To address the difficulty of diagnosing series arc faults on the main lines of low-voltage distribution systems, a diagnostic method based on random forest (RF) and light gradient boosting machine (LightGBM) is proposed. The method achieves effective fault identification by analyzing voltage signals from sampling resistors on the main line. First, utilizing a series arc fault experimental platform with multiple branches and diverse load types, the characteristics of sampling resistor voltage signals are collected and analyzed. It is found that under various operating conditions, arc fault features typically appear as pulse spikes. Subsequently, the sampling resistor signals undergo low-frequency attenuation processing to enhance the high-frequency pulse characteristics of the arc. The sudden change in abnormal pulse count serves as the arc initiation criterion, and additional time-frequency domain features such as peak-to-peak value and power spectral entropy are extracted. Finally, RF is employed to select key features, and the LightGBM algorithm is used for arc fault diagnosis. Simulation results demonstrate that the proposed method achieves a diagnostic accuracy of over 96.67% under unknown multi-load operating conditions, validating its robust generalization capability and high recognition precision in complex circuit environments.