Image segmentation and CNN-based deep learning architectures for the modelling on particulate matter formation during solid fuels combustion

Three typical solid fuels (coal, biomass, and refuse-derived fuel) were individually combusted in a lab-scale drop tube furnace. The computer Vision method was employed to extract the morphological characteristics of particulate matter (PM) and establish a dataset of 12,637 particle features. Six convolutional neural network models were developed, and three transfer learning strategies were studied. The ResNet50 model achieved a peak accuracy of 96.6 % when fine-tuned across all layers, demonstrating its exceptional capability to identify irregular, angular, agglomerated, and completely melted particles. Based on this model, the predominant form of PM 10 produced from all three fuels was irregular, ranging from 28.91 to 81.37 wt%, whereas PM 10–200 consisted primarily of 9.92 to 49.44 wt% angular, 5.10 to 39.59 wt% agglomerated, and 0.08 to 39.06 wt% completely melted forms. By combining combustion experiments and thermodynamic equilibrium calculations, it was proven that Si, Na, K, and Cl form irregular particles as the major types of PM 10 . Si Al readily forms ‘angular’ PM 10–200 , whereas collisions with fine particles in the gas phase leads to the formation of agglomerated particles of Na/K-Al-Si. Ca/Fe-Al-Si formed completely melted particles of 59.06 wt% in total. • A detection model based on deep learning for particle morphology classification was established. • The ResNet50 model coupled with a transfer learning strategy achieved a peak accuracy of 96.6 %. • Particles with irregular shape constitute the predominant morphological class of PM 10 . • PM 10–200 yielded from RDF combustion primarily comprised completely molten Ca/Fe-Al-Si.