Efficiency Improvement of Bridge Deck Condition Assessment using Deep Learning-based GPR Data Analysis Model
Introduction
Bridge decks are subjected to various detrimental environmental factors, including traffic loads, rain, and chloride infiltration, which contribute to progressive damage and deterioration. As the number of aging bridges is expected to rise, the need for effective maintenance and safety management is increasing. For the assessment of bridge deck conditions, Ground Penetrating Radar (GPR), which is a non-destructive testing method, is utilized due to its efficiency in Survey. Corrosion of reinforcing bars within concrete structures is closely linked to moisture and chloride ingress through cracks and initial defects, which influence the attenuation of signal amplitude in GPR data. By measuring the amplitude of the signals reflected from the upper reinforcing bars within the concrete, the condition of the bridge deck can be assessed based on the degree of signal attenuation.
Method
The interpretation of GPR data traditionally requires experts to manually identify hyperbolic signals generated by the upper reinforcing bars, a process that is both time-consuming and labor-intensive. To address this issue, deterministic algorithmic approaches have been used. However, due to their sensitivity to noise characteristics and the significant variability of results depending on parameter settings, deep learning-based techniques have recently been utilized. In this study, we utilize a convolutional neural network (CNN) model to detect hyperbolic signals. A supervised learning approach was adopted, with ground truth labels manually created by analyzing the hyperbolic patterns of signals. Among various image processing techniques, we employed image segmentation, and through post-processing, the final coordinates of the upper reinforcing bars were extracted.
Examples
We conducted a survey to assess the condition of the bridge deck at Yeongdongdaegyo (Bridge). Figure 1(a) shows a multi-sensor survey vehicle capable of performing both internal condition assessments and surface inspections, such as detecting cracks in the bridge deck. This vehicle is equipped with a high-resolution GPR device mounted on its lower rear section and operates at a speed of 5 km/h. The GPR system used is Hi-BrigHT from IDS GeoRadar, with an antenna frequency of 2 GHz. It is ground-coupled and configured with multi-channel capabilities. Figure 1(b) is a satellite image of Yeongdongdaegyo, with the yellow areas indicating the approximate extent of the survey. Training and validation datasets were prepared, early stopping conditions were applied, and the trained model was applied to the test dataset. Figure 2 presents the results for the test dataset, where the red box indicates manually picked results and the green box represents deep learning-based, automatically picked results. As shown, the trained model effectively and reliably detects upper reinforcement bar signals. Figure 2(c) provides a detailed comparison between manual and automatic picking results. Manual picking of large datasets has occasionally led to human errors that have, in some instances, significantly impacted the accuracy of the assessment results. In contrast, the deep learning-based interpretation demonstrated consistent and accurate results.
Conclusions
In this study, a deep learning-based model was employed to extract the location coordinates of hyperbolic signals representing the upper reinforcing bars from GPR data. The deep learning-based interpretation method significantly reduces the time required for data analysis while enhancing the consistency and accuracy of the results. The proposed approach not only improves analysis efficiency but is also expected to reduce maintenance costs and enhance the reliability of assessment results.
References
Alani, A. M., Aboutalebi, M., & Kilic, G. (2013). Applications of ground penetrating radar (GPR) in bridge deck monitoring and assessment. Journal of applied geophysics, 97, 45-54.
Kaur, P., Dana, K. J., Romero, F. A., & Gucunski, N. (2015). Automated GPR rebar analysis for robotic bridge deck evaluation. IEEE transactions on cybernetics, 46(10), 2265-2276.
