Abstract: When the high impedance fault (HIF) occurs, traditional fault identification methods have technical bottlenecks such as difficulty in feature extraction and poor flexibility in threshold selection, which leads to misjudgment in extreme fault scenarios. Therefore, an HIF identification method in distribution network based on transient time-frequency characteristics difference is proposed. First, the time-frequency distribution of transient signals of HIF and normal disturbance condition is analyzed by using Shannon entropy quantization of wavelet packet. It is found that there are significant differences between them: in frequency domain, the energy of disturbance condition signal is concentrated in low frequency, while the energy distribution of HIF signal is relatively uniform. In time domain, the signal energy of disturbance condition is concentrated in the first half of the time window, and the signal energy of HIF is evenly distributed throughout the time window. On this basis, the time-frequency domain waveform of transient signal is taken as the input sample, and the softmax classifier in the traditional convolutional neural networks (CNN) model is replaced by the support vector machine (SVM) classifier, so as to construct a CNN-SVM composite classification model which is suitable for the small sample scenario of HIF identification of distribution network. The convolution layer is used as the characteristics extractor and SVM is used as the classifier to realize HIF identification. Finally, in order to demonstrate the inherent reason of the strong adaptability of the proposed method, the LIME interpretable analysis algorithm is used to visually show the high attention area in the model training process. The model classification principle proves that the proposed method is not affected by various fault conditions; it overcomes the defect of misjudgment of traditional methods in extreme fault scenarios, and can accurately identify the 10 kΩ HIF at the end of distribution lines.