Comparative Analysis of CNN GAN Models for Defect Detection of BLDC Motors Under Low Data Conditions

This study proposes a generative adversarial networks (GAN)-based data augmentation framework to address the class imbalance and lack of defect data encountered in smart factory vision inspection systems and provides a comparative analysis of various GAN models. Using brushless direct current motor stator pin data acquired from industrial sites, normal and defective images were augmented by employing the least-squares deep convolutional GAN (LS-DCGAN) for training stability, contrastive unpaired translation GAN (CUTGAN) for maintaining structural consistency through patch-based contrastive learning, and FastGAN as a lightweight model specialized for small-scale datasets. The quality of the generated images was quantitatively evaluated using the kernel inception distance metric, which ensures high statistical reliability in limited sample environments. Experimental results demonstrate that CUTGAN outperforms other models in preserving the geometric structure of stator pins and enhancing classification accuracy. Furthermore, convolutional neural network models trained with data augmented by CUTGAN exhibited considerable improvements in accuracy and F1-score compared with those trained solely on real-world data. This research study provides a practical guideline for selecting optimal GAN models and data-augmentation strategies to establish high-performance vision inspection systems in early-stage manufacturing processes where defect samples are scarce.