Aiming　at　the　problem　of　the　high　acquisition　cost　of　high　impedance　fault　(HIF)　labeled　data　in　distribution　networks　and　the　difficulty　of　using　unlabeled　data,　this　paper　proposes　a　novel　HIF　semi-supervised　detection　method　based　on　tri-training　and　support　vector　machine　(SVM).　Unlike　supervised　learning　methods,　this　method　can　use　labeled　and　unlabeled　data　by　tri-training.　Firstly,　discrete　wavelet　transform　decomposes　the　zero-sequence　currents　into　different　wavelet　coefficients　and　extracts　special　features.　Secondly,　three　SVM　classifiers　with　different　kernel　functions　are　collaboratively　trained　to　construct　a　semi-supervised　classifier.　Finally,　the　method　is　verified　based　on　the　PSCAD/EMTDC　simulation　software.　The　simulation　results　show　that　the　proposed　method　can　utilize　massive　unlabeled　data　to　improve　fault　detection　performance　and　reflect　the　differences　between　the　classifiers　through　different　kernel　functions　of　SVM,　which　further　improves　the　effect　of　tri-training.　©　2022　IEEE.