摘要
针对某炼钢厂A号转炉炉膛火焰燃烧状态分析判定问题,提出基于改进YOLOv7深度学习网络架构结合视频火焰状态的判断方法。首先引入无参平均注意力机制(PfAAM),在不添加参数或超参数的情况下捕获通道注意力和空间注意力,并使用轻量化上采样算子(CARAFE)、功能的内容感知重组进行上采样,使YOLOv7网络获得更好的鲁棒性和准确性。优化后的网络,MAP_0.5值提升了2.4%,精确率提升了5.5%,帧率达到了51.5帧/秒。结合视频图像Matlab火焰状态分析方法,充分利用图像和特征信息增加了对火焰状态判断的可解释性和鲁棒性,并提高了算法的安全性和可信度。实验结果也表明改进算法能够有效提升对火焰燃烧状态识别的准确率和实时性。
Aiming at analyzing and judging flame combustion of No.A converter furnace in a steel mill,a judgment method based on improved YOLOv7 deep learning network architecture and video flame state was proposed.Firstly,having a parameter-free average attention mechanism(PfAAM)introduced to capture channel attention and spatial attention without adding parameters or hyperparameters,including having a lightweight up-sampling operator(CARAFE)and functional content-aware recombination adopted to upsample the YOLOv7 network so as to achieve better robustness and accuracy.In the optimized network,the MAP_0.5 value can be increased by 2.4%,the precision value increased by 5.5%,and the frame rate reached 51.5 frames per second.Through combined with video image Matlab flame state analysis method,both interpretability and robustness of the flame state judgment can be increased by making full use of image and feature information,and both security and reliability of the algorithm were improved.The experimental results also show that the improved algorithm can effectively improve both accuracy rating and real time recognition of flame combustion state.
作者
段志伟
邵女
豆全辉
徐武
DUAN Zhi-wei;SHAO Nv;DOU Quan-hui;XU Wu(School of Physics and Electronic Engineering,Northeast Petroleum University;Angang Construction Group Co.,Ltd)
出处
《化工自动化及仪表》
CAS
2024年第3期388-395,共8页
Control and Instruments in Chemical Industry
基金
国家自然科学基金(批准号:51474069)资助的课题。
关键词
深度学习
燃烧状态
火焰区域
注意力机制
转炉火焰
特征信息
火焰燃烧视频
MATLAB
sdeep learning
combustion state
flame region
attention mechanism
converter flame
characteristic information
video of flame combustion
Matlab