Aiming at the shortcomings of traditional Transformer models in rolling bearing fault diagnosis, such as insufficient structural relationship modeling capability and poor noise robustness, this paper proposes an intelligent diagnostic method based on TAGCN-Transformer. This method innovatively integrates the advantages of graph convolutional networks and Transformer through three key technical modules named as multi-scale feature extraction, dynamic weighted selection, and global feature fusion. These modules enable the precise identification of bearing fault characteristics. Firstly, a multi-scale wide convolutional kernel architecture is adopted to extract time-frequency features of rolling bearing fault vibration signals. Secondly, a dynamic weighted selection mechanism is introduced to adaptively optimize feature representation. Finally, the self-attention mechanism of Transformer is utilized to construct comprehensive feature representations. In motor bearing fault diagnosis experiments, the proposed method maintains a very high recognition accuracy under strong noise conditions, and t-SNE visualization analysis verifies its excellent feature extraction capability. Experimental results demonstrate that this method can maintain high accuracy in bearing fault identification even under strong noise interference, proving to be an effective approach for intelligent fault feature extraction and pattern recognition of rolling bearings.