Series arc fault detection based on multi-domain depth feature association

In low voltage distribution systems, series arc fault current is small and hidden, and traditional circuit protection devices cannot effectively identify it. To address this problem, a series arc fault detection method based on multi-domain depth feature association is proposed in this paper. By building an experimental platform, the current signal data on normal and arc fault states under different loads are obtained. The time-domain features, frequency-domain features and wavelet packet energy features of the current signal are extracted. To enhance the quality of the data, the importance of each domain characteristic is ranked using four distinct tree techniques, and the most useful features are chosen. A one-dimensional stacked neural network (1D-SNN) fault detection model is constructed to further extract the depth features of each domain. To achieve series arc defect detection, the depth features are combined and fed into a fully connected neural network. The Radam algorithm is used to optimize the detection model. It is then compared with Adam, SGD, and RMSprop optimization algorithms, which verifies that Radam has a better effect on the optimization of the arc detection model. Experimental results show that the average detection accuracy is 99.63%.