TAAGNet: a temporal–attribute–adaptive graph convolutional network for unsupervised cross-domain bearing fault diagnosis

Abstract In practical rolling bearing fault diagnosis, complex load variations and composite fault scenarios frequently induce substantial shifts in feature distributions, which significantly impair the generalization capability of diagnostic models. To mitigate this challenge, this paper proposes an unsupervised domain adaptation diagnostic network integrating temporal modeling, attribute enhancement, and structural alignment. The proposed method initially constructs a hybrid feature extractor based on residual one-dimensional convolutional networks and bidirectional gated recurrent units, enabling simultaneous extraction of local frequency-domain features and global temporal dependencies from vibration signals. Subsequently, fine-grained category attribute modeling and soft allocation mechanisms are incorporated to guide the model in producing semantically consistent and discriminative feature representations. Additionally, a K-nearest neighbor graph is utilized to establish structural relationships between samples. In contrast, a graph convolutional network is employed to jointly model structure and semantics, thereby enhancing the model’s class discrimination capability in the target domain. Furthermore, by incorporating attribute consistency loss, structural triplet loss, and adversarial transfer strategies, a multi-task joint optimization objective is formulated to achieve feature alignment and structural preservation. In this study, six load variation tasks and ten composite fault identification tasks were designed on two representative bearing datasets, Case Western Reserve University and Huazhong University of Science and Technology. Experimental results demonstrate that the proposed method achieves excellent diagnostic accuracy and robustness in all transfer tasks, significantly outperforming current mainstream adaptive comparison methods.