Nuclear power systems unsupervised anomaly localization considering spatiotemporal information and influence mechanism between devices

The anomaly detection and localization methods based on unsupervised clustering models are more suitable for nuclear power systems operation monitoring than supervised classification models, especially in the absence of anomalies and faults training data in reality. However, existing anomaly localization methods ignore the mutual influences' differences between devices, and the effects of thermal and volumetric inertia. A novel unsupervised anomaly localization method for nuclear power systems is proposed to address these problems. The Devices Influence Relationship Directed Matrix is constructed based on Auto-Regressive Integrated Moving Average model and thermal-hydraulic mechanism to quantify the influence degrees between devices; The Spatiotemporal Graph Convolutional Networks are combined with the Auto-Encoder to extract parameters' spatiotemporal information and reconstruct systems operation data; Finally, anomalies are located based on the parameters' data reconstruction error trends. The novel method's effectiveness was validated based on two nuclear power systems anomalies datasets. The results show that compared to other state-of-the-art methods, the novel method has accuracy rates that are approximately 5 % higher for anomaly detection and 7.5 % higher for anomaly localization, respectively, and can alert three time steps in advance.