車輛工程外文翻譯-汽車主動懸架系統(tǒng)的神經(jīng)網(wǎng)絡控制運算法則的研究【中文2904字】【PDF+中文W,中文2904字,車輛,工程,外文,翻譯,汽車,主動,懸架,系統(tǒng),神經(jīng)網(wǎng)絡,控制,運算,法則,研究,中文,2904,PDF 外文翻譯 外文資料名稱：汽車主動懸架系統(tǒng)的神經(jīng)網(wǎng)絡控 制運算法則的研究 外文資料出處：International Conference on Neural Networks and Brain, 2005. 【中文2904字】 汽車主動懸架系統(tǒng)的神經(jīng)網(wǎng)絡控制運算法則的研究 L.J.Fu, J.G.Cao 重慶工學院車輛工程系 中國重慶市楊家坪興盛路4號，400050 C. R. Liao, B. Chen 重慶技術學院車輛工程系 中國重慶市楊家坪興盛路4號，400050 E-mail: flj@cqit.edu.cn 周祥 譯 摘要:為適應不同路面狀況和汽車運行狀況,半可控懸架由從動彈簧和活動減振器組成。由于主動懸架結構復雜并且消極懸架無法滿足各種路面條件和汽車運行狀態(tài)的要求,因此半可控懸架系統(tǒng)是目前最常用的懸架系統(tǒng)。本文將著重介紹自適應神經(jīng)控制的汽車懸架循環(huán)神經(jīng)網(wǎng)絡模擬控制器。懸架系統(tǒng)神經(jīng)網(wǎng)絡不同于汽車懸架的動態(tài)參數(shù),并且還能夠為神經(jīng)自動調(diào)節(jié)控制器提供學習信號,為了檢驗控制結果，在DSP微處理系統(tǒng)基礎上為中巴安裝液壓減振器和多維控制系統(tǒng),并在各種速度和路面上進行實驗.將此控制結果和開環(huán)消極懸架系統(tǒng)進行比較,結果表明神經(jīng)網(wǎng)絡控制運算在減少微型客車振動方面表現(xiàn)的非常良好。 1．概述 汽車懸架系統(tǒng)的主要功用是支撐車身的重量,并且使汽車穩(wěn)定有效的進行轉(zhuǎn)向操縱控制,同時有效的分離路面波動對車身的影響。不同的需要導致設計的要求不同,半自動懸架由從動彈簧和需要克服不同路面狀況和汽車運行條件的阻尼離的自動減振器組成。由于主動懸架結構復雜而傳統(tǒng)的消極式懸架無法滿足不同路面狀況和汽車運行狀況的要求。因此,半自動懸架是目前最常用的懸架系統(tǒng)。半自動懸架系統(tǒng)的優(yōu)點是帶有液壓減振使車身在低動力情況下振動降低。目前,許多控制系統(tǒng)是為半自動懸架系統(tǒng)而開發(fā)的。從Karnoopp的Skyhook方法開始。這個方法主要是使緩沖器承受一定的力的作用,而這個力是與汽車全速時懸架上的質(zhì)量成一定比例的。許多調(diào)查都是用一維模型,它可以推導出模糊的控制點和控制運算法則。如LQG和活躍控制[2,3]。由于汽車懸架固有非線性特性,導致這種控制方法不能充分發(fā)揮半自動懸架的功用。為充分利用懸架系統(tǒng)的非線性功用。如模糊邏輯控制。神經(jīng)網(wǎng)絡控制和模糊神經(jīng)控制等智能化控制方法近來都已被科研人員用于非線性懸架系統(tǒng)控制[4,5]。 本文，一種神經(jīng)自適應控制控制器被用于控制汽車懸架神經(jīng)網(wǎng)絡和瞬邊的MR減振器的循環(huán)振動?？刂破鞯慕Y構設計和控制運算法則將在第2部分進行詳細敘述。懸架的循環(huán)神經(jīng)網(wǎng)絡動態(tài)模擬在第3部分進行介紹控制系統(tǒng)實驗在第4部分，第5部分是總結。 1. 汽車懸架的多維自調(diào)節(jié)控制法則 神經(jīng)模糊控制系統(tǒng)將在本文進行介紹，由圖1可知，它是由模糊神經(jīng)網(wǎng)絡和神經(jīng)網(wǎng)絡模型構成的微型客車懸架。神經(jīng)網(wǎng)絡模糊控制即自適應控制，它有學習和控制的功能。它的循環(huán)神經(jīng)網(wǎng)絡功用是用來鑒別中巴車懸架的模擬參數(shù)。圖1中的y(t)和yd(t)分別是系統(tǒng)實際輸出和系統(tǒng)理想輸出。xl(t)是系統(tǒng)實際輸出和理想輸出之間的誤差。x2(t)是系統(tǒng)實際輸出和理想輸出的誤差率xl(t)和x2(t)定義如下： xI (t) e(t)= y(t)- Yd (t) (1) X2 (t)= e(t)= e(t + 1)- e(t) (2) 圖1.懸架神經(jīng)網(wǎng)絡控制系統(tǒng)的結構 網(wǎng)絡控制系統(tǒng)：整體集的定義分別如下： = [- E,E], = [- E,E], =[-U,U].神經(jīng)模糊控制器有四層神經(jīng)元。第一層和第二層和與模糊法則相一致。第三層與推理相一致，而第四層與模糊法則相一致。, 和的集合分別分成7個子集，，，集的組成分別如下： X1 = {NB, NM, NS, ZE, PS, PM, PB} X2 = {NB, NM, NS, ZE, PS, PM, PB} U = {NB, NM, NS, ZE, PS, PM, PB} 本文，將用高斯函數(shù)解決模糊集,和模糊集的組成，其函數(shù)的第一如下： 圖2.自動懸架神經(jīng)網(wǎng)絡控制器簡圖 ，由圖2可知，輸入/輸出如下：1： 和 和 都是神經(jīng)網(wǎng)絡的輸入部分。是其重量，是其輸出部分，，都是高斯函數(shù)的重要值。神經(jīng)網(wǎng)絡控制器的學習法則是以斜率誤差信號逆向傳遞方法為基礎的。誤差逆向傳遞方法通過使函數(shù)[5]損失降至最低自動調(diào)節(jié)重量。 3．