Data-driven　fault　diagnosis　of　high　impedance　fault　(HIF)　has　received　increasing　attention　and　achieved　fruitful　results.　However,　HIF　data　is　difficult　to　obtain　in　engineering.　Furthermore,　there　exists　an　imbalance　between　the　fault　data　and　non-fault　data,　making　data-driven　methods　hard　to　detect　HIFs　reliably　under　the　small　imbalanced　sample　condition.　To　solve　this　problem,　this　paper　proposes　a　novel　HIF　diagnosis　method　based　on　conditional　Wasserstein　generative　adversarial　network　(WCGAN).　By　adversarial　training,　WCGAN　can　generate　sufficient　labeled　zero-sequence　current　signals,　which　can　expand　the　limited　training　set　and　achieve　the　balanced　distribution　of　the　samples.　In　addition,　the　Wasserstein　distance　was　introduced　to　improve　the　loss　function.　Experimental　results　indicate　that　the　proposed　method　can　generate　high-quality　samples　and　achieve　a　high　accuracy　rate　of　fault　detection　in　the　case　of　small　imbalanced　samples.　©　2021　IEEE.