基于遗传算法的BP神经网络优化在无传感器速度上的应用(IJEM-V1-N5-1)

I.J.EngineeringandManufacturing2011,5,1-10PublishedOnlineOctober2011inMECS()DOI:10.5815/ijem.2011.05.01Availableonlineat*1,GuorongLIUb,*2aDepartmentofInformationengineeringinstitute,XiangtanUniversity,Xiangtan,Hunan,ChinabHuNanInstitutionofEngineering,HuNanInstitutionofEngineering,Xiangtan,Hunan,ChinaAbstractThispaperputsforwardastrategythatthesensor-lessbeingoptimizedbytheintelligentalgorithmappliesinthesystemofdirecttorquecontrol.BasedontheBPneuralnetworkintheidentificationofDTCspeed,itoptimizestheBPneuralnetworkspeedidentifierbyusinggenericalgorithm.Itisprovedbythesimulationresultsthattheneuralnetworkhasbeengreatlyimprovedbytheapplicationofintelligentalgorithm,andthespeedsensor-lessbasedontheintelligentalgorithmhasbetterstabilityproperties.IndexTerms:DirectTorqueControl;SpeedSensor-less;BPNeural-Network;GeneticAlgorithm©2011PublishedbyMECSPublisher.Selectionand/orpeerreviewunderresponsibilityoftheResearchAssociationofModernEducationandComputerScience.1.IntroductionDirectTorqueControl(DTC)technologysinceitsinception,receivesuniversalconcernforitsinnovativethinking,excellentdynamicandstaticcharacteristics.Thespeedloopinthehigh-performanceACspeedcontrolsystemisessential[1].Inordertoachievethefeedbackcontrolofthemotorspeedandposition,thespeedsensorneedtobeinstalledinthemotor-side[2].However,thismethodincreasesthecosttothesystemandthevolume,whilereducingthereliability.Duetoitscomplexity,thespeedofidentificationoftheACspeedcontrolsystemhasbeenathornyissue.Withtherapiddevelopmentofintelligentalgorithm,sensorlesstechnologyhasbeengreatlyimproved.Fuzzylogiccontrol,artificialneuralnetworkandaseriesofintelligentcontroltheoryhasbeenwidelyapplied.Asthemostdeveloped,theBPneuralnetworkitself,hasalotofdefects[3].Therefore,weusethegeneticalgorithmstooptimizethestructureandweightsoftheBPnetworks,whichimprovesthepredictionaccuracyofthenetwork,andtheacceleratedconvergencerate,toovercometheshortcomingsoftraditionalforecastingmethods.Simulationresultsshowthatthespeedidentificationismuchbetter.*Correspondingauthor:E-mailaddress:*[e-mail];*[e-mail]2TheApplicationinSpeedSensor-lessBasedonBPNeuralNetworkOptimizedbyGeneticAlgorithm2.TheBasicEquationofDirectTorqueControlThemathematicalmodeloftheACasynchronousmotorinstationarycoordinatesystemasfollow:Voltageequation:000000sssmssssmsrmrmrrrrrrmrrriRLpLpuiRLpLpuiLpLRLpLiLLpLRLp(1)Fluxlinkageequation:0000sssmsssmrrmrrrmriLLiLLiLLiLL(2)Intheformula,variousquantitiesmeaningisasfollows:u、istandforthestator-axisvoltageandcurrentu、istandforthestator-axisvoltageandcurrenti、istandfortherotor-axisvoltageand-axiscurrentRs、LsstandfortheresistanceandinductanceofthestatorwindingsRr、LrstandfortheresistanceandinductanceoftherotorwindingsLmstandsformutualinductancebetweenstatorandrotorwindingsPForderivativeoperatorPstandsfortheDifferentialoperatorrstandsfortheRotorelectricalangularvelocitys、sstandforthestator-axisfluxand-axisfluxr、rstandfortherotor-axisfluxand-axisfluxBasedonthevoltageequationandFluxequation(2),wecangetthespeedexpression:()()()()()rsrssrmmsrsssrssrmmsrrsssLuRiLLLLpiRuRidtRLiLLLLLuRidt(3)orTheApplicationinSpeedSensor-lessBasedonBPNeuralNetworkOptimizedbyGeneticAlgorithm3()()()()()rsrsrmmsrsssrssrmmsrrsssLuRiLLLLpiRuRiRLiLLLLLuRidt(4)Fromtheformulawecanclearlyseethat,amongtheMotorspeedrandStatorvoltage、Statorcurrentthereisonekindofextremelycomplexmisalignmentmappingrelationship[4].3.TheDesignofGeneticAlgorithmOptimizestheBPSpeedIdentificationAsshowninFig.1,thegeneticalgorithmwhichoptimizestheBPneuralnetworkincludingtheBPneuralnetwork'sdetermination,thegeneticalgorithmoptimizationandtheBPneuralnetworkforecastthreeparts.Wecanclearsee,thedeterminationoftheBPneuralnetworkincludesdefinitionnetworktopologyandinitializingnetworkweightandthresholdvalue[5].Byidentifyingtheinput,output,hiddenlayernodeswecandeterminethenetworktopologyandrandomlyinitializetheweightsandthresholdvalueforthefollowingstepsforgeneticalgorithm.Byselection,crossoverandmutationofaseriesofoperations,geneticalgorithmoptimizationobtainstheoptimalweightsandthresholds.FinallyafterthemostsuperiorweightandthethresholdvaluearesubstitutedintheBPneuralnetworkstructure,wecancarryonthetrainingwiththecollecteddata,thenobtainingtheoptimizedresult.Fig.1.GAoptimizeBPNeuralNetworkFlowchart4TheApplicationinSpeedSensor-lessBasedonBPNeuralNetworkOptimizedbyGeneticAlgorithmUUIIr(k)w111w211...Fig.2.Neuralnetworkspeedidentificationnetworkarchitecture3.1.BPNeuralNetworkDesignAsthemostdevelopedBPneuralnetwork,S-functionmulti-layerBPneuralnetworkcanrealizeanynonlinearmappingfrominputtooutput,aslongasthereareenoughhiddenlayerunits.ThisthesisselectsthreeBPneuralnetworkconstitutionspeedobservationmodel,bytheformula(3)andformula(4)det