adaptive filterWord下载.docx
《adaptive filterWord下载.docx》由会员分享,可在线阅读,更多相关《adaptive filterWord下载.docx(9页珍藏版)》请在冰点文库上搜索。
2.Basicconfigurationofanadaptivefilter
Thebasicconfigurationofanadaptivefilter,operatinginthediscrete-timedomaink,isillustratedinFigure2.1.Insuchascheme,theinputsignalisdenotedbyx(k),thereferencesignald(k)representsthedesiredoutputsignal(thatusuallyincludessomenoisecomponent),y(k)istheoutputoftheadaptivefilter,andtheerrorsignalisdefinedase(k)=d(k).y(k).
Fig.2.1Basicblockdiagramofanadaptivefilter.
Theerrorsignalisusedbytheadaptationalgorithmtoupdatetheadaptivefiltercoefficientvectorw(k)accordingtosomeperformancecriterion.Ingeneral,thewholeadaptationprocessaimsatminimizingsomemetricoftheerrorsignal,forcingtheadaptivefilteroutputsignaltoapproximatethereferencesignalinastatisticalsense.
Fig.2.2Channelequalizationconfigurationofanadaptivefilter:
Theoutputsignaly(k)estimatesthetransmittedsignals(k).
Fig.2.3Predictorconfigurationofanadaptivefilter:
Theoutputsignaly(k)estimatesthepresent
inputsamples(k)basedonpastvaluesofthissamesignal.Therefore,whentheadaptivefilteroutputy(k)approximatesthereference,theadaptivefilteroperatesasapredictorsystem.
3.Adaptationalgorithm
Severaloptimizationprocedurescanbeemployedtoadjustthefiltercoefficients,including,forinstance,theleastmean-square(LMS)anditsnormalizedversion,thedata-reusing(DR)includingtheaffineprojection(AP),andtherecursiveleast-squares(RLS)algorithms.AlltheseschemesarediscussedinSection2.3,emphasizingtheirmainconvergenceandimplementationcharacteristics.TheremainingofthebookfocusesontheRLSalgorithms,particularly,thoseemployingQRdecomposition,whichachieveexcellentoverallconvergenceperformance.
3.1ErrorMeasurements
Adaptationofthefiltercoefficientsfollowsaminimizationprocedureofaparticularobjectiveorcostfunction.Thisfunctioniscommonlydefinedasanormoftheerrorsignale(k).Thethreemostcommonlyemployednormsarethemean-squareerror(MSE),theinstantaneoussquareerror(ISE),andtheweightedleast-squares(WLS),whichareintroducedbelow.
3.2Themean-squareerror
TheMSEisdefinedas
⎩(k)=E[e2(k)]=E[|d(k)−y(k)|2].
WhereRandparetheinput-signalcorrelationmatrixandthecross-correlationvectorbetweenthereferencesignalandtheinputsignal,respectively,andaredefinedas
R=E[x(k)xT(k)],
p=E[d(k)xT(k)].
Note,fromtheaboveequations,thatRandparenotrepresentedasafunctionoftheiterationkornottime-varying,duetotheassumedstationarityoftheinputandreferencesignals.
FromEquation(2.5),thegradientvectoroftheMSEfunctionwithrespecttotheadaptivefiltercoefficientvectorisgivenby
Theso-calledWienersolutionwo,thatminimizestheMSEcostfunction,isobtainedbyequatingthegradientvectorinEquation(2.8)tozero.AssumingthatRisnon-singular,onegetsthat
3.3Theinstantaneoussquareerror
TheMSEisacostfunctionthatrequiresknowledgeoftheerrorfunctione(k)atalltimek.Forthatpurpose,theMSEcannotbedeterminedpreciselyinpracticeandiscommonlyapproximatedbyothercostfunctions.ThesimplerformtoestimatetheMSEfunctionistoworkwiththeISEgivenby
Inthiscase,theassociatedgradientvectorwithrespecttothecoefficientvectorisdeterminedas
ThisvectorcanbeseenasanoisyestimateoftheMSEgradientvectordefinedinEquation(2.8)orasaprecisegradientoftheISEfunction,which,initsownturn,isanoisyestimateoftheMSEcostfunctionseeninSection2.2.1.
3.4Theweightedleast-squares
AnotherobjectivefunctionistheWLSfunctiongivenby
where0_⎣<
1istheso-calledforgettingfactor.Theparameter⎣k−iemphasizesthemostrecenterrorsamples(wherei≈k)inthecompositionofthedeterministiccostfunction⎩D(k),givingtothisfunctiontheabilityofmodelingnon-stationaryprocesses.Inaddition,sincetheWLSfunctionisbasedonseveralerrorsamples,itsstochasticnaturereducesintime,beingsignificantlysmallerthanthenoisyISEnatureaskincreases.
2.3AdaptationAlgorithms
Inthissection,anumberofschemesarepresentedtofindtheoptimalfiltersolutionfortheerrorfunctionsseeninSection2.2.Eachschemeconstitutesanadaptationalgorithmthatadjuststheadaptivefiltercoefficientsinordertominimizetheassociatederrornorm.
