A Graph Neural Network and Transformer-Based Anomaly Detector and Root Cause Analyzer in Radio Access Networks

Abstract

The emergence of 5G technology marks a major milestone in the evolution of telecommunication networks, enabling transformative applications such as augmented reality and autonomous vehicles. However, these advancements significantly increase network management complexity and raise critical challenges in failure handling, as 5G-supported applications heavily depend on ultra-reliable, high-performance, and low-latency connectivity. Consequently, automated self-healing mechanisms have become essential, using learning-based approaches to detect anomalies and perform Root Cause Analysis (RCA) with minimal human intervention. Such intelligent systems must effectively capture the inherent spatio-temporal communication patterns among mobile devices (UEs) and radio base stations, using realistic Key Performance Indicators (KPIs) derived from representative Radio Access Networks (RANs). Nevertheless, existing AI-driven solutions often rely on outdated LTE datasets or model only spatial or temporal characteristics in isolation, limiting their effectiveness in highly dynamic 5G RAN environments. This work develops a hybrid Graph Neural Network (GNN) and Transformer-based model, Simba, where the GNN captures spatial dependencies while the Transformer models temporal dynamics for joint anomaly detection and RCA. Then, we generate synthetic yet representative datasets using a calibrated Simu5G simulator aligned with 3GPP specifications. Extensive evaluations under multiple concurrent fault scenarios across both normal and peak-hour RAN operations demonstrate Simba’s superior performance over existing approaches. In particular, Simba consistently achieves an F1-score of at least 0.95 across all evaluated scenarios, representing an 18% improvement over a state-of-the-art baseline detection scheme.