Abstract
Federated learning (FL) enables distributed model training across sensor-equipped edge devices while preserving data privacy. However, its performance is often hindered by statistical heterogeneity among clients and system heterogeneity in dynamic wireless networks. To address these challenges, we propose TCS-FEEL, a topology-aware client selection framework that jointly considers user distribution, device-to-device (D2D) communication, and statistical similarity of client data. The proposed approach integrates randomized client sampling with an adaptive tree-based communication structure, where user devices not only participate in local model training but also serve as relays to exploit efficient D2D transmission. TCS-FEEL is particularly suited for sensor-driven edge intelligence scenarios such as autonomous driving, smart city monitoring, and the Industrial IoT, where real-time performance and efficient resource utilization are crucial. Extensive experiments on MNIST and CIFAR-10 under various non-IID data distributions and mobility settings demonstrated that TCS-FEEL consistently reduced the number of training rounds and shortened per-round wall-clock time compared with existing baselines while maintaining model accuracy. These results highlight that integrating topology control with client selection provides an effective solution for accelerating privacy-preserving and resource-efficient FL in dynamic, sensor-rich edge environments.