Explainable Performance Anomaly Detection and Rectification in 5G-ORAN

Owing to the increasing reliance on black-box artificial intelligence (AI) and machine learning techniques for performance anomaly detection and rectification (PADR) in Fifth-Generation (5G) Open Radio Access Networks (O-RAN), concerns have arisen regarding the transparency and trustworthiness of such AI-driven systems. To address this challenge, a transparent 5G O-RAN framework is developed that leverages explainable AI (eXAI) metrics to interpret and validate the PADR process. Within this framework, a novel Self-Learning Attention Mechanism (SLAM) integrated with a long short-term memory (LSTM) network is proposed to accurately identify and predict the key features that trigger performance anomalies. The output of the SLAM-LSTM is subsequently utilized as input to a Proximal Policy Optimization (PPO) algorithm for anomaly rectification and error mitigation. In parallel, the SLAM-LSTM outputs, together with backed-up PPO operational statistics stored in a centralized registry, are leveraged by the extensible application (xAPP) to compute eXplainable AI (eXAI) metrics-namely confidence, comprehensiveness, entropy, and sensitivity-thereby enabling identification and interpretation of the features driving performance anomalies. Extensive simulation results demonstrate that the proposed approach consistently outperforms existing baseline methods.