Multimode DALSTM model for anomaly detection of nuclear reactor core

Abstract

The nuclear reactor core is a life-critical system, whose reliable operation ensures the safety of a nuclear power plant. The reactor core temperature is a direct representation of the reactor core state. Therefore, the detection and analysis of core temperature anomalies is of great importance. However, multiple temperature variables are nonlinear and complexly coupled, showing different distribution characteristics with the variations of operating condition. For this reason, a multi-mode anomaly monitoring strategy for core temperature in nuclear power plant is proposed. To realize the nonlinear feature extraction of multiple temperature variables, a deep auto-encoder and long short-term memory (DALSTM) network are constructed. Considering the complex temperature distribution characteristics, the nonlinear features are clustered to divide the data into different subspaces. In this way, the distribution features are similar in the same subspace. Finally, local DALSTM models of core temperature in different subspaces are developed, and a multimode DALSTM monitoring strategy is designed. The monitoring statistics of the local models reflect the detailed variations of the process and achieve fine-grained detection of anomalous behavior. The effectiveness of the proposed model is verified by real operation data. The experimental results show that the method can achieve fast and accurate anomaly detection for reactor core temperature.