首页 | 本学科首页   官方微博 | 高级检索  
     检索      

基于卷积神经网络的钝体尾迹识别研究
引用本文:杜祥波,陈少强,侯靖尧,张帆,胡海豹,任峰.基于卷积神经网络的钝体尾迹识别研究[J].力学学报,2022,54(1):59-67.
作者姓名:杜祥波  陈少强  侯靖尧  张帆  胡海豹  任峰
作者单位:*.西北工业大学航海学院, 西安 710072
基金项目:国家自然科学基金(52071272,12102357);基础前沿(JCKY2018*18);陕西省自然科学基础研究计划(2020JC-18);中央高校基本科研业务费专项资金(3102021HHZY030002);河南省水下智能装备重点实验室开放基金(KL01B2101)资助项目。
摘    要:针对相同特征长度不同钝体的尾迹结构相近, 肉眼难于分辨的问题, 提出了一种基于卷积神经网络的钝体尾迹识别方法, 并在竖直肥皂膜水洞的典型钝体模型尾迹实验中获得高准确率验证. 实验平台由自建竖直肥皂膜实验装置、钝体模型(方柱、圆柱和三角柱)及图像采集系统组成, 可基于光学干涉法实现对不同速度下钝体肥皂膜尾迹的高清持续拍摄. 所建立卷积神经网络识别模型由输入层、卷积层、池化层、全连接层和分类层组成, 其中, 卷积层和池化层用于提取尾迹的深层次特征信息, 而全连接层和分类层构成识别分类模式来分类输出图像对应的钝体类型或雷诺数. 通过将9000张尾迹图像数据集导入卷积神经网络模型, 以数据驱动方式建立了具有钝体形状或雷诺数识别能力的尾迹特征识别模型. 结果表明, 该模型对相同雷诺数下识别钝体形状的准确率达97.6%(识别300张不同形状钝体尾迹图像), 对不同雷诺数下识别钝体形状的准确率达96%(识别1200张不同雷诺数尾迹图像), 即使将不同钝体形状和雷诺数下尾迹图像混放一起, 其钝体形状和雷诺数识别准确率也可以达到91%(识别1500张混放尾迹图像). 该方法为进一步利用人工智能提取流体尾迹中的物理信息提供借鉴. 

关 键 词:竖直肥皂膜水洞    卷积神经网络    钝体    尾迹识别
收稿时间:2021-08-20

WAKE RECOGNITION OF A BLUNT BODY BASED ON CONVOLUTIONAL NEURAL NETWORK
Du Xiangbo,Chen Shaoqiang,Hou Jingyao,Zhang Fan,Hu Haibao,Ren Feng.WAKE RECOGNITION OF A BLUNT BODY BASED ON CONVOLUTIONAL NEURAL NETWORK[J].chinese journal of theoretical and applied mechanics,2022,54(1):59-67.
Authors:Du Xiangbo  Chen Shaoqiang  Hou Jingyao  Zhang Fan  Hu Haibao  Ren Feng
Institution:*.School of Marine Science and Technology, Northwestern Polytechnical University, Xi'an 710072, China?.The 705 Research Institute of China State Shipbuilding Corporation Limited, Xi'an 710077, China**.Hydrodynamic Research Department, China Ship Scientific Research Center, Wuxi 214082, Jiangsu, China
Abstract:Wake structures of different blunt bodies with identical characteristic length are similar, this is quite challenging to be distinguished using solely human eyes. Here, we propose a blunt body wake recognition method based on the convolutional neural network(CNN), which is then verified to be highly accurate with various types of blunt bodies models in vertical soap-film water tunnel experiments. The experimental platform is composed of a self-built vertical soap-film device, three typical blunt body models(square cylinder, circular cylinder, and triangle cylinder), and an image acquisition system. Based on the optical interference method, this image processing modulus can realize continuous high-fidelity photography of blunt body wakes with different incoming velocities. The CNN recognition model is built up with input layer, convolutional layer, pooling layer, fully-connected layer, and classification layer.Among them, the convolutional layer and the pooling layer are used to extract the deep feature information of wakes,while the fully-connected layer and the classification layer together can finally determine the category or Reynolds numbers of the input wake image. By importing a data set with 9000 wake images into the CNN model, a wake feature recognition model capable of classifying various body shapes is established in a data-driven manner. Results show that the shape recognition accuracy is 97.6% at the same Reynolds number(300 wake images), and 96% at different Reynolds numbers(1200 wake images). Even when wake images with different shapes and Reynolds numbers are mixed together,the recognition accuracy in terms of both shape and Reynolds number can still reach 91%(1500 mixed wake images).The proposed method provides a solid reference for future applications of artificial intelligence in extracting physical information from blunt body wakes.
Keywords:vertical soap-film tunnel  convolutional neural network  blunt body  wake recognition
本文献已被 维普 等数据库收录!
点击此处可从《力学学报》浏览原始摘要信息
点击此处可从《力学学报》下载免费的PDF全文
设为首页 | 免责声明 | 关于勤云 | 加入收藏

Copyright©北京勤云科技发展有限公司  京ICP备09084417号