A Deep Learning Approach for the Optical Estimation of Particle Size Distributions of Bulk Materials

In energy-intensive industries, such as cement manufacturing, steal production, glass factories or in power plants classical fossil fuels, such as coal or gas, are replaced by more sustainable fuel sources derived from biomass and/or waste streams. However, these fuels tend to be quite volatile according to their main mechanical and chemical properties, such as particle size distribution, bulk density, etc. Therefore, real-time characterization methods need to be developed in order to allow an optimized utilization of such fuels in industry. This paper presents a visual approach for the self-acting estimation of the particle size distribution of a given fuel based on monocular image streams. The suggested methodology follows an image segmentation approach based on deep learning techniques. Due to the lack of ground truth data, an associated methodology for the automatic generation of a synthetic training data set is also presented. Different state-of-the-art segmentation models are implemented and trained on a training database for two different sustainable fuel sources. The performance is carefully validated by using different evaluation metrics. The implementation results showed that the deep learning approach produced more accurate segmented images compared to conventional image processing techniques.