A　recognition　method　of　internal　overvoltage　in　distribution　network　via　improved　CWD-CNN　is　proposed　in　order　to　solve　the　problem　that　the　overvoltage　category　is　difficult　to　identify.　By　applying　Choi-Williams　distribution(CWD)on　seven　common　overvoltage　signals,　two-dimensional　matrix　for　time-frequency　energy　characteristic　of　overvoltage　signals　was　constructed.　Then　the　classification　of　overvoltage　was　carried　out　by　means　of　convolutional　neural　network　(CNN).　The　convolution　kernel　of　the　improved　CNN　has　a　rectangular　scale,　and　it　can　extract　the　features　of　time-frequency　images　efficiently　and　quickly.　The　impact　of　the　number　of　iterations,　the　number　of　samples,　the　number　of　hidden　layers,　the　size　of　the　convolution　kernel,　on　the　optimization　results　were　analyzed　to　determine　the　optimum　parameters.　Finally,　cross-validation　was　performed　by　random　sampling　of　data　from　the　sample　base.　The　results　show　that　the　method　can　effectively　classify　the　seven　kinds　of　overvoltage　signals,　and　it　has　a　higher　recognition.　The　proposed　method　avoids　the　limitation　and　complexity　of　extracting　feature　quantity　manually.　©　2020,　Harbin　University　of　Science　and　Technology　Publication.　All　right　reserved.