Objectives  To address the challenge of low fault diagnosis accuracy in traditional neural networks with few labeled samples, a method based on contrastive learning and convolution transformer network is proposed.

Methods  First, raw monitoring data are transformed into similar sample pairs by data augmentation. These similar sample pairs are then mapped to a deep feature space by a feature extractor. A transformer network is utilized to design cross-prediction tasks for both local and global comparisons, facilitating the clustering of data with the same fault type by comparing the intrinsic similarity between the same batches of data. 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 validated using a self-built reducer test rig. The results show that accuracy of the proposed method reaches 98.38% with few labeled samples, showing significant advantages over existing methods.

Conclusions  The research results can provide the key technology for fault diagnosis of industrial equipment with few labeled samples, contributing to the advancement of intelligent manufacturing.