Use of non-linear prediction tools to assess rock mass permeability using various discontinuity parameters

Abstract

Because of complex discontinuity patterns, it is almost impossible to determine the permeability of rock masses if no proper testing methodology is used. As available in the literature, many empirical approaches to estimate the permeability of a rock mass have been proposed. There is no publication, however, that uses regression analyses and ANFIS (Adaptive Neuro-Fuzzy Inference System) modeling to determine the rock mass permeability. The purpose of the study is to develop various ANFIS and multiple regression models to estimate the rock mass permeability. To this end, a dataset including 453 cases with Lugeon test results and corresponding RQD (Rock Quality Designation), spacing of discontinuities and SCR (Surface Condition Rating) properties is employed. The data were obtained from granite, diorite, volcanic breccia, andesite and agglomerate rock masses from various dam sites and a coal mine in Turkey. Whole data were randomly divided into two parts for training and testing. Two different models were developed to estimate the rock mass permeability. The inputs of the first model are RQD and SCR (Model 1), and the inputs of the second model are discontinuity spacing and SCR (Model 2). Simple regression analyses indicate that there is no statistically meaningful relationship between the Lugeon values with discontinuity spacing and SCR. There is a statistically meaningful relationship, however, between the Lugeon values and RQD. Non-linear multiple regression analyses were implemented for two independent variables and a dependent variable because of the non-linear relationships between the inputs and the output. ANFIS was employed as a second non-linear tool to construct prediction models. According to the performance assessments of the developed models, both of the models and all of the sets are successful. ANFIS is a more successful tool than NLMR. These results show that the models developed are reliable enough and, if there is no direct test result, these models can be used in engineering projects.