Quantitative characterisation of iron ore pellets based on optical microscopy and machine learning

Purpose and goal

The goal was to evaluate the potential of a system for quantitative characterization of iron ore pellets based on automated microscopy, image analysis and machine learning. The system could, after training with annotated data, reliably identify a number of relevant phases like hematite, magnetite and metallic iron. The results were well matched with manual assessment and contributed more information than previous systems. The potential for the new approach is therefore considered to be very large, and could contribute to job saving and increased knowledge.

Expected results and effects

A relevant dataset with annotated microscopic images with iron ore pellets was created. This was then used to train and evaluate a number of classifiers. Based on this, relevant microstructures could be quantified. Amount and distribution of different phases, minerals and additives. In addition, porosity and size distribution of particles could be determined. With the help of collected and annotated data, the concept could be validated in a laboratory environment. Code optimization and user interface development were assessed as prerequisites for a commercial system.

Planned approach and implementation

Annotation of collected data was done in collaboration with application experts, which ensured the quality of training and test data. Some phases proved difficult to evaluate even for experts with optical microscopy. Additional technologies such as SEM may be necessary to determine the correct phase. Collected data can also be used to train and test new machine learning methods in the future, such as convolutional neural networks (CNN).