Organization
HonoraryChair
TaoWuShanghaiUniversityofMedicineandHealth Sciences,China
GeneralChairs
HamidoFujitaIwatePrefecturalUniversity,Japan
MoonisAliTexasStateUniversity,USA
OrganizingChair
JunSasakii-SOMETInc.,Japan
ProgramCommitteeChairs
PhilippeFournier-VigerShenzhenUniversity,China
YinglinWangShanghaiUniversityofFinanceandEconomics, China
SpecialSessionChairs
AliSelamatUniversitiTeknologiMalaysia,Malaysia XingWuShanghaiUniversity,China
JerryChun-WeiLinWesternNorwayUniversityofAppliedSciences, Norway
NgocThanhNguyenWroclawUniversityofTechnology,Poland
ProgramCommittee
MoulayA.AkhloufiUniversitédeMoncton,Canada
AzriAzmiUniversitiTeknologiMalaysia,Malaysia
HafewaBargaouiInstitutSupérieurdeGestiondeTunis,Tunisia OlfaBelkahlaDrissEcoleSupérieuredeCommercedeTunis,Tunisia LadjelBellatrecheLIAS/ISAE-ENSMA,France
ZalanBodoBabes-BolyaiUniversity,Romania ZakiBrahmiISITCOM,Tunisia FranciscoJ.CabrerizoUniversityofGranada,Spain
AlbertoCanoVirginiaCommonwealthUniversity,USA AndrewTzer-YeuChenUniversityofAuckland,NewZealand Chun-HaoChenNationalTaipeiUniversityofTechnology,Taiwan
Shyi-MingChenNationalTaiwanUniversityofScienceand Technology,Taiwan
TaiDinhJAIST,Japan
YoucefDjenouriSouthernDenmarkUniversity,Denmark AlexanderFerreinAachenUniversityofAppliedScience,Germany PhilippeFournier-VigerShenzhenUniversity,China HamidoFujitaIwatePrefecturalUniversity,Japan
AbdennaceurGhandriHigherInstituteofManagementofGabes,Tunisia SergeiGorlatchMuensterUniversity,Germany DeepakGuptaNITArunachalPradesh,India
Tzung-PeiHongNationalUniversityofKaohsiung,Taiwan
Ko-WeiHuangNationalKaohsiungUniversityofScienceand Technology,Taiwan
MiroslavHudecUniversityofEconomicsinBratislava,Slovakia DosamHwangYeungnamUniversity,SouthKorea MarcinJodłowiecWroclawUniversityofScienceandTechnology, Poland
FadouaKhennouUniversitédeMoncton,Canada YunSingKohUniversityofAuckland,NewZealand AdriannaKozierkiewiczWroclawUniversityofScienceandTechnology, Poland
MarekKrótkiewiczWroclawUniversityofScienceandTechnology, Poland
MasakiKurematsuIwatePrefecturalUniversity,Japan ThomasLacombeUniversityofAuckland,NewZealand ShihHsiungLeeNationalKaohsiungUniversityofApplied Sciences,Taiwan ArkadiuszLiberWroclawUniversityofScienceandTechnology, Poland
JerryChun-WeiLinWesternNorwayUniversityofAppliedSciences, Norway Wen-YangLinNationalUniversityofKaohsiung,Taiwan Yu-ChenLinFengChiaUniversity,Taiwan FrederickMaierUniversityofGeorgia,USA WolfgangMayerUniversityofSouthAustralia,Australia MasurahMohamadUniversitiTeknologiMalaysia,Malaysia M.RashedurRahmanNorthSouthUniversity,Bangladesh YasserMohammedAssiutUniversity,Egypt TauheedKhanMohdAugustanaCollege,USA AnirbanMondalAshokaUniversity,India
M.SaqibNawazHarbinInstituteofTechnology,China DuNguyenNongLamUniversity,Vietnam
DucNguyenVietnamMaritimeUniversity,Vietnam HienNguyenUniversityofInformationTechnology,Vietnam Ngoc-ThanhNguyenWroclawUniversityofTechnology,Poland Tat-Bao-ThienNguyenThuyloiUniversity,Vietnam ThanhBinhNguyenHoChiMinhCityUniversityofTechnology, Vietnam
MouradNouiouaHarbinInstituteofTechnology,China HoussemEddineNouriInstitutSupérieurdeGestiondeGabes,Tunisia AmmarOdehPrincessSumayaUniversityforTechnology, Jordan
SamirOuchaniLINEACT,CESI,France
P.KrishnaReddyIIITHyderabad,India
HauPhamQuangBinhUniversity,Vietnam
MarcinPietranikWroclawUniversityofScienceandTechnology, Poland
UdayRageUniversityofTokyo,Japan
ShafinRahmanNorthSouthUniversity,Bangladesh PenugondaRavikumarUniversityofAizu,Japan AndreasSpeckKielUniversity,Germany
GautamSrivastavaBrandonUniversity,Canada FeiyangTangNorwegianUniversityofScienceofTechnology, Norway
StefaniaTomasielloUniversityofTartu,Estonia HaiTranHoChiMinhUniversityofPedagogy,Vietnam JianjiaWangShanghaiUniversity,China ZhijinWangJimeiUniversity,China
YutakaWatanobeUniversityofAizu,Japan Cheng-WeiWuNationalIlanUniversity,Taiwan TakeruYokoiTokyoMetropolitanCollegeofIndustrial Technology,Japan
NurulhudaZainuddinUniversitiTeknologiMalaysia,Malaysia WeiZhangAdobe,USA HibaZuhairAl-NahrainUniversity,Iraq
Contents
IndustrialApplications
ComparativeStudyofMethodsfortheReal-TimeDetectionofDynamic BottlenecksinSerialProductionLines....................................3 NikolaiWest,JörnSchwenken,andJochenDeuse
Ultra-short-TermLoadForecastingModelBasedonVMDandTGCN-GRU...15 MeirongDing,HangZhang,BiqingZeng,GaoyanCai,YuanChai, andWenshengGan
LearningtoMatchProductCodes........................................29 YingExcellandSebastianLink
ResUnet:AFullyConvolutionalNetworkforSpeechEnhancement inIndustrialRobots....................................................42 YangyiPuandHongyangYu
SurfaceDefectDetectionandClassificationBasedonFusingMultiple ComputerVisionTechniques............................................51 MinZhu,BingqingShen,YanSun,ChongyuWang,GuoxinHou, ZhijieYan,andHongmingCai
DevelopmentofaMultiagentBasedOrderPickingSimulator forOptimizingOperationsinaLogisticsWarehouse........................63 TakutoSakuma,MinamiWatanabe,KoyaIhara,andShoheiKato