懸架循環(huán)神經(jīng)網(wǎng)絡動態(tài)模擬法則 懸架神經(jīng)網(wǎng)絡設計用于將實際輸出量通過第三層神經(jīng)網(wǎng)近似反饋給潛在的循環(huán)層，結構如圖3所示。其性能是使循環(huán)神經(jīng)網(wǎng)絡能夠自動獲知周圍環(huán)境并且據(jù)此提高其重量自動適應作用.循環(huán)神經(jīng)網(wǎng)絡輸入信號和和潛在層的邏輯反饋循環(huán)神經(jīng)的輸出量的總輸出量對等于神經(jīng)。 圖3.懸架系統(tǒng)神經(jīng)網(wǎng)絡模擬簡圖。 是循環(huán)神經(jīng)網(wǎng)絡的負荷，是潛在層邏輯循環(huán)反饋神經(jīng)的輸出神經(jīng)量，分別是輸入神經(jīng)量和反饋神經(jīng)量。激活函數(shù)是輸入函數(shù)和輸出函數(shù)的線性函數(shù)，潛在層神經(jīng)的激活是S形的函數(shù)。 它的反函數(shù)通過誤差信號定義如下： 是誤差能量的瞬時值.神經(jīng)元的突出質(zhì)量一步一步連續(xù)的自動調(diào)節(jié)直至系統(tǒng)達到 穩(wěn)定狀態(tài)，即突出質(zhì)量基本上穩(wěn)定。 從式1，2和3可知： 從4，5和6分析和分別推導出循環(huán)分子式。 突出質(zhì)量可以由下式計算得到： · 是速率參數(shù)，詳細分析循環(huán)算法獲得速率參數(shù)值是相當復雜的。根據(jù)式13得，循環(huán)神經(jīng)網(wǎng)絡質(zhì)量矢量能夠自動調(diào)節(jié)。函數(shù)如下，其變值經(jīng)過t時間可以定義為： 我們通過式13和式14可以知道誤差信號如下： 函數(shù)增量經(jīng)過t時間可以定義為： . 4 .路面測試結果分析 神經(jīng)控制運算的正確性的證明，帶有MR液壓減振器的微型客車懸架在中國已經(jīng)大量投產(chǎn)制造. 微型客車自適應懸架系統(tǒng)由一個DSP微處理系統(tǒng)，8個加速度傳感器，4個MR液壓減振器和一個輸入電壓為12v的可控循環(huán)電流控制器組成.DSP微處理器通過傳感器獲取懸架彈簧負載和空載時候的懸架振動信號.根據(jù)振動信號和本文的控制圖，DSP微處理系統(tǒng)通過調(diào)節(jié)控制信號來調(diào)節(jié)MR液壓減振器中的電磁線圈的電流。 MR液壓減振器電磁線圈產(chǎn)生的磁場能夠在壓縮沖程和反彈過程中調(diào)節(jié)MR液壓減振器中流體運行狀態(tài)。 本文描述的是以神經(jīng)網(wǎng)絡控制為基礎的微型客車懸架的路面測試，其速度分別為30，40，50 km/h.路面測試過程中微型客車以恒定的速度運行。自適應懸架分別以神經(jīng)網(wǎng)絡和消極懸架系統(tǒng)在同樣的路面和運行速度下進行測試實驗。表1的測試結果表明神經(jīng)網(wǎng)絡控制自適應懸架能夠在懸架彈簧重載和空載的條件下都能減小振動。 圖4描述的是滿載和空載時候的消極和自適應微型客車懸架在D級路面上的振動曲線圖。很明顯神經(jīng)網(wǎng)絡控制主要提高減緩振動的能力。受力曲線圖表明自適應懸架系統(tǒng)和消極懸架系統(tǒng)相比較能夠明顯減小微型客車的振動。減振器有卓越的模糊控制原理和模擬推理，帶有神經(jīng)網(wǎng)絡控制的自適應懸架系統(tǒng)遠乘舒適性能和路面穩(wěn)定保持性能。 表1 微型客車懸架路面測試結果 微型客車懸架滿載和空栽時速度變換曲線（D級路況） 圖4.微型客車振動力曲線圖 （左）滿載 （右）空載 （速度40km/h） 結論 本文中主要講述的是微型客車的一種新型的循環(huán)神經(jīng)網(wǎng)絡模型和模糊神經(jīng)控制原理.根據(jù)要求使用8個加速度傳感器和一個信號處理器?？紤]到MR減振器的復雜性，動態(tài)參數(shù)載入硬盤進行仿真.它表明自適應控制系統(tǒng)可以通過模糊神經(jīng)控制和循環(huán)神經(jīng)網(wǎng)絡懸架達到完全控制作用。由于控制法設計，增益調(diào)度策略和硬件循環(huán)仿真的開發(fā)本文限于微型客車的具體參數(shù)，在懸架參數(shù)變化的情況下此方法可以延伸到其它半主動懸架系統(tǒng).路面實驗結果表明模糊神經(jīng)控制可以有效改善微型客車行使的舒適性和穩(wěn)定性。使用DSP控制器能有效的減小整個車身的振動，包括滿載時候和非滿載時候的振動。模糊神經(jīng)控制器可以減少對對控制系統(tǒng)性能影響很大的模擬參數(shù)的變化。 參考文獻 [1] Kanopp D. (1995) Active and Semi-active Vibration Isolation,Transactions of ASME, Journal of Special 50th Anniversary Design Issue, Vol.117, p1 17-125. [2] Chantrnuwathhana, S. and Peng, H. (1999) Adaptive Robust Control For Active Suspension, proceedings of the American Control Conference, San Diego, California, pp.l702-1706 [3] Yu, F. and Crolla, D.A. (1998) an Optimal Self-Tuning Controller for Active Suspension, Vehicle System Dynamics, vol.29, pp.51-65 [4] Zadeh, A., Fahim, A., and El-Gindy, M. (1997) Neural Networks