基于自适应最优核和卷积神经网络的气液两相流流型识别方法  

Flow pattern recognition method of gas-liquid two-phase flow based on adaptive optimal kernel and convolution neural network

作　　者：翁润滢 孙斌 赵玉晓 张竟月 文英杰 WENG Runying;SUN Bin;ZHAO Yuxiao;ZHANG Jingyue;WEN Yingjie(School of Measurement and Testing Engineering,China Jiliang University,Hangzhou310018,Zhejiang,China;College of Modern Science and Technology,China Jiliang University,Hangzhou310018,Zhejiang,China;National Institute of Metrology,China,Beijing100029,China;Hangzhou Institute of Calibration and Testing for Quality and Technical Supervision,Hangzhou310019,Zhejiang,China)

机构地区：[1]中国计量大学计量测试工程学院,浙江杭州310018 [2]中国计量大学现代科技学院,浙江杭州310018 [3]中国计量科学研究院,北京100029 [4]杭州市质量技术监督检测院,浙江杭州310019

出　　处：《化工学报》2018年第12期5065-5072,共8页CIESC Journal

基　　金：国家自然科学基金项目(51475440)~~

摘　　要：为了研究气液两相流的动态特性,以及解决提取的特征值少而没有代表性导致识别率不高的传统问题,利用V锥流量计和动态差压传感器获取气液两相流在不同流型下的波动信号,采用自适应最优核算法对获取的动态信号进行时频分析,把一维时域信号转换为三维的时频谱图,能够清晰描述出管道内气液两相流的流动状态。将不同流型的时频谱图通过卷积神经网络(CNN)进行学习并自动提取相应的特征值,然后使用Softmax分类器进行训练从而实现流型识别。通过对几种常见流型进行试验与分析发现,采用时频谱图结合卷积神经网络的深度学习方法识别气液两相流流型,克服了传统流型识别方法特征值提取的不足之处,能够更贴切地描述气液两相流的动态特征。此方法可以进一步研究更多种类的流型以及空隙率等。To study the dynamic characteristics of gas-liquid two-phase flow,and to solve the traditional problem that the eigenvalues extracted are so few that is not representative and result in low recognition rate,V-cone flowmeter and dynamic differential pressure sensor were used to obtain fluctuating signals of gas-liquid two-phase flow in different flow patterns,time-frequency analysis of obtained dynamic signals were performed by adaptive optimal kernel(AOK)algorithm,to convert the one-dimensional time-domain signals to three-dimensional time-frequency spectra so as to describe clearly the flow status of gas-liquid two-phase flow in the pipeline.Time-frequency spectra of different flow patterns were learned by the convolution neural network(CNN)to extract the corresponding eigenvalue automatically,then were practiced by Softmax classifier to achieve flow pattern recognition.By experiment and analysis for several common flow patterns,adopting time-frequency spectrum and deep learning method of convolution neural network was found to identify the gas-liquid two-phase flow pattern can overcome the shortcomings of the traditional flow pattern recognition method for few eigenvalue extracted,and describe appropriately the dynamic characteristics of gas-liquid two-phase flow.This method allows further investigation of more types of flow patterns as well as porosity and more.

关 键 词：气液两相流 流型识别 算法 时频分析 神经网络 

分 类 号：TP274.2[自动化与计算机技术—检测技术与自动化装置]