HealthInformatics
PredictingInfectionAreaofDengueFeverforNextWeekThrough MultipleFactors.......................................................77 Cong-HanZheng,Ping-YuHsu,Ming-ShienCheng,NiXu, andYu-ChunChen
HospitalReadmissionPredictionviaPersonalizedFeatureLearning andEmbedding:ANovelDeepLearningFramework.......................89 YuxiLiuandShaowenQin
IntelligentMedicalInteractiveEducationalSystemforCardiovascular Disease..............................................................101 Sheng-ShanChen,Hou-TsanLee,Tun-WenPai,andChao-HungWang
EvolutionaryOptimizationforCNNCompressionUsingThoracicX-Ray ImageClassification...................................................112
HassenLouati,SlimBechikh,AliLouati,AbdulazizAldaej, andLamjedBenSaid
AnOrientedAttentionModelforInfectiousDiseaseCasesPrediction.........124 PeisongZhang,ZhijinWang,GuoqingChao,YaohuiHuang, andJingwenYan
TheDifferentialGeneDetectingMethodforIdentifyingLeukemiaPatients....137 MingzhaoWang,WeiliangJiang,andJuanyingXie
EpidemicModelingoftheSpatiotemporalSpreadofCOVID-19 overanIntercityPopulationMobilityNetwork.............................147 YuxiLiu,ShaowenQin,andZhenhaoZhang
SkinCancerClassificationUsingDifferentBackbonesofConvolutional NeuralNetworks......................................................160
AnhT.Huynh,Van-DungHoang,SangVu,TrongT.Le, andHienD.Nguyen
CardiovascularDiseaseDetectiononX-RayImageswithTransferLearning...173 NguyenVan-BinhandNguyenThai-Nghe
CausalReasoningMethodsinMedicalDomain:AReview..................184 XingWu,JingwenLi,QuanQian,YueLiu,andYikeGuo
Optimization
EnhancingaMulti-populationOptimisationApproachwithaDynamic TransformationScheme................................................199 ShengqiDai,VincentW.L.Tam,ZhenglongLi,andL.K.Yeung
AModelDrivenApproachtoTransformBusinessVision-Oriented Decision-MakingRequirementintoSolution-OrientedOptimizationModel....211 LiwenZhang,HervéPingaud,ElyesLamine,FranckFontanili, ChristopheBortolaso,andMustaphaDerras
AHybridApproachBasedonGeneticAlgorithmwithRanking AggregationforFeatureSelection........................................226 BuiQuocTrung,LeMinhDuc,andBuiThiMaiAnh
ANovelType-BasedGeneticAlgorithmforExtractiveSummarization........240 BuiThiMaiAnh,NguyenThiThuTrang,andTranThiDinh
DragonflyAlgorithmforMulti-targetSearchProbleminSwarmRobotic withDynamicEnvironmentSize.........................................253
MohdGhazaliMohdHamamiandZoolH.Ismail
VideoandImageProcessing
ImprovedProcessingofUltrasoundTongueVideosbyCombining ConvLSTMand3DConvolutionalNetworks..............................265 AminHonarmandiShandizandLászlóTóth
ImprovementofTextImageSuper-ResolutionBenefitingMulti-task Learning.............................................................275 KosukeHonda,HamidoFujita,andMasakiKurematsu
QuestionDifficultyEstimationwithDirectionalModalityAssociation inVideoQuestionAnswering...........................................287 Bong-MinKimandSeong-BaePark
NaturalLanguageProcessing
ImprovingNeuralMachineTranslationbyEfficientlyIncorporating SyntacticTemplates....................................................303
PhuongNguyen,TungLe,Thanh-LeHa,ThaiDang,KhanhTran, KimAnhNguyen,andNguyenLeMinh
ForensicAnalysisofTextandMessagesinSmartphonesbyaUnification RosettaStoneProcedure................................................315 ClaudioTomazzoli,SimoneScannapieco,andMatteoCristani
Relation-LevelVectorRepresentationforRelationExtraction andClassificationonSpecializedData....................................327 CamilleGosset,MokhtarBoumedyenBillami,MathieuLafourcade, ChristopheBortolaso,andMustaphaDerras
SAKE:AGraph-BasedKeyphraseExtractionMethodUsingSelf-attention....339 PingZhu,ChuanyangGong,andZhihuaWei
SynonymPredictionforVietnameseOccupationalSkills....................351 Hai-NamCao,Duc-ThaiDo,Viet-TrungTran,Tuan-DungCao, andYoung-InSong
ASurveyofPretrainedEmbeddingsforJapaneseLegalRepresentation.......363 Ha-ThanhNguyen,Le-MinhNguyen,andKenSatoh
MachineReadingComprehensionModelforLow-ResourceLanguages andExperimentingonVietnamese.......................................370
BachHoangTienNguyen,DungManhNguyen, andTrangThiThuNguyen
InducingaMalayLexiconfromanUnlabelledDatasetUsingWord Embeddings..........................................................382
IanH.J.Ho,Hui-NgoGoh,andYi-FeiTan
AgentandGroup-BasedSystems
Agent-BasedIntermodalBehaviorforUrbanToll..........................397 AziseOumarDiallo,GuillaumeLozenguez,ArnaudDoniec, andRenéMandiau
EntropyBasedApproachtoMeasuringConsensusinGroup Decision-MakingProblems.............................................409
J.M.Tapia,F.Chiclana,M.J.delMoral,andE.Herrera–Viedma
AdaptationofHMIsAccordingtoUsers’FeelingsBasedonMulti-agent Systems..............................................................416
AliaMaaloul,HoussemEddineNouri,ZiedTrifa,andOlfaBelkahlaDriss
PatternRecognition
AGeneralizedInvertedDirichletPredictiveModelforActivity RecognitionUsingSmallTrainingData...................................431
