Modelling of pavement performance evolution considering uncertainty and interpretability: a machine learning based framework

Machine learning (ML) based pavement performance models have gained increasing popularity in recent years due to their strong power in modelling complex relationships. However, the insufficiency of a feature selection process prior to model construction, the difficulty in explaining the black box models, and the lack of uncertainty consideration all impeded the application of the produced models in real world. To fill these gaps, this study aims to develop a new framework to model the pavement performance evolution based on the state-of-the-art ML techniques, including the BorutaShap method for feature selection, the Bayesian neural network (BNN) for model development and uncertainty quantification, and the SHapley Additive exPlanations (SHAP) approach for model interpretation. A case study of predicting the pavement transverse cracking was conducted. The two generated BNN models yielded relatively accurate predictions with the R-square of 0.86 and 0.79 for unmaintained and maintained segments, respectively. Poor data quality was found to be the dominant source of uncertainty. The model interpretation also provided some insight into the underlying influential mechanism of various factors. The framework was expected to enable the decision-makers to build more reliable and informative pavement performance models that could be integrated into the pavement management tools.