ContentslistsavailableatScienceDirect
Energy
Storage
Materials
journalhomepage:www.elsevier.com/locate/ensm
Battery
production
design
using
multi-output
machine
learning
models
Artem
Turetskyy
a,b,∗,
Jacob
Wessel
a,b,
Christoph
Herrmann
a,b,
Sebastian
Thiede
a,ba Technische Universität Braunschweig, Institute of Machine Tools and Production Technology, Chair of Sustainable Manufacturing and Life Cycle Engineering, Langer Kamp 19b, 38106 Braunschweig, Germany
b Battery LabFactory Braunschweig (BLB), Langer Kamp 8, 38106 Braunschweig, Germany
a
r
t
i
c
l
e
i
n
f
o
Keywords:Lithium-ion battery cells Data mining Machine learning Multi-output modelling Production design Cyber-physical system
a
b
s
t
r
a
c
t
Thelithium-ionbattery(LiB)isaprominentenergystoragetechnologyplayinganimportantroleinthefutureof e-mobilityandthetransformationoftheenergysector.However,LiBcellmanufacturinghasstillhighproduction costsandahighenvironmentalimpact,duetocostlymaterials,highprocessfluctuationswithhighscraprates, andhighenergydemands.AlackofaprofoundknowledgeofLiBcellproductionprocessesandtheirinfluence onthequalityandtheperformanceoftheLiBcellsmakesitdifficulttoplan,controlandexecutethe produc-tion.Therefore,asystematicapproachisnecessarytoestablishanin-depthunderstandingoftheinterlinkage ofprocessesandproducts’qualityandperformance.Thispaperpresentsamulti-outputapproachforabattery productiondesign,basedondata-drivenmodelspredictingfinalproductpropertiesfromintermediateproduct features.Thegivenconceptshowshowtheapproachcanbedeployedwithintheframeworkofacyber-physical productionsystemforcontinuousimprovementoftheunderlyingdata-drivenmodelanddecisionsupportin production.
1. Introduction
Thetransformationoftheautomotivesectortowardse-mobility to-getherwiththetransformationoftheenergysectortowardsahigher shareofrenewableenergies,heavilyreliesonavailableenergystorage technologies.Lithium-ionbatteries(LiB)havebeenthestateoftheart technologyforthelastdecades.Forthelastfiveyears,thecostsforLiB (Euro/kWh)havedecreasedandareexpectedtodecreaseevenfurther toapprox.75€/kWhin2022,comparedto400€/kWhin2013)[1]. LiBcellmanufacturingcostsarehighlysensitivetoscrapandprocess deviationssincethematerialcostshaveashareofabout75%ofthe totalmanufacturingcosts[2].Scrapratescanvary fromaround6% (housingparts)to15%(collectorfoils)[3]andleadevenupto40% ofproducedLiBcellsthatareeitherdefectiveorneed tobefixed in post-production[4].LiBcellsmakeupto50%ofCO2emissionsofan electricvehicle[5]ofwhich50%areduetotheenergydemandinLiB cellmanufacturing[6].LiBcellsrejectedattheendoftheproduction chain(formationandaging)havethehighestimpactonproductioncosts duetoaccumulatedenergydemandsandresourcesandfurtherprocess productionfailures[2].Therefore,itisofgreatimportancetohavean in-depthunderstandingofLiBcellmanufacturingandappropriate qual-itymanagementtodiscoverproductionfailuresandprocessdeviations inanearlystageofproduction.Thiscanhelptomeettheincreasing
de-∗Correspondingauthor.
E-mailaddress:a.turetskyy@tu-braunschweig.de(A.Turetskyy).
mandforenergystoragesandtoachieveeconomicandenvironmental goalsoflowermanufacturingcostsandlowerenvironmentalimpact.
TheproductionofLiBcellsiscomplex,consistingofvariousdiscrete andcontinuousprocesseswithdivergingandconvergingmaterialflows fromdifferentareasofengineering(processengineering,production en-gineering,electricengineering).Furthermore,theLiBcellisitselfa com-plexelectrochemicalsystem[2].Itconsistsofanodes,cathodes, separa-tor,electrolyte,andhousingmaterials,whereeachofthepartshasarole toplayinelectrochemicalreactionsduringcharginganddischarging processes.TheLiBcell’squalitycanbecharacterizedbyvarious differ-entproperties(chargecapacityordischargecapacityatcertainC-rates, self-discharge,stateofhealth,etc.),whichareasumofallthe struc-tures/featuresoftheusedmaterials(activematerials,conductive addi-tives,binders,etc.)andtheirstructuresadjustedduringtheproduction processes(particlesizedistribution,coatinglayerdensity,coatinglayer thickness,porosity,etc.).Thisrelationiscalledthe process-structure-propertyfunctionandisdepictedinFig.1[7].Alltheseaspectsmakeit hardtohaveaconstantlyhighproductqualitycontrolinLiBcell man-ufacturing.
Recentreviews concludethatthemostprominentapproachesuse soft-sensoringanddata-drivenapplicationsforqualitycontrolinLiBcell manufacturing[8,9].Mostoftheseapproachesaredeployedinorderto identifycriticalprocesssteps,criticalprocessparameters,orthe influ-encesofintermediateproductsonthefinalproductproperties.Theylack theintegrationinarunningLiBcellproductionforcontinuous deploy-mentandonlysomeofthemfocusontheimprovementofmanufacturing [10].Therefore,thispaperpresentsaconceptofhowadata-driven ap-proachforqualitymodellinginLiBcellproductioncanberealizedand https://doi.org/10.1016/j.ensm.2021.03.002
Received4December2020;Receivedinrevisedform7February2021;Accepted1March2021 Availableonline5March2021
Fig.1. Process-structure-propertyfunctionadaptedfrom[7].
deployedcontinuously.Thedeployedapproachaimsatimprovingthe qualityofLiBcellsbyenablingabetterproductiondesignandplanning usinginsightsfromthedata-drivenmodelling.
2. Technicalbackground
2.1. Generaloverviewoflithium-ionbatterycellproduction
Theproductionchainoflithium-ionbatterycellsconsistsof mani-folddifferentprocessesfromdifferentfieldsofengineering.Aschematic depictionoftheprocesschainwithinafactoryisdepictedinFig.2.A batterycellfactoryisshownwithtechnicalbuildingservices(TBS)and aprocesschainconsistingofanodeandcathodeelectrode manufactur-ingaswellascellmanufacturing.Eachproductionstepreceivesas in-puteithermaterialsorintermediate products(IP)aswell asprocess parameters(PP)whichadjustthecorrespondingproductionmachine. Asaresultofeachprocess,anintermediateproductismanufactured. Thetaskofeachprocessistoadjustthedesiredstructures/featuresof theseresulting intermediateproducts,whicharethenresponsiblefor theproperties/qualityofthebatterycell.Thestructures/featuresofthe
intermediateproductsaremeasuredbyintermediateproductanalytics (IPA).Furthermore,certaincellproductionstepsaresituatedinadry roomtoprotectwatersensitivecompounds.
Theelectrodes,theanode,andthecathodeareproduced individ-ually.Theelectrodematerials(activematerial,binder,conductive ad-ditive)arefirstmixedandthendispersedwithasolventintoaviscous slurry.Anintensiveprocessandincreasingstressdurationresultin adhe-sionoffragmentsofconductiveadditivesonactivematerialinfluencing therheologyandtheelectricalresistanceoftheslurry[12].Electrodes coatedwiththisslurryshowlateralowerspecificdischargeathigher C-rates[13].Insufficientpre-mixingleadstoanagglomerationof con-ductiveadditivesorofactivematerialresultinginincreasedinterfacial resistance[14].Further,insufficientmixingmayleadtotheformation of blisters/agglomeratesduringthecoatingprocess[15]. Inthe fol-lowingstep,theslurryiscoatedonafoil(cathode:aluminum,anode: copper)andthendried.Increasingdryingtemperatureandincreasing dryingrateleadbinderstomigratetotheelectrodes’surface[16,17]. Aninhomogeneousbinderdistributionin ananodecausedbyahigh dryingratemayleadtodecreasedcellcapacity[17].Thenextprocess, calendering,compressesthecoatedlayeroftheelectrode.Itimproves theelectrode’smechanicalstabilityandhomogenizesthecoatedlayer byreducingcracksinthecoating[13],butreducestheporosityleading toanincreaseofionicresistance[13].Further,modestcalenderingmay improvethewettingrateoftheelectrodesduetoalignmentofthe par-ticles[18],butcalenderingbeyondanideallevelmayleadtodecreased electrolytedistribution[18,19]aswellasreducedtortuosity[20]. Al-readyatlowcompressionratesparticledeformationcan beobserved [21],leadinguptocracksinsidetheactivematerialparticles[13].On theotherhand,electrodeswith10%compressionshowbetterlongterm performanceataC-rateof1Ccomparedtonon-compressedelectrodes [21].Later,theelectrodesarecutintopropersizescorrespondingtothe batterycellformatthatwillbebuilt.Twocuttingprocessesare com-monlyused,thedie-cuttingandthelasercutting.Theseprocesses
Fig.3. Relationbetweenprocess, interme-diate product, and the final product in theproductionoflithium-ionbatterycells, basedon[11].
encetheelectrodes’edgeindifferentways.Electrodescutwithalaser showmeltformationandaheat-affectedzoneontheiredges,whereas die-cutelectrodesshowstrongdelaminationandbendingoftheedges [22].Further,lasercuttingmayleadtoaluminumspattersonthe cath-odeside,whichmightpromotedendritegrowthonthecorresponding anodeside[22].
