基于BS-TransUNet网络的高炉雷达料线分割方法  

Segmentation method of blast furnace radar burden lines based on BS-TransUNet network

作　　者：蒋博涵 陈先中[1,2] 侯庆文[1,3] 张洁 张森 JIANG Bohan;CHEN Xianzhong;HOU Qingwen;ZHANG Jie;ZHANG Sen(School of Automation and Electrical Engineering,University of Science and Technology Beijing,Beijing 100083,China;Key Laboratory of Knowledge Automation for Industrial Processes of Ministry of Education,University of Science and Technology Beijing,Beijing 100083,China;Beijing Engineering Research Center of Industrial Spectrum Imaging,Beijing 100083,China)

机构地区：[1]北京科技大学自动化学院,北京100083 [2]北京科技大学工业过程知识自动化教育部重点实验室,北京100083 [3]北京市工业波谱成像工程技术研究中心,北京100083

出　　处：《冶金自动化》2024年第2期50-59,共10页Metallurgical Industry Automation

基　　金：国家自然科学基金面上项目(61671054)。

摘　　要：高炉雷达料线提取目前普遍采用神经网络加能量重心法的两步提取料线法,存在网络模型和机理模型混合分步计算,易受特殊环境强噪声影响的问题。本文提出了一种改进的基于语义分割的高炉料线提取BS-TransUNet算法。首先,针对高炉料面周期性形态和粒度变化以及信噪比衰减问题,在卷积神经网络(convolution neural network, CNN)和Transformer模块之间引入了空洞空间金字塔池化(atrous spatial pyramid pooling, ASPP)模块,获得料面细粒度特征;然后,将坐标注意力(coordinate attention, CA)模块集成到每次上采样之后,更全面地滤除背景噪声,抑制对无效高频纹理特征的提取;最后,将跳跃链接替换为跳跃连接融合(BiFusion)模块,进一步提高分割性能。实验结果表明,改进的算法在高炉雷达料面数据集上,平均交并比(mean intersection over union, MIoU)和F1分数分别提高了1.77%和1.46%,类别平均像素准确率(mean pixel accuracy, MPA)提高了1.97%,其中F1分数可以达到86.18%。与传统的两步提取料线法相比,在高炉恶劣环境下采用端到端的分割料线一步法,料线获取的精度和稳定性均得到了改善。Blast furnace radar burden line extraction currently commonly used neural network plus energy center of gravity method of two-step extraction of burden line method,there is a mixture of network model and mechanism model step-by-step computation,susceptible to the influence of the special environment of strong noise problem.In this paper,an improved BS-TransUNet algorithm for blast furnace burden line extraction based on semantic segmentation was proposed.Firstly,to address the problems of periodic morphology and particle size variation of blast furnace burden surface and signal to noise ratio attenuation,the atrous spatial pyramid pooling(ASPP)module is introduced between convolution neural network(CNN)and Transformer modules to obtain fine-grained features of the burden surface.Then,the coordinate attention(CA)module is integrated after each up-sampling to filter out the background noise more comprehensively and inhibit the extraction of ineffective high-frequency texture features.Finally,the jump link is replaced with the BiFusion module to further improve the segmentation performance.The experimental results show that the improved algorithm improves the mean intersection over union(MIoU)and F1 scores by 1.77%and 1.46%,respectively,the mean pixel accuracy(MPA)by 1.97%on the blast furnace radar burden surface dataset,and the F1 score can reach 86.18%.Compared to the conventional two-step extraction of burden line method,the one-step method with end-to-end split burden line in the harsh environment of a blast furnace provides improved accuracy and stability of burden line acquisition.

关 键 词：高炉雷达 料线提取 语义分割 BS-TransUNet 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TN957.52[自动化与计算机技术—控制科学与工程] TF54[电子电信—信号与信息处理]