The Role of Multispectral Scrap Characterisation and Artificial Intelligence in Efficient Steel Recycling

Recycling of steel scrap is an important component in the modern steel industry since it reducesngreenhouse gas emissions, helps to conserve primary raw materials, and facilitates the steelnindustry's decarbonization efforts. However, the composition of steel scrap is highly fluctuatingnand can contain a large range of unwanted particles, including non-ferrous metals, plastics, ornconcrete. A good characterisation of the scrap composition is therefore key for an efficient steelnrecycling process.nThis work investigates the potentials of hyperspectral imaging and image segmentation to characterisenthe scrap particles, so that said characterisation can then be used for optimisation ornadaptation of the scrap recycling process. Hyperspectral imaging covers a spectrum much largernthan then the visible light spectrum. Hence, it is easier to distinguish materials by their hyperspectralnprofile then by just the visible light profile. The image capturing in this work is done by anhyperspectral system operating in the short-wavelength infrared (SWIR, up to 2500 nm) rangesnsampled in 288 individual bands.nThe sheer volume of information in hyperspectral images, makes it impossible for humans toninspect these images in an industrial process. Therefore, Artificial Intelligence Methods, in particularnMachine Learning (ML) methods, are employed to segment the images into regions coverednby different materials. Deep Neural Networks (DNNs) have been selected here, since they do notnneed any dimensionality reduction and can directly use the 288 colour channels as their input.nThe results show that the DNNs deliver a good segmentation performance, namely that the DNNncan identify the correct object class in most cases. Moreover, when misclassifications occur, thenincorrectly assigned classes are often those whose spectral signatures are similar to the actualnmaterial, for instance, copper gets frequently misclassified as iron or paper as plastic.nThe scrap characterisation provided as the output of the image segmentation is an importantninput for the digitalisation of scrap recycling since it provides information about impurities in thensteel scrap. This information has potential future use for optimisation and automatic adaption innsteel recycling.nThis work is part of the project InSpecScrap funded by the Future Fund of the State of Styrian(“Zukunftsfonds Steiermark”) by grant No. PN1510.