第47卷第4期电网技术Vol.47No.42023年4月PowerSystemTechnologyApr.2023文章编号：1000-3673（2023）04-1478-12中图分类号：TM721文献标志码：A学科代码：470·40基于格拉姆角场变换和深度压缩模型的变压器故障识别方法刘志坚，何蔚，刘航，谢静，陶韵旭，张德春（昆明理工大学电力工程学院，云南省昆明市650000）FaultIdentificationMethodforPowerTransformerBasedonGramianAngularFieldTransformationandDeepCompressionModelLIUZhijian,HEWei,LIUHang,XIEJing,TAOYunxu,ZHANGDechun(FacultyofElectricPowerEngineering,KunmingUniversityofScienceandTechnology,Kunming650000,YunnanProvince,China)1ABSTRACT:ThispaperpresentedafaultidentificationmethodforpowertransformerbasedonGramianangularfieldtransformationanddeepcompressionmodel.Aimingattheproblemsthatone-dimensionalfaultsampleswerescarceandcouldnotbedirectlyinputintothevisualgeometrygroup(VGG)network.Firstly,theGramianangularfieldtransformationmethodwasproposedtoconvertthefaultsamplesintothree-dimensionalfeatureimages.Secondly,thenumberoffeatureimageswasexpandedbydataaugmentationmethodtomeettheinputrequirementsoffaultidentificationmethod.Thirdly,tosolvetheshortcomingsoftheVGGnetworksuchasdeeplayers,toomanyparametersandthecomplexstructure,animproveddeepcompressionmodelwasproposed.Theglobalaveragepoolinglayersofnetworkinnetwork(NiNNet)wereappliedtoreplacethefullyconnectedlayersoftheVGGnetworktogreatlyreducethenumberoflayersandparameters.AstructuredpruningmethodwasproposedtoprunethemultilayerconvolutionalkernelsoftheVGGfront-endnetworktofurtherreducetheparameters,andthedeepcompressionofthenetworkwaseventuallyachieved.ThenumericalexperimentsandperformanceevaluationresultsimplementedontransformeroilchromatographyfaultdataillustratethattheproposedmethodachievesdeepcompressionandstructuralsimplificationoftheVGGnetworkwithoutlosingtheaccuracyoffaultidentificationresults.Inaddition,thedeepcompressionmodeleffectivelyreducestherequiredstoragespacesandcomputingresources,makingitapplicable基金项目：云南省基础研究计划青年项目(202201AU070086)；云南省教育厅科学研究基...
