TransWeld: automated radiographic inspection and defect analysis framework for pressure-vessel welds

This is the authors' abstract. We don't add key points for this paper.

Radiographic inspection automation and defect analysis for industrial non-destructive testing (NDT) are introduced here for the first time. It uses transformer-based global encoding with convolutional local feature extraction to better identify small and low-contrast defects under varying illumination and noise conditions in images. Realistic defect samples are synthesised to attain class balance over six major classes of defects: porosity, crack, lack of fusion, lack of penetration, slag inclusion, and undercut to augment the GDXray+ weld dataset using a diffusion-based data generation pipeline. 2,291 annotated samples constitute the data and it maintains a split of 70:15:15 for training, validation, and testing. Results of experimental evaluations have shown that TransWeld has been able to achieve high detection performance with an overall precision of 0.9623, recall 0.8809, F1-score 0.9198, and mAP@0.5 of 0.9207; better than its ablation variants in addition to strong evidence for robust generalisation on various defect types. Such a framework offers a deployable foundation for automated weld inspection systems aligned with intelligent manufacturing and safety compliance goals inspired by Industry 4.0. Inference benchmarking shows that the model achieves approximately 39.9 FPS with a mean latency of 25.07 ms per image at 960 × 960 resolution, supporting near-real-time inspection scenarios.