A confident learning-based coordinate attention-guided fusion vision transformer for mixed-type wafer map defect detection

Wafer defect detection is crucial for quality assurance in semiconductor manufacturing. Current methods often overlook two key challenges: the adverse effects of mislabeled data on model reliability, and the significant correlation between defects and their spatial locations. To address these issues, we propose a novel confident learning-based coordinate attention-guided vision Transformer framework. Our approach includes: (1) automatic mislabel data identification and dataset cleaning using confident learning, and (2) a mixed-type wafer defect detection network that fuses convolutional operations, coordinate attention, and self-attention mechanisms. The architecture enables effective local–global feature extraction with positional awareness, and a decoupled classifier further improves detection performance. Evaluated on the clean MixedWM38 dataset (with 192 mislabeled noisy samples removed via confident learning), our framework achieves 99.60% accuracy while maintaining computational efficiency, outperforming advanced wafer defect detection methods. These results demonstrate its strong potential for industrial applications.