Abstract

MineSense Technologies equips mining shovels, loaders, and conveyor belts with X-ray fluorescence sensors to predict rock grades and other properties in real-time. The primary purpose is the prediction of material grades at the highest resolution – for each bucket at the post-blast mine face, and at high-frequency intervals on the belt. Predictions are integrated with the mine’s fleet management system for real-time bulk ore sorting during excavation and can be used in advanced process control to improve processing performance, producing a higher and more consistent recovery through a mine-to-mill optimization. Further investigation of XRF data using machine learning techniques has identified correlations between spectral features and other rock properties, such as acid-generating potential or grindability. These techniques are used to build more advanced models that yield further optimizations from the mine face to downstream processing.

For rock grade predictions using XRF, univariate linear models are calibrated by training supervised models using mine site data as the true grade response. Blast hole sample assays or the ore grade control block model are used to identify strong correlations with in-bucket XRF sensor measurements at a global, aggregated resolution. Value can then be generated locally when the model is deployed at a high bucket resolution and thus measures post-blast movement and intricate geological features that are below the detection limit of traditional ore control practices. There are further statistical considerations to factor in modelling, including the spatial misalignment of the predictors and the response, the heteroscedasticity of the conditional distribution of grade, and the spatiotemporal correlation in the variables.

A key statistical consideration arises due to the surficial nature of XRF, as the surface-to-volume relationship of the measured rocks can depend on lithology. For example, given a particular grade, it has been observed across several sites that XRF readings can differ substantially when the rocks come from skarn versus porphyry dominant areas. Under these situations, auxiliary variables are required to select rock type-specific models. To handle this, an XRF-based lithology classifier is built and validated, with predictor variables identified by leveraging geological knowledge and applying statistical feature selection techniques. The classifier is then used as the first step in a multi-stage grade modelling pipeline in which separate grade models are selected for relevant lithologies.

Additional benefits can be extracted from a stand-alone application of a lithology classifier. For instance, the optimal grinding and flotation operational settings for skarn and porphyry ore can differ significantly. Thus, preemptive identification of the rock-type can be used to implement a preventative rather than reactive method of operation, leading to greater mill processing efficiency.

This presentation will provide an overview of the methodology used in developing and validating lithology classifiers and the corresponding lithology-based ore grade models. Case studies across several sites will be explored, with the improvement in ore extraction on both mine plans and XRF-based grade models invariant to lithology demonstrated.