Spatial-Temporal Evolutionary Graph Attention Network for Bearing Remaining Useful Life Prediction

Currently., most deep learning-based rolling bearing remaining useful life (RUL) prediction methods lack interpretability and fail to fully exploit feature structural information and temporal sequence information. To address these issues., this paper proposes a spatial-temporal evolutionary graph attention network for precise RUL prediction of bearings. Firstly., canonical variable analysis (CVA) is utilized to reduce the feature set obtained from the time domain., frequency domain., and time-frequency domain. Subsequently., an adaptive stage partitioning is achieved based on kernel density estimation (KDE). Furthermore., a directed spatial-temporal graph (STG) containing both structural and sequential information is innovatively proposed., which adequately reflects the degradation process of bearings. Finally., combining graph attention network (GAT) and gate recurrent unit (GRU)., evolutionary-GAT (E-GAT) is constructed to explore RUL prediction from both local and global perspectives. The effectiveness and superiority of the proposed method are verified through accelerated bearing life experiments.