Domain generalization for rotating machinery real-time remaining useful life prediction via multi-domain orthogonal degradation feature exploration

The domain adaptation-based approach for remaining useful life (RUL) prediction has gained significant attention in addressing the challenges of cross-domain RUL prediction, characterized by distribution discrepancies between training and testing data. However, highly relying on the availability of target data limits its applicability in real-time RUL prediction scenarios, where accessing target data in advance is often very difficult. To tackle this issue, a domain generalization network is proposed for predicting RUL under unknown operating conditions. The foundation of this method is adaptively fusing the degradation features of multiple source domains to represent the degradation features of the test data based on the similarity between the test data and the multi-source data. This process emphasizes focusing on source data that exhibits high similarity to the test data, enabling the model to leverage task-relevant source degradation information while ignoring task-irrelevant degradation cues. Simultaneously, the discrepancies in marginal and conditional distributions across multiple source domains are mitigated through the proposed label consistency constraints and sample pairing strategy. These strategies enhance cross-domain transferability and facilitate the acquisition of generalized predictive knowledge. Extensive experiments in cross-domain RUL prediction under unknown operating conditions, conducted on one real dataset and two public datasets, validate the efficacy of the proposed methodology.