Rail fastener detection of heavy railway based on deep learning  

作　　者：Yuan Cao Zihao Chen Tao Wen Clive Roberts Yongkui Sun Shuai Su 

机构地区：[1]National Engineering Research Center of Rail Transportation Operation and Control System,Beijing Jiaotong University,Beijing 100044,China [2]School of Electronic and Information Engineering,Bejing Jiaotong University,Beijing 100044,China [3]Birmingham Centre for Railway Research and Education,University of Birmingham,Birmingham B152TT,UK [4]Frontiers Science Center for Smart High-speed Railway System,Beijing Jiaotong University,Beijing 100044,China

出　　处：《High-Speed Railway》2023年第1期63-69,共7页高速铁路（英文）

基　　金：supported by the National Key R&D Program of China(Grant 2021YFF0501102);National Natural Science Foundation of China(Grant U1934219);National Science Fund for Excellent Young Scholars(Grant 52022010);National Natural Science Foundation of China(Grant 52202392,Grant 62120106011).

摘　　要：Image detection based on machine learning and deep learning currently has a good application prospect for railway fault diagnosis,with good performance in feature extraction and the accuracy of image localization and good classification results.To improve the speed of locating small target objects of fasteners,the YOLOv5 framework model with faster algorithm speed is selected.To improve the classification accuracy of fasteners,YOLOv5-based heavy-duty railway rail fastener detection is proposed.The anchor size is modified on the original basis to improve the attention to small targets of fasteners.The CBAM(Convolutional Block Attention Module)module and TPH(Transformer Prediction Head)module are introduced to improve the speed and accuracy issues.The rail fasteners are divided into 6 categories.Experiment comparisons show that before the improvement,the MAP@0.5 value of all categories are close to the peak of 0.989 after the epoch of 150,and the F1 score approaches 1 with confidence in the interval(0.2,0.95).The improved mAP@0.5 value approached the highest value of 0.991 after the epoch of 75,and the F1 score approached 1 with confidence in the interval(0.01,0.95).The experiment results indicate that the improved YOLOv5 model proposed in this paper is more suitable for the task of detecting rail fasteners.

关 键 词：Rail fasteners Fault diagnosis Heavy haul railways Deep learning YOLO5 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] U216.3[自动化与计算机技术—控制科学与工程]