基于注意力机制的滚动轴承剩余使用寿命预测方法  

Remaining useful life prediction of rolling bearings based on attention mechanism

作　　者：卢瑾 张永平[1] LU Jin;ZHANG Yong-ping(School of Information Engineering,Yancheng Institute of Technology,Yancheng 224002,China)

机构地区：[1]盐城工学院信息工程学院,江苏盐城224002

出　　处：《机电工程》2023年第4期516-521,551,共7页Journal of Mechanical & Electrical Engineering

基　　金：国家自然科学基金资助项目(62076215);江苏省产学研合作项目(BY2022502)。

摘　　要：现有的轴承振动信号特征的提取方法过分依赖于专家的经验,同时在轴承的寿命预测过程中,存在因序列过长而导致的记忆力退化等问题,为此,结合卷积神经网络-注意力机制网络(CNN-attention)和基于注意力机制的Encoder-Decoder方法,提出了一种滚动轴承剩余使用寿命(RUL)的预测模型(方法)。首先,利用快速傅里叶变换(FFT)方法,将滚动轴承的初始振动信号转换成频域幅值信号;然后,设计了一种基于注意力机制的模型:其中,利用CNN-attention进行了退化特征提取,利用基于注意力机制的Encoder-Decoder网络进行了RUL预测,并进一步在远距离信号传输中解决了循环神经网络记忆衰退的问题;最后,为了验证特征提取模型以及寿命预测模型的有效性,采用PHM 2012轴承退化数据集,通过轴承加速退化PRONOSTIA实验平台进行了实验,并将其所得结果与未采用注意力机制模型的预测结果以及其他文献方法所得结果进行了对比。实验结果表明:与其他方法相比,基于注意力机制模型的方法平均绝对误差分别降低了29.41%、32.00%、29.56%、32.34%,平均得分分别提高了0.39%、0.98%、0.82%、15.46%。研究结果表明:在轴承RUL预测方面,基于注意力机制的轴承剩余使用寿命预测模型(方法)是有效的。Aiming at the problems that the existing feature extraction methods of bearing vibration signals relied too much on expert experience and the memory degradation caused by too long sequences in life prediction,a prediction model(method)for the remaining useful life(RUL)of rolling bearings was proposed by combining convolutional neural network-attention network(CNN-attention)and the Encoder-Decoder method based on the attention mechanism.First,the initial vibration signal of the rolling bearing was converted into a frequency domain amplitude signal using the fast Fourier transform(FFT)method.Then,a model based on attention mechanism was designed,in which degenerate feature extraction was performed by using CNN-attention,RUL prediction was performed by using the Encoder-Decoder network based on attention mechanism,and the problem of memory decay in the recurrent neural network in long-distance signal transmission was further solved.Finally,in order to verify the validity of the feature extraction model as well as the lifetime prediction model,the experiments were conducted with the PHM 2012 bearing degradation data set using the PRONOSTIA experimental platform for accelerated bearing degradation.And the obtained predictions were compared with the predictions without the attention mechanism model and the results of other literature methods.The experiment results show that compared with the results obtained by other methods,the average absolute errors of the method based on the attention mechanism model are reduced by 29.41%,32.00%,29.56%,32.34%,and the average scores are increased by 0.39%,0.98%,0.82%and 15.46%.The research results show that,in terms of bearing RUL prediction,the attention mechanism-based bearing remaining useful life prediction model(method)is effective.

关 键 词：剩余使用寿命 卷积神经网络-注意力机制网络 编码器-解码器模型 退化特征提取 滚动轴承寿命预测模型 记忆力退化 

分 类 号：TH133.33[机械工程—机械制造及自动化] TP183[自动化与计算机技术—控制理论与控制工程]