Signal fault identification in nuclear power plants based on deep neural networks

Abstract

In large and complex industrial systems such as nuclear power plant (NPP), numerous instrumentation signals are collected in order to support operators to make proper decisions. Especially, the importance of reliable instrumentation signals is more emphasized under harsh conditions, as wrong decisions for such systems could result in massive casualties and financial loss. Under harsh conditions, multiple instrumentation signals can become faulty simultaneously, and these signals should be separated from normal signals as they could induce excessive confusion. However, identification of faulty signals under NPP emergency situation is challenging, since plant conditions and corresponding measurements can be diversified due to many factors. In this study, the authors proposed variational autoencoder (VAE)-based signal fault identification method under NPP emergency situations. In detail, VAE is pre-trained to reconstruct normal signal set, and it identifies faulty signals by comparing element-wise deviation between input and output and finding elements within faulty signal set that induces reconstruction error. To validate the proposed method, experiments based on simulation data were conducted. The results shown that the proposed method can effectively identifies faulty signals under various emergency situations in NPP with acceptable performance, without prior knowledge on plant conditions.