摘 要:锁紧板偏移故障是货运列车频发的典型故障之一,针对其平均识别精确度较低的问题,将目标检测YOLOv4 模型应用于改善货运列车部位锁紧板图像检测。首先,对锁紧板偏移、正常图像进行Mosaic 数据增强,以解决数据集样本较少问题。其次,使用k-means 聚类算法,得到更优的初始anchor 的位置,以提高故障检测精确度。最后,通过非极大值抑制获取得分最高的检测结果。实验表明,通过使用目标检测YOLOv4 能够精确地实现锁紧板故障的检测。
关键词:故障检测;锁紧板;YOLOv4
DOI:10.19850/j.cnki.2096-4706.2021.20.012
中图分类号:TP391.4 文献标识码:A 文章编号:2096-4706(2021)20-0042-03
WANG Miao, ZHANG Longxin
(Hunan University of Technology, Zhuzhou 412007, China)
Abstract: Locking plate offset fault is one of the typical faults frequently occurring in freight trains. Aiming at the problem of low average recognition accuracy, the object detection YOLOv4 model is used to improve the image detection of part locking plate of freight trains. Firstly, Mosaic data enhancement is performed on the offset of the locking plate and the normal image to solve the problem of less samples in the data set. Secondly, K-means clustering algorithm is used to obtain a better initial anchor position, so as to improve the accuracy of fault detection. Finally, the detection result with the highest score is obtained by non-maximum suppression. Experiments show that the use of target detection YOLOv4 can accurately detect the fault of locking plate.
Keywords: fault detection; locking plate; YOLOv4
[1] 黄东林. 关于提高TFDS 动态检查质量的探讨 [J]. 铁道车辆,2018,56(11):37-39+6.
[2] 孙国栋,徐威,梁永强,等. 基于形状上下文的列车挡键丢失图像识别算法 [J]. 铁道科学与工程学报,2014,11(6):127-131.
[3] 雷蕾,宋冬利,张卫华,等. 基于HCRD 的列车锁紧板偏转自动识别方法 [J]. 铁道科学与工程学报,2021,18(7):1895-1902.
[4] BOCHKOVSKIY A,WANG C Y,LIAO H Y M. YOLOv4:Optimal Speed and Accuracy of Object Detection [J/OL].arXiv:2004.10934 [cs.CV].[2021-08-27].https://arxiv.org/abs/2004.10934.
[5] 张雨金,杨凌帆,葛双冶,等. 基于Kmeans-SVM 的短期光伏发电功率预测 [J]. 电力系统保护与控制,2018,46(21):118-124.
作者简介:王苗(1998—),女,汉族,湖南耒阳人,硕士在读,研究方向:深度学习和故障检测;张龙信(1983—),男,汉族,湖南浏阳人,博士,副教授,CCF 会员(34915M),研究方向:深度学习和大数据分析。