Improving Off-System Bridge Monitoring: An AI-Driven Methodology for Enhanced Condition Prediction

Abstract

Off-system bridges, especially in rural and underserved areas, play a vital role in connecting communities to essential services but often suffer from neglect and underinvestment. These structures are frequently excluded from comprehensive inspection programs, resulting in limited data and a high risk of undetected deterioration. This study addresses the critical challenge of class imbalance in bridge condition datasets by introducing a machine learning framework that integrates Boruta feature selection, Ridge regularization, TomekLinks sampling, and a Generative Adversarial Network trained with focal loss (GAN-FL). The model aims to enhance the accuracy and interpretability of bridge deck condition predictions. Despite improvements in performance for majority classes, minority class predictions remained a challenge, with certain conditions (e.g., 4, 5, and 8) showing low F1-scores. However, the use of GAN-FL led to measurable gains in balancing recall and precision across classes. The proposed approach demonstrates the potential to support risk-informed maintenance decisions, prioritize inspections, and improve safety outcomes for off-system bridges. Future work will focus on improving performance for underrepresented classes and increasing model transparency to support adoption by transportation agencies.