埃森哲-未來的工廠：使用知識圖和生成式AI構建數字孿生的案例研究.pdf

《埃森哲-未來的工廠：使用知識圖和生成式AI構建數字孿生的案例研究.pdf》由會員分享，可在線閱讀，更多相關《埃森哲-未來的工廠：使用知識圖和生成式AI構建數字孿生的案例研究.pdf（34頁珍藏版）》請在三個皮匠報告上搜索。

1、Factory of the FutureA Case Study Building Digital Twins using Knowledge Graphs and Gen AIDatabricks2023AgendaState of Digital TwinsTwin of Twins w/Knowledge Graphs and Data ProductsWhats Next?Twins w/Semantic Layer and Generative AI1_DAIS_Title_SlideDigital Twinsw/Generative DesignWhat the industry

2、 is looking forEnd to End AutomationAbility to capture and utilize corporate/tribal knowledge along with automation of mundane/non skilled tasks to best utilize current resources and alleviate the engineering skills gapIntelligent OptimizationInsights to support early decision making;access to desig

3、n information,costs,schedules,and risks are key to successful design creation,project planning,and resource planning to avoid rework in later project phasesImprove QualityProactive actions not reactive.Identify causes of cost over-runs and schedule extensions with domain knowledge data to create des

4、igns and plans which leverage prior project performanceProjects with high quality delivered within time and cost constraints with resilience to changeEvolution in manufacturing twin19801990200020152020Traditional DesignParametric DesignGenerative DesignComputer Aided DraftingParametric ModellingComp

5、utational ModellingOption GenerationDesign OptimizationIn Traditional Design,a designer uses basic techniques like sketching to describe the idea that is inside the mindIn Parametric Design,the user defines relationships between traditionally drawn or sculpted elementsIn Computational Modelling,the

6、user explicitly describes a process to create a design outcomeWith Option Generation,the user explores a variations of computed rules given different starting points for the calculationsWith Design Optimization,the user defines explicit goals,and a computational or parametric model is automatically

7、explored for states that fit those goalsKey DriversKey DriversCompetitive MarketEnhanced Use of AI&MLComputational PowerCost ReductionTime SavingMinimizing WasteBring competitive designs,processes and services to the market quicklyArtificial Intelligence-based algorithms produce a wide range of desi

8、gn alternatives evaluated with Machine Learning Increased productivity with thousands of the right iterations delivered quicklyOptimized designs with less raw material usageReduced design time and faster processing of dataReduced need of iterative physical prototyping&selection of optimized processI

9、llustration Purpose OnlyTraditional approach to design generally results in only one output?Parametric DesignGenerative DesignApproachXXXXXXX134XX2XXGenerative design explores a larger design space to unlock optimal solutionsParametric DesignGenerative DesignApproachLayout OptimizationSchedule Optim

10、izationXXXXXXX134XX2XXEquipment SizingProcess OptimizationApply a knowledge graph to generative design to optimize processesParametric DesignGenerative DesignVia Knowledge GraphApproachLayout OptimizationSchedule OptimizationXXXXXXX134XX2XXEquipment SizingProcess OptimizationCombine AI with generati

11、ve capabilities to produce optimized deliverablesParametric DesignGenerative DesignVia Knowledge GraphApproachAutomated generation of deliverablesEnd to End ProcessAI-PoweredIncluding Gen AI1_DAIS_Title_SlideTwin of Twins w/Knowledge GraphsCase Study:Warehouse AutomationWarehouseDistribution Center1

12、234512345Distribution Center Automation Use Cases:Describes Warehouse and Distribution Center archetypes and use casesAutomation Vendor Execution:Automation vendor capabilitiesAI-Enabled Orchestration:Knowledge graph powers an AI engine to optimize warehouse operationsWarehouse Simulation:Vendor-neu

13、tral process simulation and probabilistic models to uncover bottlenecks and enable what-if scenarios3rd Party Twin&Data Integration:3rd party vendors(i.e.,Dematic,KOI Reader,Worlds,Locanis,etc.)to collect real-time data and insightsChallenges with Existing Digital Twin(s)Automation and robotics equi

