System-level Anomaly Detection for Nuclear Power Plants Using Variational Graph Auto-encoders

Abstract

In order to efficiently identify the system-level anomalies of the nuclear power plants (NPP), a Variational Graph Auto-encoder (VGAE) anomaly detection method with coarse-grained feature input is proposed to solve the large-scale unlabeled and multi-source coupled operational data of NPP. First, detrended cross-correlation analysis (DCCA) was used to quantitatively evaluate the correlation between variables, the multivariate coupling networks were constructed, and the weakly correlated edges were removed. Second, the first-order difference sequences of variables were symbolized. The symbolic values represent the fluctuation characteristics of the variables at the corresponding time. Based on the above input, a semi-supervised learning VGAE model was established, the reconstruction loss of real-time operating data is considered as the anomaly detection indicator. Finally, it was proved by a case of an actual NPP circulating water system. The results show that comprehensive consideration of the fluctuation characteristics and correlation characteristics of the operating data can effectively improve the accuracy of the anomaly detection model. Compared with traditional network evaluation indicators, such as network structure entropy, the proposed reconstruction loss has higher sensitivity and accuracy, and can realize early anomaly detection.