Incorporate Real-World Repair Events When Training Deep Learning Models to Improve Bridge and Culvert Condition Rating Accuracy.
Lessons Learned from an Intelligent Bridge Management Tool Developed in Colorado Using Physics-Guided Data-Driven Artificial Intelligence (AI) Technologies.
Colorado, United States
i-BM: An Intelligent Bridge Management Tool for Bridge and Culvert Deterioration Forecasting and Anomaly Detection based on Physics-Guided Deep Learning
Summary Information
Bridge and culvert deterioration poses significant safety and maintenance challenges for transportation agencies. Previous data-driven deterioration models either had limited accuracy or were too complex for practical applications. This study proposed an intelligent bridge management tool (i-BM) that used deep learning (DL) models to serve as a standalone, web-based, and user-friendly software application to facilitate bridge engineers in Colorado Department of Transportation (CDOT). Specifically, i-BM combined data-driven DL models with physics-based deterioration models that describe deterioration as a function of material properties, stress conditions, and exposure environments, to forecast future bridge element condition ratings and detect performance anomalies. The tool was evaluated using multi-modal datasets covering all bridges and culverts in Colorado, including bridge performance records, traffic data, and weather data.
This study identified several lessons learned, including:
-
Incorporate real-world repair events to improve model accuracy and generalization. Repair and maintenance affect deterioration patterns, and ignoring them introduces noise and bias. Using clean, intervention-considered datasets that explicitly account for repair events can improve the accuracy and generalization capability of bridge and culvert condition rating models.
-
Combine physics-based and data-driven models for better performance. Purely data-driven models may capture correlations but lack physical consistency, while physics-based models may not fully reflect real-world variability. Integrating both approaches could capture underlying deterioration mechanisms while employing large-scale historical data, improving both prediction accuracy and interpretability.
-
Leverage long-term historical inspection data and integrate multi-source data. Using long-term historical inspection data, combined with multi-source datasets, can provide temporal depth and contextual completeness, and yield more robust and realistic predictions.
-
Incorporate expert knowledge to enhance infrastructure anomaly detection model performance. Infrastructure data are often noisy, incomplete, or inconsistent, making fully automated modeling challenging. This project demonstrated that integrating active learning and user feedback (e.g., allow engineers to validate anomalies and iteratively refine models) could enhance performance.
