Eventos Anais de eventos
COBEM 2021
26th International Congress of Mechanical Engineering
Bubble Pattern Recognition from Particle Image Velocimetry (PIV) Images using a Deep-Learning-Based Image Processing Technique
Submission Author:
Rafael Franklin Lazaro de Cerqueira , SC
Co-Authors:
Rafael Franklin Lazaro de Cerqueira, Marco Antônio Cerutti, Emilio Paladino
Presenter: Rafael Franklin Lazaro de Cerqueira
doi://10.26678/ABCM.COBEM2021.COB2021-0893
Abstract
This work presents a deep-learning image processing tool for the analysis of bubbly flows using Particle Image Velocimetry (PIV) technique. The measurement in gas-liquid bubbly flows is a challenging task, mostly due to the laser light's dispersion caused by the gas-liquid interfaces. A common solution is the use of fluorescent seeding particles associated with a bandpass filter for the laser light, which consists of the PIV technique combined with laser-induced fluorescence (LIF), known as PIV/LIF. When using the additional LIF setup, the interfaces are not fully removed from the acquired images, once they still reflect the fluoresced light. This is not beneficial for the PIV acquisitions, since the measurement technique is based on the movement of tracer particles located in the liquid phase. In the past few years, several authors proposed image processing techniques to alleviate this problem and enable the use of the PIV technique for the characterization of gas-liquid bubbly flows. The present work aims to use this drawback, the appearance of the gas-liquid interface in the PIV acquisitions, as a tool to completely characterize the gas phase. Due to the high noise level from the seeding particles into the dispersed bubble interface, which appears as bright particles in the acquired PIV images, traditional computer vision techniques algorithm fails to detect the dispersed bubbles in the images. Hence, the current work presents a novel deep-learning-based imaging processing technique based on the use of a U-Net Convolutional Neural Network (CNN), employed for semantic image segmentation tasks, and a Faster-RCNN, a region-based object detection convolutional neural network. Additionally, a second CNN is used to reconstruct the dispersed gas bubble interface through ellipsoids. Thanks to the bubble reconstruction capability of the proposed technique, a Labelled Object Velocimetry (LOV) algorithm can be incorporated into the developed framework to calculate the instantaneous velocities of the dispersed bubbles from the two recorded PIV frames.In this way a full characterization of the liquid and gas flow field can be performed only from the PIV acquisitions.The developed methods are tested with PIV acquisitions from a set of upward laminar and turbulent vertical bubbly gas-liquid flows. Ensemble average bubble velocity profiles are calculated in different experimental conditions, showing that the global void fraction affects the gas velocity profiles.
Keywords
PIV, bubbly flow, LOV, Deep learning
