Abstract: Objectives To address the issue of low fault diagnosis accuracy of traditional neural networks with few labeled samples, a method based contrastive learning and convolution transformer network is proposed. Methods First, raw monitoring data are transformed into similar sample pairs by data augmentation. At the same time, the similar sample pairs are mapped to the deep feature space by a feature extractor. Then, the transformer network is used to contrast locally and globally, thereby clustering representations with similar samples. Finally, the downstream classification network is trained with few labeled samples to improve the diagnostic performance of the proposed model. Results The effectiveness of the proposed method is verified by a self-built reducer test rig. The results show that the proposed method can accurately realize reducer faults with few labeled samples with an accuracy of 98.35%. Compared with the existing method, the accuracy is improved by at least 5.29%. Conclusions The research results can provide a reference for fault diagnosis with few labeled samples.