AcousticSourceLocalizationviaPredictingPropagatedWavesonComplexSystemsusingTimeReversalandDeepNeuralNetworks

Abstract

Anovelmethodforthelocalizationtoidentifyacousticsourceswasproposedbyutilizingtimereversalfordispersivewavesanddeepneuralnetworksbasedonthewaveprediction.Thestructuralvibrationsincomplexstructureswithmulti-linkedpathswerepredictedusingthewaveapproachafterconsideringdispersivepropagationcharacteristics.Inordertoconsiderlinkedconditionsbetweenmultiplepaths,thevibrationsinthecomplexstructureswereanalyzedaswavescoupledbothintransverseandtorsionaldirections.Thecouplingeffectsondispersivevibrationswasverifiedbysimplebeammodels.Thenumericalprocedurefortimereversalwasproposedtoidentifytheimpactlocationinthecomplexstructureshavingarbitrarypaths.Theproposedmethodwasappliedforexperimentsinanactualvehiclestructure.Thelocationsoftherattlesourceswereidentifiedfromthefocusedpointofflexuralvibrationsanalyzedbytheproposednumericaltimereversalprocedure.Giventhecoupledwaveapproach,theproposedmethodwasappliedforlocalizationofdispersivesignalsinarbitrarycomplexstructures. Thedeepconvolutionalneuralnetworks(DCNNs)wereproposedthroughthefeatureconstructionforthelocalizationandthesimulativelearningtechnique.Thefeaturesfordetectingthelocationofsourceswereextractedbyperformingacross-cepstralanalysisandimage-mappingprocess.Groupsofcomplexcross-cepstrumswerecalculatedbyusingwavesmeasuredbythreecloselyspacedtransducers.Theproposedfeatureconstructionallowstoclassifyingthesourcelocationsregardlessofspectraldensityofthesources.Eachgroupwastransformedintored,green,andblue(RGB)channelsbypixelmapping.Theimagepatternswereinfluencedbythesourcelocation.AsimulativelearningtechniquewasproposedinthisstudyandpresentedtotraintheDCNNwithoutrepetitiveexperiments.InordertogeneratethelearningdatafortheDCNN,thepropagatedwaveswerepredictedforvarioussourcelocationsandconditions.Themethodwasverifiedbyperformingexperimentsinananechoicroom.ThemappedimagesofthemeasuredacousticwaveswereclassifiedbyusingtheDCNNtodetectthelocationoftheacousticsources.Thesourceswereaccuratelydeterminedbyusingonlysmallmicrophonesirrespectiveofthetypeofacousticsourceandwithreducedeffectsfromthebackgroundnoise. TheproposedDCNNforlocalizationofthesourcesonthecomplexstructureswaspresented.TheDCNNforthecomplexstructureswasachievedthroughthecoupledwavepredictionandfeatureconstructionfortheflexuralwaves.ThefeaturesfortheDCNNwerecomprisedbyutilizingmeasuredvibrations.Themodifiedcepstralanalysisusingenvelopsofthespectrumswaspresentedinordertoextractthefeaturesforthesourcelocations.ThedatafortrainingoftheDCNNweregeneratedbychangingconditionsofsourcesandstructuresinthecoupledwaveprediction.Thefourchangedconditionswerecorrespondedtolocationandspectraldensityofsources,thicknessandwidthofstructures.Thenetworksweretrainedbythegenerateddata,andthesourcelocationswereidentifiedfor150classes.Theproposednetworkswereverifiedbythevibrationmeasurementsontheactualvehicle.Thelocationsofsourcesinthecomplexstructurewerepredictedthroughthenetworkswithonlysimplecomputationalcosts.Thedeepneuralnetworkwiththesimulativelearningtechniqueandthefeatureconstructionsprovidestheoptimalperformanceforlocalizationofsourcesonthecomplexstructuresevenundersparsedatafortraining.Giventhewaveapproachforthestructures,theproposedmethodmakesitpossibletoeasilyadoptvariouscomplexsystems.