Multi-sensor bearing remaining life prediction method based on bidirectional temporal convolutional capsule network fusion attention mechanism

Abstract Aiming at the shortcomings of existing deep learning life prediction methods in long time series feature extraction and the impact of the complexity of bearing working conditions on life assessment, a data fusion remaining useful life prediction method based on the bi-directional temporal convolutional capsule fusion attention mechanism (Bi-TCC) is proposed. Bi-TCC employs a dual-channel structure to process forward and backward time-series data separately in order to comprehensively capture the full life cycle characteristics of bearing, especially the late degradation trends. The forward channel adopts an improved temporal convolution network (TCN) to extract shallow time series information, and combines the multi-scale attention mechanism to focus on both local and global features, and enhances the spatial hierarchy learning through a one-dimensional (1D) causal capsule network. The reverse channel optimization Improving the TCN block and introducing a self-attention mechanism to improve the ability to capture key features in long time series, and further utilizes the 1D causal capsule network to increase sensitivity to the later declining stage. In addition, Bi-TCC fuses multi-sensor data to make full use of multi-source information in order to enhance the accuracy and robustness of prediction. The experiments on Prognostics and Health Management 2012 and XJTU-SY datasets validate the effectiveness of Bi-TCC and demonstrate its superiority under different operating conditions. The results show that compared with existing advanced methods, Bi-TCC has significant improvements in both accuracy and robustness, demonstrating excellent potential for engineering applications.