Anomaly sign detection by monitoring thousands of process values using a two-stage autoencoder

Abstract

In a large-scale plant such as a nuclear power plant, thousands of process values are measured for the purpose of monitoring the plant performance and the health of various systems. It is difficult for plant operators to constantly monitor all of the process values. We present a data-driven method to comprehensively monitor a large number of process values and detect early signs of anomalies, including unknown events, with few false positives. In order to learn the complex changing internal state of a nuclear power plant and accurately predict the normal process values, we created a two-stage autoencoder composed of a time window autoencoder and a deviation autoencoder, which is a deep learning network structure corresponding to the characteristics of the process values. We assessed performances of the two-stage autoencoder with simulated process values of a nuclear power plant, a 1,100 MW boiling water reactor having 3,100 analog process values. In situations where it is difficult to predict the normal state (rapid operation mode change, transient state, and small fluctuations in the process values), the two-stage autoencoder properly predicted the normal process values and showed excellent performances with early detection and zero false positives, except for one case. The two-stage autoencoder would be an effective solution for comprehensive plant monitoring and early detection of anomaly signs.

Keywords

• plant monitoring
• big data
• anomaly sign detection
• machine learning
• deep learning
• autoencoder
• multivariate time series