LKABs core business is the mining and processing of iron ore for the steel industry. Over the years, LKAB has developed a unique market offering for its customers, namely blast furnace pellets. These have now become LKABs most vital product with an annual production capacity at its plants in northern Sweden, of28 milliontons per annum. The main benefit of using pellets over standard iron ore products at steel mills, is lower furnace energy requirements. An addition benefit is that the pellets contain extra minerals, such as olivine, which provide improved high-temperature properties.
The R&D team at LKAB are continuously searching for improved ways toproducemore effective aswell ascustomized furnace pellets. Part of this process includes manually inspectingtheformation of different iron oxide phases and microstructures in sample pellets. This is a time-consuming process that must be done by experts, who examine sample pellets embedded in epoxy, through an optical microscope. Some minerals can be identified by color and intensity, such as magnetite and hematite, while others are based onpelletsurface texture. This process lends itself perfectly to machine learning.
Our goal was to automate this process. The first step was to create a dataset of high-quality microscopy images of representative iron ore pellets. Experts at LKAB provided us with sample pellets embedded in epoxy and access to automated microscopes each equipped with high-resolution digital cameras at their facility.
High-magnificationimages of the sample pellets wereacquired,and multiple images were pieced together into large mosaic images to create a complete view of each pellet. Depending on the size of the pellet,the images ranged between 500 and 900 mega pixels.
Once the image library wascreated,the images were annotated in close collaboration with analysis experts at LKAB.Regions from each class weremarkedin a set of images and used to train thepixelClassifier.The Classifier was subsequently used to classify the images and the annotation was improved by correcting the classification. This was repeated until a satisfactory classification was achieved. The Classifier was then evaluated on a set of images that were not included in the training process.
The last step was to extract and quantify relevant information from the classified images and create a visual format that could be easily interpreted by all members of thepelletanalysis team at LKAB. The result was a series of graphs depicting a mineral map, the microstructure and the mineral content of a pellet.
Thanks to the automated mapping of sample pellets, experts can now dedicate their time to analyzing the results, rather than inspecting the pellets. This has speeded up the analysis process considerably. Furthermore, analysis is no longer based on one persons viewpoint, it is now standardized across all pellet samples based on the collective experience and advise of the R&D team.
The automated and quantitative characterization of microstructures can be used to generate the necessary data to better understand the effects of pellet additives and the different process parameters. Additionally, it is now possible to automate a quantitative study of the reaction mechanisms in the blast furnace based on pellet microstructures. With this knowledge, pellet properties can be optimized for different customer applications.
Thanks to a close collaboration with Data Ductus, we were able to define the machine learning project and move through development to implementation quickly and smoothly, saysJohan Sandberg,Section Manager, Process & Product Development atLKAB. Automating the inspection of pellets has freed up time for the R&D team to focus on more valuable and interesting work.