Abstract: Mining transformers operate for a long time in complex mine environments and are susceptible to factors such as humidity, dust, and high loads, which can pose potential risks of transformer failure. The article proposes a fault diagnosis model for mining transformers based on voiceprint signals, which combines improved wavelet threshold denoising with CNN for multi-level analysis and extraction of voiceprint features. Firstly, perform wavelet threshold denoising on the voiceprint signals collected by the sensor. Next, extract the main frequency and fault frequency of the transformer under no-load and load conditions, and perform feature extraction and CNN parameter setting. Finally, the experimental results show that this method exhibits high diagnostic accuracy in complex environments, effectively improving the timeliness and accuracy of mining transformer fault warning, and has good practical application value and promotion prospects.