Prediction of axial performance of FRP-confined concrete based on data-driven models: A novel transfer learning approach

Abstract

Fiber-reinforced polymer (FRP) composites are widely used to enhance the performance of concrete structures and strengthen damaged components. The ultimate condition of FRP-confined concrete is critical for designing the type, thickness, and number of layers of FRP. However, the existing models for predicting the ultimate axial strain show low accuracy and poor applicability due to limited data. Therefore, the investigation developed a novel transfer learning approach to create a highly accurate and reliable artificial neural network (ANN) model for predicting the ultimate axial strain. Firstly, through optimizing the initial weights of the ANN and appropriately addressing missing data, the reliable model for predicting compressive strength (ANN-Strength) was established. Then, the knowledge learned from ANN-Strength was transferred via fine-tuning to develop the model for predicting ultimate axial strain (ANN-S-FT). Furthermore, this study employed the bagging ensemble ANN approach and synthesis data approach based on tabular generative adversarial network (TGAN) to develop predictive models. The comparative analysis revealed that ANN-S-FT exhibits the highest predictive accuracy, robustness, and generalization ability, which achieves RMSE, MAE, and R ² values of 0.239, 0.160, and 0.973, respectively. Additionally, it significantly outperforms existing empirical models in terms of predictive performance. Finally, incorporating GUI design, this study created software for predicting the ultimate axial stress and strain of FRP-confined concrete. In summary, the developed models for predicting the ultimate condition of FRP-confined concrete can guide the development and structural design of FRP-confined concrete and serve as an alternative to traditional empirical models. Moreover, the transfer learning approach presented in this study offers a novel solution for the poor accuracy of models due to the lack of experimental data in engineering.