Research Seminars @ Illinois

AUTOMATED INSPECTION AND ASSESSMENT OF DAMAGED STEEL BRIDGE GIRDERS USING COMPUTER VISION AND MACHINE LEARNING

Advisor: Professor Ahmed Elbanna

Abstract

Steel bridge girders are often subjected to over-height vehicle strikes that induce significant permanent deformations affecting their structural performance and load-carrying capacity. After such strikes, affected bridges need to be promptly inspected and evaluated to determine whether their structural integrity has been compromised and whether traffic restrictions, emergency repairs, or bridge postings are necessary. Current inspection practices rely heavily on manual field measurements that are labor-intensive and often result in traffic disruptions and increased operational costs. In addition, finite element models can be used to provide detailed analysis of the damage, but they are typically developed for specific case studies and are time-consuming and computationally demanding to build and analyze. These limitations create significant challenges that prevent bridge inspectors and engineers from implementing rapid assessment workflows and making informed decisions after strike events. Accordingly, there is a pressing need for efficient, automated, and reliable approaches that can rapidly quantify damage and assess the structural condition of damaged steel girders after over-height vehicle strikes. 

This research presents a unified framework for automated inspection and assessment of damaged steel girders that integrates computer vision, machine learning, and explainable artificial intelligence to enable a rapid and quantitative evaluation of damaged girders. The developed framework consists of three novel methodologies: (1) an innovative vision-based inspection methodology that automatically reconstructs deformation measurements of damaged girders from inspection images without requiring calibration targets or prior camera information; (2) a novel data-driven prediction methodology that estimates the residual load-carrying capacity of damaged girders based on geometric properties and damage measurements, providing fast and accurate prediction models without the need for time-consuming and computationally demanding numerical simulations; and (3) an explainable artificial intelligence methodology that identifies the most influential parameters governing structural response, enabling engineers to understand how design parameters and damage magnitudes affect the predicted capacity while promoting transparent evaluation and informed decision-making.

The results of the framework demonstrate the original contributions and capabilities of the developed methodologies for advancing bridge inspection and assessment procedures. The vision-based approach provides a practical and deployable inspection pipeline capable of quantifying deformation measurements directly from bridge imagery without the need for manual measurements or expensive scanning equipment. The trained machine learning models provide accurate and computationally efficient surrogates for detailed finite element analyses, allowing rapid estimation of residual load-carrying capacity for use in post-impact assessment. The explainable AI methodology increases confidence in model predictions by identifying the most influential input parameters and improving model interpretability to support maintenance and rehabilitation decisions. These contributions and capabilities of the developed framework are expected to provide bridge inspectors and engineers with practical tools that can improve the efficiency, accuracy, and reliability of condition evaluation of damaged steel girders.