With the development of power system, arc fault detection becomes particularly important. In order to improve the detection efficiency and accuracy of arc faults, this paper proposes a semi-supervised learning method combining virtual adversarial samples and boundary uncertainty sampling to solve the problem of insufficient labeled data and low cost of unlabeled data. First, the model's resistance to unfamiliar disturbance is enhanced by introducing virtual adversarial samples to enhance the training set. Secondly, the boundary uncertainty sampling is used to improve the judgment ability of the model in complex cases, so that the model can diagnose faults more accurately in practical applications. The experimental results show that the proposed method is more accurate than the traditional supervised learning method in arc fault detection. Especially in the case of incomplete data labels, the diagnostic performance of the model is significantly improved by rational use of unlabeled data. This finding validates the effectiveness and potential of semi-supervised learning in dealing with electrical fault diagnosis problems with practical application value.