个人简介
2007年毕业于新加坡国立大学并获博士学位，2008年加入数学院系统所统计科学研究室工作至今。主要研究方向为工业统计，侧重于系统可靠性与软件可靠性理论方法，同时面向我国民用航天质量与可靠性实际需求。主要研究工作发表于工业统计及质量可靠性领域的重要期刊，包括IEEETransactionsonReliability,IIETransactions,JournalofQualityTechnology,ReliabilityEngineeringandSystemSafety等。胡庆培博士现任中国现场统计研究会理事，中国现场统计研究会可靠性分会常务理事、中国运筹学会可靠性分会常务理事。
近期论文
查看导师新发文章
（温馨提示：请注意重名现象，建议点开原文通过作者单位确认）
Residuallifepredictionforcomplexsystemswithmulti-phasedegradationbyARMA-filteredhiddenMarkovmodel
Theperformanceofcertaincriticalcomplexsystems,suchasthepoweroutputofgroundphotovoltaic(PV)modulesorspacecraftsolararrays,exhibitsamulti-phasedegradationpatternduetotheredundantstructure.Thispatternshowsadegradationtrendwithmultiplejumppoints,whicharemixedeffectsoftwofailuremodes:asoftmodeofcontinuoussmoothdegradationandahardmodeofabruptfailure.Bothmodesneedtobemodeledjointlytopredictthesystemresiduallife.Inthispaper,anautoregressivemovingaveragemodel-filteredhiddenMarkovmodelisproposedtofitthemulti-phasedegradationdatawithunknownnumberofjumppoints,togetherwithaniterativealgorithmforparameterestimation.Thecomprehensivealgorithmiscomposedofnon-linearleast-squaremethod,recursiveextendedleast-squaremethod,andexpectation–maximizationalgorithmtohandledifferentpartsofthemodel.TheproposedmethodologyisappliedtoaspecificPVmodulesystemwithsimulatedperformancemeasurementsforitsreliabilityevaluationandresiduallifeprediction.Comprehensivestudieshavebeenconducted,andanalysisresultsshowbetterperformanceovercompetingmodelsandmoreimportantlyallthejumppointsinthesimulateddatahavebeenidentified.Also,thisalgorithmconvergesfastwithsatisfactoryparameterestimatesaccuracy,regardlessofthejumppointnumber.
AReliabilityAssessmentApproachforSystemswithHeterogeneousComponentInformation
Reliabilityassessmentofcomplexsystemsisanimportantyetdifficulttask.Thedifficultyariseslargelybecauseofheterogeneouscomponent-leveldata,e.g.,lifetimedata,degradationdata,component-levelassessmentresultandpriorinformation.Amethodforreliabilityassessmentisdevelopedinthispaperforsystemsplaguedbydegradationandlifetimedata.Ourframeworkdividescomponentlevelinformationintotwotypes.Thesetwotypesarethencombinedusingasystematicapproach.Thispaperdescribesthemethodalongwithsomeapplicationexampleswhichdemonstratetheapproachcapableofovercomingseveraldifficultiesassociatedwithconventionalreliabilityassessmentapproaches.
StrategicAllocationofTestUnitsinanAcceleratedDegradationTestPlan
Degradationisoftendenedintermsofthechangeofakeyperformancecharacteristicovertime.Itiscommontoseethattheinitialperformanceofthetestunitsvariesanditisstronglycorrelatedwiththedegradationrate.
Reliabilitydemonstrationtestforload-sharingsystemswithexponentialandWeibullcomponents
DegradationModeling,Analysis,andApplicationsonLifetimePrediction
Degradationsignalsprovidemoreinformationforproductlifestatusthanfailuredata,whenspecificdegradationmechanismcanbeidentified.Modelingandanalysiswiththedegradationsignalishelpfultoextrapolateforproductlifetimeprediction.Inthischapter,comprehensivereviewhasbeenconductedfordifferentkindsofmodelingandanalysisapproaches,togetherwiththecorrespondinglifetimepredictionresults.Furthermore,discussionsoverrelatedissueslikeproductinitialperformancearepresented.
