Multi-sensor AI for Surface Fault Detection

Pavement asset owners and researchers face persistent challenges in predicting the performance of asphalt pavements when moving from controlled laboratory conditions into the complex interactions of real traffic and weather environments. Recent advancements can leverage three complementary approaches to close this gap: predictive modeling with gradient boosting decision trees (GBDT), causal discovery to uncover the effect of traffic and age on raveling, and explainable AI methods such as SHAP (SHapley Additive exPlanations) to make model outputs transparent. These advanced data-driven methods combine predictive modeling, causal reasoning, and explainable artificial intelligence. Finally, the framework can be grounded in micromechanics so models respect physics as well as patterns.GBDT has been shown to forecast pavement deterioration more effectively than traditional techniques such as linear regression. It captures nonlinear interactions among material properties such as binder content, air voids, and aggregate structure. To ensure predictions are not treated as “black box” outputs, GBDT models are paired with SHAP, which highlights the most influential factors and provides engineers with clear insight into key parameters. Causal discovery (CD) helps researchers identify causal links between traffic composition and surface ravelling. By moving beyond simple correlations, CD methods reveal time-series interactions between mixed traffic loads that accelerate the deterioration process. This approach shows that deterioration is not driven by traffic volume alone but by the interplay of pavement age, site-specific conditions, and vehicle mix.Together, these three approaches form a cohesive framework for robust deterioration prediction with GBDT, causal discovery to uncover loading mechanisms, SHAP-backed explanations for transparency, and micromechanics-informed neural networks (MINN) to guide intervention strategies. The result is a pavement management approach that is proactive rather than reactive.