基于改进卷积神经网络的受电弓滑板缺陷识别方法  

Defects Recognition Method of Pantograph Contact Strip Based on Improved Convolutional Neural Network

作　　者：王科理 石春珉 王克俊 程传彬 李勇 孙飚 WANG Keli;SHI Chunmin;WANG Kejun;CHENG Chuanbin;LI Yong;SUN Biao(Postgraduate Department,China Academy of Railway Science,Beijing 100081,China;Standards&Metrology Research Institute,China Academy of Railway Sciences Corporation Limited,Beijing 100081,China;China Railway Test&Certification Center Limited,Beijing 100081,China)

机构地区：[1]中国铁道科学研究院研究生部,北京100081 [2]中国铁道科学研究院集团有限公司标准计量研究所,北京100081 [3]中铁检验认证中心有限公司,北京100081

出　　处：《铁道机车车辆》2024年第4期99-105,共7页Railway Locomotive & Car

摘　　要：受电弓滑板监测装置(5C)利用高清成像装置获得机车受电弓滑板图像,以确保受电弓运行状态良好。针对传统方法检测精度低的问题,提出基于卷积神经网络的受电弓滑板缺陷识别方法。替换YOLO v5网络激活函数提高模型训练速度和泛化能力,通过数据增强方法平衡正负样本的比例。通过对现场5C数据进行试验,表明该方法的准确率达到97.33%,召回率达到了88.77%,F1分数达到了92.85%,证明了其在实际铁路场景中的应用价值。The pantograph contact strip monitoring device(5C)uses high-definition imaging devices to obtain images of the locomotive's pantograph contact strip to ensure that the pantograph is in good operating condition.Aiming at the problem of low detection accuracy in traditional methods,a method for identifying pantograph contact strip plate defects based on convolutional neural networks is proposed.Replacing the YOLO v5 network activation function improves the training speed and generalization ability of the model,balances the proportion of positive and negative samples through data enhancement methods.On-site 5C experiments data show that the accuracy rate of this method reaches 97.33%,the recall rate reaches 88.77%,and the F1 score reaches 92.85%,demonstrating the application value in actual railway scenarios.

关 键 词：受电弓 滑板 目标检测 图像识别 卷积神经网络 

分 类 号：U264.34[机械工程—车辆工程]