Abstract Regardless of whether the occurrence of foams in industrial processes is desirable or not, the knowledge about the characteristics of their formation and morphology is crucial. This study addresses the measuring of characteristics in foam and the trailing bubbly liquid that result from air bubble entrainment by a plunging jet in the environment of industry-like bottling processes of non-carbonated beverages. Typically encountered during the bottling of fruit juices, this process configuration is characterized by very fast filling speeds with high dynamic system parameter changes. Especially in multiphase systems with a sensitive disperse phase like gas bubbles, the task of its measurement turns out to be difficult. The aim of the study is to develop and employ an image processing capability in real geometries under realistic industrial conditions, e.g. as opposed to a narrow measurement chamber. Therefore, a typically sized test bottle was only slightly modified to adapt an endoscopic measurement technique and to acquire image data in a minimally invasive way. Two convolutional neural networks (CNNs) were employed to analyze irregular non-overlapping bubbles and circular overlapping bubbles. CNNs provide a robust object recognition for varying image qualities and therefore can cover a broad range of process conditions at the cost of a time-consuming training process. The obtained single bubble and population measurements allow approximation, correlation and interpretation of the bubble size and shape distributions within the foam and in the bubbly liquid. The classification of the measured foam morphologies and the influence of operating conditions are presented. The applicability to the described test case as an industrial multiphase process reveals high potential for a huge field of operations for particle size and shape measurement by the introduced method.

中文翻译：

---

通过内窥镜成像和卷积神经网络表征装瓶过程中快速增长的泡沫

摘要 不管泡沫在工业过程中的出现是否合乎需要，了解泡沫的形成和形态特征是至关重要的。本研究解决了在非碳酸饮料的类似工业装瓶过程的环境中测量泡沫和由插入式射流夹带气泡所产生的尾随气泡液体的特性。通常在果汁装瓶过程中遇到，这种工艺配置的特点是灌装速度非常快，系统参数变化很大。特别是在具有敏感分散相（如气泡）的多相系统中，其测量任务变得困难。该研究的目的是在现实工业条件下，例如与狭窄的测量室相反，在真实几何形状中开发和使用图像处理能力。因此，通常尺寸的测试瓶仅稍作修改以适应内窥镜测量技术并以微创方式获取图像数据。两个卷积神经网络 (CNN) 被用来分析不规则的非重叠气泡和圆形重叠气泡。CNN 为不同的图像质量提供了强大的对象识别，因此可以以耗时的训练过程为代价覆盖广泛的过程条件。获得的单气泡和总体测量允许近似、关联和解释泡沫内和起泡液体中的气泡大小和形状分布。给出了所测泡沫形态的分类和操作条件的影响。所描述的测试案例作为工业多相过程的适用性揭示了通过引入的方法进行粒度和形状测量的巨大操作领域的巨大潜力。