基于波形分析的汽車電控系統(tǒng)故障診斷技術(shù)(文獻翻譯)

的汽車電控系統(tǒng)故障診斷技術(shù)(文獻翻譯)基于波形的汽車信號診斷及機械研究郭紅，雅閣.克羅斯曼，伊璐.墨菲，馬克.科爾曼，電機與電子工程師聯(lián)合會成員摘要斷的解決方案中采用的是小波變換。汽車發(fā)動機診斷往往涉及多個信號的分析。的信號段中，提取特征向量及訓(xùn)練集合學習診斷特性。模糊系統(tǒng)按其診斷理論，實驗結(jié)果也呈現(xiàn)出來了。1介紹“扔掉汽車部件迅速找到導(dǎo)致車輛故障的根本原因。在20世紀80診斷技術(shù)，可以分為三大類：1)車載診斷軟件和自檢過程。電子控制單元（ECU）軟件可合并自檢過程，在檢測到故障時可以存儲故障代碼。2）使用板外的診斷工具。當檢查車輛獲取診斷數(shù)據(jù)時，掃描儀或掃描工具ECU自檢，也可以記錄從主板上的車輛傳感器驅(qū)動時的連續(xù)輸出信號。3）關(guān)閉車載診斷站。這些工具結(jié)合從車輛ECU和傳感器下載數(shù)據(jù)，離板診斷在車輛上所使用的復(fù)雜傳感器。同樣，這項技術(shù)剩下的全部任務(wù)是解釋數(shù)據(jù)。CPU和信單元之間有數(shù)據(jù)連接，這也是標準所在的地方（ISO9141從ECU候，診斷技術(shù)遠遠落后于數(shù)據(jù)采集技術(shù)。知識，許多有經(jīng)驗的技術(shù)人員仍然可以找到故障。該系統(tǒng)的例子包括策略引擎（HPIDEA（菲亞特研究中心）和MDS（戴姆勒-奔信號A信號B圖1-1信號改變圖ECU確定故障車輛的狀況。然而這里討論的來自動力總成控制模塊（PCM）的信號開發(fā)的方法，這個信號系統(tǒng)足夠被用來其他多種信號故障診斷問題。PCM改變控制策輛的重量、主動配件等提供物理反饋系統(tǒng)進一步改變運行狀態(tài)。在我們的系統(tǒng)中，我們依賴動力系統(tǒng)的信號獲得到這些物理事件。例如，圖1-1展示了一個簡化的節(jié)氣門位置（TP）和每分鐘轉(zhuǎn)數(shù)（RPM）信號之間的關(guān)系。TP突然上升和下降，而RPM模仿這種行為，但更順暢。這種簡化是不完全1-1表示出同的重要信號，并且有許多信號之間存在一定的循環(huán)。PCM或的信號。當考慮的車輛因素依賴于信號的關(guān)系時，會造成不能丟失的信號信息，否則依賴于這些信息某些故障無法進行診斷。的影響，以及到目前為止我們的工作和我們未來的目標。2系統(tǒng)概述我們已經(jīng)開發(fā)出的系統(tǒng)是一個多層次的診斷系統(tǒng)（參見圖2-1有更詳細討論。第一層和第五層在這里我們僅僅做下簡要地討論。步的討論。2-2顯示了使用已被分割的該模塊的TP信號。這些分部有三個目的。首先，他們把信號1數(shù)據(jù)轉(zhuǎn)換層2分割層3特征向量建設(shè)層4超級功能載體構(gòu)建成圖2-1診斷系統(tǒng)的框圖圖2-2時間-樣本圖如圖2-2所示為分割TP信號的例子。線段在ADSAS上升表示開始加速，高穩(wěn)定段表示巡航，減速表示下降，低穩(wěn)定段表示閑置。這種分割是通過我們的系統(tǒng)自動完成的?！霸胍簟⒍蝺?nèi)的變化率或運動模式。為了提取信息，我們以一個緊湊的形式使用先進的信主信號的功能從基準信號選擇功能，以形成一個“超級特征向量。這種超級的特征矢量也轉(zhuǎn)速上升在TP類似的上升所造成的RPM的功能的超級特征向量RPM。隨著TP使用此信息來區(qū)分轉(zhuǎn)速信號、TP和那些不正常的反應(yīng)。4層將產(chǎn)生的超向量送入系統(tǒng)進行訓(xùn)“壞良好的樣本。后面三章是對2-4層比較詳細地論述。3信號分割的車輛狀態(tài)是怠速狀態(tài)、巡航、加速和減速。TP信號是進行分割的不錯選擇，因為它的行為密切關(guān)系到模仿車輛的這四種狀態(tài)。TP信號在發(fā)生信號上升的期TPTP車信號時，車輛狀態(tài)是一致的。第二節(jié)舉了一個例子，這個例子采用了典型的TP信號分割算法。使用TPTP信號與其他從相同的車輛的角的信號覆蓋用TP段,TP段演示了如何映射到其他信號。這樣問題就變成了如何對TP信號進行最佳區(qū)分。我們提出了一種基于多分辨率分析（MRA波變換系數(shù)的算法，以幫助找到區(qū)段界限。3.1基于小波變換的多分辨率分析的信號分割我們的自動分割算法基于MRA于許多其它領(lǐng)域。應(yīng)于在TP信號的上升沿和下降沿。小波函數(shù)是正確選擇，與該小波系數(shù)信號相以便非常順利地發(fā)生變化。我們實施小波變換采用快速小波變換（FWT）算法。FWT的第一階段以原始圖3-1時間-樣本圖如圖3-1所示，TP段的信號應(yīng)用到轉(zhuǎn)速信號和點火提前角信號。逼近系數(shù)同樣標記為CA中，我們選擇采用母小波DB1，因為它的細節(jié)系數(shù)表明急劇變化的信號，表示過FWT取得細節(jié)系數(shù)中使用DB1能明顯地改變對應(yīng)的波峰和波谷。3.2分割算法分割主要有我們在下面描述的四個步驟，這些步驟詳情如下：步驟1創(chuàng)建近似的區(qū)段界限。這一步在一系列的細節(jié)系數(shù)級別中選擇小波段劃分成多個較小的段進行進一步的詳細分析。步驟2合并相同的狀態(tài)區(qū)段：第1步后，一些相鄰的段可能具有相同的狀態(tài)，各段連接在一起。步驟3微調(diào)區(qū)段界限：這一步是在一個小鄰域，通常是一個或兩個樣本，步驟4再次合并相同狀態(tài)區(qū)段：步驟3中的這一過程可能會造成相鄰區(qū)段的相同狀態(tài)，所以我們重新運行第2步。忽略了更精細級別系數(shù)，我們將錯過小的變化。算法遍歷在給定具體水平的每個系數(shù)，跟蹤連續(xù)的0序列（采用β確定）狀態(tài)，而統(tǒng)稱他們?yōu)榉€(wěn)態(tài)。每次系數(shù)的值在0點位置：路徑過渡態(tài)和收縮巡航狀態(tài)。這個規(guī)則是基于這樣的事實：TP轉(zhuǎn)換往往表示車輛狀態(tài)變化的開始。有一些TP改變時，車輛反應(yīng)這種變化之間的延遲時間。在過渡態(tài)結(jié)束時的填充為車輛的反應(yīng)留下一些時間。00除了短的巡航區(qū)段，并進一步為過渡態(tài)做鋪墊。的移動保證我們所有車輛的狀態(tài)。此外，我們可以更精確地定位區(qū)段界限。這里是基本的遞推層次的級別。本的各部分。4特征提取有屬性。對于車輛診斷工程和信號處理的理論知識，我們發(fā)現(xiàn)它有以下有用功能：（1）分部狀態(tài)?？臻e0，加速度1，2，巡航或減速3。（2）區(qū)段長度。一般情況下，在一個段內(nèi)的信號特性依賴于該段的長度。例如，一個很長的加速狀態(tài)，可能會導(dǎo)致在轉(zhuǎn)速比上一個短而明顯的變化。（3）最大和最小的信號。在一個段內(nèi)的信號的最大和最小值是有價值的用于檢測的邊界條件，如TP低于其閑置的門檻值。