In recent years, the construction of hollow slab bridges with cylindrical voids has increased due to their benefits in weight reduction and cost efficiency. However, issues such as insufficient concrete cover thickness and sedimentation at the pavement-slab boundary have been identified. Traditionally, these issues are confirmed by chiseling concrete, which is invasive and provides limited information. With the aging of bridges in Japan, there is a growing need for more efficient and non-destructive inspection methods. While electromagnetic radar (GPR) is being utilized, the visual inspection of the vast amount of data it generates remains labor-intensive and error-prone.
This study aims to streamline structural inspections by visualizing GPR waveforms, employing the YOLO (You Only Look Once) object detection algorithm, and using Python scripts for data filtering and analysis. This approach focuses on detecting areas with insufficient cover thickness and sedimentation. Currently, engineers must visually inspect thousands of images when assessing cover thickness using GPR data, making it nearly impossible to avoid oversight. Our findings suggest that YOLO can effectively automate this process.
The YOLO object detection algorithm demonstrated high accuracy in detecting voids when tested with actual data. In suspected areas, voids with insufficient cover thickness were successfully identified using YOLO-detected coordinates. This indicates the potential for automating the inspection process using deep learning techniques.
Comparative performance analysis between YOLOv7 and the newly developed YOLOv8 showed that YOLOv8 detected more suspected areas of insufficient cover thickness (8 regions) than YOLOv7 (6 regions). Additionally, YOLOv8 identified several regions missed by YOLOv7, highlighting its superior accuracy and reliability.