Abstract:
Next-generation nuclear reactors pose challenges for effective system monitoring and data management, necessitating the differentiation of anomalous sensor data from noise. While digital twin models hold promise, fundamental analyses using simplified models are needed. This study proposes a foundational approach utilizing simulated data from PCTRAN-generated datasets to emulate reference pressurized-water reactor (PWR) conditions and introduces typical sensor performance and anomalies. The method employs data partitioning, linear regression for preprocessing, and a K-means clustering algorithm for anomaly detection, achieving over 95% precision in identifying anomalies. A parametric study using Monte Carlo Sampling on the anomaly detection algorithm’s input values reveals critical factors such as window size’s impact on accuracy and computational time. Utilizing the Risk Analysis Virtual Environment (RAVEN) tools, a Pareto optimal frontier is determined to balance accuracy and execution time. Sensitivity and uncertainty analysis highlight window size as a critical factor. While further refinement is necessary for practical application, these techniques show promise for enhancing nuclear system monitoring and data management. © 2024 by The United States Government.
