1、智能交叉科学与工程DOI:10.15961/j.jsuese.202200975基于深度学习的废钢分类评级方法研究肖鹏程1,2,3，徐文广1，常金宝2，朱立光3,4*，朱荣3，许云峰5(1.华北理工大学 冶金与能源学院，河北 唐山 063210；2.河钢集团有限公司，河北 石家庄 050000；3.北京科技大学 冶金与生态学院，北京 100083；4.河北科技大学 材料科学与工程学院，河北 石家庄 050000；5.河北科技大学 信息科学与工程学院，河北 石家庄 050000)摘要：废钢是现代钢铁工业重要的铁素来源，是钢企实现碳中和的重要原料。不同级别的废钢价格悬殊，其质量直接影响钢企的生产成
2、本和产品质量。因此，废钢入炉前的分类和评级问题，受到钢企的普遍重视和高度关注。针对传统人工方法在废钢的分类评级中所出现的效率低、安全性和公正性差等问题，基于深度学习中的卷积注意力机制和加权双向特征融合网络构建废钢分类评级模型。首先，搭建废钢质量查验物理模型，模拟货车卸载废钢的生产作业场景，采用高分辨率视觉传感器采集不同类别的废钢图像。其次，设计了一种结合注意力与特征融合的废钢验质深度学习模型，将卷积注意力模块（convolutional block attention module，CBAM）加入主干网络对采集的废钢图像数据集进行特征提取，聚焦并保留图像的重要特征；使用双向特征金字塔（bidi
3、rectional feature pyr-amid network，BiFPN）平衡多尺度特征信息，进行多尺度特征融合。最后，在模型预测阶段，利用所构建的废钢质量验质模型进行废钢类别和质量判级，验证模型的精确性与检测效率。基于自制废钢验证数据集，与主流的目标检测模型Faster RCNN、YOLOv4、YOLOv5系列以及YOLOv7进行性能比较。实验结果表明：本研究构建的废钢质量验质模型识别判级的准确率Acc达到了86.8%，所有类别平均精度mAP为89.2%，均高于对比的目标检测模型，在准确性、实时性以及识别评级效率方面可满足实际生产应用，解决废钢分类评级过程中的诸多难题，实现废钢的智能
4、验质和公正判定。关键词：再生钢铁原料；废钢智能评级；深度学习；注意力机制中图分类号：TP274+.5文献标志码：A文章编号：2096-3246（2023）02-0184-10Classification and Rating of Scrap Steel Based on Deep LearningXIAO Pengcheng1,2,3，XU Wenguang1，CHANG Jinbao2，ZHU Liguang3,4*，ZHU Rong3，XU Yunfeng5(1.College of Metallurgy and Energy,North China Univ.of Sci.and Te
5、chnol.Univ.,Tangshan 063210,China;2.HBIS Group Co.,Ltd.(HBIS),Shijiazhuang,050000,China;3.Metallurgical and Ecological Eng.School,Univ.of Sci.and Technol.Beijing,Beijing 100083,China;4.College of Materials Sci.and Eng.,Hebei Univ.of Sci.and Technol.,Shijiazhuang 050000,China;5.College of Info.Sci.an
6、d Eng.,Hebei Univ.of Sci.and Technol.,Shijiazhuang 050000,China)Abstract:Steel scrap is an important source of ferrite for the modern steel industry and an important raw material for steel companies to achievecarbon neutrality.The price of different grades of scrap varies greatly and its quality dir
7、ectly affects the production cost and product quality ofsteel enterprises.Therefore,the classification and grading of scrap before feeding into the furnace has received widespread attention and greatconcern from steel enterprises.To address the problems of low efficiency,poor safety,and fairness in
8、the classification and rating of scrap by tra-ditional manual methods,a scrap classification and rating model(CCBFNet)based on the spatial and channel attention mechanism and weightedbidirectional feature fusion network was proposed in the paper.Firstly,a physical model of scrap quality checking was
9、 built to simulate the pro-收稿日期:2022 09 09基金项目:国家自然科学基金项目（51904107）；河北省自然科学基金项目（E2020209005；E2021209094）；河北省高等学校科学技术研究项目（BJ2019041）；河北省“三三三人才工程”资助项目（A202102002）；唐山市人才资助重点项目（A202010004）作者简介:肖鹏程（1985），男，副教授，博士.研究方向：凝固理论、炼钢连铸工艺及钢铁智能制造技术.E-mail：*通信作者:朱立光，教授，E-mail：网络出版时间:2023 03 06 15:08:05 网络出版地址:https
10、:/ http:/http:/ 第 55 卷 第 2 期工 程 科 学 与 技 术Vol.55 No.22023 年 3 月ADVANCED ENGINEERING SCIENCESMar.2023duction operation scene of unloading scrap by trucks,and high-resolution vision sensors were used to collect the images of different types ofscrap.Secondly,a deep learning model combining attention and
11、 feature fusion was designed for scrap quality inspection in the model trainingstage,and the spatial and channel attention module(CBAM)was added to the backbone network to extract features from the collected scrap im-age dataset,focusing and retaining the important features of the images;then,a weig
12、hted Bidirectional Feature Pyramid Network(BFPN)wasused.Secondly,the multi-scale feature fusion was performed by balancing the multi-scale feature information using the Bidirectional Feature Pyr-amid Network(BiFPN).Finally,in the model prediction stage,the constructed scrap quality verification mode
13、l CCBFNet was used for scrap cat-egory and quality grading to verify the accuracy and detection efficiency of the model.Based on the homemade scrap validation dataset,the per-formance of CCBFNet was compared with the mainstream target detection Faster RCNN,YOLOv4,YOLOv5 series,and YOLOv7.The experi-
14、mental results showed that the Acc of CCBFNet reaches 86.8%and the mAP is 89.2%,which are higher than the compared target detection mod-els.The proposed CCBFNet can fully meet the actual production applications in terms of accuracy,real-time and recognition rating efficiency,solve many difficulties
15、in the process of scrap classification and rating,and realize the intelligent quality inspection and fair determination ofscrap.Key words:recycling ironsteel materials;scrap intelligent rating;deep learning;attention mechanism 废钢是一种可替代铁矿石的绿色资源，也是现代钢铁行业的重要原料12。目前钢铁生产主要有两种路线，即采用铁矿石与废钢为主要原料的高炉碱性氧炉（blas
16、t furnacebasic oxygen furnace，BFBOF）路线和采用废钢为主要原料的电弧炉（electric arcfurnace，EAF）路线35。根据世界钢铁协会数据，2021年全球粗钢产量为19.52 亿t，较上年增加3.8%，电炉钢产量同比增长14.4%，达5.63 亿t。其中，美国的钢铁产量约2/3来自废钢，废钢来源包括生产和成型废钢、加工废料和消费后废钢6。欧盟的钢铁生产份额约54%来自废钢78，日本20%以上的粗钢是由废钢制成9。尽管当前中国废钢占钢铁工业原料比例相对较少，但随着中国碳达峰、碳中和战略的进一步落实以及社会钢铁材料的逐步更新，中国将成为继美国、欧盟和日本之后废钢消费量最大的国家1011。废钢消费量与进口量的迅速持续增长使其回收验质方面的压力与日俱增。现今，大多数钢铁企业对废钢等级的判定主要由人工进行，存在着危险性高、评级精度低、公正性易受到质疑等诸多问题。不同级别的废钢价格差异较大，其质量直接影响钢企的生产成本和产品质量。因此，废钢的分类和评级成为钢企生产成本和产品质量控制的重要一环。目前，国内外对于废钢分类评级的人工智能算法研究较少，与之相似
