一种基于TC-CAE的轴承寿命预测方法  

Bearing life prediction based on the method of TC-CAE

作　　者：李海浪 刘永志 邹益胜[1] 刘彦涛 宋小欣[1] LI Hailang;LIU Yongzhi;ZOU Yisheng;LIU Yantao;SONG Xiaoxin(School of Mechanical Engineering,Southwest Jiaotong University,Chengdu 610031,China)

机构地区：[1]西南交通大学机械工程学院,成都610031

出　　处：《振动与冲击》2022年第14期105-113,189,共10页Journal of Vibration and Shock

基　　金：国家重点研发计划资助项目(2020YFB1708000)。

摘　　要：在预测轴承的剩余使用寿命(remaining useful life,RUL)时,能否有效提取退化特征是实现准确预测的关键之一。轴承的个体异质性和工况差异性导致退化特征曲线不同,同一特征的变化趋势在不同轴承上是具有差异的,从而导致训练轴承建立的RUL预测模型与测试轴承不匹配。在提取特征时应当考虑轴承的个体差异性,减少轴承特征的个体差异性有利于提升预测精度。为了促进同一特征在不同轴承上的趋势一致性,减少退化特征的轴承个体差异性,提出了一种基于趋势一致性约束卷积编码(trend consistency convolutional auto-encoder,TC-CAE)的轴承寿命预测方法。通过构造趋势一致性约束,并和卷积自编码相结合,形成了TC-CAE特征提取模型。预测流程为先用TC-CAE模型在频域信号内提取特征,再用长短期记忆网络(long short-term memory,LSTM)预测。在一个轴承公开数据集上进行试验,试验结果表明,相比于普通卷积自编码方法的预测结果,该方法的综合平均误差降低了21.1%,相比于特征评价方法和卷积神经网络方法分别降低了35.6%和25.9%。When predicting the remaining useful life(RUL)of a bearing,the ability to effectively extract the degradation features is one of the keys to achieve accurate prediction.The individual heterogeneity of the bearing and the difference in operating conditions lead to different degradation characteristic curves,and the change trend of the same characteristic is different on different bearings,which leads to the mismatch between the RUL prediction model established according to the training bearing and the test bearing.The individual variability of the bearing should be considered when extracting features,so,reducing the individual variability of bearing features is beneficial to improve the prediction accuracy.In order to promote the trend consistency of the same feature on different bearings and reduce the individual bearing differences in degraded features,a bearing life prediction method based on trend consistency convolutional auto-encoder(TC-CAE)was proposed.By constructing trend consistency constraints and combining with convolutional autoencoding,a TC-CAE feature extraction model was presented.The prediction process is using the TC-CAE model to extract features of the signal in frequency domain,and then using long short-term memory(LSTM)to predict.The experimental results on a bearing public data set show that:compared with the prediction results of the ordinary convolutional autoencoding method,the comprehensive average error of the method is reduced by 21.1%,and compared with the feature evaluation method and the convolutional neural network method,it is reduced by 35.6%and 25.9%respectively.

关 键 词：轴承 寿命预测 特征提取 趋势一致性 

分 类 号：TH17[机械工程—机械制造及自动化]