This study proposes an unsupervised learning-based outlier detection algorithm that helps detect defects early even in unlabeled situations to increase yield in semiconductor manufacturing processes. This research is very important because serious class imbalance problems occur in real industrial set- tings, and it is impossible for engineers to manually check all data. To overcome the limitations of existing models, we combined deep embedded clustering (DEC) and cluster adaptive scoring techniques. This is a method of detecting outliers more precisely by evaluating how much the data deviates from each cluster based on the Z-score. The experimental results using the UCI-SECOM data set showed that the defective rate of the clusters judged as abnormal by the model proposed in this paper was 17.86%. This is approximately 2.69 times higher than the overall average defect rate of 6.64%. Additionally, by using a decision tree, we explicitly extracted in the form of rules that Sensor 217, Sensor 99, etc. were the key sensors causing the anomaly. In conclusion, this study provides guidelines for effectively classifying risk groups and analyzing their causes even in unlabeled environments.
Shin et al. (Sat,) studied this question.