One of the factors that most affects the safety of structures in or on rock masses is the shear behaviour of rock discontinuities. Some analytical models predicting the shear behavior of rock joints have been developed on the basis of experimental data obtained from direct shear tests performed under different boundary stiffness conditions. However, the use of these analytical models sometimes become difficult due to lack of laboratory test results, or even the difficulty of obtaining some of their parameters. The objective of this paper is to present a model that can be used to predict the shear behaviour of soft clean joints using the artificial neural network (ANN) known as perceptron. Results from direct shear tests conducted on idealized saw-tooth synthetic rock joints under different boundary conditions were considered in the development of the proposed ANN model. The following parameters were considered as input for the ANN model training: boundary normal stiffness, asperity height, initial asperity angle, initial normal stress and the horizontal displacement The output of the ANN model is the shear stress at a particular horizontal displacement. The developed ANN model has an A:5–15–5-1 architecture, where each number represents the number of neurons per layer. The coefficients of correlation between the actual test results and the model output values used to evaluate the behaviour of the neural model during training and validation were 0.999 and 0.997, respectively. The shear strength obtained by applying this kind of model can be considered more advantageous for use in practice than that obtained from the analytical models. One of the advantages of this tool is that, once the synaptic weights and bias of the model are known, the prediction of the shear behaviour of the clean joints of soft rocks can be made using a simple spreadsheet with only parameters which represent the initial roughness of the joint and the boundary conditions (CNL or CNS), rather than several model parameters. Besides, it was observed that the results obtained from the proposed ANN model presented a better fit to the experimental data when compared to the results of some of the existing analytical models.

This content is only available via PDF.
You can access this article if you purchase or spend a download.