Modelo de rede neural artificial para previsão do comportamento cisalhante de descontinuidades rochosas

Abstract

The artificial intelligence (AI) has been widely used in engineering due to its capacity to interpret and process complex information. One of these AI technologies is the artificial neural network (ANN) which is based on the functioning of the human central nervous system and its ability to learn and recognize patterns. This work aims to present an ANN model capable of predicting the shear strength of rock discontinuities. The shear strength of rock discontinuity is one of the most important factors governing the mechanical behavior of rock mass whose definition sometimes requires expensive laboratory procedures not always available. Moreover, the existing analytical models have several limitations regarding not consider all the variables which influence the shear strength of rock joints or needing shear testes. Therefore, nine ANN architectures were tested considering the following inputs: the normal boundary stiffness, the ratio between infill thickness and asperity amplitude, the initial normal stress, the joint roughness coefficient, uniaxial rock compressive strength, infill friction angle, and the horizontal displacement. As outputs the shear strength and dilation of rock discontinuities. The architecture with the best performance is the 7-20-10-2 and with 500,000 iterations with a correlation of 99% for training data and 96% in the validation data. The results show a nice fitting for the ANN model output data with experimental data. As the analytical models made so far for infilled joints are only capable of predicting the peak shear strength, the ANN comes with a handful tool for predicting shear strength with velocity and low cost.