14、pment need to be properly orchestrated by process and tools to produce positive ROI The company invested heavily in automation point-solutions on their warehouse,product flow and fulfillment accuracy were measured 20%below target;largely due to complexities in properly orchestrating resolutions to u

15、nplanned eventsVendor-neutral simulation capabilities are needed to reduce potential vendor bias during the design/procurement processThe company purchased a fleet of AGVs based on vendor-provided simulation and recommendations,resulting in over-purchasing and significant under-utilization.Roughly 1

16、/3 of the AGVs were idle,with additional“spare”unitsProcess digitization and automated data capture of completed workflows is a key step towards enabling autonomous systemsThe company struggled with maintaining accurate product and equipment availability metrics and were challenged in effectively or

17、chestrating timely resolutions to unplanned events.Operators captured most of their updates on paper,leading to poor visibility and poor data for optimization analysisSummary of ChallengesTwins not aligned with processDifficult to enable interoperability across twinsBiased twins need baseline w/real

18、&synthetic dataGaps in DataWhat were seeing at one manufacturer“Twin of Twins”unlocks synergies across twinsNowSiloed digital twins that dont talk to each other leading to fragmented decision-making.FutureAll twins and enterprise data connect to form a Twin of Twins,eliminating siloes.Bringing toget

19、her people,processes,products,materials,equipment and technology for system wide interoperability and actionable insights.Aligns with the concept of data meshInternal DataExternal DataPhysical LandscapeTwinsSystem of RecordsTwin of Twins(Powered by Knowledge Graph)AI&Analytics(Including LLM,ML)OneAP

20、IJetson&MetropolisPredictSimulateAutomateOperateVisualizePhysical TwinStorage Digital TwinAGV Digital TwinMachine Digital TwinDifferent Twins for Different NeedsComputer Vision Automated Data CollectionYard OperationsSynchronization of warehouse operationsCustomized solutions for unique situationsAu

21、tomated Storage&Retrieval SystemLive Data Stream to automate critical variations and process improvementYard Check InAI-Powered Automated Queuing and OptimizationAutomated Product MovementProduct PickingHigh-density storage system to move and house productsSite-Wide OperationsIncreased Data Quality&

22、CollectionIncreased Yard EfficiencyIncreased EfficiencyIncreased Versatility with Complex IssuesIncreased ThroughputIncreased Efficiency&SafetyExample:Distribution Center Value INBOUNDOUTBOUNDOUTBOUNDOUTBOUNDPRODUCT XA Knowledge Graph captures Domain NodesThe digital representation of our physical a

23、ssetNodes have types whose properties are outlined by the ontology EdgesFormalizes semantic relationships within graphCan point from node to node or node to dataAlso outlined in the ontologyDistributionCenterYaskawa Picking RobotDistributionCenterYaskawaPicking RobothasDevicehasDeviceManfacturing&Re

24、source Capability Ontology MaRCO,Jrvenp et al.Production AreaWarehouseMovingAutonomousMobile RobotPicking RobotChips LineFactory FloorSite,Area,Line,Cell,StationOrdering and position are represented in ontologyStations have individual devices or device combinations Device Combinations Complex device

25、s are decomposed into other smaller Complex Devices and Individual Devices Individual Devices are the baseline Device.Device CapabilitiesAll devices have a capabilityCapabilities describe what a device can do,and the constraints to that actionDevice Combinations have capabilities with more rigid rul

26、es based on individual component capabilitiesProduct1Product3Product2StorageTrailerWarehouseStorageTrailer ID:BALEAllocationAreaA11011WWhere digital twin is a data productE.g.,Truck Unload&Put-awayDigital TwinPut-Away Logic(Simulation)Unload Schedule(Simulation)Trip ID1Trip ID2Trip ID3Unload schedul

27、eBy running this module of the simulation,we can generate an unload schedule subject to the missing blocks that affect unloads such as audits and crews working on other tasks.The model is object-oriented;this means that each entity of the model is an individual object that can be replicated.While de

