Toyota Prius HEV neurocontrol and diagnostics

NeuralNetworks21(2008)

458–465

http://doc.guandang.net/locate/neunet

2008SpecialIssue

ToyotaPriusHEVneurocontrolanddiagnostics$

DanilV.Prokhorov

ToyotaTechnicalCenter,AdivisionofToyotaMotorEngineeringandManufacturingNorthAmerica(TEMA),AnnArbor,MI48105,UnitedStates

Received10August2007;receivedinrevisedform20November2007;accepted11December2007

Abstract

Aneuralnetworkcontrollerforimprovedfuelef ciencyoftheToyotaPriushybridelectricvehicleisproposed.Anewmethodtodetectandmitigateabatteryfaultisalsopresented.TheapproachisbasedonrecurrentneuralnetworksandincludestheextendedKalman lter.Theproposedapproachisquitegeneralandapplicabletoothercontrolsystems.c2008ElsevierLtd.Allrightsreserved.

Keywords:RNN;Neurocontrol;Batterydiagnostics;Faultmitigation;HEV;Control;NNmodel;NNcontroller;EKF

1.Introduction

Hybridpowertrainshavebeengainingpopularityduetotheirpotentialtoimprovefueleconomysigni cantlyandreduceundesirableemissions.Controlstrategiesofthehybridelectricvehicle(HEV)aremorecomplexthanthoseoftheinternalcombustionengine-onlyvehiclebecausetheyhavetodealwithmultiplepowersourcesinsophisticatedcon gurations.Themainfunctionofanycontrolstrategyispowermanagement.Ittypicallyimplementsahigh-levelcontrolalgorithmwhichdeterminestheappropriatepowersplitbetweentheelectricmotorandtheenginetominimizefuelconsumptionandemissions,whilestayingwithinspeci edconstraintsondrivability,reliability,batterychargesustenance,etc.

ComputationalintelligencetechniqueshavepreviouslybeenappliedtoHEVpowermanagementbyvariousauthors.Arule-basedcontrolwasemployedinBaumann,Washington,Glenn,andRizzoni(2000).FueleconomyimprovementwithafuzzycontrollerwasdemonstratedinSalman,Schouten,andKheir(2000)andSchouten,Salman,andKheir(2002),relativetootherstrategieswhichmaximizedonlytheengineef ciency.Anothersystemforimprovingfueleconomyintheformoffuzzyrule-basedadvisorwasproposedinSyed,Filev,andYing

$Anabbreviatedversionofsomeportionsofthisarticleappearedin

Prokhorov(2007)aspartoftheIJCNN2007ConferenceProceedings,publishedunderIEEcopyright. Tel.:+17349951017.

E-mailaddress:dvprokhorov@http://doc.guandang.net.c2008ElsevierLtd.Allrightsreserved.0893-6080/$-seefrontmatter

doi:10.1016/j.neunet.2007.12.043

(2007).Theadvisoreitherletsthedriverinputsthroughintact(acceleratorandbrakepositions),oradjustsitslimitstoprovideadvantageouscorrectionsevenforafuelef ciencymindeddriver.

Anintelligentcontrollercombiningneuralnetworksandfuzzylogicwhichcouldadapttodifferentdriversanddrivecycles(pro lesoftherequiredvehiclespeedovertime)wasstudiedinBaumann,Rizzoni,andWashington(1998).Recentlyaneurocontrollerwasemployedinahybridelectricpropulsionsystemofasmallunmannedaerialvehiclewhichresultedinsigni cantenergysaving(Harmon,Frank,&Joshi,2005).Thereferencescitedaboveindicateasigni cantpotentialforimprovingHEVperformancethroughmoreef cientpowermanagementbasedonapplicationofcomputationalintelligence(CI)techniques.ThoughtheToyotaHEVPriusef ciencyisquitehighalready,thereisapotentialforfurtherimprovement,asillustratedinthispaper.

Unliketraditionalhybridpowertrainschemes,seriesorparallel,thePriushybridimplementswhatiscalledthepowersplitscheme.ThisschemeisquiteinnovativeandhasnotbeenstudiedextensivelyyetfromthestandpointofapplicationofCItechniques.ThePriuspowertrainusesaplanetarygearmechanismtoconnectaninternalcombustionengine,anelectricmotorandagenerator.Ahighlyef cientenginecansimultaneouslychargethebatterythroughthegeneratorandpropelthevehicle(Fig.1).Itisimportanttobeabletosettheengineoperatingpointtothehighestef ciencypossibleandatsuf cientlylowemissionlevelsofundesirableexhaustgasessuchashydrocarbons,nitrogenoxidesandcarbonmonoxide.

D.V.Prokhorov/NeuralNetworks21(2008)458–465

459

Fig.1.ThePriuscarandthemaincomponentsoftheToyotahybridsystem.

Themotorisphysicallyattachedtotheringgear.Itcanmovethevehiclethroughthe xedgearratioandeitherassisttheengineorpropelthevehicleonitsownforlowspeeds.Themotorcanalsoreturnsomeenergytothebatterybyworkingasanothergeneratorintheregenerativebrakingmode.

Asinthepreviouswork(Prokhorov,2006;Prokhorov,Puskorius,&Feldkamp,2001),Iemployrecurrentneuralnetworks(RNN)ascontrollersandtrainthemforrobustnesstoparametricandsignaluncertainties(knownboundedvariationsofphysicalparameters,referencetrajectories,measurementnoise,etc.).Iintendtodeploythetrainedneurocontrollerwith xedweights.Itisstilldesirabletohaveapossibilitytoin uencetheclosed-loopperformanceincasesomedegreeofadaptivityisneeded,e.g.,whenanintermittentfaultinthesystemoccurswhichtemporarilymakessigni cantchangesinitsperformance(untilrepairsaremade).Itmaybenothelpfultoadaptweightsofthecontrollerbecause(1)itwouldcompromiseitsalreadytrainedweights,i.e.,itslong-termmemory,whichisundesirableintheintermittentfaultcase,and(2)adaptationinstronglynonlinearsystemscancausebifurcations.Itmaybesafertoaugmentthe xed-weightRNNcontrollerbysimplermeansforadaptation.

Thispaperisstructuredasfollows.Innextsection2describemainelementsoftheoff-linetraining.TheapproachpermitsmetocreateanRNNcontrollerwhichisreadyfordeploymentwith xedweights.IdescribemycontrolexperimentsinSection3.IthenproposeaNNforbatterydiagnosticsanddiscusswaystomitigateabatteryfaultinSection4.Batteryfaultmitigationiscarriedoutbyin uencinginputsoroutputsofthe xed-weightNNcontroller,withadiagnosticNNcontinuouslymonitoringtheclosed-loop

performance.

Fig.2.StepsofmyprocessforNNcontrollertrainingandveri cation.

2.Off-linetraining

Iadopttheapproachofindirectormodel-basedcontroldevelopmentforoff-linetraining.ThePriussimulatorisahighlycomplex,distributedsoftwarewhichmakestraininganeurocontrollerdirectlyinthesimulatordif cult.Iimplementedanapproachinwhichthemostessentialelementsofthesimulatorareapproximatedsuf cientlyaccuratelybyaneuralnetworkmodel.TheNNmodelisusedtotrainaneurocontrollerbyeffectivelyreplacingthesimulator;thiscon gurationisalsoknownastheparallelide …… 此处隐藏：28573字，全部文档内容请下载后查看。喜欢就下载吧 ……