Wafer Bin Maps (WBM) are essential for diagnosing process issues and monitoring yield in semiconductor manufacturing, yet conventional supervised models suffer from class imbalance and cannot detect defect patterns outside predefined labels. This stud...
Wafer Bin Maps (WBM) are essential for diagnosing process issues and monitoring yield in semiconductor manufacturing, yet conventional supervised models suffer from class imbalance and cannot detect defect patterns outside predefined labels. This study introduces an integrated framework that combines preprocessing, open-set recognition, deep-learning classification, and active learning to overcome these limitations. C-means filtering and Radon transform features were used to construct rotation-invariant representations of WBM data. An OC-SVM discriminator, enhanced with EEOC-SVM, identified unknown defect patterns, while a ResNet50 model with ImageNet pretrained weights classified known patterns under imbalanced conditions. Unknown samples were clustered using DBSCAN, manually labeled, and incorporated into iterative model updates via active learning. Results show that the proposed pipeline improves feature robustness, reliably detects unseen defect types, and significantly enhances classification performance. This work provides a practical and scalable approach for adaptive defect-pattern analysis in real semiconductor yield-management environments.