Series AC Arc Fault Detection Model Based on Hybrid Time and Frequency Analysis

Series arc fault cannot be protected by a common purpose circuit breaker for its lower fault current amplitude compared with its normal one, hence arc fault circuit interrupters are required for avoiding potential fire hazards. This paper presents an ac series arc fault detection method based on hybrid time and frequency analysis and softmax classification neural network (HTFSCNN). An experimental platform capable of automatically recording normal and arc fault current waveforms is designed in order to collect data set. In this paper, four indicators in time-domain and six indicators in frequency-domain are selected as inputs of the HTFSCNN classifier, according to the characteristics of the current waveforms and frequency spectra. The conjugate gradient method was applied to train the backwards-propagation algorithm. The loss function was cross entropy and the output function was softmax. Experimental results show that this method can effectively separate the fault currents from the normal ones with accuracy of 98.74% under seven loads specified in IEC standards. Finally, the trained HTFSCNN model was implanted into a microcontroller and its feasibility is verified.