Development and application of probabilistic pavement performance model based on Bayesian neural networks and Bayesian optimisation

Predicting pavement performance is crucial for reasonable roadway maintenance. Accurate forecasts can help agencies optimise resource allocation, extend pavement service life, and reduce maintenance costs. However, incomplete pavement material records, missing historical information, and the integration of multi-source databases result in uncertainty in predictive model development. To address this issue, this study proposed a probabilistic pavement performance prediction method for Guangxi Province using Bayesian Neural Networks (BNNs) combined with SHapley Additive exPlanations (SHAP). A variational Bayesian neural network with a dual hidden layer structure was established, and repeated stochastic forward passes were used to sample from the learned weight posterior, approximate the predictive distribution, and quantify prediction uncertainty. The performance of traditional machine learning models, including gradient boosting and random forest, was compared with that of the BNN model. The results showed that the BNN model performed better in prediction and demonstrated stronger robustness. SHAP analysis was employed to identify the key factors influencing pavement performance. Finally, a graphical user interface was developed as a decision-making tool for pavement maintenance, establishing a technical chain from research to practical application.