《BCG-工业分销商价值创造者2023（英）_市场营销策划_重点报告202301202_doc.docx》由会员分享，可在线阅读，更多相关《BCG-工业分销商价值创造者2023（英）_市场营销策划_重点报告202301202_doc.docx（15页珍藏版）》请在三一办公上搜索。
1、GenerativeAPsRoleintheFactoryoftheFutureDECEMBER08,2023ByDanielKupper,KristianKuhlmann,MonikaSaundersjJohnKnapp,Kai-FredericSeitzjJuIianEnglberger,TilmanBuchner,andMartinKleinhansREADINGTIME:15MINGenerativeAlisoneoftoday,shottestbusinesstopics,withcompaniesexploringitspotentialapplicationsandbenefit
2、sacrossindustriesandfunctions,includingmanufacturing.Butdespitetherecentbuzz,manufacturersshouldrecognizethatsimplyapplyingtoolslikeChatGPTontheirownwillnotrevolutionizefactoryoperations.InsteadofreplacingtraditionalAl,GenAIofferscomplementaryusecasesintheareasofassistance,recommendations,andautonom
3、ythatpavethewaytothefactoryofthefuture.Itdoessothroughitscapacitytogeneratecontent,suchastextandimages,tailoredtospecifictasksorinquiries.(SeeuHowGenAIWorks.n)Howgenaiworks-TodiscusstheapplicationsofGenAI,itisessentialtofirstdefinehowitdiffersfromttclassica,machinelearning(ML).ClassicalMLalgorithmsd
4、iscernpatternswithinobserveddata,enablingthemtogeneralizetheseinsightstonew,previouslyunseendata.Forinstance,anMLmodelmightbetrainedusingspecifictextfragmentssuchasoperatorincidentreportsinwhichmachinebreakdowndescriptionsareclassifiedintospecificrootcausessuchas,endoftoolinglifeoroperatorerror.Base
5、donthistraining,themodelcanprocesspreviouslyunseentextfragmentsofincidentreportsandjudgewhatcausedtheincident.Thebasisforsuchmodelsmaybedeepneuralnetworks,supportvectormachines,orothermethods.GenAItakesthisapproachfurther.Beyondmerelyclassifyingexistingtext,itcangeneratenewtextbasedonspecifiedcriter
6、ia-suchasoperatorinstructionsthatoutlineaprocesstoresolveaparticularrootcauseofamachinebreakdown.AlthoughtheprogressionfromclassicalMLtoGenAImightseemincremental,itposesafundamentaltechnicalchallenge.InclassicalML,themodelmerelyneedssuffcienttrainingtoconfidentlycategorizeatextfragment.Incontrast,Ge
7、nAImustconstructatextfragmentfromindividualwordsandletters,ensuringthatitisgrammaticallycorrect,comprehensible,andaccuratelyrepresentstheprocess.ThenumberofpotentialoutputsfromGenAIisvirtuallylimitless.Consideringthatthereareroughly170,000Englishincurrentuse,amerefive-wordtexthasmorethan140septillio
8、npotentialcombinations.Ontheotherhand,onlyafractionofthemwouldbegrammaticallycorrectandunderstandable.Amongthose,anevensmallerfractionwouldaccuratelydescribeagivenprocesstofixtherootcauseofamachinebreakdown.Consequently,themarginforerrorinGenAImodelsisincrediblynarrow,necessitatingextremelyprecisemo
9、dels.Toattainthisprecision,GenAImustuseufoudationalmodels5*insteadofbeingtrainedonlyoncontext-specificdata.Foundationalmodelsaretrainedonextensivedatasets,suchasallavailabletextorimagesonline,andaresubsequentlyfine-tunedforspecificapplications.Thesemodelscanbelargelanguagemodels(suchasOpenAsGPT-4orA
10、mazonQ)orimageorspeechmodels.Theyseemtogainanunderstandingofrealityfromtheextensivedatasets.However,foundationalmodelsareobservationallearnersthatdonotapplylogicorreasonashumansdo.ThismeansthatthereisnoguaranteeofplausibleoraccurateresultsfromGenAI.Inouroperatorinstructionexample,thefoundationalmode
11、lfirstlearnswhatconstitutescomprehensibleandaccuratetext,withprocessdescriptionsbeingasmallsubset.Next5themodelisfine-tunedbylearningwhatoperatorinstructionslooklikeandhowtheycorrelatewithgivenmachinebreakdownrootcauses.However,thereisnoassurancethatthemodelwillcreatecorrectorhigh-qualityoperatorins
12、tructions.Ergonomicsillustratestheproblem.BecausetheGenAImodellacksinsightintotheprocessthatfixestherootcauseandthepeoplewhoaretheoperators,itmightoverlookpotentiallimitations,suchasinfeasiblemovementsorinaccessiblespaces.Asaresult,aqualityassessmentisalwaysrequiredtoensurethattherecommendedremediat
