Establishment and extension of digital aggregate database using auxiliary classifier Wasserstein GAN with gradient penalty

•Digital aggregate database was established by using deep learning methods.•The generated aggregates via ACWGAN-gp have a very close distribution of angularity to the real aggregates.•An algorithm based on MATLAB was adopted for particle image segmentation and edge traction. For road construction, t...

Full description

Saved in:

Bibliographic Details
Published in:	Construction & building materials Vol. 300; p. 124217
Main Authors:	Wang, Chonghui, Li, Feifei, Liu, Quan, Wang, Hainian, Benmoussa, Pia, Jeschke, Sabina, Oeser, Markus
Format:	Journal Article
Language:	English
Published:	Elsevier Ltd 20-09-2021
Subjects:	ACWGAN-gp Aggregate morphology Angularity Discrete element Generative adversarial networks Machine learning Pavement Aggregate morphology Discrete element Pavement Machine learning Generative adversarial networks ACWGAN-gp Angularity
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	•Digital aggregate database was established by using deep learning methods.•The generated aggregates via ACWGAN-gp have a very close distribution of angularity to the real aggregates.•An algorithm based on MATLAB was adopted for particle image segmentation and edge traction. For road construction, the morphological characteristics of coarse aggregates such as angularity and sphericity have a considerable influence on asphalt pavement performance. In traditional aggregate simulation processes, images of real coarse grains are captured, and their parameters are extracted manually for reproducing them in a numerical simulation such as Discrete Element Modeling (DEM). Generative Adversarial Networks can generate aggregate images, which can be stored in the Aggregate DEM Database directly. In this paper, it has been demonstrated that applying Auxiliary Classifier Wasserstein GANs with gradient penalty (ACWGAN-gp) is reliable and efficient for the establishment of an aggregate image database. In addition, the distribution of original images was compared with that of images generated based on ACGAN and ACWGAN-gp models. Generated images were validated through obtaining identifiable edge coordinates and represented as DEM input in the simulation process. The results prove that the ACWGAN-gp approach can be used for generating aggregate images for the DEM database. It successfully generates high-quality images of aggregates with a representative distribution of morphologies used for DEM simulation. This work shows convenience and efficiency for machine learning applications in the road construction field.
ISSN:	0950-0618 1879-0526
DOI:	10.1016/j.conbuildmat.2021.124217