Fault Detection and Diagnosis of Nuclear Power Plant Using Deep Learning Architecture

Being a safety-critical system, safety is extremely important for any nuclear power plant (NPP). Therefore, to maintain the safety of NPPs at an acceptable level, preventive measures are necessary to deal with potential issues. During plant operation, faults and failures can occur in sensors, equipment, and processes which can have impact on the performance of the plant. These faults are more prominent in aged NPPs because of their vulnerability to aging-related faults [1]. Hence there is need to monitor the status of the plant during operation. To do this, fault detection and diagnosis (FDD) techniques are developed and used in NPPs. FDD is the process of detecting and identifying unexpected behavior in a system. One of these techniques is data-driven methods, which comprises of artificial neural network (ANN), multivariate state estimation technique (MSET), principal component analysis (PCA), and autoassociative kernel regression (AAKR) [2]. However, the possibility and applicability of the deep learning – the current trend in the field of machine learning, to FDD of NPPs is not yet explored. Therefore, this work seeks to propose and apply deep learning techniques to FDD of NPPs. Deep learning originated from artificial neural network, and it is a branch of machine learning algorithms that use a cascade of many layers of non-linear processing units for feature extraction and transformation. It based on a set of algorithms that attempt to model/learn high level abstractions and hierarchy representations in data. There are various deep learning architectures, which include Restricted Boltzmann Machine (RBM) based deep belief network (DBN), Convolutional Neural Network (CNN), deep Auto-encoders, and Recurrent Neural Network (RNN). Recently, deep learning has been successfully adopted in various areas such as computer vision, automatic speech recognition, natural language processing, audio recognition and bioinformatics [3, 4], where they have been shown to produce state-of-the-art results on various tasks. To verify the applicability of the proposed deep learning model, we used the NPP simulation data for accident detection and identification. The verified model showed high performance applicability to FDD of NPPs.