System and Component Anomaly Detection Using LSTM-VAE

In the event of an accident at a nuclear power plant, the operators have to take appropriate actions after carrying out the diagnosis of the accident. However, the accident diagnosis can cause the human error because complex procedures have to be performed quickly within a limited time. Accordingly, researches using artificial intelligence are actively being conducted to reduce the occurrence frequency of human errors that may occur in diagnostic tasks. Most studies use a supervised learning strategy to assist operators in diagnostic tasks using artificial intelligence. However, there is a problem that the supervised learning strategy cannot be handled properly when untrained data is input. Therefore, this paper aims to provide information to operators by adopting an unsupervised learning strategy that does not cause such a problem. Therefore, we intend to detect abnormalities in the systems and components of nuclear power plants by utilizing long short-term memory variational autoencoder, an artificial intelligence methodology. It is expected that the results of detecting anomalies in systems and components will help operators in diagnosing and mitigating accidents.