The safety of high-speed maglev trains operating at 600 km/h is crucial, as any surface defect in key components, like electromagnets, could significantly impact operational safety. Currently, surface defect detection for electromagnets primarily relies on manual inspection. However, deep learning-based methods still face challenges in industrial environments, including data scarcity, low accuracy, and poor real-time performance. This paper proposes a two-stage defect detection method for maglev electromagnets: The first stage rapidly localizes and crops the electromagnet, while the second stage precisely locates defects. To address the issue of scarce defect samples, we use StarGAN v2 for data augmentation and a hard-negative mining strategy to reduce false positives. The first stage employs a lightweight network to reduce model complexity, achieving a 97% reduction in parameters and a 370% improvement in CPU inference speed. The second stage model incorporates multiple architectures such as high-order relationship modeling, multi-branch dynamic fusion modules, and denormalized Transformer to improve cross-scale modeling. Experimental results show a 6.2% increase in mAP_50-95 and the highest F₁ scores on on the PCB and GC10-DET public datasets, demonstrating strong feasibility for industrial deployment.