基于卷积神经网络的大尺寸气泡体积二维图像测定方法

胡宁宁; 党卓然; 张牧昊; 唐可; MamoruIshii

doi:10.13832/j.jnpe.2021.06.0038

基于卷积神经网络的大尺寸气泡体积二维图像测定方法

doi: 10.13832/j.jnpe.2021.06.0038

1.
中国核动力研究设计院，成都，610213
2.
成都理工大学核技术与自动化工程学院，成都，610059
3.
普渡大学核工程学院，美国西拉法叶，IN 47907-2017

基金项目: 国家建设高水平大学公派研究生项目（201706050102）

详细信息

作者简介:
胡宁宁（1990—），女，工程师，现主要从事反应堆运行与检修工作，E-mail: 804075335@qq.com

通讯作者:
张牧昊，E-mail: zhangmuhao@cdut.edu.cn

中图分类号: TL33
计量
- 文章访问数: 489
- HTML全文浏览量: 92
- PDF下载量: 53
- 被引次数: 0
出版历程
- 收稿日期: 2020-10-10
- 修回日期: 2020-11-16
- 刊出日期: 2021-12-09

Measurements of Large Bubble Volume Based on 2-D Images Processing Applying Convolutional Neural Network

1.
Nuclear Power Institute of China, Chengdu, 610213, China
2.
College of Nuclear Techonology and Autornation Engineering, Chengdu University of Technology, Chengdu, 610059, China
3.
School of Nuclear Engineering, Purdue University, West Lafayette, IN 47907-2017, USA

摘要

摘要: 由于表面张力与惯性力作用，静止流场内较大尺寸气泡[500<气泡雷诺数（Re）<2000]形成不规则几何形状，造成二维图形处理方法等效球体或椭球获取三维体积的方式误差较大。此外，由于不规则界面的散射和反射，引起二维图像处理中边界模糊，难以辨识。本文以高速摄像机获得的静止流场内大尺寸气泡二维灰度图像作为卷积神经网络（CNN）的输入，以图像内气泡二维投影面积及实验获得三维体积训练网络，并用训练好的网络预测气泡体积。实验采用小气泡叠加法获得真实气泡体积，与网络预测结果进行对比。结果表明，该方法与传统图像处理方法相比，不需要对气泡形状进行假设，提高了对大尺寸气泡的适用性。
- 气泡等效体积直径 /
- 卷积神经网络（CNN） /
- 图像处理
Abstract: Large bubbles (500<Re<2000) in static flow field have the irregular geometrical shape due to the effects of surface tension and inertia force, which brings the enormous error to spherical or ellipsoid equivalent method used in 2D image processing to obtain 3D volume. Besides, due to the refraction and reflection on the irregular surface, the bubble boundary in 2D image is blurred and unrecognizable. The present research applied high-speed video camera to obtain 2D gray scale images of large bubbles in static flow field and used the images as the input of the convolutional neural network (CNN). The bubble 2D projected area and volume obtained in experiments were applied to train the CNN. Finally well-trained CNN was applied to predict the bubble volume. In experiment, actual bubble volume was obtained by small bubble superposition and compared with the predicted value of CNN. The results show that comparing with the traditional image processing method, the proposed method needs no assumption for bubble shape and improves the applicability for large bubble.
- Bubble equivalent volume diameter /
- Convolutional neural network（CNN） /
- Image processing

HTML全文

图 1 实验装置示意图

Figure 1. Schematic Diagram of Experiemntal Facility

下载: 全尺寸图片幻灯片

图 2 图像处理方法流程图

Figure 2. Flow Chart of Image Processing Method

下载: 全尺寸图片幻灯片

图 3 训练集图像处理示例

Figure 3. Examples of Image Processing Applied by Obtaining Training Set

下载: 全尺寸图片幻灯片

图 4 CNN结构

Figure 4. Structure of CNN

下载: 全尺寸图片幻灯片

图 5 网络训练及验证误差

Figure 5. Network Training and Verification Error

下载: 全尺寸图片幻灯片

图 6 CNN预测气泡投影面积与实验数据的误差

Figure 6. Error of Bubble Projected Area between Prediction of CNN and Experimental Data.

下载: 全尺寸图片幻灯片

图 7 气泡等效体积直径预测结果与实验数据的相对误差

Figure 7. Relative Error of CNN Prediction of Bubble Equivalent Volume Diameter to Experimental Data

下载: 全尺寸图片幻灯片

参考文献(14)

[1]	LUNDE K, PERKINS R J. Shape oscillations of rising bubbles[M]. Dordrecht: Springer, 1998: 387-408.
[2]	CLIFT R, GRACE J R, WEBER M E. Bubbles, drops, and particles[M]. Mineola: Dover Publications, 2005: 340.
[3]	ISHII M. One-dimensional drift-flux model and constitutive equations for relative motion between phases in various two-phase flow regimes: ANL-77-47[R]. Illinois: Argonne National Lab., 1977.
[4]	TRIPATHI M K, SAHU K C, GOVINDARAJAN R. Dynamics of an initially spherical bubble rising in quiescent liquid[J]. Nature Communications, 2015(6): 6268. doi: 10.1038/ncomms7268
[5]	LAU Y M, DEEN N G, KUIPERS J A M. Development of an image measurement technique for size distribution in dense bubbly flows[J]. Chemical Engineering Science, 2013(94): 20-29. doi: 10.1016/j.ces.2013.02.043
[6]	FU Y C, LIU Y. Development of a robust image processing technique for bubbly flow measurement in a narrow rectangular channel[J]. International Journal of Multiphase Flow, 2016(84): 217-228. doi: 10.1016/j.ijmultiphaseflow.2016.04.011
[7]	TOMIYAMA A, CELATA G P, HOSOKAWA S, et al. Terminal velocity of single bubbles in surface tension force dominant regime[J]. International Journal of Multiphase Flow, 2002, 28(9): 1497-1519. doi: 10.1016/S0301-9322(02)00032-0
[8]	MAJUMDER S K, KUNDU G, MUKHERJEE D. Bubble size distribution and gas–liquid interfacial area in a modified downflow bubble column[J]. Chemical Engineering Journal, 2006, 122(1-2): 1-10. doi: 10.1016/j.cej.2006.04.007
[9]	LI Z C, SONG X M, JIANG S Y, et al. The lateral migration of relative large bubble in simple shear flow in water[J]. Experimental Thermal and Fluid Science, 2016(77): 144-158. doi: 10.1016/j.expthermflusci.2016.04.013
[10]	LEWANDOWSKI B, ULBRICHT M, KREKEL G. An automated image analysing routine for estimation of equivalent diameter in high-speed image sequences with high accuracy and its validation[J]. Experimental Thermal and Fluid Science, 2018(98): 158-169. doi: 10.1016/j.expthermflusci.2018.05.016
[11]	王红一. 水中气泡图像处理方法及运动特性研究[D]. 天津: 天津大学, 2011.
[12]	CANNY J. A computational approach to edge detection[M]. Los Altos: Morgan Kaufmann, 1987: 203.
[13]	WONG S C, GATT A, STAMATESCU V, et al. Understanding data augmentation for classification: when to warp?[J]. arXiv, 2016, 45(2): 529-534.
[14]	COUVERT A, ROUSTAN M, CHATELLIER P. Two-phase hydrodynamic study of a rectangular air-lift loop reactor with an internal baffle[J]. Chemical Engineering Science, 1999, 54(21): 5245-5252. doi: 10.1016/S0009-2509(99)00246-8