Personalized federated learning for remaining useful life prediction under scenarios of fragmented out-of-distribution data

Accurate Remaining Useful Life (RUL) prediction model relies on full-lifecycle degradation features of the equipment. However, fragmented out-of-distribution (OOD) data due to specific working condition, equipment service time and communication packet loss inevitably affect the prediction accuracy. This study proposes a personalized federated RUL prediction method for fragmented OOD data scenarios, aiming to integrate OOD data fragments provided by different clients. In this means, a federated prediction model can be established to capture the full-lifecycle degradation features by incorporating fragmented OOD data. We focus on establishing a correctable cycle-consistent alignment mechanism driven by health state similarity to solve the challenging problem arisen by inter-client spatiotemporal heterogeneity. A novel health assessment index based on the quantile of hypothesis test is designed to capture the degradation feature required in the cycle-consistent alignment mechanism. Once new fragmented OOD data is available, a personalized federation strategy is developed by designing an adversarial mechanism between degradation features involved in the previous old OOD data and the new OOD data, such that previous degradation features can be further extended to a more full degradation feature. The superiority of the proposed method in RUL prediction was validated on fragmented OOD data collected on benchmark bearing prognostic system (BPS) platform.