Abstract: In order to improve the accuracy of transformer voiceprint fault diagnosis under small sample conditions,a transformer voiceprint diagnosis model based on Mel spectrogram and IW-GAN is proposed. Firstly,the Mel spectrogram of transformer sound signal is extracted,and then the spectrogram is input into IW-GAN for sample expansion. Among them,IW-GAN uses a more expressive transformer network,and the discriminator uses SN-CNN that satisfies Lipschitz continuity constraints. This improvement enables IW-GAN to stably generate diverse and high-quality samples;Finally,the expanded data is input into different classifiers for fault classification. Experimental results show that the proposed method effectively expand the transformer fault voiceprint data and significantly improve the overall performance of transformer voiceprint fault diagnosis under small sample conditions. This method significantly improves the recognition accuracy of different classifiers,especially the recognition accuracy of CNN classifier has been improved by 6.9%.