HIGH PROFILE – Comex

obtained, Kolacz says. This number can be even higher if more sensors are employed in the identification process. In addition, this neural network analysis must be carried out in very limited time intervals ranging from 20-30 milliseconds, thus it always requires structure optimisation.

XRT sensor response indicating Zn / Pb ore inclusions( yellow / green) in waste rock material( blue)

information can be useful for detecting low concentrations in disseminated orebodies.

Data fusion by AI models

Output signals from different sensors can be processed by AI models to provide the best mineral or material identification, according to Kolacz. Identification works by detecting features hidden in the convolutional layers of the neural network. The correct selection of neural networks in every case considers overfitting, underfitting and robustness analysis. For the first two analyses, the goal is to identify networks capable of inferring general material patterns with minimal noise mapping ability. In turn, robustness analysis aims to identify such a structure that, in the event of changes in input data, can generate the correct responses for the device’ s input material evaluation system. This is especially important when the geological structure of the sorted particles changes in time due to deposit variations. The approximate number of network training processes performed in the process of network structure optimisation exceeds 50,000 structures. Thanks to such a complex process, satisfactory results are

Sorting results

The choice of sensors for sorting is dependent on the types of sorted materials. Nevertheless, the XRT sensor can be very useful, especially for metal ores, where the whole particle volume is analysed. The combination of this sensor with either colour analysis( RGB camera) or hyperspectral SWIR analysis can be very effective in identifying some minerals or mineral groups, as shown below.

Single XRT sensors

The XRT sensors can be used for many applications without combining them with other sensing techniques but employing AI models. One of the applications of preconcentration of the Zn / Pb ore in the Comex industrial sorter CXR-1000 is an example. In this case the feed material had 1.8 % Zn and 0.8 % Pb and the upgraded product fraction provided 10.96 % Zn and 5.4 % Pb. The yield of the product fraction was only 13 %, which meant 87 % of the input stream was removed as the waste fraction, without any further

www. comex-group. com