Ultrasonic tomography with deep learning for detecting embedded components and internal damage of concrete structures

Ultrasonic tomography is a powerful nondestructive technique for evaluating internal defects in concrete structures. This study presents a deep learning–enhanced approach utilizing a nanoscale object detection model to automate the localization and quantification of internal defects and embedded structural components, including reinforcement bars and ducts. Controlled concrete samples containing artificial defects of varying shapes and depths, along with embedded rebars and ducts, were designed. Ultrasonic signals were collected using a MIRA A1040 tomograph and reconstructed into 3D volumes via Synthetic Aperture Focusing Technique (SAFT). These volumes were converted into 2D slices and segmented using Chan-Vese segmentation and morphological post-processing. A partial histogram matching procedure unified color scales across segmented slices, minimizing color-related biases before model training. Segmentation-assisted labeling provided robust ground truth annotations, resulting in 7220 labeled images. The trained AI model accurately detected delaminations, rebars, and ducts (both grouted and ungrouted), achieving a mean Average Precision (mAP@0.5) of 0.73 and an Average Intersection-over-Union (IoU) of 0.80. Testing on real-world bridge data demonstrated the model's generalization to unseen conditions. Key innovations include automated segmentation-based labeling, robust color standardization via histogram matching, and a lightweight deep learning model optimized for real-time deployment on resource-constrained devices. This integrated approach has the potential to reduce manual interpretation and subjective variability, providing an effective, scalable NDT/E solution for rapid assessment and monitoring of concrete infrastructure through advanced ultrasonic imaging combined with standardized, machine learning-based defect detection.