Experimentos de previsão da carga de ruptura em estacas carregadas axialmente à compressão utilizando redes neurais artificiais (RNA)

• Autor(a) principal: Jesus, Bruno Vieira de
• Data de Publicação: 2019
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UERJ
• Texto Completo: http://www.bdtd.uerj.br/handle/1/11498
• topic: Civil Engineering; Estacaria (Civil Engineering); Resistance of materials; Artificial Neural Networks (ANNs); Failure load; Piles Static load test; Dynamic load test; Engenharia civil; Estacaria (Engenharia civil); Resistência de materiais; Redes Neurais Artificiais (RNAs); Carga de Ruptura; Estacas; Prova de Carga Estática; Ensaio de Carregamento Dinâmico; CNPQ::ENGENHARIAS::ENGENHARIA CIVIL

This work presents experiments of prediction failure load in piles compression axially loaded using artificial neural networks (ANNs). The database used in the research has 342 elements, and consists of static compression load test, dynamic load tests and standard penetration tests results. The software used was MATLAB®, version R2017a. Four network models are used to try to predict the failure load from data of the pile s length and cross-sectional area and the NSPT s values. In each model, training and testing are perform. Subsequently, each architecture is validated, presenting the best performing networks. The different models search a refinement of the network, and the generalization capacity of the same. In Model 01 the best performing network presents Rtraining = 0.96 and Rtest = 0.81. In Model 02, there was a refinement in the network, obtains Rtraining = 0.99 and Rtest = 0.89. In Model 03, the best network presents Rtraining = 0.95 and Rtest = 0.93 and in Model 04 the network obtains Rtraining = 0.95 and Rtest = 0.96. Although the good correlation coefficients presented, all the networks present high percentage errors in the training and test results, as well as in the validation results, when the data obtained (output) with the expected data (target) were compared. The results suggest that the networks were not able to generalize a solution obtaining a final result with satisfactory answers. Some factors that may interfere with network performance are the quality of the database information, the diversity of soils, which can lead to different outputs for the same input data, and the uncertainties inherent in soil behavior and the quality of the tests analyzed.