A Series Arc Fault Detection Method Based on Multi-layer Convolutional Neural Network

Arc faults in low-voltage distribution networks are one of the important causes of fires. The current value of the series arc is generally too small to reach the setting value of the overcurrent breaker. Under certain load conditions, it is usually difficult to identify the series arc fault because its current has the similar waveforms as the normal working current. In order to solve this problem, a time domain visualization convolutional neural network (TDV-CNN) is proposed based on a multi-layer convolutional neural network. First, the high frequency signals of series arc faults are acquired with a high frequency coupled filter circuit and a high speed data acquisition system. Second, the high-dimensional features of the arc image are extracted by constructing a multi-layer convolutional neural network. Third, the abstract feature extraction of the arc fault data by the convolutional neural network algorithm is visualized in the form of the time-domain grayscale image. Finally, by comparing with and analyzing the other machine learning prediction algorithm, the proposed methodology is believed to be reliable for the series arc detection with relatively higher accuracy and also has important potential application in other fault diagnosis. © 2020, Power System Technology Press. All right reserved.