AnApproachforReliabilityDemonstrationTestBasedonPower-LawGrowthModel
Reliabilitydemonstrationtest(RDT)isacriticalandnecessarystepbeforetheacceptanceofanindustrialsystem.Generally,aRDTfocusesondesigningatestplanthroughwhichonecanjudgewhetherthesystemreliabilityindicesmeetspecificrequirements.TherearemanyestablishedRDTplans,butfewhaveincorporatedthereliabilitygrowthaspectsofthecorrespondingproducts.Inthispaper,weexamineacomprehensivetestplanthatinvolvesinformationconcerningthereliabilitygrowthstage.AnapproachforRDTundertheassumptionofthepower-lawmodelisproposed.Itcombinesdatarelatedtothegrowthstagewiththosepertainingtotheteststageoftheproducttoreducethecostofthetest.Throughsimulationstudiesandnumericalexamples,weillustratethecharacteristicsofthetestplanandsignificantreductionintestcoststhroughourapproach.
DesignandRiskEvaluationofReliabilityDemonstrationTestforHierarchicalSystemswithMultilevelInformationAggregation
Asreliabilityrequirementsbecomeincreasinglydemandingformanyengineeringsystems,conventionalsystemreliabilitydemonstrationtesting(SRDT)basedonthenumberoffailuresdependsonalargesampleofsystemunits.However,formanysafetycriticalsystems,suchasmissiles,itisprohibitivetoperformsuchtestingwithlargesamples.Toreducethesamplesize,existingSRDTmethodsutilizetestdatafromeithersystemlevelorcomponentlevel.Inthispaper,anaggregation-basedSRDTmethodologyisproposedforhierarchicalsystemsbyutilizingmultilevelreliabilityinformationofcomponents,subsystems,andtheoverallsystem.Analyticalconditionsareidentifiedfortheproposedmethodtoachievelowerconsumerrisk.TheperformancesofdifferentSRDTdesignstrategiesareevaluatedandcomparedaccordingtotheirconsumerrisks.Anumericalcasestudyispresentedtoillustratetheproposedmethodologyanddemonstrateitsvalidityandeffectiveness.
Softwarereliabilitygrowthmodelingandanalysiswithdualfaultdetectionandcorrectionprocesses
Computersoftwareiswidelyappliedinsafety-criticalsystems.Theever-increasingcomplexityofsoftwaresystemsmakesitextremelydifficulttoensuresoftwarereliability,andthisproblemhasdrawnconsiderableattentionfrombothindustryandacademia.Mostsoftwarereliabilitymodelsarebuiltonacommonassumptionthatthedetectedfaultsareimmediatelycorrected;thus,thefaultdetectionandcorrectionprocessescanberegardedasthesameprocess.Inthisarticle,acomprehensivestudyisconductedtoanalyzethetimedependenciesbetweenthefaultdetectionandcorrectionprocesses.ThemodelparametersareestimatedusingtheMaximumLikelihoodEstimation(MLE)method,whichisbasedonanexplicitlikelihoodfunctioncombiningboththefaultdetectionandcorrectionprocesses.Numericalcasestudiesareconductedundertheproposedmodelingframework.TheobtainedresultsdemonstratethattheproposedMLEmethodcanbeappliedtomoregeneralsituationsandprovidemoreaccurateresults.Furthermore,thepredictivecapabilityoftheMLEmethodiscomparedwiththatoftheLeastSquaresEstimation(LSE)method.ThepredictionresultsindicatethattheproposedMLEmethodperformsbetterthantheLSEmethodwhenthedataarenotlargeinsizeorarecollectedintheearlyphaseofsoftwaretesting.
