人人文库网
换一换
首页 人人文库网 > 资源分类 > PDF文档下载
 

外文资料--Neural network prediction.pdf

  • 资源ID:93293       资源大小:365.14KB        全文页数:7页
  • 资源格式: PDF        下载权限:游客/注册会员    下载费用:1
验证码快捷下载 游客一键下载
会员登录下载
三方登录下载: 微信快捷登录 支付宝快捷登录   QQ登录   微博登录  
下载资源需要1
邮箱/手机:
验证码: 获取验证码
温馨提示:
支付成功后,系统会自动生成账号(用户名为邮箱或者手机号,密码是验证码),方便下次登录下载和查询订单;
支付方式: 微信支付    支付宝   
验证码:   换一换

 
友情提示
2、PDF文件下载后,可能会被浏览器默认打开,此种情况可以点击浏览器菜单,保存网页到桌面,既可以正常下载了。
3、本站不支持迅雷下载,请使用电脑自带的IE浏览器,或者360浏览器、谷歌浏览器下载即可。
4、本站资源下载后的文档和图纸-无水印,预览文档经过压缩,下载后原文更清晰   

外文资料--Neural network prediction.pdf

perancealjiceFrictionmaterialormanceThesmaterial’ssynergistandbytraining18differentneuralnetworkarchitectureswiththefivedifferentlearningalgorithms.TheoptimalneuralmodelofdiscbrakeoperationhasbeenshowntobidforpredictingthebrakefactorCvariationofthecolddiscbrakeoverawiderangeofbrake’soperatingregimesandfordifferenttypesingsystxandmanifold.berelatih,andhumiditygbrakeebrakdistance,pedalfeel,discwear,andbrakeinducedvibrations[4].theeinofThesynergeticeffectsofallingredientsincludedinafrictionARTICLEINPRESSContentslistsavailableatScienceDirectsevier.com/locate/tribointTribologyIntTribologyInternational4220091074–1080ditions,iscomplicatedbythefactthatthetribologyatthefrictionE-mailaddressdaleksendricmas.bg.ac.yuD.Aleksendric.Forinstance,thevibrationsgeneratedattheinterfacebetweenthematerial,forthespecificmanufacturingconditions,determinethefinalfrictionmaterialcharacteristicsandaccordinglyaffectthebrakesystem’sperance.Improvementandcontrolofanautomotivebrake’sperance,underdifferentoperatingcon-0301-679X/-seefrontmatter2009ElsevierLtd.Allrightsreserved.doi10.1016/j.triboint.2009.03.005C3Correspondingauthor.Tel.381113370346;fax38113370364.foroverallperanceofavehicle.Thisisbecauseitplayscrucialrolesinvariousaspectsofthebrakeperancesuchasstoppingfrictionmaterialsandbrakingconditions[11]whichbothaffectthebrakingsystem’sperance.stablefrictioncoefficient,lowwearrate,nonoise,lowcost,andenvironmentfriendly[3].Thefrictionmaterialintheautomotivebrakesystemhasbeenconsideredasoneofthekeycomponentsaffectedbythewidediversityinmechanicalpropertiesofcompositematerial’singredients[7–10].Thatiswhy,achangfrictioncoefficientishighlydependentontheingredientsvaluesandstabilityofthefrictioncoefficientoverdifferentbrake’soperatingconditionsdefinedbychangingappliedpressureand/orslidingspeedand/ortemperature.Thefrictionbehaviourofautomotivebrakesisdeterminedbythecharacteroftheactivesurfacesofthediscandpadandthirdbodiesbetweenthesesurfaces[2].Thebrakesrequirefrictionmaterialswithhigherandoperatingregimes.Therefore,thebrake’speranceisprimarilyinfluencedbythecontactsituationbetweenacastironbrakediscandthecompositefrictionmaterial.Thecontactsituationisadditionallycomplicatedbythefactthatfrictionmaterialsarecomplexpolymercompositesandmaycontainover20differentingredients.Hencethecontactsituationcanbesignificantly1.IntroductionThedemandsimposedonabrakofoperatingconditions,arecomplethatthefrictioncoefficientshouldstablefrictionforce,reliablestrengtareneededirrespectiveoftemperature,wearandcorrosion,etc[1].Thebrakinmostlydeterminedbythefoundationrequirementsimposedonautomotiv2009ElsevierLtd.Allrightsreserved.em,overawiderangeItisexpectedvelyhighbutalsogoodwearresistance,age,degreeofsystemperanceisassembly.Thebasicesarerelatedtothetwobodiesinfrictionareresponsibleforvariousnoisessuchassquealing,juddering,hammering,hooting,etc[5].Ontheotherhand,theperancecriteriahaveincreasedandhavebecomemoresensitivetobraking[6].Anautomotivebrake’sfrictionbehaviourresultsfromthecomplexinterrelatedphenomenaoccurringatthecontactofthefrictionpairduringbraking.Thesecomplexbrakingphenomenaaremostlyaffectedbythetribochemicalpropertiesofthecompositematerialasthefrictionelement,thebrakediscasthemetalliccounterface,andtheconditionsimposedbythebrake’soffrictionmaterial.NeuralnetworkpredictionofdiscbrakeDraganAleksendrica,C3,DavidC.BartonbaAutomotiveDepartment,UniversityofBelgrade,FacultyofMechanicalEngineering,KrbSchoolofMechanicalEngineering,UniversityofLeeds,LS29JT,UKarticleinfoArticlehistoryReceived28November2007Receivedinrevised3March2009Accepted16March2009Availableonline24March2009KeywordsNeuralnetworkPredictionDiscbrakeperanceabstractAnautomotivebrake’sperfcontactofthefrictionpair.propertiesofthefrictionregimes.Inthispaper,thecompositionandmanufacturinvariationhavebeenmodelleparameters,determinedbyconditions5parameters,variation,havebeenpredicted.journalhomepagewww.elMarije16,11120Belgrade35,Serbiaresultsfromthecomplexinterrelatedphenomenaoccurringattheecomplexbrakingphenomenaaremostlyaffectedbythetribochemicalingredients,thebrakediscproperties,andthebrake’soperatingiceffectsofthefrictionmaterial’sproperties,definedbyitsgconditions,andthebrake’soperatingregimesonthediscbrakefactorCdbymeansofartificialneuralnetworks.Theinfluencesof26thefrictionmaterialcomposition18ingredients,itsmanufacturingthebrake’soperatingregimes3parametersonthebrakefactorCTheneuralmodelofthediscbrakecoldperancehasbeendevelopedernationalARTICLEINPRESSnumberofneuronsinhiddenlayers,respectively;IScalscaledvalueICurrcurrentvalueIMaxmaximumvalueIMinminimumvalueOLinlinearizedoutputvalueOCurrcurrentoutputvalueOMaxmaximumoutputvalueD.Aleksendric,D.C.Barton/TribologyInternational4220091074–10801075interfacehasastochasticnatureaffectedbyvariationsoftherealcontactarea,transferlayeration,changingpressure,tem-perature,andspeedconditions,aswellasdeationandwearofthecomponents.Theareaofrealcontactbetweenthepadandthediscisfarfromconstant[1],verysmallcomparedtothetotalcontactarea[2],andhighlydependentonchangesofpressure,temperatures,deation,andwear.Takingintoconsiderationthatverycomplexandhighlynon-linearphenomenaareinvolvedinthebrakingprocess[2,11],completeanalyticalmodelsofbrakeoperationaredifficultifnotimpossibletoobtain.Incontrasttoclassicalanalyticalapproaches,itisarguedinthispaperthatartificialneuralnetworkscanbeusedtomodelthecomplexnon-linear,multidimensionalfactorsthatcaninfluenceabrake’sperance.Aspointedoutbymanyresearchers[12–15],forexample,artificialneuralnetworksareapromisingfieldofresearchinpredictingexperimentaltrendsandarecapableofconsiderablesavingsintermsofcostandtimecomparedwithclassicalanalyticalmodels.Inordertoimproveabrakingsystemoperation,itisdesirablethatthebrakesshouldbemorepreciselycontrolledversuschangesofcoefficientofthefriction.Consequently,thebrakeperanceshouldbecalibratedforthespecificbrakeoperatingregimesandafrictionpair’scharacteristics[15–17].Inthispaper,artificialneuralnetworkshavebeenusedodellingandpredictingthediscbrake’sfrictioncharacteristicsi.e.thebrakefactorCvariationtakingintoconsiderationthefollowinginfluen-cingfactorsifrictionmaterialcomposition,iimanufacturingparametersoffrictionmaterial,andiiibrake’soperatingconditions.Therearemanycomplexinfluencesoffrictionmaterialcomposition,itsmanufacturingconditions,andbrakeoperatingregimesonthewearresistanceandnoisypropensityofadiscFthenumberofoutputsfxactivationfunctionNomenclatureCbrakefactorTbrakingtorquepapplicationpressuredcpistondiameterreeffectivebrakediscradiusFtypeoffrictionmaterialFTtypeoffrictionmaterialusedforthetestdatasetA[C–D–E]BFneuralnetworkarchitectureAthenumberofs;Bthenumberofhiddenlayers;C,D,EthebrakebutinthispaperourattentionhasbeenfocusedonpredictionofthediscbrakefactorCasoneofthemostimportantperanceofthediscbrakeoperation.2.ExperimentalsInordertobetaughtaboutthediscbrakeoperationi.e.brakeperanceasafunctionofdifferenttypesoffrictionmaterialandbrake’soperatingconditions,theartificialneuralnetworkshavetobetrainedwithappropriatedata.Theprocessofmodellingofadiscbrakeoperationbymeansofartificialneuralnetworksisnottrivialandmanycriticalissueshavetoberesolved.Thefollowingoperationshavetobeconsiderediselectionofadatagenerator,iidefinitionoftherangesanddistributionofdata,iiidatageneration,ivdatapre-processing,vselectionoftheneuralnetwork’sarchitectures,viselectionofthetrainingalgorithms,viitrainingoftheneuralnetworks,viiivalidationandaccuracyuation,andixtestingoftheartificialneuralnetworks.Thepreliminarystepindevelopmentoftheneuralmodelofadiscbrakeoperationistheidentificationofthemodelsandoutputs./outputidentificationdependsonthemodelobjectivesandchoiceofthedatagenerator.Forthepurposesofthispaper,theparametersaredefinedbythefrictionmaterialcomposition,itsmanufacturingprocessconditions,andthediscbrakeoperatingconditions.ThebrakefactorChasbeentakenastheoutputparameterandusedforrepresentingthediscbrakeperance.ThebrakefactorCcorrespondstochangesofthefrictioncoefficientinthecontactoffrictionpairduringbrakingC2m.ThebrakefactorCiscalculatedfromthemeasuredvariationofthebrakingtorqueandapplicationpressureduringthebrakingcycle,andknownvaluesofthepistondiameterandeffectivebrakediscradiusaccordingtoexpression1C4Tpd2cpre1Thetypeofdatageneratordependsontheapplicationandtheavailability.Inthiscase,thedatageneratorhasbeenasingle-endfull-scaleinertialdynamometer,developedatthelaboratoryforfrictionmechanismandbrakingsystemsFRIMEKSAutomotiveDepartment,FacultyofMechanicalEngineering,UniversityofBelgrade.Obviously,thetestingologyneedstobechosenaccordingtotherangeanddistributionofdatathataregoingtobecollected.Table1presentsthetestingologyusedfortheoutputdatageneration.Thebraketestingconditions,aftertheburnishingprocedure,havebeenchoseninordertoidentifytheinfluencesofappliedhydraulicpressureandinitialequivalentOMinminimumoutputvalueBRBayesianRegulationlearningalgorithmBRabcdneuralmodelBRBayesianRegulationlearningalgorithm;athenumberofs;bthenumberofneuronsinthefirsthiddenlayer;cthenumberofneuronsinthesecondhiddenlayer;dthenumberofoutputsvehiclespeedonthefinalcoldperanceofthediscbrakeforthedifferenttypesoffrictionmaterial[18].Thesedatahavebeenusedfortraining,validation,andtestingoftheneuralnetworksinordertoestablishthefunctionalrelationshipbetweenthediscbrakeoperatingconditions,thetypeofthefrictionmaterial,andthebrakefactorCvariationastheoutput.Itisobviousthattherangesanddistributionofthesdatafortraining,validation,andtestinghavetobepredefined.TheneuralmodelofdiscbrakeoperationtakesintoconsiderationtheTable1Testingology.TestconditionsAppliedpressurebarInitialspeedkm/hTemperature1CNumberofbrakingeventsInitialburnishing4090o100150Brakingregimes20,40,60,80,10020,40,60,80,100o10025ARTICLEINPRESSD.Aleksendric,D.C.Barton/TribologyInternational4220091074–10801076Table2Theselectionandrangesofrawmaterialsforthefrictionmaterialcompositionsvol.RawmaterialsF1–F9trainingandvalidationdatasetFT1testdatasetFT2testdatasetPhenolicresin17–252517Ironoxide3–553Barites26–151526Calciumcarbonate1–331Brasschips1–331Aramid2–662Mineralfibre10–16109Vermiculite4–884Steelfibre4–114Glassfibre2–442Brasspowder1–221Copperpowder1–331Graphite7–337Frictiondust5–225MolybdenumDisulphide1–331Aluminiumoxide2–332Silica1–221Magnesiumoxide8–228Table3Rangesofmanufacturingparametersforthefrictionmaterials.threegroupsofdataithefrictionmaterialcomposition,iiitsmanufacturingconditions,andiiithebrake’soperationregimes.Therangesanddistributionofdatarelatedtothebrake’soperationregimesisdefinedbythetestingologyTable1.Ontheotherhand,choiceoftherangesanddistributionofthemanufacturingandespeciallythecompositionparametersofthefrictionmaterialsisamuchmoredifficulttask.Forthetrainingandvalidationdatasetsation,eachingredientinthecompositionofthefrictionmaterialanditsmanufacturingparametershavebeenselectedwithinarangeF1–F9aspresentedinTables2and3.FromTables2and3,itcanbeseenthatelevendifferenttypesoffrictionmaterialwereproducedasadiscpadassembly,mountedonthefrontbrakeaxlestaticweightof730kgofasmallpassengercarYugoFlorida1.4andtestedusingthesingleendfull-scaleinertialdynamometer.Thediscpadswiththefrictionsurfaceareaof32.4cm2andpadthicknessof16.8mmweredesignedforthebrakewithaneffectivediscradiusof101mmandfloatingcalliperpistondiameterof48mm.Thecompositionandmanufacturingparametersforeachtypeoffrictionmaterial,aspresentedinTables2and3,werecompletelydifferentfromoneanother.ResultsobtainedduringbraketestingwithfrictionmaterialsF1–F8wereusedfortrainingtheneuralnetworks,whileresultswiththefrictionmaterialF9wereusedforvalidatingthecapabilitiesoftheartificialneuralnetworks.Thevolumepercentagesofthefrictionmaterial’singredients,usedfortheneuralnetworks’trainingandvalidationF1–F9overtherangespresentedinTable2,havebeenrandomlyselected.Thelearningalgorithmselected[19,20].ThelearningabilityoftheManufacturingparametersF1–F9trainingandvalidationdatasetFT1testdatasetFT2testdatasetSpecificmouldingpressureMPa45–654070Mouldingtemperature1C155–170170155Mouldingtimemin6–11116Heattreatmenttemperature1C200–250200250Heattreatmenttimeh12–5125neuralnetworktoextenditspredictivepowerfordataoutsideofthetrainingdatasetisessentialinimplementationoftheartificialneuralnetworksforpredictingdiscbrakeperance.Itisaclearthatsufficient/targetpairshavetobestoredinthetrainingdataset./outputdatahavebeenobtainedbyulation,manufacturing,anddynamometertestingofelevendifferentfrictionmaterialsrepresentingalargedatasetthatcanbeusedfortraining,validation,andtestingtheneuralnetwork.Thetotalnumberofoutputresults,obtainedbythedynamometertestingforeachtypeoffrictionmaterial,is25accordingtotheadoptedtestingologyTable1.Thismeansthat275/outputpairsareavailablefortheneuralnetworktraining,validation,andtesting.Thetotalnumberof275/outputpairshasbeendividedintothreesets,200/outputpairsfortheneuralnetworktraining,25pairsforvalidation,and50pairsfortheneuralnetworktesting.Sincethebestneuralnetworkarchitectureandalearningalgorithmareunknowninadvance,atrailanderrorhasbeenemployedtofindoutthebestnetworkcharacteristicsatchingtheparticular/outputrelationship.BasedonMatLab6.5Rel.13,thefollowingnetworksarchitectureshavebeeninvestigatedinthisapplicationione-layeredstructures26[1]11,26[2]11,26[3]11,26[5]11,26[8]11,iitwo-layeredstructures26[1–1]21,26[2–2]21,26[3–2]2manufacturingparameters,presentedinTable3,havebeenalsorandomlyselectedinthecaseoffrictionmaterialsdenotedasF1–F9.Furthermore,theaccuracyofthetrainedneuralnetworksforpredictingthediscbrakeperancehasbeentestedusingthe/outputdatastoredinthetestdataset.ThetestdatasethavebeenobtainedbyproducingtwonewtypesoffrictionmaterialsFT1andFT2whoseparameterswerecompletelydifferentfromthosestoredinthetrainingandvalidationtestdatasets.Thevolumepercentageofingredients,presentedinTable2,usedforthecompositionoffrictionmaterialsFT1andFT2weremostlyselectedtocorrespondtotheupperandlowerboundvaluesofthespecifiedranges.ThemanufacturingparametersofthefrictionmaterialsFT1andFT2havebeenalsosimilarlyselectedregardingtherangesspecifiedinTable3.Theonlydifferenceisrelatedtothespecificmouldingpressureswhich,inthecaseoffrictionmaterialsFT1andFT2,wereoutoftherangeusedanufactur-ingthefrictionmaterialsdenotedasF1–F9seeTable3.Thesueswereselectedinordertotesttheneuralmodelabilitiestoextenditspredictivepowerfordataeitherattheendsoftherangesorcompletelyoutsideoftherangesusedforthetrainingdatasetcreation.3.NeuralnetworkmodellingBasedonTables1–3,neuralmodellingofthediscbrakeoperationhasbeenperedfor26parameters18parametersrelatedtothefrictionmaterial’scomposition,5parametersrelatedtothemanufacturingconditions,and3parametersrelatedtothebraketestingconditions,andoneoutputparameterbrakefactorC.Neuralmodellingofthediscbrakeoperationisacomplextaskandtheappropriatearchitec-tureoftheneuralnetworkaswellasthelearningalgorithmneedtobeproperlydetermined.Thearchitectureofanartificialneuralnetworkconsistsofadescriptionofhowmanylayersanetworkhas,thenumberofneuronsineachlayer,eachlayer’stransferfunctionandhowthelayersareconnectedtoeachother.Thebestarchitecturetousedependsonthekindofproblemtoberepresentedbythenetwork.Thebestneuralnetworksetisaffectedbytherepresentationalpowerofthenetworkandthe1,26[5–2]21,26[8–2]21,26[8–4]21,26[10–5]21,andiiithree-layeredstructures26[3–2–2]31,26[4–3–2]31,26[4–2–2]31,26[5–2–2]31,26[8–2–2]31,26[8–4–2]31.Thesenetworkarchitectureshavebeentrainedbythefollow-ingtrainingalgorithmsLevenberg–Marquardt,BayesianRegula-tion,ResilientBackpropagation,ScaledConjugateGradient,andGradientDecent.Thesigmoidactivationfunctionhasbeenusedbetweentheandhiddenlayersseeexpression2fx11eC0 