An experimental platform of photovoltaic (PV) system arc fault detection was built, and the current when the system was in operation was collected as the basis for arc detection. Afterwards, a method of combining the time- and frequency-domain features was used to detect the arc fault. Analysis results show that the changes in average time-domain current signals were significant in cases with and without arc fault. Moreover, in the frequency-domain, by decomposing the arc current and normal current into 5 layers using wavelet decomposition, the d5 frequency-band energy can be used to effectively distinguish among the normal, shadowed and fault running states. Finally, the arc fault was detected by combining the time- and frequency-domain features with the secondary arc fault detection method by means of threshold. Detection results show that the proposed method is effective and feasible.