通过卷积神经网络(Convolutional Neural Network, CNN)处理轴承一维时域或频域信号，难以提取具有代表性的非线性特征信息，且易忽略低层次信息。针对这一问题，基于多尺度特征提取，引入一种特征注意力机制，提出一种基于卷积双向长短期记忆网络(MSAMCNNBiLSTM)的轴承剩余寿命预测方法。基于西安交通大学（Xi′an Jiao Tong University，XJTU）轴承数据集中的3组数据对MSAMCNNBiLSTM、LSTM、CNNLSTM和MSAMCNNLSTM 4种方法的预测误差进行对比分析。结果表明：MSAMCNNBiLSTM方法在3组数据集中的预测误差均小于其他3种方法，说明该模型能同时学习数据中的低层次与高层次信息，可有效提高轴承的剩余寿命预测精度。

Processing the onedimensional time and frequency domain signals of bearings by convolutional neural network (CNN) was difficult to extract the representative nonlinear characteristic information, and easy to ignore the lowlevel information. To solve this problem, a feature attention mechanism was introduced based on multiscale feature extraction, and a prediction method of bearing remaining useful (RUL) life based on convolutional bidirectional long and short term memory network (MSAMCNNBiLSTM) was proposed. Based on three groups of data in the Xi′an Jiaotong University (XJTU) bearing data set, the prediction errors of four methods including MSAMCNNBiLSTM, LSTM, CNNLSTM and MSAMCNNLSTM were compared and analyzed. The results show that the prediction errors of the proposed MSAMCNNBiLSTM method in the three data sets are smaller than that of the other three methods, indicating that the model can learn the low level and high level information in the data at the same time, and can effectively improve the prediction accuracy of the remaining useful life of bearings.

TH133, TP183

国家自然科学基金（51976131，52006148，52106262）；上海市IV类高峰学科-能源科学与技术-上海非碳基能源转换与利用研究院建设项目资助