For a wide variety of domestic appliances, the fault current waveforms among different types of appliances may be similar to normal current waveforms, which leads to the problem that traditional methods of fault arc identification cannot detect effectively, this paper presents a series low voltage fault arc identification method which combines time-frequency domain analysis and random forest which is suitable for a variety of typical loads working independently or mixed. Based on the correlation coefficients between the collected load spectra and the pure resistance load spectra, the loads are divided into switched-supply loads and non-switched-supply loads, then two random forest models are trained to identify the faults. A total of 33 723 sets of normal and fault current samples were collected to verify the proposed detection method, which proves that the proposed method can improve the recognition rate of fault arc.