In recent years, electrical fires occur frequently. Arc fault is a important causes of electrical fires. In this paper, considering the characteristics of the low-voltage customer scenarios, the research of non-invasive arc fault detection is carried out. First, the aggregated load current waveform data is acquired at the entrance of the customer's power supply. Then, the amplitude and phase information of the fundamental and each harmonic wave of the total current is obtained by harmonic analysis. Next, the total current and the current characteristic parameter matrix obtained from training are used together to construct the objective function and form a multi-load current decomposition model. Finally, the intelligent optimization algorithm is adopted to optimize the solution to obtain the operating state of each appliance (including the fault states), and identify the arc fault and analyze its causes. In addition, this paper carries out the simulation experiment of arc fault for common appliances of actual low-voltage users in the laboratory, and the experimental results show that the proposed non-invasive arc fault recognition method is effective.