基于深度学习的连铸坯低倍质量评级  

Evaluation on macro quality of strand based on deep learning method

作　　者：宋翰凌 孟晓亮 罗森 王卫领 朱苗勇 SONG Hanling;MENG Xiaoliang;LUO Sen;WANG Weiling;ZHU Miaoyong(School of Metallurgy,Northeastern University,Shenyang 110819,China)

机构地区：[1]东北大学冶金学院,辽宁沈阳110819

出　　处：《冶金自动化》2023年第2期73-81,共9页Metallurgical Industry Automation

基　　金：国家重点研发计划项目(2021YFB3702000)。

摘　　要：连铸坯低倍质量检测是评价连铸坯质量的重要手段,正广泛应用于连铸生产过程。然而,钢厂大多采用人工方法对连铸坯低倍质量进行评价。这种人工方法依赖于评级人的经验,缺乏检验的一致性、客观性及准确性。为了准确地对连铸坯凝固组织及中心偏析评级,以U-Net网络为基础,集成了残差模块、金字塔池化模块(pyramid pooling module,PPM)以及注意力模块,提出了APR-UNet模型。模型中,残差模块可避免深层网络出现退化问题;PPM可聚合不同区域的上下文信息,加强模型感受野,以提高网络获取全局信息的能力;注意力机制可抑制无用信息,提高模型对铸坯凝固组织及缺陷的分割精度及模型鲁棒性。使用相同数据集分别训练U-Net模型和APR-UNet模型。试验表明,对连铸坯等轴晶区分割,APR-UNet模型的交并比(inter-section over union,IOU)达93.54%,较U-Net模型提高了1.06%;对中心偏析的评级,APR-UNet模型的评级成功率达91.2%,较U-Net模型提高了3.6%。APR-UNet模型有效改善了原模型分割结果中出现的过分割现象,在连铸坯凝固组织及缺陷的评级方面具有很大潜力。Macro quality inspection of continuous casting strand is an important means to evaluate the quality of continuous casting strand,which is widely used in continuous casting process.However,most steel mills use manual method to evaluate the macro quality of continuous casting strand.This artificial method relies on the experience of the raters and lacks the consistency,objectivity and accuracy of the inspection.In order to accurately rate solidification structure and central segregation of continuous casting strand,an APR-UNet model was proposed based on U-Net,which integrates residual module,pyramid pooling module(PPM)and attention module.In this model,the residual module can avoid the degradation of the deep network.The PPM can aggregate the context information of different regions,strengthen the receptive field of the model,and improve the ability of the network to obtain global information.The attention module can suppress useless information and improve the segmentation accuracy and robustness of the model for solidification structure and defects of casting strand.The same dataset was used to train the U-Net model and the APR-UNet model respectively.The results show that the intersection over union(IOU)of APR-UNet model is 93.54%for the equiaxed crystal segmentation of continuous casting strand,which increases 1.06%than U-Net.For the rating of central segregation,the rating success rate of APR-UNet model is 91.2%,which increases 3.6%than U-Net.The APR-UNet model can effectively improve the over-segmentation phenomenon in the segmentation results of the original model,and has great potential in the solidification structure and defect rating of continuous casting strand.

关 键 词：数字化 深度学习 铸坯质量评级 低倍检验 U-Net APR-UNet 

分 类 号：TF777[冶金工程—钢铁冶金] TP18[自动化与计算机技术—控制理论与控制工程]