Physics-Based Sensitivity and Explainable AI for Condition Monitoring of SMR Feedwater Control Valves

This study proposes a diagnostic framework for feedwater control valves in Small Modular Reactors (SMRs), integrating physics-based sensitivity analysis with machine learning (ML) prediction and explainable AI (XAI). Experimental measurements of valve opening (LVDT), flow rate (Q), and differential pressure (△P) were analyzed to compute △Q / △LVDT and △P / △LVDT, enabling classification into Normal, Warning, Inactive, and Abnormal states. A multi-layer perceptron (MLP) regression model was developed to predict Q and △P, with performance compared to linear, polynomial, and random forest models. The MLP showed stable tracking and low error variance across operating ranges. SHAP (SHapley Additive exPlanations) analysis revealed that LVDT contributions to Q and △P predictions were consistent with sensitivity analysis results, validating the physical reliability of the model and enabling root cause identification. The proposed approach enhances both reliability and interpretability, addressing the limitations of black-box ML models. Although validated under limited experimental conditions, it can be extended to various valve types and operating scenarios. Future work will focus on real-time optimization, comparative studies of XAI methods, and integration with digital twins for broader industrial application.