Chemical Industry and Engineering Progress ›› 2024, Vol. 43 ›› Issue (2): 688-695.DOI: 10.16085/j.issn.1000-6613.2023-1354

• Column: multiphase flow test • Previous Articles     Next Articles

Improved residual network based on attention mechanism for flame temperature field reconstruction

SHAN Liang1(), ZHOU Rongxing1, HONG Bo1, YANG Wenqi1, KONG Ming2()   

  1. 1.Key Laboratory of Electromagnetic Wave Information Technology and Metrology of Zhejiang Province, College of Information Engineering, China Jiliang University, Hangzhou 310018,Zhejiang,China
    2.College of Metrology and Measurement Engineering, China Jiliang University, Hangzhou 310018,Zhejiang,China
  • Received:2023-08-08 Revised:2023-11-10 Online:2024-03-07 Published:2024-02-25
  • Contact: KONG Ming

基于注意力机制的改进残差网络火焰温度场重建

单良1(), 周荣幸1, 洪波1, 仰文淇1, 孔明2()   

  1. 1.中国计量大学信息工程学院,浙江省电磁波信息技术与计量检测重点实验室,浙江 杭州 310018
    2.中国计量大学计量测试工程学院,浙江 杭州 310018
  • 通讯作者: 孔明
  • 作者简介:单良(1979—),女,教授,硕士生导师,研究方向为信号处理、光电检测。E-mail:lshan@cjlu.edu.cn
  • 基金资助:
    国家自然科学基金(51874264)

Abstract:

The method of reconstructing the flame temperature field based on convolutional neural network has been widely used in recent years, but the traditional convolutional neural network model is prone to overfitting or model degradation as the number of network layers increases, resulting in large reconstruction errors. This paper proposed an improved method, which used the ResNet18 network to reconstruct the flame temperature field, and introduced the attention mechanism and local importance pooling to optimize the extracted content, realized the full use of known information, and reduced the reconstruction error. The experimental results showed that after introducing the local importance pooling and attention mechanism at the same time, the average relative error of temperature field reconstruction was 0.13%, and the maximum relative error was 0.75%. Compared with the initial ResNet18 network, the average relative error was reduced by 31.58%. The maximum relative error was reduced by 34.21%. The influence of the two factors on the reconstruction accuracy was verified by ablation experiments. The results showed that the temperature field reconstruction accuracy after adding two improved modules at the same time was better than that after adding a single improved module, and the local importance pooling module had a significant effect on the accuracy improvement.

Key words: temperature field, residual network, attention mechanism, pooling

摘要:

基于卷积神经网络重建火焰温度场的方法近年来已被广泛采用,但是传统卷积神经网络模型随着其网络层数的增加极易出现过拟合或者模型退化的现象,导致重建误差较大。本文提出一种改进的方法,使用ResNet18网络进行火焰温度场重建,并引入注意力机制和局部重要性池化,优化提取内容,实现已知信息的充分利用,减少重建误差。实验结果表明,同时引入局部重要性池化和注意力机制后,温度场重建的平均相对误差为0.13%,最大相对误差为0.75%;相较于初始ResNet18网络,平均相对误差减少了31.58%,最大相对误差减少了34.21%。通过消融实验验证了两种因素对重建精度的影响,结果表明:同时加入两个改进模块后的温度场重建精度要优于加入单个改进模块后的精度,局部重要性池化模块对精度提升的作用更大。

关键词: 温度场, 残差网络, 注意力机制, 池化

CLC Number: 

京ICP备12046843号-2;京公网安备 11010102001994号
Copyright © Chemical Industry and Engineering Progress, All Rights Reserved.
E-mail: hgjz@cip.com.cn
Powered by Beijing Magtech Co. Ltd