Abnormal event detection, identification and isolation in nuclear power plants using LSTM networks. (October 2021)

• Record Type: Journal Article
• Authors: Wang, Meng-Die; Lin, Ting-Han; Jhan, Kai-Chun; Wu, Shun-Chi
• Is Part Of: Progress in nuclear energy. Volume 140(2021)
• Publication Date: 2021-10
• Subjects: Long short-term memory (LSTM) -- Deep learning -- Event detection -- Event identification -- Unseen event isolation; Nuclear energy -- Periodicals; Nuclear engineering -- Periodicals; 333.7924
• DOI: 10.1016/j.pnucene.2021.103928
• Languages: English
• ISSNs: 0149-1970

Abstract

Abstract: Increases in concerns regarding system safety and reliability challenge nuclear energy's attractiveness among the public. To alleviate such concerns, being able to prevent a developing event from escalating into a severe accident is indispensable, which requires an abnormal event to be identified in its early stage. In this study, several long short-term memory (LSTM)-based networks for abnormal event detection, identification, and isolation are proposed to help maintain the safe operations of nuclear power plants (NPPs). With the proposed model for predicting normal operation sensing readings, an abnormal event is detected if the discrepancies between the acquired and predicted readings exceed a preset threshold. Through LSTM's superior capability in time series analysis, the process information for sensing reading generation and the interrelations among the sensors in the event recordings can be extracted to enable valid event identification. Via the proposed autoencoders for sensing reading reconstruction, the plausible type for the ongoing event can be further verified to prevent an unseen event from being wrongly linked to any class in the event set. Results from experiments utilizing data of 13 event classes generated by a Maanshan NPP simulator illustrate the efficacy of the proposed models. Highlights: LSTM-based networks for abnormal event detection, identification, and isolation in nuclear power plants. An abnormal event is detected if discrepancies between the acquired and predicted readings exceed the preset threshold. LSTM's superior capability in revealing discriminant information in time series enables valid event identification. Linking an unseen event to a collected event class can be avoided through the proposed autoencoders.