JiaxunGuo,ManarAmayri,WentaoFan,andNizarBouguila
DeepfakeDetectionUsingCNNTrainedonEyeRegion....................443 DavidJohnson,TonyGwyn,LetuQingge,andKaushikRoy
FaceAuthenticationfromMaskedFaceImagesUsingDeepLearning onPeriocularBiometrics...............................................452
JeffreyJ.HernandezV.,RodneyDejournett,UdayasriNannuri, TonyGwyn,XiaohongYuan,andKaushikRoy
AnOptimizationAlgorithmforExtractiveMulti-documentSummarization BasedonAssociationofSentences.......................................460
Chun-HaoChen,Yi-ChenYang,andJerryChun-WeiLin
ASpatiotemporalImageFusionMethodforPredictingHigh-Resolution SatelliteImages.......................................................470
VipulChhabra,R.UdayKiran,JuanXiao,P.KrishnaReddy, andRamAvtar
Security
WHTE:WeightedHoeffdingTreeEnsembleforNetworkAttackDetection atFog-IoMT..........................................................485 ShilanS.Hameed,AliSelamat,LizaAbdulLatiff,ShukorA.Razak, andOndrejKrejcar
AnImprovedEnsembleDeepLearningModelBasedonCNN forMaliciousWebsiteDetection.........................................497 NguyetQuangDo,AliSelamat,KokChengLim,andOndrejKrejcar
Intrusion-BasedAttackDetectionUsingMachineLearningTechniques forConnectedAutonomousVehicle......................................505 MansiBhavsar,KaushikRoy,ZhipengLiu,JohnKelly, andBalakrishnaGokaraju
DetectionofAnti-forensicsandMalwareApplicationsinVolatileMemory Acquisition...........................................................516 ChandlorRatcliffe,BiodoumoyeGeorgeBokolo,DamilolaOladimeji, andBingZhou
MalwareClassificationBasedonGraphConvolutionalNeuralNetworks andStaticCallGraphFeatures..........................................528 AttilaMesterandZalánBodó
ModellingandDiagnosis
The Java2CSP DebuggingToolUtilizingConstraintSolving andModel-BasedDiagnosisPrinciples...................................543 FranzWotawaandVladAndreiDumitru
FormalModellingandSecurityAnalysisofInter-OperableSystems..........555 AbdelhakimBaouya,SamirOuchani,andSaddekBensalem
SocialNetworkAnalysis
Content-Context-BasedGraphConvolutionalNetworkforFakeNews Detection.............................................................571 HuyenTrangPhan,NgocThanhNguyen,andDosamHwang
Multi-classSentimentClassificationforCustomers’Reviews................583 CuongT.V.Nguyen,AnhM.Tran,ThaoNguyen,TrungT.Nguyen, andBinhT.Nguyen
xviContents
TransportationandUrbanApplications
MM-AQI:ANovelFrameworktoUnderstandtheAssociationsBetween UrbanTraffic,VisualPollution,andAirPollution..........................597 KazukiTejima,Minh-SonDao,andKojiZettsu
Two-StageTrafficClusteringBasedonHNSW............................609 XuZhang,XinzhengNiu,PhilippeFournier-Viger,andBingWang
ExplainableOnlineLaneChangePredictionsonaDigitalTwin withaLayerNormalizedLSTMandLayer-wiseRelevancePropagation.......621 ChristophWehner,FrancisPowlesland,BasharAltakrouri, andUteSchmid
AnAgendaontheEmploymentofAITechnologiesinPortAreas: TheTEBETSProject..................................................633 AdorniEmanuele,RozhokAnastasiia,RevetriaRoberto, andSuchevSergey
ModellingandSolvingtheGreenShare-a-RideProblem....................648 ElhemElkoutandOlfaBelkahlaDriss
MachineLearningTechniquestoPredictRealTimeThermalComfort, Preference,Acceptability,andSensationforAutomationofHVAC Temperature..........................................................659 YaaT.Acquaah,BalakrishnaGokaraju,RaymondC.TesieroIII, andKaushikRoy
NeuralNetworks
SeriallyDisentangledLearningforMulti-LayeredNeuralNetworks..........669 RyotaroKamimuraandRyozoKitajima
DetectingUseCaseScenariosinRequirementsArtifacts:ADeepLearning Approach............................................................682 MunimaJahan,ZahraShakeriHosseinAbad,andBehrouzFar
HybridDeepNeuralNetworksforIndustrialTextScoring...................695 SidharrthNagappan,Hui-NgoGoh,andAmyHui-LanLim
BenchmarkingTrainingMethodologiesforDenseNeuralNetworks..........707 IsaacTonkin,GeoffHarris,andVolodymyrNovykov
ProposingNovelHigh-PerformanceCompoundsbyNestedVAEsTrained IndependentlyonDifferentDatasets......................................714
YoshihiroOsakabeandAkinoriAsahara
Clustering
MonotonicConstrainedClustering:AFirstApproach.......................725 GermánGonzález-Almagro,PabloSánchezBermejo,JuanLuisSuarez, José-RamónCano,andSalvadorGarcía
ExtractiveTextSummarizationonLarge-scaleDatasetUsingK-Means Clustering............................................................737 Ti-HonNguyenandThanh-NghiDo
Multi-GranularLargeScaleGroupDecision-MakingMethodwithaNew ConsensusMeasureBasedonClusteringofAlternativesinModifiable Scenarios.............................................................747
JoséRamónTrillo,IgnacioJavierPérez,EnriqueHerrera-Viedma, JuanAntonioMorente-Molinera,andFranciscoJavierCabrerizo
OptimalUserCategorizationfromaHierarchicalClusteringTree forRecommendation...................................................759
