ConvNeXt网络及Stacked BiLSTM-Self-Attention在轴承剩余寿命预测中的应用  

Remaining useful life prediction of bearings based on ConvNeXt network and Stacked BiLSTM-Self-Attention mechanism

作　　者：张印文 王琳霖[1] 薛文科 梁文婕 ZHANG Yinwe;WANG Linlin;XUE Wenke;LIANG Wenjie(School of Artificial Intelligence,Shenyang Aerospace University,Shenyang 110000,China)

机构地区：[1]沈阳航空航天大学人工智能学院,辽宁沈阳110000

出　　处：《机电工程》2024年第11期1977-1985,1994,共10页Journal of Mechanical & Electrical Engineering

基　　金：辽宁省科技厅重点研发计划项目(2022JH2/101300211);沈阳市科技局项目(20230152)。

摘　　要：在滚动轴承剩余使用寿命预测方面,采用传统方法时存在鲁棒性差、精度低等各种问题。近些年来深度学习的发展为解决这些问题提供了新的思路。为了进一步提高对轴承寿命的预测精度,提出了一种基于ConvNeXt网络、堆叠双向长短时记忆网络(SBiLSTM)和自注意力机制(Self-Attention)的滚动轴承寿命预测方法。首先,采用连续小波变换(CWT)构造了振动信号的时频图,以更好地捕捉信号的时域和频域特征;然后,将得到的时频图输入到构建的ConvNeXt网络中,通过卷积、池化和层归一化等操作,对时频图的关键特征进行了提取;最后,将提取后的特征输入到SBiLSTM-Self-Attention模块中,进一步提取了时序信息和特征权重分配数据,利用PHM2012挑战数据集进行了验证,通过实验分析了该方法的均方根误差(RMSE)和平均绝对误差(MAE)。研究结果表明:相较于现有技术方法,该方法的平均RMSE为0.031;与其他三种方法,即卷积神经网络(CNN)、深度残差双向门控循环单元(DRN-BiGRU)和深度卷积自注意力双向门控循环单元(DCNN-Self-Attention-BiGRU)相比,其平均RMSE值分别下降了79%、74%和55%,MAE值分别下降了78%、73%和53%,说明该方法在滚动轴承剩余寿命预测中有较好的性能。Traditional methods have various problems such as poor robustness and low accuracy in the remaining service life of rolling bearings.In recent years,the development of deep learning has provided new ideas for solving these problems.In order to further improve the accuracy of predicting bearing life,a rolling bearing life prediction method based on ConvNeXt network,stacked bidirectional long shortterm memory network(SBiLSTM)and self-attention mechanism(Self-Attention)was proposed.Firstly,continuous wavelet transform(CWT)was used to construct the time-frequency map of the vibration signal,in order to better capture the time-domain and frequency-domain characteristics of the signal.Then,the obtained time-frequency map was input into the constructed ConvNeXt network,and key features of the time-frequency map were extracted through operations such as convolution,pooling,and layer normalization.Finally,the extracted features were input into the SBiLSTM-Self-Attention module for further extraction of temporal information and feature weight allocation.The PHM2012 challenge data set was used for experimental verification.The root means square error(RMSE)and mean absolute error(MAE)of the proposed method were experimentally analyzed.The results show that,comparing with existing technical methods,the average RMSE of this method is 0.031.Comparing with the other three comparison methods,convolutional neural network(CNN),deep residual networkbidirectional gated recurrent unit(DRN-BiGRU)and deep convolutional neural network-self attention-bidirectional gated recurrent unit(DCNN-Self Attention-BiGRU),its average RMSE values respectively decrease by 79%,74%and 55%,the average MAE values respectively decrease by 78%,73%and 53%.This method has achieved good performance in predicting the remaining life of rolling bearings.

关 键 词：滚动轴承 剩余寿命预测 ConvNeXt网络 堆叠双向长短时记忆网络 自注意力机制 深度学习 连续小波变换 

分 类 号：TH133.3[机械工程—机械制造及自动化]