外文翻譯--內(nèi)嵌于可編程控制器的先進(jìn)控制算法 英文文版.pdf

ControlEngineeringPracticeembeddedVrancLjubljana,bNovaGoricaPolytechnic,NovaGorica,Sloveniaidenticationstepstoprovidereliableoperation.Thecontrollermonitorsandevaluatesthecontrolperformanceoftheclosed-loopsystem.Thecontrollerwasimplementedonaprogrammablelogiccontroller(PLC).Theperformanceisillustratedonaeldtestinindustrialapplications,assummarisedbelow:ARTICLEINPRESS/locate/conengpracC3Correspondingauthor.Tel.:+38614773994；1.Becauseofthediversityofreal-lifeproblems,asinglenonlinearcontrolmethodhasarelativelynarrow0967-0661/$-seefrontmatterr2005ElsevierLtd.Allrightsreserved.doi:10.1016/j.conengprac.2005.05.006fax:+38614257009.E-mailaddress:samo.gerksicijs.si(S.Gerksic).applicationforcontrolofpressureonahydraulicvalve.r2005ElsevierLtd.Allrightsreserved.Keywords:Controlengineering；Fuzzymodelling；Industrialcontrol；Model-basedcontrol；Nonlinearcontrol；Programmablelogiccontrollers；Self-tuningregulators1.IntroductionModerncontroltheoryoffersmanycontrolmethodstoachievemoreefcientcontrolofnonlinearprocessesthanprovidedbyconventionallinearmethods,takingadvantageofmoreaccurateprocessmodels(Bequette,1991；Henson&Seborg,1997；Murray-Smith&Johansen,1997).Surveys(Takatsu,Itoh,&Araki,1998；Seborg,1999)indicatethatwhilethereisaconsiderableandgrowingmarketforadvancedcon-trollers,relativelyfewvendorsofferturn-keyproducts.Excellentresultsofadvancedcontrolconcepts,basedonfuzzyparameterscheduling(Tan,Hang,&Chai,1997；Babuska,Oosterhoff,Oudshoorn,&Bruijn,2002),multiple-modelcontrol(Dougherty&Cooper,2003；Gundala,Hoo,&Piovoso,2000),andadaptivecontrol(Henson&Seborg,1994；Hagglund&Astrom,2000),havebeenreportedintheliterature.However,thereareseveralrestrictionsforapplyingthesemethodsdINEAd.o.o.,Ljubljana,SloveniaeComputerTechnologyInstitute,Athens,GreecefUniversityofChemicalTechnologyandMetallurgySofia,Sofia,BulgariaReceived23April2004；accepted15May2005AbstractThispaperpresentsaninnovativeself-tuningnonlinearcontrollerASPECT(advancedcontrolalgorithmsforprogrammablelogiccontrollers).Itisintendedforthecontrolofhighlynonlinearprocesseswhosepropertieschangeradicallyoveritsrangeofoperation,andincludesthreeadvancedcontrolalgorithms.Itisdesignedusingtheconceptsofagent-basedsystems,appliedwiththeaimofautomatingsomeofthecongurationtasks.Theprocessisrepresentedbyasetoflow-orderlocallinearmodelswhoseparametersareidentiedusinganonlinelearningprocedure.Thisprocedurecombinesmodelidenticationwithpre-andpost-cUniversityofLjubljana,FacultyofElectricalEngineering,Ljubljana,SloveniaAdvancedcontrolalgorithmslogiccontrollerSamoGerksica,C3,GregorDolanca,DamirSasoBlazicc,IgorSkrjancc,ZoranMarinsRobertKinge,MinchoHadjiskiaJozefStefanInstitute,inaprogrammableica,JusKocijana,b,StankoStrmcnika,ekd,MihaBozicekd,AnnaStathakie,f,KostaBoshnakovfSlovenia()friendmaticindustfromling,procedcontrollermonitorstheresultingcontrolperformanceARTICLEINPRESSanonlinearprocessmodel.Themodelisobtainedoperatingprocesssignalsbyexperimentalmodel-usinganovelonlinelearningprocedure.ThisThefromforimplementationonPLCoropencontrollerrialhardwareplatforms.controllerparametersareautomaticallytunedfeatuadaptedssioningofthecontrollerissimpliedbyauto-experimentationandtuning.AdistinguishingreofthecontrolleristhatthealgorithmsaremetecommiTheASPECTcontrollerisanefcientanduser-lyengineeringtoolforimplementationofpara-r-schedulingcontrolintheprocessindustry.Theused,thesensorreadings,specichardwareplatformsareetc.isdemandedtoeldofapplication.Therefore,moreexiblemethodsoratoolboxofmethodsarerequiredinindustry.2.Newmethodsareusuallynotavailableinaready-to-useindustrialform.Customdesignrequiresconsider-ableeffort,timeandmoney.3.Thehardwarerequirementsarerelativelyhigh,duetothecomplexityofimplementationandcomputationaldemands.4.Thecomplexityoftuning(Babuskaetal.,2002)andmaintenancemakesthemethodsunattractivetononspecialisedengineers.5.Thereliabilityofnonlinearmodellingisofteninquestion.6.Manynonlinearprocessescanbecontrolledusingthewell-knownandindustriallyprovenPIDcontroller.Aconsiderabledirectperformanceincrease(nancialgain)isdemandedwhenreplacingaconventionalcontrolsystemwithanadvancedone.Themain-tenancecostsofaninadequateconventionalcontrolsolutionmaybelessobvious.Theaimofthisworkistodesignanadvancedcontrollerthataddressessomeoftheaforementionedproblemsbyusingtheconceptsofagent-basedsystems(ABS)(Wooldridge&Jennings,1995).Themainpurposeistosimplifycontrollercongurationbypartialautomationofthecommissioningprocedure,whichistypicallyperformedbythecontrolengineer.ABSsolvedifcultproblemsbyassigningtaskstonetworkedsoftwareagents.Thesoftwareagentsarecharacterisedbypropertiessuchasautonomy(operationwithoutdirectinterventionofhumans),socialability(interactionwithotheragents),reactivity(perceptionandresponsetotheenvironment),pro-activeness(goal-directedbe-haviour,takingtheinitiative),etc.ThisworkdoesnotaddressissuesofABStheory,butrathertheapplicationofthebasicconceptsofABStotheeldofprocesssystemsengineering.Inthiscontext,anumberoflimitshavetobeconsidered.Forexample:initiativeisrestricted,ahighdegreeofreliabilityandpredictability,insightintotheproblemdomainislimitedS.Gerksicetal./ControlEngineerin2ureisbasedonmodelidenticationusingtheandreactstodetectedirregularities.Thecontrollercomprisestherun-timemodule(RTM)andthecongurationtool(CT).TheRTMrunsonaPLC,performingallthemainfunctionalityofreal-timecontrol,onlinelearningandcontrolperformancemonitoring.TheCT,usedonapersonalcomputer(PC)duringtheinitialcongurationphase,simpliesthecongurationprocedurebyprovidingguidanceanddefaultparametervalues.Theoutlineofthepaperisasfollows:Section2presentsanoverviewoftheRTMstructureanddescribesitsmostimportantmodules；Section3givesabriefdescriptionoftheCT；andnally,Section4describestheapplicationofthecontrollertoapilotplantwhereitisusedforcontrolofthepressuredifferenceonahydraulicvalveinavalvetestapparatus.2.Run-TimeModuleTheRTMoftheASPECTcontrollercomprisesasetofmodules,linkedintheformofamulti-agentsystem.Fig.1showsanoverviewoftheRTManditsmainmodules:thesignalpre-processingagent(SPA),theonlinelearningagent(OLA),themodelinformationagent(MIA),thecontrolalgorithmagent(CAA),thecontrolperformancemonitor(CPM),andtheoperationsupervisor(OS).2.1.Multi-facetedmodel(MFM)TheASPECTcontrollerisbasedonthemulti-facetedmodelconceptproposedbyStephanopoulus,Henning,andLeone(1990)andincorporatesseveralmodelformsrequiredbytheCAAandtheOLA.Specically,theMFMincludesasetoflocalrst-andsecond-orderlocallearningapproach(Murray-Smith&Johansen,1997,p.188).Themeasurementdataareprocessedbatch-wise.Additionalstepsareperformedbeforeandafteridenticationinordertoimprovethereliabilityofmodelling,comparedtoadaptivemethodsthatuserecursiveidenticationcontinuously(Hagglund&As-trom,2000).Thenonlinearmodelcomprisesasetoflocallow-orderlinearmodels,eachofwhichisvalidoveraspeciedoperatingregion.Theactivelocalmodel(s)isselectedusingaconguredschedulingvariable.Thecontrollerisspecicallydesignedforsingle-input,single-outputprocessesthatmayincludeameasureddis-turbanceusedforfeed-forward.Additionally,theapplicationrangeofthecontrollerdependsontheselectedcontrolalgorithm.Amodularstructureofthecontrollerpermitsuseofarangeofcontrolalgorithmsthataremostsuitablefordifferentprocesses.ThegPractice()discrete-timelinearmodelswithtimedelayandoffset,ARTICLEINPRESSS.Gerksicetal./ControlEngineerinwhicharespeciedbyagivenschedulingvariables(k).Themodelequationofrstorderlocalmodelsisyk1C0a1；jykb1；jukC0dujc1；jvkC0dvjrj,(1)whilethemodelequationofsecondordermodelsisyk1C0a1；jykC0a2；jykC01b1；jukC0dujb2；jukC01C0dujc1；jvkC0dvjc2；jvkC01C0dvjrj,2wherekisthediscretetimeindex,jisthenumberofthelocalmodel,y(k)istheprocessoutputsignal,u(k)istheprocessinputsignal,v(k)istheoptionalmeasureddisturbancesignal(MD),duisthedelayinthemodelbranchfromutoy,dvisthedelayinthemodelbranchfromvtoy,andai,j,bi,j,ci,jandrjaretheparametersofthejthlocalmodel.ThesetoflocalmodelscanbeinterpretedasaTakagiSugenofuzzymodel,whichinthecaseofasecondordermodelcanbeexpressedintheFig.1.Run-timemodulegPractice()3followingform:yk1C0Xmj1bjka1；jykC0Xmj1bjka2；jykC01Xmj1bjkb1；jukC0dujXmj1bjkb2；jukC01C0dujXmj1bjkc1；jnkC0dnjXmj1bjkc2；jnkC01C0dnjXmj1bjkrj,3wherebj(k)isthevalueofthemembershipfunctionofthejthlocalmodelatthecurrentvalueoftheschedulingvariables(k).Normalisedtriangularmembershipfunc-tionsareused,asillustratedinFig.2.overview.ARTICLEINPRESSTheschedulingvariables(k)iscalculatedusingcoefcientskr,ky,ku,andkv,usingtheweightedsumskkrrkkyykkuukC01kvvk.(4)Thecoefcientsareconguredbytheengineeraccord-ingtothenatureoftheprocessnonlinearity.2.2.OnlineLearningAgent(OLA)TheOLAscansthebufferofrecentreal-timesignals,preparedbytheSPA,andestimatestheparametersofthelocalmodelsthatareexcitedbythesignals.ThemostrecentlyderivedparametersaresubmittedtotheMIAonlywhentheypassthevericationtestandareprovedtobebetterthantheexistingset.TheOLAisinvokedupondemandfromtheOSorautonomously,whenanintervaloftheprocesssignalswithsufcientexcitationisavailableforprocessing.Itprocessesthesignalsbatch-wiseandusingthelocallearningapproach.Anadvantageofthebatch-wiseconceptisthatthedecisiononwhethertoadaptthemodelisnotperformedinreal-timebutfollowingadelaythatallowsforinspectionoftheidenticationresultbeforeitisapplied.Thus,bettermeansforcontroloverdataselectionisprovided.Aproblemofdistributionofthecomputationtimerequiredforidenticationappearswithbatch-wiseprocessingofdata(opposedtotheonlinerecursiveprocessingthatistypicallyusedinadaptivecontrollers).Thisproblemisresolvedusingamulti-taskingoperationsystem.SincetheOLAtypicallyrequiresconsiderablyFig.2.FuzzymembershipfunctionsoflocalmodelsintheMFM.S.Gerksicetal./ControlEngineerin4morecomputationthanthereal-timecontrolalgorithm,itrunsinthebackgroundasalow-prioritytask.Thefollowingprocedure,illustratedinFig.3,isexecutedwhentheOLAisinvoked.2.2.1.CopysignalbufferThebufferofthereal-timesignalsismaintainedbytheSPA.WhentheOLAisinvoked,therelevantsectionofthebufferiscopiedforfurtherprocessing.2.2.2.ExcitationcheckAquickexcitationcheckisperformedatthestart,sothatprocessingofthesignalsisperformedonlywhentheycontainexcitation.Ifthestandarddeviationsofthesignalsr(k),y(k),u(k),andv(k)intheactivebufferarebelowtheirthresholds,theexecutioniscancelled.2.2.3.CopyactiveMFMfromMIATheonlinelearningprocedurealwayscomparesthenewlyidentiedlocalmodelswiththeprevioussetofparameters.Therefore,theactiveMFMiscopiedfromtheMIAwhereitisstored.Adefaultsetofmodelparametersisusedforthelocalmodelsthathavenotyetbeenidentied(seeSection2.3).2.2.4.SelectlocalmodelsAlocalmodelisselectedifthesumofitsmembershipfunctionsbj(k)overtheactivebuffernormalisedbytheactivebufferlengthexceedsagiventhreshold.Onlytheselectedlocalmodelsareincludedinfurtherprocessing.2.2.5.IdentificationThelocalmodelparametersareidentiedusingthefuzzyinstrumentalvariables(FIV)identicationmethoddevelopedbyBlazicetal.(2003).Itisanextensionofthelinearinstrumentalvariablesidenticationprocedure(Ljung,1987)forthespeciedMFM,basedonthelocallearningapproach(Murray-Smith&Johansen,1997).Thelocallearningapproachisbasedontheassumptionthattheparametersofalllocalmodelswillnotbeestimatedin