Aiming at the problems of imbalanced sample sizes across different fault categories in industrial robot harmonic reducers and the limited information obtained from single-source signals, which result in low diagnostic accuracy, a fault diagnosis method for harmonic reducers is proposed based on federated learning under multi-source imbalanced data. This method performs wavelet transform on the multi-source signals of different users to convert one-dimensional signals into two-dimensional images, constructing a time-frequency dataset. An improved data augmentation method is then applied to balance the dataset. The efficient channel attention is introduced, and the output of the residual branches is weighted by learnable weights, which can enhance the model′s adaptability to residual information of different input signals and the ability to extract key features of the data. A modified multimodal variational autoencoder is used to mine the complementary information among multi-source signals for feature fusion, and adopting the focal loss function as the training loss function, the model can pay more attention to category samples with high misclassification frequencies, thus constructing personalized local models for multi-user. The server aggregates the local model parameters at the multi-user and updates the global model, multi-user local island privacy data is protected through federated learning, so as to achieve the fault diagnosis of harmonic reducers under multi-source imbalanced data. The effectiveness of the proposed method is verified by building a signal acquisition experimental platform for harmonic reducers. The proposed method can effectively extract the features from multi-source imbalanced data of harmonic reducers and achieve information fusion, achieving an average fault diagnosis accuracy of 98.8%, outperforming the compared methods.