资源描述:
《Nonlinear Principal Component Analysis Using Autoassociative Neural Networks》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库。
1、NonlinearPrincipalComponentAnalysisUsingAutoassociativeNeuralNetworksMarkA.KramerLaboratoryforIntelligentSystemsinProcessEngineering,Dept.ofChemicalEngineering,MassachusettsInstituteofTechnology,Cambridge,MA02139Nonlinearprincipalcomponentanalysisisanoveltechniqueformultivariatedataanal
2、ysis,similartothewell-knownmethodofprincipalcomponentanalysis.NLPCA,likePCA,isusedtoidentifyandremovecorrelationsamongproblemvariablesasanaidtodimensionalityreduction,visualization,andexploratorydataanalysis.WhilePCAidentifiesonlylinearcorrelationsbetweenvariables,NLPCAuncoversbothlinea
3、randnonlinearcorrelations,withoutrestrictiononthecharacterofthenonlinearitiespresentinthedata.NLPCAoperatesbytrainingafeedforwardneuralnetworktoperformtheidentitymapping,wherethenetworkinputsarereproducedattheoutputlayer.Thenetworkcontainsaninternal“bottleneck”layer(containingfewernodes
4、thaninputoroutputlayers),whichforcesthenetworktodevelopacompactrepresentationoftheinputdata,andtwoadditionalhiddenlayers.TheNLPCAmethodisdemonstratedusingtime-dependent,simulatedbatchreactiondata.ResultsshowthatNLPCAsuccessfullyreducesdimensionalityandproducesafeaturespacemapresemblingt
5、heactualdistributionoftheunderlyingsystemparameters.introductionEngineersareoftenconfrontedwiththeproblemofextract-theidentificationtask(sincealeakisonlypossiblewhentheinginformationaboutpoorly-knownprocessesfromdata.Dis-netflowisnonzero)andreducesthedimensionalityofthecerningthesignifi
6、cantpatternsindata,asafirststeptoprocessrecognitionproblembysubstitutingasingleindicatorfortwounderstanding,canbegreatlyfacilitatedbyreducingdimen-ormoreindividualmeasurements.Ingeneral,selectionoffea-sionality.Thesuperficialdimensionalityofdata,orthenumberturessuchasthenetflowratedepen
7、dsontheultimateappli-ofindividualobservationsconstitutingonemeasurementvec-cation.However,incaseswheretheultimateapplicationoftor,isoftenmuchgreaterthantheintrinsicdimensionality,thedataisnotknowninadvance,asuitableobjectiveforthenumberofindependentvariablesunderlyingthesignifi