A nation's social and economic progress depends on the reliability of its infrastructure. Bridges within a transportation network are crucial for social cohesion and economic growth. Transportation agencies use bridge condition predictions to plan maintenance, rehabilitation, and repair (MR&R) projects. Current techniques provide point estimates but lack uncertainty quantification (UQ), reducing reliability for risk-based decisions. Effective data-driven decisions require reliable UQ of predictions. Although machine learning (ML) models lack calibrated uncertainty estimates, rigorous Bayesian techniques are computationally intensive and require restrictive data assumptions. This study introduces a conformal prediction methodology for assessing uncertainty in ML-based bridge deck condition predictions, offering a practical UQ technique. This method provides prediction regions with valid coverage probabilities, supporting risk-informed bridge management. It impacts transportation infrastructure management by quantifying uncertainties in condition predictions for assets like pavements and railroads. The proposed method enhances confidence in data-driven models for MR&R decisions in transportation. This study offers an uncertainty estimation approach that overcomes current limitations and is efficiently applicable to various ML models without restrictive assumptions, enriching predictive analytics for informed asset management decisions.
Uncertainty Quantification in Bridge Deck Condition Predictions
State of the Industry
