Series arc fault caused by aging damage of the line insulation layer and poor electrical contact seriously threatens the power safety of low voltage distribution systems. And it is difficult to detect and extinguish series arc fault for its characteristics of small current, high temperature and strong concealment. Because of the above reasons, a method based on wavelet feature and deep learning is proposed for detecting series arc fault. Firstly, series arc fault experiments platform was built to collect the current signals under typical resistive loads, inductive loads and resistive-inductive loads. Secondly, after transformed by wavelet transform, collected signals were decomposed to construct training sets and test sets. Finally, the arc fault was identified by the improved AlexNet model, and the test results were output. The experimental results show that the accuracy of this method for serial arc fault identification is almost 95.58%, about 10.58 percentage points higher than using AlexNet model.