Predicting blast-induced slope stability using a hybrid field-monitoring and machine learning surrogate model

Abstract

This study evaluates the impact of mine production blasts on ubiquitous jointed slopes using field data and three machine learning algorithms: an ensemble of trees (LSBoost), Gaussian Process Regression (GPR), and Support Vector Machines (SVM). Rather than relying on pure data driven approach, this study develops predictor equation for field monitored peak horizontal acceleration (PHA) to generate blast induced factor of safety (FoS) in FLAC/SLOPE to train machine learning models. PHA was used to formulate a predictor equation for the horizontal seismic coefficient based on the distance of blast (DoB) and maximum charge per delay (MCD). Using this coefficient, a dynamic slope stability analysis was conducted by varying eight input parameters namely, density, cohesion, friction angle, joint angle, joint cohesion, joint friction, DoB, and MCD generating 243 FoS datasets. Among the algorithms, GPR proved to be the best-fitting model, achieving an R2 of 0.99 and the lowest error metrics (RMSE, MSE, MAE) during training and testing. Compared to the other two models, the SVM model marginally indicated the effect of DoB and MCD on the FoS. Consequently, SVM was utilized to investigate the specific effects of these blast parameters on the factor of safety.