文献类型：期刊文献

中文题名：基于深度可分离卷积神经网络轴承剩余寿命预测

英文题名：BEARING REMAINING LIFE PREDICTION BASED ON DEEP SEPARABLE CONVOLUTIONAL NEURAL NETWORK

作者：徐海铭[1];夏乔阳[1];李勇[1];章兰珠[1]

机构：[1]华东理工大学机械与动力工程学院,上海200237

年份：2022

卷号：44

期号：4

起止页码：763

中文期刊名：机械强度

外文期刊名：Journal of Mechanical Strength

收录：CSTPCD;;Scopus;北大核心:【北大核心2020】；CSCD:【CSCD_E2021_2022】；

基金：上海市自然科学基金项目(20ZR1414400)资助。

语种：中文

中文关键词：深度可分离卷积;注意力机制;神经网络;初始故障点;剩余寿命预测

外文关键词：Deep separable convolution;Attention mechanism;Neural network;Incipient fault point(IFP);Remaining useful life(RUL)prediction

摘要：为进行轴承剩余寿命(Remaining Useful Life,RUL)预测,采用小波-谱峭度分析方法,首先对轴承振动序列信号进行小波包分解,并以谱峭度作为指标,确定故障特征频带并进行信号重构,然后,根据其频谱特征判断轴承是否发生故障,最终确定轴承振动序列信号的初始故障点(Incipient Fault Point,IFP)。在此基础上,将引入注意力(Attention)机制的一维深度可分离卷积神经网络用于轴承初始故障点之后振动信号特征的提取,相比传统卷积神经网络,深度可分离卷积层可减少网络训练参数个数,加快网络训练速度。实验结果表明,注意力机制的引入使网络能够聚焦信号中关键的特征,为重要特征赋予较大权重,避免人工处理特征的不足,利于有效特征提取,最终预测结果好于SVR、CNN、RNN等常用数据驱动方法。
In order to predict remaining useful life(RUL)of bearings,the wavelet-spectral kurtosis analysis method is used.Firstly,the bearing vibration sequence signal is decomposed by wavelet packet,the spectral kurtosis is chosen to determine the fault characteristic frequency band and reconstructed the signal.Then,determine whether the bearing is faulty according to its spectral characteristics.Lastly,the incipient fault point(IFT)of the bearing vibration sequence signal is determined.On this basis,the one-dimensional deep separable convolutional neural network with attention mechanism is used for the extraction of bearing vibration signal features.Compared with traditional convolutional neural networks,deep separable convolutional layers can reduce the number of network training parameters and speed up network training.The experimental results show that the introduction of the attention mechanism enables the network to focus on the key features in the signal,assign greater weight to important features,avoid the shortage of manual processing features,and facilitate effective feature extraction.The final prediction results are better than common data-driven methods such as SVR,CNN,and RNN.