摘要
为了提高VD(vaccum degassing)精炼非真空过程中对吹氩等级检测的识别速度和准确性,提出1种改进YOLOv8n模型的VD精炼非真空过程中吹氩等级检测模型。首先,基于YOLOv8n模型,在主干网络层中引入坐标注意力机制(coordinate attention,CA),利用坐标注意力机制的特点使基础模型在识别精确度方面有所提升;进一步利用快速网络(FasterNet)内部的设计模式,使用其网络模型内部的FasterBlock模块替换基础模型主干网络和颈部网络中特征提取模块C2f内部的Bottleneck处理模块,以期增加原基础模型的训练速度和识别速度;其次,由于目前没有钢包精炼吹氩等级识别的相关训练权重,通过借助标注工具和视频转换工具将收集到的钢厂钢包精炼图像集进行标注与划分;然后,根据不同的训练权重结果开展不同改进模型和基础模型的对比试验和消融试验。试验结果表明,改进后YOLO8n模型识别精确度(precision)提高了11个百分点,平均精度均值(PmA)提升了6个百分点,每秒图像处理速度(FPS)提高了6.4帧/s,这说明改进后的模型在识别精确度和检测速度上均有显著提升。最后,将改进模型在VD精炼初期-破渣期-喂线期的图像集上进行大量识别检测试验,结果证明,改进后模型能够满足VD生产过程中的实时吹氩等级检测,这为VD精炼过程中非真空情况下规范化的流量控制提供了有效的智能化检测模型。
In order to improve the identification speed and accuracy of argon blowing grade detection in VD refining non-vacuum process,an improved YOLOv8n model for argon blowing grade detection in VD refining non-vacuum process was proposed.First,based on the YOLOv8n model,Coordinate Attention was introduced from the backbone network layer,and the characteristics of the attention mechanism was used to increase the recognition accuracy of the basic model.In order to increase the training speed and recognition speed of the origin model.Furthermore,the FasterBlock module inside the FasterNet was used to replace the Bottleneck processing module inside the feature extraction module C2f in the backbone network and neck network of the basic model,in order to increase the training speed and recognition speed of the original basic model.Secondly,since there was no relevant training weight for the recognition of argon blowing grade for ladle refining,the collected image set of ladle refining was marked and divided by annotation tools and video conversion tools.Then,according to the results of different training weights,the comparison test and ablation test of different improved models and basic models were carried out.The test results show that the improved YOLO8n model recognition accuracy(precision)is increased by 11 percent point,the average accuracy(PmA)is increased by 6 percent point,and the image processing speed per second(FPS)is increased by 6.4 f/s.This indicates that the improved model has significantly improved the recognition accuracy and speed.Finally,a large number of identification tests were carried out on the image set of the initial stage for VD refining-slag breaking stage-line feeding stage.The results show that the improved model can meet the real-time argon blowing grade detection in VD production process,and provide an effective intelligent detection model for standardized flow control under non-vacuum conditions in VD refining.
作者
孙大元
史利民
朱晓风
孔令种
臧喜民
杨杰
SUN Dayuan;SHI Limin;ZHU Xiaofeng;KONG Lingzhong;ZANG Ximin;YANG Jie(College of Materials and Metallurgy,University of Science and Technology Liaoning,Anshan 114051,Liaoning,China;Institute of Automation,Chinese Academy of Sciences,Beijing 100190,China;No.1 Steelmaking Plant,Jiangyin Xingcheng Special Steel Co.,Ltd.,Wuxi 214429,Jiangsu,China;Shenyang University of Technology,Shenyang 110870,Liaoning,China)
出处
《钢铁》
北大核心
2025年第8期94-103,共10页
Iron and Steel
基金
国家自然科学基金资助项目(U24A20100)。
