EMU fault detection algorithm based on GAN+YOLO model under limited data

Abstract EMU fault detection under the TEDS system faces persistent challenges due to data scarcity and frequent missed detections of small or cross-scale fault targets. This paper proposes a dual-enhanced detection framework that integrates high-fidelity data generation with a structurally optimized lightweight detector to address these issues. On the data side, we present DefectGAN-SLS, a style-transfer-based generation network incorporating a self-light suppression branch to reduce overexposure artifacts and preserve structural details. Compared with the original DefectGAN, this improvement lowers the FID score from 96.33 to 92.98, yielding more realistic synthetic samples. When used for model training, the inclusion of synthetic data improves detection performance. The mAP50:95 increases from 0.459 (real data only) to 0.514 with DefectGAN, and further to 0.523 with DefectGAN-SLS, particularly benefiting low-sample fault categories. On the model side, we introduce EFD-YOLOv11n, an enhanced detector based on YOLOv11n. It includes a cross-scale detection head to fuse shallow high-resolution and deep semantic features, alongside a lightweight attention module that suppresses redundancy and emphasizes key regions. A coarse-to-fine feature aggregation structure is also incorporated to improve multi-scale representation, enhancing localization accuracy for small and dense targets while maintaining real-time efficiency. Experiments show that EFD-YOLOv11n improves mAP50:95 by 3.8% (from 0.504 to 0.523) with nearly unchanged parameters (from 2.6M to 2.9M) and moderately increased FLOPs (from 6.3G to 11.3G), while maintaining a high inference speed of 416 FPS. This demonstrates an excellent efficiency-accuracy trade-off, making EFD-YOLOv11n a lightweight and practical solution for real-world EMU fault detection.