Accident diagnosis algorithm with untrained accident identification during power-increasing operation

Abstract

To ensure the safety of nuclear power plants (NPPs) from accidents or anomalies, regulatory bodies provide procedures that describe safety regulations that must be followed. However, even if well-designed procedures are provided to operators, diagnostic activity in an emergency scenario is classified as an extremely demanding task. Moreover, the diagnosis of accidents occurring under various operation modes, such as power increasing, is expected to be extremely difficult, owing to the diverse behaviors and availability of systems and components. With regard to such emergency response issues, artificial neural network-based methods are regarded as one of the most promising approaches, because of their noticeable achievements. However, regarding the application of neural networks, in the case of an untrained accident, there is no capability to answer "do not know." This study aims to develop algorithms that can cover various NPP operation modes and deal with untrained accidents. To address the various NPP operation modes, the major changes that can affect the plant states are classified. Furthermore, to deal with untrained accidents, the applied diagnostic algorithms use long short-term memory and an autoencoder. Following this, this paper presents the implementation and test results of the accident diagnosis algorithms.