第2卷第2期Vol.2No.22023年4月JournalofArmyEngineeringUniversityofPLAApr.2023串行集成CNN模型的两种机械故障诊断算法周静1,孙强1,黄蔚1,张庆2,石昌友1(1.陆军工程大学通信士官学校,重庆400035;2.96721部队,四川宜宾644000)摘要:表征机械故障的数据种类和数量多样,数据质量参差不齐,故障信息的价值密度也较低,对机械故障诊断提出了挑战。由于浅层神经网络模型自学能力较弱,无法达到精确诊断故障的要求,因此对故障特征提取和故障诊断模型进行了深入研究。利用时间序列排序转换和连续小波变换,分别构建了基于时间序列图像和时频图的两种2D-CNN故障特征数据集,结合卷积神经网络(convolutionalneuralnetwork,CNN)和梯度提升决策树(gradientboostingdecisiontree,GBDT)算法的优点,提出了基于GBDT的CNN-Z和CNN-F两种故障诊断算法。通过实验,两种算法的故障诊断精度分别为94.33%和98.62%。与传统CNN算法相比较,实验结果展示了所提的两种故障诊断算法的精度更高,另外基于GBDT的CNN-F故障诊断算法的误差收敛时间为前者的三分之一,验证了所提算法的有效性和精确性。关键词:机械故障;卷积神经网络;基分类器;集成学习中图分类号:TP277DOI:10.12018/j.issn.2097-0730.20220310001TwoMechanicalFaultDiagnosisAlgorithmsBasedonSerialIntegratedConvolutionalNeuralNetworkModelZHOUJing1,SUNQiang1,HUANGWei1,ZHANGQing2,SHIChangyou1(1.CommunicationsNCOAcademy,ArmyEngineeringUniversityofPLA,Chongqing400035,China;2.Unit96721ofPLA,Yibing644000,China)Abstract:Thevarioustypesandhugeamountofdatathatcharacterizemechanicalfailures,aswellastheunevenqualityofthedataandthelowvaluedensityoffailureinformation,poseabigchallengetome-chanicalfailurediagnosis.However,theshallowneuralnetworkmodelcannotmeettherequirementsofaccuratefaultdiagnosisduetoitsweekself-learningability.Therefore,theresearchgroupconductedin-depthresearchonfaultfeatureextractionandfaultdiagnosismodels.Withtimeseriespermutationtrans-formandcontinuouswavelettransform,two2D-CNNfaultfeaturedadasetswereconstructedbasedonimagesoftimeseriespermutationandtime-frequencyimagesrespectively.Combinedwiththeadvantagesofconvolutionalneuralnetwork(CNN)andgradientboostingdecisiontree(GBDT)algorithm,twofaultdiagnosisalgorithmsofCNN-ZandCNN-...
