Particle classification of iron ore sinter green bed mixtures by 3D X-ray microcomputed tomography and machine learning

The iron ore sintering process needs to be optimised to decrease its energy intensity and emissions of carbon and atmospheric pollutants, while continuing to produce sinter of sufficient quality for current and future low carbon blast furnace operations. Ideally, the sinter structure and mineralogy should be related back to the particle-level structure of the iron ore types mixed from different mine sources. This particle-level detail can be visually obtained by 3D X-ray micro-Computed Tomography (micro-CT), but requires subsequent algorithms to individually identify and classify particles and identify the relationship between ore sources and sinter quality. In this study, individual particles in sinter green — beds comprising a mixture of coking coal, fluxes, return fines and 5 iron ore samples from different mine sources are identified and classified in high resolution micro-CT images using a machine learning algorithm and associated data processing workflow. Coking coal, fluxes, and return fines are first segmented from iron ores based on their X-ray attenuation and texture. By imaging individual samples from each iron ore source, reliable training data is readily obtained from particle isolation with Convolutional Neural Networks (CNNs) guided by Trainable Weka Segmentation (TWS). Supervised machine learning is then applied to the datasets of isolated particles to produce a per-particle segmented digital sinter green bed image. A collection of geometric, texture, and greyscale features are computed for the particles and used to train a gradient boosting classifier. Tests are then performed on unseen subsets of the single ore source data, on a stratified mixture, and on a random mixture. An accuracy over 90% is achieved for iron ores that are morphologically domain-distinct in their feature space, while lower accuracy in the order of 40%–80% is achieved between iron ore particles that derive from different mine sources, but are domain-similar, suggesting similar mineralogy. The effect of limited training domain, the visual/morphological/feature space similarities and the resulting domain shift in data between training and testing are carefully analysed to identify major sources of similarity. This per-particle multilabel classification of sinter green bed mixtures distinguishes both similar and distinct ores from different mines, and provides a high resolution, accurately characterised digital twin analogue of mixed iron ore sinter green beds. This allows for future detailed analysis of sinter quality, energy intensity, and carbon emissions during the metallurgical process, all of which could be optimised to produce cleaner, higher quality iron.