A hybrid machine learning framework with GAN-based data augmentation for predicting strain properties of fiber-reinforced repair mortar

Sihao Zhou, Zixuan Sun, Wenliang Li, Junyuan Guo, Dandan Sun, Linglin Xu, Kai Wu,
A hybrid machine learning framework with GAN-based data augmentation for predicting strain properties of fiber-reinforced repair mortar,
Journal of Building Engineering,
Volume 114,
2025,
114140,
ISSN 2352-7102,
https://doi.org/10.1016/j.jobe.2025.114140.
(https://www.sciencedirect.com/science/article/pii/S2352710225023770)
Abstract: Predicting the strain of repairing mortar is essential to assess the reliability of a repaired structure. Traditional lab-based mortar tests are complex and time-consuming. The current scarcity of quality datasets limits machine learning (ML) model accuracy and robustness. In this paper, a hybrid machine learning framework was proposed by utilizing stochastic gradient descent (SGD) to predict the strain properties of fiber-reinforced calcium sulphoaluminate cement-based repair mortars with polyvinyl alcohol fiber (PVAF) and steel fiber (SF). Addressing the inherent data scarcity for the determined materials, we implemented a generative adversarial network (GAN)-based data augmentation strategy. The results reveal a consistent improvement in model performance on the augmented dataset, with notable enhancements observed in SVR and XGBoost models, underscoring the effectiveness of data augmentation in enhancing model generalization. The GAN method improves the model prediction performance and adaptability to sparse data regions, vital for enhancing model robustness and reliability in practical applications. The hybrid model, integrating predictions from multiple models, achieves stable and accurate predictions across datasets, demonstrating significant practical value in dynamic data environments. This methodology significantly boosts the accuracy of predictive models and concurrently enriches the interpretability of their outcomes. It offers a robust tool for forecasting the strain characteristics of fiber-reinforced repair composites.
Keywords: Fiber-reinforced repairing materials; Mechanical properties; Machine learning; Generative adversarial network; Stochastic gradient descent