Aplicações de redes neurais artificiais para previsão do comportamento cisalhante em descontinuidades de maciços rochosos

• Ano de defesa: 2019
• Autor(a) principal: Leite, Ana Raquel Sena
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Geotecnia; Redes neurais (Computação); Zonas de cisalhamento (Geologia); Discontinuities; Dilation; Shear strengt
• Link de acesso: http://www.repositorio.ufc.br/handle/riufc/49667

The strength of rock masses is significantly influenced by the presence of discontinuities, having their shear behavior as ruler in several stability issues. Although there are established models for obtaining the values of resistance, these are not valid with increasing complexity of the problem; because presence of padding, or the presence of confining systems. The existing models that measure these and other particularities have limitations, mainly due to the lack of sufficient laboratory tests to validate them and the difficulty in obtaining some of their parameters. In this context, it was proposed to use artificial neural networks (RNA) of the multilayer perceptron with error backpropagation to predict the behavior shear of filled and unfilled rock discontinuities under constant normal loading (CNL) and constant normal stiffness (CNS). For this purpose, two models were developed to shear behavior prediction: a first model for shear stress acquisition and dilatance with the course of horizontal displacement and a second model for resistance peak shear. The input parameters generally used were: normal contour stiffness (kn); fill thickness to roughness amplitude ratio (t / a); normal tension initial (no); rock roughness index (JRC); uniaxial rock compressive strength sound (c); basic friction angle (b), fill friction angle (fill); and the displacement horizontal (h), but only for the first model. For prediction of shear stress and with horizontal displacement, after analysis of several models, the 8-20-10-5-2 architecture Gm in 500 thousand iterations and correlation coefficient in the 99.3% training and 99.0% test. As the most relevant input variable, the model Gm had a no and kn for shear stress and c, t / a and JRC for swelling. The model had good agreement with the graphical comparisons, experimental data, hypothetical joints and literature models. For the peak shear strength, after the evaluation of the models proposed, G1m was best obtained, with architecture 7-30-1 in 797,660 iterations. Your Correlation coefficient in the training phase was 99.5% and in the test phase 98.0%. The model presented good correlation with experimental data, good interpretation of variables rulers in hypothetical joints and good interpolation with literature models. G1m model had as input variables with greater influence the no, c, kn and JRC.