基金项目:中石油科技中青年创新基金项目(编号:05E7040)收稿日期:2022-11-22基于改进OTSU-CNN的轴承智能故障诊断*张伟,鲍泽富※,李寿香,徐浩,张迪(西安石油大学机械工程学院,西安710065)摘要:针对传统故障诊断方法在小样本数据集下诊断准确率低且故障特征提取难的问题,提出了一种改进大津阈值分割算法(OTSU)和卷积神经网络(CNN)相结合的智能故障诊断方法。首先,对采集到的振动信号进行希尔伯特变换(Hillbert)得到信号的包络谱,同时使用小波变换对包络谱信号处理,获取二维特征时频图;其次,建立最大类间方差目标函数模型,通过算术优化算法(AOA)得到时频图的最佳分割阈值,再将变换后的阈值分割图像作为CNN的输入得到最优训练模型,最后得到分类结果。试验结果表明:所提方法相比于传统OTSU方法,所提取的故障特征更为突出,为模型提供了优秀的训练样本;在较小数据样本条件下,所提方法的准确率达99.01%,远高于传统故障诊断方法,且模型有着良好的泛化能力。关键词:OTSU;故障特征提取;卷积神经网络;时频图中图分类号:TH133.33文献标志码:A文章编号:1009-9492(2023)03-0222-06IntelligentFaultDiagnosisofBearingBasedonImprovedOTSU-CNNZhangWei,BaoZefu※,LiShouxiang,XuHao,ZhangDi(SchoolofMechanicalEngineering,Xi′anShiyouUniversity,Xi′an710065,China)Abstract:Aimingattheproblemthatthetraditionalfaultdiagnosismethodhaslowdiagnosisaccuracyanddifficultfaultfeatureextractionundersmallsampledatasets,anintelligentfaultdiagnosismethodcombiningtheOtsuthresholdsegmentationalgorithm(OTSU)andtheConvolutionalNeuralNetwork(CNN)wasproposed.Firstly,theHillberttransformwasperformedonthecollectedvibrationsignaltoobtaintheenvelopespectrumofthesignal,andthewavelettransformwasusedtoprocesstheenvelopespectrumsignaltoobtainatwo-dimensionalfeaturetime-frequencymap.Secondly,themaximumbetween-classvarianceobjectivefunctionmodelwasestablished,theoptimalsegmentationthresholdofthetime-frequencygraphwasobtainedbythearithmeticoptimizationalgorithm(AOA),andthenthetransformedthresholdsegmentationimagewasusedastheinputofCNNtoobtaintheoptimaltrainingmodel,andfinallytheclassificationresultwasobtained.Theexperimentalresultsshowthatcomparedw...