Characterizing Coke Particle Size Distribution Using CondInst Instance Segmentation and LightGBM Regression Model

Coke particle size distribution (PSD) is crucial for assessing the quality of furnace coke. Current methods relying on manual or semiautomated techniques exhibit inefficiencies and inaccuracies due to irregular shapes and adhesion issues among particles. Thus, a novel architecture integrating CondInst instance segmentation and regression prediction models is proposed for investigating the detailed features and PSD of coke particles. First, a split-attention (SA) mechanism is integrated into the feature extraction process to sharpen the focus on relevant particle features while suppressing irrelevant information. Meanwhile, a feature alignment module (FAM) is applied to maintain precise shapes and sizes by addressing spatial misalignment during feature fusion. Second, an efficient multiscale attention (EMA) is introduced to capture advanced feature and optimize particle masks, improving the accuracy of segmenting adhesion areas. Finally, an estimation model based on light gradient boosting machine (LightGBM) is established to map the captured 2-D features to particle thickness, enabling the characterization of the PSD. Experimental results demonstrate that the proposed model performs better than other segmentation models, and the estimation model can stably characterize the PSD of coke, with a mean relative error (MRE) below 8% for distinct particle size ranges. This provides a novel and efficient way to rapidly analyze coke PSDs from images.