A PV DC series arc fault has the characteristics of randomness and concealment, and it is easily affected by the external environment and internal noise of the PV system, making it difficult to detect. The current time-frequency domain features extracted by wavelet transform can identify an arc fault very well, but it faces the problem of optimal wavelet base selection. Based on the collection of a large amount of arc fault data, this paper proposes an optimal wavelet base selection method for the extraction of commonly used arc fault characteristic indicators through wavelet transform analysis and comparative experiments. By this method, the bior4.4 wavelet base is determined to be the optimal wavelet base for extracting arc fault features, and the time-frequency domain features are constructed based on bior4.4 stationary wavelet transform. Through comparative experiments, it is found that the time-frequency domain feature based on bior4.4 can significantly improve the identification of an arc fault, and shows the suppression effect on normal noise signals. To reflect the characteristics of arc faults from multiple angles, it complements time-domain features, combines with time-frequency domain features to form a current feature library, and uses the random forest algorithm to realize the diagnosis of arc faults. The accuracy rate of arc fault detection reaches 98.58%, and the misjudgment rate of normal signal is only 0.76%.