Decentralized Federated Learning for GNN-Based Channel Estimation With DM-RS in O-RAN
Sajedeh Norouzi, Eric Samikwa, Mostafa Rahmani, Torsten Ingo Braun, Alister Graham Burr · 2025
This paper introduces GrINet, a Graph-based neural NETwork for channel estimation and Interpolation using Demodulation Reference Signals (DM-RS) to enhance estimation accuracy in wireless communication systems. GrINet models each Resource Element (RE), the smallest resource unit in 5G New Radio, as a node in a graph, with edges connecting DM-RS-enriched nodes, enabling effective modeling and processing of complex channel conditions. Building on this, we propose decentralized Federated GrINet (FGrINet), a hierarchical framework that combines Federated Learning (FL) at the base station (BS) level with decentralized collaboration across BSs. Locally, each BS employs FL to optimize CE models using data from its connected users under unique channel conditions. Globally, BSs share and aggregate their locally trained models in a decentralized manner, enabling collaborative learning without relying on centralized orchestration. This two-tiered approach allows BSs to operate in favorable conditions to assist others, enhancing adaptability to diverse channel environments. FGrINet can be implemented as an xApp architecture, aligning with O-RAN’s goals of using distributed machine learning for intelligent, real-time RAN optimization. Our simulation involving multiple BSs, diverse channel profiles, and varying user mobility demonstrates that FGrINet reduces local training time, enhances CE accuracy, and achieves low mean squared error (MSE).