WeiSongandSiqiLiu
Classification
APreliminaryApproachforusingMetricLearninginMonotonic Classification.........................................................773
JuanLuisSuárez,GermánGonzález-Almagro,SalvadorGarcía, andFranciscoHerrera
DeepLearningArchitecturesExtendedfromTransferLearning forClassificationofRiceLeafDiseases...................................785 HaiThanhNguyen,QuyenThucQuach,ChiLeHoangTran, andHuongHoangLuong
HeightEstimationforAbrasiveGrainofSyntheticDiamonds onMicroscopeImagesbyConditionalAdversarialNetworks................797 JoeBrinton,ShotaOki,XinYang,andMaikoShigeno
PatternMiningandTsetlinMachines
FastWeightedSequentialPatternMining.................................807 ZhenqiangYe,ZiyangLi,WeibinGuo,WenshengGan,ShichengWan, andJiahuiChen
ParallelHighUtilityItemsetMining.....................................819 GaojuanFan,HuaiyuanXiao,ChongshengZhang,GeorgeAlmpanidis, PhilippeFournier-Viger,andHamidoFujita
TowardsEfficientDiscoveryofStablePeriodicPatternsinBigColumnar TemporalDatabases...................................................831
HongN.Dao,PenugondaRavikumar,P.Likitha, BathalaVenusVikranthRaj,R.UdayKiran,YutakaWatanobe, andIncheonPaik
CyclostationaryRandomNumberSequencesfortheTsetlinMachine.........844 SveinAndersTunheim,RohanKumarYadav,LeiJiao,RishadShafik, andOle-ChristofferGranmo
LogicsandOntologies
EvolutionofPrioritized EL Ontologies...................................859 RimMohamed,ZiedLoukil,FaiezGargouri,andZiedBouraoui
AComparisonofResourceDataFrameworkandInductiveLogic ProgramingforOntologyDevelopment...................................871 DurgeshNandini
MDNCaching:AStrategytoGenerateQualityNegativesforKnowledge GraphEmbedding.....................................................877 TiroshanMadushankaandRyutaroIchise
Robotics,GamesandConsumerApplications
ApplicationofaLimitTheoremtotheConstructionofJapaneseCrossword Puzzles..............................................................891 VolodymyrNovykov,GeoffHarris,andIsaacTonkin
NonImmersiveVirtualLaboratoryAppliedtoRoboticsArms................898 DanielaA.Bastidas,LuisF.Recalde,PatriciaN.Constante, VictorH.Andaluz,DayanaE.Gallegos,andJoséVarela-Aldás
AnImprovedSubject-IndependentStressDetectionModelApplied toConsumer-gradeWearableDevices....................................907 Van-TuNinh,Manh-DuyNguyen,SinéadSmyth,Minh-TrietTran, GrahamHealy,BinhT.Nguyen,andCathalGurrin
WDTourism:APersonalizedTourismRecommendationSystemBased onSemanticWeb......................................................920
KaiyuDai,PengfeiJi,XiaoruiZuo,andDaixinDai
AuthorIndex .........................................................935
ComparativeStudyofMethods fortheReal-TimeDetectionofDynamic BottlenecksinSerialProductionLines
NikolaiWest1(B) ,JörnSchwenken1 ,andJochenDeuse1,2
1 InstituteofProductionSystems,TechnicalUniversityDortmund,Leonhard-Euler-Str.5, 44227Dortmund,Germany nikolai.west@tu-dortmund.de
2 CentreforAdvancedManufacturing,UniversityofTechnologySydney,11Broadway,Ultimo, NSW2007,Australia
Abstract. Capacity-limitingbottlenecksinmanufacturingsystemsformtheideal startingpointformeasuresofimprovement.However,theinherentvariability ofmodernsystemsleadstodynamicbottleneckbehavior,causingthemtoshift betweenstations.Numerousmethodsforthedetectionofshiftingbottlenecksexist inliterature.Inthispaper,wepresentandcomparethreemethods:BottleneckWalk (BNW),ActivePeriodMethod(APM),andanadaptationofInterdepartureTime Variances(ITV).Thecomparativestudydeploysthemethodsinaserialproduction linewithsevenstationsandeightbuffers.Wevarytheindividuallocationsofthe bottlenecksbyaddingmoreprocesstime.Tocomparethemethods,wedetermine theoverallaverageratioofagreementbetweenthethreedetectionmethods.APM andITVhavethehighestagreementatanaverageof80.10%.PairingswithBNW achievesignificantlylowerratesofagreement,with56.33%forITV,and62.03%% whencomparedtotheAPM.
Keywords: Bottleneckanalysis · Bottleneckdetection · Materialflowsimulation
1Introduction
Accordingtothe TheoryofConstraints (TOC),theperformanceofamaterialflowsystem oramanufacturingsystemisinevitablylimitedbyjustonestation,whichtheTOCcallsa bottleneck [1].Liketheweakestlinkinachain,suchabottlenecklimitstheoutputofthe entiresystem.Focusingallimprovementeffortsontheselinksofthechainconstitutes themostefficientapproachtoincreasingtheoverallperformanceofthesystem[2]. Sinceoptimizationmeasuresofnon-bottleneckstationshavenoimpactonthesystemic output,onlyimprovementsatthebottleneckleadtoaquantifiableenhancement.Despite ‘bottleneck’beingacommonlyusedtermingenerallinguistics,thereisnoaccepted definitioninscientificliterature[3].Thecomparativestudyinthispapershowsthat differentmetricsforbottleneckidentificationalsoledtovaryingbottleneckdefinitions adoptedbytherespectiveauthors.Therefore,wefollowametric-independentdefinition andrefertoabottleneckinageneralfashionas“theresourcethatrestrictsasystemic outputuptoaspecificlimitation”[4, 5].