Thecutelectrodesaredriedinatemperaturechamberandarethen usedtobuildanelectrode-separatorassembly(ESA)duringthe assem-blyprocess.Duetothedifferentgeometriesofbatterycells,thereare threemainprocesses usedfor this,winding,stacking, andZ-folding. Duringthewindingprocess,theelectrodesheetsarewindedwiththe separatorintoanESAcalledjellyroll.Thewindingoftheelectrodes mayleadtocrackformationduetobindingstresswhichincreaseswith theelectrodes’ coatedlayerthickness[23].During thestacking pro-cess,electrodesheetsandseparatorsheetsarestackedovereachother. Theadvantageofthestackingprocesscomparedtothewinding pro-cessisamoreuniformmechanicalstressappliedtotheelectrodesheets, whichallowstheuseofthickerelectrodecoatingsandthereforehigher achievableenergydensitieswithoutpotentialcoatinglayerfailures[2]. Sincethestackingprocessisdiscrete,itisinferiorintermsof productiv-ityandprecisioncomparedtothewindingprocess[2].Further, stack-ingprocessesathigherprocessspeedsmayresultinamisalignmentof thestacked electrodesthatleadstodecreaseddischarge capacityand astrongercapacitylossovercharginganddischargingcycles[24–26]. DuringZ-folding,theelectrodesarestackedintoanendlesszigzag-shape oftheseparator.Z-foldedESAshowsalowerriskofshortsduetothe misalignmentofthestackedelectrodesheetsandstillhasthehigh en-ergydensityofanormallystackedESA[2].AftertheassemblyofESA, theelectrodearrestersareweldedtotabs(usuallyultrasonicorbylaser welding)inordertoprolongthemandtoensuretheconnectionofall electrodesofthesametypetoeachother.Mechanicalstressduring cal-enderingmayleadtowrinklingofthetabsandthereforehasanegative impactonthequalityoftheweldingprocess[27].TheESAisthenput intoacase(hardcaseorpouchbag),theso-calledhousingprocess,then filledwithelectrolyteandfinallythecaseissealed.Thebatterycellis storedinatemperature-controlledenvironmenttoensurethewettingof theporouselectrodecoatingandtheseparator.Anun-wettedelectrode materialis notusedduringcharginganddischarging,leadingto un-derutilizationoftheelectrodes’capacityandanincreaseinelectrolyte resistance[18].Thelastprocessesareformationandaging.Duringthe formation,thecharginganddischargingof thebatterycell,acertain amountoftheinitialcapacityofthebatterycellisirreversiblyconsumed inordertoformaninterfacelayerbetweentheelectrodesandthe elec-trolytecalledthesolidelectrolyteinterface(SEI)layer[28].TheSEI layeraffectscapacityloss,self-discharge,ratecapability,cyclelife,and
safetyofthebatterycell[28].Ithasbeenobservedthatanincreasing self-dischargeduringtheagingprocessmayleadtoalowermaximum capacityofthebatterycell[29].
AsimplificationofthementionedinteractionsisdepictedinFig.3. Productionprocesssteps(PS)dependingontheirsetupofPPandstate variables(SV)influencetheintermediateproductfeatures/structures (IPF)suchasparticlesizedistribution,coatinglayerthickness, overlap-pingrateoftheelectrodesheets,etc.TheseIPFsthendeterminehowthe qualityormultiplefinalproductproperties(FPP)ofthebatterycell(e.g. maximalcapacity,self-discharge,capacitylossofcyclelife,etc.)willbe. StudiesdescribetheinfluenceoftheproductionprocessesontheIPF.In order toassureandcontrolthequalityofproducedbatterycells,the influenceofIPFs,whicharemeasuredduringoraftereachproduction step,onmultipleFPPsofinterest,needtobedeterminedanddescribed. Thiswouldenableabetterassessmentofthequalityoftheintermediate productconcerningthequalityofthefinalproduct,leadingtoabetter productionunderstandingandcontrol.
2.2. Dataanalyticsregardingqualityinmanufacturing
A literaturereviewof qualitymanagement inmanufacturingwas conducted withregard toapplications of data analytics,data-driven methods,andmachinelearningmethodsforpredictingthequalityof afinalproduct.Manyofthempredictwhetherthefinalproductisokay (OK)ornot-okay (NOK)usingthemachine,sensor,andintermediate producttestingdata.Baietal.2017predictaqualityscoreof produc-tionbatchesusingadjustableandnon-adjustableprocessparametersas inputandcompare“shallow” anddeeplearningartificialneural net-works(ANN)asmodellingtechniques[30].Lieberetal.2013introduce ausecasefromthesteelindustryutilizingfeatureextraction/generation ofcontinuoussensordatabyapplyingfeatureselectionand dimension-alityreductiontobuildamodelfortheclassificationofOKandNOK steelbars[31].Weissetal.2014showanapplicationof regressions, decisiontrees,andrandomforestsforpredictingthespeedofawafer beforeitsqualitycontrolinmanufacturing.Thespeedofthewafercan beconsideredinthiscontextasanFPP.Arif,SuryanaandHussinpredict thequalityofaproductinamulti-stagemanufacturingsystemby us-inganadaptedcascadequalitypredictionmethod[32].Thisapproach predicts thequalityofeach intermediate productbyconsidering the adjustablePPofthecurrentworkstationandthequalityofitafterthe previousworkstation.Therefore,thequalityofthefinalproductis deter-minedbyallpreviousworkstationsandintermediateproductqualities. Themodellingtechniquesaredecisiontrees.Leeetal.2018deploya cyber-physicalproductionsystemwithaforecastingmodeltopredict thequality(OK/NOK)ofproductsinmetalcastingprocesses[33].The
Fig. 4. Integration of the intended goal of thisworkwithinthetechnicalbackgroundand stateofresearch.
modellingtechniquesaredecisiontrees,randomforests,andANNand thefeaturesarederivedfromtemperaturesensorsofthemold.
Recently,data-drivenapproachesforpredictingthequalityofLiB, morespecificallypredictingtheFPP ofLiB,wereintroduced.Schnell etal.2018compared linearmodels,tree-basedmodels,andANN to-wardstheirpredictabilityofthemaximalcapacityofbatterycellsbased on datafrom cellmanufacturing,excluding electrodemanufacturing [34].Theresultsshowacorrelationbetweencellmass,electrolytemass, andcellmax.capacity.Koransetal.2019performedanotherdata min-ingapproachonlyoncellmanufacturingdataalsoinordertopredict cellmax.capacity[35].Again,acorrelationbetweenelectrolytemass andthemax.capacityhasbeenidentified,butalsowiththeanodeand cathodelength.Thedifferencein thoseapproacheswas thatSchnell etal.2018usedstackingprocessesforcellassemblyandKornasetal. 2019windingprocesses,wheretheelectrodesusuallyhavelongersheets [35,36].Furthermore,Kornasetal.2019usedaprioriexpertknowledge fordataminingandmodelling[35].Turetskyyetal.2020applieddata miningondatafromthewholebatterycellmanufacturing(including electrodemanufacturingaswellascellmanufacturing)andusedsolely tree-basedmethodstopredictmax.capacityfromIPFdata.The influ-encingfactorsonthemax.capacitywereidentifiedbyfeatureselection methodsandwereevaluatedbythefeatureimportancecalculatedfrom thetree-basedmethodsandcategorizedbyproductionsteps(laser cut-ting,cellassembly,dispersing,andcalendering).Thiedeetal.2019also useddatafromthewholebatterycellmanufacturingbutusedtheIPF datatopredictnotonlythemax.capacity,butalsotheformationloss andthecapacitylossafter400charginganddischargingcycles[37]. Theinfluencingfactorswerealsoidentifiedbyapplyingfeature selec-tion,butthemodellingtechniquewaslinearregression.Theinfluencing factorswerealsocategorizedbyprocessstepsandtheirimpactswere evaluated bythecoefficient factorsofthe linearregression,andthe statisticalsignificancewasevaluatedbythep-valuecalculatedbythe T-test.
Furtherstudiesintroduced applicationsof machinelearning tech-niquesinLiBelectrodemanufacturing,withthefocusonelectrode for-mulation.Thesestudiespresentapproachescomplimentarytotheone presentedbeforepredictingthequalityofLiBcells.Cunhaetal.2020 comparedthreemachinelearningalgorithmsdecisiontrees,supporting vectormachines,anddeepneuronalnetworkstoidentifyandquantify theinterdependenciesbetweenelectrodeslurry(suspension)IPFs(e.g. viscosity,solid-to-liquidratio,etc)andcoatedelectrodeIPFs(massload andporosity)[38].AstudypresentedbyLiuetal.2020,alsofocusedon identifyingandquantifyinginterdependenciesbetweenelectrodeslurry IPFs,processparameters,andIPFsofacoatedelectrode[39].Liuetal. 2020used Gaussianprocessregression-basedmachine learning
algo-rithmswithautomaticrelevancedeterminationkernelstoquantifythe influencesofthethreeidentifiedelectrodeslurryIPFsandoneprocess parameteroncoatedelectrodemassload[39].Duquesnoyetal.2020 presentedamethodologycombiningexperimentalresults,insilico elec-trodestructures,derivedfromusingadata-drivenstochasticelectrode generator,andmachinelearningalgorithmSureIndependentScreening andSparsifyingOperator(SISSO)[40].Withthismethodology, Duques-noyetal.2020wereabletolinkanddescribetheinfluencesofelectrode structures ofuncalendered electrodestoLiBcellperformance-related electrodestructuresaftercalendaring,includingtheinfluenceof pro-cessparameters[40].
2.3. Researchdemand
Abatterycellasafinalproductischaracterizedbymorethanone quality property (FPP: max.capacity, agingbehavior, self-discharge, etc.),whichareinfluencedbyaplethoraoffeaturesandstructuresof in-termediateproducts(IPF:particledistributions,coatinglayerthickness, porosity,etc.)thatareshapedduringproductionprocesses.LiB produc-tionresearchfocusesstronglyontherelationsbetweenprocessesand IPFsandhowcertainIPFscanbeachievedthroughsuitableprocess pa-rameters(seeFig.4).Recentapplicationsofdataminingandapproaches drivenbyexpertknowledgeaimatassessingLiBcellqualityprediction, focusingonlyononetargetvaluepermodel,throughIPFandprocess parameters[11,34,35].However,inapreviousstudy,amodel predict-ingFPPbasedonIPFs,stillpredictingoneFPPpermodel,wasdeployed asaqualitygateapproachpredictingtheFPPafterupdatingthemodel withmeasuredIPFsduringLiBcellproduction[41].Nonetheless,most oftheapproachesdonotprovideautilizationordeploymentoftheir approachestoshowhowitcanbenefittheLiBcellproduction. Further-more,theyshowrelationsbetweenIPFsandFPPsinretrospectivebased onpreviouslyacquireddataanddonotderiveIPFsbasedonachieved FPPs.Therefore,thisworkpresentsamulti-outputmodelpredicting sev-eralLiBcellFPPsfromonesetofIPFsandprovidesanapproachofthe utilization/deploymentofthemulti-outputpredictionmodelinLiBcell production.Thisutilization/deploymentisachievedbypredictingIPFs basedon FPPsenablingadata-drivenbatteryproductiondesign(see Fig.4).