x2Alinearactivationfunctionfx1xwasemployedbetweenthehiddenandoutputlayer.Pre-processingoftheparameterswascarriedoutbeforetheneuralnetworktraining.Thus,18parametersrelatedtothefrictionmaterialulationwerepresentedtothenetworkinpercentbyvolume,and5manufacturingparametersand3testingconditionswerescaledintherangeof0–1accordingtoexpression3ICurrC0IMaxtwotypesoffrictionmaterialsFT1andFT2havebeenfirstlyproducedandtestedusingtheinertialfull-scalebrakedynam-between20and60barand60and100bar,inordertobetterillustratethecomplexityofrealchangesofdiscbrakeperanceinfluencedbythefrictionmaterialFT1underthespecifieddiscbrakeoperationregimes.FromFig.1,thegeneraltrendofthediscbrakeperanceisevidentforappliedpressuresbetween20and60barandinitialspeedsbetween20and100km/h.ThebrakefactorCincreasesintherangeof20–40barforinitialspeedsbetween20and60km/h.Thediscbrakeperanceisrelativelyconstantoverthewholerangeofinitialspeeds20–100km/hforfurtherincreasesofappliedpressurefrom40to60bar,seeFig.1.Forinitialspeedsbetween80and100km/h,thebrakefactorChasbeenrelativelyconstantinthesamerangeofappliedpressures20–60bar.Contrarytorelativelyconstantdiscbrakeperance,intherangeofappliedpressuresbetween40and60barandinitialspeedsbetween20and100km/h,thediscbrakeperancehasbeendecreasedbyfurtherincreasingofappliedpressureto100barFig.2.Obviously,themeasureddiscbrakeperancehasbeendifferentlyaffectedbythefrictionmaterialpropertiesinsynergywithchangesofthebrakeoperationregimesFigs.1and2.FromFigs.1and2,itcanbeseenthreedifferentrangesofdiscbrakeoperationversusappliedpressuresandinitialspeedsexistabetween20and40bar,bbetween40and60bar,andcbetween60and100bar.AccordingtoFigs.1and2,themeasureddiscbrakeperfor-ARTICLEINPRESS0.840.860.880.90.920.940.96BrakefactorCv20v40v60v80v100Fig.3.Predicteddiscbrakeperanceversusappliedpressures20–60barandinitialspeeds20–100km/hfrictionmaterialFT1.D.Aleksendric,D.C.Barton/TribologyInternational4220091074–10801077ometer.ThecompositionandmanufacturingparametersoffrictionmaterialsFT1andFT2havebeencompletelyunknowntotheneuralmodels.TheperanceofthediscbrakeequippedwiththefrictionmaterialFT1isshowninFigs.1and2versusapplicationpressureandinitialspeedchanges.Themeasureddiscbrakeperance,expressedasthebrakefactorCvariation,hasbeendividedintotworangesdependingonappliedpressure,0.840.860.880.90.920.940.960.98120PressureapplicationbarBrakefactorCv20v40v60v80v10060504030Fig.1.Measureddiscbrakeperanceversusappliedpressures20–60barandIScal1IMaxC0IMin3Ontheotherhand,theoutputparameterbrakefactorChasbeenlinearizedbyexpression4OLin07502OCurrC0OMaxOMaxC0OMin44.ResultsanddiscussionAftertheirtrainingandvalidation,theneuralnetworkshavebeenemployedforpredictingtheperanceofthediscbrakeequippedwiththetwonewtypesofdiscpadsFT1andFT2.Intotal90differentneuralmodelshavebeentested18differentneuralnetworkstrainedbythefivelearningalgorithmsinordertouatetheircapabilitiesforpredictingthediscbrakefactorCvariationasinfluencedbythedifferenttypesoffrictionmaterialsunderthespecificbrakingregimes.Asmentionedabovethenewinitialspeeds20–100km/hfrictionmaterialFT1.0.840.860.880.90.920.940.960.98160PressureapplicationbarBrakefactorCv20v40v60v80v100100908070Fig.2.Measureddiscbrakeperanceversusappliedpressures60–100barandinitialspeeds20–100km/hfrictionmaterialFT1.0.98120Pressureapplicationbar60504030mancehasbeenstronglyaffectedbytheoperatingconditionsfor

注意事项

本文(外文资料--Neural network prediction.pdf)为本站会员(英文资料库)主动上传,人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对上载内容本身不做任何修改或编辑。 若此文所含内容侵犯了您的版权或隐私,请立即通知人人文库网(发送邮件至[email protected]或直接QQ联系客服),我们立即给予删除!

温馨提示:如果因为网速或其他原因下载失败请重新下载,重复下载不扣分。




关于我们 - 网站声明 - 网站地图 - 资源地图 - 友情链接 - 网站客服 - 联系我们

网站客服QQ:2846424093    人人文档上传用户QQ群:460291265   

[email protected] 2016-2018  renrendoc.com 网站版权所有   

备案号:苏ICP备12009002号-5