`reetal./SignalProcessing88(2008)468–491G.Merce
469
5.
6.
Adaptationofthepropagatormethodtoidenti cation...................................474
4.1.1.Problemformulation......................................................4744.1.2.DiscussiononthematrixSconstruction........................................4744.1.3.Minimizedcriterioninthenoisefreecase.......................................4754.1.4.Noiseeffects............................................................4764.2.EstimationfromtherecursivePI/POMOESPschemeupdate..............................477
4.2.1.RPM2algorithm........................................................4774.2.2.COPMalgorithm........................................................4784.3.EstimationfromtheordinaryMOESPschemeupdate...................................479
4.3.1.COIVPM,IVPMandEIVPMalgorithms.......................................4794.3.2.Improvementoftheextendedvariablemethod:theEIVsqrtPMalgorithm................4804.4.Summaryandcomments........................................................482Simulationexamples...............................................................4855.1.MIMOtime-invariantandtime-varyingsystems.......................................485
5.1.1.Time-invariantcase.......................................................4855.1.2.Slowlytime-varyingcase...................................................4865.1.3.Time-varyingwithabruptchangecase.........................................4875.2.In uenceoftheSmatrix........................................................488Conclusion......................................................................489References......................................................................489
4.1.
1.Introduction
Subspace-basedapproachforsystemidenti ca-tionisaparticular eldofexperimentalmodellingwhichhasreachedade nitematurityfromnowon,asshownbynumerousreferencearticles[1–5]andvariousapplicationpapers[6–10].Of inesubspaceidenti cationmethodsareindeedattractivesinceastate-spacerealizationcanbedirectlyestimatedfrominput/output(I/O)datawithoutnonlinearoptimization(generallyrequiredbythepredictionmethods[11]).Furthermore,thesetechniquesarecharacterizedbytheuseofrobustnumericaltoolssuchastheRQfactorizationandthesingularvaluesdecomposition(SVD).Interestingfromanumericalpointofview,thebatchsubspacemodelidenti ca-tion(SMI)algorithmsarenotusableforonlineimplementationbecauseoftheSVDcomputationalcomplexity.Indeed,inmanyonlineidenti cationscenarios,itisimportanttoupdatethemodelastimegoesonwithareducedcomputationalcost.Consequently,itwasnecessaryto ndSVDalter-nativealgorithmsinordertoapplythesubspaceconceptinarecursiveframework.First,someworksproposedadaptationsofSMImethodsinordertoupdatetheSVD[12,13].Unfortunately,thesetechniqueshadthedrawbackofrequiringthatthedisturbancesactingonthesystemoutputswerespatiallyandtemporallywhite,whichisobviouslyrestrictiveinpractice.Then,recursivesubspacemodelidenti cation(RSMI)methodsbasedontheYang’scriterionandtheprojectionapproximationsubspacetracking(PAST)costfunction[14]wereintroduced.Thesespeci ctechniquesweredevel-opedtotrackthesubspacespannedbytheextendedobservabilitymatrixinacoloureddisturbancesframework[15–18].Moreprecisely,instrumentalvariableadaptationsofthePASTtechniquetotheproblemofRSMIwereconsidered,thefocusbeingcomputationalef ciency[15]orestimationaccuracy[16,17].Morerecently,inordertoprovidesolutionstotheapproximationinducedbythereductionofthefourthorderYang’scriteriontoaquadraticcostfunction,newdevelopmentsintheRSMIclassofalgorithmshavebeenputforward[19–21].Theproposedmethodsarebasedontheadaptationofaparticulararraysignalprocessingtechniquetotherecursivesubspaceidenti cationproblem:thepro-pagator[22].Themainadvantageofthisapproachoverthepreviousconceptionliesintheuseofalinearoperatorandquadraticcriteriawhichleadtorecursiveleastsquaresimplementationsforthealgorithms.Thischaracteristichasnewlyallowedtheanalysisoftheconvergencepropertiesofthedevelopedtechniquesfortherecursiveupdateofthesubspaceestimates[23].Ithasalsobeenusedwithsuccesstoonlinetrackthemodalparametersofairplanesduringtest ights[24].
热门文档