摘要
为实现快速准确地检测轮对踏面缺陷,针对轮对踏面噪声干扰大、传统检测算法特征融合不充分的问题,提出一种基于残差注意力的YOLO-v5列车轮对踏面缺陷快速检测方法。首先,针对噪声干扰大的问题,设计了一个残差注意力降噪模块,以有效提升模型检测准确率,并使用Grad-CAM类激活映射技术验证残差注意力模块降低噪声干扰的作用;其次,针对特征融合不充分和模型容易产生漏检的问题,使用一种双向特征金字塔特征融合模块,对主干网络提取的特征进行高效融合,从而有效地降低检测漏检率;最后,采集了数百幅轮对踏面真实缺陷图像,并与5种经典检测模型进行对比,验证了算法的优越性。试验结果表明,该算法能够达到77.9%的准确率和72.3%的召回率,同时所提算法的图像检测速度能达到125幅/秒,模型权重仅为15.1 MB。该模型能快速准确地检测出剥离和凹陷2种缺陷,可便捷地应用于实际的轮对踏面实时缺陷检测场景。
In respect of large noise interference of wheel tread and insufficient feature fusion of traditional detection algorithms,in order to achieve fast and accurate detection of wheel tread defects,a method for rapid detection of wheel tread defects for YOLO-v5trains based on residual attention was proposed.First,for the large noise interference,a residual attention noise reduction module was designed to effectively improve the accuracy of model detection,and the Grad-CAM activation mapping technology was used to verify the effect of the residual attention module on reducing noise interference.Secondly,in view of insufficient feature fusion and the model being prone to omission,a feature fusion module of bidirectional feature pyramid was used to efficiently fuse the features extracted by the backbone network,thereby effectively reducing the omission rate.Finally,hundreds of real defect images of wheel treads were collected and contrasted with five classic detection models to verify the superiority of the algorithm.The results show that the algorithm can achieve 77.9%accuracy and 72.3%recall;in addition,the proposed algorithm can detect 125 images per second,and the model weight is only 15.1 MB.The model can quickly and accurately detect such two defects as peeling and sagging,and can be easily applied to the actual real-time defect detection scenario of wheel tread.
作者
张昌凡
徐逸夫
何静
杨皓楠
ZHANG Changfan;XU Yifu;HE Jing;YANG Haonan(School of Electrical and Information Engineering,Hunan University of Technology,Zhuzhou,Hunan 412007,China)
出处
《机车电传动》
北大核心
2022年第6期1-9,共9页
Electric Drive for Locomotives
基金
国家自然科学基金项目(52172403,62173137)。
