Abstract

This work investigates the applicability of artificial intelligence techniques to open stope stability assessment in the Stability Graph method of open stope design. Eleven supervised learning models are applied to the refined open stope database consisting of 316 historical cases to classify the open stope stability states. Supervised learning algorithms used in this study include Naïve Bayes (NB), K-Nearest Neighborhood (KNN), Multinomial Logistic Regression (MLR), Random Forest (RF), Support Vector Machine (SVM), Extreme Gradient Boosting (XGB), Light Gradient Boosting (LightGB), Cat Boosting (CatBoost), Bagging, Artificial Neural Network (ANN), and Multi-layer Perception Neural Network (MLPNN). The modified stability number constituents and shape factor are used as input to these models, specifically, measure of block size defined by RQD/joint set number, inter-block shear strength defined by joint roughness number/joint alteration number, stress factor (A), joint defect factor (B), gravity factor (C) and hydraulic radius are taken as input parameters. The outcomes of the algorithms are the classified open stope stability states such as stable, unstable, and failure. To eliminate the bias in model performance cross-validation, particularly, repeated stratified ten-fold is applied to tackle the imbalanced stope stability state distribution in the database. The performance of each model is assessed and compared using accuracy, precision, recall, and f1-score criteria. The accuracy of the NB, KNN, MLR, RF, SVM, XGB, LightGB, CatBoost, Bagging, ANN, and MLPNN on testing set are 67%, 76.6%, 73.4%, 78.7%, 74.5%, 76.6%, 79.8%, 73.4%, 75.5%, 73.4%, and 75.5%, respectively. NB-based model shows a relatively poor performance, while the highest accuracies are reached by RF and LightGB-based models. On the refined stability graph stability zones, around 18% and 2 % of stable points are located on the unstable and failure zones, respectively; 18.8% and 26.1 % of unstable points are on the stable and failure zones, respectively; 22.2% of failure points are on the unstable zone. Considering this, the accuracy of artificial intelligence techniques (except Naïve Bayes) ranged between 73.4% and 79.8% is valid. Thus, artificial intelligence approaches are reliable and feasible to apply for open stope stability prediction, reinforces the idea that the stability state of stope surfaces in the Stability Graph is more appropriate than the deterministic definite approach.  It is well known that the learning ability of the artificial intelligence techniques improves with increased data. However, it is suggested to include additional input parameters such as nearby fault existence, blasting effect and stope stand-up time into artificial intelligence models to predict the stope stability more accurately. This is because the considerable influence of these factors on the stope stability is empirically and theoretically proven but have not been accounted for in the stability graph components and hence have not been included in the input parameters of the artificial intelligence models used in this paper. Considering including the additional input parameters along with increasing data size (quality data) will enhance the performance of the artificial intelligence techniques in predicting open stope stability, as well, these steps will also improve the empirical open stope stability approach in open stope design.