多尺度卷积神经网络的电阻层析成像算法  

MULTI-SCALE CONVOLUTIONAL NEURAL NETWORK ALGORITHM FOR ELECTRICAL RESISTANCE TOMOGRAPHY

作　　者：仝卫国[1] 曾世超 张立峰[1] Tong Weiguo;Zeng Shichao;Zhang Lifeng(Department of Automation,North China Electric Power University,Baoding 071003,Hebei,China)

机构地区：[1]华北电力大学自动化系,河北保定071003

出　　处：《计算机应用与软件》2024年第5期177-182,共6页Computer Applications and Software

基　　金：国家自然科学基金项目(61773160)。

摘　　要：针对电阻层析成像(ERT)经典算法(LBP、Landweber等)在复杂流型情况下成像精度低的问题,提出一种基于多尺度卷积神经网络(Multi-scale Convolutional Neural Network, MS-CNN)的电阻层析成像图像重建算法。根据气液两相流流型特点,构建有限元模型得到20 000组包含“边界电压向量-电导率分布”的数据集。在典型卷积神经网络Resnet50和Vgg16的基础上,构建针对ERT图像重建问题的MS-CNN。仿真实验结果表明,与Landweber迭代算法和单尺度卷积神经网络算法相比,MS-CNN算法的ICC分别提升了0.715和0.023,RIE分别降低了0.812和0.057。抗噪性测试与静态测试结果表明,MS-CNN算法具有良好的图像重建结果和鲁棒性。Aimed at the problem of low imaging accuracy of classical algorithms(LBP,Landweber,etc.)for electrical resistance tomography(ERT)in complex flow patterns,an image reconstruction algorithm based on multiscale convolutional neural network(MS-CNN)for electrical resistance tomography is proposed.According to the characteristics of gasliquid twophase flow pattern,a finite element model was built to obtain 20,000 data sets containing"boundary voltage vectorconductivity distribution".On the basis of typical convolutional neural networks Resnet50 and VGG16,MS-CNN for ERT image reconstruction was constructed.The simulation results show that compared with Landweber iterative algorithm and singlescale convolutional neural network algorithm,the ICC of MS-CNN algorithm is improved by 0.715 and 0.023,and the RIE is decreased by 0.812 and 0.057 respectively.The antinoise test and static test results show that the MS-CNN algorithm has good image reconstruction results and robustness.

关 键 词：卷积神经网络 计量学 电阻层析成像 Landweber 电导率分布 

分 类 号：TP274[自动化与计算机技术—检测技术与自动化装置]