Predicting uniaxial compressive strength of different rocks using principal component analysis and deep neural network

Uniaxial compressive strength (UCS) is one of the most practical parameters of rock mechanics. It is an important and basic geomechanical factor in the design of tunnels, dams, and underground drilling. The direct method for determining the UCS in the laboratory is expensive and time-consuming. Therefore, several empirical equations have been developed to estimate the UCS from the results of index and physical tests of rock. Nevertheless, numerous empirical models available in the literature often make it difficult for mining engineers to decide which empirical equation provides the most reliable estimate of UCS.  This work aims to estimate the UCS of rocks using a machine learning-based approach. More specifically, a deep neural networks (DNN) model is designed to predict the UCS from the physical and mechanical characteristics of rocks. ۲۲۱ different rock block samples were collected from various areas of Iran. The physical and mechanical properties include Dry density (ρ), P-wave velocity ( ), Point load test ( ), Brazilian tensile strength (BTS), and water absorption ( ). In order to reduce the dimension of the input features, before the DNN model, principal component analysis (PCA) is employed. A combination of the PCA and the proposed DNN model is found to be efficient and useful in predicting UCS. The mean square error of the proposed method with and without the feature reduction stage was ۰.۰۰۶۸ ± ۰.۰۰۱ and ۰.۰۰۶۷ ± ۰.۰۱۳, respectively.

Keywords:

Physical properties , Mechanical properties , Uniaxial compressive strength (UCS) , Deep neural network (DNN)