Federated Learning with Hybrid Clustering for Radio Link Failure Detection in 5G Networks

Radio Link Failures (RLFs) pose a critical challenge in fifth-generation (5G) networks, especially in millimetre-wave deployments where links are vulnerable to blockages and environmental fluctuations. Conventional centralized detection schemes suffer from scalability, latency, and privacy limitations. In contrast, Federated Learning (FL) is hindered by heterogeneous data distributions across sites, leading to biased updates and unstable convergence. This paper introduces an FL framework for distributed RLF detection that leverages a Hybrid Graph Attention Network with Discrete Particle Swarm Optimization GAT-DPSO clustering to form coherent clients. Base stations are grouped into clusters based on Key Performance Indicators (KPIs) similarity and spatial proximity, resulting in balanced and representative training groups. Each cluster utilizes a Transformer-based model to capture both short-term and long-term dependencies in KPI sequences, thereby providing improved robustness under heterogeneous conditions compared to recurrent alternatives. Experimental results confirm that the proposed approach produces compact and well-separated clusters, achieving superior detection performance compared to baseline methods.