基于深度学习的电磁离合器皮带轮表面微小缺陷检测方法  

Detection Method for Surface Micro Defects of Belt Wheel of Electromagnetic Clutch Based on Deep Learning

作　　者：丁建雄 万延见 柯海森 张砀砀 叶建甬 DING Jianxiong;WAN Yanjian;KE Haisen;ZHANG Dangdang;YE Jianyong(College of Mechanical and Electrical Engineering,China Jiliang University,Hangzhou Zhejiang 310018,China)

机构地区：[1]中国计量大学机电工程学院,浙江杭州310018

出　　处：《传感技术学报》2024年第7期1154-1161,共8页Chinese Journal of Sensors and Actuators

摘　　要：针对电磁离合器皮带轮表面缺陷形态不一,尺寸小所带来的检测效率低精度差等问题,通过改进YOLOv4提出Pulley-YOLOv4模型进行检测。提出了改进空间金字塔池化模块,能够更有效地提取微小缺陷特征且更加准确地定位。在颈部网络与主干特征提取网络之间添加全局注意力机制,使模型能够抑制不重要特征,关注感兴趣的目标。实验表明,Pulley-YOLOv4模型平均精度均值为98.54%,FPS指数为15,综合实际生产速度和精度的要求,所提出的Pulley-YOLOv4在皮带轮表面微小缺陷检测方面具有明显优势并且满足实时性要求。Targeting at the problems of low detection efficiency and poor accuracy caused by different shapes and small size of surface defects of electromagnetic clutch pulleys,the Pulley-YOLOv4 model is proposed by improving YOLOv4 for detection.An improved spatial pyramid pooling module is proposed,which can extract small defect features more effectively and locate them more accurately.A global attention mechanism is added between the neck network and the backbone feature extraction network so that the model can suppress unimportant features and focus on interesting targets.Experiments show that the average accuracy of the Pulley-YOLOv4 model is 98.54%,and the FPS index is 15.Combined with the requirements of actual production speed and accuracy,the Pulley-YOLOv4 proposed has obvious advantages in the detection of small defects on the surface of the pulley and meets the real-time requirements.

关 键 词：皮带轮 缺陷检测 深度学习 YOLOv4 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]