Adaptive Beam Tracking in 5G/6G mmWave Networks: A Clustered Federated Learning Approach

Millimeter wave (mmWave, 30–100 GHz) communication is essential for meeting the high data throughput demands of 5G/6G networks. However, mmWave signals are highly susceptible to attenuation and blockage, necessitating directional beamforming antennas and efficient beam tracking algorithms. Traditional machine learning-based approaches, such as centralized learning (CL) and federated learning (FL), face significant challenges. While CL achieves fast convergence, it suffers from high computational costs and privacy concerns. Conversely, FL addresses these issues by enabling distributed model training but struggles with slow convergence and suboptimal accuracy due to data heterogeneity. To overcome these limitations, this paper introduces a novel clustered federated learning (CFL) framework for beam tracking. CFL leverages the benefits of FL while grouping users with similar data distributions, enabling the training of a single model per cluster. This approach reduces communication overhead, accelerates training, and improves accuracy. We analyze key attributes influencing user clustering and their impact on learning efficiency. Numerical results demonstrate that CFL significantly outperforms traditional CL and FL methods, achieving an 18% accuracy improvement over FL (specifically, the FedAvg algorithm) and an 11% enhancement over CL. These findings highlight the potential of CFL for enhancing beam tracking in mmWave systems, offering a more adaptive and privacy-preserving solution for future wireless networks.