Mutual-learning based self-supervised knowledge distillation framework for remaining useful life prediction under variable working condition-induced domain shift scenarios

Abstract

Domain shifts induced by variable working conditions, including both multiple steady and time-varying working conditions, result in inconsistent degradation patterns and pose significant challenges for remaining useful life (RUL) prediction. To address the above issue, we propose a self-supervised knowledge distillation framework based on mutual learning for RUL prediction under variable working conditions. The proposed framework employs a teacher-student architecture, facilitating knowledge transfer through self-supervised pseudo-labels. A mutual learning-based training strategy is developed to prevent over-adaptation to the source domain and promote domain generalization. Additionally, during student model training, a feature-level domain adversarial training strategy is implemented to improve cross-domain feature decoupling and ensure the learning of domain-invariant features. The above two components complement each other: adversarial learning aligns marginal distributions (variable working conditions), while pseudo-label learning refines conditional alignment (normal and fast degradation stages), allowing the model to adapt more effectively to complex degradation scenarios. Furthermore, we incorporate a sparse attention mechanism for efficient feature extraction, significantly reducing computational complexity while maintaining robust performance. The RUL prediction experiments under multi steady conditions and time-varying conditions are carried out on two life-cycle bearing datasets respectively. Comparative results demonstrate the superiority and practicality of our proposed method.