Digital Twin Framework for Nuclear Reactor Monitoring and Anomaly Detection

Artificial intelligence and digital twin technologies are increasingly being explored for reactor process analysis, anomaly detection, predictive maintenance, and operator decision support in nuclear engineering. However, many existing systems remain limited by static model behavior, reduced robustness to changing operating conditions, and insufficient integration of complementary monitoring functionalities within unified monitoring workflows. In this paper, a digital twin architecture is proposed for reactor condition assessment under multivariate operating conditions. The architecture supports integration of temporal analysis and reconstruction-based anomaly detection within a layered reactor-monitoring workflow, while defining adaptive recalibration support, uncertainty-estimation and explainability modules, and physics-informed integration components intended for future extension of the analytical monitoring architecture. Experimental validation is performed using the publicly available PWR Abnormality Dataset, containing multivariate measurements from a pressurized water reactor environment. The implemented and experimentally validated component corresponds to the data-driven monitoring module of the proposed architecture. A limited proof-of-concept adaptive threshold recalibration experiment under simulated operating drift is additionally performed to illustrate one adaptive monitoring mechanism. The comparative evaluation includes temporal data partitioning, sliding-window sequence generation, controlled synthetic perturbations, anomaly score normalization, temporal smoothing, threshold optimization, and evaluation using Precision, Recall, F1-score, and ROC-AUC metrics. Several anomaly detection approaches are comparatively evaluated, including temporal residual analysis, robust multivariate monitoring, PCA-based reconstruction, anomaly score fusion, and classification using Gradient Boosting. The results show that PCA-based reconstruction achieves the highest F1-score, while score-level classification achieves the highest ROC-AUC value. The performed experiments validate the monitoring layer of the proposed architecture and demonstrate the feasibility of integrating multiple anomaly-detection approaches within a unified reactor-monitoring workflow. The study contributes a digital twin architecture for reactor monitoring together with comparative validation of its analytical monitoring layer under controlled multivariate reactor-monitoring conditions.