The principal component analysis network (PCANet), as a network model based on the deep subspace learning framework, has demonstrated remarkable performance in various application domains. However, in the field of rolling bearing fault diagnosis, the PCANet algorithm suffers from issues such as inaccurate reflection of data structural information, poor robustness, and limited generalization ability. To address these issues, this paper proposes a novel rolling bearing fault diagnosis method based on the PCANet algorithm and data augmentation. The proposed method utilizes the L21-norm to learn the frequency domain sparse structure of the rolling bearing vibration signals, effectively suppressing noisy data and enhancing the robustness of the model. Moreover, through the data augmentation processing, the method significantly increases the variability between different classes of the training samples, thereby greatly improving the generalization ability of the model. Finally, experimental results demonstrate that the proposed method significantly enhances the robustness and generalization ability of the PCANet model, enabling accurate identification of different types of the rolling bearing faults.