3. Approach
3.1. Overviewandframework
ThegoalistoestablishasystemfordeterminingneededIPFsderived fromdesiredFPPsoftheLiBcellsusingadata-drivenmodel(seeFig.5).
Fig.5. Basicprinciplesandgoalfortheconcept.
Fig.6.Proposedframeworkwithcyber-physicalproductionsystemlogic.
IPFsandFPPsareacquiredintheproductionofLiBcellsandareusedto buildamulti-outputdata-drivenmodelpredictingFPPsfromIPFs,the qualitypredictionmodel.ThismodelcanthenbeusedtodetermineIPFs basedonchosenFPPs,batteryproductiondesign.Thisprovidesatool forabetterunderstandingofLiBcellproductionplanninganddesign.
Toreachthedescribed goal,theapproach isbasedon the cyber-physicalproductionsystem(CPPS)logic(seeFig.6)[42].ACPPSisa frameworkcomprisingofelementsneededtodescribeinteractions be-tweentherealworld(physicalworld)andcomputer-basedapplications (cyberworld)suchassimulationordataanalytics,thatareusedto
sup-portdecisionmaking andtoenableimprovementsin thereal world. Theseelementsformafeedbackloopfromi)thephysicalworld,toii) dataacquisition,overiii)cyberworld,andiv)decisionsupportandback tothephysicalworld.Thephysicalworlddescribesthesysteminthereal world.Dataacquisitiondescribeshowthedatafromthephysicalworld isgathered,merged,andstored.Thecyberworlddescribes computer-basedapplicationsthatrunonthedataacquiredfromthephysicalworld togenerateresults.Theseresultsareusedtosupportdecisionmaking andimplementationof measuresin thephysical world.Thephysical worldwasalreadydescribedinChapter2:Technicalbackground.Data acquisitionwillbe presentedbrieflysinceithasalreadybeen exten-sivelypresentedinpreviousstudies[11,43].Themainfocuslieson Cy-berWorldwiththemulti-outputqualitypredictionmodelandthe bat-terycellproductiondesignaswellasDecisionSupportfordata-driven batteryproductiondesign.
3.2. Dataacquisition
Inthis phase,allrelevantdatasourcesareidentified.Asdepicted inFig.2,alldatacanbesubdividedintocertaindatasubsets:(1)input andoutputdata,(2)processparameters(PPs),orstatevariables(SVs)as wellas(3)IPFs.Besidestheidentificationofthedatasources,knowing thedifferentdatatypesisnecessary.Here,discreteandcontinuousdata aswellastimeseriesareofrelevance.Afteranalyzingwhichdataof whattypeandsourceistobeimplementedinlaterdeploymentphases, anacquisitionstrategyneedstobeputintoplace.Ingeneral,data ac-quisitionstrategiescan bedividedintotwotypes:(1)automaticand (2)manual.Theautomateddataacquisitionstrategycanbeenabledby automationtechnologiessuchassupervisorycontrolanddata acquisi-tionsystem(SCADA)ormanufacturingexecutionsystem(MES),which automaticallyacquiresdatafromdifferentsources,suchassensorsand programmablelogiccontrollers(PLCs).Anotherwayofautomateddata acquisitioncanbeachievedwithinformationtechnology(IT)software thatsupportsvariouscommunicationprotocols(e.g.OPCUA)andlogs thedataintoadatabase.Therearenostandardizedrestrictions con-cerningtheuseofcertaindatabases.Inindustry,SQL(StructuredQuery Language)databasesarewidelyspread,butalsotheuseofNoSQL(Not onlySQL)databasesisemerging.Alldatathatcannotbeautomatically acquiredhastobeimplementedviamanualdataacquisitionstrategies. Thiscanincludeoff-lineacquireddatafrome.g.productanalytics, op-erations,orsimulationsandcanbeintegrated,e.g.viaweb-interfaces, intothedatabase.
3.3. Cyberworld
3.3.1. Qualitypredictionmodel
Inthisphase,adata-basedmodelisderived,whichcanpredict cho-senFPPsofLiBcellsfromIPFsoftheirintermediateproducts.Therefore, thedataneedstobeprepared:transformed,cleaned,andifnecessary, thenumberoffeaturesformodeltraininganddeploymentcondensed (featureselection).Itisnecessarytobringthedataintosuchaformthat themodelscanutilizeitinthebestway.Ifthedataisnotstoredinthe appropriateform,itneedstobetransformed.Thedatacanalsohave missingorconstantvaluesorvaluesoflowvarianceandthereforecan becleanedinordertoobtainagoodqualityforthemodellingstep.Some modelscanperformbetterifthedatahasacomparablerangeandthe datacanbescaled(standardization,normalization,etc.).Models perfor-mancecanalsobeenhancediftheamountofauto-correlatedfeatures isreducedandthereforeeachfeaturecanrepresentapieceofunique unbiasedinformation.
AmodelforFPPpredictionisderivedbasedonIPFdata.Often mod-elscanperformbetterifthenumberoffeaturesmismuchsmallerthan theamountofdatasetsn(n>>m).Therefore,afeatureselectioncanbe performed,basedeitheronpurelystatisticalmethods(T-test,ANOVA, etc.)ormachinelearningmethods(forwardwrappers,backward wrap-pers,etc.).Sincethepremiseistoobtainamulti-objectiveprediction
model,amulti-objectivefeatureselectionmethodisrequiredinorderto selectasetofIPFswhicharerelevanttothemodelchosenFPPs.Machine learningalgorithmsthatarecommonlyusedforpredictionaremore of-tenspecialformsofmethodslikelinearregression,decisiontrees,and artificialneuralnetworks(ANN).Formodeltraining,twoconceptsare mostoftenused:80/20%datasplitandk-foldcross-validation.Inthe firstone,thedataissplitintotwo,adatasetwith80%ofthedatafor modeltrainingandadatasetwith20%ofthedataformodeltesting.It isnotunusual,however,tousea90/10%split,iftheamountofdatais low.Thesecondonesplitsthedataintoksets,ofwhichk-1setsareused formodeltrainingandoneformodelvalidation.Duringthisprocedure, k-modelsarebuiltandvalidated,soeverysetisusedasavalidationset once.Thebestpracticeistofirstsplitthedataintotrainingandtestset andthentousethetrainingsetfork-foldcross-validation.The cross-validatedmodelisevaluatedbasedontheindividualscorevaluesper trainedmodelorasameanscoreforthevalidationandtestdatasets.
Toenhancethepredictabilityofthemodel,ahyperparameter opti-mizationcanbeperformedduringthecross-validation. Hyperparame-tersareparametersdefiningcertainfunctionsinamodel.Thesecanbe alphas,ofaregularizationterm,thenumberofhiddenlayersinANN, orthenumberoftreesinrandomforests(RF).Hyperparametershelp tomodifymodelsandincreasetheirperformance,thoughtheycanlead toanoverfit,averyclosepredictionoftraineddatabythemodelthat leadstoapredictionlossforfurtherdata.Hyperparameterscanbe cho-senusingagridsearchoranoptimizationalgorithm.Gridsearchis suit-ableforcategoricalhyperparameterslikeactivationfunctionorsolvers, butalsointegervalueslikethenumberofneuronsinahiddenlayer. Optimizationalgorithmssuitnumericalvaluesliketheaforementioned alphas.
3.3.2. Batteryproductiondesign
In this step, the quality prediction model is used for a quality-orientedbatterycellproductiondesign.ThedesiredFPPvaluesare se-lectedandIPFvaluesaredetermined.Severaldifferentapproachescan be chosentoderivetherightvalues forIPFsinorder toachieve the desired valuesforFPPs.TheneededIPFvalues canbederivedusing variousapproaches,e.g.byusinganoptimizationalgorithmoragrid searchapproach.Kornasetal.2019usedamulti-variateoptimization withadesirabilityfunctioninordertoderivegoodvaluesof influenc-ingvariablesforbatterycapacityandbatteryweight[10]. Asimilar approachwasalsoperformedbyMeyeretal.2019formaximizingthe qualityofcutelectrodesheets[44].Anoptimizationapproachworksby definingacostfunction(e.g.desirabilityfunction)basedonthetarget valueandthevariationofinputparameterstodeterminetheoptimal subsetthatresultsintheminimalcostfunctionvalue.Thesubspaceof theparametersdoesnotneedtobedefined,butthealgorithmsusually acceptboundariesfortheparameters.Inthecaseofagridsearch,no op-timizationalgorithmisneededandthesubspaceoftheinputparameters isdefinedbyvaryingthem.Thebestparametersubsetisderivedwhen thedesiredvalueforthetargetisobtained.Theseapproachescanbe appliedtothequalitymodel,wheretheIPFsaretheinputparameters thatarevariedandthetargetsaretheFPPs.However,anFPPtarget valuewithatolerance/intervalis selected,becausethefluctuationof thebatterycellproductionprocessesmakesithardtoachievean ex-actmatchforFPPsofproducedLiBcells.WithintheintervalofFPPs, differentsubsetsofIPFscanbedetermined.Thesedifferentsubsetsare evaluatedcloselywithregardtotheusedqualitypredictionmodeland theuseddata.
3.4. Decisionsupport
This step explains howthe deploymentof thequality prediction modelandtheprocedureofbatteryproductiondesigncanbeusedto enabledecisionsupportinplanningandoperatingbatteryproduction asdepictedinFig.7.Thequalitypredictionmodelisdeployedwith ac-cesstothedatabase,whichstoresallpreviouslyacquireddataofIPFs
Fig.7.Decisionsupportinplanningandoperationofbatteryproduction.
andFPPs.Batteryproductiondesignisdeployedwithaconnectionto thequalitypredictionmodel.Furthermore,aproductionprocess simula-tionisusedtopredictPPsbasedonIPFsderivedfrombatteryproduction design.