©SpringerNatureSwitzerlandAG2022
H.Fujitaetal.(Eds.):IEA/AIE2022,LNAI13343,pp.3–14,2022. https://doi.org/10.1007/978-3-031-08530-7 1
1.1OntheDynamicNatureofBottlenecks
Asmentionedbefore,abottleneck’slocationcanchangeovertimeduetovariability. Variabilityreferstoafluctuationofproductandprocessvariablesandaffectsallrealworldmanufacturingsystems.Thus,literatureclassifiesabottleneckaccordingtoits behaviorinbeing static or dynamic [5–7].Staticbottleneckshaveafixedpositioninthe valuestream.Theyonlyaffectonestationduringtheentireobservationperiod.Their occurrencecanoftenbetracedbacktodesignflawswhichcauseasingularrestrictionof thematerialflow.Incontrast,theplaceofoccurrenceofdynamicbottlenecksisvariable. Suchshiftsoccureitherduetorandomeventsorduetograduallychangingconditions inthemanufacturingsystem[8].
Despitethisdistinction,inpracticeoftenonestationisreferredtoasmainbottleneck. However,staticassumptionsonlyapplytoverysimplesystemsandarerarelyobserved inpractice[7].Inthispaper,wethereforeexaminemethodsforidentifyingdynamic bottlenecks.Still,themethodsarealsosuitedtodetectstaticbottlenecks.
1.2TheNeedforReal-TimeBottleneckDetection
Insteadofpointwisemeasurementswithmanualefforts,dynamicbehaviorrequires continuousrecordingandevaluationtoallowtargetedmeasuresofoptimization[6].The primarygoalistodetectthebottleneckthatiscurrentlyaffectingthesystem.Similarto thewaybottleneckscanbedifferentiatedaccordingtotheirbehavior,detectionmethods canbedividedintothetwogroupsof AverageValueMethods (AVM)and Momentary ValueMethods (MVM)[4, 9].AVMusedefinedperiodsasbasisfortheanalysis.They determinebottlenecksusingtheaveragevaluesofdifferentproductionmetrics.AVMfor bottleneckdetectionarebasedonavarietyofmetrics,suchasthe OverallThroughput Effectiveness [10],the degreeofUtilization [11],or InterdepartureTimesVariances [12].Thesemethodsarelesssuitabletohandledynamicsystemsandtendtoachieve relativelylowdetectionconfidence.Incontrast,MVMuseasingularobservation.They enableanidentificationofshiftingbottlenecksindynamicsystems[13].Examplesfor MVMmethodsarethe ArrowMethod [14, 15],the ActivePeriodMethod [7],orthe TurningPointMethod [16, 17].Thisselectiondoesnotclaimtobecomprehensive,asa multitudeofmethodsandmethodvariantsexists.Wereferto[4, 12, 18]foranextensive reviewofsuchmethods.
Insummary, BottleneckDetection,asthefirstphaseofaholisticbottleneckanalysis [4],aimstodetectshiftingbottlenecks.Therefore,itisessentialtofirstdetectthemomentarybottleneckbeforecalculatingaveragesoftheoverallsystemicimpact.Anymethod thatusesaveragestodetectthebottlenecksislikelytointroduceerrorsinthedetection ofshiftingbottlenecks[13].ForthestudyinSec. 4,weusetwoMVMapproachesand adaptanAVMapproachformomentarybottleneckdetection. Theremainderofthispaperisdividedintofourparts.First,webrieflydescribea selectionofthreemethodsforreal-timebottleneckdetectionthatusedifferentmetrics. Next,weintroducethedesignofthecasestudy,followedbyanoverviewoftheresults foraserialproductionline.Finally,wediscussandcomparetheresultsofthestudy,and lastlyprovideabriefoutlookwithrecommendationsforfutureresearch.
2RelatedWorkonBottleneckDetection
AccordingtotheTOC,anymeasureofimprovementfirstrequiresadeterminationof thepositionofthebottleneck.Asmentionedbefore,thereisalargenumberofpotential methodsfordetectingbottlenecks.Wedonotaimforacomparisonofallavailable methods,butagainrefertotherespectiveliteratureinstead[4, 12, 13, 18, 19].
Throughourselectionofthefollowingthreedetectionmethods,wepromotethe utilizationofdifferentkeyfiguresfromthemanufacturingsystem.TheBottleneckWalk isfocussesprimarilyonexaminingthelevelsoftheproductionbuffers.TheActive Periodmethodusesmachinestatestoidentifybottlenecks.Lastly,InterdepartureTime Variancedetectionevaluatesoftheprocesstimesoftheworkstations[4].
2.1DetectionUsingBottleneckWalkwithBufferLevels
The BottleneckWalk (BNW)isamethodfortheidentificationofdynamicbottlenecksin seriallineswithfinitebuffers[20].BNWiscloselyrelatedtothe ArrowMethod [14].It servesasahands-onmethodforbottleneckdetectionthatrequirespractitionerstotakea tourthroughthemanufacturingline.Whilewalkingalongtheline,saidobserverwrites downinventorylevelsandprocessstatesofthelineintoadefineddatasheet.These statesarerecordedaccordingtotheobservations,whiledistinguishingthethreestates processing, breakdown and waiting.Forwaiting,BNWrequiresasecondsubdivision: eitherastationcanbeblockedduetoafullsubsequentbuffer,oritcanbestarvingdue toanemptyprecedingbuffer.Inaddition,BNWconsidersthelevelofthebuffers,where aclearlydefinedbuffercanbefilledbetween0%and100%withregardtoitsmaximum capacity.Ifthecurrentlyobservedbufferlevelislowerthanonethirdofthemaximum capacity,thebottleneckislocatedupstream.Ifthelevelishigherthantwothirdsofthe capacity,theBNWpinpointsthestationdownstream.BNWusesanarrow-basedsystem todetermineabottleneckstation.Astationwitharrowspointingtoitfrombothsides representsthebottleneck[20].
