A Novel Partial Cross-Domain Remaining Useful Life Prediction Method for Aero-Engine Based on Health Indicator Weighted Distribution Alignment

Abstract Recently, domain adaptation methods have gained significant attention in the field of aero-engine remaining useful life prediction. These methods effectively address distributional shifts between training and testing datasets caused by changes in operating conditions or engine versions. However, existing methods often assume that the target domain covers data from the engine's complete run-to-failure cycle. In practice, collecting complete data from initial operation to failure is difficult for newly deployed aero-engines. The issue of incomplete data in the target domain becomes a significant challenge for the application of domain adaptation methods. To address this problem, we propose a novel health indicator weighted distribution alignment strategy to measure domain discrepancy. This strategy evaluates the degradation state similarity between source and target samples using health indicator slopes. By assigning instance-level weights to the source samples based on degradation state similarities, the model's attention during domain alignment is optimized, promoting knowledge transfer among samples with similar degradation states. Finally, cross-condition experiments on the commercial modular aero-propulsion system simulation dataset and cross-platform experiments on the turboshaft engine dataset are conducted. The experimental results demonstrate that the proposed method effectively addresses real-world partial cross-domain remaining useful Life prediction.