Reconstruction-Driven Induction Thermography for AI-Assisted Surface Defect Detection in Welded Structures

Welding is widely used to join steel components in energy storage systems, wind power generation systems, and transportation infrastructure. During service, surface defects in steel welds can promote rapid crack propagation and sudden fracture. To address this issue, this study presents an induction thermography framework for post-weld quality inspection of surface-breaking defects in welded steel structures. First, a yoke excitation structure was optimized and evaluated through numerical simulation and experimental validation, and its performance was compared with those of a straight coil and a three-loop coil. The results show that the yoke configuration maintains sufficient heating intensity in the weld region while improving the spatial uniformity of the thermal response. Furthermore, a speed-based reconstruction method was developed to reduce the effects of specimen motion and coil occlusion. By aligning thermal responses from the same physical locations in sequential thermograms, the method reconstructs a continuous temperature field over an extended inspection area. Consequently, the reconstructed images retain clear crack-related thermal features and provide stable inputs for automated analysis. In addition, DeepLabv3+, DSCA-UNet, and feature pyramid network (FPN) were used as representative segmentation models to evaluate the suitability of the reconstructed thermograms for automated defect extraction. On the current laboratory dataset, the three models achieved precision values of 90.4%, 88.6%, and 92.6%, respectively. These results indicate the potential of the proposed framework, while further validation with larger datasets and natural weld defects is still required.