Abstract

MineSense Technologies equips shovels, loaders, and belts with XRF sensors to provide real-time measurements of material from excavation to downstream processing. The primary purpose of the measurements is to predict the grade of material at the highest resolution, namely, for each bucket at the mine face and at high-frequency intervals on the belt. The bucket grade predictions from ShovelSense© are aggregated to classify individual trucks as ore or waste, with classifications synchronously integrated with the mine’s fleet management system to reroute trucks to the correct destination. Thus, ore that would have otherwise gone to waste can be recovered, and conversely the amount of waste material sent to the mill is reduced. The belt predictions from BeltSense© can then be integrated into Advanced Process Control (APC) systems in the mill to further improve mine-to-mill processing and tracking. The key advantage of these technologies is the large scale of the datasets generated by the high resolution of measurements, which enables detection of an often complex and variable ore body post-blast in contrast to the coarse pre-blast block model. Overall, this achieves higher economic and environmental efficiency by reducing grade variability and increasing the head grade of the mill feed – ultimately allowing the mine to extract more ore at a higher rate of recovery.

Grade prediction captured by a spectral element response is, however, not the extent of predictive information held within a full XRF spectrum. Further investigation of these readings using key mine datasets and the application of machine learning techniques has led to the discovery of relationships between features of the spectra and other rock properties, which can be used to build more advanced models that yield additional economic and environmental benefits. Methodologically, these relationships are identified by engineering problem-specific feature sets, by leveraging physical and geological domain knowledge to extract informative predictor variables from the spectrum that are both directly and indirectly related to elemental composition. Potential challenges include confounding factors such as the spatiotemporal correlation of these properties and spatial drift in the conditional distribution of the response. Robust validation of these advanced models is imperative to prove value and can be conducted quantitatively by tracking relationships between predictions and relevant downstream mill processing KPI’s, and qualitatively by collaborating with subject matter experts at site for regular model reviews.

This presentation will provide a holistic overview of the methodology in building advanced rock property prediction models using XRF data, as well as specific example use cases. One such use case is waste sorting by the quantification of acid-generating potential, which can result in both increased waste material for construction and a reduction in acid drainage from trucks routed to incorrect destinations. Other use cases with significant impact on mill processing performance include the prediction of copper mineral speciation, ore leachability, and grindability. When combined with bulk ore sorting via grade prediction, these models provide an opportunity to optimize the entire mine-to-mill pipeline and thus enact more sustainable and profitable mining practices.