1. Introduction

The utilization of both plastic particles and biomass in a fluidized bed system has become increasingly popular and is investigated in many studies due to their great potential for sustainable energy conversion processes such as combustion, gasification, and pyrolysis, to reduce greenhouse gas emissions and abundant supply of the raw material1–3. Most of these recycled raw materials are physically processed into granular materials of non-spherical shape. This poses additional complications in establishing accurate predictions of their behavior during applications since the non-spherical particle-particle interaction and particle-fluid interaction behave rather differently compared to spherical particles4. Such difference is mainly related to particle orientation and corresponding differences in particle stresses. For example, non-spherical particles with high surface-to-volume ratios can have 20 % - 120 % higher average shear particle stress than that of spheroidal paricles5; a free-falling cylinder will align itself with the axis parallel to the flow for moderate Reynolds numbers6; non-spherical particles of large aspect ratio are difficult to be fluidized due to interlocking between particles7. Spherical particles are usually used to assist the fluidization of non-spherical particles in typical industrial applications8. Due to the difference in particle physical properties, such as shape, size, and density, the non-spherical particle can affect the fluidization of spherical particles, such as the transition of fluidization regimes9, changes of fluctuation frequency10, and so on. Moreover, the non-spherical particles can separate from spherical particles, which leads to segregation. However, segregation is not preferable because non-uniform mixing can significantly decrease the bed performance. Understanding the dynamics of non-spherical particles and the effects of non-spherical particles on binary fluidization is critical for the understanding of fundamental physics and the associated industrial applications.
In situ experimental studies on non-spherical particle dynamics in fluidized beds were seldomly reported in the literature as related experiments to resolve particle-scale information such as position, orientation, velocity, and size are rather costly and complicated. Buist et al.11 measured the translational and rotational velocities of cylindrical particles with varying elongation ratios using magnetic particle tracking (MPT) in a cylindrical fluidized bed with 17.4 cm in diameter. Fotovat et al.12 investigated the biomass particle shape factor on the biomass distribution and velocity profiles in a spherical-particles-assisted binary fluidization system using the radioactive particle tracking (RPT) method. Chen et al.13 measured the 3D particle position and velocity of a single tagged cylindrical particle over a long period in the binary fluidized bed using X-ray particle tracking velocimetry (XPTV). Vollmari et al.14 investigated the distribution and orientation statistics of non-spherical particles in a rectangular bed using the in-house image analysis algorithms. Studying the dynamics of particles through processing images obtained from a high-speed and high-resolution camera is a relatively straightforward method of acquiring quantitative data compared to other complex measurement techniques discussed above.
Efforts have been made to increase the capability of particle scale imaging techniques to higher particle volume fractions by improving the particle detection algorithms or limiting the sample depth. In the processing of the image for non-spherical particle and spherical particle binary fluidization, the main challenge is the segmentation of the non-spherical particles from the spherical particle and background. Segmentation is a process to partition an image into multiple parts or segments. Classical image segmentation methods include histogram thresholding, edge detection based, relaxation, and semantic and syntactic approaches15. Each approach has its advantages and limitations. For example, region-based segmentation separates the objects into a different region based on an automatically or manually determined threshold value. It is simple, fast, and performs well when the target object and background have high contrast. However, the accuracy of this method becomes very low when the contrast is low and there is a large overlap. Some studies employing classical image segmentation are available in the literature. Yin et al.16 applied an image multilevel thresholding approach using the k-means algorithm to identify clusters in a fluidized bed riser. Jiang et al.17 employed a particle-mask correlation segmentation approach to detect the particle geometric center in a 2D fluidized bed. In recent years, the machine learning approach demonstrates great promise in the field of particle-scale data extraction from images in multiphase flow research18–24. For example, Yevick et al.18 measured particle size and positions from analyzing the holographic video microscopy data using machine learning techniques based on support vector machines (SVMs) in real-time on low-power computers. Shao et al.24 developed a convolutional neural network (CNN)-based approach for measuring the 3D particle distributions using digital in-line holography.
The aim of this study is to experimentally investigate the cylindrical particle dynamics and their impact on binary fluidization using image processing and pressure signal analysis. In section 2, the experiment setups and the methodology of the machine learning-enabled image processing methods are introduced. In section 3, the results and discussion of pure LDPE sphere fluidization behavior, effects of cylindrical particles mass fraction, superficial gas velocity, and sphere inventory on the cylinder dynamics and fluidization behavior were presented. In section 4, the results were summarized and further discussion was presented.