过程监控 利用tlpp
JournalofProcessControl19(2009)

288–296
ContentslistsavailableatScienceDirect
JournalofProcessControl
journalhomepage:doc.guandang.net

t
Batchprocessmonitoringwithtensorfactorization
KunlunHu,JingqiYuan*
DepartmentofAutomation,ShanghaiJiaoTongUniversity,800DongchuanRoad,200240Shanghai,PRChina
articleinfoabstract
Abatchprocessmonitoringmethodusingtensorfactorization,tensorlocalitypreservingprojections(TLPP),isproposed.Inmanyexistingvector-basedmethodsonbatchprocessmonitoringsuchasMPCAandMLPP,abatchdataisrepresentedasavectorinhigh-dimensionalspace.Butvectorizingbatchdatawillleadtoinformationloss.Essentially,abatchdataispresentedasasecondordertensor,oramatrix.Inthiscase,tensorfactorizationmaybeusedtodealwiththetwo-waybatchdatamatrixdirectlyinsteadofperformingvectorizingprocedure.Furthermore,tensorrepresentationhassomeadvantagessuchaslowmemoryandstoragerequirementsandlessestimatedparametersfornormaloperatingcondition(NOC)model.Ontheotherhand,differentfromprincipalcomponentanalysis(PCA)whichaimsatpreservingtheglobalEuclideanstructureofthedata,theTLPPaimstopreservethelocalneighborhoodinformationandtodetecttheintrinsicmanifoldstructureofthedata.Consequently,TLPPmaybeusedto ndmoremeaningfulintrinsicinformationhiddenintheobservations.TheeffectivenessandadvantagesoftheTLPPmonitoringapproacharetestedwiththedatafromabenchmarkfed-batchpenicillinfermentationandtwoindustrialfermentationprocesses,penicillinandcephalosporin,respectively.
Ó2008ElsevierLtd.Allrightsreserved.
Articlehistory:
Received26November2007
Receivedinrevisedform18March2008Accepted31March2008
Keywords:
Tensorfactorization
TensorlocalitypreservingprojectionsBatchprocessmonitoringFaultdetectionanddiagnosisDynamicmonitoringMovingdatawindow
1.Introduction
Batchprocesseshavebeenutilizedtoproducehighvalue-addedmaterialsinpharmaceutical,biochemistry,foodandsemiconduc-torindustries.Processmonitoringandfaultdiagnosisofsuchpro-cessesareextremelyimportanttoensurehighqualityproductsandplantsafety.
NomikosandMacGregor rstlydevelopedthebatchprocessmonitoringapproachesusingmultiwayprincipalcomponentanal-ysis(MPCA)[1,2]andmultiwaypartialleastsquares[3].Subse-quently,severaltechniquesusingmultivariatestatisticalanalysishavebeenpresentedforthemonitoringandfaultdiagnosisofbatchprocesses.LouwerseandSmilde[4]usedparallelfactoranal-ysisandTucker3modelsforbatchprocessmonitoring.Recently,themethodsbasedonindependentcomponentanalysis[5–7]havebeenproposedtodecomposetheobserveddataintolinearcombi-nationsofstatisticallyindependentcomponentsandtodealwithhigher-orderstatistics.Besides,Luetal.[8]presentedatwo-dimensionaldynamicprincipalcomponentanalysistomodelandmonitorsimultaneouslythetime-andbatch-wisedynamicsforbatchprocesses.CamachoandPicó[9,10]usedmulti-phaseprinci-palcomponentanalysisforonlinemonitoringofbatchprocesses.Zhaoetal.[11]proposedamethodbasedondissimilarityanalysisforbatchprocessmonitoringusingmovingdatawindowtech-
*Correspondingauthor.Tel./fax:+862134204055.E-mailaddress:jqyuan@sjtu.edu.cn(J.Yuan).
0959-1524/$-seefrontmatterÓ2008ElsevierLtd.Allrightsreserved.doi:10.1016/j.jprocont.2008.03.003
nique.HuandYuan[12]presentedabatchprocessmonitoringbasedonmultiwaylocalitypreservingprojections(MLPP)inordertoextracttheintrinsicgeometrystructureoftheobserveddataand ndingmoremeaningfullow-dimensionalinformationhiddeninthehigh-dimensionalobservations.In[13],theperformanceofglo-bal,localandtimeevolvingmodelsforbatchprocessmonitoringhavebeenevaluatedusingtwoperformanceindices:theoveralltypeIerrorandtherelativeactionsignaltime.
Inthispaper,abatchprocessmonitoringapproachwithtensorlocalitypreservingprojections(TLPP)isproposed.Manyofprevi-ousmultivariatestatisticalanalysistechniquestreatabatchdataasavector,whilethetensorlocalitypreservingprojections[14]treatsabatchdataasasecondordertensor(ormatrix).Mostoftheexistingmonitoringmethodsbasedontraditionaltechniquesneedtounfoldthree-waybatchdatasettotwo-wayarraybeforedevelopingNOCmodel,whilethemonitoringmethodbasedonTLPPiscapableofdealingwiththree-waybatchdatasetdirectly.Ontheotherhand,differentfromPCAwhichaimsatpreservingtheglobalEuclideanstructureofthedata,TLPPpreservesthelocalneighborhoodstructureofthedataset[14,15].Hence,theTLPPisabletoextractmoremeaningfulinformationfromthebatchdataset.
Theorganizationofthepaperisasfollows.MPCAandMLPPmonitoringaredescribed rstly,followedbytheintroductiontothetensorlocalitypreservingprojections.Then,themonitoringap-proachwithTLPPispresentedinmoredetail.Finally,themonitor-ingresultsofTLPP,MLPPandMPCAarepresentedandcomparedwiththedatafromabenchmarkfed-batchpenicillinfermentation
Batch process monitoring... 9页 2财富值 Programming computer vis... 44页...Non-negative tensor factorization with applications to statistics and computer ...
Gaussian Process and Kernelization of Tensor-Based Model for Multiway Data Analysis...TensorFactorization/Decomposition ? GaussianProcesses ? ProbabilisticKernel...
Batch process monitoring... 9页 2财富值 Pion Form Factor in the ... 暂无...Transition Form Factor with Tensor Current within the kT Factorization Approac...
Batch process monitoring... 9页 2财富值 Pion Form Factor in the ... 暂无...Transition Form Factor with Tensor Current within the $k_T$ Factorization ...
C. Controlling sparseness in nonnegative tensor factorization Abstract. Non-negative tensor factorization (NTF) has recently been proposed as sparse and efficien...
Approximate L0 constrained Non-negative Matrix and Tensor Factorization Abstract — Non-negative matrix factorization (NMF), i.e. V ≈ WH where both V,...
[2009]Blind multispectral image decomposition by 3D nonnegative tensor factorization [2009]Blind multispectral image decomposition by 3D nonnegative tensor factoriz...
[2009]Blind multispectral image decomposition by 3D nonnegative tensor factorization [2009]Blind multispectral image decomposition by 3D nonnegative tensor factoriz...
热门文档