Abstract: The underground production environment in coal mines is harsh, and key equipment such as gas pumps, ventilators, and coal shearers often operate continuously, making them susceptible to faults. Currently, end-to-end audio data fault diagnosis methods heavily depend on data labeling for model training and update. Although large amounts of raw data can be collected, these data are typically unlabeled and cannot be directly used for model training. Factors such as sudden changes in equipment operating conditions or equipment reconfiguration may cause data distribution changes, leading to decreased model performance. To address these issues, an underground coal mining equipment audio signal fault diagnosis method based on improved transfer learning is proposed. First, Mel-Frequency Cepstral Coefficients (MFCC) features were extracted from the audio signals of coal mining equipment to capture key information about the equipment's operational status, generating a 2D fault feature coefficient map. Then, a fault diagnosis network model based on improved transfer learning was established, using the improved Maximum Mean Discrepancy (MMD) and multi-kernel joint MMD as metrics. The joint distribution distance was calculated using pseudo-labels, and label information was mapped through multiple linear transformations to match features and reduce data distribution differences, achieving simultaneous alignment of both marginal and conditional distributions. Experimental results showed that the proposed method achieved high-accuracy fault diagnosis under unlabeled conditions, with an accuracy rate of 96.99% and a standard deviation of 0.014. In model noise resistance experiments, the fault diagnosis model based on improved transfer learning maintained 80% diagnostic accuracy under low signal-to-noise ratio conditions (e.g., 10 dB), demonstrating strong noise robustness.