汽车后雨刮装配中支撑塑料帽在线缺陷检测  

Online Defect Detection of Support Plastic Caps in Automotive Rear Wiper Assembly

作　　者：陈淳 陈燚涛 柳雄 宋志峰 刘旺 CHEN Chun;CHEN Yi-tao;LIU Xiong;SONG Zhi-feng;LIU Wang(School of Mechanical Engineering and Automation,Wuhan Textile University,Wuhan Hubei 430200,China)

机构地区：[1]武汉纺织大学机械工程与自动化学院,湖北武汉430200

出　　处：《武汉纺织大学学报》2022年第2期16-20,共5页Journal of Wuhan Textile University

摘　　要：针对支撑塑料帽这种薄壁、小型零件在汽车后雨刮器齿轮连杆机构自动装配中存在裂纹、破损及漏装等多类型缺陷,导致人工很难在线识别、传统机器视觉辨识度不高的问题,本文在基于YOLO v3网络多目标缺陷检测算法的基础上,采用k-means聚类方法,使用平均重叠度重新选取目标区域候选边界框尺寸加快了目标检测识别速度及精度;同时使用Mish激活函数替换YOLO v3网络普遍采用的Leaky ReLU激活函数,强化了网络学习能力。实验结果表明,相较于传统的YOLO v3算法,本文改进的YOLO v3具有更高的检测精度和检测速度。该在线缺陷检测方法成功应用在实际汽车后雨刮自动装配线中,满足了支撑塑料帽在线缺陷检测的速度和精度要求,有效地提高了汽车后雨刮器齿轮连杆机构总成装配质量。To address the problem that it is difficult to manually identify multiple types of defects such as cracks, breakage and leakage in the automatic assembly of the rear wiper gear linkage mechanism of automobiles, and that the traditional machine vision recognition is not high,this paper adopts the k-means clustering method based on the multi-target defect detection algorithm of YOLO v3 network, and re-selects the target area using the average overlap. The candidate bounding box size speeds up the target detection and recognition speed and accuracy;meanwhile, the Mish activation function is used to replace the Leaky ReLU activation function commonly used in YOLO v3 network to strengthen the network learning ability. The experimental results show that the improved YOLO v3 in this paper has higher detection accuracy and detection speed compared with the traditional YOLO v3 algorithm. The online defect detection method is successfully applied in the actual automotive rear wiper automatic assembly line, which meets the speed and accuracy requirements of online defect detection of the support plastic cap and effectively improves the assembly quality of the automotive rear wiper gear linkage mechanism.

关 键 词：机器视觉 目标检测 卷积神经网络 YOLO v3 

分 类 号：TP368.1[自动化与计算机技术—计算机系统结构]