Abstract: It is difficult to identify the arc fault effectively when the loads on the user side have become increasingly complex, which blocks the development of fault monitoring and pre-warning inspection. In this paper, the time–frequency analysis and identification of series arc fault was studied based on the generalized S-transform. Firstly, the differences in time-frequency features of arc faults among 3 signal processing methods, STFT (Short-time Fourier Transform), wavelet transform and generalized S-transform were compared, highlighting the advantages of generalized S-transform in processing high-frequency features of nonlinear loads. After that, the bi-Gaussian generalized S-transform was used to receive time-frequency features of the nonlinear loads and construct image feature samples. Finally, the samples are trained and classified by 2D-CNN (two-dimensional Convolutional Neural Network), and the recognition effectiveness was verified by the accuracy and clustering analysis. The overall accuracy is 96.81%, of which involves various domestic loads, providing a reference for the follow-up arc fault monitoring and inspection research.