不均衡小样本下多特征优化选择的生命体触电故障识别方法
文档名：不均衡小样本下多特征优化选择的生命体触电故障识别方法
摘要：针对现有的剩余电流保护装置无法有效识别触电事故的问题,该文提出了一种不均衡小样本下多特征优化选择的生命体触电故障识别方法.首先通过变分自编码器(VAE)对实验收集到的生命体触电小样本数据进行增殖以实现正负样本均衡;然后在时域上提取能够反映波形动态变化特性的23个特征量,并利用高斯核Fisher判别分析(GKFDA)与最大信息系数(MIC)法从中选择最优表达特征组;最后,提出基于遗忘因子的在线顺序极限学习机(FOS-ELM)算法实现生命体触电行为的鉴别.实验结果表明,所提方法利用不均衡小样本触电数据集就可以训练出一个优秀的分类模型,诊断准确率可达98.75%,诊断时间仅为1.33ms.其优良的性能结合在线增量式学习分类器设计,使得模型具备新知识学习能力,具有极好的工程应用前景.
Abstract：Theexistingresidualcurrentdevice(RCD)operatesbasedontheamplitudeoftheresidualcurrent,butifthethresholdisnotreasonablyset,theRCDispronetorejectormisoperate.Therefore,identifyingbiologicalelectric-shockfaultsfromgroundingfaultsisacrucialapproach.Currentresearchonlyselectsoneorseveralfeatureswithoutfollowingproperfeatureselectionrules.Furthermore,machinelearningmethodsrequireacertainnumberofsamplestotrainthemodeltoensurealgorithmaccuracyandstability.However,obtainingalargenumberofbiologicalelectric-shocksamplesischallengingduringactualexperiments,andthealgorithmmodelcannotlearnthewaveforminrealsettings.Tosolvetheaboveproblems,abiologicalelectric-shockfaultidentificationmethodbasedonmulti-featureoptimizationselectionunderunbalancedsmallsamplesisproposed.Firstly,variationalauto-encoders(VAE)isadoptedtomultiplytheelectric-shocksmallsampledatacollectedbyexperimentstoachievepositiveandnegativesamplebalance.Duetothecomplexityanddangerofthescenes,itisdifficulttoobtaintheactualelectric-shocksamples.Theproblemofsmallsampleswillleadtolowaccuracyandpooreffectivenessofthetrainingmodel,andtheunbalancedsampleswillleadtodeviationsinthepredictionresultsofthemodel,resultinginpooridentificationaccuracyofafewtypesofsamples.Therefore,afewsamplesareenhancedbyintroducingVAEtoimprovetheeffectivenessofthemodel.Secondly,23featureswhichcanreflectthedynamiccharacteristicsofthewaveformareextractedintimedomain,theoptimalexpressionfeaturegroupisselectedfromthembyGaussiankernelFisherdiscriminantanalysis(GKFDA)andmaximalinformationcoefficient(MIC).Throughdataanalysis,variousindexfeaturescanbeextractedfromthechangingformsofbiologicalelectric-shockwaveforms.Theadditionofhigh-qualityfeatureswillimprovethediagnosticaccuracyoftheclassifiertoacertainextent,buttheintroductionofbadandredundantfeatureswillincreasetherunningtimeofthealgorithmandreducethediagnosticaccuracyoftheclassifier.Therefore,GKFDAandMICarecombinedtoperformfeaturescoringforeachfeature,andtheoptimalexpressionfeaturegroupisselectedintuitivelyandindependentlybasedonthescoringresults,whichcouldimprovethefeaturequalityandreflecttheregularityoffeatureselection.Finally,aforgetting-factor-basedonlinesequentialextremelearningmachine(FOS-ELM)algorithmisinvestigatedtoidentifytheelectric-shockbehavior.Thereareabundantelectric-shockscenesintherealenvironments.Theescapebehaviorsoflivingobjectsduringelectricshockwillhaveagreatinfluenceontheelectric-shockwaveform,whichmakesitdifficultforthetraditionaloff-lineclassifiertohaveadaptability.Theonlinesequentialextremelearningmachine(OS-ELM)hasanonlinelearningmechanismthatallowsonlineupdatesfornewsampleswithoutthehistoricaldata.TheforgettingfactorisintroducedtoformFOS-ELM,aimingtofurthersolvetheshortcomingofslowlearningspeedofOS-ELM,sothatitcanquicklyadapttochangesofenvironmentalsampleswithhigherlearningefficiency.Theexperimentaldataofconventionalgroundingfaultandbiologicalelectric-shockfaultin12sceneswerecollectedfortheverificationoftheproposedalgorithm.Theresultsshowthatthediagnosisaccuracyoftheproposedmodelcanreach98.75%,amongwhichall40conventionalgroundingfaultsamplesarecorrectlyjudgedwithanaccuracyof100%,whileonly1of40actualbiologicalelectric-shockfaultsamplesiswrongwithanaccuracyof97.5%.Fromtheperspectiveoftime,theaverageonlinelearningtimeis1.378ms,andtheaveragediagnosistimeisonly1.33ms.
作者：高伟 饶俊民 全圣鑫 郭谋发 Author：GaoWei RaoJunmin QuanShengxin GuoMoufa
作者单位：福州大学电气工程与自动化学院福州350108;智能配电网装备福建省高校工程研究中心福州350108福州大学电气工程与自动化学院福州350108
刊名：电工技术学报
Journal：TransactionsofChinaElectrotechnicalSociety
年，卷(期)：2024, 39(7)
分类号：TM773
关键词：剩余电流保护装置生命体触电故障多特征优化选择基于遗忘因子的在线顺序极限学习机(FOS-ELM)不均衡小样本
Keywords：Residualcurrentprotectiondevicebiologicalelectric-shockfaultmulti-featureoptimizationselectionforgetting-factor-basedonlinesequentialextremelearningmachine(FOS-ELM)unbalancedsmallsample
机标分类号：TP391.6TP183TG315.4
在线出版日期：2024年4月12日
基金项目：福建省自然科学基金资助项目不均衡小样本下多特征优化选择的生命体触电故障识别方法[
期刊论文]电工技术学报--2024, 39(7)高伟饶俊民全圣鑫郭谋发针对现有的剩余电流保护装置无法有效识别触电事故的问题,该文提出了一种不均衡小样本下多特征优化选择的生命体触电故障识别方法.首先通过变分自编码器(VAE)对实验收集到的生命体触电小样本数据进行增殖以实现正负样本...参考文献和引证文献
参考文献
引证文献
本文读者也读过
相似文献
相关博文
不均衡小样本下多特征优化选择的生命体触电故障识别方法Biological Electric-Shock Fault Identification Method Based on Multi-Feature Optimization Selection under Unbalanced Small Sample
不均衡小样本下多特征优化选择的生命体触电故障识别方法.pdf
页:
[1]