Innovations Series

AI-Driven Multi-Domain Feature Analysis of GPR Signals for Predictive Bridge Deck Assessment

This research proposes the development of a machine learning framework for predicting bridge deck deterioration through comprehensive analysis of ground penetrating radar (GPR) data. Traditional GPR-based evaluation methods rely primarily on single-feature metrics—most notably the amplitude of reflections from reinforcing steel—which limits predictive accuracy and fails to leverage the full informational depth of the signal. To address this, the proposed framework will systematically extract and analyze features across time and frequency domains.

Efficiency will be prioritized through early-stage feature screening using LASSO and elastic net regression, allowing the research to identify and focus on 15-20 high-value characteristics while managing computational overhead. A central benchmark throughout the project will be the study by Varnavina et al., which reported correlation coefficients of r=0.54 and r=0.46 using amplitude-only analysis. This benchmark provides a well-defined baseline against which all improvements will be measured.

The modeling strategy will follow a pragmatic progression beginning with multiple linear regression and advancing through Random Forest to gradient boosting techniques. Random Forest will serve as the primary approach due to its interpretability, feature importance rankings, and natural variance estimates for uncertainty quantification. The methodology targets 15-20% improvement over baseline methods, aiming to reduce prediction error from 1.2 to 1.0 inches RMSE in concrete removal depth.

The presentation will employ case-based learning using actual bridge deck data with hydrodemolition ground truth. Attendees will work through progressive modeling approaches and create uncertainty maps for risk-based decisions. Practical deliverables will include feature extraction algorithms, Python code templates, and decision frameworks immediately applicable to field assessments. The framework will transform GPR from deterministic screening to probabilistic decision support, providing calibrated 95% confidence intervals for maintenance planning. Validation across three bridge decks will demonstrate robust generalization, ensuring practitioners can confidently apply these techniques while understanding when predictions are most reliable.

1 hour CE credit

Registration

This is a virtual event on Zoom. All registrants MUST have a free Zoom account to access the webinar.

Members - FREE
Nonmembers - $25

About the Speaker

Babak Enami Alamdari, PhD Candidate
Beyond Visual Inspection: AI-Enhanced GPR for Next-Generation Bridge Deck Assessment
University of Illinois Chicago
Babak Enami Alamdari is a Ph.D. candidate in Civil Engineering at the University of Illinois Chicago, specializing in structural engineering and non-destructive evaluation. His research focuses on developing AI-powered GPR analysis methods for bridge deck assessment, working with an interdisciplinary team on drone-mounted GPR systems. With extensive experience in experimental testing and computational modeling of concrete structures, his work combines structural engineering expertise with advanced machine learning techniques to enhance infrastructure evaluation methods.