Fault Diagnosis of DCS SMPSs in Nuclear Power Plants Based on Machine Learning

Distributed control system (DCS) is the core digital instrumentation and control (I&C) equipment and research related to its predictive maintenance is highly valued by the industry. Switched-mode power supply (SMPS) circuit modules are widely used in DCS boards, and their fault can cause the board to fail and may even disrupt the safe and economical operation of the nuclear power plant (NPP). In this study, a machine learning-based board-level fault diagnosis method is proposed for an SMPS circuit module of the DCS board in an NPP. Support vector machine based on particle swarm optimization (PSO-SVM) and one-dimensional convolutional neural network (1D-CNN) models are developed based on traditional machine learning and deep learning, respectively. Furthermore, wavelet packet transform is used for circuit fault feature extraction of the PSO-SVM model. Based on the aging samples of aluminum electrolytic capacitors (AECs) obtained from the accelerated life test (ALT) and their aging process data, the waveform data of the output voltage of SMPS under the corresponding fault modes are obtained via circuit simulation and hardware experimental tests. The influence of aging of the two output filter capacitors on the SMPS output is analyzed, and the feasibility of the two fault diagnosis methods is verified. All developed fault diagnosis models exhibited good diagnostic performance. The research results can provide application reference with practical engineering significance for the predictive maintenance of DCS boards.