翻譯部分英文原文：KNOWLEDGEDISCOVERYFROMFINITEELEMENTSIMULATIONDATAJI-LONGYIN,DA-YONGLI,YING-CIftTNWANG,YING-HONGPENGInstituteofKnowledge-basedEngineering,SchoolofMechanical,ShanghaiJiaotongUniversity,Shanghai,200030,ChinaE-MAIL:yinjilongsjtu,,,Abstract:Knowledge-basedengineering(KBE)andfiniteelementanalysis(FEA)havebeenusedwidelyinsheetmetalformingarea.However,theacquisitionofknowledgekeepsbottleneckwhenbuildingknowledgebaseinKBE.Also,toproperlyunderstandtheresultsoftheFEAandconsequentlychoosetheappropriatedesign,alotofknowledgeandexperienceareneeded.FEAcangeneratemassivedata,inwhichlargeamountsofusefullyimplicitknowledgeishidden.Thus,knowledgeacquisitionfromthemisprospectivetoeasetheabovedifficultiesbyapplyingKnowledgeDiscoveryinDatabases(KDD)technology.Inthisstudy,thecharacteristicsoftheFEAdataarediscussedfirstly.ThenaframeworkofknowledgediscoveryfromFEAdataisproposed.Correspondingly,adata-miningalgorithmnamedfuzzy-roughalgorithmisdevelopedtodealwiththeFEAsimulationdata.Finally,thestampingprocessofasquare-cuppartwasstudiedasanexample.Theproposedknowledgediscoveryprocessisappliedtoobtainsomeuseful,implicitproductionruleswithefficiencymeasure.TheresultshowsthatknowledgediscoveryfromFEAsimulationdataisvaluable.Keywords:Knowledgediscovery；NumericalSimulation；Fuzzyset；Roughset；Ruleinduction1.IntroductionNowadays,KBEiswidelyusedinengineeringarea,whichintegratesartificialintelligencewithCAXsystemandconnectsengineeringdesignwithCAXsystemwithoutinterruption1.Greatly,aKnowledge-BasedEngineeringSystems(KBES)performancedependsonthescaleoftheknowledgebaseitpossesses.Knowledgeacquisitionremainsasthemaindifficultandcrucialproblem.Manualacquisitionneedshardworkofknowledgeengineersanddomainexperts,togetherwiththetightcorporationbetweenthem.Thequalityofacquiredknowledgeisusuallypoor.Therefore,thereisanurgentneedfornewknowledgeacquisitiontechniquesandtoolstoextractusefulknowledgefromtherapidlygrowingvolumesofdata.KDDisthenon-trivialprocessofidentifying2valid,novel,potentiallyuseful,andultimatelyunderstandablepatternsindata.Itcanacquireimplicitandusefulknowledgeinlarge-scaledatasetsandhasmadegreatsuccessincommercialareas.Ithasexpandedtoengineeringdisciplines.TheoverallKDDprocessincludesdataselection,datapreprocessing,datatransformation,datamining,interpretationandevaluation,asshowninFigure1.Recently,numericalsimulationhasbecomethethirdmodeofsciencecomplementingtheoryandexperimentinalmostalloftheengineeringareas.FEAisthemostcommoncomputersimulationmethodinsheetmetalforminganalysis3.FEAsimulationsgeneratevastquantitiesofdata.TohelpthedesignersunderstandtheoutputofFEA,visualizationtechniquesareoftenusedtodisplaytheresults.However,thescaleoftheresultdataissolargethatvisualizationisfarfromsufficientresultdescription.Designershavetointerpretanalysisresultstodeterminewhetheradesignschemeisacceptable.Thisisalaboriousanderror-proneprocess,andrequiresasignificantamountofexperienceandexpertise.Ontheotherhand,themassiveresultdataimpliesmuchusefulknowledge,buttheyaresimplystoredawayondisksandneveranalyzedeffectively.SoextractingtheimplicitengineeringknowledgefromFEAresultsisverymeaningfulandurgent.Inthisstudy,thecharacteristicsoftheFEAdataarestudiedfirstly.ThenaframeworkforknowledgediscoveryfromFEAsimulationdataisproposed.Accordingtothecharacteristicsofthedata,afuzzy-roughalgorithmisdeveloped.Finally,toverifythevalidityoftheframeworkandthealgorithm,thestampingprocessofasquarecupisanalyzedandtheconclusionisgiven.2.FrameworkofKnowledgeDiscoveryfromFEASimulationData2.1.CharacteristicsofFEASimulationDataThoughitisthesuccessofKDDincommercialareathatinterestsusinknowledgediscoveryfromFEAdata4,5,thereismuchdifferencebetweenthem.Firstly,simulationdataareusuallystoredinaflatfileorspecialformatdatabase,whilebusinessdataareoftenstoredincommercialdatabase6.TheaccessibilityandqueryofdataismoredifficultforFEAsimulationdatafilethanforcommercialdatabase.ToaccessthedatafromvariousCAXsystems,aspecialinterfacetoolkitmustbeused.Secondly,mostbusinessdatabasescontainstructureddataconsistingofwell-definedfields.Eachvalueofthatattributeprovidesforthetargetlabel.However,FEAdataareintheformofmeshdatawithoutlabels.Valuesatameshpointarerealandcanbeelement-centered,node-centeredoredge-centered7.Obviously,theyaresemi-structuredorunstructured.Domainknowledgemustbeusedtoidentifythepatternfeature.Thirdly,unlikeinbusinessorproduction,thegenerationoftheFEAdatadoesnotrelyonexternaleventsandcanbecontrolledcompletely.Thusthedesignofexperiments(DOE)canbeappliedByDOEtechniques,fewersimulationdataisneededtoacquiremoreknowledge.Comparisonbetweensimulationsalsocanbemadetounderstandthedependenceofoutputdataonthedesignparameterspace.Therefore,amodifiedframeworkforknowledgediscoveryfromFEAsimulationdatamustbedevelopedandanappropriatedata-miningalgorithmmustbedesignedtofitthecharacteristicsofFEAdata.2.2.TheProposedFrameworkAccordingtothecharacteristicsofFEAdata,amodifiedknowledgediscoveryframeworkisproposedasshowninFigure2.Thetotalframeworkiscomposedoffourparts:productdesignanddevelopment,data-collection,knowledgediscovery,knowledgemanagementandreuse.Productdesignandprocessdevelopmentisthesequenceofactivitiestoturnopportunitiesandideasintosuccessfulproducts.Eachdesignwillbeexaminedbysimulationmethodorexperimentbeforeobtainingasuccessfulproduct.Tostudytherelationbetweenthedesignparametersandproductsperformance,DOEtechnologycanbeused.Intheiterativeprocessofproductdevelopment,largeamountofFEAsimulationdatarelatedtodesignparametersaregenerated.Thesedataareusuallystoredintoflatfilesorspecialformatdatabasesdispersedlyandcanbeusedasthedatasourceforknowledgediscovery.Duetothediversityofthedata,therefore,thesecondpartoftheframework,adatacollectorisusedtocollectthesedataandtransformsthemintoaunifieddatabase.ItshouldintegratevarioustoolstoexchangedataamongdifferentCAX(CAD/CAEKAM)softwareandknowledgediscoverysystem.Thethirdpartisknowledgediscovery,aniterativeprocessincludingfivebasicsteps:domainunderstanding,dataselectionandintegration,datapre-processing,ruleinduction,knowledgeevaluationandinterpretation.Indomainunderstandingstage,everydatasetsconnotativemeaningandthemechanismbywhichtheyinteractshouldbeknownclearly.Theselecteddatawillbeusedandanalyzedtogiveananswertotheproblemunderconsideration.ToimprovethequalityofthedataforDMalgorithm,datapre-processingmustbedone.Inruleinduction,intelligentmethodsareappliedinordertoextractdatapatterns.Productionrulesareselectedastheknowledgerepresentationforminthisstudyduetotheirmodularity,simplicityandexpandability.Thedataminingprocessmayberefinedandsomeofitsstepsbeiteratedseveraltimesbeforetheextractedknowledgecanbeused.Thefourthpartoftheframeisknowledgemanagementandreuse.Theminedknowledgeiscleanedupfirsttoeliminatetheredundancyandconflictsbeforestoringintoknowledgebase.Themindedknowledgecanbeappliedinthreeways.Firstly,itcanhelpdesignersunderstandsimulationresultclearly.Secondly,itcanbeusedasheuristicknowledgeinsearchingoptimaldesign.Thirdly,itcanbeusedasaknowledgeauto-acquisitiontooltohelpknowledgeengineersinbuildingknowledgebase.Theframeworkitselfisalsoaniterativeprocess.Minedknowledgecanbereused,verifiedandrefreshedinthenextdesignloops.NewFEAsimulationdataaregeneratedandcanbeappendedintodatabaseasdatasourcefornextknowledgediscovery.Thus,theknowledgebasewillbecomemoreefficientandeasiertobeused.3.Fuzzy-roughsetsalgorithmTherough-settheory(RST)proposedbyPawlakhasbeenusedwidelyinknowledgereasoningandknowledgeacquisition9.SincethebasicRSTalgorithmcanonlyhandlenominalfeatureindecisiontable,mostpreviousstudieshavejustshownhowbinaryorcrisptrainingdatamaybehandled10.ToapplyingtheRSTalgorithmonrealvaluedataset,discretizationoftenhastobeappliedasthepreprocessingsteptotransformthemintonominalfeaturespace11.Inthisstudy,animprovedalgorithmnamedfuzzy-roughsetsalgorithmisdevelopedbyintegra