Intelligent railway component inspection during construction and maintenance via YOLO-GNN architecture

The safety and reliability of railway systems depend on the integrity of track components, a challenge exacerbated by the heterogeneity and aging of legacy U.S. infrastructure. Ensuring component quality during construction and maintenance requires precise, context-aware inspection. While deep learning has enabled automated visual inspection, most models cannot differentiate genuinely missing components from intentionally unused installation points, a task demanding structural and spatial reasoning. This paper presents YOLO-GNN (YG-Net), a hybrid framework that integrates YOLO11-seg for rail component detection and segmentation with a graph neural network module, EGATFormer, for spatial-context-aware inference of missing spikes. EGATFormer captures baseplate-level relationships through edge-guided attention masking. To support benchmarking, the Rail Components Dataset (RCD) is introduced, containing 2100 real-world images and over 41,000 annotated components. YG-Net achieves 55.90% mAP, 50.99% Seg-mAP, and 14.76 ms/image inference, while EGATFormer improves missing-spike classification F1-score from 80.59% to 89.21%, approaching the 95.11% upper bound.