A Feasibility Study on The Implementation of Neural Network Classifiers for Open Stope Design

Assessing the stability of stopes is essential in open stope mine design as unstable hangingwalls and footwalls lead to sloughing, unplanned stope dilution, and safety concerns compromising the profitability of the mine. Over the past few decades, numerous empirical tools have been developed to dimension open stope in connection with its stability, using the stability graph method. However, one of the principal limitations of the stability graph method is to objectively determine the boundary of the stability zones, and gain a clear probabilistic interpretation of the graph. To overcome this issue, this paper aims to explore the feasibility of artificial neural network (ANN) based classifiers for the design of open stopes. A stope stability database was compiled and included the stope dimensions, rock mass properties, and the stope stability conditions. The main parameters included the modified stability number (N’), and the stope stability conditions (stable, unstable, and failed), and hydraulic radius (HR). A feed-forward neural network (FFNN) classifier containing two hidden layers (110 neurons each) was employed to identify the stope stability conditions. Overall, the outcome of the analysis showed good agreement with the field data; most stope surfaces were correctly predicted with an average accuracy of 91%. This shows an improvement over using the existing stability graph method. In addition, for a better interpretation of the results, the associated probability of occurrence of stable, unstable, or caved stope was determined and shown in iso-probability contour charts which were compared with the stability graph. The proposed FFNN-based classifier outperformed the conventional stability graph method in terms of accuracy and better prabablistic interpretation. It is suggested that the classifier could be a reliable tool that can complement the conventional stability graph for the design of open stopes.