Decisionsupportintheplanningofbatteryproductionstartswiththe customerandproductionplannerdefiningthedesiredFPPs/targetFPPs thatareusedbythequalitypredictionmodelandbatteryproduction designtogeneratepotentialIPFsthatareneededtoproduceabattery cellwithdesiredFPPs(seeFig.7).Theprocessexpertthatmightexclude unreasonableIPFvaluesthenverifiesthesepotentialIPFs.Thenthe po-tentialIPFsareusedinaproductionprocesssimulationtogeneratePPs thatarerequiredtoobtainneededIPFs.
Decisionsupportintheoperationofbatteryproductiontakesoff at thepointwherethefirstintermediateproductisproduced(seeFig.7). TheacquiredactualIPFsaretransferredtothequalitypredictionmodel andupdatethepredictionforFPPsthatmightresultafterfinishingthe productionthewayithasbeenplanned.Iftherehavebeenany devi-ationsfromtheplannedIPFvalues,thebatteryproductiondesigncan reevaluatethecurrentsituationusingactualIPFsasconstantvaluesand varyingtheremainingIPFsinsuchawaythattheinitiallydesired tar-getFPPscanbereached.ThenewpotentialIPFsarethenagainverified bytheprocessexpertandusedbytheproductionprocesssimulationto generatenewPPsfortheupdatedIPFs.WiththesePPs,theremaining processesareexecuteduntiltheproductionisfinishedoranew devia-tionoccurs.
4. Casestudyqualitysystem
ThecasestudywasconductedinthefacilitiesoftheBattery Lab-FactoryBraunschweig(BLB),aresearchLiBcellproductionlinewith industry-scale productionmachinesand processes,coveringall steps
frommaterialpreconditioningtoformationandaging.Allinvestigated LiBcellsweremanufacturedinBLB.
4.1. Dataacquisition
Thebasisforthelaterdeployedqualitypredictionmodelisthe ac-quiredIPFandFPPalongwiththeLiBmanufacturing.TheBLBoperates ahighlyflexibleproductionline,consistingofsemi-continuousand dis-creteproductionprocessesincludingconverginganddivergingmaterial flows.Furthermore,someprocessand(intermediate)productanalytics areperformedin-lineandothersoff-line,resultingindifferentdatatypes andsources.Therefore,ahighlyflexibledataacquisitionisrequired. Thedataacquisitionstrategyforthisworkisbuiltuponthelegacy pre-sentedin[11].Anadaptedstructureservingthepurposeofthispaper isdepictedinFig.8.Dependingonthedifferentdatatypes,presented inFig.3,aspecificdataacquisitionstrategyneedstobeapplied.
ThebottomlayerinFig.8depictstheautomateddataacquisition, consistingof sensorsandPLCs(e.g. in-linethickness). These entities representthedifferentdatasourcesaccessibleforautomatedacquisition measures.This data,includingauniqueIDandaspecifictimestamp, is acquiredviaindustrialcommunication protocolssuchasOPCUA, Profinet,ModbusbyNode.js,ajavascriptlibrary,andloggedintoa SQLdatabase,whereitisprocessedandisaccessible.Theoff-linedata fromsourcessuchasintermediateproductanalyticsandfinalproduct analyticsareacquiredmanuallyviaawebinterfaceintoanengineered datawarehouse.Thisisenabledbyawebserverwithavisualinterface. ThisprocessisdepictedintheupperlayerofFig.8.
4.2. Predictivequalitymodel
Themodellingpartwas performedonz-foldedpouch bagbattery cellswith15anodesandcathodes(ESC– electrodeseparator compos-ite)andwith10cathodesandanodes.AnodematerialwasPVDF,SFG6L,
Fig.8.ThedataacquisitionstrategyoftheBatteryLabFactoryBraunschweig(BLB)basedon[11].
GraphiteC65andSMGA4andthecathodematerialwasPVDF,SFG6L, GraphiteC65,NMC111,andNMP.Thegeometryoftheelectrodesheets was150mmx110mmforanodesand145mmx105mmforcathodes. Theseparatorwas fromSeparion (ceramicseparatord=28μm)and theelectrolytewasBra-003.AsmodellingsoftwarePython3.8.5was usedwiththefollowinglibraries:Numpy1.14.5,Pandas0.25.1[45], andscikit-learn0.21.2[46]formodelling,andhyperopt[47,48]for hy-perparameteroptimizationandSALib[49]forsensitivityanalysis.The computerusedformodellingwasaDellprecisionminitower3000series withanIntel® CoreTM i56500processor(3,2GHz,6MBcache)and
32GBDDR42.133MHzRAM(Random-AccessMemory).Modellingwas solelyperformedusingtheCPU(centralprocessingunit).Forthecase study,191LiBcellswithatotalamountof1029IPFswereused.The IPFsconsistedofdatafromalltheprocessesuptotheformation.The formationandagingsetupwasnottakenintoaccount,sinceitwasnot varied.DataacquiredduringformationandagingwereconsideredFPP, e.g.chargingordischargingcapacitiesundercertainchargeordischarge rates,innerresistances,self-discharge,etc.Furthermore,theproduced LiBcellsweretheresultsofcasestudiesperformedbyvaryingcalender, lasercutting,andz-foldingprocessparameters,acasestudy investigat-ingthereproducibilityofLiBcellproduction,andacasestudyvarying theelectrolyteamount[22,25,26,29].Therefore,expectedresultswere consideredtobeheavilyinfluencedbytheperformedproduction vari-ation,meaninginfluentialIPFsofprocessessuchasmixing,dispersing andcoatingmightappearthathaveaminorinfluenceontheFPPs.
4.2.1. Developmentofthequalitypredictionmodel
Theprocessofthequalitypredictionmodeldevelopmentisdepicted inFig.9andisexplainedinthefollowing.Theacquiredrawdatawas storedinaSQL databaseandwasfurthertransformed formodelling purposesintothefirstnormalform,whereeach cellwasa rowand eachIPForFPPwasacolumn.Thus,eachcellcouldberegardedasa datasetcontainingallcorrespondinginformationofitsIPFsandFPPs. Afterthetransformationstep,themanualcleaningoftheIPFdatawas performed.IPFswereremovedthatweremeasuredduringtheprocess onlyifacorrespondingIPFexistedthatwasmeasuredrightafterthe endoftheprocess.
Afterthemanualdatacleaning,afeatureengineering/feature gen-erationstepwasperformed.Newfeaturesweregeneratedbasedonthe existingonestogiveabetterinsightintoproductiondata.Asan exam-ple,afeaturenamed‘cathodeamount activematerial[g]’was calcu-latedforeachcellbasedonthecoatinglayergrammage,thegeometry
Fig.9. Developmentprocessofthequalitypredictionmodel.
ofcutelectrodes,andthenumberofcutelectrodesinthecell.This fea-turegivesameaningfuldistinctionbetween15and10ESCbatterycells. ‘Cathode/anodetheo.capacityratio[%]’istheratioofthecalculated theoreticalcapacityofthecathodedividedbythetheoreticalcapacity oftheanodeinthebatterycell.Thetheoreticalcapacitiesarebasedon
Table1
Selectedlinearandnon-linearIPFsubsetswithcorrespondingprocessesoftheiracquisition.
Process step Linear IPF subset Non-linear IPF subset
Mixing/dispersing ‘anode mixing mean PSD [m-6]’, ‘anode suspension viscosity fitting value’, ‘anode suspension dry weight [%]’, ‘cathode amount active material [g]’ 1, ‘cathode/anode theo. capacity ratio [%]’2
‘anode mixing std PSD right [-]’, ‘anode suspension viscosity fitting value’, ‘cathode mixing PSD d10 [m ∗ 10-6]’, ‘cathode amount active material [g]’ 1, ‘cathode/anode theo. capacity ratio [%]’ 2 Coating/drying ‘cathode temperature during conductivity measurement [°C]’,
‘cathode amount active material [g]’ 1, ‘cathode/anode theo. capacity ratio [%]’ 2
‘cathode uncalendered porosity value 418-10 [-]’, ‘cathode uncalendered electrode thickness [m ∗ 10-6]’, ‘cathode amount active material [g]’ 1,
‘cathode/anode theo. capacity ratio [%]’ 2 Calendaring ‘cathode total porosity calculated [ml]’,
‘cathode total porosity measured [ml]’, ‘cathode adhesion strength [kPa]’, ‘cathode adhesion Fmin [N]’,
‘electrolyte per thickness per compartment [ml/mm/comp]’ 3∗ , ‘electrolyte porosity volume ratio’3
Laser cutting ‘anode delamination mean [m ∗ 10-6]’, ‘anode melt formation [m-6]’, ‘cathode mean pyrometer FWHM [s]’, ‘cathode mean pyrometer T150 [s]’, ‘cathode amount active material [g]’ 1, ‘cathode/anode theo. capacity ratio [%]’ 2
‘anode mean pyrometer FWHM [s]’, ‘cathode mean pyrometer FWHM [s]’, ‘anode delamination mean [m ∗ 10-6]’, ‘cathode mean pyrometer integral TS [Ks]’, ‘anode mean pyrometer T150 [s]’, ‘cathode mean chamfer angle [°]’, ‘cathode amount active material [g]’ 1, ‘cathode/anode theo. capacity ratio [%]’ 2 Assembly (Z-folding) ‘electrodes overlapping rate [%]’ 4,
‘electrodes overlapping rate manual [%]’ 5, ‘cathode amount active material [g]’ 1
‘electrodes overlapping rate [%]’ 4, ‘cathode amount active material [g]’ 1, ‘cathode/anode theo. capacity ratio [%]’ 2
Filling ‘battery cell thickness per compartment [mm/comp]’, ‘electrolyte per thickness per compartment [ml/mm/comp]’ 3, ‘electrolyte porosity volume ratio’ 3
1 ‘cathodeamountactivematerial[g]’isadjustedbytherecipeinmixing,thecoatinglayergrammageincoating/drying,thegeometryofcutelectrodesheets,
andbytheamountofESCinassembly(Z-folding)
2 ‘cathode/anodetheo.capacityratio[%]’isadjustedbytherecipeinmixing,thecoatinglayergrammageincoating/drying,thegeometryofthecutelectrode
sheets,andbytheamountofESCinassembly(Z-folding)
3 ‘electrolyteperthicknesspercompartment[ml/mm/comp]’and‘electrolyteporosityvolumeratio’areadjustedincalendaringandinelectrolytefilling 4 ‘electrodesoverlappingrate[%]’calculatedbasedonvaluesmeasuredbytheZ-folder
5 ‘electrodesoverlappingratemanual[%]’measuredmanuallyaftertheprocess
thenumberofactivematerialsofbothelectrodes.Thisfeature charac-terizesthemisbalancebetweenplannedcathodecapacityandplanned anodecapacity.Anothertwofeaturesrepresenttheinfluenceof calen-daringandelectrolytefillingvariations.‘Electrolyteperthicknessper compartment[ml/mm/comp.]’containstheinformationofhowmuch electrolyteisinthecellcomparedtoitsthicknessandthenumberofits compartments(opposinganodeandcathodesides).‘Electrolyteporosity volumeratio’isaratiooftheelectrolytevolumedividedbythevolume ofthetotalporosityofthewholeelectrodematerial(anodesand cath-odestogether).Itistherelativevalueofhowmuchmoreelectrolyteis presentinthecellthanintheelectrodematerialpores.Manyother fea-turesweregenerated,butthesewerethemostprominentchoseninthe featureselectionstep.
