In non-destructive evaluation of concrete bridge decks, ground penetrating radar (GPR) represents a critical assessment tool for transportation agencies nationwide. However, the interpretation of GPR amplitude data remains constrained by subjective threshold selection processes that introduce significant variability in deterioration mapping outcomes. This research proposes an innovative dual-phase artificial intelligence (AI) framework to address this industry-wide challenge by establishing systematic, objective threshold determination methodologies for GPR data analysis.
The developed framework integrates computer vision and machine learning technologies through a sequential approach: (1) implementation of a modified U-Net convolutional neural network architecture optimized for GPR B-scan semantic segmentation, and (2) development of a reinforcement learning system using Deep Q-Networks that iteratively refines amplitude thresholds by referencing multimodal ground truth datasets. This methodology transforms traditional subjective amplitude analysis into a standardized, data-driven procedure with quantifiable confidence metrics.
Preliminary validation employing datasets from field investigations indicates improvements in interpretation consistency across varying GPR acquisition configurations (ground-coupled and air-launched systems). The ongoing comprehensive validation protocol encompasses multiple in-service concrete bridge decks with diverse deterioration profiles and environmental exposure conditions. Performance evaluation metrics focus on deterioration boundary precision, classification stability, and correlation with established ground truth verification methods including core samples, chloride concentration measurements, and hydrodemolition results.
The research directly addresses current deficiencies in nondestructive testing (NDT) practice by establishing objective thresholding parameters adaptable to various field conditions and equipment configurations. Practical applications for NDT professionals include improved repeatability in deterioration assessment, enhanced decision support for maintenance planning, and reduced dependency on analyst experience levels for reliable interpretation.