Rapid, stochastic updating of resource knowledge with sensor information for high-resolution reconciliation and rapid decision-making.

This project focuses on rapidly updating resource knowledge with upstream (drill) and downstream (belt) sensor information, in collaboration with partner organisations, Think & Act Differently by BHP, Maptek and Petra Data Science. The upstream and downstream information contains the resource knowledge close to the mining stage, allowing for rapid reconciliation of the ore attributes for both upstream and downstream use.

Resource models are generally constructed from directly observed data (e.g., grades of drill cores) that have relatively high accuracy. Resource models, however, are limited by the scale on which the data are collected. As mining progresses, more information becomes available on different scales from various types and sources of data. Drill cuttings from blast holes provide in situ data on smaller scales. Sensors on drilling rigs and on conveyor belts provide data in near real-time, as do draw-point sensing and Load Haul Dump (LHD) sensors.

All these types of data – drill core, drill sensors, draw-point sensors, LHD sensors, belt sensors – measure different volumes in different ways and have different levels of accuracy, all of which must be accounted for when integrating the various types of data. Although most data will be quantitative (e.g., ore grades), some will be qualitative (e.g., structures such as geology or domain types) and non-additive (e.g., geometallurgical attributes related to comminution and processing). Integrating these types of data requires a stochastic approach to account for the different levels of accuracy (uncertainty) and the different volumes of measurement. To enable rapid decision-making, the resource model must be updated with newly acquired data in near real-time.

This project involves data assimilation for updating primary resource variables, as well as multi-Gaussian transformation and Gibbs sampling for converting complex data into variables suitable for data assimilation. Machine learning then predicts the non-additive response variables, such as geometallurgical attributes, using updated models of the primary resource variables.