Channel Coefficient Estimation in Massive MIMO System using RIS and Federated Learning Techniques

Abstract

Reconfigurable Intelligent Surfaces (RIS) enhance the spectral efficiency and coverage of massive MIMO systems by establishing programmable reflection channels. However, the passive nature of RIS elements results in high-dimensional cascaded channels that make accurate channel state information (CSI) acquisition challenging and pilot-intensive. Traditional centralized deep learning approaches address this complexity at the expense of high feedback signaling overhead and compromised user data privacy. To overcome these limitations, this paper proposes a decentralized channel estimation framework that couples comb pilot-assisted interpolation with Convolutional Neural Networks (CNNs) orchestrated via Federated Learning (FL). In this design, individual user terminals extract spatial channel features by training local CNN estimators directly on initial pilot measurements. Rather than transmitting raw channel observations to the base station, clients exchange only local gradient parameters, which are consolidated using Federated Averaging (FedAvg) to iteratively update a global estimation model. Numerical simulations demonstrate that the proposed FL-CNN framework significantly decreases normalized mean square error (NMSE) across diverse SNR regimes, closely tracking centralized MMSE benchmarks as the client count increases. Furthermore, the framework achieves fast RMSE convergence matching centralized training while cutting uplink signaling loads by nearly an order of magnitude. These results confirm that federated learning delivers a private, bandwidth-efficient, and scalable channel estimation pathway for next-generation RIS-assisted networks.