A novel remaining useful life prediction method under multiple operating conditions based on attention mechanism and deep learning

Abstract

Remaining useful life (RUL) prediction is a key technique for supporting predictive maintenance. Accurate RUL prediction plays an important role in maintenance decisions. However, RUL prediction has two challenges: first, it is difficult to capture long-term dependencies effectively; second, the accuracy and efficiency are not satisfied under multiple operating conditions. A novel RUL prediction model that integrates bidirectional temporal convolution and improved Informer (ABiTCI) is proposed with consideration of multiple operating conditions. First, the bidirectional temporal convolution network (BiTCN) is designed with efficient channel attention (ECA). The degradation features from different channels can be extracted by weighting feature contributions. Second, the Informer with sparse pyramid temporal self-attention is designed to capture degradation information from different time steps. Finally, the effectiveness of the proposed method is verified by different datasets of aircraft engines. Compared with the present methods, the results show that the root mean square errors (RMSEs) have been reduced by 20.84 %–50.38 %, 16.29 %–41.49 %, and 36.96 %–59.53 % on the CMAPSS-FD002, CMAPSS-FD004, and NCMAPSS datasets, respectively. It demonstrates that the ABiTCI model performs well for RUL prediction under multiple operating conditions.