The　existing　residual　current　device　(RCD)　operates　based　on　the　amplitude　of　the　residual　current,　but　if　the　threshold　is　not　reasonably　set,　the　RCD　is　prone　to　reject　or　misoperate.　Therefore,　identifying　biological　electric-shock　faults　from　grounding　faults　is　a　crucial　approach.　Current　research　only　selects　one　or　several　features　without　following　proper　feature　selection　rules.　Furthermore,　machine　learning　methods　require　a　certain　number　of　samples　to　train　the　model　to　ensure　algorithm　accuracy　and　stability.　However,　obtaining　a　large　number　of　biological　electric-shock　samples　is　challenging　during　actual　experiments,　and　the　algorithm　model　cannot　learn　the　waveform　in　real　settings.　To　solve　the　above　problems,　a　biological　electric-shock　fault　identification　method　based　on　multi-feature　optimization　selection　under　unbalanced　small　samples　is　proposed.　Firstly,　variational　auto-encoders　(VAE)　is　adopted　to　multiply　the　electric-shock　small　sample　data　collected　by　experiments　to　achieve　positive　and　negative　sample　balance.　Due　to　the　complexity　and　danger　of　the　scenes,　it　is　difficult　to　obtain　the　actual　electric-shock　samples.　The　problem　of　small　samples　will　lead　to　low　accuracy　and　poor　effectiveness　of　the　training　model,　and　the　unbalanced　samples　will　lead　to　deviations　in　the　prediction　results　of　the　model,　resulting　in　poor　identification　accuracy　of　a　few　types　of　samples.　Therefore,　a　few　samples　are　enhanced　by　introducing　VAE　to　improve　the　effectiveness　of　the　model.　Secondly,　23　features　which　can　reflect　the　dynamic　characteristics　of　the　waveform　are　extracted　in　time　domain,　the　optimal　expression　feature　group　is　selected　from　them　by　Gaussian　kernel　Fisher　discriminant　analysis　(GKFDA)　and　maximal　information　coefficient　(MIC).　Through　data　analysis,　various　index　features　can　be　extracted　from　the　changing　forms　of　biological　electric-shock　waveforms.　The　addition　of　high-quality　features　will　improve　the　diagnostic　accuracy　of　the　classifier　to　a　certain　extent,　but　the　introduction　of　bad　and　redundant　features　will　increase　the　running　time　of　the　algorithm　and　reduce　the　diagnostic　accuracy　of　the　classifier.　Therefore,　GKFDA　and　MIC　are　combined　to　perform　feature　scoring　for　each　feature,　and　the　optimal　expression　feature　group　is　selected　intuitively　and　independently　based　on　the　scoring　results,　which　could　improve　the　feature　quality　and　reflect　the　regularity　of　feature　selection.　Finally,　a　forgetting-factor-based　online　sequential　extreme　learning　machine　(FOS-ELM)　algorithm　is　investigated　to　identify　the　electric-shock　behavior.　There　are　abundant　electric-shock　scenes　in　the　real　environments.　The　escape　behaviors　of　living　objects　during　electric　shock　will　have　a　great　influence　on　the　electric-shock　waveform,　which　makes　it　difficult　for　the　traditional　off-line　classifier　to　have　adaptability.　The　online　sequential　extreme　learning　machine　(OS-ELM)　has　an　online　learning　mechanism　that　allows　online　updates　for　new　samples　without　the　historical　data.　The　forgetting　factor　is　introduced　to　form　FOS-ELM,　aiming　to　further　solve　the　shortcoming　of　slow　learning　speed　of　OS-ELM,　so　that　it　can　quickly　adapt　to　changes　of　environmental　samples　with　higher　learning　efficiency.　The　experimental　data　of　conventional　grounding　fault　and　biological　electric-shock　fault　in　12　scenes　were　collected　for　the　verification　of　the　proposed　algorithm.　The　results　show　that　the　diagnosis　accuracy　of　the　proposed　model　can　reach　98.75%,　among　which　all　40　conventional　grounding　fault　samples　are　correctly　judged　with　an　accuracy　of　100%,　while　only　1　of　40　actual　biological　electric-shock　fault　samples　is　wrong　with　an　accuracy　of　97.5%.　From　the　perspective　of　time,　the　average　online　learning　time　is　1.378　ms,　and　the　average　diagnosis　time　is　only　1.33　ms.　©　2024　China　Machine　Press.　All　rights　reserved.