Abstract—The rapid evolution of 5G networks introduces complex operational challenges, particularly the “sleeping cell” phenomenon, where base stations appear functional in management systems yet fail to serve users. Although, traditional monitoring systems rely on static thresholds that struggle to adapt to the non-stationary, multi-periodic nature of 5G traffic, leading to excessive false alarms and missed degradations. This paper proposes a Hybrid Artificial Intelligence (AI)-Based Framework for robust sleeping cell detection that reconciles domain expertise with neural sensitivity. The framework integrates 3 components: a deterministic rule-based engine for high-confidence anomaly identification, a multi-task self-supervised temporal backbone to learn representations of normal behavior, and a supervised ensemble trained on domain-aware synthetic anomalies to address label scarcity. A key contribution is the introduction of a rule-priority fusion mechanism that ensures critical service-loss alarms take precedence while machine learning modules identify subtle, non-linear degradations. Evaluated on 126,801 hourly Key Performance Indicator (KPI) records from 194 live 5G cells, the proposed approach achieves an F1-Score of 0.728 and a The Receiver Operating Characteristic-Area Under the Curve (ROC-AUC) of 0.856, with 98.7% agreement with expert rules and the discovery of 11,375 previously undetected degradations. The results demonstrate a practical and scalable solution for proactive fault management in real-world 5G networks.
 
Keywords—5G networks, sleeping cell detection, time-series anomaly detection, self-supervised learning, Temporal Convolutional Attention Network (TCAN), ensemble learning