Bearing fault diagnosis has evolved from machine learning to deep learning, addressing the issues of performance degradation in deep learning networks and the potential loss of key feature information. This paper proposes a fault diagnosis method for rolling bearing faults based on ICEEMDAN combined with the Hilbert transform (ICEEMDAN-Hilbert) and a residual network (ResNet). Firstly, the collected fault vibration signals are classified as fault samples and randomly sampled with a fixed length. The IMF components obtained by decomposing the bearing fault vibration signals using ICEEMDAN are able to maximize the restoration of fault vibrations. Then, the IMF components are transformed from one-dimensional time-domain signals to two-dimensional time-frequency domain images using Hilbert transformation. The RGB color images can be directly used in deep learning models without the need for manual labeling of a large amount of data, thereby avoiding the loss of key feature information. The ResNet network incorporates the attention mechanism (CBAM) structure for the precise extraction of fault features, enabling a more detailed classification of fault features. Additionally, the residual network effectively addresses the problem of performance degradation in multi-layer network models. Finally, transfer learning is applied in the deep learning network by freezing the training layer parameters and training the fully connected layer. This effectively solves the problem of insufficient data in real operating conditions, which hinders deep training of the model, while also reducing the training time. By combining the ResNet network with the convolutional block attention module (CBAM) structure, the model completes the recognition and training of time-frequency images for rolling bearing faults. The results demonstrate that the ResNet with CBAM model has strong fault feature extraction capabilities, achieving higher accuracy, 7−12% higher than other conventional network models, and exhibiting superior diagnostic performance compared to other deep learning models.