An interpretable transfer bayesian method for remaining useful life prediction

Accurate and interpretable remaining useful life (RUL) prediction under domain shifts with streaming multi-sensor data remains challenging for field equipment. To address this challenge, an interpretable transfer Bayesian method is developed, integrating three key components: dynamic pseudo-domain generation (DPG), online Bayesian updating, and a pseudo-domain-based ensemble strategy. First, the DPG algorithm transforms source domains into pseudo-domain units whose degradation trajectories closely match those of the target domain. Second, a dual-scale distance is proposed to identify the cumulatively selected optimal pseudo-domain units, which are then utilized in the Bayesian updating of the degradation model. Third, adaptive weights are assigned to multi-sensor features based on the selected optimal pseudo-domain units, thereby improving RUL prediction accuracy and robustness. Finally, simulation studies and experimental validation on two real-world Stirling cryocooler datasets demonstrate that the proposed method outperforms existing methods in terms of both accuracy and robustness.