A deep active learning framework for defect classification of wafer bin maps under noisy labels

In modern semiconductor manufacturing, accurately classifying wafer bin map (WBM) defect patterns is essential for ensuring productivity. While recent studies increasingly employ deep learning-based approaches, their effectiveness often depends on large-scale labeled datasets that are costly to obtain. Active learning (AL) offers a practical solution by querying the most informative samples, thereby reducing labeling costs. However, existing AL strategies cannot be effectively utilized due to the existence of noisy labels from human annotation errors, which often leads to incorrect decision boundaries, confirmation bias, and performance deterioration. To address these limitations, this work proposes a hybrid deep AL framework for WBM defect classification under noisy labels. The proposed framework presents three novel techniques. First, a coverage-based diversity sampling identifies candidate samples that provide broad, non-redundant coverage of the unlabeled pool. Second, a Bayesian-based uncertainty sampling ranks the candidates based on information gain. Third, a loss-based noise filtering mechanism using a Gaussian mixture model distinguishes clean samples from noisy ones. Instead of discarding noisy samples, their neighborhoods are marked as unexplored, allowing subsequent diversity sampling to revisit and mitigate confirmation bias. The effectiveness of the proposed framework is validated using a real-world WBM dataset under AL with a noisy oracle setup. The comprehensive experimental results demonstrate that the proposed framework substantially outperforms conventional AL baselines and state-of-the-art AL methods under different label noise rates. Extensive ablation studies also verify the effects of the proposed framework’s techniques on robustness and label efficiency. • Deep active learning is utilized for wafer bin map defect classification. • Coverage-based diversity sampling is performed to select representative candidates. • Efficient Bayesian epistemic uncertainty is employed with Monte Carlo dropout. • Loss-based Gaussian mixture modeling is applied to filter noisy from clean labels. • The proposed method outperforms existing methods under noisy label conditions.