Page 107 - 2023年第54卷第2期
P. 107
[22] GUJ,WANGZ,KUEN J,etal.Recentadvancesinconvolutionalneuralnetworks[J].PatternRecognition,
2018,77:354 - 377.
[23] ALEXKRIZHEVSKY,ILYASUTSKEVER,GEOFFREYEHINTON.ImageNetclassificationwithdeepconvolu
tionalneuralnetworks[J].CommunicationsoftheACM,2017,60(6):84 - 90
[24] 卞景艺,刘秀丽,徐小力,等.基于多尺度深度卷积神经网络的故障诊断方法 [J].振 动 与 冲 击,2021,
40(18):204 - 211.
[25] CHETC,DUANH F,LEEPJ.Transientwave - basedmethodsforanomalydetectioninfluidpipes:Areview
[J].MechanicalSystemsandSignalProcessing,2021,160:107874.
[26] BRUNONEB,KARNEYBW,MECARELLIM,etal.Velocityprofilesandunsteadypipefrictionintransient
flow [J].JournalofWaterResourcesPlanningandManagement,2000,126(4):236 - 244.
'
[27] V?TKOVSKYJP,BERGANTA,SIMPSONAR,etal.Systematicevaluationofone - dimensionalunsteadyfric
tionmodelsinsimplepipelines[J].JournalofHydraulicEngineering,2006,132(7):696 - 708.
[28] 郭新蕾,杨开林,郭永鑫,等.管道系统泄漏检 测 的 瞬 变 水 击 压 力 波 法 [J].应 用 基 础 与 工 程 科 学 学 报,
2011,19(1):20 - 28.
[29] GUOX,YANGK,GUOY.Leakdetectioninpipelinesbyexclusivelyfrequencydomainmethod[J].SCIENCE
CHINATechnologicalSciences.2012 ,55(3):743 - 752.
[30] WANGX,LINJ,KERAMATA,etal.Matched - fieldprocessingforleaklocalizationinaviscoelasticpipe:An
experimentalstudy [J].MechanicalSystemsandSignalProcessing,2019,124:459 - 478.
Multi - scaleconvolutionalneuralnetworkmodelforpipelineleakdetection
TANZhen,GUOXinlei,LIJiazhen,GUOYongxin,PANJiajia
(StateKeyLaboratoryofSimulationandRegulationofWaterCycleinRiverBasin,
ChinaInstituteofWaterResourcesandHydropowerResearch,Beijing 100038,China)
Abstract:Howtodetectpipelineleakseffectivelyisoneofthekeyandhotresearchtopicstobesolvedinthecon
structionofwater - savingsociety.Recently ,pipelineleakdetectionmethodsbasedondeeplearninghavebeende
velopedrapidly.Thispaperproposesapipelineleakdetectionmodelforpipelinesystemsbasedonmulti - scaleone -
dimensionalconvolutionalneuralnetwork(MS1DCNN)fortheproblemofinadequateextractionofleakfeaturesby
traditionalsingle - scaleconvolutionalneuralnetwork.Themodeladoptsmultipleconvolutionpathswithdifferent
convolutionkernelsizestoextractthefeaturesofpipelineleaksandtoclassifytheleakinformation.Basedonthe
classicalpipelinesystemlayout ,inordertovalidatethemodel,thetransientflowmodelwasusedtogeneratethree
waterpressuredatasetswhichwereusedtoidentifytheleaklocation ,leakageandunsteadyfrictioncoefficientre
spectivelyundervariouspipelineleakcases.Thesamplenumberswere39601 ,3980and4900.Themodelhas
beencomparedwithotherdeeplearningmethodsortraditionalmachinelearningmethodsliketraditionalone - di
mensionalconvolutionalneuralnetwork(1DCNN),BPneuralnetwork,supportvectormachine(SVM)andk -
NearestNeighbor (KNN).TheresultsshowthattheMS1DCNNmodelachieves99.96%,98.49% and100% of
theclassificationaccuracyforleaklocation ,leakageandunsteadyfrictioncoefficientunderthedatasets.Theaver
agepredictionaccuracyis0.31%,2.00%,1.27% and22.80% higherthanothermodels,respectively.Theaver
ageF 1 scoresofMS1DCNNmodelforeachdatasetsare99.2%,97.02% and100% respectivelyinthenoiseenvi
ronmentwithSNRof - 4 - 12dB.Theyare0.61%,2.3%,2.78% and28.59% higherthanthoseof1DCNN,BP
neuralnetwork ,SVM andKNN,respectively,whichprovesthepredictionperformanceofMS1DCNNmodel.The
modelpresentedinthispaperisapplicableforthesynchronouspredictionofpipelineleakageparametersandun
steadyfrictioncoefficient.
Keywords:pipelineleakdetection;transientflowmodel;unsteadyfrictioncoefficient;multi - scaleleakfeature;
convolutionalneuralnetwork
(责任编辑:李福田)
— 2 3 1 —