Ternary semantic segmentation of hybrid particles in fluidized beds via dual CBAM-DeepLabV3+: Enhanced characterization of nanoparticle agglomeration

Accurate characterization of nanoparticle agglomeration in mixed-particle micro fluidized beds remains challenging because conventional binary semantic segmentation cannot effectively distinguish nanoparticles from Geldart B particles. To address this issue, a ternary semantic segmentation framework based on an improved DeepLabV3+ architecture is developed, in which pixels are classified into background, nanoparticles, and Geldart B particles. Dual convolutional block attention modules (CBAM) are incorporated to enhance multi-scale feature extraction and boundary recognition under complex fluidization conditions. Compared with baseline models, the proposed framework improves the mean intersection over union (mIoU) by 4.00%, enabling reliable pixel-level separation of mixed particles. Based on the high-precision segmentation results, the evolution of nanoparticle agglomeration characteristics and radial distribution behavior under different superficial gas velocities is quantitatively analyzed. The results show that introducing Geldart B particles suppresses excessive agglomerate growth, reduces the average equivalent diameter, and promotes the formation of more compact and morphologically regular agglomerates. Enhanced local hydrodynamic disturbance and heterogeneous particle interactions further promote agglomerate restructuring and modify the competition between aggregation and breakup processes during fluidization. Overall, the proposed framework provides an effective tool for quantitative investigation of agglomeration behavior and mesoscale flow structure evolution in complex gas–solid fluidized systems.