28、veloping the model we create these objects in a way that can be reused at other locations.During the development,logics defined in the conceptual model are translated into code to mimic the system behavior.Put-away logic will be the responsible of storing the pallets in its right location when an in

29、bound truck arrives to the warehouse.Use CasesTwo use-cases using the same simulation models:Trip Activation Schedule:the simulation model is used as a“brain”for the orchestrator generating the best trip activation schedule considering all the activities that happen inside the warehouseStrategical W

30、hat-If scenarios:test different scenarios and simulate longer periods of time to see the impact that changes in the system would haveSimulation:Optimal Layout Design and Trade-offsArea Simulation Phase 0 1 Site Simulation Phase 1 1+Coverage:Collection of modeled assets that can simulate outputs in a

31、 selected areaPurpose of Use:Simulate operations flow and metrics through selected area based on minor to moderate changes in layoutReuse:Assets and area(e.g.,behavior of forklift or expansion,etc.)can be reused to expand coverage,or enhanced to increase purpose of useCoverage:Collection of modeled

32、areas and processes that make up the entire siteReuse:Assets are modeled as modular components with input/output that can potentially be combined to create new scenarios.Previously modeled sites can accelerate design/simulation of similar new sitesLayout Optimization:Simulate effectiveness of wareho

33、use layout with moderate to significant changes in designNew Location:Re-useable assets and services that accelerates layout design for new locations(with similar processes and goals)Simulated Data Generation:Generate simulated data based on modeled assets to create a simulation environment Orchestr

34、ation:Optimal Execution of OperationsSingle Rule-based ScenarioRule-based Scenario CatalogScenario Prioritized for MLOperations AnalyticsProcesss Step:e.g.,time it takes to load a 52ft trailer for order X.Each step contains a model that simulates an output based on inputs and boundary conditionsInpu

35、ts:consists of system,user,device,and app dataOutputs:may require validation based of physical constrainsScenario:e.g.,collection of steps that describes the process to handle a Live LoadRelationship between steps are contextualized by capturing operator knowledge 23Configured Scenario:collection of

36、 pre-configured scenarios to select and executeEarly Indicators:predictors and early warning for warehouse issues,to aid operations planning and improve results Operations AnalyticsOperations Dashboard:querying and display of aggregated metrics that show operations trends Prioritized Scenarios:selec

37、t high-value,complex scenarios that repeat often,evolved from rules-based to stochastic-based modeling and MLIntegration of Digital Twins w/Phygital World Real-time VisibilityKnowledge Graph and OrchestrationProcess and Asset SimulationAutomation+Connected Shop FloorRetrieves&contextualizes data for

38、 use case/KPIsTrailer InactivityTrailer Fill TimeDESIGN:Enables optimal system design choices through simulationsWhat-if ScenariosDesign Trade-OffsSynthetic Training DataOPERATIONS:Provides impacts of unexpected eventsTripsCrewProductSensors,Systems,and Analytics that captures current state of opera

39、tionsInformation Systems,Automation Systems,and Associates executing recommendationsOperations DataContextualized data and design constraintsActions,sequencing,and insightsSimulation results and operations impactsCaptured data and feedbackArchitecture1_DAIS_Title_SlideWhats Next w/Generative AILLM N

40、atural Language QuerySELECT machine_resource FROM machinesWHERE pm_date=2023-07-02 or pm_date=2023-07-03 or pm_date=2023-07-04 Which machines are scheduled for preventive maintenance this weekend?Machines MR2&MR5 are scheduled for preventive maintenance this weekend.Application stores conversation c

41、ontext for use in the next query to better interpret natural language.LLM extracts key entities the user may be discussing i.e.,intents.User expresses question in natural language.LLM enriches query response with original user prompts.Application may run this process iteratively.LLM creates a SQL qu

42、ery around these entities&sends to the data warehouse.Data warehouse runs query and returns results.“Which machines are scheduled for preventive maintenance this weekend?”MachinesScheduledPreventiveMaintenanceNextWeekendTermsCandidate Entities&IntentsMachine:MR1Machine:MR2Machine:MR3Preventive Maint

