The　accurate　identification　of　over-voltages　is　the　primary　task　when　disposing　over-voltage　accidents.　Because　of　the　difficulty　in　identifying　high-dimensional　features　by　shallow　classifiers,　we　proposed　an　internal　over-voltage　identification　method　based　on　atomic　decomposition　(AD)　and　convolution　neural　network　(CNN).　The　AD　algorithm　was　used　to　decompose　the　three-phase　voltages　of　the　bus,　and　the　optimal　atoms　were　reconstructed　according　to　the　frequency　to　obtain　the　characteristic　atomic　spectrum　in　this　method.　Then　characteristic　atomic　spectrum　was　input　into　the　CNN　to　realize　the　identification　of　seven　typical　internal　over-voltages.　The　method　introduced　in　this　paper　was　verified　through　the　simulation　and　physics　experiment　platform.　The　results　show　that,　compared　with　the　shallow　learning　support　vector　machine　and　extreme　learning　machine,　CNN　has　stronger　self-learning　ability;　and　the　proposed　method　has　a　higher　recognition　rate　and　stronger　adaptability　than　the　traditional　low-dimensional　features　combined　with　shallow　learning　algorithms.　The　identification　method　can　be　well　applied　to　identify　the　internal　over-voltage　in　the　distribution　network.　The　research　can　provide　a　reference　for　the　identification　of　over-voltage　in　distribution　network.　©　2019,　High　Voltage　Engineering　Editorial　Department　of　CEPRI.　All　right　reserved.