[2F15-18] Advanced Operation & Management Technology for Nuclear Power Plant

[2F15] Development of Abnormal Sign Detection System using AI for Nuclear Power Plant

(6)Building an Architecture that Incorporates the Knowledge of Engineers into a Detection Algorithm and Enhances Detection Performance

*Susumu Naito¹, Yasunori Taguchi¹, Yuichi Kato¹, Kouta Nakata¹, Isaku Nagura², Shinya Tominaga², Ryota Miyake², Toshio Aoki², Chikashi Miyamoto², Yusuke Terakado² (1. TOSHIBA, 2. TOSHIBA ESS)

Keywords:Two-stage autoencoder, Plant performance monitoring, Plant health monitoring, Anomaly detection, Deep learning

In a nuclear power plant, thousands of process values are measured in order to monitor plant performance and the health of various systems. It is difficult for plant operators to constantly monitor all of the process values. To address this, we are developing a detection algorithm based on deep learning to monitor them and to enable early detection of anomaly signs including unknown events with few false detections. In actual process data, process values that have no physical relationship may coincidentally show similar fluctuation trends. As a result, the algorithm mistakenly learns that there is a physical relationship between them, which can cause false positives. The presence or absence of a physical relationship cannot be determined by process data alone. Therefore, we improved the algorithm so that engineers can easily assign the relationship by simply dividing the process values into two classifications. The effect of this improvement was verified using plant simulator data. It was confirmed that false positives due to incorrect learning were eliminated and detection performance was enhanced.