Development of anomaly detection system using AI Anomaly detection AI technology, two-stage autoencoder
内藤 晋 田口 安則 加藤 佑一 中田 康太 （東芝） 名倉 伊作 富永 真哉 三宅 亮太 青木 俊夫 宮本 千賀司 高戸 直之 （東芝ESS）
anomaly detectionautoencoderdeep learningmultivariate time series dataplant monitoring
In large-scale power plants such as nuclear and thermal power plants, thousands of sensors are installed to monitor the performance and health of the plant. However, it is difficult for operators to monitor all sensor values at all times. Therefore, Toshiba has developed a unique AI technology that comprehensively monitors thousands of sensor values and detects possible abnormalities earlier than humans. In a power plant, there are major fluctuations in sensor values during operation and minute fluctuations in sensor values. It has been difficult to learn these fluctuations with enough accuracy to identify minute abnormalities buried in them. Toshiba has developed a technology to detect signals buried in the fluctuations with high accuracy by enabling learning through a deep learning network structure that corresponds to the characteristics of the plant's sensor values. The effectiveness of this technology was confirmed by using simulation data of about 3,000 sensors generated with a nuclear plant simulator.