Inautomatedcleaning,IPFswithahighamountofmissingvaluesor entirelyconsistingofconstantvalueswereremovedinordernottoaffect themodellingnegatively.LiBcellswithahighamountofmissingvalues wereremovedaswellsincethesedatasetsalsomighthaveanegative influenceontherelevanceofsomeIPFs.FPPoutlierswereremoved ex-ceeding1.5timestheinterquartilerangeiftheycouldnotbeexplained bytheIPFdatasincetheyrepresentanunexplainablevariancewithin FPPdata.AllFPPandIPFdataweremin-max-scaledbetween0and1, inordertoachievebettercomparabilitybetweenIPFs.IPFswithalow variancewereremoved,astheyrepresentdatawithasmallvariation andthereforewereconsideredtohaveaninsignificantinfluenceon tar-getvalues.ThecorrelationofIPFswithinthedatasetswasdetermined usingPearson’sr.Withinthecorrelationcoefficients,29clusterswitha thresholdofr≥0.68wereidentified,meaningthatwithintheclusters theIPFsstronglycorrelatetoeachother(r≥0.68).ThreeFPPswere se-lectedforthecasestudy:maximaldischargecapacityobtainedafterthe thirddischargecycle(max.capacity),self-dischargeduringtheaging step(self-discharge)andstateofhealthafter400chargeanddischarge cycles(SOHafter400cycles).Foreachcluster,oneIPFwaschosenwith
thehighestcorrelationtooneoftheFPPs,leadingtoasubsetof50IPFs. Aftertheseprocedures,theresultingdataconsistedof155LiBcellsand 50 IPFs.Datapreparationstepswiththebiggestdatareductionwere manual datacleaning(508removed IPFs),removalof columnswith ahighamountofmissingvalues(142removedIPFs),andremovalof highlyco-correlatedIPFs(234removedIPFs).
FortheselectionofthemostrelevantIPFsabackwardwrapperwas used, arecursivefeatureelimination(RFE)algorithm[50]. The algo-rithmworksiteratively.Duringeachiterationstep,apredefined num-beroffeatureswasexcludedbasedonafeaturerankingattribute.Two differentapproacheswereadoptedusingtheRFE,selectionof anIPF subsetwithlinearinterrelationtotheFPPs,andselectionofanIPF sub-setwithanon-linearinterrelationtotheFPPs.Fortheselectionofthe linearfeaturesubset,thefeaturerankingattributeswerethecoefficients oftheLasso-Larsregression(leastabsoluteshrinkageandaselection op-eratormodelwithaleast-angleregressionfit).Fortheselectionofthe non-linearfeaturesubset,thefeature-rankingattributewasthefeature importanceprovidedbyanRFregressor,adecisiontree-based ensem-blelearningalgorithm.Inordertoachievethemulti-objectivityofthe featureselection,featureselectionwasperformedforeachFPP sepa-ratelyon thegivenIPF data. ChosenIPFs weremergedandthe du-plicates wereremoved. Thepredefinednumber of featurestobe se-lectedforeachFPPwassetto6inordertokeepthetotalamountof selected featureslowfor modelling(n>>m).Thenumberoffeatures above 6resultedin anincreaseof totalfeaturestobetoolargeboth forlinearandnon-linearsubsets,withnofurthermodelperformance increase.Thenumberoffeaturesbelow6resultedinlowermodel per-formance.ThefinalsubsetsoflinearIPFsconsistedof17featuresandfor non-linearIPFsof16.TheseIPFsubsetsarelistedinTable1.Though highlycorrelatedIPFswereremoved,somecanbe seeninthelinear IPF subset. This is due totheselectionof IPFs toeach of thethree FPPs from theclusters of thehighly correlatedIPFs. Insome cases,
Table2
ModelhyperparametersformodelsbasedondifferentIPFsubsets.
IPF subset Model Hyperparameters Hyperparameter values
LLR Linear Alpha 0.0001 ∗
Non-linear Alpha 0.0001 ∗
RF Linear Number of trees 32
Non-linear Number of trees 32
ANN Linear Activation functionSolverAlphaToleranceHidden layer neurons ReLUL-BFGS0.0014 ∗0.0005 ∗(13,35,29) Non-linear Activation functionSolverAlphaToleranceHidden layer neurons ReLUL-BFGS0.0292 ∗0.0007 ∗(16,11,42) ∗roundedto10-4
forallthreeFPPsthesamefeaturewasselected,andinsomedifferent IPFs.
Afterthefeatureselection,IPFandFPPdataweresplitinto90/10 %withtrainingandvalidationdatabeingwithin90%and10%for latermodeltesting.Thissplitwaschosenduetothelowamountofdata. Lasso-Larsregression(LLR),RF,andANNmulti-layerPerceptron regres-sorswereusedwith5-foldcross-validation(CV)asatrainingapproach formodelling.These modelswereused fordemonstrationreasonsas theyrepresentthemoststandardmachinelearningpredictionmodels currentlyused.Specialformsofthesemodelsfitthesameapplication.
Asafurthersteptoimprovethemodels,hyperparametersofthe mod-elswereoptimizedduringcross-validationusingapossibly-stochastic functionwiththeTreeofParzenEstimators(TPE)asasearchalgorithm [51].Onlyfortwohyperparameters,tobeexplainedlater,agridsearch algorithmwasused. Aslossfunction,minimizedfortheoptimization algorithms,therootmeansquareerrorwaschosen(RMSE).Thespace fortheoptimizationwaskeptconsiderablysmall,toeasepossible modi-ficationofdatapreparationandmodellingstepsandtokeepthemodels frombecomingtoocomputationallyexpensive.ForLLR,the regulariza-tionparameteralphawasoptimizedwithinthespacebetween0.0001 and0.1.Lowervalueswouldleadtoareducedregularizationand there-forecouldleadtoanoverfitandhighervaluescouldleadtoan under-fit,bothinsufficientforthepredictionofdata.ForRF,thenumberof treeswasoptimizedwithinthespacebetween1and50.Basedonthe authors’experience,RFisacomputationallyexpensivealgorithmand thenumberoftreesshouldbelimitedifthealgorithmsshouldbeused forfurtherinvestigation.ForANN,theactivationfunction,solverfor weightsoptimization,theregularizationparameteralpha,thetolerance fortheANNweightsoptimization,andthenumberofneuronsinthe hiddenlayerswerechosenashyperparametersforfurtheroptimization. Theactivationfunctionandthesolverwerechosenusingagridsearch. Fortheactivationfunction,rectifiedlinearunit(ReLU),tanh,anda sig-moidfunctionwereconsidered.Forthesolver,thefollowingalgorithms wereconsidered:astochasticgradientdescent(sgd),aniterative op-timizationalgorithmapproximatinggradientdescentstochastically,an AdaptiveMomentEstimation(adam),aspecialversionofsgdfor ma-chinelearningapplications,anda limited-memory Broyden-Fletcher-Goldfarb-Shannonalgorithm(L-BFGS),aquasi-Newtonoptimization al-gorithm.TheregularizationparameteralphaandthetoleranceforANN weightsoptimizationwereoptimizedwithinthespacebetween0.0001 and0.1.Thenumberofhiddenlayerswaschosentobe3,butthe num-berof neuronsineach hiddenlayerwas optimizedwithinthespace between1and50.Table 2showsthehyperparametervalues chosen forthecorrespondingmodels.BesideshyperparametersofANN, hyper-parametersofLLRandRFarealmostsimilarforlinearandnon-linear featuresubsets.ThedifferencebetweenhyperparametersinANNlinear andnon-linearsubsetsisduetothedifferentnumberoffeaturesinthe IPFsubsetsandbecausetheANNperformsdifferentlyifitsnumberof neuronsinthehiddenlayersischanged.
