In semiconductor manufacturing

wafer map defect pattern recognition plays a key role in tracing the source of defects.Machine learning has made significant progress in this field in recent years.However

problems such as imbalanced defect categories and the high cost of expert annotation still exist.Therefore

a semi-supervised learning method incorporating non-uniform rotation(NU-R)is proposed to construct efficient prediction models at a low annotation cost.First

the dataset is expanded by a hybrid data enhancement method based on NU-R

which changes the relative positions of the dies while maintaining the invariance of the defect patterns.Second

three uncertainty metrics are combined to filter out the most valuable samples

and an iterative classifier model based on active learning and pseudo-labeling is designed

which combines expert knowledge and high-confidence prediction results to accelerate learning.Experiments on the WM-811K dataset show that the accuracy of the scheme is improved by 2.1% compared with the traditional data enhancement methods

and the effectiveness of the proposed data metrics and iterative strategy is verified.