基于改进YOLO v5的钢轨内部伤损B显图像识别与分类

Recognition and classification of internal defects in railway tracks based on improved YOLO v5

钢轨正常工作是铁路安全运营的重要保障,开展钢轨损伤检测研究具有重要的意义。钢轨的损坏、裂纹和疲劳可能引发脱轨和事故,威胁乘客和货物的安全。传统的人工钢轨探伤存在判伤时间长、漏报率高等问题。基于改进的YOLO v5目标检测算法,对铁路行业的钢轨超声波B显图像进行伤损识别与分类研究。在YOLO v5网络结构中引入卷积注意力机制模块,基于一种改进的损失评价函数,以提高模型训练的速度和鲁棒性。同时,在样本集构建方面,改进马赛克(Mosaic)数据增强方法,随机粘贴一个或多个样本小目标于另外样本对应出波区形成新样本,扩充B显图像数据集;基于数据迁移技术,将基于第3方钢轨超声波B显样本数据训练得到的模型参数作为模型训练的初始化参数以提升模型泛化性能。对实际标准样轨进行实验,可成功识别全部4类典型钢轨伤损,平均精度均值(mean average precision,mAP)mAP@0.5∶0.95达到了76.9%,验证了研究成果良好的快速性、准确性、鲁棒性和泛化性。

The normal operation of steel rails is crucial for ensuring safe railway operations,making research on steel rail damage detection highly significant.Damages,cracks,and fatigue in steel rails can lead to derailments and accidents,posing a threat to the safety of passengers and cargo.Traditional manual steel rail inspections suffer from long inspection times and high miss rates.This study focuses on the identification and classification of damages in ultrasonic B-scan images of railway steel rails using an improved YOLO v5 object detection algorithm.The YOLO v5 network structure is enhanced by incorporating a convolutional block attention module(CBAM)and utilizing an improved loss evaluation function to improve training speed and robustness.In terms of dataset construction,the Mosaic data augmentation method is improved by randomly pasting one or more small target samples onto other samples in the corresponding wave area,expanding the B-scan image dataset.Additionally,data transfer techniques are employed to initialize the model training with the parameters obtained from training on third-party steel rail ultrasonic B-scan sample data,thereby enhancing model generalization performance.Experimental results on actual standard rail samples demonstrate successful identification of all four typical types of steel rail damages,achieving a mean average precision(mAP)of 76.9%for mAP@0.5:0.95.This validates the research findings in terms of their speed,accuracy,robustness,and generalization capabilities.