Uncertainty-Aware Survival Modeling of Pavement Deterioration with Connected Vehicle Data

Sammanfattning

Survival models are commonly used to predict when pavements will need maintenance, yet the uncertainty surrounding these predictions is rarely quantified. This gap poses challenges for road agencies that must make multi-million–dollar decisions based on point estimates without knowing how reliable these predictions are.This study propagates parameter uncertainty into time-to-maintenance predictions using high-frequency connected vehicle (CV) data. Data were collected weekly from 3,707 Swedish road sections (3,357 deterioration events) and analyzed using Cox and Weibull models with identical covariate structures. CV derived indicators were treated as time-dependent covariates, and prediction uncertainty was quantified using section-level bootstrap resampling. Including CV indicators improved model fit and provided additional information beyond static road attributes. The Weibull model produced a median time-to-maintenance of 17.7 weeks with a 95% bootstrap confidence interval of 16.9 –18.7 weeks. In contrast, the Cox model could not estimate a median within the study period, illustrating that the predicted maintenance timing strongly depends on model specification. Overall, the results show that uncertainty quantification is essential for risk-aware pavement management. The framework developed here supports a shift from deterministic scheduling toprobabilistic, uncertainty-aware planning, enabling more informed and efficient maintenance decisions.