Forthe evaluationof the models, theirprediction is depictedin Fig.10.Ithelpstounderstandhowgoodthepredictiononthelevel ofsinglebatteriesis,tofindweaknessesofthecorrespondingmodels. Forexample,thepredictionofmax.capacityofthelast13cellsispoorly fortheLLR,slightlybetterforANN,andmuchbetterforRF.Itisnotan
indicatorthatRFisoverallbetter,butitcouldmeanthatRFpredictsthe max.capacityof10ESCcellsbetterthantheothermodels.Furthermore, thisdepictioncangiveafirstglimpseintohowgoodamodel general-izesorwhetheritmighthaveanoverfit.Table3showsthemeanmodel scoresanderrorsofthevalidationdatasetsofthe5-foldCVandthetest datasets.Thescore displayedisthecoefficientofdetermination(R2),
whichreflectsthevarianceofthepredictedtargetvaluesbasedonthe inputvaluesofthemodel.Theerrorsarethemostcommonerrorsused in modelling,meanabsoluteerror (MAE)andRMSE.Additionally,it showsthepredictiontimefor100datasetsforeachmodel.Thesescores anderrorshelptounderstandthequalityofthemodel.Thescoresand theerrorsarethemeanvaluesof 5different validationdatasetsthat werecreatedduringCVandofthetestdataset.Theyshowhowgood amodelisbasedondifferenttrainingandvalidationdata.Amongthe validationdata,thescoresanderrorsshowlittledifferencebetweenthe modelswiththeANNbuiltonthenon-linearfeaturesubsethavinga betterperformancebyasmallmargin.LLRshows,asalreadyseenin Fig.10,clearlowerperformanceforR2.Thefeaturesusedtobuildthe
LLR areall putintolinearrelationwithinthemodel,contrarytoRF andANN.ItindicatesthatthebatterycellIPFsmighthaveanon-linear interrelationwithitsFPPsandthattheLLRismissingoutonthis infor-mation.Theresultsofthetestdatashowintotalforallmodelsaslightly betterperformance.Thiscouldhappenbecauseoftherandomizedsplit oftrainingandtestdata.Sincebatterycellsareproducedinbatches, theIPFsofthesebatterycellsareoftensimilarforthewholebatch(e.g. slurryIPFslike viscosity,coatedlayerthicknessof anelectrodecoil, etc.).Ifthemodellearnshowtopredictabatterycellbatch,theycan betterpredictbatterycellsfromthesamebatchinthetestdataset.For thetestdata,thebestmodelappearstobetheRFwithalinearfeature subset.Nonetheless,theANNwithanon-linearfeaturesubsetshowsa constantlygoodperformance.Additionally,thepredictiontimeshows thatLLRandANNpredictthefastest,0.02sand0.03srespectively.RF needsforthesameamountofdataalmost20timeslonger.Forfurther investigations, theANNmodelbasedon anon-linear IPFsubsetwas used.
TobetterunderstandtheinfluenceofIPFsthatwereusedtomodel theFPPs,intheANNmodelbasedonanon-linearIPFsubset,aSobol sensitivityanalysis[52–54]wasperformedtocalculatefirst-, second-orderandtotal-ordersensitivityindices.Thesensitivityindicesofthe first-order describethecontributionofasinglemodelinputvariable (single IPF) to the variance of a single model output (single FPP). Second-orderindicesdescribethejoint contributionoftwomodel in-putvariables(twoIPFs)tothevarianceofasinglemodeloutput(single FPP).Total-orderindicesdescribethecontributionofasinglemodel in-putvariable(singleIPF)tothevarianceofasinglemodeloutput(single FPP)includingthefirst-,second-andallhigher-orderinteractions.The indicesaresetbetween0and1,with1havingaverystronginfluence and0havingnoinfluenceatall.Thesesensitivityindicesaredepicted foreachIPFandFPPinFig.11.Eachheatmaprepresentsthesensitivity indicestothecorrespondingFPPasdescribedinthetitle.The diago-nallaneofeachheatmap,whereIPFrowsandIPFcolumnsareequal (e.g.row1andcolumn1arebothIPF1),arethesensitivityindicesof thetotal-order,describingthetotalinfluenceoftheIPFonthe corre-spondingFPP.Valuesabovethediagonalline,whererowsdonotequal
Fig.10. BatterycellsandpredictionsofLasso-Larsregression(LLR)inthefirstrow,randomforestregressor(RF)inthesecondrowandartificialneuralnetwork (ANN)inthethirdrowusingalinearandanon-linearIPFsubsetformax.capacity,self-dischargeandstateofhealth(SOH)after400cycles.
columns,arethesensitivityindicesofthesecond-orderdescribingthe influenceofthetwoIPFsonthecorrespondingFPP.Insteadof depict-ingthefirst-orderindices,thesensitivityanalysiswasperformedlater onbyvaryingoneinputvariableatatime.ThechangesintheFPPsare displayedseparatelyinFig.12.
Thestrongestinfluencesthatcan beseenaremostly ofthe total-orderwithsensitivityindicesvaluesofupto0.74,whereasthe second-ordersensitivityindicesareonlyupto0.14.Mostnotableinfluenceson themax.capacity(indicesabove0.1)aretheIPF15(0.74),amountof cathodeactivematerial(‘cathodeamountactivematerial[g]’)andIPF 2(0.09),themeanvalueofthefullwidthathalfmaximum(FWHM) temperaturepeakmeasuredwithapyrometerduringlasercuttingof cathodesheets(‘cathodepyrometerFWHM[s]’).Forthemax. capac-ity,itisclearthatthemoreactivematerialisinthebatterycellthe higheristhepotentialcapacityandthataninfluenceoncathodesheets likeduringlasercuttingmightimpact thecapacityaswell.Most no-tableinfluencesontheself-discharge(indicesaroundorabove0.1)are
IPF2(0.11),IPF4(0.09),IPF5(0.16),IPF7(0.27),IPF8(0.1),IPF 13(0.25),IPF15(0.4),IPF16(0.08)andasecond-orderinfluenceof IPF7andIPF15(0.14).IPF4isthemeanvalueofthedelamination ofcutanodesheets(‘anodedelaminationmeanm-6’).IPF5isthe in-tegraloftemperaturepeakovertimemeasuredwithapyrometer dur-inglasercuttingofcathodesheets(‘cathodemeanpyrometerintegral TS [Ks]’).IPF7is theamountof electrolytedividedbybatterycells’ thicknessandthenumberofitscompartments(‘electrolyteper thick-nesspercompartment[ml/mm/comp.]’).IPF8istheratiobetweenthe electrolytevolumeandthetotalbatterycellelectrodeporosityvolume (‘electrolyteporosityvolumeratio’).IPF13istheoverlappingrate be-tweencathodeandanodesheetsmeasuredinpercentageduringthe z-foldingprocess(‘electrodesoverlappingrate[%]’).IPF16istheratio betweenthetheoreticalcapacitiesofcathodesincomparisontoanodes of thewhole batterycell(‘cathode/anode theo.capacityratio[%]’). TheseIPFsindicateapotentialinfluenceoftheelectrodecutting pro-cessontheself-dischargeofbatterycells.Thesecond-orderinfluenceis
Table3
Meanscoresanderrorsofvalidationandtestdatasubsetsofthe5-foldCVformodelsbasedondifferent IPFsubsets.R2– coefficientofdetermination,RMSE– rootmeansquareerror,MAE– meanabsoluteerror.
Model IPF subset 5-fold CV validation data Test data Time of predicting 100 datasets R 2 RMSE MAE R 2 RMSE MAE
LLR Linear 0.74 0.09 0.06 0.80 0.08 0.06 0.02s Non-linear 0.76 0.09 0.06 0.78 0.08 0.06 0.02s RF Linear 0.81 0.08 0.05 0.88 0.06 0.04 0.42s Non-linear 0.80 0.08 0.05 0.84 0.07 0.05 0.43s ANN Linear 0.81 0.07 0.05 0.83 0.07 0.05 0.03s Non-linear 0.84 0.07 0.05 0.85 0.07 0.05 0.03s
Fig.11. ResultsofthesensitivityanalysisoftheANNbasedonnon-linearIPFsubsetsusingsensitivityindices.Total-ordersensitivityindicesareonthediagonal laneswhereIPFnumbersareequal,thesecond-orderindicesareinthefieldsoftheuppertrianglerepresentingthesecond-orderinteractionbetweenIPFsinthe columnsandtherows.
Fig.12. ResultsofthesensitivityanalysisforallFPPsoftheANNmodelwithnon-linearIPFsubsetsbyvaryingonevariableatatime.
between‘electrolyteperthicknesspercompartment[ml/mm/comp.]’ and‘cathodeamountactivematerial[g]’.Sincethecompressionrateof thecathodeswasvariedforsomebatterycellswithinthecasestudy, theporosityisdifferentforthosecells.Theamountofelectrolytewas variedaswell,leadingtotheassumptionthatthewettingdegreeand theabundanceofelectrolyteinrelationtotheamountofcathode ma-terialmayinfluencetheself-dischargeaswell.Mostnotableinfluences ontheSOHafter400cycles(indices aroundorabove0.1) areIPF4 (0.08),IPF10(0.09),IPF12(0.08),IPF13(0.64)andIPF16(0.11).IPF 10istheuncalenderedcathodethickness(‘cathodeuncalendered elec-trodethickness[m-6]’).IPF12istheD10oftheparticlesizedistribution (PSD)ofcathodeslurrymeasuredduringthemixingprocess(‘cathode mixingPSDd10[m-6]’)meaningthat10%ofparticleshavesmaller and90%ofparticleshavelargerdiametersthanD10.Asindicatedby insilicowork,PSDmayimpacttheperformanceandthedegradationof batterycells[55].Furthermore,itwasshownexperimentallythat de-creasingelectrodes’overlappingrate(orelectrodedepositionaccuracy) influencethebatterycells’dischargecapacitynegatively[25].High dis-proportionsbetweenanodeandcathodecapacitiesplayaroleinaging processesofcapacitydegradation[56].
ThoughthesensitivityindicesshowwhethercertainIPFsinfluence theFPPs,theyonlyshow whetheritisastronginfluenceornot.The informationonwhethertheinfluenceisapositiveone(increasingIPF valuesleadtoanincreaseinFPPvalues)oranegativeone(increasing
IPF valueslead toadecrease inFPP values)isnotgiven. Therefore, theinputdatafortheANNmodelbasedonanon-linearIPFsubsetwas variedwithonevariableatatime(seeFig.12)toseetheirinfluence onallFPPs.Thediagramsshowthenon-linearnatureofsomeIPFsand underlinetheirimpactontheFPPs.ThisdepictionofFPPsensitivity to-wardsvariousIPFscanbeusedtofurtherstrengthentheunderstanding ofmodelledinterrelationsandhelptoevaluatethemodel.Itcanalready provideaglimpseofwhethertheinterrelationsappearastheywere ex-pectedorwhetherthemodel,althoughithasagoodscoreandalow errorrate,representstheinterrelationswrongly.
