Centrifugal pumps, as critical equipment in nuclear power plants, play a key role in the cooling water circulation. Misalignment and imbalance faults in the centrifugal pump rotor can lead to performance degradation, increased vibration, and equipment damage, significantly affecting the safety and reliability of nuclear power plants. In the process of centrifugal pump rotor fault diagnosis, the difficulty in acquiring samples and the limited amount of data can result in an imbalance between the number of normal and fault state samples in the dataset. To address this issue, this paper proposes a rotor fault diagnosis method based on the Multi-Scale Spatial Convolutional Neural Network (MSSCNN) model. The MSSCNN model enhances fault diagnosis accuracy by learning features from imbalanced datasets using its multi-branch spatial convolutional network architecture. To further improve model performance, data augmentation techniques are employed to generate additional rotor fault data, thereby optimizing the training process. A comparison with other data generation models demonstrates that the proposed method, when combined with the MSSCNN model, significantly improves fault diagnosis accuracy on datasets with varying degrees of imbalance and outperforms multiple evaluation metrics. Experimental results show that the MSSCNN model can effectively address the data imbalance problem and achieve high accuracy in rotor fault diagnosis, offering promising application potential.