RNN-Based online anomaly detection in nuclear reactors for highly imbalanced datasets with uncertainty

• Published in: Nuclear engineering and design 2020-08, Vol.364 (C), p.110699, Article 110699
• Main Authors: Kim, Minhee, Ou, Elisa, Loh, Po-Ling, Allen, Todd, Agasie, Robert, Liu, Kaibo
• Format: Article
• Language: English
• Subjects: Anomalies; Condition monitoring; Error analysis; Historical account; Internet; Neural networks; Nuclear engineering; Nuclear reactors; Nuclear safety; Recurrent neural networks; Uncertainty
• ISSN: 0029-5493; 1872-759X
• DOI: 10.1016/j.nucengdes.2020.110699

•Several advanced techniques are proposed to preprocess data from nuclear power plants.•Neural networks capture complex relationships between signals and operating condition.•Anomaly detection performance can be improved using data from abnormal operations. Accurate online condition monitoring and anomaly detection are crucial in nuclear applications to optimize economic performance and minimize safety risks. To achieve this goal, several major challenges exist which must be addressed. First, multi-sensor signals are often collected in the form of complex, multivariate time series. Second, relatively few anomaly records are available to train detection models. Lastly, the recorded data may contain uncertainties resulting from various sources, such as operator-induced variability and measurement error. In this paper, a recurrent neural network-based approach is proposed to tackle these issues by effectively utilizing historical data obtained during both normal and abnormal operations. Several advanced data preprocessing techniques are developed to improve the training process of the proposed neural network. The efficiency and sensitivity of the proposed method are evaluated on the multi-sensor signal measurements and operational reports obtained from a real case study. The results demonstrate much improved detection accuracy and practicality of the proposed method over conventional approaches.