Prediction of rock strain using soft computing framework

摘要

This study presents a comparative analysis of conventional soft computing techniques in predicting strain of a rock sample fitted with several strain gauges in horizontal and vertical directions. For this purpose, a total of 2040 experimental test data was obtained from an experimental setup. Six conventional soft computing techniques, namely relevance vector machine, genetic programming, multivariate adaptive regression spline, minimax probability machine regression, emotional neural network, and extreme learning machine were used. These models were trained and validated with 70% and 30% observations of the main dataset, respectively. Experimental results demonstrate that most of the employed models have attained the most accurate prediction of rock strain. Overall, the result of the RVM model is significantly better than those obtained from other soft computing methods employed in this study. In the testing phase, the RVM model attained 94.0% and 99.8% accuracies (in terms of R-2 value) against horizontal and vertical directions, respectively. Based on the experimental results, the RVM model has the potential to be a new alternative to assist geological/geotechnical engineers to estimate the rock strain in the design phase of civil engineering projects.