Abstract: Targeting the low classification accuracy of convolutional neural networks（CNN）during high impedance fault in distribution network,the paper proposes a fault classification method for the distribution network by combining CNN and SVM. Firstly,the fault data is converted into a time-frequency spectrum grayscale image and is entered into the CNN as a training set. Then support vector machine（SVM）instead of softmax classifier in the CNN is used to build the CNN-SVM model,and the hyper-parameters of the SVM are optimized by grid search algorithm. Finally,a variety of numerical examples are used to verify the superiority of the proposed method. The results of the example analysis show that the CNN-SVM model has higher classification accuracy than the traditional CNNSoftmax model in case of the high impedance fault,and has better adaptability under the conditions of changing the neutral grounding mode of main transformer,network structure change,noise interference and single-phase arc grounding.