石英坩埚内壁缺陷检测平台搭建与算法研究  

CONSTRUCTION OF PLATFORM FOR DEFECT DETECTION AND ALGORITHM RESEARCH ON INNER WALL OF QUARTZ CRUCIBLES

作　　者：赵谦[1] 许东巍 缪正丽 郑轩 赵曼 Zhao Qian;Xu Dongwei;Miao Zhengli;Zheng Xuan;Zhao Man(School of Communication and Information Engineering,Xi’an University of Science and Technology,Xi’an 710054,China;Xi’an Dishan Vision Technology Limited Company,Xi’an 712044,China)

机构地区：[1]西安科技大学通信与信息工程学院,西安710054 [2]西安地山视聚科技有限公司,西安712044

出　　处：《太阳能学报》2025年第3期421-427,共7页Acta Energiae Solaris Sinica

基　　金：陕西省科技厅工业攻关(2022GY-115);陕西省教育厅服务地方企业(22JC050);陕西省科学技术协会青年人才托举计划(XXJS202206);陕西省科技成果转化计划(2023-YD-CGZH-29)。

摘　　要：目前石英坩埚缺陷主要使用人工目检,仅靠人工无法完成准确的分类以及全量计数。该文通过使用六自由度机械臂、旋转台搭建一套石英坩埚缺陷检测平台,结合背光源、高速相机等设备获取高清晰度的坩埚缺陷图像。同时提出改进的YOLOv5s坩埚缺陷检测算法,可识别杂质黑点、气泡、白斑等多种类型缺陷。具体而言,该算法首先应用K-均值聚类技术自适应生成最适合坩埚缺陷数据集的锚框;随后增加针对微小缺陷的检测层,以提高对小目标的识别能力;最后引入全维动态卷积(ODConv)和高效通道注意力机制(ECA),优化模型对关键目标区域的关注度,同时保持较低的计算开销。实验结果表明,在自建的石英坩埚缺陷数据集中,提出的改进算法mAPa0.5为98.88%,检测速度达到138帧/s,可达到工业检测要求。At present,the defects of quartz crucible mainly use manual eye inspection,which can not complete the accurate classification and full count only by manual.In this paper,a set of quartz crucible defect detection platform was built by using a 6-DOF manipulator and a rotating table,and high resolution images of crucible defects were obtained by combining backlight source and high speed camera.Simultaneously,the paper proposes an improved YOLOv5 algorithm for detecting crucible defects,capable of identifying various defect types such as impurity black spots,bubbles,and white spots.Firstly,K means clustering algorithm is used to generate anchor box suitable for crucible defect data sets,and then small target detection layer is added to improve the detection effect of small targets.Finally,omni-Dimensional dynamic convolution(ODConv)and the efficient channel attention(ECA)are used to make the network pay more attention to the targets to be detected without increasing too much computation.The experimental results show that in the self-built quartz crucible defect data set,the improved algorithm mAP@0.5 is 98.88%and the detection speed reaches 138 fps,which can meet the requirements of industrial detection.

关 键 词：单晶硅 坩埚 缺陷 深度学习 YOLOv5s 小目标检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TN307[自动化与计算机技术—计算机科学与技术]