Development of a Robust Stope Stability Prediction (RSSP) model using machine learning techniques

Abstract

Assessing stopes' stability is essential in open stope mine design, as unstable hangingwalls, footwalls, and crowns lead to dilution and safety concerns, thereby compromising the mine's profitability. In the past, several empirical methods, predominantly the stability graph method, were used to assess stopes' stability. However, one of the stability graph method's main challenges is the subjectivity inherent in determining the stability zones. Since stability graphs are empirical, their accuracy is dependent on the database used, and no two geological shields may be the same. This research aimed to address the stability graph method's challenges by exploring the feasibility of using machine learning techniques to predict open stopes' stability status in some selected Ghanaian mines. A comprehensive stability database was established for the research based on available data on block and stope notes, design and planning, surveying, geological and other geotechnical reports. In all, a total of 1 067 data was obtained from Newmont GoldCorp Ahafo Mine (NGAM), Kinross Chirano Gold Mine Limited (CGML) and AngloGold Ashanti Obuasi Mine (AGAG), with each company contributing 292, 303 and 472 case histories, respectively. 283 of the database were caved cases; 336 cases were observed as stable, while the remaining 448 were unstable. In this research, thirteen machine learning algorithms were usedDecision Trees Classifier (DTC), Random Forest Classifier (RFC), K Nearest Neighbours Classifier (KNN), Gaussian Process Classifier (GPC), Extra Trees Classifier (ETC), Bootstrap Aggregation Classifier (BAC), Adaptive Boosting Classifier (ABC), Gradient Boosting Classifier (GBC), Quadratic Discriminant Analysis (QDA), Stochastic Gradient Boosting (SGB), Support Vector Classifier (SVC), Neural Network Classifier (NNC) and Multi-Layer Perceptron Classifier (MLP). Using the Python programme's train-test split technique, 853 representing 80% of the database were used for training the models, while the remaining 214 (20%) were used in evaluating the performance of each of the models. Eleven performance metrics were used to evaluate the reliability and generalisability of all thirteen models. This assessment revealed that seven models (RFC, GBC, KNN, GPC, MLP, NNC and BAC) performed remarkably well and can be used as alternative stope design tools. However, based on a 95% confidence interval, the RFC model was the selected model for the study areas considered. The findings of this study could significantly assist mining engineers in designing more reliable open stopes.