Self-data-driven methods for Remaining Useful Life (RUL) prediction are promising where failure data is scarce. However, conventional batch-update approaches are computationally inefficient for online scenarios and their fixed models fail to capture multi-stage degradation. To overcome these limitations, this paper proposes a novel two-stage framework. The first, offline stage uses a Phase Adaptive Expectation Maximization (PAEM) algorithm, which identifies degradation phases to achieve robust parameter initialization for a library of candidate models. The second, online stage employs an Entropy-Driven Particle Filter (EDPF) to adaptively fuse these models in real-time, tracking time-varying dynamics without reusing historical data. Validation on the XJTU-SY bearing dataset demonstrates that the framework significantly improves prediction accuracy and stability over traditional methods, providing a computationally efficient and robust solution for online RUL prediction.

RUL prediction;Self-data-driven method;Particle filtering;Online Scenarios;Time-Varying Degradation