It is difficult to identify the arc fault effectively when the loads in the user-side are more complicated, blocking the development of low-voltage monitoring and pre-warning inspection. In this paper, series arc fault signals are acquired according to IEC 62606. The main time-frequency features can be strengthened more effectively by the generalized S-transform with bi-Gaussian window, meanwhile the power spectrum density (PSD) determination allows for the detection of imperceptible high-frequency harmonics energy reflections, increasing the rate of arc fault diagnosis and suitable for the arc fault monitoring of nonlinear loads. The final samples are trained and classified by two-dimensional Convolutional Neural Network (CNN) and the overall accuracy of identification is 98.13%, of which involves various domestic loads, providing a reference for the follow-up arc fault monitoring and inspection research.