量，直到取得令人滿意的結(jié)果（95-1005采用模糊邏輯的信號故障診斷種不確定性迫使我們，采用模糊診斷方法找到一個解決方案。5-1表示的是模糊學習組件和模糊推理成分。則和模糊歸屬函數(shù)（MSF圖5-1模糊推理過程圖的最具挑戰(zhàn)性的部分是產(chǎn)生一個有效的知識的基礎(chǔ)。6實施和實驗前面各節(jié)中所描述的算法被整合到一個單一的系統(tǒng)即高級診斷信號分析系A(chǔ)DSASADSAS描述了車輛信號診斷的兩個實驗。6.1先進的信號診斷分析系統(tǒng)ADSASWin32API的兼容Windows95/98/NT。該系統(tǒng)有兩個目標。首先，它提供了一個強大而靈活截圖是一個通用系統(tǒng)的主要窗口截圖，顯示了一個包含小波變換系數(shù)的信號。6.2實驗結(jié)果在本節(jié)中，我們描述了使用ADSAS的兩組進行了多個信號的診斷實驗。ADSAS的任務(wù)是由上述第五節(jié)所述的模糊智能系統(tǒng)根據(jù)學到的知識標記所有異常的信號段。第一組信號診斷實驗的目的是TPTP和關(guān)閉節(jié)氣門位置油門（TPCT）信號。第二組實驗中涉及檢測RPM的障礙，使用RPM與TP。GOOD（較低的值）表示正常的段，而差的值（高值）或UNKNOWN指示異常段。我們產(chǎn)生一個游泳池上面顯示的信號和11個異常節(jié)段的正常部分。測試結(jié)果的解釋如下：（1）實驗編號（2）在訓(xùn)練集的數(shù)據(jù)（3）訓(xùn)練數(shù)據(jù)的分類錯誤數(shù)（4）產(chǎn)生的模糊規(guī)則數(shù)（5）測試集合中的數(shù)據(jù)數(shù)量（6）正常段中的測試集的編號（7）異常在測試集的異常（8）誤報數(shù)（好標記為壞）（9）點火故障（壞標記為好）ADSAS相同的格式。在實驗3133對一個產(chǎn)生7729段和16段異常，它沒有產(chǎn)生任何誤報。然而，它丟失了五個壞部分。在實驗4129分類而與規(guī)則沖突。對產(chǎn)生63個模糊規(guī)則的模糊智能系統(tǒng)進行了測試，它正確地檢測到所有異常節(jié)段，并標明只有六個部分異常。驗1-3和實驗4124段，這是在工程診斷方面優(yōu)異的成績。第二組信號診斷實驗使用RPM和TP檢測故障：RPM使用TP作為參考信號。結(jié)果是下面這些實驗中的一個。TP和RPM將不再重復(fù)。在訓(xùn)練和測試集，ADSAS正確標記了所有壞的分部，并沒有產(chǎn)生任何誤報。III顯示的模糊變量和它們所示，使用相同的格式。在實驗3133模糊系統(tǒng)表示一些規(guī)則的沖突，由于少量的數(shù)據(jù)的不一致性。產(chǎn)生77個模糊規(guī)29段和16然而，它錯過了五個壞分部。在實驗4129分類的規(guī)則沖突。產(chǎn)生63個模糊規(guī)則的模糊智能系統(tǒng)。當系統(tǒng)進行了測試，正常使用49段和7個異常節(jié)段，它正確地檢測到異常節(jié)段，并標明只有六個分部正常異常。在所有實驗中，我們只有三個錯誤分類的訓(xùn)練數(shù)據(jù)。點火故障的數(shù)目是零，1-3和6在實驗4124檢測所有異常節(jié)段，這是在工程診斷優(yōu)異的成績。第二組信號診斷實驗進行檢測絆腳石使用RPM和TP：RPM的使用TP作為參考信號。結(jié)果示于下面這些實驗中的一個。TP和RPM的典型行為，在前面已進行了描述，這里將不再重復(fù)。87結(jié)論ADSAS現(xiàn)實世界中的汽車發(fā)動機診斷問題。對異常信號區(qū)段95-100％正確的區(qū)分率這ADSAS括信號展示，分割，特征提取，學習測試幾方面。雖然我們只使用信號對（TP，TPCTTPRPMADSAS的良好的試驗研究平臺。8感謝作者衷心地感謝阿爾-米爾斯和斯菈蒂齊先生，感謝他們在福特汽車公司高級汽車故障診斷與設(shè)計部中對他們項目的大力支持。AutomotiveSignalDiagnosticsUsingWaveletsandMachineLearningHongGuo,Member,IEEE,JacobA.Crossman,Member,IEEE,YiLuMurphey,SeniorMember,IEEE,andMarkColemanAbstract—Inthispaper,wedescribeanintelligentsignalanalysissystememployingthewavelettransformationinthesolutionofvehicleenginediagnosisproblems.Vehicleenginediagnosisofteninvolvesmultiplesignalanalysis.Thedevelopedsystemfirstpartitionsaleadingsignalintosmallsegmentsrepresentingphysicaleventsorstatesbasedonwaveletmulti-resolutionanalysis.Second,byapplyingthesegmentationresultoftheleadingsignaltotheothersignals,thedetailedpropertiesofeachsegment,includinginter-signalrelationships,areextractedtoformafeaturevector.Finally,afuzzyintelligentsystemisusedtolearndiagnosticfeaturesfromatrainingsetcontainingfeaturevectorsextractedfromsignalsegmentsatvariousvehiclestates.Thefuzzysystemappliesitsdiagnosticknowledgetoclassifysignalsasabnormalornormal.TheimplementationofthesystemisdescribedandexperimentresultsarepresentedI.INTRODUCTIONTODAY'Svehiclesarebecomingmoreandmorecomplexwithincreasedreliabilityonelectronicsandon-boardcomputers.Asaresult,faultdiagnosisonthesevehicleshasbecomeincreasinglychallengingwithagreaternumberofpartsandcontrollersinteractinginalargenumberofcomplexand,sometimes,poorlyunderstoodways.Correspondingly,thejobofvehiclediagnosishasbecomemoredifficult,especiallyfornonroutinefaults.Often,technicianscannotevenpinpointtherootcauseofadifficultfaultandfindthemselvesreplacingpartsinthehopethatthegivenpartisthesourceoftheproblem.This“throwingpartsatthevehicle”approachincreasescarmanufacturerwarrantycostsandleadstodissatisfiedcustomers.Therefore,carmanufacturersarefindingitnecessarytodevelopanewbreedofelectronicdiagnostictechnologythatcanhelpleadquicklytotherootcauseofavehiclefault.Duringthe1980s,therapidintroductionofelectronicenginemanagementtechniquesgreatlyimprovedtheperformanceofthevehicleengine,while,conversely,makingenginediagnosisthemostdifficultpartofvehiclediagnosis.Vehiclediagnosistoolsandtechniquescanbedividedintothreeclasses[7]:1)On-boarddiagnosticsoftwareandselftestroutines.AnElectronicControlUnit’s(ECU)softwaremayincorporateself-testroutinesthatcanstorethefaultcodewhenafaultisdetected.2)On-boarddiagnosticdataaccessedusinganoff-boarddiagnostictool.Whenavehicleisinspected,ascannerorascantoolcanbeconnectedtothediagnosticterminaloftheon-boardcomputer.ThesetoolscaneithersimplycollectfaultcodesfromtheECUselftest,ortheycanrecordcontinuoussignaloutputsfromtheon-boardvehiclesensorsduringdriving.3)Off-boarddiagnosticstations.ThesetoolscombinedatadownloadedfromthevehicleECUandsensorswithoff-boarddiagnosticsensorsmoresophisticatedthanareavailableonthevehicle.Again,thetechnicianisleftwiththefullresponsibilityofinterpretingthedataThereareseverallimitationstoon-boarddiagnostics.First,on-boardsoftwaremustbeintegratedwithvehiclespecifichardware,whichmeansdifferentvehiclescannotsharethesamesoftwareordiagnosticmethods.Second,theerrorcodesprovidedbytheon-boardsoftwaredonotprovideenoughdetailsregardingthefaulttoallowdiagnosis.Third,theknowledgestoredinthesystemisfixedunlessthemanufacturerupdatesitwithcostlyreplacements.Finally,becauseofthelimitedcomputingresourcesofavehicle(slowprocessorandlessinformationstoragespace),it’sdifficulttodomuchmorethanlimitcheckingtypediagnostics.Advancedsignalanalysistechniquessuchassignaltransformationsormachinelearningtechniquesarenotavailable.WiththerapiddevelopmentoftheCPUandsignalprocessing,off-boarddiagnostictechniquesaremorepromisingthanon-boarddiagnostics.Standardsareinplace(ISO9141)forthedatalinkfromtheon-boardcomputertotheoff-boardunit,sodatacanbecollectedfromtheECUandanalyzedoff-linebypowerfulcomputers.Unfortunately,atthistime,diagnostictechniqueslagfarbehinddatacollectiontechniques.Vehiclediagnosistechniquescanbedividedintotwoclasses:model-basedandmodel-free.Model-basedtechniquesemploymathematicalmodelsofthedynamicsofthevehiclecomponentstoanalyzethebehaviorofvehiclesystems[2],[10],[12].Whilethesemodelsmayusefulforexaminingsimplifiedversionsofeachoftheenginecomponents,wedonothaveaccuratemodelsforarealvehiclewithmanyinteractivecomponents.Model-freesystemsareknowledge-based,incorporatingprofessionalknowledgefromengineerswithoutexactinformationregardingthedetailsofsystemdynamics.Therationalebehindthisapproachisthatmanyexperiencedtechnicianscanfindfaultseventhoughtheydonothaveextensiveknowledgeofthemechanicalorelectricaldynamicsofthevehicle.ExamplesofsuchsystemincludeStrategyEngine(HP),TestBench(CarnegieGroup),IDEA(FiatResearchCenter),andMDS(Daimler-BenzResearch)[25].Fig.1.Inmultiplesignalsystems,changesinonesignalgenerallyresultinchangesinoneormoreothersignals.In(a)signalBisthesmoothedversionofsignalA,asimplerelationship.In(b)eachedgeisindicatingthefeaturethesignalatthetailendiscausinginthesignalattheheadend.Noticethateachsignaleffectsandiseffectedbymultipledifferentsignals.Inthispaper,wedescribeanoff-boardandmodel-freediagnosticsystemforidentifyingfaultyvehiclebehaviorthroughanalysisofECUsignals.ThesignalsdiscussedherecomefromthePowertrainControlModule(PCM)oftheECU,however,themethodsdevelopedaresufficientlygeneraltoallowforuseinothermulti-signalfaultdiagnosisproblems.Inatightlycoupledsystemsuchasavehiclepowertrain,inputsandoutputsfromeverycomponenteffectmostoftheothercomponentsofthesystem.Forexample,thedriverpressingthethrottlecausesanincreaseintheairflowtotheengine.ThePCMchangesthecontrolstrategymodifyingfueldeliveryandsparktiming.IncreasedfuelandairincreasesRPM,which,inturn,dramaticallychangesthebehaviorofthetransmissionandothercomponentssuchasalternatoroutput.Furthermore,therearefeedbackloopsinthesystem.Theon-boardcontrollermonitorsoutputexhaustquality,gearchangesandairflowchangesfurthermodifyingsystembehaviortokeepperformanceatamaximumandexhaustpollutionataminimum.Outsidefactorssuchasroadquality,roadgradient,vehicleweight,activeaccessories,videphysicalfeedbacktothesystemfurtheralteringbehavior.Inoursystem,wecapturethesephysicaleventsanddependenciesthroughthepowertrainsignals.Forexample,Fig.1(a)demonstratesasimplificationoftherelationshipbetweenthethrottleposition(TP)andrevolutionperminute(RPM)signals.TPmakesasuddenriseandfallwhileRPMmimicsthisbehaviorbutmoresmoothly.Thissimplificationisnotcompletelyaccuratebutdemonstratesthekeypointthatimportantphysicalrelationshipscanbeseenthroughthevehiclesignals.Fig.1(b)showsamoretypicalsetofrelationshipsbetweenfourdifferentsignals.Eachcircleisasignalandeachedgeindicatesafeaturethatthetailsignalinfluencesintheheadsignal.Theserelationshipsareoftencomplex,includefivetotendifferentimportantsignals,andhavemanycyclicdependenciesbetweensignals.Wenoteseveralimportantissuesrelatedtousingsignalstodiagnoseavehicle.First,wemustdifferentiatebetweenabadsignalandbadvehiclebehaviorreflectedinthesignal.AbadsignalisgenerallycausedbyabadPCMorabadsensor.Badvehiclebehaviorcanbecausedbyanyofanumberoffactors,physicalorelectronic.Oursystemdetectssignalfeaturesthatindicatebadvehiclebehavior,whetheritiscausedbybadelectronicpartsorphysicalfaults.Second,wenotethatnotallofthephysicaldependenciespresentintheactualvehiclecanbemodeledwithcorrespondingsignals.Forinstance,thereisnosignaltoindicateroadbumpiness,aphysicalfactorthatcaneffectvehicleand,therefore,signalbehavior.Tohandletheseunknownconditionswetrainwithvehicledatainseveralconditionswhileavoidingextremedrivingconditions(e.g.,off-roadracing).Finally,thesamesignalsarenotavailablefromallvehicles.Whenconsideringbehaviorthatdependsonsignalrelationships,thiscanleadtoaninabilitydiagnosecertainfaultsthatdependoninformationpresentinthemissingsignal.Inthispaper,wefocusondevelopingtechniquesofdecomposingmultiplesignals,diagnosticfeatureextraction,andintelligentdiagnosis.Thepaperisorganizedasfollows.InSectionIIwebrieflyintroducethediagnosticsystem.InSectionIII,anautomaticsegmentationalgorithmbasedonwaveletmulti-resolutionanalysisisintroduced.InSectionIVwediscusshowwecanprocessandcombinefeatureandsegmentinformationtoformfeaturevectorssuitableforinputintoamachinelearningsystem.SectionVdescribeshowafuzzy-basedmachinelearningsystemcanbeusedtolearngoodandbadsignalbehavior.SectionVIdescribestheimplementeddiagnosticsystemandtheencouragingexperimentalresultswe’veobtained.Finally,SectionVIIdiscussestheimpactofourworkthusfarandourfuturegoalsforthisresearch.II.SYSTEMOVERVIEWThesystemwehavedevelopedisamulti-layereddiagnosticsystem(seeFig.2).Here,wepresentabriefoverviewofeachlayeranditsgoals.Layers2,3,and4arediscussedinmoredetailinlatersections.Layers1and5arediscussedonlybriefly.Thefirstlayertranslatesthedataintoaformatsuitableforprocessing.Thislayerisrelativelysimpleandisnotdiscussedfurther.Thesecondlayerautomaticallypartitionsthesignalintosegmentsusingeitherwaveletfeaturesfromthatsignalorthesegmentsofanothersignal.Fig.3showsaTPsignalthathasbeensegmentedusingthismodule.Thesesegmentshavethreepurposes.First,theydividethesignalintoregionsthatrelatetosomephysicalvehiclestate,e.g.,accelerationoridle.Ifweknowthegeneralphysicalstateofthevehiclewecaneliminatemanypossiblefaultsandbehaviorsthatweknowcannotoccurinthegivenphysicalstate.Second,segmentationleadstoanaturalclusteringofthesignaldata.Signalbehaviorwithinagivensegmentisgenerallyverysimilartobehaviorinothersegmentsofthesamestate.Thisleadstomoreconsistenttrainingandtestdata.Third,usingthesegmentswecanisolatefaultlocationwithinasignal.Thiscanleadtoeasierfaultidentification.Finally,thesegmentationstrikesanicebalancebetweenanalyzingtheoriginalsignalasawhole,whichwouldresultinenormousamountsofsuperfluousdata,andanalyzingthesignalinonepiece,whichwouldresultinaverycomplexfeaturevector.Segmentationallowsustoexamineimportantdetailsofthesignalswithoutoverloadingthesystemwithdata.ExternalVehicledata↓︱︱︱TPsignalRPMsignalSPARKAVDsignal↓↓↓︱︱∣TPsegmentationRPMsegmentationSPARKAVDsegmentation↓↓↓∣︱︱∣TPFeatureRPMFeatureSPARKAVDFeature↓↓↓4.“super”FeaturevectorconstructionlayerFig.2.Diagramoftheproposeddiagnosticsystem.Eachlayerhandlesdifferentsignalsindividuallythusenablingthediagnosisofeachsignaltobecustomized.Time[samples-55ms]Fig.3.ExampleofsegmentationontheTPsignal.InADSAS,thesegmentlinesarecolor-codedtoindicatethebeginningofacceleration(rising),cruise(steadyhigh),deceleration(falling),andidle(steadylow)segments.Thesegmentationwasdoneautomaticallybyoursystem,ADSAS(seeexperiments).Fig.4.ExampleofafewofthefeaturesextractedfromtheTPsignalshowninFig.3.Startandendindicatethebeginningandendingsampleofthesegment.Statesare0-3indicatingidle,acceleration,cruise,anddeceleration,respectively.F:[4.00;20.0]istheenergyoftheFouriercoefficientsoverthefrequencyrange4-20Hzforthegivensegment.Minimum,maximumandaveragearecalculatedontheoriginalsignal.TheWE_DB1LncolumnsaretheenergyoftheDB1(i.e.,Haary)waveletdetailcoefficientsatthenthleveloverthegivensegment.(TakenfromADSAS—seeexperiments)Thethirdlayerextractsfeaturesfromeachsegmentandcombinesthesefeaturesintoafeaturevectorcompatiblewiththemachinelearningsystem.Thesefeaturevectorshighlightimportantaspectsofthesegmentsthatcanbeusedtodefinegoodandbadbehavior.Forexample,wemaywanttolookatthe“noise”inthesignal,therateofchangewithinasegmentorthemotionpattern.ToextractthisinformationinacompactformweuseadvancedsignalprocessingtechniquesincludingwaveletandFouriertransformsaswellasstatisticaldataregardingthesignalitself.Fromthetransformationcoefficientsandthestatisticaldatawechoose,foreachsignal,thedataelementsthatbestrepresentthegivensignal’sfeatures.Finally,wecomposeourfeaturevectorfromthisdata.Fig.4showsexamplesoffeaturevectorsextractedfromtheTPsignalshowninFig.3.Thefourthlayerintroducesthenotionoftimeandsignaldependencyintothesystem.Inthislayer,wechooseaprimarysignalforanalysisandselectasetofreferencesignalsthathavesomecausalrelationshipwithagivenprimarysignal.Wethencombinethefeaturesofourprimarysignalwithselectedfeaturesfromthereferencesignalstoforma“superfeaturevector.”Thissuperfeaturevectormayalsoincludefeaturesfromprevioussegments,thusincorporatingtimedependencyintothesystem.Forexample,asmentionedearlier,anRPMriseisgenerallycausedbyasimilarriseinTP.Furthermore,becauseofphysicalinertia,behaviorofRPMinonesegmentisusuallyverycloselytiedtothebehaviorintheimmediatelyprevioussegment.ThuswecreateasuperfeaturevectorforRPMthatcontainsRPM’sfeaturesforthecurrentandprevioussegmentalongwiththestateoftheTPsignal(increasing-acceleration,decreasing-deceleration,stable-cruise/idle,etc).AmachinelearningsystemcanusethisinformationtodifferentiatebetweenRPMsignalsthatreactnormallytoTPandthosethatdonot.Thefifthandfinallayerconsistsofamachinelearningsystem.Themachinelearningsystemistrainedtorecognizefaultsfromonetypeofsignalatatime,whichresultsinseparateknowledgebasesforeachsignaltype.Currently,weuseafuzzylearningsystem,butthislayercanbegeneralizedtoneuralnetworksorothersuitablemachinelearningsystems.EachofthesupervectorsproducedinLayer4isfedintothelearningsystemfortraining.Atthistime,thetrainingissupervisedsoweprovideatargetoutputthatthesystemattemptstomatch.Wenotethat,aswithmanycomplexdiagnosisproblems,weoftendon’thavemanyverified“bad”datasamplessowemainlytrainwithgoodsamples.ThefollowingthreesectionsdiscusseachoftheLayers2-4inmoredetail.III.SIGNALSEGMENTATIONThesignalsegmentationalgorithmwedevelopedpartitionsasignalintotimeslicesrepresentingdifferentvehiclestates.Thevehiclestatesweconsiderareidle,cruise,acceleration,anddeceleration.TheTPsignal,inparticular,isagoodcandidateforsegmentationbecauseitsbehaviorcloselymimicsthesefourstatesofthevehicle.IntheTPsignal,signalrisesoccurduringperiodsofacceleration,declinesoccurduringperiodsofdecelerationandtherelativelyflatTPsignalindicatescruiseoridle.FromTPwecanmakeagoodestimateofthevehicle’sstateatanygiventimeandpartitionthesignalintosegmentsrepresentingtimeperiodswhenthevehiclestateisconsistent.Fig.3above(SectionII)showsanexampleofatypicalTPsignalsegmentedusingoursegmentationalgorithm.UsingTPastheleadingsignal,weapplythesegmentsobtainedfromautomaticallysegmentingtheTPsignaltoothersignalsfromthesamevehiclerecordinginordertolabelthecorrespondingvehiclestatesinthosesignals.InFig.5weshowtheRPMandSPARKADV(SparkAdvance)signalsoverlaidwiththeTPsegmentsfromFig.3,demonstratinghowtheTPsegmentsmaptoothersignals.TheissuethenbecomeshowtobestsegmenttheTPsignal.Weproposeanapproachbasedonmulti-resolutionanalysis(MRA)belowthatuseswaveletcoefficientstohelpfindsegmentboundaries.A.UsingWaveletsandMulti-ResolutionAnalysisinSignalSegmentationOurautomaticsegmentationalgorithmisbasedonMRAandusesthediscretewavelettransform(DWT)[9],[18],[19]toisolatefeaturesinmultiplescales.TheDWThasbeenusedrecentlyformanyotherapplicationsincludingimagecompression[23],patternrecognition[4],speechprocessing[13],signaldetection[17],astronomy[3],andmodelestimation[5].Findingsegmentboundariesisatypeofedgedetectionproblem.Inparticular,thestatesaccelerationanddecelerationcorrespondtorisingandfallingedgesintheTPsignal,respectively.Withthecorrectchoiceofmotherwaveletfunction,thewaveletcoefficientvaluesassociatedwithasignalcanbeusedtoidentifytheseedges[18].Furthermore,thewaveletcoefficientstendtoisolatesignalfeaturesin(suchasedges)byscale.Thisallowsustotunethesegmentationtoavoidcertainedgesoccurringfromrandomnoiseorverysmoothchanges.WeimplementtheDWTusingtheFastWaveletTransformation(FWT)algorithm[18],[19].ThefirststageofFWTstartswiththeoriginalsignal.Bypassingthesignalthroughspeciallowandhighpassfiltersassociatedwiththemotherwaveletfunction,weobtaindetailwaveletcoefficientscorrespondingtothehigh-frequencydetailsofthesignalandapproximationcoefficientscorrespondingtothesignalminusthedetails.Thesetwosetsofcoefficientsarethendownsampledby2beforethealgorithmcontinuesbyrepeatingtheprocessontheapproximationcoefficientsasweshowinFig.6Time[samples-55ms]Fig.5.SegmentsfromtheTPsignalinFig.3appliedtotheRPMandSPARKADVsignals.WelabelthedetailcoefficientsasCDwhereisthelevelofdecomposition(numberofiterationsofthefilteringanddown-samplingsteps).TheapproximationcoefficientsaresimilarlylabeledasCA.Asbecomeslarger,thecorrespondingdetailcoefficients,CDindicatecoarser(largerscale)detailsinthesignal.High-frequencydetailscanbefoundinthefinedetaillevels(e.g.,1-3),whilelow-frequencydetailscanbefoundinthecourserdetaillevels(e.g.,4-6).Specificallywhatthedetailcoefficientsindicateisdependentonthemotherwaveletfunction.Inourcase,wehavechosentousethemotherwaveletDB1(DaubauchiesOne)becauseitsdetailcoefficientsindicatesharpchangesinasignalindicatingtransitionstates(accelerationordeceleration).Specifically,usingDB1sharpchangescorrespondtopeaksandvalleysinthedetailcoefficientsobtainedfromtheFWT.B.SegmentationAlgorithmThesegmentationhasfourmajorstepswedescribebelow.Detailsofeachofthesestepsfollows.Step1)Createapproximatesegmentboundaries.Thisstepusesthewaveletcoefficientsfromarange,[1,k],ofdetailcoefficientlevelstoplacesegmentboundariesveryclosetothecorrectlocationinthesignal.Weusearecursive,multi