Time Series Analysis With Combined Learning Approach for Anomaly Detection in Nuclear Power Plants

Abstract Timely and accurate anomaly detection is of utmost importance in ensuring nuclear safety to guarantee the normal operation of nuclear facilities and to prevent severe accidents. Due to the complexity of the dataset from the condition-based monitoring, advanced data-driven deep learning algorithms with their state-of-the art performance, are deployed for automatic anomaly detection in nuclear power plants (NPPs). Nevertheless, in the collected dataset, the knowledge about whether some certain periods are in an abnormal or normal state is available, while for a greater portion of the time, only data in unknown operating states can be acquired, rendering the general supervised learning (SL) and unsupervised learning (UL) suboptimal learning methods. Therefore, a combined learning approach is deployed in the proposed time series analysis framework, allowing the utilization of additional domain knowledge within the dataset to train the framework from SL to UL seamlessly. Additionally, due to the intricate inherent relationships among different instruments and the low frequency of anomaly occurrences, the implicit influential weights of each dimension vary when identifying different kind of anomalies. To tackle this challenge, an attention mechanism is integrated into the anomaly detection model for highlighting the prominent abnormal features and hence enable it to address the subtle impact of each dimension to enhance the ability of anomaly detection task. Experimental evaluations on the real-world multivariate benchmark dataset Security Water Treatment (SWaT) and on the high temperature gas-cooled reactor (HTGR) anomaly cases dataset based on the analytical code ‘ACCORD’, demonstrate the feasibility and effectiveness of the proposed approach.