Wafer classification is a critical task for semiconductor manufacturing. Most conventional algorithms require a large-scale perfectly labeled dataset to train accurate classifiers. In practice, it is usually difficult or even impossible to collect perfect labels without errors, and the classification accuracy in the presence of imperfectly labeled data would degrade. To facilitate robust wafer classification with noisy labels, we propose a novel self-boosting co-teaching (SB-CT) approach. Specifically, we iteratively correct the wrong labels by using the predictions of two classifiers that are jointly trained with noisily labeled data. To make the proposed method of practical utility, we develop a novel method to accurately estimate the noise rate, and adopt a probability scaling technique to further improve the classification accuracy. As demonstrated by the experimental results based on two industrial datasets, the proposed SB-CT approach achieves superior accuracy over other conventional methods.

Classification is an important machine learning technique that attracts growing interests in various manufacturing applications. Learning an accurate classifier generally requires a large-scale perfectly-labeled training dataset. However, such "golden" labels are not only expensive but also difficult to collect in practice. To facilitate accurate classification in the presence of noisy labels, we propose a novel hybrid method based on active learning and data cleaning. Specifically, we first train an initial classifier with noisily- labeled data. Based on its prediction outcomes, a set of most informative samples is queried for manual annotation. To effectively correct other incorrect labels, we further self-label the unqueried samples based on the true labels provided by human experts and the estimated labels predicted by the initial classifier. As demonstrated by the experimental results based on two industrial datasets, the proposed approach achieves superior accuracy over other conventional methods.

In recent years, classification techniques have been broadly adopted for a variety of smart manufacturing applications, including system health maintenance, defect detection and diagnosis, etc. However, most classification methods require a large set of training data that are accurately labeled by human experts in a specific domain. Collecting these training data is time-consuming and prohibitively expensive in practical applications. To overcome this challenge, we develop a classification framework using imperfectly labeled data. First, a statistical model is proposed to derive a set of probabilistic labels with consideration of labeling errors. Next, an accurate classifier is trained from these inaccurate labels. As demonstrated by the experimental results of two industrial examples, the proposed framework achieves superior classification accuracy over other conventional approaches.