Monitoring rock damage subjected to cracks is an important stage in underground spaces such as radioactive waste disposal repository, civil tunnel, and mining industries. Acoustic emission (AE) technique is one of the methods for monitoring rock damage and has been used by many researchers. To increase the accuracy of the evaluation and prediction of rock damage, it is required to consider various AE parameters, but this work is a difficult problem due to the complexity of the relationship between several AE parameters and rock damage. The purpose of this study is to propose a machine learning (ML)-based prediction model of the quantitative rock damage taking into account of combined features between several AE parameters. To achieve the goal, 10 granite samples from KAERI (Korea Atomic Energy Research Institute) in Daejeon were prepared, and a uniaxial compression test was conducted. To construct a model, random forest (RF) was employed and compared with support vector regression (SVR). The result showed that the generalization performance of RF is higher than that of SVRRBF. The R2, RMSE, and MAPE of the RF for testing data are 0.989, 0.032, and 0.014, respectively, which are acceptable results for application in laboratory scale. As a complementary work, parameter analysis was conducted by means of the Shapley additive explanations (SHAP) for model interpretability. It was confirmed that the cumulative absolute energy and initiation frequency were selected as the main parameter in both high and low-level degrees of the damage. This study suggests the possibility of extension to in-situ application, as subsequent research. Additionally, it provides information that the RF algorithm is a suitable technique and which parameters should be considered for predicting the degree of damage. In future work, we will extend the research to the engineering scale and consider the attenuation characteristics of rocks for practical application.
Size Effect on Acoustic Emission Characteristics of Coal-Rock Damage Evolution
