The electric arc fault is an important cause of electrical fire. When the series arc fault occurs in the low-voltage circuit, the traditional feature extraction method cannot fully express all the data features of the time domain signal. The feature expression ability of arc fault is limited, which may bring high false alarm rate and miss alarm rate of detection results. To solve this problem, an arc fault detection method based on the self-normalized convolutional neural network is proposed. This method intercepts the current time series of different kinds of loads according to half period. Then, they are normalized. The two-dimensional images of arc faults and normal operation are generated by the grayscale data transformation. The gray transformation features of arc faults are extracted by using the convolutional neural network. The arc fault convolution features are identified by multi-layer full connection layer fitting calculation of the following sampling information. The evaluation shows that the accuracy of the proposed method is 99. 67% , which is better than the traditional convolutional neural network and has good generalization performance.