Hybrid prediction model for the friction performance of asphalt pavements by combining multi-scale decomposition and transformer network

Abstract

The monitoring of asphalt pavements is key to ensuring traffic safety in modern road environments. This study introduces a hybrid model combining two-dimensional (2D) Fourier transform and a transformer network to predict the skid resistance of asphalt pavements. Precise surface texture data were collected using the Laser Scanner-40s (LS-40s) three-dimensional (3D) laser texture measurement device, and skid resistance tests were conducted using the British pendulum tester. The collected data were denoised using bilateral filtering and the median absolute deviation method. Random rotation cropping was used for texture segmentation, resulting in 5650 sets of 3D texture data. A 2D Fourier transform was applied for the multi-scale decomposition of the surface textures. Three network architectures—Visual Geometry Group 16 (VGG16), Residual Network 34 (Resnet34), and Transformer—were employed for feature extraction and then compared. The transformer network outperformed the other two network architectures, with an average absolute error of 3.7178 and a coefficient of determination of 0.9461. The proposed method enables a rapid, accurate, non-contact prediction of the skid resistance of asphalt pavements, suitable for large-scale pavement performance assessments.