The　accurate　identification　of　fault　types　is　the　primary　measure　when　disposing　grid　faults.　Based　on　singular　value　decomposition　(SVD)　of　time-frequency　matrix　and　support　vector　machine　(SVM),　a　method　to　identify　distribution　network　faults　is　proposed.　By　applying　local　characteristic-scale　decomposition　(LCD),　Hilbert　transform　as　well　as　band-pass　filter　on　the　voltage　waveforms　of　grid　bus　and　current　waveforms　of　low-voltage　inlet　line　of　the　main　transformer,　a　group　of　time-frequency　matrices　are　constructed.　Then　these　time-frequency　matrices　are　decomposed　by　SVD　to　obtain　corresponding　singular　spectra,　of　which　the　distributed　parameters　are　extracted　as　feature　vectors.　By　taking　these　vectors　as　the　input　of　multilevel　SVM,　grid　faults　can　be　identified.　The　results　of　simulation　and　experiments　under　typical　working　conditions　show　that　the　accurate　rate　of　proposed　method　can　reach　90%,　revealing　that　the　identification　of　various　faults　can　be　realized　effectively.　In　addition,　the　proposed　method　is　highly　adaptable　and　practicable.　©　2017,　High　Voltage　Engineering　Editorial　Department　of　CEPRI.　All　right　reserved.