Input data dimensionality reduction of abnormality diagnosis model for nuclear power plants

Abstract

Nuclear power plants are diagnosed by operators according to the alarms and plant parameters that can be identified in the main control room. The operators are trained to conduct tasks in any cases by following an operating procedure. When a component has in malfunction, the operator must choose the appropriate abnormal operating procedures to stabilize the plant. However, the operators take a high burden because this task requires complex judgement with large amounts of information in a short time. To support the operators, this paper studied the data preprocessing methods to develop the nuclear power plant abnormal state diagnosis system using deep learning algorithms. A nuclear power plant simulator was used to produce training data which includes more than 2800 variables recorded in the given time. It is necessary to reduce the dimensionality of the generated data to achieve the best estimation of the training. There are two ways to reduce the dimensionality of the data: feature selection and feature extraction methods. Abnormal operating procedures of the advanced pressurized water reactor 1400 were analyzed to select parameters related with abnormal events. On the other hand, principal components analysis was used as one of the feature extraction methods. Preprocessed data through two methods were trained by the same deep learning algorithm, gated recurrent unit. The data selected by humans and the data extracted by considering the relationship among the variables showed different performance for diagnosing the plant state. The results showed that it is advantageous for the developing diagnosis model to learn and judge through the feature extraction method.