Category-aware trajectory domain adaptation network for cross-machine fault diagnosis

Given the difficulty of obtaining complete and well-labeled fault data from a single machine, it is particularly important to achieve fault diagnosis by leveraging data from other machines or test rigs. To address the critical challenges posed by distribution discrepancies in cross-machine fault diagnosis, this study proposes a novel category-aware trajectory domain adaptation network (CTDAN). The proposed method integrates multi-kernel maximum mean discrepancy and covariance alignment by label smoothing, thereby mitigating the adverse effects of mislabeled samples and achieving both marginal and conditional distribution adaptation between the source and target domains. Furthermore, a category-induced fine-grained alignment module is introduced to dynamically update class centroids, enhancing the discriminative power of the learned features. To further tackle the complex nonlinear discrepancies in cross-machine scenarios, based on the optimal transmission theory, implement cross-domain fault signature sharing to complete the trajectory adaptation. Experiments on four benchmark bearing datasets demonstrated that CTDAN significantly outperforms state-of-the-art methods in cross-machine diagnostic tasks, achieving an average performance improvement of approximately 26%. In addition, ablation studies and parameter sensitivity analyses validated the complementary advantages of the proposed modules.