Graph Neural Network-based C-RAN Monitoring for Beyond 5G Non-Terrestrial Networks

Non-Terrestrial Networks (NTN) expands traditional terrestrial service offering connectivity to remote and extreme geographical areas, to address 5G service everywhere at any time. In the virtualized and software-defined networking (SDN) landscape, the integration of NTN with the terrestrial network requires efficient network management strategies for resilient service deployment. Within the deployment of a dis-aggregated next-generation radio access network (NGRAN) with splitting functionality, one of the virtual network functions (VNF) can be incorporated into a large constellation of regenerative Low Earth Orbit (LEO) satellites as a payload. The limited computational resources of LEO satellites should be compensated for through efficient network function and link failure detection strategies. A graph neural network (GNN) model is developed to address these challenges. In this work, the NGRAN is split among centralized (gNB-CU) and distributed units (gNB-DU) in the end-to-end 5G network using a Kubernetes cluster. The gNB-CU is deployed in the terrestrial network, while the gNB-DU is a payload of regenerative LEO satellite constellations. A graph convolution network (GCN), a type of GNN, is used to learn the graph representation data of both gNB-CU and gNB-DU, incorporating traffic information. GCN identifies link failures within the disaggregated NGRAN's F1 interface and determines the efficient traffic routing path. The end-to-end NTN delay computed after link failure detection exhibits better performance compared to the default routing. The trained GCN model exhibits 85% link failure detection accuracy between the gNB-CU and gNB-DU, thus demonstrating a reduced end-to-end delay in the emulated network infrastructure.