Abstract:

This study proposed a multi-parameter extraction method for particle-containing droplets based on the DeepViT deep learning model and rainbow scattering, allowing simultaneous and accurate measurement of the host droplet size and volume fraction of inclusions from the extinction rainbow pattern. The basic composition and implementation methods of the DeepViT model were introduced, including training data preprocessing and network hyperparameters setting. Then, the rainbow optical system and typical measurement signals of particle-containing droplets were demonstrated, and the measurement results of this method under different host droplet sizes and inclusion volume fraction were analyzed and compared with the measurement values of the extinction rainbow method. The relative error of the droplet size measured by this method under the inclusion volume fraction of 0—0.3% was within ±0.5%, while the maximum relative error of the extinction rainbow method was about 2%. The maximum absolute error of measuring the inclusion volume fraction under the droplet size of 120—140μm was less than 0.01%. The proposed DeepViT-based method could quickly achieve high-precision in-situ parameter measurements of dynamic heterogeneous droplets such as particle-containing droplets, providing a new idea for the development of particle-containing droplet measurement techniques.

Key words: deep learning, rainbow scattering, particle-containing droplet, droplet size, inclusion volume fraction, multiphase flow

摘要：

提出了一种基于DeepViT深度学习模型和彩虹散射的含颗粒液滴多参数提取方法，实现了从消光彩虹图像中对宿主液滴粒径和内含物体积分数的同时精确测量。介绍了DeepViT模型的基本组成和实现手段，包括训练数据预处理以及网络超参数的设置。随后，展示了含颗粒液滴彩虹光路系统和典型测量信号，分析了该方法在不同粒径和体积分数工况下的测量结果，并与消光彩虹法的测量值对比。本方法在0~0.3%体积分数条件下测量的粒径相对误差均在±0.5%以内，而消光彩虹法最大相对误差约为2%；在120~140μm粒径范围条件下测量内含物体积分数的最大绝对误差小于0.01%。所提出的基于DeepViT方法可快速地实现对动态含颗粒液滴这类非均质液滴的高精度原位参数测量，为含颗粒液滴测量技术的发展提供新思路。

关键词: 深度学习, 彩虹散射, 含颗粒液滴, 液滴粒径, 内含物体积分数, 多相流