Multiscale Spatiotemporal Attention Network for Remaining Useful Life Prediction of Mechanical Systems

Remaining useful life (RUL) prediction plays a critical role in mechanical systems. RNN-based methods have achieved unprecedented success. However, these methods neglect spatial dependencies among sensors and suffer from long-term dependency learning. To break through these limitations, a novel multiscale spatiotemporal attention network (MSAN) is proposed for predicting the RUL of aircraft engines. In the MSAN, a multiscale discrete wavelet transformation (MDWT) is first constructed to obtain a multiscale subseries set. Then, an adaptive spatiotemporal feature extraction module is proposed to mine both long-term and spatial dependencies and form holistic spatiotemporal features by a collaborative spatiotemporal learning module (CSLM). Finally, a versatile fusion module is developed to integrate holistic spatiotemporal features for RUL prediction. The MSAN is validated on C-MAPSS datasets, and the experimental results demonstrate that the MSAN can better perform prediction tasks than existing state-of-the-art (SOTA) methods.