Federated Learning and MEC for Disaggregated RAN Monitoring in the 5G Non-Terrestrial Networks

The disaggregation of the Next Generation Radio Access Network (NG-RAN) is essential for optimizing resource consumption in 5G Non-Terrestrial Networks (NTN), particularly within Low Earth Orbit (LEO) satellite constellations operating in regenerative mode. By splitting the access network into Central Unit (CU), Distributed Unit (DU), and Radio Unit (RU), disaggregated NG-RAN architectures offer flexibility in satellite payload design and address capacity constraints inherent in NTN environments. However, this disaggregation, combined with satellite mobility and limited processing power, presents challenges in ensuring network resilience and addressing Quality of Service (QoS) requirements. This work proposes a novel framework that integrates Federated Graph Neural Network (FedGNN) with Multi-access Edge Computing (MEC) to monitor disaggregated NG-RAN components in 5G NTN networks. FedGNN enables decentralized model training across terrestrial and satellite nodes without sharing raw data, ensuring privacy, reducing communication overhead, and enhancing fault detection capabilities. MEC brings computation closer to the network edge, facilitating real-time traffic monitoring and adaptive routing decisions. Using a Graph Convolution Network (GCN), the framework detects link failures, optimizes traffic paths, and dynamically adapts to network conditions, ensuring robust and efficient operations across the NTN. We implement the proposed framework within a Kubernetes cluster comprising terrestrial and satellite edge nodes. The deployment includes local GCN models hosted on MEC nodes, supported by the Mosaic 5G Operator (M5G) for virtualized network function orchestration and a Flexible Radio Access Network (FlexRAN) controller for centralized NG-RAN control. This integration ensures efficient management of resources over the F1 interface between gNB Centralized Unit (gNB-CU) and gNB Distributed Unit (gNB-DU) network functions across the Satellite Radio Interface (SRI) (F1-over-SRI). The results demonstrate significant improvements, including fault detection precision of 93.5%, latency reductions from 30 ms to 15 ms across the F1-over-SRI, throughput enhancements from 70% To 90%, and traffic routing optimization, reducing average hops from 5 to 4.1. Scalability tests confirm efficient convergence times and robustness under dynamic satellite mobility, with reductions in latency and packet loss. The results validate the effectiveness of the FedGNN and MEC framework in addressing the challenges of disaggregated NG-RAN architectures in dynamic and resource-constrained 5G NTN environments.