基于动态加权卷积长短时记忆网络的滚动轴承剩余寿命预测方法

Residual life prediction method of rolling bearing based on DW-CNN-LSTM networks

现有数据驱动方法在滚动轴承剩余使用寿命预测中,因不能有效提取对轴承退化过程敏感的特征信息而导致预测精度不足。为此提出一种基于动态加权卷积长短时记忆网络(DW-CNN-LSTM)的滚动轴承剩余寿命预测方法。对滚动轴承振动信号进行小波包分解,将获得的小波包系数矩阵通过可训练参数动态加权层进行动态加权,来实现对轴承退化的表征信息进行有效筛选,以增强轴承振动特征学习能力;利用卷积神经网络的自适应挖掘数据深层特征能力,从动态加权后的小波包系数矩阵中提取对轴承退化过程敏感的特征集;借助长短时记忆网络(LSTM)预测时间信息序列的优势,由双层LSTM进一步提取其高维退化特征,来提高滚动轴承剩余寿命预测精度。对XJTU-SY轴承数据和IMS轴承数据的试验结果表明,所提DW-CNN-LSTM方法相比于经典的长短时记忆网络方法,其均方根误差指标平均降低了61.08%,预测准确度平均提高了9.95%,模型训练时间平均减少了44.14%,获得了较满意的寿命预测精度和鲁棒性。

In prediction of residual service life of rolling bearing,the existing data-driven methods can’t effectively extract feature information sensitive to bearing degradation process to cause lack of prediction accuracy.Here,a prediction method of residual life of rolling bearing based on dynamic weighted convolution neural network and long-short term memory(DW-CNN-LSTM)network was proposed.Firstly,vibration signal of rolling bearing was decomposed with wavelet packet,and the obtained wavelet packet coefficient matrix was dynamically weighted with the dynamic weighting layer of trainable parameters to effectively screen the characterization information of bearing degradation,and enhance the learning ability of bearing vibration features.Then,using CNN’sadaptive mining ability for deep features of data,the feature set sensitive to bearing degradation process was extracted from the dynamically weighted wavelet packet coefficient matrix.Finally,with help of advantages of LSTM in predicting time information series,bearing high-dimensional degradation features were further extracted using the dual-layer LSTM to improve the prediction accuracy of residual life of rolling bearing.Test results of XJTU-SY bearing data and IMS bearing data showedthat compared with the classical LSTM method,the RMS error index of the proposed DW-CNN-LSTM method is reduced by 61.08%,itsmean prediction accuracy is improved by 9.95%,its mean model training time is reduced by 44.14%,so the results achieve more satisfactory life prediction accuracy and robustness.