama, J.; Sebas ião, R.; Gjo gje ik, D. Reg ession ees om da a s eams wi h d i de ec ion. In P oceedings o he In e na ional Con e ence on Disco e y Science, Po o, Po ugal, 3–5 Oc obe 2009; pp. 121–135. 28. Ikonomo ska, E.; Gama, J.; Dže oski, S. Online T ee-based Ensembles and Op ion T ees o Reg ession on E ol ing Da a S eams. Neu ocompu ing 2015,150, 458–470. [C ossRe ] Senso s 2019,19, 5370 18 o 18 29. Ikonomo ska, E.; Gama, J.; Zenko, B.; Dze oski, S. Speeding-up Hoe ding-based Reg ession T ees wi h Op ions. In P oceedings o he 28 h In e na ional Con e ence on Machine Lea ning, Washing on, DC, USA, 28 June–2 July 2011; pp. 537–544. 30. Bol on, W. P og ammable Logic Con olle s; Newnes: Ox o d, UK, 2015. 31. Yeh, C.C.M.; Zhu, Y.; Ulano a, L.; Begum, N.; Ding, Y.; Dau, H.A.; Sil a, D.F.; Mueen, A.; Keogh, E. Ma ix P o ile I: All Pai s Simila i y Joins o Time Se ies: A Uni ying View ha Includes Mo i s, Disco ds and Shapele s. In P oceedings o he 16 h In e na ional Con e ence on Da a Mining (ICDM), Ba celona, Spain, 12–15 Decembe 2016; pp. 1317–1322. 32. Mueen, A.; Viswana han, K.; Gup a, C.; Keogh, E. The Fas es Simila i y Sea ch Algo i hm o Time Se ies Subsequences unde Euclidean Dis ance. 2019. A ailable online: h p://www.cs.unm.edu/~mueen/ Fas es Simila i ySea ch.h ml (accessed on 15 No embe 2019). 33. Gama, J.; Gabe , M. Lea ning om Da a S eams: P ocessing Techniques in Senso Ne wo ks; Sp inge : Be lin, Ge many, 2007. 34. Babcock, B.; Babu, S.; Da a , M.; Mo wani, R.; Widom, J. Models and Issues in Da a S eam Sys ems. In P oceedings o he 21s ACM Symposium on P inciples o Da abase Sys ems (PODS ’02), Madison, WI, USA, 3–5 June 2002; pp. 1–16. 35. Akidau, T.; B adshaw, R.; Chambe s, C.; Che nyak, S.; Fe nández-Moc ezuma, R.J.; Lax, R.; McVee y, S.; Mills, D.; Pe y, F.; Schmid , E.; e al. The Da a low Model: A P ac ical App oach o Balancing Co ec ness, La ency, and Cos in Massi e-scale, Unbounded, Ou -o -o de Da a P ocessing. P oc. VLDB Endow. 2015 , 8, 1792–1803. [C ossRe ] 36. Ped egosa, F.; Va oquaux, G.; G am o , A.; Michel, V.; Thi ion, B.; G isel, O.; Blondel, M.; P e enho e , P.; Weiss, R.; Dubou g, V.; e al. Sciki -lea n: Machine Lea ning in Py hon. J. Mach. Lea n. Res. 2011 , 12, 2825–2830. 37. K eps, J. Ques ioning he Lambda A chi ec u e. 2014. A ailable online: h ps://www.o eilly.com/ ada / ques ioning- he-lambda-a chi ec u e/ (accessed on 3 No embe 2019). 38. Rak hanmanon, T.; Campana, B.; Mueen, A.; Ba is a, G.; Wes o e , B.; Zhu, Q.; Zaka ia, J.; Keogh, E. Sea ching and Mining T illions o Time Se ies Subsequences unde Dynamic Time Wa ping. In P oceedings o he 18 h In e na ional Con e ence on Knowledge Disco e y and Da a Mining (KDD), Beijing, China, 12–16 Augus 2012; pp. 262–270. 39. Hoppens ed , B.; Reiche , M.; Kamme e , K.; Spiliopoulou, M.; P yss, R. Towa ds a Hie a chical App oach o Ou lie De ec ion in Indus ial P oduc ion Se ings. In P oceedings o he CEUR Wo kshops o he EDBT/ICDT 2019 Join Con e ence (CEUR-WS.o g 2019), Lisbon, Po ugal, 26 Ma ch 2019. © 2019 by he au ho s. Licensee MDPI, Basel, Swi ze land. This a icle is an open access a icle dis ibu ed unde he e ms and condi ions o he C ea i e Commons A ibu ion (CC BY) license (h p://c ea i ecommons.o g/licenses/by/4.0/).