Inthisusagemode,BNWrequiresamanualcheckofthesystemstatetobeperformed severaltimesaday.Foradata-drivenapplicationinflexiblemanufacturingsystems,the methodologywasthereforeadaptedforreal-timemonitoring[5].Focusingonvirtual bufferlevelsandneglectingthestationstates,allowsmakingequivalentstatementsabout thebottleneckstation.Givenasufficientavailabilityofdata,thisadoptionoftheBNW allowsdeterminingabottleneckatanytime.
2.2DetectionUsingActivePeriodMethodwithMachineStates
Accordingtothe ActivePeriodMethod (APM),abottleneckisthestationwithinthe materialflowsystemthathasbeenworkingthelongestwithoutanyinterruptions.APM reliesonasimilarunderstandingofmachinestatesasBNW.Workstationsareconsidered active whentheyareprocessingproductsasdefinedbytheirproductionprogram,orwhen theyareotherwisebusydueto repair , setup or maintenance operations.Incontrast,a stationiscalled inactive ifitiswaitingduetobuffer-relatedstarvationorblockage[6, 7].APMconsidersastation blocked ifthedownstreambufferisfilledtothemaximum. Thenitisunabletotransferanotherparttothefollowingstations.Fortheoppositecase,
iftheupstreambufferisemptyanditcannotsupplyanotherparttothenextstation, thestationisconsidered starved [14].Whilewesimplyrefertostationstatesinthis paper,APMallowsagroupingofstatesfordifferententitiesofaproductionsystem.For example,foraprocessingmachine,thestates‘working’,‘inrepair’,‘changingtools’ or‘beingserviced’areallconsideredactivestates.Forafactoryworker,‘working’or ‘beingonscheduledbreak’areactivestates,while‘plannedorunplannedwaiting’are consideredinactivestates.Regardlessoftheobservedentity,thelongestactiveoperating periodthenmarksthebottleneck[7].
APMalsoincludestheshiftingstateofabottleneckbydeterminingwhetherastationisthesolebottleneckorashiftingbottleneck.Shiftingstatesoccurattheoverlapof thecurrentandthesubsequentbottleneckperiods.Anaccuratedistinctionbetweensole bottleneckandshiftingbottleneckrequiresaretrospectiveviewofthestates.Several activestationsatanobservationtimesimplyleadtoanidentificationofshiftingbottlenecks.Despitethisdisputablelimitation,APMallowsnearreal-timedeterminationof bottleneckstationsbasedoncurrentconditionsofsaidstations.
2.3DetectionUsingInterdepartureTimeVariancewithProcessTimes
Thethirdmethodutilizes InterdepartureTimeVariances (ITV)todetectdynamicbottlenecks.Asamethod,ITVisbasedontheassumptionthatthebottleneckofamanufacturingsystemhasthelowestvarianceintheinterdeparturetimesofallproducts, i.e.thelowestITV[12].Themethodreliesonachainofconsiderations:Ifamachine inthemanufacturingsystemrequiresalonger effectiveprocessingtime tocomplete products,thismachinetendstobemoreutilized.Higherutilizationisassociatedwith ahigherutilizationrate,whichleadstoalongerqueueinfrontofthemachine,i.e.a higherbufferlevel.Analogously,thiscanleadtoalesshighlyfilledbufferbehindthe consideredmachine.Thisbufferbehaviorleadstolessfrequent starvation or blocking ofthemachine,andtheproportionofidlestatesiscorrespondinglylower.Similarly, theproportionofactivemachinestatesissignificantlyhigher,whichinturnleadsto othermachinesstarvingdownstreamandblockingupstream.Duetotheserecurring interruptions,thevarianceofthetimesbetweenthecompletionofproducts,i.e.theITV, increasesfortheothermachines.Abottleneckcanthenbeunambiguouslydetermined asthemachinethatachievesthelowestITVoftheentiresystem[12].Inadditiontothis logical-argumentativereasoning,theassumptioncanalsobedescribedinmathematical termsusingthe linkingequation,inaccordancewith[11].Thecalculationsarefurther elaboratedby[12]withreferenceto[11].
ForITVcalculation,theprocesstimemustbedeterminedasthedifferencetothetime stampofthenextproduct,startingfromastation-specifictimestampperstation.The varianceofthesedistancesisthencalculatedforanaggregationinterval.Whenselecting theintervallength,itisimportanttoselectasufficientlylargeandrepresentativesetof observationtimes.Atthesametime,alossofinformationduetoexcessiveaggregation mustbeprevented.Toenablereal-timedetectionlikeaMVM,thevariancecanalsobe determinedatanytime,usingaslidingwindowwithadefinedlength.Thisallowsa bottleneckdetectionthatutilizesthecurrentsystemstate.
3DesignoftheComparativeStudyforBottleneckDetection
Tocomparetheselectedbottleneckdetectionmethods,webuildaserialproductionline withsevenworkstations Mi andeightbuffers Bi asshowninFig. 1.Allstationsare connectedthroughabuffer,withthefirstbuffer B0 andthelastbuffer B7 formingthe systemboundaries.Fortheboundaries,weassumeinfinitesupplyanddemand.Weset amaximumcapacityofallbuffers BCBi to5units.Eachstationhasadefinedprocess time ptMi ,althoughweintroducevariabilityintothesystembymeansofanappropriate distributionfunction.Thedistributionfunctionusedtodeterminetheindividualprocess timesisshowninFig. 2.Theright-skeweddistributioncorrespondstoamanufacturing systemwithoccasionalequipmentdowntimes.
Fig.1. Layoutoftheserialproductionlinewithexemplary20%-bottlenecksatM3andM6
Sinceweneedtoapplythedetectionmethodsinsystemswithdynamicbottlenecks, weintentionallyinducebottleneckstatesbyapplyingapercentagefactortotheprocess timesoftwoselectedstations.Figure 1,forexample,showsasystemconfigurationin whichstations M3 and M6 have20%higherprocesstimes.Weshowthecorresponding distributionofprocesstimesinredinFig. 2.Forunchangedstationswithaninitial processtimeof10,theaverageprocesstimeaftervariabilityadjustmentis 17.94 units. Similarly,formodifiedstationswithaninitialprocesstimeof12,theaveragetime aftervariabilitypenaltyis 21.53.Duetounlimitedboundaries, M1 doesnotstarvewhile M7 isneverblocked.Tocomparedifferentbottlenecksituations,weuseallbottleneck combinations,for M2 to M6 ,includingsingularbottleneckstateswithjustoneaffected station.Thisresultsin15uniquecombinationsand25intotal.
