摘要
大丝束碳纤维生产过程中的人为操作失误以及工艺缺陷会导致大丝束表面出现长短丝、毛球、接头、滞浆等缺陷,这些缺陷会影响产品质量,严重的甚至会使得生产过程存在安全隐患,而生产现场视觉系统只具备缺陷检出和图片保存能力,没有实现缺陷的识别分类。因此,从大丝束碳纤维表面缺陷检测的实际应用场景出发,提出一种基于YOLOv5s的轻量化丝束表面缺陷检测算法:首先,将MobileNetV2引入原模型主干网络进行轻量化改进,并插入ODConv动态卷积模块提升模型性能;其次,使用Dyhead动态检测头结构替换原有检测头,在不过度增加计算量的同时提升模型性能;接下来,引入CARAFE上采样算子替换原有最邻近上采样操作,提升模型特征信息聚合效果;最后,在自制大丝束碳纤维表面缺陷数据集上进行消融实验及对比实验。实验结果表明,所提轻量化改进算法相比其他3种经典模型具有更高的运行速度和检测精度,为解决大丝束碳纤维表面缺陷检测问题的研究提供了一种新的方法和思路。
Human operation errors and process defects in the production of large tow carbon fiber will lead to defects such as long and short filaments,hairballs,joints and stuck pulp on the surface of large tow,thereby affecting product quality and even causing security risks in the production process.The on-site vision system,however,is merely capable of detecting defects and saving pictures without realizing the classification and positioning of defects.Therefore,based on the actual application scenario of surface defect detection of large tow carbon fibers,this paper proposed a lightweight tow surface defect detection algorithm based on YOLOv5s.Firstly,MobileNetV2 was introduced into the backbone network of the original model for lightweight improvement,and ODConv dynamic convolution module was inserted to improve the model performance.Secondly,the Dyhead dynamic detection head structure was used to replace the original detection head,which could improve the model performance without excessive increase in computation.Next,CARAFE upsampling operator was introduced to replace the original nearest upsampling operation to improve the aggregation effect of model feature information.Finally,the ablation experiment and comparison experiment were carried out on the dataset of the self-made large tow carbon fiber surface defect.The experimental results show that the improved lightweight algorithm proposed in this paper has higher running speed and detection accuracy than that of the other three classical models,providing a new method and idea for solving the surface defect detection problem of large tow carbon fibers.
作者
李卓尘
刘新
李敏
白华
LI Zhuo-chen;LIU Xin;LI Min;BAI Hua(Beijing Research Institute of Automation for Machinery Industry Co.,Ltd.,Beijing 100120)
出处
《制造业自动化》
2025年第6期114-125,共12页
Manufacturing Automation