-scalealgorithmtodividelargesegmentsintomultiplesmallersegmentsforfurtherdetailedanalysis.Step2)Combinesamestatesegments:AfterStep1,someadjacentsegmentsmayhavethesamestateandthisstepjoinsthesegmentstogether.Step3)Finetunesegmentboundaries:Thissteplooksatasmallneighborhood,typicallyoneortwosamples,aroundthesegmentboundariesandshiftstheboundariestomoreoptimallocations.Thisstepalsore-movesanysteadystates(idleorcruise)thataretooshorttobesignificant.Step4)Combinesamestatesegmentsagain:TheprocessinStep3maycreateadjacentsamestatesegments,sowere-runStep2Thefirststepusesarecursiveproceduretoexamineincreasinglyfinerlevelsofdetailinthesignal.Segmentsbeginaslarge,inexactboundariesandarefine-tunedasfinerlevelsofdetailcoefficientsareexamined.Wehavefoundthatifweignorethecoarselevelcoefficients,wecannotidentifycertainsmoothchangesinthesignal.Ontheotherhand,ifweignorethefinerlevelcoefficients,wemisssmallchanges.ThefollowingparametersareusedtocontrolthesegmentationresultsEachrecursiveinstantiationofthealgorithmfocusesononesectionofthesignalrepresentedastheindexes,[begin,end),intothesignal’sdetailwaveletcoefficientsatagivenlevel.Theassumedstateofthesegmentwhenthealgorithmbegins,isalsopassedintothealgorithm.Thealgorithmiteratesthrougheverycoefficientatthegivendetaillevel,trackingconsecutivesequencesof0s(determinedusingβ),positivevaluesornegativevalues.Thesesequencesrepresentthestatescruise,deceleration,andacceleration,respectively,ifthemotherwaveletfunctionusedisDB1.WeuseSTEADY,DECEL,andACCELtodenotethethreestates,respectively.Notesthatwedonotdifferentiatebetweenidleandcruise,callingthembothSTEADY.Idlestatesaredeterminedlaterusinginformationfromanothersignal.Everytimethevalueofthecoefficientschangesbetween0,positive,andnegative,wehaveastatechangeandwelocatetheendpointofthepreviousstate.Theendpointisfoundusingthelocationandstateofasegmenttooptimizetheendpointlocationaccordingtothefollowingrule:padtransitionstatesandshrinkcruisestates.ThisruleisbasedonthefactthatTPtransitionstendtoindicatethebeginningofavehiclestatechange.ThereisalwayssometimedelaybetweenwhenTPchangesandthevehiclereactstothischange.ThepaddingattheendoftransitionstatesleavessomeroomforthereactiontimeofthevehicleSteadystatesarehandledslightlydifferentlythanbothaccelerationanddecelerationstates.Ifthevalueofacoefficientis0,weincrementacounter.Ifthenumberofconsecutive0sisoveragivenlimit,thestateisconsideredSTEADY.Thiseliminatesshortcruisesegments(e.g.,“steps”)andfurtherpadstransitionstates.Oncewehaveidentifiedastate’sbeginandendboundaries,werecursivelycalltheSegmentproceduretofurthersub-dividethesegmentasnecessary.Thisisnecessarybecausethewavelettransformtendstoisolatefeaturesofdifferentscalesatdifferentlevelsofthedetailcoefficients.Itisnotuncommonforatransitionstatetoappearinonelevelofdetailcoefficientsandbetotallyabsentinanother.Movingtofinerlevelsofcoefficientsguaranteeswefindallsignificantvehiclestates.Furthermore,wecanmorepreciselylocatesegmentboundaries.ThebasecaseoftherecursionoccursatlevelLE,here.Itindicatesthatweareoperatingontheoriginalsignalinsteadofthecoefficients.Here,wesimp