Forabetterunderstandingofthestudiesresult,weshowasingleexampleofour simulationresultsinFig. 3,where M3 and M6 areagainsetto20%-bottlenecks.Inall followingvisualizations,thedashedlinemarksthebottleneckstationsofthescenario. Werunallsimulationsfor20,000units,precededby5,000unitsforinitialsettlingofthe system.TheY-axisshowsthebottleneckstationidentifiedbytherespectivedetection method,withthedetectionshownusingthebluemarkers.InFig. 3,forexample,a shiftingbottleneckoccursafterabout9,500units,thebottleneckat M3 shiftsto M6 Moreover,sinceadetectionismadeforeachpointintime,unambiguousvisualization isnotalwayspossible.Intheexample,after14,500timeunits,thebottleneckrepeatedly alternatesbetweenstation M3 and M6 foranintervalofapprox.1,000units.
Right-skewed distribution of process times
Fig.2. Distributionofprocesstimesfor bottleneckandnone-bottleneckstations
Visualization of detected bottleneck stations
Fig.3. Exemplarystructureofthedetection resultsfora20%-BNat M3 and M6
4DetectionResultsusingBNW,APMandITV
AllvisualizationsinSec. 4 usea20%factorforthebottleneckstations.Thelocationofthe suchstationsisvariedbetween M3 and M6 ,leadingto25combinationsforeachdetection method.Tocomparetheresults,weusematrixplotsthatcontainall25combinations. Thediagonalofthematrixshowssimulationswithjustonebottleneck.Allotherfields showsimulationswithtwobottlenecks.Sinceweusethesamerandomseedtodetermine theprocesstimevariability,theresultmatrixissymmetrical.Assuch,thesameresult isobtainedfora M2 -M4 systemasfora M4 -M2 system.Thevisualizationlogicofeach tileinthematrixfollowstheformatpresentedinFig. 3
4.1BottleneckDetectionwithBottleneckWalk
Figure 4 showstheresultsofbottleneckdetectionusingtheBNW.Asdescribedabove, thematrixissymmetrical.Inthecasesonthediagonal,inwhichonlyonestationhas anincreasedprocesstime,theBNWidentifiesthisstationasabottleneckinmostcases. Intheothercases,thestationsmarkedwithdashedlines,whicharethestationswith anincreasedprocesstime,arealternatinglyidentifiedasbottlenecks.Inadditionto theapparentlycorrectlyidentifiedbottlenecksituations,allplotsshowascatteringof selectivelyidentifiedbottlenecksatallstations,rangingfrom M1 to M7 .
4.2BottleneckDetectionUsingtheActivePeriodMethod
Next,weidentifiedbottlenecksusingtheAPM.Forthesakeofclarity,welimitthe useoftheAPMtoanidentificationofsolebottlenecks.Insituationswithashifting bottleneck,werefertothestationwiththelongestactiveperiodasbottleneck.Only afteraninterruptionofthisstation,theothershiftingbottlenecktakesoverthelabel ofsolebottleneck.TheresultsofthisinvestigationareshowninFig. 5.Again,clearly identifiedbottlenecksareevidentinthefivescenariosinthediagonalofthematrix. Intheremainingscenarios,changingbottlenecksbetweentheexpectedstationscanbe
Study results for bottleneck detection using the Bottleneck Walk with 20%-BN BN(M4)BN(M5)BN(M6)
BN(M6)BN(M5) BN(M4)BN(M3)BN(M2)
BN(M2)BN(M3)
Fig.4. Matrixplotofallbottleneckcombinations(20%),usingtheBottleneckWalk
Study results for bottleneck detection using the Act. Period Method with 20%-BN BN(M4)BN(M5)BN(M6)
BN(M4)BN(M3)BN(M2)
BN(M6)BN(M5)
BN(M2)BN(M3)
Fig.5. Matrixplotofallbottleneckcombinations(20%),usingtheActivePeriodMethod identified.Itisnoticeablethatshiftsinthebottleneckstationsoccurmorefrequentlyif thestationsarelocateddirectlyafteroneanotherorareonlyashortdistanceapartinthe
valuestream.Althoughtherearealsoisolatedpointwiseidentifiedbottleneckstations, theseoccurmuchlessfrequentlythanpreviouslyforBNW.Overall,theresultsshowa clearimageoftheexpectedsystembehaviorinallscenarios.
4.3BottleneckDetectionUsingInterdepartureTimeVariances
Figure 6 showsthebottleneckstationsdeterminedusingITV.SinceITVintheoriginal proposalisanAVM,weadaptedthemethodformomentarybottleneckdetection.As alreadymentionedinSec. 3,weuseaslidingwindowtocalculatetime-dependent variancesforeachpointoftimeinthesimulations.Weuse5,000asthelengthofthe varianceinterval,whichcorrespondstotheperiodoftheinitialsettlingofthesystem. Then,foreachpointoftime,theITVwithinthisslidingwindowisused,whilethelowest ITVmarkstherespectivebottleneckstation.Onceagain,theindividualbottlenecksare reliablyidentified,withonlyaminoranomalyintheM5-M5scenario.Intheother scenarios,theinfluencedstationsareagainidentifiedasthemainbottlenecks,with shiftingstatesoccurringhereaswell.Additionally,itisnoticeablethatthefrequency oftheshiftingisnotasstronglyinfluencedbythedistancesbetweenthestationsinthe valuestreamasseenfortheAPMdetection.Thenumberofindividualidentificationsis slightlyhigherthanitwasforAPM,butsignificantlylowerthanforBNW.Insummary, theresultsoftheITVmethodshowaslightlylessclearresultthantheAPM,butare considerablymoreuniformthanthedetectionofBNW.