13、ionispracticalfromanergonomicperspective.GenAsgreatertechnicalcomplexityelevatestheimportanceofestablishingarobusttechnologicalfoundationtoharnessitscapabilitieseffectively.Withanumberofarchetypespossible,manufacturersmustunderstandthefactorsthatdetermineanoptimalchoice.Theycanapplythisknowledgetoin
14、tegrateGenAIintofactoryoperations,consideringthevalue-addingapplications,change-andpeople-relatedinitiatives,andrequiredtechnologicalinfrastructure.ManufacturersArePrioritizingGenAIforitsDisruptivePotentialBCGrecentlysurveyedmanufacturerstounderstandtheirperspectiveontechnologydevelopments.(SeeAbout
15、theSurvey/*)Regardlessoftheiraffinityfordigitaltechnology,manufacturingexecutivesrankedAl(includingGenAI)firstamongtechnologiesthatcouldpositivelydisrupttheiroperations.(SeeExhibit1.)ThepotentialROIwarrantstheirenthusiasm.ABCGanalysisfoundthattheuseofAlcouldenhanceshop-floorproductivitybymorethan20%
16、.ABOUTTHESURVEYBCGconductedaglobalsurveyfromJanuarythroughMarch2023toassessthelatesttechnologiesinthemanufacturingindustry.Approximately1,800respondentsfrom15countriesspanningNorthandSouthAmerica,Europe,andAsiatookpartinthestudy,witheachcountrycontributingmorethan100completedsurveys.Theparticipantsr
17、epresentabroadarrayofproductionindustries,includingautomotive,capitalgoods,consumergoods,energy,IT,healthcare,andmaterials.Exhibit1-ManufacturingExecutivesRankArtificialIntelligence,IncludingGenAI,astheMostDisruptiveTechnologyTechnologiesthatcouldpositivelydisruptmanufactung,rankedbymanUtactunngexec
18、utivesExecutiveswhohavealc for digital technologiesArtificial Intelbgence (including GenAI)2 1 Industry 4.0 and cyber-physical systems3 Additive manufactunng4 Industnal metaverse5 VlrtUal and augmented realityExecutives who have a higher affinityfor digital technologiesArtificial intelligence (inclu
19、ding GenAI)Z Industnal metaverseJ Mb3 i Industry 4.0 and cyber-physical systems4 1 Additive manufactunng5 Virtual and augmented realitySource:BCGglobalsurveyof1,800manufacturingexecutives.Foroneautomotivesupplier,deploymentofAlresultedina21%boostinproductivity,withROIobservedbetweenonetothreeyearsac
20、rossvariousapplications.OneapplicationwasanAl-poweredscrapadviserthatprovidedoperatorswithoptimizedparameters,cuttingscrapratesby25%.Anotherwasapumphealthmonitorthatnearlyeliminatedbreakdownsofacriticalproductionpump,enhancingoverallequipmenteffectivenessbymorethansevenpercentagepoints.Athirdapplica
21、tioninvolvedimplementinganAl-drivenvisualqualityinspectionsystemthatdetectedaestheticdefectsinproducts.Thissystemreducedtheneedforqualitycontrolstaffby65%whileimprovingtheaccuracyoftheinspections.Altoolshaveavarietyoftechnologicalunderpinningsandapplications.TraditionalAlapplications,builtonmachinel
22、earning(ML)anddeeplearning(DL),havegainedtractioninrecentyears,primarilytofacilitatedataanalysis,classification,clustering,andranking.Thesecapabilitiessupporttaskssuchasanomalydetectionandpatternidentification.Incontrast,thegroundbreakingaspectofGenAItoolslikeChatGPTistheircapacitytocreatevarioustyp
23、esofnewcontentsuchastext,code,andimages一inresponsetopromptsinmultipleformats.HowGenAIHelpsPavethePathtotheFactoryoftheFutureAlthoughGenAIintroducesarangeofinnovativefeatures,itisnotideallysuitedfortaskssuchasanomalydetection,productionanalytics,orsetpointoptimization.Forthesetasks,thetraditionalAlth
24、athasbeenavailableinrecentyearsisbetterequipped.Nonetheless,GenAIhasacomplementaryrolethatwillsignificantlyaidmanufacturers*effortstorealizethefactoryofthefuture.Itsuniquecapabilitiescanenablemanufacturerstoautomateandenhancefactoryactivitiesandsupporttheirworkforceinnovelways.Manufacturershavebeenu
