Predicting Uniaxial Compressive Strengths of Brecciated Rock Specimens using neural networks and different learning models

Abstract

Calculation of the uniaxial compressive strength (UCS) of Breccia rock specimens (BRS) is required for the correct determination of material strengths of marble specimens. However, this procedure is expensive and difficult since destructive laboratory tests (DLT) are needed to be done. Therefore, the results of non-destructive laboratory tests (NDLT) combined with different features that are extracted by using image processing techniques can be used instead of DLT to predict UCS of BRS. The goal of this study is to predict the results of DLT by using the results of NDLT, extracted features and artificial neural networks (ANN). Unfortunately, having enough number of specimens for training of ANN is often impossible since the preparation of the standard BRS is extraordinarily difficult. Hence, it is very important to use a learning methodology that prevents deficient evaluation practices. Therefore, different well-known learning methodologies are tested to train the ANN. Then, their effects on error estimation for our small size sample set of BRS are evaluated. The results of simulations show the importance of learning strategies for accurate evaluation of an ANN with a low error rate in prediction of UCS of BRS.