Uma abordagem baseada em dados para a predição da resistência ao cisalhamento de vigas de concreto armado reforçadas com EB-FRP

Abstract

The use of Fiber-Reinforced Polymer (FRP) has become established as a promising solution for strengthening concrete structures due to its remarkable properties, such as durability, lightweight, corrosion resistance, and speed of installation. However, the application of this material faces challenges as it is a relatively recent method, particularly regarding shear strengthening, design optimization, and the development of accurate models for strength prediction. In this context, recent publications on the use of machine learning have demonstrated the potential to overcome these hurdles, although limitations such as the scarcity of reference data and the need for robust models that capture the complexity of real structural behavior still exist. Given this background, this study aimed to propose strength prediction models for externally bonded FRP-strengthened reinforced concrete beams using linear regression models and artificial neural networks (ANN). As fundamental stages of the research, a comprehensive database was constructed from experimental studies, covering various strengthening scenarios. Through statistical analysis, dimensionality reduction was performed, identifying the most significant physical parameters and developing multiple linear regression mathematical models for shear strength prediction. Furthermore, employing machine learning techniques, an ANN was developed along with the optimization of the base model's hyperparameters. The best-performing linear model presented a mean absolute error up to 66.97% lower than international codes, with a coefficient of determination (R²) of 0.74. In the hyperparameter optimization and regularization stage of the base ANN model, predictive models were obtained with performance significantly superior to existing standards, reaching an R² in the order of 0.90. The application of LASSO regularization resulted in a model with maximum generalization capacity and physical consistency, achieving an R² of 0.92 during blind testing.