基于深度学习的气液两相流气泡分割算法  

Bubble segmentation algorithm for gas-liquid two-phase flow based on deep learning

作　　者：郭春雨[1,2] 汪永号[1] 韩阳 GUO Chunyu;WANG Yonghao;HAN Yang(School of Ship Engineering,Harbin Engineering University,Harbin 150001,China;Qingdao Innovation and Development Center of Harbin Engineering University,Qingdao 266000,Shandong China)

机构地区：[1]哈尔滨工程大学船舶工程学院,黑龙江哈尔滨150001 [2]青岛哈尔滨工程大学创新发展中心,山东青岛266000

出　　处：《华中科技大学学报（自然科学版）》2023年第11期157-164,共8页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金青年基金资助项目(52001090)。

摘　　要：针对气液两相流中重叠气泡难以检测与分割的问题,基于深度学习框架,在YOLACT模型算法的基础上,将有效通道注意力机制(ECA)和空洞卷积引入特征提取网络,提出一种ECA-YOLACT气泡检测与分割算法,以增加重叠气泡的边缘提取能力.为获得气泡数据集,基于气泡生成对抗网络(BubGAN)生成数据集,同时开展气液两相流试验工作,采用数据增强的方式进行扩充,完成对网络模型训练.为验证算法的可行性,针对气泡的不同空隙率开展模型算法实验.实验结果表明:基于改进的YOLACT气泡检测与分割算法,在测试集上进行验证的准确率为89.49%,召回率为97.51%,平均准确率为96.80%.Aiming at the problem of difficult detection and segmentation of overlapping bubbles in gas-liquid two-phase flow,based on a deep learning framework,and on the basis of the YOLACT(you only look at coefficients)modeling algorithm,by introducing efficient channel attention(ECA)and void convolution into the feature extraction network,an ECA-YOLACT bubble detection and segmentation algorithm was proposed to increase the edge extraction ability of overlapping bubbles.To obtain the bubble dataset,based on bubble generation adversarial network(BubGAN),the dataset was generated and meanwhile the gas-liquid two-phase flow test work was carried out,which was expanded by using data enhancement to complete the training of the network model.To verify the feasibility of the algorithm,modeling algorithm experiments were carried out for different void fractions.Experimental results show that based on the improved YOLACT bubble detection and segmentation algorithm,which is validated on the test set,the accuracy is 89.49%,and the recall is 97.51%,with the average accuracy of 96.80%.

关 键 词：气液两相流 重叠气泡分割 注意力机制 空洞卷积 空隙率 数据增强 气泡生成对抗网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] U661.1[自动化与计算机技术—计算机科学与技术]