InfluencesofIPFsonFPPsmodelledbytheANNbasedona non-linearfeaturesubsetaredepictedinFig.12.Theyshow,asalready de-scribedbythesensitivityindices,stronginfluencesof‘cathodeamount activematerial[g]’onmax.capacityandself-discharge,of‘electrodes overlappingrate[%]’ onself-dischargeandSOHafter400cycles,of ‘Cathode meanpyrometerFWHM [s]’on max. capacity, of ‘cathode meanpyrometerintegral[Ks]’onself-discharge,andof‘electrolyteper thicknesspercompartment[ml/mm/comp.]’onself-discharge.The ‘an-odedelaminationmean[m-6]’showsinthissensitivityanalysisan in-fluenceonmax.capacity,despiteasensitivityindexof0.04,although thesensitivityindicesof0.09and0.08indicateits influenceon self-dischargeandSOHafter400cyclesrespectively. ‘Electrolyteporosity volumeratio’,‘cathodeuncalenderedelectrodethickness[m-6]’, ‘cath-ode mixingPSD d10[m-6]’and‘cathode/anode theo.capacityratio
Fig.13. Processofbatteryproductiondesign.
[%]’,donotshowinthisanalysisdespitetheirsensitivityindicesshow noinfluencesonself-dischargeorSOHafter400cycles.However,the sensitivityindicesdepictedinFig.11donothavetomatchthe influ-encesshowninFig.12,becausethesensitivityindicesareofthe total-orderandthereforeincludethefirst-orderinfluenceaswellasall higher-orderinfluences.ThismeansthattheseIPFsnotshowingstrong influ-encesinFig.12probablyhavestronginteractionswithotherIPFs. Fur-thermore,itisimportanttounderstandthateventhoughsome interre-lationsappearlogicalandcomprehensible,othersarederivedbasedon availabledata.ThismeansthatiftheIPFswerenotvariedtotheirfull range,theinterrelationbetweentheseIPFsandtheFPPsmightbe mod-elledwithastrongbias.Therefore,anyinterpretationofthedepicted interrelationsbetweennotvariedIPFsandFPPsshouldberegardedas anindicationofapossibleinterrelationwithintherangesofthedata.
Thedevelopedmodelsshow,likemostmachinelearningmodels,a limitationinitspredictabilitytonewandunexperienceddata.Contrary tophysicalmodels,thequantitativeinterrelationbetweeninputand out-putdataislearnedwithinthetrainingprocessofamachinelearning model.Itishighlyinfluencedbythemodels’architecture,inputand tar-getvaluescales,andtheirvariablespace.Variationofinputdatacanbut doesnothaveto,covertheentirevariablespaceofthatinput,whereas anunvariedinputdatamightslightlyfluctuatearoundcertainvalues. Whenvariedandunvariedinputdataistransformedandrescaled,itis difficultforthemodeltodistinguishwhichinputreallyvariesstrongly andwhichdoesnot.Therefore,inorderforthemodeltopredictnew dataornewvariations,thesevariationsneedtobefirstintroducedtothe model.Thesamerecommendationistruefortheextensionofvariable spaces.Eventhebestmachinelearningmodel,givenalimitedamount ofdata,mightnotextrapolatethepredictionwellenoughtopredictdata outsideknownvariablespace.
4.2.2. Batteryproductiondesign
TheprocessofbatteryproductiondesignisdepictedinFig.13and describedinthefollowing.ItrequiresboththerangeofIPFsthatisbased onexistingdataaswellasthetargetvaluesandtherangefordesired FPPs.TherangeofIPFsis neededtogeneratebatterysamplesbased onacombinationofpossibleIPFvalues.Targetvaluesandranges of desiredFPPsareneededtoselectthegeneratedbatteriesthatliewith theirpredictedFPPswithinthoseranges.Furtherclusteringisusedto consolidatesampledbatterycellswithsimilarIPFvalues.Theseclusters arethencomparedtorealbatterycells.
AgridsearchapproachwaschosentogenerateIPFdataandtouse thisdatatopredictFPPs.SincenotallparametersoftheLiBcell
pro-ductionwerevariedandthereforenotallpossiblecombinationsofthe selectedIPFswereavailabletothemodel,theparametersubspacewas restricted.Aquantiletransformerwasusedtotransformunivariate dis-tributedIPFdataintothedatasubspaceofavailableIPFs.Thenthe trans-formedIPFvalueswereusedtopredicttheFPPs.Intotal10800datasets ofLiBcellsamplesweregenerated.InFig.14,thegeneratedandthereal batterycellsaredepicted.Twoclustersofrealbatterycellscanbeseen veryclearly,onewithcellsthathave15ESCandtheotheronewithcells thathave10ESC.Althoughthequantiletransformerhelpstoreproduce therestrictedparametersubspace,somenewbatterycells’ character-isticscanbeseenofthegeneratedbatterycellsarrangedbetweenthe twoclustersoftherealbatterycells.Thesenewcharacteristicscanbe newcombinationsofdifferentIPFvaluesthatdonotappearinthe orig-inaldata.Incasethesenewcharacteristicsareofinterest,itneedsto beverifiedbyprocessandbatteryexpertswhethertheyarephysically plausible.
Toprovidecomparabilityandanevaluationoftheconcept,seven batterycells wereexcludedfrommodeltraining andtesting,i.e. the modelhasno informationofthese cells.Four of thesecells have15 cathode-anodecompartmentsandthreeofthemonly10.Basedonthese cells,thetargetvaluesforFPPsandtheirtolerance/intervalwere cho-sen.Thetargetvaluesweresettothemeanvaluesofthecellsandthe tolerance/intervaltoplusthreetimesthestandarddeviationandminus threetimesthestandarddeviation(seeTable4).Basedonthiscriterion, generatedcellswhoseFPPvaluesliewithinthedefinedintervalwere selectedforfurthersteps.
TheirIPFsweredeterminedandcomparedtotheIPFsfortheseven LiBcells.ThiscomparisoncanbeseeninFig.15,ahistogramforeach IPF depictinggeneratedandrealbatterycells.Fig.15shows10 ESC generated batterycellsin blue,15ESCgeneratedbatterycellsin or-ange,10ESCrealbatterycellsingreen,and15ESCrealbatterycellsin gray.Intotal54generatedbatterycellswerefoundwithinthedefined intervals:26for15ESCand28for10ESC.Itcanbeseenthat,though somegeneratedbatterycellsshowsimilarIPFvaluestorealbatterycells, notallofthemdo.Thisdepictionofrealandgeneratedbatterycellsin Fig.15 showsthatthegenerated batterycellsaretightlydistributed aroundspecificvalues,partiallyupto12or25perhistogram.This in-dicatesapossiblepatternwithinthedataofgeneratedbatterycells.
TheIPFvaluesofthegeneratedLiBcellsthatwerefoundwithinthe intervalsdonotentirelycorrespondtotheIPFvaluesofthewithheld LiBcells.Thisisduetothefactthatthemodelisabletopredict sim-ilaroutputvaluesbasedondifferentinputvalues,i.e.whenoneinput variableislower,itcanbecompensatedbyanotheronebeinghigher. Therefore,therecanbedifferentarrangementsofIPFvaluesleadingto almostsimilarresults.Thiswasinvestigatedbyidentifyingthree clus-tersfor15ESCcellsandtwofor10ESCcellsusingaGaussianmixture modelontheIPFdata.Thebestamountofclusterswasidentifiedusing theBayesinformationcriterion.Theidentifiedclustersforcellswith15 ESCarec_0_15ESC,c_1_15ESCandc_2_15ESCandforcellswith10ESC arec_0_0ESCandc_1_10ESC.Intotalthereare26chosencellswith15 ESC(_0_15ESC:10cells,c_1_15ESC:12cellsandc_2_15ESC:4cells)and 28with10ESC(c_0_0ESC:14cellsandc_1_10ESC:13cells).
TheIPFsofeachclusterandofthesevenwithheldLiBcellsare de-pictedinFig.16 forvalidation purposes.Inordertoutilizethe pro-cedureofbatteryproductiondesign,thegeneratedclustersofpotential IPFsneedtobevalidated.Outgoingfromtheavailabledatabase,itneeds tobeevaluatedwhethertheIPFvaluesarecomprehensibleorwhether itisnotpossibletoachievethiscombinationofvalues.Incomparison torealcells(thewithholdcells),itshowsthatindeedsomeoftheIPF valuesweremetbythegeneratedLiBcellsintheclusters,butalsothat some haveslightlydifferentvaluesdue tothecompensationthrough otherIPFs.Forexample,the‘cathodeuncalenderedelectrodethickness [m∗10-6]’valuesofc_0_15ESCandc_1_15ESCdonotfitthevaluesof
realbatterycellsandthevaluesofall15ESCclustersdonotentirely fittherealbatterycellvaluesfor‘cathode/anodetheo.capacity[%]’. Further deviationsof c_0_15ESCfromrealbatterycellsarein‘anode
Fig.14. PredictedFPPs(max.capacity,self-discharge,andSOHafter400cycles)ofgeneratedIPFdata.
Table4
FPPrangesoftheexcludedbatterycells.
FPP cells with 15 cathode-anode compartments cells with 10 cathode-anode compartments mean standard deviation range mean standard deviation range
max. capacity [Ah] 9.115 0.011 9.080 – 9.148 5.83 0.040 5.705 – 5.946
self-discharge [%] 0.151 0.016 0.102 – 0.199 0.229 0.005 0.214 – 0.243
SOH after 400 cycles [%] 0.929 0.003 0.921 – 0.937 0.870 0.002 0.866 – 0.875
mixingPSDright[-]’and‘anodesuspsension viscosityfitting value’. Deviationsofc_1_15ESCfromrealbatterycellscanbeseenin‘cathode uncalenderedporosityvalue418-10[-]’andin‘cathodemeanchamfer angle[°]’.Theclusterc_2_15ESChasthehighestdeviationsin‘cathode meanpyrometerFWHM[s]’,‘cathodeuncalenderedporosityvalue 418-10[-]’and‘anodemixingstdPSDright[-]’.Theclusterswith10ESC showalsoonlytwohigherdeviationsfromreal10ESCbatterycellsin ‘cathodeuncalenderedporosityvalue418-10[-]’and‘anode
suspsen-sionviscosityfittingvalue’.Itfurtherneedstobeconsideredthatnotall oftheIPFvaluesthatareusedinthequalitypredictionmodelwere var-ied.Thismeansthattheirrangeanddistributionistheresultofprocess fluctuation.Inordertoruleoutthestrongsensitivitytothesevalues completely,thesevaluesneedtobevaried.Anothermeasurecouldbe thescalingofthesevaluesduringthedatapreprocessingbetweenthe realminimumandtherealmaximumtoindicatethattheirfluctuation isminimalcomparedtotherealvaluerange.