Thealgorithmsseenherecanbegroupedintothreefamilies,namelytheLMS,theDR,andtheRLSclassesofalgorithms.Eachgrouppresentsparticularcharacteristicsofcomputationalcomplexityandspeedofconvergence,whichtendtodeterminethebestpossiblesolutiontoanapplicationathand.
2.3.1LMSandnormalized-LMSalgorithms
DeterminingtheWienersolutionfortheMSEproblemrequiresinversionofmatrixR,whichmakesEquation(2.9)hardtoimplementinrealtime.OnecanthenestimatetheWienersolution,inacomputationallyefficientmanner,iterativelyadjustingthecoefficientvectorwateachtimeinstantk,insuchamannerthattheresultingsequencew(k)convergestothedesiredwosolution,possiblyinasufficientlysmallnumberofiterations.
TheLMSalgorithmissummarizedinTable2.1,wherethesuperscripts.andHdenotethecomplex-conjugateandtheHermitianoperations,respectively.
TheLMSalgorithmisverypopularandhasbeenwidelyusedduetoitsextremesimplicity.Itsconvergencespeed,however,ishighlydependentontheconditionnumberpoftheinput-signalautocorrelationmatrix[1–3],definedastheratiobetweenthemaximumandminimumEigenvaluesofthismatrix.
IntheNLMSalgorithm,whenυ=0,onehasw(k)=w(k−1)andtheupdatinghalts.Whenυ=1,thefastestconvergenceisattainedatthepriceofahighermisadjustmentthentheoneobtainedfor0<
υ<
1.
2.3.2Data-reusingLMSalgorithms
Asremarkedbefore,theLMSalgorithmestimatestheMSEfunctionwiththecurrentISEvalue,yieldinganoisyadaptationprocess.Inthisalgorithm,informationfromeachtimesamplekisdisregardedinfuturecoefficientupdates.DRalgorithms[9–11]employpresentandpastsamplesofthereferenceandinputsignalstoimproveconvergencecharacteristicsoftheoveralladaptationprocess.
Asageneralizationofthepreviousidea,theAPalgorithm[13–15]isamongtheprominentadaptationalgorithmsthatallowtrade-offbetweenfastconvergenceandlowcomputationalcomplexity.Byadjustingthenumberofprojections,oralternatively,thenumberofdatareuses,oneobtainsadaptationprocessesrangingfromthatoftheNLMSalgorithmtothatofthesliding-windowRLSalgorithm[16,17].
2.3.3RLS-typealgorithms
ThissubsectionpresentsthebasicversionsoftheRLSfamilyofadaptivealgorithms.Importanceoftheexpressionspresentedherecannotbeoverstatedfortheyallowaneasyandsmoothreadingoftheforthcomingchapters.
TheRLS-typealgorithmshaveahighconvergencespeedwhichisindependentoftheEigenvaluespreadoftheinputcorrelationmatrix.Thesealgorithmsarealsoveryusefulinapplicationswheretheenvironmentisslowlyvarying.
ThepriceofallthesebenefitsisaconsiderableincreaseinthecomputationalcomplexityofthealgorithmsbelongingtotheRLSfamily.
ThemainadvantagesassociatedtotheQR-decompositionRLS(QRD-RLS)algorithms,asopposedtotheirconventionalRLScounterpart,arethepossibilityofimplementationinsystolicarraysandtheimprovednumericalbehaviorinlimitedprecisionenvironment.
2.5Conclusion
Itwasverifiedhowadaptivealgorithmsareemployedtoadjustthecoefficientsofadigitalfiltertoachieveadesiredtime-varyingperformanceinseveralpracticalsituations.Emphasiswasgivenonthedescriptionofseveraladaptationalgorithms.Inparticular,theLMSandtheNLMSalgorithmswereseenasiterativeschemesforoptimizingtheISE,aninstantaneousapproximationoftheMSEobjectivefunction.Data-reusealgorithmsintroducedtheconceptofutilizingdatafrompasttimesamples,resultinginafasterconvergenceoftheadaptiveprocess.Finally,theRLSfamilyofalgorithms,basedontheWLSfunction,wasseenastheepitomeoffastadaptationalgorithms,whichuseallavailablesignalsamplestoperformtheadaptationprocess.Ingeneral,RLSalgorithmsareusedwheneverfastconvergenceisnecessary,forinputsignalswithahighEigenvaluespread,andwhentheincreaseinthecomputationalloadistolerable.AdetaileddiscussionontheRLSfamilyofalgorithmsbasedontheQRdecomposition,whichalsoguaranteesgoodnumericalpropertiesinfinite-precisionimplementations,constitutesthemaingoalsofthisbook.Practicalexamplesofadaptivesystemidentificationandchannelequalizationwerepresented,allowingonetovisualizeconvergenceproperties,suchasmisadjustment,speed,andstability,ofseveraldistinctalgorithmsdiscussedpreviously.