ADQ—Anomaly Detection and Quantification From Delayed Neutron Monitoring Data of Nuclear Power Plants
Yogita Yogita, Durga Toshniwal, Pramod Gupta, Vikas Khurana, Pushp Upadhyay · IEEE Sensors Journal · 2023
Identifying anomalies and quantifying their severity from delayed neutron monitoring data are of great significance to ascertain the reason of anomalous behavior and for predictive maintenance of nuclear power plants (NPPs). An existing anomaly detection technique is primarily aimed to identify anomalies and does not consider quantification of anomalies. In this work, an anomaly detection and quantification (ADQ) technique has been proposed. It detects anomalies based on the fact that abnormal behavior occurs rarely in the real world as opposed to normal behavior. Thereby, the dissimilarity of an anomalous data instance happens to be large from a normal data instance in comparison with dissimilarity between normal data instances and anomalies in itself. This has been exploited by the proposed technique by employing the concept of candidate anomaly pair. It computes a quantified value for severity of given anomaly as intensity in reference to underlying normal behavior, representative of which is extracted from data itself. The proposed technique has been applied on delayed neutron monitoring data of 28 sensors for an NPP, which was provided by Nuclear Power Corporation of India Ltd. (NPCIL), Mumbai, India. The detected anomalies are categorized as point and interval anomalies, and sensors are ranked based on them. The performance of the proposed technique has been found highly promising on validating results with domain experts. The proposed technique has outperformed other three state-of-the-art techniques in terms of anomaly detection rate (ADR), false alarm rate (FAR), and precision.