Bubble Detection in Gas-Solid Separation Fluidized Beds Based on Deep Learning

In dry coal beneficiation with gas-solid fluidized beds, bubble dynamics critically affects fluidization stability and separation efficiency. This study proposed an improved YOLOv8-based bubble detection model to achieve accurate and real-time bubble monitoring. The model integrates a Multi-Head Self-Attention (MHSA) mechanism to enhance global feature extraction, a multiscale feature fusion structure (BiFPN-CONCAT) to achieve efficient feature integration, and an Involution-based Decoupled Head to improve detection precision while reducing computational complexity. Experimental validation demonstrated that the proposed model achieved superior detection performance, with precision, recall, and mAP@0.5 being 99.1, 96.0, and 95.5, respectively, outperforming the YOLO series and mainstream detectors such as Faster R-CNN and Mask R-CNN. Moreover, as was revealed through the analysis of 120 experimental data sets, average bubble area was strongly negatively correlated with ash content $\left(r=-0.72\right)$ , while bubble number was positively correlated with clean coal yield $\left(r=+0.90\right)$ . A regression model based on bubble features achieved a coefficient of determination of $R^2=0.89$ , confirming their predictive value for separation performance. These findings demonstrate that the proposed model not only ensures high-precision bubble detection but also provides new insights into the coupling between fluidization dynamics and the beneficiation efficiency. This study offers theoretical and technical support for intelligent dry coal separation systems.