and Fuzzy Logic Applications to Vehicle System, International Journal of Vehicle Design, vol.18 (2), pp.132-193 [5] Wuwei Chen, James K. Mills and Le Wu,(2003) Neurofuzzy and Fuzzy Control of Automotive Semi-Active Suspensions, International Journal of Vehicle Autonomous Systems, vol.1 (2), pp.222-236 7 StudyonNeuralNetworksControlAlgorithmsforAutomotiveAdaptiveSuspensionSystemsL.J.Fu,J.G.CaoSchoolofAutomobileEngineering,ChongqingInstituteofTechnology,XingshengRoadNo.04Yangjiaping,Chongqing,China400050E-mail:Abstract-Thesemi-activesuspension,whichconsistsofpassivespringandactiveshockabsorberinthelightofdifferentroadconditionsandautomobilerunningconditions,isthemostpopularautomotivesuspensionbecauseactivesuspensioniscomplicatedinstructureandpassivesuspensioncannotmeetthedemandsofvariousroadconditionsandautomobilerunningconditions.Inthispaper,aneurofuzzyadaptivecontrolcontrollerviamodelingofrecurrentneuralnetworksofautomotivesuspensionispresented.Themodelingofneuralnetworkshasidentifiedautomotivesuspensiondynamicparametersandprovidedlearningsignalstoneurofuzzyadaptivecontrolcontroller.Inordertoverifycontrolresults,amini-busfittedwithmagnetorheologicalfluidshockabsorberandneurofuzzycontrolsystembasedonDSPmicroprocessorhasbeenexperimentedwithvariousvelocityandroadsurfaces.Thecontrolresultshavebeencomparedwiththoseofopenlooppassivesuspensionsystem.Theseresultsshowthatneuralcontrolalgorithmexhibitsgoodperformancetoreductionofmini-busvibration.I.INTRODUCTIONThemainfunctionsofautomotivesuspensionsystemaretoprovidesupporttheweightofautomobile,toprovidestabilityanddirectioncontrolduringhandlingmaneuversandtoprovideeffectiveisolationfromroaddisturbances.Thesedifferenttasksleadtoconflictingdesignrequirements.Thesemi-activesuspension,whichconsistsofpassivespringandactiveshockabsorberwithcontrollabledampingforceinthelightofdifferentroadconditionsandautomobilerunningconditions,isthemostpopularautomotivesuspensionbecausetheactivesuspensioniscomplicatedinstructureandconventionalpassivesuspensioncannotmeetthedemandsofdifferentroadconditionsandautomobilerunningconditions.Simi-activesuspensionwithvariablemagnetorheological(MR)fluidshockabsorbershassomeadvantagesinreducingautomobilevibrationatrelativelowcastandpower.Sofar,thereareanumberofcontrolmethodsthathavebeendevelopedforsemi-activesuspension,startwithskyhookmethoddescribedbyKarnoopp,etal.lThismethodattemptstomaketheshockabsorberexertaforcethatisproportionaltotheabsolutevelocitybetweensprungmasses.SomeinvestigationsuseC.R.Liao,B.ChenSchoolofAutomobileEngineering,