Multimodal Anomaly Detection in Nuclear Power Plants Using Explainable Artificial Intelligence for Enhanced Safety and Reliability

Abstract

The integration of artificial intelligence (AI) into nuclear power plant (NPP) operations offers transformative potential for enhancing safety, reliability, and operational decision-making. This study presents a multimodal anomaly detection framework combining sensor measurements, inspection imagery, textual logs, and cybersecurity data, processed through hybrid deep learning models and Explainable AI (XAI) techniques. Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Transformer models were employed to learn baseline operational patterns and detect deviations indicative of equipment faults, human errors, or cyber threats. The framework was trained on a hybrid dataset comprising 15,000 normal operational instances and 3,500 labeled synthetic anomalies derived from simulated Supervisory Control and Data Acquisition (SCADA) environments. Evaluation metrics indicate that hybrid fusion achieved a precision of 0.94, recall of 0.92, F1-score of 0.93, and an area under the ROC curve (AUC-ROC) of 0.96, outperforming early and late fusion strategies by 6–10%. SHAP and LIME analyses provided interpretable insights into feature contributions, achieving an Explanation Satisfaction Index (ESI) of 0.89, reflecting strong operator trust. The results demonstrate that AI-driven multimodal anomaly detection, coupled with explainability, enables proactive fault identification, reduces false positives, and enhances operator situational awareness, providing a robust foundation for next-generation nuclear safety management.