Convolutional Neural Network–Based Friction Model Using Pavement Texture Data

AbstractPavement friction and texture characteristics are important to road surface safety. Despite extensive studies conducted in the last decades, the relationship between pavement texture and surface friction has not been fully understood. This paper implements deep learning (DL) techniques to in...

Full description

Saved in:
Bibliographic Details
Published in:Journal of computing in civil engineering 2018-11, Vol.32 (6)
Main Authors: Yang, Guangwei, Li, Qiang Joshua, Zhan, You, Fei, Yue, Zhang, Aonan
Format: Article
Language:English
Subjects:
Artificial neural networks
Convolution
Friction
Friction resistance
Machine learning
Neural networks
Pavements
Skid resistance
Surface layers
Surface treatment
Technical Papers
Texture
Training
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:AbstractPavement friction and texture characteristics are important to road surface safety. Despite extensive studies conducted in the last decades, the relationship between pavement texture and surface friction has not been fully understood. This paper implements deep learning (DL) techniques to investigate the application of pavement texture data for pavement skid resistance and safety analysis. High speed texture profiles and grip tester friction data are collected in parallel on high friction surface treatment (HFST) sites including various types of lead-in and lead-out pavement sections distributed in 12 states of the United States. FrictionNet, a convolutional neural network (CNN)–based DL architecture, was developed to predict pavement friction levels directly using texture profiles. This architecture is composed of six artificial neuron layers: two convolution layers, three fully connected layers, and one output layer, with 606,409 tuned hyperparameters. There were 50,400 pairs of texture and friction data sets gathered for training, whereas another 12,600 pairs were gathered for validation and testing. The input of FrictionNet is the spectrogram of original texture profile for 1 m segments, and the output is the corresponding friction level ranging from 0.2 to 1.0. FrictionNet achieves 96.85% accuracy for training, 88.92% for validation, and 88.37% for testing in friction prediction. The result demonstrates the potential of using DL methods for highway speed noncontact texture measurements for pavement friction evaluation at the network level.
ISSN:0887-3801
1943-5487
DOI:10.1061/(ASCE)CP.1943-5487.0000797