Study results for bottleneck detection using the Int. Time Variance with 20%-BN BN(M4)BN(M5)BN(M6)
BN(M2)BN(M3)
Fig.6. Matrixplotofallbottleneckcombinations(20%),usingInterdepartureTimeVariances
5Comparison
Wefirstaddressthecomparisonwiththe20%bottlenecksasshownbefore.Sincethe processtimeincreasehasasignificanteffectonthesystembehavior,wealsocompare theresultsfor10%to100%increases,focusingontheagreementofthemethods.
5.1Comparisonof20%-BottleneckResults
WiththehelpoftheBNW,itisrelativelysimpletorecognizebottlenecksituationsinthe systemandtodetectbottlenecksaccordingly.However,alternatingbottlenecks,between twostationswithincreasedprocesstimes,occurmuchmorefrequently,whencompared toAPMandITV.Furthermore,onlythevisualizationoftheBNWshowsanextensive scatteringofselectivelyidentifiedbottlenecks.Thesevariationsoccurredmostlikelydue totheratioofaverageprocesstimeandmaximumbuffercapacity.Withacapacitylimit of5unitsassumedinthecasestudy,thelowerandupperdecisionlimitsareclosetoeach other.Onthisscale,onethirdrespectivelytwothird,areunreliablecapacitylimitsdueto therule-of-thumbsassumptions.ItispossiblethatBNWismoresuitableforbottleneck detectioninsystemsofadifferentnaturewithachangeinthisratio.However,inthe setupwitha20%increaseinprocesstimeatbottleneckstations,theBNWperformsthe worst.
APMhasproventobeamorerobustmethodforbottleneckdetection.Itshowsalmost nopunctualdeviationsanddesignatesmainlytheinfluencedstationsasbottlenecks. However,thisobservationhastobequantifiedtosomedegree.Byfocusingoncontinuous activeperiods,itisinthenatureofAPMthatlongerperiodsofutilizedstationslead toacontinuousbottleneckidentification.Themethodisthuspartiallyprotectedfrom short-termandvariability-induceddeviations.ItmayevenbearguedthatAPMisa memory-basedmethodduetothisimplicituseofpastsystems,effectivelybecominga hybridofAVMandMVM.Regardingthedecreasingfrequenciesofbottleneckshifting formorewidelyseparatedstationsinthevaluestream,thisislikelyduetotheincreasing overallworkloadbetweenstations.Withonlytwostationswithhigherprocesstimes, thestationsinbetweenactlikeasharedbuffer,makingstarvationorblockingconditions lesslikelytooccur.Summarizing,theAPMisprovingeffectiveinidentifyingshifting bottlenecks.
LiketheAPM,themodifiedITVhasalsobeenshowntobewellsuitedforbottleneck detection.TheadvantageofusingITVisthatbottleneckscanbeidentifiedselectively, especiallyinlongervaluestreams.Thephenomenonofasplitbufferathighworkloads betweentwostationsislesslikely.Adisadvantageisthatimmediatebottlenecksare subjecttoaslightlyhigherdelayduetotheaverage-baseddeterminationofthevariance. Inaddition,thelengthoftheselectedvarianceintervalhasasignificantimpacton thequalityofthedetection.Thedeterminationofasuitableintervallengthishighly application-dependent.Inordernottoexceedthescopeofthiswork,wehadtorefrain fromanadditionalvariationofthisparameter.
Weprovidethesimulationmodel,thegenerateddataandallvariationsofbottleneck processtimehere:github.com/nikolaiwest/2021-bottleneck-detection-ieaaie.
5.2ResultsforVaryingBottleneckProcessTimes(10%to100%)
Todeterminethecorrectnessofadetectedbottleneck,awaytoidentifyitaccording toasinglesourceoftruthisrequired.Suchatruthcanthenbeusedtocrosscheckthe degreeofcorrectidentification.However,sincesuchmeasuresalwaysdependonthe selectedbottleneckdetectionmethod,wehavenowaytodetermineanabsolutetruth regardingthebottleneckstation.Instead,wehavetocomparethestatementsofthethree previouslyusedmethodsBNW,APMandITV.Allagreementratiosthereforeconsistof pairingsofmethods.Ineachcase,wedeterminetheaverageratioofagreementfromthe 25scenarios.Thiscalculationisperformedforbottleneckstationsofvaryingincrease ofprocesstimes.Wecompareanextraof10to100%ofthebasicprocesstime.Table 1 summarizestheresults.
Table1. Averageratioofagreementbetweentwodetectionmethodsfordifferentadditionstothe stations’bottleneckprocesstimes
Theratiosofagreementaregivenaspercentages,with100%correspondingtoa fullagreementand0%tonoagreement.BNWandITVhavethelowestdegreeof agreementat39.26%withabottleneckincreaseof10%,andanaverageof56.33%, APMandITVhavethehighestagreementat90.55%witha90%bottleneckincrease, whilereachinganaverageof80.10%.Throughthisrepresentation,aclearrankingofthe methodpairsbecomesapparent.ITVandAPMalwayshaveahigheragreementratio thanBNWandAPM,whereastheratioforBNWandAPMisalwaysaboveBNWand ITV.ThesedeviationsareprobablyduetothesporadicdeviationsoftheBNW.Although themethodrecognizestheinfluencedstationsasbottlenecksinmanycases,itisalso frequentlywrongduetothesporadicallyrecognizedbottlenecks.
Furthermore,allpairingsshowatendencyofanincreasedratioofagreementwith moreadditionalprocesstimeforbottlenecks.Thisincreasecanbeattributedtoatendencytowardsstaticbottlenecksinthefrontsectionsofthevaluestream.Duetothe increasingdifferencesinprocesstimesonbottlenecks,thedurationandfrequencyof earlybottlenecksincreases.Consequently,thedetectionofsuchbottlenecksbecomes easierforallthreemethods,leadingtoanincreasingtherateofagreement.