Augmentation of 3D virtual aggregate database using deep convolutional Wasserstein generative adversarial networks
Dong Feng, Jinchao Guan, Chaoliang Fu, Yuanyuan Hu, Frédéric Otto, Alvaro García Hernandez, Pengfei Liu
Abstract
• An original 3D virtual aggregate database was established using stereo vision. • Six DC-WGANs were used to augment the original 3D virtual aggregate database. • A multi-criteria evaluation approach was proposed to assess the augmented database. • The feasibility of applying lightweight processing to the database was studied. The generalization capability of asphalt mixture numerical models incorporating virtual aggregates is often constrained by the limited morphological diversity of the virtual aggregate database, which reduces the ability of the model to reliably predict the mechanical response of asphalt pavements. To address this, a novel framework based on deep convolutional Wasserstein generative adversarial networks (DC-WGANs) is proposed in this study to augment the 3D virtual aggregate database. The original 3D virtual aggregate database was created using stereo vision and multi-view matching techniques. Six DC-WGAN variants were subsequently employed for data augmentation, and a multi-criteria evaluation approach was introduced to comprehensively assess both model performance and the augmented databases. In addition, aggregate packing and Superpave Gyratory Compaction (SGC) simulations were performed to investigate the effect of data augmentation on the prediction of mechanical response. Finally, the feasibility of applying lightweight processing to the database augmented by the best-performing DC-WGAN variant was investigated. Results show that the DC-WGAN-div model achieved the best overall performance, as evidenced by its lowest Wasserstein Distance and Fréchet Inception Distance values and morphological fidelity in terms of true sphericity, angularity index, and fractal dimension. Additionally, the augmented aggregate database can reproduce the mechanical response of asphalt mixtures with high fidelity, as evidenced by the highly consistent aggregate packing density and gyration-height curve. Moreover, the lightweight database, although incurring minor quality loss, substantially improved training efficiency. These findings demonstrate the feasibility of incorporating artificial intelligence-driven data augmentation techniques into virtual aggregate modeling.