深度学习目标检测算法在货运列车车钩识别中的应用  

Application of deep learning target detection algorithm in freight train coupler recognition

作　　者：郭忠峰[1] 张渊博 王赫莹[1] 任仲伟 GUO Zhongfeng;ZHANG Yuanbo;WANG Heying;REN Zhongwei(Key Laboratory of Intelligent Manufacturing and Industrial Robot of Liaoning Province,Shenyang University of Technology,Shenyang 110870,China;Guizhou Institute of Technology,Guiyang 550009,China)

机构地区：[1]沈阳工业大学辽宁省智能制造与工业机器人重点实验室,辽宁沈阳110870 [2]贵州理工学院,贵州贵阳550009

出　　处：《铁道科学与工程学报》2020年第10期2479-2484,共6页Journal of Railway Science and Engineering

基　　金：辽宁省教育厅资助项目(LQGD2017034)。

摘　　要：铁路货运列车的自动摘钩是实现驼峰作业自动化的重要环节,为了完成货运列车自动摘钩工作,需要实现车钩的快速准确识别。通过当前广泛应用的YOLO_v2网络模型,研究针对货运列车在正常工作条件下车钩的识别问题,通过K-means聚类算法对YOLO_v2网络模型中anchor的个数进行调整优化,找出适用于本次车钩识别的最优anchor个数以及宽高维度,并通过训练自制具有明显目标特征数据集来获取更加准确的权重。结果表明改进YOLO_v2模型在精确度上达到92.6%;在召回率上达到了91.8%;在FPS上达到45帧/s,改进的YOLO_v2模型达到了预期设计目标。Aiming at the complex problem that the EMU(Electric Multiple Units) trains operation need to consider punctuality, energy saving, safety and comfort, the operation time was regarded as the standard of passenger satisfaction, and the energy consumption was seen as the standard of the railway company satisfaction. Meanwhile, considering the influence of electrical phases in order to make the train operation more in line with the actual situation, a multi-objective optimization model for the train operation was established, which was constrained by safety, track characteristics and passenger comfort and so on. Then, a kind of algorithm combining NSGA-Ⅱ(Non-dominated Sorting Genetic Algorithm-Ⅱ) with golden ratio technology was proposed to solve the problem of uneven individuals distribution in the solution space when using NSGA-Ⅱ to optimize. The tests of the algorithms show that the golden ratio NSGA-Ⅱ algorithm has better distribution and convergence than the NSGA-Ⅱ algorithm. Finally, taking CRH3 of a certain section of Wuhan-Guangzhou line as a simulation case, some simulation results are shown, which further indicate that the model and the proposed algorithm are feasible.

关 键 词：改进YOLO_v2模型 货运列车 车钩识别 K-MEANS聚类算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]