基于注意力的高炉流态化料面多尺度检测算法

Multi-scale detection algorithm for fluidized burden surface of blast furnace based on attention

中文关键词  流态化  特征提取  注意力机制  多尺度卷积核  料面

英文关键词  fluidization  feature extraction  attention mechanism  multi-scale convolutional kernel  burden surface

中文摘要

      针对高炉料面图像经常发生多物理形态周期转变导致料线追踪精度下降问题, 研究了一种基于注意力的多尺度卷积核流态化料面检测算法(MKAD). 构建了雷达数据集–灰度图像–料形可视化的一类特征提取框架, 在卷积层采用通道和空间双注意力机制, 获得不同尺度的精细化颗粒流态化特征; 使用多尺度卷积核自适应方法提取 并融合喷涌料面多尺度颗粒物特征, 实现跨通道特征融合. 在南钢3#高炉和武钢7#高炉进行了实验和综合评估, 精确率分别可达83.01% 和86.50%, 与峰脊锐化方法相比, 分别实现了1.41%和4.9%的性能提升, 上述融合特征提取框架显著增强了料面检测的鲁棒性.

英文摘要

      The phenomenon of multi-physical morphology periodic transformation often occurs in the blast furnace(BF), which leads to the decrease of tracking accuracy of stock line. Aiming at this situation, a multi-scale convolutional kernel fluidized burden surface detection algorithm based on attention is proposed. A kind of feature extraction framework of radar dataset-gray image-burden shape visualization model is constructed, and the channel and spatial dual attention mechanism is used in the convolutional layer to obtain fine particle fluidized characteristics at different scales. The multi-scale convolution kernel adaptive method is used to extract and fuse the feature of multi-scale particles on the gushing burden surface, which realizes the cross-channel feature fusion. Experiments and comprehensive evaluations were carried out in No.3 BF of Nanjing iron and steel and No.7 BF of Wuhan iron and steel. The accuracy rates were 83.01% and 86.50%, respectively. Compared with peak and ridge sharpening method, the performance was improved 1.41% and 4.9%, respectively. The above fusion feature extraction framework significantly enhanced the robustness of burden surface detection.