Remaining Useful Life Prediction of Aeroengines via Target-Domain Mutual Information and Fine-Grained Alignment

Abstract

Predicting the remaining useful life (RUL) of aeroengines across different operating conditions, traditional cross-domain learning struggles to capture the intrinsic semantics of the target domain, resulting in prediction results that are highly sensitive to uncertainty in the target-domain distribution. Furthermore, feature alignment relying solely on a single source domain often fails to adequately capture the fine-grained characteristics of complex degradation processes, leading to source-domain cognitive uncertainty. To address these challenges, a RUL prediction method for aeroengines is proposed, integrating target-domain mutual information constraint with fine-grained adversarial alignment (FAA). The proposed method aims to construct a domain-adversarial neural network framework under an unlabeled target-domain setting to perform domain-invariant feature extraction. By developing a target-domain mutual information constraint, a structural penalty is imposed on the predicted output distribution to reduce uncertainty in the target-domain distribution. Also, an FAA strategy is designed, which incorporates a loss-weight warming mechanism to achieve multilevel domain alignment across the convolutional, temporal, and predictive feature layers, thereby completely eliminating the source-domain cognitive uncertainty. Finally, experimental results demonstrate that the proposed method achieves superior predictive performance compared to baseline methods in the prediction uncertainty mitigation task, validating its effectiveness and robustness.