Staged domain adaptation with transformer for bearing remaining useful life prediction

Accurate prediction of Remaining Useful Life (RUL) is essential to optimize operational efficiency in engineered maintenance systems. Conventional transfer learning approaches often suffer from negative transfer when applied across the entire equipment lifecycle, primarily due to stage-specific distribution shifts. Forcing global alignment between early-stage normal data and late-stage fault data distorts the degradation trajectory, leading to significant prediction errors. Additionally, health indicators are difficult to accurately model in the bearing degradation process. To address this, we propose a Stage-Aware Adversarial Domain Adaptation Framework. Unlike conventional methods that treat the lifecycle as homogeneous, our approach introduces a theoretically grounded staging mechanism driven by CAE-derived Health Indicators (HI). By explicitly decoupling domain alignment into distinct degradation phases, we ensure feature alignment occurs only within semantically consistent operational regimes. Furthermore, domain adversarial neural networks (DANN) and maximum mean discrepancy (MMD) enhance cross-phase adaptability. The experimental results have been validated with two datasets, and show an improvement in prediction accuracy above 10% compared to both traditional and transfer learning methods, offering new methods for domain alignment for transfer learning in this field.