文档名：基于优化长短期记忆神经网络的IGBT寿命预测模型
摘要：为了防止绝缘栅双极型晶体管(IGBT)突发性失效而影响电力电子设备安全可靠运行,急需对IGBT剩余寿命做出精确预测,这对现有预测模型在高准确性和低不确定性方面提出了挑战.该文提出一种优化模型,该模型通过利用逐次变分模态分解(SVMD)技术来提取退化特征,并采用贝叶斯方法优化长短期记忆(LSTM)神经网络的超参数以提高预测性能.首先,该模型通过SVMD技术将退化特征数据分解为多个模态后将有用模态重构从而提取和增强退化特征;其次;利用贝叶斯优化方法通过高斯过程(GP)代理模型和期望改进(EI)采集函数对LSTM预测模型超参数实现全局寻优;最后,基于SVMD特征提取技术和贝叶斯优化LSTM网络的预测模型通过实际IGBT退化特征数据证明了模型的有效性和优越性.结果表明,所提模型与传统优化模型相比,提高了13%的寿命预测准确性,并减少了34%的预测不确定性.
Abstract：Insulatedgatebipolartransistors(IGBTs)arethecorecomponentsofpowerelectronicsystemsforconvertingandcontrollingelectricalenergy.However,thereliabilityofIGBTislowerthanexpectedduetothecomplexenvironmentandoperatingconditions,andthesuddenfailureofIGBTwillleadtounplanneddowntimeoftheentiresystem.Therefore,assessingtheremainingusefullifetime(RUL)ofIGBTwillhelpguideregularmaintenanceandreduceeconomiclosses.TopreventthesuddenfailureofIGBT,itisurgenttoaccuratelypredicttheRULofIGBT,butmostexistingmethodshavelowpredictionaccuracyandhighuncertainty.Therefore,thispaperproposesanIGBTlifepredictionmodelbasedonoptimizedlongshort-termmemory(LSTM).Startingfromthetwocoresofthedata-drivenmodel,"data"and"model"areoptimizedandupgraded,whichcaneffectivelyimprovetheaccuracyandreducetheuncertaintyofthemodelprediction.Firstly,theoriginalconditionmonitoring(CM)dataoftencontainmanycontaminateddatathatappearabnormalduetoenvironmentalinterferenceandlimitationsofmeasurementtechnology.Meanwhile,CMdatamayalsoappearabnormalwhenIGBTdevicesdegradeorfail,containingimportantinformationtocharacterizethedegradationandfailureofIGBT.Itcannotbeprocessedsimultaneouslywithcontaminateddata.TheproposedmodelextractsandenhancesdegradedfeaturesbydecomposingtheIGBTdegradeddataintomultiplemodesusingthesuccessivevariationalmodedecomposition(SVMD)techniqueandthenreconstructingtheusefulmodes.Secondly,selectingthemodel'shyperparameterswillgreatlyaffectthemodel'slearningabilityandtrainingeffect.Traditionally,theselectionofhyperparametersbytheempiricaltrial-and-errormethodhascontingencyandrandomness,seriouslyaffectingtheperformanceofthemodel.TheproposedmodelusestheBayesianoptimization(BO)methodtorealizetheglobaloptimizationofmultiplehyperparametersinthemodelthroughtheGaussianprocess(GP)proxymodelandexpectationimprovement(EI)acquisitionfunction.Finally,theeffectivenessandsuperiorityoftheLSTMpredictionmodelbasedonSVMDandBOareverifiedwithrealdata.TheresultsshowthatthepredictedRULisnotclosetotherealRULbytheBO+LSTMmethodandcannotevenmeetthe30%errorrequirementatCMis160cycles.Incontrast,theerrorsoftheconventionalLSTMandRNNmethodsarelarge,whilethepredictedRULerrorsusingtheproposedmodelmeettherequirementsforallCMcycles.Inaddition,theevaluationoftheoverallperformanceofthemodelshowsthatasanimprovementontheRNN,theaveragerelativeaccuracy(YARA)oftheLSTMmethodimprovesfrom34.65%ofRNNto50.53%,andtheaveragewidthofpredictioninterval(WAPI)reducesfrom365.3cyclesto272cycles.Incomparison,theBO+LSTMmethodhasabetterpredictionperformance.Furthermore,theYARAoftheproposedmodelimprovesto90.91%,andtheWAPIdecreasesto169.3cycles,whichisthebestperformanceamongseveralmodels.Quantitativeanalysisshowsthattheproposedmodelimprovesthelifetimepredictionaccuracyby13%andreducesthepredictionuncertaintyby34%comparedtotheBO+LSTMmodel.Theconclusionscanbedrawn:(1)TheBOalgorithmisusedtooptimizethehyperparametersoftheLSTM,whichimprovesthepredictionaccuracyofthemodel.(2)TheSVMDisusedtoextractthedegradedfeaturesoftheIGBT,whichreducestheuncertaintyandimprovestheaccuracyofthemodelprediction.(3)Comparedwithothermodels,theproposedmodelcanmaintainahighpredictionaccuracywithlessCMdata,anditslong-termpredictionperformanceisbetter.
作者：任宏宇 余瑶怡 杜雄 刘俊良 周君洁Author：RenHongyu YuYaoyi DuXiong LiuJunliang ZhouJunjie
作者单位：输配电装备及系统安全与新技术国家重点实验室(重庆大学)重庆400044
刊名：电工技术学报
Journal：TransactionsofChinaElectrotechnicalSociety
年，卷(期)：2024, 39(4)
分类号：TN322.8TM46
关键词：IGBT 可靠性 寿命预测 模态分解 失效分布
Keywords：IGBT reliability lifetimeprediction modedecomposition failuredistribution
机标分类号：TP391TP183TM73
在线出版日期：2024年3月5日
基金项目：国家自然科学基金，中央高校基本科研业务费专项资金资助项目基于优化长短期记忆神经网络的IGBT寿命预测模型[
期刊论文] 电工技术学报--2024, 39(4)任宏宇 余瑶怡 杜雄 刘俊良 周君洁为了防止绝缘栅双极型晶体管(IGBT)突发性失效而影响电力电子设备安全可靠运行,急需对IGBT剩余寿命做出精确预测,这对现有预测模型在高准确性和低不确定性方面提出了挑战.该文提出一种优化模型,该模型通过利用逐次变分...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
基于优化长短期记忆神经网络的IGBT寿命预测模型 IGBT Lifetime Prediction Model Based on Optimized Long Short-Term Memory Neural Network
基于优化长短期记忆神经网络的IGBT寿命预测模型.pdf
- 文件大小:
- 1.56 MB
- 下载次数:
- 60
-
高速下载
|
|
