Semisupervised Domain Adaptation for Wafer Map Defect Recognition via Cross-Alignment Network

Wafer map defect recognition is a crucial step in semiconductor manufacturing and measurement. Semisupervised domain adaptation (SSDA) aims to leverage abundant labeled data in the source domain to achieve accurate classification in the target domain with limited labeled data, effectively addressing the challenge of insufficient labels in wafer map datasets. However, a significant obstacle lies in the label mismatch among wafer map datasets, a problem that most current SSDA methods struggle to mitigate. To solve this problem, this article proposes a new SSDA algorithm named cross-alignment network (CAN). CAN decouples the feature subspace and label space using different encoders and classifiers. To facilitate knowledge transfer between the source and target domains, we propose cross-alignment learning (CAL). CAL consists of sample-wise cross-alignment and class-wise cross-alignment, designed to ensure the consistency of encoders’ outputs and facilitate the formation of robust clusters. We also integrate consistency regularization to enforce uniformity in outputs across different views of a sample. Additionally, we introduce a dynamic threshold strategy that tailors thresholds for each class based on its learning effect, thereby enhancing the model’s efficiency and data utilization. Empirically, our method outperforms the current state-of-the-art SSDA methods on public wafer map datasets. Furthermore, detailed experiments are conducted to verify the effectiveness of each component in CAN.