Due to the low time?frequency resolution of S?transform (ST), it is difficult to accurately extract the characteristics of series fault arcs. In this paper, an improved incomplete S?transform (IIST) time?frequency analysis method was developed, where the center frequency measure is selected as the standard to screen the main frequency points in low and high frequency bands, and the Gaussian window coefficients corresponding to such low and high frequency bands is introduced, respectively. Firstly, a standard series arc fault test acquisition platform was built to collect current signals under different loads. Secondly, IIST was used for time?frequency analysis of the signals, and the corresponding features of low and high frequency bands were extracted to form the feature vector sample set. Finally, a fault arc diagnosis model was constructed to classify and identify the sample set. The results show that the recognition accuracy of this feature extraction method in support vector machine (SVM) is up to 98.29%, validating that the current fault features can be extracted effectively. By adding comparison experiments, the influence of different feature extraction methods and the SVMs with different kernel functions on diagnosis results was explored, which further verified the effectiveness of IIST and SVM fault diagnosis methods.