基于多任务级联的动车裙板螺栓缺陷检测算法  

Multi-task cascading-based defect detection method for skirt bolts of EMU

作　　者：徐文辉[1,2] 钟胜[1,2] 邹旭[1,2] 何顶新[1] XU Wenhui;ZHONG Sheng;ZOU Xu;HE Dingxin(School of Artificial Intelligence and Automation,Huazhong University of Science and Technology,Wuhan 430074,China;National Key Laboratory of Multispectral Information Intelligent Processing Technology,Huazhong University of Science and Technology,Wuhan 430074,China)

机构地区：[1]华中科技大学人工智能与自动化学院,湖北武汉430074 [2]华中科技大学多谱信息智能处理技术全国重点实验室,湖北武汉430074

出　　处：《实验技术与管理》2023年第10期63-69,共7页Experimental Technology and Management

基　　金：国家自然科学基金项目(62176100)。

摘　　要：为了提高动车检修效率和准确性,设计了一种基于多任务级联的动车裙板螺栓缺陷检测算法。首先结合螺栓缺陷特征的先验知识,在YOLOv3的基础上引入注意力机制,采用通道级拼接方式引入螺栓的边缘特征图,引导检测网络学习鲁棒的螺栓缺陷特征,检测螺栓是否缺失;然后对螺栓局部区域进行语义分割,获得防松标记线信息,并基于这些信息判断是否存在螺栓松动和标记线缺失等缺陷。实验结果表明,该检测算法显著提升了动车裙板螺栓缺陷的检测性能,与YOLOv3相比,平均准确率提升11.3%,平均召回率提升13.6%。To improve the efficiency and accuracy of EMU Inspection and Repairment,a multi-task cascadingbased defect detection method for skirt bolts is proposed.Firstly,based on the prior knowledge of bolt defect features,an attention mechanism is introduced on YOLOv3 by injecting edge feature maps through channel splicing,which leads the detection network to learn robust bolt defect features and detect whether bolts are missing.Then,semantic segmentation is applied to the local areas of the bolts,and the marked lines are detected and fitted.With the information on the marked lines,the other defects(including loose bolts and missing marked lines)can be detected.The experimental results show that the detection method significantly improves the performance of defect detection for the skirt bolts of EMU.Compared to YOLOv3,the average accuracy is improved by 11.3%,and the average recall is improved by 13.6%.

关 键 词：动车检修 螺栓缺陷检测 多任务级联 注意力引导 YOLO 

分 类 号：TP274.5[自动化与计算机技术—检测技术与自动化装置]