基于深度学习的铁路列车关键零部件图像故障检测  

Vision-based fault detection for key components of railway train based on deep learning

作　　者：李萍[1] 吴斌方[1] 刘默耘 张杨[1,2] 林凯 孙国栋 LI Ping;WU Binfang;LIU Moyun;ZHANG Yang;LIN Kai;SUN Guodong(School of Mechanical Engineering,Hubei University of Technology,Wuhan 430068,China;Department of Computer Science,Nanjing University,Nanjing 210023,China;School of Mechanical Science and Engineering,Huazhong University of Science and Technology,Wuhan 430074,China)

机构地区：[1]湖北工业大学机械工程学院,湖北武汉430068 [2]南京大学计算机科学与技术系,江苏南京210023 [3]华中科技大学机械科学与工程学院,湖北武汉430074

出　　处：《铁道科学与工程学报》2019年第12期3119-3125,共7页Journal of Railway Science and Engineering

基　　金：国家自然科学基金资助项目(51775177);大学生创新创业训练项目(201710500038)

摘　　要：提出一种基于深度学习中卷积神经网络的列车关键零部件图像故障视觉检测算法。首先,引入故障区域复合提议网络和一组先验包围盒来生成高质量的故障区域;然后,采用线性非极大值抑制算法来保留最合适的故障区域并去除冗余;最后,结合故障区域复合提议网络,提出一种多尺度故障检测网络来进行故障区域分类和精确检测。本文将提出的算法在多个典型列车故障的数据库中进行实验,结果表明,本算法检测精度高,检测速度为每张图像0.246s,检测性能明显优于现有的最先进的方法,能更好地应用于实际工程中。This paper proposes a visual detection method for fault detection of key components of railway train based on convolutional neural network in deep learning. Firstly, the multi-region proposal network with a set of prior bounding boxes was introduced to achieve high quality fault proposal generation. Then, a linear non-maximum suppression method was applied to retain the most suitable anchor while removing redundant boxes. Finally, a powerful multi-level region-of-interest(RoI) pooling was proposed for fault zone classification and accurate detection. The proposed method was attempted in several typical train faults databases, and the experimental results indicate that the proposed method can achieve high accuracy with 0.246 s per image including all steps, substantially outperforming the state-of-the-art methods. The proposed method can be used for actual fault detection of freight train images.

关 键 词：铁路列车 故障检测 卷积神经网络 多尺度 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]