DS-RAN-XAP: AI-Driven Dual Stage Explainable Anomaly Prediction for Beyond 5G Networks

In this work, we introduce Dual-Stage Radio Access Network eXplainable Anomaly Prediction (DS-RAN-XAP), a novel Artificial Intelligence (AI)-driven and explainable framework for predictive Anomaly Detection (AD) in RAN environments. Designed to proactively monitor mobile connectivity, the framework integrates three key components: (i) multivariate time-series forecasting of network telemetry, (ii) connectivity classification between Radio Access Technologies (RATs), and (iii) unsupervised AD to detect degradations in network behavior. To ensure transparency and interpretability, the framework leverages SHapley Additive exPlanations (SHAP) to identify which Key Performance Indicators (KPIs) mostly affect AI model outputs. Evaluated on real-world multivariate RAN data, DS-RAN-XAP demonstrates strong performance across all stages. The classification component achieves an F1-score of up to 89.2%, while predictive classification retains high accuracy across medium and long-term forecasting horizons. The AD module, based on an Autoencoder (AE) architecture, achieves an F1-score of over 75% on forecasted data, validating its ability to generalize to future connectivity conditions. SHAP explainability further enhances model trust by offering insight into the KPIs contributing most to anomalous behavior. Thanks to its modular architecture, DS-RAN-XAP allows for flexible integration of alternative AI models at each stage. Thus, it supports deployment across diverse network infrastructures and operational objectives, aligning with the goal of predictive, explainable intelligence in next-generation mobile networks.