25、singtraditionalAltosupporttransparency,predictability,andself-controlledsystems,whicharethekeycharacteristicsandmaturityphasesofthisenvisionedfactory.(SeeThreeLevelsofDigitalMaturityintheFactory.)NowGenAscontent-creationcapabilitiesenablethreetypesofusecasesinmanufacturingassistancesystems,recommend
26、ationsystems,andautonomoussystems-thatcorrespondtothesematuritylevels.(SeeExhibit2.)THREELEVELSOFDIGITALMATURITYINTHEFACTORY-Thefirstlevelofdigitalmaturityinvolvesusingdatatocreatetransparencyaboutwhathasalreadyhappenedinthefactory,suchasthroughadigitalperformancedashboarddisplayingKPIs.Thenextlevel
27、leveragesdata,analytics,andAltoanticipatefutureevents.Operatorsusetheinsightstostabilizeprocessesforinstance,theyreceivealertswhenapatternsuggestsanimminentbreakdownandcanthenconductpreventivemaintenance.Themostadvancedlevelischaracterizedbyself-controlledsystems,suchasautonomousmobilerobots.Exhibit
28、2-GenAISupportsthePathtotheFactoryoftheFutureFactoryofthefutureDigitalmaturity Assistance systemsDoing (manual) tasks more efficientlyStability and predictabilityGenAI use case types Recommendation systemsFinding the best solutions anddoing the right thingsSelf-control andstructural optimizationG ；匕
29、廿1 j(3)Autonomous systemsSolutions that can self-adaptto new environmentsResearth stageSource:BCGanalysis.AssistanceSystems.GenAIapplicationsinthiscategoryraisetheefficiencyofhands-ontaskssuchasprogrammingormachinemaintenance.Forexample,automationengineerstraditionallyhadtomanuallyprogramandcodeauto
30、mationsolutionsformachinesandprogrammablelogiccontrollers(PLCs).ButaGenAItool,akintoChatGPT,canusetextinputstoautomaticallygeneratecodeorcodeblocks.Thisreducesthetimeandeffortrequiredforautomationengineering,therebydecreasingtherelatedexpenses.Anengineerwouldneedtofocusonlyonreviewing,adjusting,andf
31、inalizingthecode.flGenAI,scontent-creationcapabilitiesenablethreetypesofusecasesassistancesystems,recommendationsystems,andautonomoussystemsthatcorrespondtomaturitylevelsinthefactoryofthefuture.Asimilar,yetevenmorepowerful,impactisachievedbyusingGenAItocodifytheknowledgeandexperienceofworkerswhohave
32、decadesoffactoryexperiencebutmaylackskillsinmodelingandanalytics.GenAIassistsintransformingworkers*experience-basedintuitionandknow-howintodata-driven,validatedrecommendations.Forinstance,itcanvalidateaworker,sinsightsonadjustingmachineparametersorresolvingproductionanomaliesbybuildingamodelinaprogr
33、amminglanguagesuchasPythontoderivesuchinsightswithdataandanalytics.Thismodelisthenavailabletoassistwithfutureproblemsolvingandanalysis.RecommendationSystems.GenAItoolscanproviderecommendationsthathelpworkerstoidentifythebestmethodsforspecifictasks.Animprovedapproachtopredictivemaintenanceprovidesapr
34、imeillustrationofhowGenAItoolswillcomplementtraditionalML/DL-basedAl.Inthepast,manufacturerssoughttopreventbreakdownsbyperformingmaintenanceaccordingtofixedcyclesorperiods,ortheymaderepairswhenbreakdownsoccurred.WiththeadventofML/DL-basedAl,manufacturerswereabletousedatafromdifferenttypesofsensorsto
35、identifypatterns,predictbreakdowns,andthenproactivelyconductmaintenance.GenAIenhancesthispredictivemaintenanceapproachbyautomaticallycreatingtextorimagesthatprovidestep-by-stepinstructions,includinglistsofrequiredspareparts.Suchasystemallowsthemaintenancestafftospendmoretimeperformingtasksinsteadofp
36、reparinginstructionsenhancingproductivityandreducingcosts.EveninexperiencedtechnicianswouldbeabletorepairormaintainequipmentmoreeffectivelywithsupportfromaGenAItool.flGenAIassistsintransformingworkers*experience-basedintuitionandknow-howintodata-driven,validatedrecommendations.AutonomousSystems.Them
