Development of an AI Agent for Failure Monitoring, Diagnosis, and Remaining Useful Life Prediction of Capacitors in Nuclear Safety System Signal Processors

Open Access
Article
Conference Proceedings
Authors: Sangjun Park, Chang Hwoi Kim
Abstract

This study proposes an AI agent that performs failure monitoring, diagnosis, and Remaining Useful Life (RUL) prediction for the electrolytic capacitors of nuclear power plants (NPPs) safety system signal processors and explains the situation and evidence to operators. An unscented Kalman filter (UKF) applied to estimate the thermal aging degradation state and detects abnormal states from the residuals of its estimates. A Large Language Model (LLM) converts each event into diagnostic information classified by severity, cause, evidence, and recommended action. For verification, the artificial dataset generated from actual measured data of accelerated thermal aging test. The artificial date set created by applying the random error to measured value and the intentional fluctuation is used for abnormal stat. The normal state produced no false alarm over eight weeks, whereas the abnormal state at week 9 was detected immediately. A two week bias signal and a step shaped degradation were discriminated as an anomalous state and a sustained state change, using only the persistence of the residual. Twenty eight of the failed specimens were detected at 9 week, and the RUL prediction error converged to 3.4 weeks. In another test cases, the AI agent alarmed 46 cycles ahead of end of life and temporarily fluctuation was not reflected the RUL as it is not affect the degression factor of an thermal aging. These results suggest that providing operators with verifiable monitoring and diagnosis explanations, together with the limits of the monitoring system, can improve the operating reliability of NPP safety systems.

Keywords: Nuclear Safety Systems, Electrolytic Capacitor, Physical Model Based State Estimation, Kalman Filter, AI Agent, Monitoring, Diagnosis, Prognostics, Human Error Reduction

DOI: 10.54941/ahfe1008401

Cite this paper
Downloads
0
Visits
1
Download PDF

More from this volume

← Designing and Implementing a Total Pilot Performance Course for Undergraduate Aviation StudentsFrom Emerging Technologies to Operational Capability: Human-Centred Transition Pathways for AI Implementation →
View all articles in Human Factors in Design, Engineering, and Computing →