With the rise of collaborative computing, AI can be rapidly federated across disparate devices and data centers. However, data falls under privacy protection and cannot be developed through joint information collection. Therefore, the industry proposes using federated learning to train data. Even with the most advanced federated learning algorithms, the prediction performance of the model remains low. To address this, we propose FedDDS, a federated optimization method suitable for collaborative computing. This method unites a large number of scattered clients into a group and applies data augmentation, which helps alleviate data heterogeneity among clients and significantly improves global model accuracy. Specifically, we first represent each client with a data class distribution and group the clients based on Bray-Curtis dissimilarity. Then, based on data category statistics, a Gaussian mixture model is constructed for data augmentation. Within each group, sparse categories are augmented using a Gaussian mixture model. Clients within a group are trained sequentially, and the group model is trained using simulated data. Finally, the central server aggregates the individual group models to generate a new global model for the next round of global training. Experiments show that on highly heterogeneous datasets such as FMNIST and CIFAR, FedDDS significantly outperforms DynaFed, FedAvg, FedProx, and Scaffold in both accuracy and convergence speed.