Many studies have been conducted to improve wafer bin map (WBM) defect classification performance because accurate WBM classification can provide information about abnormal processes causing a decrease in yield. However, in the actual manufacturing field, the manual labeling performed by engineers leads to a high level of uncertainty. Label uncertainty has been a major cause of the reduction in WBM classification system performance. In this paper, we propose a class label reconstruction method for subdividing a defect class with various patterns into several groups, creating a new class for defect samples that cannot be categorized into known classes and detecting unknown defects. The proposed method performs discriminative feature learning of the Siamese network and repeated cross-learning of the class label reconstruction based on Gaussian means clustering in a learned feature space. We verified the proposed method using a real-world WBM dataset. In a situation where there the class labels of the training dataset were corrupted, the proposed method could increase the classification accuracy of the test dataset by enabling the corrupted sample to find its original class label. As a result, the accuracy of the proposed method was up to 7.8% higher than that of the convolutional neural network (CNN). Furthermore, through the proposed class label reconstruction, we found a new mixed-type defect class that had not been found until now, and we detected new types of unknown defects that were not used for learning with an average accuracy of over 73%.

由于精确的WBM分类可以提供有关导致产量下降的异常过程的信息，因此已经进行了许多研究来改善晶圆仓图（WBM）缺陷分类性能。但是，在实际的制造领域中，工程师进行的手动标记导致高度不确定性。标签的不确定性一直是WBM分类系统性能下降的主要原因。在本文中，我们提出了一种类别标签重构方法，用于将具有各种模式的缺陷类别细分为几个组，为无法分类为已知类别的缺陷样本创建新类别并检测未知缺陷。所提出的方法执行了暹罗网络的判别特征学习，并在学习的特征空间中基于高斯均值聚类对类标签重构进行了反复交叉学习。我们使用真实的WBM数据集验证了所提出的方法。在训练数据集的类别标签被破坏的情况下，该方法可以通过使损坏的样本找到其原始类别标签来提高测试数据集的分类准确性。结果，该方法的精度比卷积神经网络（CNN）的精度高7.8％。此外，通过提出的类别标签重构，我们发现了一种新的混合类型缺陷类别，该类别迄今为止尚未发现，