Friction-1D Transformer:A 1D Vision Transformer Hybrid Model for Skid Resistance Prediction on Asphalt Pavements
The friction performance of road surfaces is an important indicator of highway safety,which is closely related to road texture.In this paper,a set of three-dimensional denoising methods for road texture data with different noise levels was proposed,which could effectively remove noise while preserving local texture features.In addition,a skid resistance prediction model named Friction-1D Transformer was developed based on the design concept of Vision Transformer(VIT)for skid resistance evaluation of asphalt pavements.Unlike traditional convolutional neural networks,the Friction-1D Transformer utilized positional encoding structures and multi-head attention mechanisms to extract global features from one-dimensional(1D)texture signals along the direction of vehicle movement,thus accurately predicting the friction performance of road surfaces.Comparative analysis with four other models,namely random forest(RF),k-nearest neighbor(k-NN),residual neural network(Resnet),and VIT,shows that Friction-1D Transformer has higher accuracy and faster training speed,with only 3 915 914 parameters.This research validates the feasibility of using raw texture data for training and skid resistance prediction,and the proposed model structure may facilitate the development of the testing technology for skid resistance of road surfaces.
NETL
NSTL
万方数据
