The looseness fault of transformer winding leaves a huge hidden danger for the safety and stability of power system, but there is a lack of practical and effective diagnosis methods at present. Therefore, a method of transformer winding looseness voiceprint recognition based on 50 Hz frequency multiplying wavelet time-frequency entropy and RUSBoost is proposed. Firstly, according to the characteristics of transformer voiceprint, the 50 Hz frequency multiplying wavelet time-frequency entropy is proposed to extract the characteristics of voiceprint signal. Then, aiming at the problem of sample imbalance caused by less transformer fault samples, a pattern recognition method based on RUSBoost model is proposed. Finally, the effectiveness of this method is verified on the basis of field measurement data. The research results show that the proposed method can realize reliable diagnosis of transformer winding faults with different degrees of looseness, and the average recognition rate reaches 98.9%. The recognition accuracies of 75% looseness and 100% looseness with fewer samples are as high as 97.2% and 94.6%, respectively. The total recognition accuracy is improved by at least 3.3% compared with traditional models such as RF, DT, KNN and SVM, and the recognition accuracies of 75% looseness and 100% looseness with fewer samples are improved by at least 2.8% and 2.5%, respectively.