A hierarchical refined representation with degradation-guided alignment network for rolling bearings remaining useful life prediction under unseen conditions

Abstract

Domain generalization (DG) has been widely applied in bearing remaining useful life (RUL) prediction owing to its capacity for generalizing prior diagnostic knowledge toward unseen operating conditions. However, numerous DG methods are often constrained by the limited domain-invariant feature generalization capabilities and information redundancy issues. Additionally, common methods used to measure distances between data distributions may encounter the gradient vanishing problem. To address these challenges, a hierarchical refined representation with degradation-guided alignment network for bearing RUL prediction is proposed. Firstly, a hierarchical refined representation module that combines a Top-N sample-level filter mechanism with a channel gating strategy to enhance domain-invariant representations. Subsequently, a degradation-guided distribution alignment module employs RUL-aware Sinkhorn divergence to further extract the invariant features of each source domain for fine-grained alignment. Finally, comprehensive multi-source DG experiments on the IEEE PHM datasets and the XJTU-SY datasets demonstrate the proposed method’s effectiveness for RUL prediction under unseen operating conditions.