Chebyshev change-point detection with meta-learning adaptive boundary correction for few-shot multi-stage rolling bearing RUL prediction

Rolling bearings exhibit multi-stage degradation patterns. Traditional prediction methods struggle to accurately locate stage boundaries and generalize poorly under few-shot conditions. This paper proposes a stage-aware remaining useful life (RUL) prediction method that combines Chebyshev inequality-based state identification with few-shot adaptive boundary correction to achieve reliable stage division. First, a health indicator (HI) time series is constructed for the bearings. Chebyshev inequality is used to identify state change points and remove outliers, thereby obtaining stable degradation trends and initial stage partitions. Second, a meta-learning-based boundary correction module adaptively fine-tunes the initial segmentation boundaries, improving adaptability to diverse degradation modes. Finally, stage probability information is fused with temporal features and fed into an RUL regression network for precise RUL estimation, where a weighted smoothing strategy is adopted for transition intervals to eliminate abrupt jumps of predicted values near stage boundaries. Experiments on the IEEE PHM 2012 bearing dataset verify that the proposed method significantly improves prediction accuracy, demonstrating its effectiveness and practicality.