Nuclear power plants (NPPs) require rigorous monitoring systems for their safety and efficiency, thus extensive data are acquired continuously from instrumentation and control systems. Anomaly detection is one of the most widely used machine learning approaches for monitoring data, especially when available data for model training are limited or imbalanced, as are data from NPPs.

This research presents an anomaly detection system for centralized online monitoring in NPPs that is composed of two modules: data reconstruction and anomaly determination. Considering the large feature dimensions of the data, and leveraging their sequential characteristics in the time domain, four different autoencoder models for data reconstruction, long short-term memory, convolutional neural network, fully connected neural network, and principal component analysis are employed and compared.

Two anomaly determination methods are presented and analyzed from the perspective of the characteristics of residuals from the data reconstruction models. The developed system is validated with simulation data containing 239 process variables (sensors) from different subsystems in a NPP.

This paper highlights the effectiveness of simulations not only in overcoming the limited amount of data acquired from real plants in malfunctioning status, but also for evaluating the performance of the given models in a more quantitative way by comparing the performance at different malfunction levels. Weighted areas under the receiver operating characteristic curves are suggested as metrics for the validation of the given models and methods, and performance metrics, which can reflect engineers’ preferences, are demonstrated as well.