This study explores data-driven anomaly detection methods for analyzing sensor failures in the Advanced Gas Reactor (AGR) nuclear fuel irradiation experiments. Specifically, we examine failures pertaining to thermocouples (TCs), which are critical for monitoring and controlling in-reactor temperatures during reactor operation. Failures were primarily observed during abrupt power transitions and manifested as sensor drop-outs, drifts, or unexplained behavior. We applied three time-series analysis techniques—rolling mean (RM), matrix profile (MP), and vector auto-regression (VAR)—to detect anomalies in TC data prior to failure events. The RM method effectively highlighted deviations aligned with reported failures, while the MP method provided partial early warning. VAR shows potential for capturing multivariate dependencies, but requires further calibration for identifying minor failure precursors. Our findings demonstrate that conventional statistical tools can aid in anomaly detection but have limited predictive power for subtle precursors to anomalies, which may require a multivariate approach with richer training data. We propose future directions, including synthetic data generation, real-time surrogate modeling, and multi-modal feature integration. This work provides a foundation for applying robust anomaly detection frameworks to mission-critical sensor systems in experimental settings.
Kajihara et al. (Wed,) studied this question.