Characterisation of ore and bulk solid systems by use of multivariate image analysis and deep learning neural networks

The Challenge

Despite advances in sensor technology, machine vision techniques struggle to provide the level of visual field definition needed to support decision making in mining and minerals processing.

Key Findings

Convolutional neural network (CNN) processing is able to characterise visually complex fields such as those defining froth flotation systems, ore textures, and drill core imagery
Pretrained CNNs can be used for interpreting mineral-processing-related imagery, with performance further improved through transfer learning and fine-tuning.
Soft sensor technologies applying this approach can deliver near-perfect accuracy in mining applications

Benefit to Western Australia

The image processing technologies developed in Yihao’s work could support implementation of automated control processes in many areas of material processing. This approach could deliver efficiencies in key areas of the WA mining and minerals sector including froth flotation monitoring, rapid ore-sorting and particle size analysis.

Supervisor

Professor Chris Aldrich

Thesis

Characterization of Ore and Bulk Solid Systems by Use of Multivariate Image Analysis and Deep Learning Neural Networks