Stage Division and CNN-BiLSTM-Attention for Remaining Useful Life Prediction of Rolling Bearings

The prediction of the remaining useful life (RUL) of rolling bearings is one of the current research hotspots. However, due to the influence of internal degradation and external interference, the whole life cycle of rolling bearing is extremely complex and changeable, and the prediction accuracy of the existing methods needs to be improved. In order to realize the accurate RUL prediction of complex degradation process, a deep learning model based on stage division is proposed in this paper. Firstly, the feature indexes are extracted from time domain, frequency domain and time-frequency domain, and the canonical variate dissimilarity analysis (CVDA) that can handle dynamic data is used for feature reduction to obtain statistics containing streamlined degradation information. Secondly, the kernel density estimation (KDE) and expert knowledge are used to divide the whole life cycle into multiple stages based on the trend and fluctuation of statistics. Then, with the combination of convolution neural network (CNN) and Bi-directional long short-term memory (Bi-LSTM), the CNN-BiLSTM-Attention network is proposed, in which the attention mechanism is developed to highlight the focus. Moreover, the Bayesian optimization algorithm is used to adjust the network hyperparameters. Finally, the effectiveness is verified by using the accelerated life test data of rolling bearings.