<p>Federated Learning (FL) is an effective distributed machine learning framework that addresses the problem of data silos. However, due to differences among clients in terms of computing power, network latency, and data distribution (data and system heterogeneity), the training times vary. This leads to low efficiency in federated learning, which synchronously waits for all clients to complete. Asynchronous federated learning can reduce waiting times, but it often suffers from unfairness during the training process, resulting in a global model that is only suitable for a subset of clients, thereby diminishing its generalization ability. In this paper, we propose a novel framework, SAFL-KCS, to address the problems of system and data heterogeneity. First, our framework adopts a semi-asynchronous communication method, which allows clients to upload their locally trained models before completing training, thus effectively shortening the server’s waiting time and addressing system heterogeneity. Second, before each round of global training, we select representative clients for global training through a K-means clustering-based client selection method, which addresses data heterogeneity. Finally, we design a model update strategy for the central server to optimize the framework and accelerate the convergence speed of the global model. We compare our approach with current well-known synchronous and asynchronous federated learning methods and conduct extensive experiments on various training models and datasets. Experimental results demonstrate that, in the presence of system and data heterogeneity, our framework outperforms current state-of-the-art federated learning schemes based on synchronous, asynchronous, and semi-asynchronous communication methods, achieving higher accuracy and faster convergence speed.</p>