43、enance Date:2rd July,SatPreventive Maintenance Date:3rd July,SunPreventive Maintenance Date:4th July,MonLarge Language ModelData WarehouseSQLrunResults“Machines MR2&MR5 are scheduled for preventive maintenance this weekend.”enrichstoreLLM Natural Language AnalyticsPrediction ModelWhich machines are

44、scheduled for preventive maintenance this weekend?Machines MR2&MR5 are scheduled for preventive maintenance this weekend.“What will happen if I delay the preventive maintenance for MR2 by one week?”WillHappenDelayPreventiveMaintenanceMR2OneWeekTermsCandidate Entities&IntentsWill Happen:PredictionMR2

45、:MachineCurrent Maintenance Date:4th JulyNew Maintenance Date:11th JulyLarge Language ModelAI ApplicationModelrunResults“If you reschedule the preventive maintenance for machine MR2 by 1 week,the probability of failure increases from 0.7%to 14%.”enrichstoreWhat will happen if I delay the preventive

46、maintenance for MR2 by one week?If you reschedule the preventive maintenance for machine MR2 by 1 week,the probability of failure increases from 0.7%to 14%.LLM creates/calls predictive model around these entities&sends to AI application.AI application runs the predictive model on the data from the d

47、ata warehouse and returns results.Application stores conversation context for use in the next query to better interpret natural language.LLM extracts key entities the user may be discussing i.e.,intents.User expresses question in natural language.LLM enriches query response with original user prompt

48、s.Application may run this process iteratively.LLM w/Knowledge Graph for ContextLLM creates SparQL query around these entities&sends to the semantic layer.Semantic layer runs query and returns requested information from the source of truth.LLM curates recommendations based on the query response.Spar

49、QL QueryWhich machines are scheduled for preventive maintenance this weekend?Machines MR2&MR5 are scheduled for preventive maintenance this weekend.“How will the breakdown of MR2 impact my production floor and business?How can I mitigate the risks?”Breakdown MR2 Impact Production Floor Business How

50、CanMitigateRisksTermsCandidate Entities&IntentsBreakdown:EventMR2:MachineMR2:ObjectImpact Production Floor:SubjectImpact Business:SubjectHow Can Mitigate Risks:RecommendationLarge Language ModelSemantic LayerSparQLrunResults“Below are the impact,risks&their mitigation recommendations in”enrichstoreW

51、hat will happen if I delay the preventive maintenance for MR2 by one week?If you reschedule the preventive maintenance for machine MR2 by 1 week,the probability of failure increases from 0.7%to 14%.Application stores conversation context for use in the next query to better interpret natural language

52、.LLM extracts key entities the user may be discussing i.e.,intents.User expresses question in natural language.LLM enriches query response with original user prompts.Application may run this process iteratively.How will the breakdown of MR2 impact my production floor and business?How can I mitigate

53、the risks?Below are the impact,risks&their mitigation recommendations in case of breakdown of MR2:Production delay for product P3&process PR2 by 3 daysOrder 1 of customer 1 will be delayed by 2 daysAGV2 will remain idle for 3 daysStorage overflow in S3Recommendations:Inform customer C1 that order wi

54、ll be delayed by 2 daysRun early maintenance on AGV2Re-route 30%of S3 storage to S2Knowledge GraphIntelligent RecommendationscurateFactory of Future&Digital TwinThe digital twin is the virtual representation of a physical object or system across its life-cycle.It uses real-time data and other source

55、s to enable learning,reasoning,and dynamically recalibrating for improved decision making.This means creating a virtual representation that is the mirror counterpart(or twin)of a physical thing or process.The manufacturing twin is the cornerstone of“digital continuity”and the foundation of digital m

56、anufacturing.Site AssessmentEngineering analysis&edge designVendor CollaborationPre-config site outreachSite PreparationDesign Twin Plant(or Design Twin Plant(or Product)Product)Onsite deploymentTesting&sign-offincident&service managementMaintain,Service,UpgradeAdd Edge Use CasesManufacturing Twin(or Capital Build or ProductManufacturing Twin(or Capital Build or Product-view)view)Service TwinService TwinStagingImage engineeringThank You