Bidirectional gradient-guided perturbation framework for remaining useful life prediction under unseen operating conditions

Existing domain generalization-based methods for predicting remaining useful life (RUL) are limited in their ability to proactively explore unknown domains. To address these issues, this paper proposes a novel bidirectional gradient-guided perturbation framework for RUL prediction under unseen operating conditions. The proposed framework constructs a bidirectional perturbation mechanism and actively generates gradient-guided domain-confused samples during training, thereby expanding the distribution coverage of the training data to encompass a wider range of domain variations and enabling the model to learn more robust domain-invariant degradation features. Specifically, the domain-adversarial perturbation module generates input perturbations by maximizing the loss of the domain classifier, thereby actively simulating domain shifts and explicitly enhancing the model’s sensitivity to domain variations. In addition, this module integrates a prediction consistency constraint and a gradient variance matching constraint, ensuring that the RUL predictor maintains stable outputs under domain changes, thereby improving its stability under domain shifts. ... mehrThe task-adversarial perturbation module, in turn, generates input perturbations by maximizing the loss of the RUL predictor while keeping domain labels unchanged, thereby improving the domain classifier’s sensitivity to true domain boundaries and facilitating the generation of more discriminative domain-adversarial perturbations. Comparative experimental results on three life-cycle datasets validate the superior domain generalization capability of the proposed method.