In view of the problem that the traditional insulation state detection methods of transformers and medium-voltage switchgear rely on manual labor, this paper based on the audible sound recognition method, by mixing the discharge fault sound of power equipment with the sound of normal conditions and environmental noise to make a sample set, in order to simulate the real operating environment of power equipment. After the fault sound is preprocessed, the spectrogram is used to extract the short-time frequency and energy distribution features of the sound, and the spectrogram data set is constructed. Combined with the improved convolutional neural network, the discharge fault detection is realized. By adding the attention mechanism, adjusting the exponential decay learning rate, the number of data set samples, the audio sampling rate and other ways to further improve the accuracy of the network model, the final design of the network model identification accuracy up to 99.2%, compared with other detection methods have obvious advantages, can realize the online detection of discharge faults.