Abstract: The power transformer makes abnormal sounds when mechanical failures occur. The mechanical fault identification based on voiceprint signals has become a current research hotspot due to its high accuracy, timely detection, and non-invasiveness. However, voiceprint signals are easily affected by noise and difficult to obtain, and the processing speed is slow. Therefore, how to achieve rapid and accurate identification of mechanical fault based on voiceprint signals in the presence of strong noise and small sample sizes has become a current research difficulty. To address the aforementioned issues, this paper first incorporates physical principles and empirical knowledge to extract feature and builds an improved Transformer network, significantly enhancing the noise resistance. Then, a convolutional autoencoder for model compression is constructed to shorten the training time. Finally, this paper employs cross-modal Transfer Learning by pretraining the model on the ImageNet-1k dataset to address the issue of limited training samples. Compared to traditional time-series deep learning methods, the proposed method achieves higher accuracy in a high-noise environment (S_NR=−16 dB). Experimental results demonstrate significant improvements in accuracy, robustness, and generalization. This work provides a reliable solution for implementing power transformer mechanical fault identification based on voiceprint signals in complex environments.