Exploring Deep Learning for Dig-Limit Optimization in Open-Pit Mines

This paper explored a convolutional neural network to assess the clustering of selective mining units (SMUs), generated by a genetic algorithm (GA) approach. The purpose of the GA was to optimize a set of dig limits in a mine bench. Dig limits are boundaries to separate/split SMUs as per waste and other processing categorizations so they can be feasibly and profitably be extracted by existing mining infrastructure. This was achieved by involving the cluster design and the affiliated decision variables into existing mining optimization frameworks that enhance the process's efficiency and thereby increase profitability. However, the catch being that such computation comes costly both from a monetary and time consumption perceptive. This paper aimed to address this very cost issue and how to overcome this with the use of deep learning. Specifically applying the statistical learning algorithms to assess the cluster quality of the GA computed dig limit, as current assessment methodology consumes up to 70 percent of the total computation time. Short-term mine planning applications need to be directly usable by mine operators' personnel and must be generated quickly to be useful in dynamic mining environments. This requirement shall save up time and cost. A case study was conducted on a bench with multiple destinations and 420 SMUs to test whether a convolutional neural network (CNN) can make predictions on clustering quality and discover the best CNN architecture for the task. The case study showed that using a CNN does ensure prediction accuracy and can speed up the time of computation time by 3900%.