Establishment of probabilistic prediction models for pavement deterioration based on Bayesian neural network

ABSTRACT

The process of pavement deterioration involves uncertainties, and neural networks have been widely used in pavement performance prediction due to their high accuracy. However, the overwhelming majority of current performance prediction models based on neural networks are deterministic. Therefore, this study combined Bayesian theory and neural networks to establish a Bayesian neural network (BNN)-based probabilistic model for predicting pavement deterioration. The proposed model was built on the pavement data in Shanxi Province, China. This study first refined data using the K-Nearest Neighbour and empirical methods, and then selected input features based on correlation coefficient methods. Using the refined data, the deterministic neural network model was established to obtain the prior probability distribution of weights, and then the BNN-based probabilistic model was developed. Compared with the sole neural network model, the BNN-based model not only retains comparable prediction accuracy to the neural network model, but also incorporates uncertainties. The BNN-based model is also theoretically superior to the Markov-based probabilistic model because the former can incorporate all factors and does not need to classify performance values into states. The BNN-based model shall provide more reliable prediction results of pavement deterioration and help engineers make more reasonable maintenance decisions.

KEYWORDS:

• Pavement deterioration
• prediction
• probabilistic model
• Bayesian neural network
• data cleaning

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Additional information

Funding

This work was supported by the Science and technology projects of Shanxi Provincial Department of Transportation under Grant [numbers 8521002647 and 2018-1-32].