Desenvolvimento de modelos neuro-fuzzy para previsão do comportamento cisalhante das descontinuidades rochosas

Abstract

8ABSTRACTThe rupture mechanism of a rock mass may be strongly related to the constituent discontinuities. This is due to the fact that the shear strength properties of these structures are lower than those of intact rock. Due to the influence that rock discontinuities have, models have been developed with the objective of providing predictions of their shear behavior. However, the analytical models can present disadvantages in their use, such as the non-consideration of important factors thatinfluence the shear behavior of rock discontinuities, or even the difficulty of calculating certain parameters inherent to the formulations. As an alternative to analytical models, other methodologies have been used in Rock Mechanics, highlighting intelligent systems that use artificial neural networks, or neuro-fuzzy controllers. In this context, in the present work,neuro-fuzzy systems were developed to predict the shear behavior of clean and filled rock discontinuities,by means of estimates of the dilation and shear stress as a function of sheardisplacement. In the development of the models, data from 116 direct shear tests presented by several authors were used, generating a set of 2098 graphic points referring to the measurement of dilation and shearstress as a function of sheardisplacement. Several model structures belonging to different classes of data have been established and, through the tests and evaluations carried out, the systems that provided the best results have the boundary normal stiffness,the ratio betweenthe infill thickness and theasperity height,the initial normal stress,joint roughness coefficient,uniaxial compressive strength of the intact rock,basic friction angle of the intact rock,infillfriction angleand the sheardisplacementas input variables. With the dilation prediction model, values of coefficient of determinationof 0.99 were calculated for the training and test phases. In the case of the shear stress prediction model, values of coefficient of determination of 0.97 and 0.96 were obtained for the training and test phases, respectively. The estimates of the defined systems showed a satisfactory correlation with the experimental data used in their development, in addition to being compatible with the results provided by other existing models.