基于改进YOLOv8的地铁列车焊缝缺陷轻量化检测方法
Lightweight detection method for weld defects of metro train based on improved YOLOv8
李先旺 1贺岁球 1贺德强 1孙海猛 1吴金鑫 1单晟2
作者信息
- 1. 广西大学 机械工程学院,广西 南宁 530004
- 2. 中车株洲电力机车研究所有限公司,湖南株洲 412001
- 折叠
摘要
针对现有的地铁列车车体焊接质量检测技术存在检测模型较大、检测精度和效率较低的问题,提出一种基于改进YOLOv8的焊缝缺陷轻量化检测方法.首先,利用相控阵超声波检测仪采集对接焊缝内部缺陷图像,通过图像预处理制作成焊缝缺陷数据集.然后,在YOLOv8模型基础上,利用Inner-SIoU优化原有损失函数、采用C2f-PConv替换C2f模块、引入大型可分离核注意力(LSKA)模块和挤压激励(SE)注意力机制,建立了基于改进YOLOv8的地铁列车车体焊缝缺陷质量检测模型,以提高焊缝缺陷特征提取和多尺度特征融合的能力.最后,利用改进的YOLOv8模型对焊缝缺陷数据集进行训练和测试.结果表明,改进的YOLOv8模型大小为7.91 M,对于焊缝缺陷的检测精度达到98.30％,检测速度达到138.9帧/s,与YOLOv8原始模型相比,模型更小,检测精度更高.
Abstract
To address the issues of the current metro train body welding quality inspection technology,such as the large size of the detection model and low detection accuracy and efficiency,a lightweight detection method of weld defects based on improved YOLOv8 was proposed.Firstly,images of internal defects in butt welds were collected by using a phased array ultrasonic detector,and a welding defect dataset was created through image preprocessing.Then,based on the YOLOv8 model,the original loss function is optimized using Inner-SIoU,the C2f module was replaced with C2f-PConv,and the LSKA module and SE attention mechanism were introduced to establish a defect quality detection model for metro train body welding seams based on improved YOLOv8.The proposed model could enhance the capability of extracting features from weld seam defects and facilitating multi-scale feature fusion.Finally,the improved YOLOv8 model was trained and tested on the weld defect data set.Experimental results show that the size of the improved YOLOv8 model was 7.91 M,the detection accuracy of weld defects reaches 98.30％,and the detection speed reaches 138.9 frames per selend.Compared with the original YOLOv8 model,the model is smaller and has higher detection accuracy.
关键词
地铁列车/焊缝缺陷检测/YOLOv8/轻量化/相控阵超声波检测Key words
metro train/weld defect detection/YOLOv8/lightweight/phased array ultrasonic testing引用本文复制引用
基金项目
国家自然科学基金联合重点项目(U22A2053)
广西创新驱动发展专项项目(桂科AA23073019)
出版年
2024
NETL
NSTL
万方数据