This paper proposes an intelligent fault diagnosis method for rolling bearings based on Gramian Angular Difference Field (GADF) and Improved Dual Attention Residual Network (IDARN). The method converts original vibration signals into 2D-GADF feature images for network input, and incorporates dual attention mechanisms and group normalization (GN) to enhance feature integration and reduce bias caused by data discrepancies. The model is trained to classify faults, achieving an average identification accuracy of 99.2% and 97.9% on two different datasets. The method outperforms other deep learning approaches in fault diagnosis, particularly under imbalanced data conditions. The IDARN model uses a 34-layer residual structure with channel and spatial attention mechanisms to extract and refine features, and GN to adapt to different data distributions. The model was tested on the CWRU bearing dataset and bearing failure experimental equipment, demonstrating strong performance in fault classification. The method effectively handles imbalanced data and improves classification accuracy by focusing on key fault features. The results show that the proposed method achieves high accuracy and robustness in fault diagnosis of rolling bearings.This paper proposes an intelligent fault diagnosis method for rolling bearings based on Gramian Angular Difference Field (GADF) and Improved Dual Attention Residual Network (IDARN). The method converts original vibration signals into 2D-GADF feature images for network input, and incorporates dual attention mechanisms and group normalization (GN) to enhance feature integration and reduce bias caused by data discrepancies. The model is trained to classify faults, achieving an average identification accuracy of 99.2% and 97.9% on two different datasets. The method outperforms other deep learning approaches in fault diagnosis, particularly under imbalanced data conditions. The IDARN model uses a 34-layer residual structure with channel and spatial attention mechanisms to extract and refine features, and GN to adapt to different data distributions. The model was tested on the CWRU bearing dataset and bearing failure experimental equipment, demonstrating strong performance in fault classification. The method effectively handles imbalanced data and improves classification accuracy by focusing on key fault features. The results show that the proposed method achieves high accuracy and robustness in fault diagnosis of rolling bearings.