Study of bubble behavior in a gas–solid dense-phase fluidized bed based on deep learning

[Display omitted] •An automatic bubble segmentation method based on deep learning is proposed.•Achieving a pixel accuracy of 97.95% and 80.91% MIoU for bubble image segmentation.•The results were significantly improved compared to threshold segmentation methods.•Coalescence is the primary cause of the increase in bubble size. Fluidization characteristics and hydrodynamics of gas–solid fluidized beds are controlled by bubble properties. However, the existing conventional image segmentation methods are unable to effectively resolve the problems of overlap and background interference between small bubbles, and the segmentation performance for bubbles must still be enhanced. To study the bubble behavior of a gas–solid dense-phase fluidized bed, a two-dimensional fluidized bed with different air distributors was adopted. The application of deep learning to digital image analysis technology is to improve the precision and automation of bubble recognition. The DeepLab V3+ model incorporates multi-scale information, which is well-compatible with uncertain bubble boundary information while enhancing the segmentation effect. A pixel accuracy of 97.95% and a mean intersection over union (MIoU) of 80.91% for multi-group bubble image segmentation were achieved. In comparison to conventional threshold-based recognition methods, the accuracy of the deep learning method increased by up to 40.31% and the MIoU by up to 53.88%, demonstrating the great potential and application value of semantic segmentation in the study of the complex motion behavior of gas–solid sorting fluidized bed bubbles. The influence of the air distributor on the behavior of bubbles is verified. The results indicate that air distributor aperture directly influences bubble diameter, aspect ratio, and shape factor, with coalescence being the primary cause of bubble enlargement.