Dynamic Sub-Model Aggregation and Clustering for Intelligence Data via Hierarchical Federated Learning with Pre-Training

Abstract

To address the challenges of slow convergence speed, poor dynamic adaptability, and low communication efficiency in intelligence data processing, a dynamic integration and clustering method for intelligence data based on an improved federated learning algorithm is proposed. First, an improved federated learning algorithm combining decomposition and combination is designed, where the global model is decomposed into multiple sub-models for local training, and a dynamic combination strategy is applied to integrate these sub-models, thereby improving the adaptability and accuracy of the global model. Then, a pre-training mechanism is introduced to initialize the global model using feature information from historical data, enhancing the model's initialization performance in dynamic data environments and accelerating convergence. Experiments are conducted on the MNIST and CIFAR-10 datasets, with comparisons made against baseline methods including FedAvg, FedProx, and ScaFFL. The results show that the proposed algorithm achieves accuracies of 98.69% and 90.26% on the pathological heterogeneity client, and 98.14% and 89.87% on the actual scenario heterogeneity client, respectively, on the two datasets. The normalized mutual information values of the proposed intelligence dynamic data integration and clustering method are 0.91 and 0.79, respectively. In a practical medical Internet of Things scenario test, the running time and memory usage of the proposed method are 18.23s and 1681MB, respectively. Our research denotes that the designed method can effectively improve the quality of dynamic integration of intelligence data and reduce resource consumption, providing a feasible solution for efficient processing of multi-source heterogeneous intelligence data.