Trend-augmented and temporal-featured Transformer network with multi-sensor signals for remaining useful life prediction

Deep learning method has obtained abundant achievements in remaining useful life (RUL) prediction, which can steer the preventive maintenance decision-making for improving the reliability of industrial systems. However, the existing deep models often fail to consider mechanical degradation rules, and can not capture the temporal and featured dependencies effectively. To address such issues, this study proposes an improved Transformer network for RUL prediction with multi-sensor signals. Specifically, the trend augmentation module (TAM) and time-feature attention module (TFAM) are embedded into the traditional Transformer model. In TAM, a bidirectional gated recurrent unit (Bi-GRU) network is used to extract the hidden temporal information and a novel distance function is presented to improve the attention distribution. In TFAM, attention calculations are performed sequentially in both the feature and time dimensions to synthetically capture both the feature and time dependencies. Two benchmark experiments are conducted with the CMAPSS and Milling datasets respectively. The results indicate that the proposed approach outperforms state-of-the-art approaches and possesses deep interpretability.