DefectGAN: Synthetic Data Generation for EMU Defects Detection With Limited Data

Deep learning (DL)-based defect detection provides various robust automated solutions for maintaining the daily operations of electric multiple units (EMUs). However, a sophisticated detection model typically requires extensive, high-quality data. Therefore, effective detection of certain low-frequency defects can be challenging because it is impractical to collect a sufficient amount of data. To address the lack of targeted training data, a novel synthetic dataset generation method for defect detection of EMUs is proposed based on generative adversarial networks (GANs) and a 3-D CRH380A model. By using synthetic data, defect detection accuracy is significantly improved, even when the scope of real data is limited. The synthetic dataset is employed in this article to train three DL detection models (YOLO v5, Mask R-CNN, and TOOD), and comprehensive experiments are conducted to verify the proposed DefectGAN. The experimental results illustrate that the mixed dataset (synthetic and real data) can significantly improve defect detection accuracy for EMUs with limited data. With a few dozen real images, only mixing the synthetic images generated by the proposed DefectGAN into training datasets can increase the
${F}1$
score from 0 to 0.8. This work will significantly improve the accuracy of defect detection of existing trouble of moving EMU detection system (TEDS) during data scarcity, facilitate automated operation and maintenance of EMUs, and reduce overall operating costs.