A Data Augmentation Method for Establishing a Relationship Model Between Composition and Viscoelastic Properties of Asphalt Binder
Yajie Wang, Liyan Shan, Jiaming Yan, Yunze Pang
Abstract
Testing asphalt binder’s viscoelastic properties is complex and resource-consuming, contrasting with readily available composition data. Establishing a connection between these two aspects is vital, but limited high-quality data poses challenges. Data augmentation through virtual data generation can address this issue. Traditional Generative Adversarial Networks (GAN) struggle with the intricacy of asphalt binder data, so we propose a hybrid VAE-GAN model, combining a Variational Autoencoders (VAE) with GAN, specifically designed for this data type. We evaluated eight machine learning algorithms to determine the best one for linking asphalt binder composition and viscoelastic properties. Next, we constructed a progressive data augmentation framework to efficiently integrate the generated data. Our findings confirm that the VAE-GAN effectively creates realistic virtual data, with support vector regression (SVR) emerging as the top choice for building the relationship model. Integrating the augmented data improved prediction accuracy by around 30% compared to using only real data, verifying the efficiency of our virtual data generation approach. To enhance the practicality of the relationship model, we created a user-friendly graphical user interface (GUI). This study streamlines asphalt performance evaluation and adaptable design, addressing data limitations in materials research, thus supporting sustainable and low-carbon development goals. By simplifying complex processes and improving data utilization, our work paves the way for smarter material designs in various industries.