Comparación entre regresión tipo LASSO y redes neuronales en la predicción del esfuerzo de fisuración y su elongación asociada del UHPFRC sometido a tracción directa

Abstract

The purpose of this research is to model the direct traction behavior of ultra-high-performance fiber-reinforced (UHPFRC) concrete. For this analysis, LASSO-type regression methods and neural networks were used to predict the tension and elongation that cause the first crack in the concrete. The following statistical indexes were used for the validation of the developed models: Mean absolute error (MAE), root of mean quadratic error (RSME), relationship between RSME and standard deviation of measured data (RSR), mean normalized bias error (NMBE), Nash-Sutcliff efficiency coefficient (E), and multiple determination coefficient (R2). Reinforcement fibers are added to the UHPFRC concrete blend design to increase direct tensile strength. These fibers are evenly distributed and provide ductility properties to ultra-high performance concrete as these fiber-free concrete have a fragile behavior. In the training of the preventive models, 934 random data of the UHPFRC direct traction behavior were used with information on the σcc and εcc parameters, representing the tension state corresponding to the strain of the UHPFRC. During the algorithmic development of the models, these parameters will be coded as Y1 and Y2 respectively. In order to construct an accurate model with adequate results, the detection and treatment of atypical values was necessary. At the end of this process, 196 data were deleted from the database, leaving 738 for training and testing of LASSO regression models and neural networks. In addition, the data was partitioned to facilitate training and testing and to check the efficiency of the neural network and LASSO regression. In this way, 75% of the available data were used for model training, with the remaining 25% being used for model validation. <br> As a conclusion of this research, it follows that the most accurate tool for predicting variables Y1 and Y2, with R2 values of 0,9218 and 0,8075 respectively being reached in the validation subsets. LASSO regression achieved R2 values for these same variables of 0.6771 and 0.6579 respectively, clearly lower than those achieved by neural network models.