• 목록

Monitoring the wheel condition of Overhead Hoist Transport (OHT) systems is crucial for maintaining production efficiency and safety in semiconductor manufacturing processes. However, collecting defect data in real manufacturing environments is challenging, and the defects often appear in very small regions within the images, limiting the performance of existing machine learning models. This study proposes a method to train robust defect detection models from limited real data by leveraging generative AI technology to augment OHT wheel defect images.

The proposed technique first generates various virtual defect patterns resembling real defects and seamlessly synthesizes them onto OHT wheel images using appropriate image processing techniques, creating realistic defective wheel images. Domain adaptation techniques are then applied to minimize the domain gap between the generated synthetic images and real images. The augmented dataset is utilized to train a convolutional neural network-based defect detection model, and its performance is evaluated in an actual manufacturing environment.

Experimental results demonstrate that the proposed method achieves higher accuracy compared to models trained solely on real data, highlighting the effectiveness of generative AI-based data augmentation in addressing the limited real data problem.