An Improved Neural Network Model for Enhancing Rutting Depth Prediction

Abstract

Rutting is the main distress form of asphalt pavement, and its prediction accuracy is directly related to the reliability of the designed road. This research developed a neural network model to improve the prediction ability about the rutting of a pavement performance criterion and compared it with the multiple linear regression model and the existing neural network model. The neural network model is developed using the Keras module from the TensorFlow package in Python. Two reports generated by the National Cooperative Highway Research Program project 01-37A and the Long-Term Pavement Performance website records have been used as data sources for training the neural network model, which are reliable data preserved after years of monitoring. The input variables include the pavement thickness, service time, average annual daily traffic of trucks and the deformation of the asphalt concrete layer, granular base layer and subgrade layer. This experiment used 440 samples, of which 352 samples (80%) were used for model training and 88 samples (20%) for testing. The training results of the model reveal that the neural network model is significantly better than the multiple linear regression model, and the newly built neural network model performs better than another similar neural network in predictive performance. For the multiple linear regression model, the correlation coefficient R2 value between the measured and predicted in the testing set increased from 0.265 to 0.712. In contrast, it promotes from 0.867 to 0.902 for the neural network model.

Keywords

• mechanistic-empirical pavement design guide (mepdg)
• multiple linear regression (mlr)
• neural network
• nonlinear factors
• overfitting
• rutting prediction