Methods for bearing fault diagnosis based on transfer learning have shown notable success in enhancing feature adaptation across domains and improving model performance. However, existing methods still exhibit limitations in extracting highly discriminative features, which result in blurred feature distinctions between different fault categories in both source and target domains, weakening the model’s classification discriminative capability. Moreover, in scenarios with limited labeled data, models are prone to cross-domain catastrophic forgetting, further constraining improvements in diagnostic performance. To address these issues, a transfer learning model for bearing fault diagnosis based on angular marginal mapping methods and dynamic pairing strategies is proposed in this paper. First, a temporal-aware multiattention fusion network is designed, which effectively aligns features and extracts key information through cross-domain fusion and temporal attention mechanisms. Next, an angular marginal mapping method is proposed, whereby category boundary overlap is minimized by the angular distance between fault features being increased. Finally, a dynamic pairing strategy is proposed, which balances the model’s learning of features from different domains by adaptively optimizing the distance between positive and negative sample pairs. Validation through three case studies and extensive experiments demonstrates superior diagnostic performance over eight state-of-the-art methods with only 5% labeled data.
Zhou et al. (Wed,) studied this question.