A graph-aware meta network for domain generalization in cross-condition bearing RUL prediction

Driven by advances in sensor technology, data-driven remaining useful life (RUL) prediction has become an important tool for bearing predictive maintenance. However, variations in operating conditions in industrial applications can cause feature distribution shifts, thereby reducing the accuracy and robustness of prediction models. To address feature distribution shifts across operating conditions, this paper proposes a graph-aware meta network based on domain generalization for bearing RUL prediction across conditions. The proposed method converts monitoring sequences into graph structured inputs and employs a temporal encoder to learn node-level degradation representations. It further integrates prior structural knowledge, globally shared structure, and task-adaptive structure to build multi-level dynamic adjacency relationships, and introduces an attention pooling mechanism to obtain graph-level degradation representations. During training, a meta-learning strategy with inner and outer loop updates is designed to dynamically adapt the graph generation parameters, enabling task-specific topology to adjust automatically with changing operating conditions and thereby improving model generalization. Cross-condition experiments on a public bearing accelerated life dataset demonstrate that the proposed method achieves superior prediction error metrics across multiple transfer tasks and delivers stable performance under varying operating conditions.