ChongqingInstituteofTechnology,XingshengRoadNo.04Yangjiaping,Chongqing,China400050E-mail:chenbao(linearsuspensionmodel,whichislinearizedaroundtheoperationalpoints,andcontrolalgorithmarederivedusinglinearmodels,suchasLQGandrobustcontrol2,3.Thesecontrolmethodscannotmakeafullexploitationofsemi-activesuspensionresourcesbecauseofautomotivesuspensionisinherentnon-linearperformance.Inordertoimproveperformanceofnonlinearsuspensionsystem,someintelligentcontroltechniques,suchasfuzzylogiccontrol,neuralnetworkscontrolandneurofuzzycontrol,havebeenrecentlyappliedtononlinearsuspensioncontrolbyresearchers4,5.Inthispaper,aneurofuzzyadaptivecontrolcontrollerisappliedtocontrolsuspensionvibrationviamodelingofrecurrentneuralnetworksofautomotivesuspensionandcontinuouslyvariableMRshockabsorbers.Thecontrollerstructuresdesignandneurofuzzycontrolalgorithmsarepresentedinsection2.Arecurrentneuralnetworksdynamicsmodelingofsuspensionareshownrespectivelyinsection3.Thecontrolsystemexperimentationsaregiveninsection4andsomeconclusionsarefinallydrawninsection5.HI.NEUROFUZZYADAPTIVECONTROLALGORITHMSFORAUTOMOTIVESUSPENSIONSTheneurofuzzycontrolsystempresentedinthispaper,showninFigure1,iscomposedofaneurofuzzynetworkandarecurrentneuralnetworkmodelingofmini-bussuspension.Theneurofuzzynetworkisdefinedasadaptivecontroller,whichhasfunctionoflearningandcontrol.Thefunctionofrecurrentneuralnetworkistoidentifymini-bussuspensionmodelparameters.y(t)andyd(t)aresystemactualoutputandsystemdesireoutputrespectivelyinFigure1.xl(t)issystemerrorofsystemactualoutputbetweensystemdesireoutput,x2(t)issystemerrorrateofsystemactualoutputbetweensystemdesireoutput.xi(t)andx2(t)aredefinedasfellows:xI(t)e(t)=y(t)-Yd(t)(1)X2(t)=e(t)=e(t+1)-e(t)(2)0-7803-9422-4/05/$20.00C2005IEEE1795Fig.1.structureofneuralnetworkscontrolsystemforsuspensionnetworkscontrolsystem.Theglobalsetsoflinguisticvariablesaredefinedrespectivelyasfellows:-=-E,E,1=-AtJuU-U,U.Theneurofuzzycontrollerhasfourlayersne-urons,inwhichthefirstandthesecondlayerscorrespondtothefuizzyrulesif-part,thethirdlayercorrespondstotheinferenceandtheforthlayercorrespondstothefuzzyrulesthen-part.Thesetsxl,x2anduarerespectivelydivinedintosevenfuzzysubsetsofwhichfuzzysetsX1,X2Uarecomposedasfallowsrules:X1=NB,NM,NS,ZE,PS,PM,PBX2=NB,NM,NS,ZE,PS,PM,PBU=NB,NM,NS,ZE,PS,PM,PBInthispaper,theGaussianmembershipfunctionareusedinelementsoffuzzysetsX1X2andtheelementsoffuzzysetUisdefinedasfollowingmembershipfunctionci(u)J0(otherwise)0(3)=I(3)k=1,2,3.49j=13,23,3.74949Layer4:(4)-(3)wkand0(4)=I(4)/0(3)k=1k=1Wherexl(t)x2(t)aretheinputsofneuralnetworks,wkisweightofneuralnetwork,0(4)iStheoutputofneuralnetworksinwhich0(4)=U,ai,b,jarethecentralvaluesofGaussianmembershipfunction.Learningalgorithmsoftheneuralnetworkscontrollerisbasedongradientdescentbymeansoferrorsignalback-propagationmethod.Theerrorback-propagationalgorithm.saccomplishsynapticweightadjustmentthroughminimizationofcostfunction5.m.ALGORITHMFORRECURRENTNEURALNETWORKSSUSPENSIONDYNAMICALMODELINGArecurrentneuralnetworkdesignedtoapproximatetotheactualoutputofsuspensiony(t)isthree-layerneuralnetworkwithonelocalfeedbackloopinthehiddenlayer,whosearchitecturesareshowninFigure3.Thepropertythatisofprimarysignificanceforrecurrentneuralnetworkistheabilityofthenetworktolearnfromitsenvironmentandtoimproveitsperformancesbymeansofprocessofadjustmentsappliedtoitsweights.TherecurrentnetworkwithinputsignalII(t)=u(t)andI2(t)=y(t-1)hasoutputy(t)bylocalfeedbackloopneuroninthehiddenlayerwhoseoutputsumisSj(t)correspondingtotheneuronjth.(3)Fig.2.schematicofneuralnetworkscontrollerforadaptivesuspensionWhereU*Eu.Theinput/outputispresentedasfollowsaccordingtoFigure2.Layer1:I(1)x(t)andO)xi(t)i=1,2Layer2:I-2)(t)-ai)2/b2andO.epx()i=1,2j=1,2,3.7Layer3:I13)=tu(X2Q)IandFig.3.schematicofneuralnetworksmodelingofsuspensionsystem(4)Sy()=,w.*i(t)+WJD_Xj(t-_1)i1=(i(t)+wjXj(t_lqyj(t)=1wxi(t)j=l(5)(6)1796wherewI,wareweightoftherecurrentneuralnetwork,Xj(t)isoutputofneuronwithlocalfeedbackloopneuroninthehiddenlayer,p,qareinputneuronnumberandfeedbackneuronnumberrespectively.Theactivationfunctionforbothinputneuronsandoutputneuronsislinearfunction,whiletheactivationforneuronsinthehiddenlayerissigmoidfunction.heobjectivefunctionE(t)canbedefmedinthetermsoftheerrorsignale(t)as:E(t)=_y(t)-.y(t)2=1e2(t)(7)22Thatis,E(t)istheinstantaneousvalueoftheerrorenergy.Thestep-by-stepadjustmentstothesynapticweightsofneuronarecontinueduntilthesystemreachsteadystate,i.e.thesynapticweightsareessentiallystabilized.DifferentiatingE(t)withrespecttoweightvectorwyields.aE(t)_8=-e(t)0Y()(8)Fromexpression(1),(2)and(3),differentiatingA(t)0DIwithrespecttotheweightvectorw1w,-,w,-Yrespectivelyyields.aS(t)=x(t)As(t)woax1Q)-(WaXI(t)aWjJaWjFrom(4),(5)and(6),analyzingvalueofsynapticweightisdeterminedbyw(t+1)=w(t)+q*e(t)89(t)(12)whereqtheleaning-rateparameter,Adetailedconvergenceanalysisoftherecurrenttrainingalgorithmisrathercomplicatedtoacquiretheleaning-rateparametervalue.Accordingtoexpression(13),theweightvectorwforrecurrentneuralnetworkcanbeadjusted.WeestablishatheLyapunovfunctionasfollowsV(t)=1/2*e2(t),whosechangevalueAV(t)canbedeterminedaftersometiterations,inthesensethat(13)Wehavenoticedthattheerrorsignale(t)aftersometiterationscanbeexpressedasfollowsfromexpression(13)and(14),ae(t)ao(t)ae(t)ae(t)-,Aw=-qe(t)=77e(t),theawOwawOwLyapunovfunctionincrementcandeterminedaftersometiterationsasfollows(14)Mtt)=-q-&(t)+v2.e(t)-=-V(t)where(t)22jt16(t)2A=10()lpq2-5l0(t)ll2ql2-77O220w(9)?7maxa(t)29ifqf2,thenAV(t)O,wax1(t)DandaWjx1(t)uxiyieldsrespectivelyrecurrentformulas.ax1(t)a-fS(t)FX.x(tt1)1ax1(O)=,WjD=axi(t)aNiafS(t)+wat-i)&4LaNiax1(o)(11)avn=0Havingcomputedthesynapticadjustment,theupdatednamelytherecurrenttrainingalgorithmisconvergent.IV.ROADTESTANDRESULTSANALYSESTomakeademonstrationthevalidityofneuralcontrolalgorithmproposedinthepaper,anexperimentalmini-bussuspensionwithMRfluidshockabsorberhasbeenmanufacturedinChina.Themini-busadaptivesuspensionsystemconsistsofaDSPmicroprocessor,8accelerationsensors,4MRfluidshockabsorbers,and1controllableelectriccurrentpowerwithinputvoltage12V.TheDSPmicroprocessorreceivessuspensionvibrationsignalinputfromaccelerometersmountedrespectivelysprungmassandun-sprungmass.Inaccordancewithvibrationsignalandcontrolschemeinthispaper,theDSPmicroprocessoradjustsdampingofadaptivesuspensionbyapplicationcontrolsignaltothecontrollableelectriccurrentpowerconnectedtoelectromagneticcoilinMRfluidshockabsorbers.MagneticfieldproducedbytheelectromagneticcoilinMRfluidshockabsorberscandvarydampingforceinbothcompressionandreboundbyadjustmentofflow1797II,&V(t)=12(t+1)-e2(t2behaviorsofMRfluidsindampingchannels.Raodtestonmini-busadaptivesuspensionbasedneuralnetworkscontrolpresentedinthispaperarecarriedoutinDclassroadsurfacesrespectivelyinrunningvelocity30,40,50km/h.Duringroadtest,experimentalmini-busrunseachtestconditionataconstantspeed.Thetestexperimentsofadaptivesuspensionwithneuralnetworksandpassivesuspensionsystemwerecarriedoutrepeatedlyundersameroadsurfaceandrunningvelocity.TestresultslistedinTable1haveshownthattheadaptivesuspensionwithneuralnetworkscanreducevibrationpowerspectraldensitiesofbothsprungmassandun-sprungmass.Figure4isthemin-bussuspensionvibrationpowerspectraldensitiesofbothsprungmassandun-sprungmasswithpassiveandadaptivesuspensionsystembyDclassroadsurface.Itisclearthatneuralnetworkscontrolimprovesperformancesofmini-bussuspensionwithmainlyimprovementsoccurringaboutsprungmassresonancepeak.Thepowerspectraldensitiesindicatethattheadaptivesuspensionsystemwithneuralnetworkscontrolcanreducemini-busvibrationgreatlycomparedwithpassivesuspension.Ifexcellentfizzycontrolrulesandrationalmodelingofshockabsorberandsuspensioncanbeobtained,theadaptivesuspensionsystemwithneuralnetworkscontrolwillimprovefartherridecomfortandroadholdingandhandlingstabilityofautomobileinthefuture.TABLEImin-bussuspensionroadtestresults:sprungmassandun-sprungmassaccelerationr.m.s.Values(Dclassroad)Speed30(1km/h)40(1m/h)50(kmlh)PassiveControlreducePassiveControlreducePassiveControlreduce|mass10.37560.325213.40.41400.344916.70.46940.396615.5masspg1.60111426610.91.89751.660312.52.34682.065212.0massIC,-4a|1-#,-t0ri-0110.1.lo1Fry-0Qgco1okaId-ela.r10f1FrcqvOFig.4.min-bussuspensionvibrationpowerspectraldensitiesofsprungmass(left)andun-sprungmass(right)withcontrolandpassive(runningspeed40km/h)V.CONCLUSIONSInthispaper,anewrecurrentneuralnetworks-orientedsuspensionmodelandneurofuzzycontrolschemesforthemini-bussuspensionsystemwereinvestigated.Upontherequirementofusing8accelerationsensors,aDSPcontrollerwithgainschedulingwasdeveloped.ConsideringthecomplexityoftheMRfluidshockabsorber,theactuatordynamicshasbeenincorporatedduringthehardware-in-the-loopsimulations.Itwasdemonstratedthattheadaptivecontrolsystemcould1798achieveacompetitivecontrolperformancebyadoptingtheneurofuzzycontrolschemesandrecurrentneuralnetworks-orientedsuspension.Becausethecontrollawdesign,thegainschedulingstrategy,andthehardware-in-the-loopsimulationmethoddevelopedinthispaperarerestrictedtoamin-bussuspensionsystemwithspecificparameters,theentirestrategycanbeextendedtoothersemi-activesystemifsuspensionparametersarechanged.Roadtestresultsshowthatneurofizzycontrollercaneffectivelyimprovemini-busridecomfortandroadholding.ItisfeasibletoemployDSPcontroltosuppresswholevehiclevibration,includinginsprungmassvibrationandun-sprungmassvibration.Theneurofuzzycontrollershowssomerobustcapabilityandcanminimizeinfluencesonsuspensionmodelparameterschanges,whichareimportantfactorstoimprovecontrolsystemperformance.REFERENCES1KanoppD.(1995)ActiveandSemi-activeVibrationIsolation,TransactionsofASME,JournalofSpecial50thAnniversaryDesignIssue,Vol.117,pp117-125.2Chantrnuwathhana,S.andPeng,H.(1999)AdaptiveRobustControlforActiveSuspension,proceedingsoftheAmericanControlConference,SanDiego,California,pp.l702-17063Yu,F.andCrolla,D.A.(1998)AnOptimalSelf-TuningControllerforActiveSuspension,VehicleSystemDynamics,vol.29,pp.51-654Zadeh,A.,Fahim,A.,andEl-Gindy,M.(1997)NeuralNetworksandFuzzyLogicApplicationstoVehicleSystem,InternationalJournalofVehicleDesign,vol.18(2),pp.132-1935WuweiChen,JamesK.MillsandLeWu,(2003)NeurofuzzyandFuzzyControlofAutomotiveSemi-ActiveSuspensions,InternationalJournalofVehicleAutonomousSystems,vol.1(2),pp.222-2361799