| Title |
A Human-in-the-Loop Methodology for Dynamic ClassExpansion in Open-set Wafer Bin Map Classification |
| Authors |
박인영(Inyoung Park) ; 김지영(Jeeyoung Kim) |
| DOI |
https://doi.org/10.5573/ieie.2026.63.6.23 |
| Keywords |
Wafer bin map classification; Open-set recognition; One-class SVM; Human-in-the-Loop; Dynamic class expansion |
| Abstract |
In semiconductor manufacturing, previously unseen wafer defect patterns frequently emerge due to process variations, equipment aging, and environmental change. Conventional Wafer Bin Map classifiers, however, are trained under a closed-set assumption and consequently suffer from severe performance degradation when confronted with such unknown patterns in real-world settings. To address this issue, we propose a Human-in-the-Loop (HITL) Open-set Wafer Bin Map classification methodology incorporating dynamic class expansion. The proposed method integrates feature embeddings from a MobileNetV2 backbone with One-Class SVM for distinguishing between known patterns and potential unknowns, while softmax-based uncertainty estimation identifies ambiguous cases even within the known class set. Through HITL feedback, domain experts validate labels, assign new defect classes when necessary, or reject noisy samples. Based on this feedback, the classifier head is dynamically expanded, and incremental learning is performed iteratively across multiple phases. Experimental results on the WM-811K dataset demonstrate that the proposed methodology effectively learns rare defect types initially treated as unknown, progressively enhancing both overall and per-class accuracy across phases. Furthermore, the HITL-guided updates refine the decision boundaries of the classifier, enabling robust adaptation to emerging defect patterns in real-world manufacturing environments. |