Restricted self-supervised and active learning network for imbalanced few-shot wafer maps defect recognition

Wafer maps (WMs) defect recognition is a crucial step to guarantee the quality of semiconductors. Deep learning-based methods have substantially improved the performance of WMs defect recognition. However, their great performance commonly relies highly on a large amount of labeled data, which is unsuitable for WMs defects with limited samples. Moreover, the quantity of different types of WMs defects is severely imbalanced in the semiconductor production process. Hence, in this paper, a restricted self-supervised and active learning network (RSSALNet) is proposed for few-shot WMs defect recognition. Firstly, a few-shot learning method that leverages self-supervised feature restrictions is proposed to overcome the influence of limited data availability, which utilizes the features from supervised learning (SL) to improve self-supervised learning (SSL). Secondly, an adaptive feature alignment attention (AFAA) with local feature weighting (LFW) is proposed to facilitate the feature integration of SL and SSL to further enhance defect feature representation. Finally, to tackle the issue of imbalanced data, a novel sample selection method, active optimal sampling (AOS), is developed to automatically select the samples with high prototype value of different classes for model training. The performance of RSSALNet is validated on two typical WMs defect datasets. The accuracy on the two WMs datasets improves by 2.79%, 2.85% and 1.90%, 2.05% under 1-shot and 5-shot setting, compared with the suboptimal method. The testing results show that RSSALNet can effectively address the problem of WMs defect recognition with limited and imbalanced data.