基于深度学习的废钢分类评级方法研究

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,提出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CCBFNet。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,在模型训练阶段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;使用加权双向特征金字塔（Bidirectional Feature Pyramid 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CCBFNet进行废钢类别和质量判级,验证模型的准确性与检测效率。基于自制废钢验证数据集,与主流的目标检测Faster R-CNN、YOLOv4、YOLOv5系列以及YOLOv7进行性能比较。实验结果表明：CCBFNet识别判级的平均准确度达到了90%,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完全满足实际生产应用,解决废钢分类评级过程中的诸多难题,实现废钢的智能验质和公正判定。

Classification and Rating of Scrap Steel Based on Deep Learning

Steel scrap is an important source of ferrite for the modern steel industry and an important raw material for steel companies to achieve carbon neutrality. The price of different grades of scrap varies greatly and its quality directly affects the production cost and product quality of steel enterprises, therefore, the classification and grading of scrap before feeding into the furnace has received widespread attention and great concern from steel enterprises. To address the problems of low efficiency, poor safety, and fairness in the classification and rating of scrap by traditional manual methods, CCBFNet, a scrap classification and rating model based on spatial and channel attention mechanisms in deep learning and weighted bidirectional feature fusion network, is proposed. Firstly, a physical model of scrap quality checking is built to simulate the production operation scene of unloading scrap by trucks, and high-resolution vision sensors are used to collect The images of different types of scrap are collected using high-resolution vision sensors. Secondly, a deep learning model combining attention and feature fusion is designed for scrap quality inspection in the model training stage, and the spatial and channel attention module (CBAM) is added to the backbone network to extract features from the collected scrap image dataset, focusing and retaining the important features of the images; secondly, a weighted Bidirectional Feature Pyramid Network (BFPN) is used. Secondly, the multi-scale feature fusion is performed by balancing the multi-scale feature information using the Bidirectional Feature Pyramid Network (BiFPN). Finally, in the model prediction stage, the constructed scrap quality verification model CCBFNet is used for scrap category and quality grading to verify the accuracy and detection efficiency of the model. Based on the homemade scrap validation dataset, the performance is compared with the mainstream target detection Faster R-CNN, YOLOv4, YOLOv5 series, and YOLOv7. The experimental results show that the average accuracy of CCBFNet recognition rating reaches 90% and the mAP is 89.2%, which are higher than the compared target detection models, and can fully meet the actual production applications in terms of accuracy, real-time and recognition rating efficiency, solve many difficulties in the process of scrap classification and rating, and realize the intelligent quality inspection and fair determination of scrap.