To solve the problem that the identification of a high impedance grounding fault (HIF) in distribution networks is easily affected by noise, and the fact that it is difficult to use unlabeled data, a semi-supervised identification method of a high resistance grounding fault based on wavelet denoising and random forest is proposed. Different from supervised learning only using labeled data, the method can make full use of labeled and unlabeled data by collaborative training. First, the wavelet threshold denoising algorithm is used to filter the noise of zero-sequence currents. Secondly, the occurrence of an HIF can be detected by the peak and valley fault triggering algorithm. Then, applying wavelet transform to zero-sequence currents, the wavelet coefficients are extracted as fault features. Finally, two random forests are collaboratively trained with selected features to construct a semi-supervised classifier to detect the HIF. The simulation results show that the proposed method can use fully the key features in unlabeled data in the existed fault cases in distribution network to improve the accuracy of fault calssification. It has strong reliability and flexibility. This work is supported by the Science and Technology Project of China Southern Grid Co., Ltd. (No. GDKJXM 20198281).