In the DC power system, arc fault caused by insulation defects and loose joints is easy to cause fire, explosion, and other safety accidents, but it is not easy to extinguish and detect. Therefore, it is of great significance to study the electrical characteristics of the DC arc and propose effective detection methods. Detection of DC arc faults in photovoltaic systems has attracted widespread attention from scholars at home and abroad. In this paper, several experimental conditions that may affect arc characteristics are analyzed. By using the photovoltaic strings and the arc generator, a lot of field experiments are carried out to obtain the current data during arc fault in an actual photovoltaic system under different conditions of current, cable lengths, and arc gaps. According to the experimental data under different conditions, several characteristic variables from both the time domain and the frequency domain are extracted to identify the arc fault. In this paper, three machine learning methods in the field of artificial intelligence: Random Forest, Support Vector Machine, and Bayes Classifier, are used to learn the experimental samples. Through training, a model is obtained in each method to distinguish between normal current and arc fault current. Then, the three classifier models are validated by hundreds of sets of experimental data under normal and arc conditions, and the classifier with the highest accuracy is found. According to the results, the proposed method can accurately identify arc fault conditions and normal working conditions. Among them, the random forest algorithm has the highest recognition accuracy.