Ageneralmodelingandanalysisframeworkforsoftwarefaultdetectionandcorrectionprocess
Softwarereliabilitygrowthmodelingplaysanimportantroleinsoftwarereliabilityevaluation.Toincorporatemoreinformationandprovidemoreaccurateanalysis,modelingsoftwarefaultdetectionandcorrectionprocesseshasattractedwidespreadresearchattentionrecently.Inmodelingsoftwarecorrectionprocesses,theassumptionoffaultcorrectiontimeisrelaxedfromconstantdelaytorandomdelay.However,stochasticdistributionoffaultcorrectiontimebringsmoredifficultiesinmodelingandcorrespondingparameterestimation.Inthispaper,aframeworkofsoftwarereliabilitymodelscontainingbothinformationfromsoftwarefaultdetectionprocessandcorrectionprocessisstudied.Differentfrompreviousextensionsonsoftwarereliabilitygrowthmodeling,theproposedapproachisbasedonMarkovmodelotherthananonhomogeneousPoissonprocessmodel.Also,parameterestimationiscarriedoutwithweightedleast-squareestimationmethod,whichemphasizestheinfluenceoflaterdataontheprediction.Twodatasetsfrompracticalsoftwaredevelopmentprojectsareappliedwiththeproposedframework,whichshowssatisfactoryperformancewiththeresults.
Proportionalhazardmodelingforhierarchicalsystemswithmulti-levelinformationaggregation
Reliabilitymodelingofhierarchicalsystemsiscrucialfortheirhealthmanagementinmanymission-criticalindustries.Conventionalstatisticalmodelingmethodologiesareconstrainedbythelimitedavailabilityofreliabilitytestdata,especiallywhenthesystem-levelreliabilitytestsofsuchsystemsareexpensiveand/ortime-consuming.Thisarticlepresentsasemi-parametricapproachtomodelingsystem-levelreliabilitybysystematicallyandexplicitlyaggregatinglower-levelinformationofsystemelements;i.e.,componentsand/orsubsystems.AninnovativeBayesianinferenceframeworkisproposedtoimplementinformationaggregationbasedontheknownmulti-levelstructureofhierarchicalsystemsandinteractionrelationshipsamongtheircomposingelements.Numericalcasestudyresultsdemonstratetheeffectivenessoftheproposedmethod.
Lowerconfidencelimitforreliabilitybasedongroupeddatausingaquantile-fillingalgorithm
Theaimofthispaperistoproposeanapproachtoconstructinglowerconfidencelimitsforareliabilityfunctionandinvestigatetheeffectofasamplingschemeontheperformanceoftheproposedapproach.Thisisaccomplishedbyusingadata-completionalgorithmandcertainMonteCarlomethods.Thedata-completionalgorithmfillsincensoredobservationswithpseudo-completedatawhiletheMonteCarlomethodssimulateobservationsforcomplicatedpivotalquantities.TheBirnbaum–Saundersdistribution,thelognormaldistributionandtheWeibulldistributionareemployedforillustrativepurpose.Theresultsofthreecasesofdata-analysisarepresentedtovalidatetheapplicabilityandeffectivenessoftheproposedmethods.Thefirstcaseisillustratedthroughsimulateddata,andthelasttwocasesareillustratedthroughtworeal-datasets.
Studyofanimputationalgorithmfortheanalysisofinterval-censoreddata
Inthisarticle,aniterativesingle-pointimputation(SPI)algorithm,calledquantile-fillingalgorithmfortheanalysisofinterval-censoreddata,isstudied.ThisapproachcombinesthesimplicityoftheSPIandtheiterativethoughtsofmultipleimputation.Thevirtualcompletedataareimputedbyconditionalquantilesontheintervals.Thealgorithmconvergenceisbasedontheconvergenceofthemomentestimationfromthevirtualcompletedata.Simulationstudieshavebeencarriedoutandtheresultsareshownforinterval-censoreddatageneratedfromtheWeibulldistribution.FortheWeibulldistribution,completeproceduresofthealgorithmareshowninclosedforms.Furthermore,thealgorithmisapplicabletotheparameterinferencewithotherdistributions.Fromsimulationstudies,ithasbeenfoundthatthealgorithmisfeasibleandstable.Theestimationaccuracyisalsosatisfactory.
Robustrecurrentneuralnetworkmodelingforsoftwarefaultdetectionandcorrectionprediction
Softwarefaultdetectionandcorrectionprocessesarerelatedalthoughdifferent,andtheyshouldbestudiedtogether.
京公网安备 11010802027423号