<p>Welding serves as a critical joining technology for subway train body components, where welding quality directly determines operational safety. Current welding defect detection methods face performance limitations due to low inter-class separability, high intra-class feature dispersion among defect samples, and persistent class imbalance in industrial scenarios. To address these challenges, a Context-guided Real-time Detection Transformer (CGRT-DETR) network specifically designed for class-imbalanced welding defects is proposed in this study. The network architecture combines a context-aware anchor attention mechanism with intra-scale feature interaction modules, enabling simultaneous capture of localized defect patterns and global spatial-semantic relationships. A dynamically weighted cross-scale fusion strategy is further developed to synthesize discriminative defect representations by integrating multi-scale features under contextual guidance, effectively resolving inter-class similarity and intra-class variation challenges. To mitigate class imbalance impacts, an IoU-modulated focal loss function is introduced, dynamically adjusting classification penalties based on localization confidence and category distribution statistics. Experimental results demonstrate the superiority of CGRT-DETR over state-of-the-art methods on both proprietary and public datasets. On the welding defect benchmark, the proposed method achieves 96.2% mean average precision at 30 frames per second with only 10.49 million parameters, exhibiting an optimal balance between detection accuracy and computational efficiency for practical industrial deployment.</p>