37、ostadvancedsystemsuseGenAsabiltytoattainself-regulationcapabilitiesandtoadapttounfamiliarsituationsDevelopersarecurrentlyexploringGenAIsolutionsthatwouldallowmachinerytoself-adaptto,forexample,newenvironments.Considertheautonomousdeploymentofrobots.Today,operatorsmustmanuallyexecutecertainmaterialha
38、ndlingactivities,especiallynonrepetitiveprocesses,becauseautomatingsuchactivitiesisverydifficultandrequiresintensiveeffort.Lookingahead,GenAI,incombinationwithreal-worldroboticsdatasets,willenablemultimodalrobotstotranslateoperators*languageprompts(forinstance,uBringmesparepart47/1,)intoasequenceofa
39、ctionsthatthesystemthenexecutestoperformmaterial-handlingtasks.Thisadvancewouldreducetheneedfortask-andenvironment-specifictraining,datalabeling,andfrequentretraining.Itthushasthepotentialtoreduceengineeringexpenses,replacemanualactivities,andboostproductivity.Anotherexampleisthegenerationofsyntheti
40、ctrainingdataforAl-basedsystems,suchasqualitycontrolusingcomputervision.Thisapplicationexpeditestherampingupofthesystemsbyeliminatingtheneedtocollectrealtrainingdataduringproduction.DeployingtheRightTechnologyFoundationTosuccessfullyimplementandscaleAlinmanufacturing,itisnotenoughtopinpointvalue-add
41、edapplications.Itisalsoimperativetosetuptherightfoundationswithrespecttobothpeopleandtechnology.Thepeople-relatedrequirementsforGenAIdeployment,withrespecttobuildingthecapabilitiesneededtodevelopandoperateAl-basedapplications,aresimilartothoserequiredfortraditionalAldeployments.However,thetechnology
42、foundationismorecomplexforGenAI,makingitcrucialtoemphasizethetechnologyinfrastructureandoperatingmodel.ThetypicaltechnologyinfrastructureforthescalableuseofAlinmanufacturingconsistsofsixbuildingblocks:datasourcing,dataprocessing,GenAIapplications,computing,connectivity,andcybersecurity.GenAIaffectst
43、herequirementsforseveralofthese.(SeeuTheBuildingBlocksoftheTechnologyInfrastructure.1,)THEBUILDINGBLOCKSOFTHETECHNOLOGYINFRASTRUCTUREThefirstthreetechnologyinfrastructurebuildingblocksaddressdatasourcingandprocessingandtheGenAIapplications.Thelastthree-computing,connectivity,andcybersecurityplayfaci
44、litatingrolesacrossthetechstack.(Seetheexhibit.)GenAITechnologyInfrastructureintheManufacturingEnvironmentEna(emcM lyw*DataapplicabonI tc cfccAl appbcanom nd uter VnerfacGws purpose IfPt*feComputing4 r.,. ，, Btaran ldnoMMkMMoJMMmMMkMM*W0PWWNVtMM We(ADnefW MFKWVL wcqpMiinBf*nMctmMOCud UVfHantWVdHIfVi
45、ttIttcxun,tfctM)r andrvtpoMFitww.VWfUSOfl 0ecnon*MrngTranMCtlonal UytrOtherIyltem*MESRhm4m Bndm4AafemnPLCVMT 4vKtBackups, mwfSourerCG114iymNote:fRP-mfpmmoure*rann(.oT-internet/Thtftft,MES-ZAMtK50CMDon0*mlPtXpre(ranmAbtelferycontrolanddataMquiunon,SCM-Mp0yCMNIm*n*f三wn;TMS-tran4porunonEanag11wnfUm;WMS
46、-Wamhut11MtwntvyUm,NwCAJppandftnt*aaerc9tratyt*mndr*M*t*ftot4unirfof(fb*xhon*lmodimiftbrvtrdlAddmoiMiOmMarXWtotuMtMInUtraC8ofdocumnu4RequiretComMfUnMrofct11p*oceuan(umna11raptcsPrOCeMmIumtXiuirdraJimanaffnntaHor,Mpon*andIvmtsIrtofmMtonaccuDataSourcing.BecauseAltoolsarefundamentallydata-dependent,man
47、ufacturersneedtohavetherequisitesources.Datacomesfromthreelayers: FieldLayer.Sensors,PLCs,andinternet-of-things(IoT)devices. ControlLayer.Manufacturingexecutionsystem(MES)andsupervisorycontrolanddataacquisition(SCADA)systems. TransactionalLayer.Forexample,theenterpriseresourceplanning(ERP)system.DataProcessing.Thisinfrastructurepre-processes,processes,anddistributesdataforanalyticsandAl.Componentsincludedataingestion(suchasedgeanalyticsandIoThubs),processing(forexample,cleaning,filtering,andContextualization),storage(suchasdat
