Optimization of Mining–Mineral Processing Integration Using Unsupervised Machine Learning Algorithms

Abstract

Traditional ore-waste discrimination schemes often do not take into consideration the impact of fluctuations of the head grade, which have on the performance of mineral processing facilities. This research introduces the use of target grades for processing destinations as an alternative to cut off grade-based methods and models, in each processing destination, the losses due to deviation from targets via the Taguchi loss function. Three unsupervised learning algorithms, k-means clustering, CLARA and k-mean-based approximate spectral clustering, are presented to group mine planning blocks into clusters of similar grades with different processing destinations. In addition, a technique considering uncertainties associated with block grades is proposed to generate new sequences that reduce variation in processing capacities across the life of mine (LoM). The case study in this paper involves the treatment of a realistically large mining dataset. The results showed that clustering methods outperform cutoff grade-based method when divergence from target grades is penalized and that reclassification of blocks based on data from geostatistical simulations could achieve smoother capacities for processing streams across the LoM.