Designing　a　neural　network　using　simulation　waveforms　for　high-impedance　fault　(HIF)　detection　is　a　challenging　task,　particularly　in　case　of　applications　on　distribution　networks.　Based　on　the　characteristics　of　HIF　transient　zero　sequence　currents　(TZSCs),　a　novel　methodology　is　proposed　for　HIF　detection　in　the　time　domain,　which　includes　1-D　variational　prototyping-encoder　(1-D-VPE)　and　decision　tree　(DT)　algorithm.　The　proposed　HIF　detection　scheme　consists　of　four　steps:　First,　the　TZSCs　collected　from　simulations　are　applied　to　train　the　1-D-VPE　neural　network,　whose　encoder　is　used　as　the　feature　extractor.　Then,　the　extracted　features　are　utilized　to　train　the　DT　classifier.　After　that,　the　well-trained　encoder　and　DT　are　employed　to　test　field　TZSCs.　Finally,　the　1-D-VPE　neural　network　is　fine-tuned　with　some　of　the　field　TZSCs.　The　performance　of　the　proposed　method　is　analyzed　using　waveform　samples　collected　from　the　PSCAD/EMTDC　simulation　platform　and　the　actual　distribution　networks.　The　results　show　that　the　proposed　method　provides　good　classification　performances　on　both　simulation　and　field　data　at　high　noise　levels　without　transformation　between　different　domains,　and　the　classification　accuracy　can　be　improved　by　fine-tuning　with　some　field　TZSCs.