Abstract: Aiming to address the low diagnosis accuracy of existing methods in processing two-way vibration signals and their inability to utilize inter-informational correlations, this paper proposes a fault diagnosis method based on a dual-channel attention residual network with wavelet-based denoising (CWT-CBAM-ResNet34). First, an optimized wavelet threshold denoising algorithm is applied to denoise the bearing vibration signal, followed by a two-dimensional transformation using continuous wavelet transform (CWT). The convolutional block attention module (CBAM) is then employed to optimize the traditional residual network, enhancing its capability to extract and integrate both local and global information. Furthermore, a feature fusion module based on a self-attention mechanism is introduced to fuse features across different channels, thereby fully leveraging informational correlations. This approach enables comprehensive information fusion for fault diagnosis across the data, feature, and decision layers. Experimental results show that the average diagnosis accuracy of the proposed method on a self-built dataset and the bearing dataset of Huazhong University of Science and Technology (HUST) reaches 99.85% and 98.33%, respectively, confirming the method’s high accuracy and robustness in fault diagnosis.