Citas bibligráficas
Tovar, J., (2023). Modelos explicables basados en machine learning para el diseño por corte y predicción del modo de falla de vigas de concreto presforzado [Tesis, Universidad de Ingeniería y Tecnología]. https://hdl.handle.net/20.500.12815/353
Tovar, J., Modelos explicables basados en machine learning para el diseño por corte y predicción del modo de falla de vigas de concreto presforzado [Tesis]. PE: Universidad de Ingeniería y Tecnología; 2023. https://hdl.handle.net/20.500.12815/353
@misc{renati/230555,
title = "Modelos explicables basados en machine learning para el diseño por corte y predicción del modo de falla de vigas de concreto presforzado",
author = "Tovar Lagos, Jhon Paulino",
publisher = "Universidad de Ingeniería y Tecnología",
year = "2023"
}
Prestressed Concrete (CP) beams have complex shear behavior because they involve many influential parameters. For this reason, the design codes present different equations, analytical and empirical, that allow the shear design of these beams. However, different studies have shown that these equations cannot be generalized to other design cases, they are very conservative and do not predict the failure mode. The objective of this thesis was the development of a framework to obtain Machine Learning models for the prediction of shear design and the prediction of the failure mode of CP beams. To this end, a data set of 670 experimental shear strength tests of prestressed beams with and without transverse reinforcement was compiled and processed. Then, different predictive models were implemented and compared, including individual learning models and ensemble learning models. In addition, feature selection methods were applied, a grid search was implemented to find the optimal hyperparameters and it was validated with a comprehensive test set. With these methods was possible to reduce the number of input features to half of the total without loss of performance. The final models achieved high accuracy for shear strength prediction (R 2=0.980) and failure mode classification (F1_score=0.851). After that, the Shapley Additive exPlanations technique was implemented to explain and evaluate the importance of the different features in the model predictions. The final models shown to be more accurate and generalizable than current design equations and finite element analysis. Finally, the models were implemented in a user interface to be used in the early phase of CP structure design.
Este ítem está sujeto a una licencia Creative Commons Licencia Creative Commons