Dynamic distance polarization regularized optimal transport for cross domain fault diagnosis of rotating machinery

In recent years, optimal transport (OT) theory has garnered substantial attention in the cross-domain intelligent fault diagnosis of rotating machinery. OT theory is defined by the minimal cost required for transporting one distribution to another, even in the face of notable domain disparities. However, current OT-based fault diagnosis methods have limitations: distance-regularized method such as Wasserstein distance lack explicit modeling of the OT plan, and the learned OT plan cannot establish the relationship between fault clustering structures; coupling-regularized method such as entropy-regularized OT will cause the OT plan to be too dense, and the uncertainty of cross-domain transport plans will increase accordingly. To tackle these challenges, a cross-domain fault diagnosis framework based on dynamic distance polarization regularization for optimal transport is developed, which framework can explicitly model the structure of the OT transport plan and mitigate the complexity of cross-domain transport processes. Firstly, a distance polarization mechanism based on large-margin learning has been constructed and integrated into the optimal transport plan. By enforcing intra-class and inter-class thresholds, the optimal transport scheme is guided toward bipolar-oriented optimization, and a clear boundary between similar and dissimilar transport pairs is established, thereby reducing the uncertainty in the transport process. Secondly, a dynamic threshold mask strategy has been proposed, through which diagnostic category information is incorporated into the polarization process, thereby capturing intra-class/inter-class correspondences and ensuring the correct transport of diagnostic knowledge. Finally, to ensure target-domain Classifier diagnostic accuracy, source-domain samples are mapped to the target domain via Barycenter mapping, enabling the model to learn discriminative class boundaries in the target representation space and enhance diagnostic accuracy for unlabeled target data. The cross-domain diagnosis model is validated on two bearing datasets under different loads and speeds. The results demonstrate that its diagnostic performance is significantly improved compared with existing transfer learning models. The related work has been open-sourced at: https://github.com/Pear-so/OT_Fault-Diagnosis.