A Mask Disentanglement Domain Generalization Method for Bearings Remaining Useful Life Prediction Under Unknown Operating Conditions

Domain adaptation has proven to be a successful method for predicting the remaining useful life (RUL) prognostic tasks in cross-domain scenarios. However, due to the reliance on target data, these methods are difficult to deploy to RUL prediction tasks under unknown operating conditions (OCs). Additionally, existing RUL prediction methods under unknown OCs try to learn domain-invariant features, ignoring domain-specific features that are beneficial for enhancing generalization to target tasks. To eliminate dependency on target data and consider label-relevant information in domain-specific features, a mask disentanglement domain generalization (MDDG) framework is proposed. Specifically, a binary mask is used to disentangle latent features into label-relevant and label-irrelevant features, instead of using multiple encoders. Then, an information theory-based loss is designed to ensure that label-relevant features contain sufficient degradation information, while label-irrelevant features have enough domain information. Finally, a pairing loss is developed to reduce the difference in label-relevant features between samples with similar labels from different domains, facilitating generalization to unknown target domains. Experiments conducted on a run-to-failure bearing dataset demonstrate the feasibility and superiority of the proposed approach.