Fig.15. IPFsofthegeneratedFPPswithinthedefinedintervalsandtheIPFsofthesevenwithheldLiBcells.
4.3. Decisionsupport
Thoughthemodelandtheaccesstothedatabasearepurely pro-grammedinpython,themodelisdeployedonaweb-basedplatform. Themanual data cleaninghasbeen applieddirectly tothe accessed data,i.e.not allavailabledataisaccessed,onlythedesiredone. All further steps run within thepython script automatically,giving the modellingprocedure thepossibility tobe deployedin anautomated way.Theexecutionofthemodellingprocedurecanbedonemanually orcan be triggeredbytime-dependingevents,e.g. aftera fullbatch of batterycells is producedand alldata is availableor periodically at certaintime points.The visualization of modelresults is embed-ded in a web-basedapplication displaying theperformance andthe scoresofmodels.Theselectionoftheseeminglybestmodelcanbedone manually orautomated,e.g. whencertainscore criteriaarefulfilled orbyselectingthebestperformingmodeldepending onthescore of choice.Thebatteryproductiondesignisappliedonthesameweb-based platform.
Forafurtherexplanation,theclustersc_0_15ESC,c_1_15ESC,and c_2_15ESCandthewithheldrealbatterycellswith15 ESCareused. Fig.17depictstheseclustersandtherealbatterycellsforcomparison asradarcharts(realbatterycellswith15ESCinFig.17A,c_0_15ESC
inFig.17B,c_1_15ESCinFig.17C,andc_2_15ESCinFig.17D).This representationofthepotentialIPFvaluessupportstheevaluationof ob-tainedresultsbytheprocessexperts.Processexpertscanusethe avail-abledatabasisandtheinformationaboutthevariedIPFsinorderto maketheirdecisionwhetherthosevaluesarecomprehensibleornot.As anexample,valuesofthefeature‘cathodeamountactivematerial[g]’ (IPF15),werevariedandarecomprehensible.Moreactivematerialwill leadtoahighercapacityofthecell.ThisIPFisalmostthesameforthe 15ESCcells(generatedandreal).Ontheotherhand,thevaluesofthe feature‘cathodeuncalenderedelectrodethickness[m∗10-6]’(IPF10)
werenotvariedandthereforethesevaluesneedtoberegarded care-fully.JudgingbythegeometryoftheradarchartareasinFig.17, clus-terc_1_15ESCshowsthebiggestdifferencecomparedtotherealcells, c_0_15ESCandc_2_15ESC,withc_2_15ESCevenfewerdifferencesthan c_0_15ESC,thesmallestone.Theclusterc_2_15ESCshowsonlyafew IPFshavingsmallerdifferencescomparedtorealbatterycells,withthe biggestofthembeinginvariedvalues‘CathodemeanpyrometerFWHM’ (IPF2),‘cathodeuncalenderedporosityvalue418-10[-]’(IPF3),‘anode mixingstdPSDright[-]’(IPF6),‘electrolyteperthicknessper compart-ment[ml/mm/comp]’(IPF7),‘cathodeuncalenderedelectrode thick-ness[m∗10-6]’(IPF10)and‘cathode/anode theo.capacity[%]’ (IPF
Fig.16. IPFsofthefiveidentifiedclusterswiththecorrespondinggeneratedcellsandthewithheldLiBcells.
similarityoftheradarchartareashapes,thecomparisonbetweenthe valuescanbetterbevisualizedbyFig.16.
Aftertheverificationbytheprocessexperts,thechosenIPFsfrom thebatteryproductiondesigncannowbeusedinproductionprocess simulationinordertoderiveandachievetheneededprocess parame-ters[57].Basedontheproductionprocesssimulation,theuncertainties ofachievableIPFscanbederivedandbedirectlytransferredbyusing thequalitypredictionmodeltouncertaintiesofachievableFPPs[58]. Furthermore,apredictionofenergyandmaterialflowscanbemadeto planfurtherdetailsoffuturebatterycellproduction,allowingto uti-lizenotonlyproductqualitygoalcriteriabutalsoenvironmentaland economicgoalcriteria[59,60].
WiththestartoftheproductionusingthederivedPPs,thecontrolof theIPqualitywithregardtodesired/targetFPPsisachievedby utiliz-ingthemethodologyofdata-drivencyber-physicalsystemsforquality gates[41].Thismethodologyusesthequalitypredictionmodelandthe allowedtolerancesforIPFstopredictarangewithinwhichthevaluesfor targetFPPslie.Thismethodologyneedstobeupdatedwiththe multi-outputqualitypredictionmodelpresentedinthiswork.Witheverynew intermediateproduct,therangeoffutureFPPsnarrows.IftheIPFvalue
liesoutsidethetolerance,thequalitypredictionmodelupdatesits pre-dictionwithnewlysetIPFvaluesandthebatteryproductiondesignfinds suitablevaluesforremainingIPFsthatwerenotadjustedyet.Afterthe newIPFsareverifiedbytheprocessexpert,theproductionsimulation modelgeneratesnewPPsfortheremainingprocesses,leadingtheIP qualitybacktothedesireddirection.Ifnoneof thenewlygenerated IPFsarereasonable,thentheproductionmanagercandecidetocancel theproductionandsavethefurthercostsofproducingaproductwith undesiredquality.
5. Summaryandoutlook
This paper presented an approach for battery production design based onamachinelearningmodelforthedeterminationof IPFsin ordertoobtaindesiredFPPsoflithium-ionbatterycells.Thepurposeof theapproachistodetermineneededIPFs/intermediateproduct struc-turesfortheprocessstepsinordertoachieveacertainqualityofthefinal product,thebatterycell.Theconceptaimstoutilizetheapproachwithin theframeworkof a cyber-physicalproductionsystem forcontinuous dataacquisitionandmodellearningtoenhancethemodel’s
predictabil-Fig.17. Radarchartofidentifiedclustersincomparisontorealbatterycells.
ityandthereforetoimprovetheprecisionoftheapproach.Theconcept wasappliedtoacasestudywithinthepilotlineoftheBattery LabFac-toryBraunschweig.Intotal155lithium-ionbatterycellswereusedto buildanartificialneuralnetworkmodelforthepredictionoffinal prod-uctpropertiesfromintermediateproductfeatures.10800batterycell samplesweregeneratedinsilicoandusedtodeterminetheneeded in-termediateproductfeaturevaluesthatwerecomparedtoseven(4with 15ESCsand3with10ESCs)lithium-ionbatterycellsthatwere with-heldfromthemodelling.Itcouldbeshownthatthemodelpredicts bat-terycellsthatwerevalidatedwithrealcellsandshownewintermediate productfeaturesubsetsthatmayleadtobatterycellswithcomparable performance.
Thoughthemodelisbasedonacomparablylowamountof data, theapproachshowsautilizationofmachinelearningmethodsfor bat-terycellmanufacturingimprovementbysupportingproduction plan-ningandoperation.Themodelneedsfurthervalidationandtraining withmoreavailabledatainordertoshowsignificantresults. Neverthe-less,itshowsgreatpotentialfordeploymentinbatterycellproduction helpingtocontrolandtovaryproductionprocessestowardsmore cus-tomizablebatterycells.
Althoughtheapproachwasappliedtopouch bagbatterycells,it appliestoallotherbatterycellgeometries.Furthermore,theapproach appliestoacertainextendtosimilartasksinotherareasofproduction engineeringinwhichthequalityofthefinalproductcanbedetermined initsperformance,purity,etc.ratherthanbydistinctionwhetheritis OKandNOK(e.g.semiconductors).
Furtherresearchwillfocusonproducingmorebatterycellsand im-provingthemodel.Specifically,thenextstepswillfocusondetermining IPFsthatwouldleadtoanimprovementofoneormoreFPPsand vali-datingtheseresultsbyproducingthebatterycellswiththedetermined IPFs.Theinfluenceofthenon-variedinputofthemodelalsoneedsto befurtherinvestigatedandvalidated.Additionally,thevariablespace ofalreadyacquiredIPFswillbeextendedtobetterunderstandtheir in-fluencesandtobetterquantifythem.Morefinalproductpropertiesand
theusageofdifferentmaterialsneedtobeaddedtothemodel,tohave highercustomizability.Concerningsustainability,suchtargetvaluesas energydemandandcostscouldbeusedtoextendthemodel,butlead toanevenhigherdatademand.
Authorcontributions
ArtemTuretskyy:conceptualization,methodology,formalanalysis, investigation,resources,datacuration,writing-originaldraft, writing-reviewandediting,visualization,fundingacquisition;
Jacob Wessel: conceptualization, methodology, writing-original draft,writing-reviewandediting,visualization;
ChristophHerrmann:writing-reviewandediting,funding acquisi-tion,supervision,projectadministration;
SebastianThiede:conceptualization,methodology,writing-review andediting,fundingacquisition,supervision;
DeclarationofCompetingInterest
Theauthorsdeclarethattheyhavenoknowncompetingfinancial interestsorpersonalrelationshipsthatcouldhaveappearedtoinfluence theworkreportedinthispaper.
Acknowledgments
TheauthorsthanktheBMWI-FederalMinistryofEconomicAffairs andEnergyforsupportingtheprojectDALION4.0-DataMiningas Ba-sisforcyber-physicalSystemsinProductionofLithium-ionBatteryCells (03ETE017A).Theauthorsextend their gratitudetotheir colleagues from theDALION4.0teamforrunningtheexperimentsand produc-ingbatterycells.Further,theauthorswouldliketothankOlafWojahn forsupportingtheminquestionsandchallengesregardingautomation andIT.
