近些年来，由于城市化的进一步推进，为了满足城市居民的出行需求，城市轨道交通得到了前所未有的发展。由此导致了地铁列车负荷量进一步加大，列车走行部的零部件更容易出现破裂、丢失等隐患。为保证人民群众的出行安全，需要对地铁列车的安全巡检提出更高的要求。传统的地铁列车巡检以人工为主，具有成本高，效率低下等特点。本文基于巡检现场采集的地铁列车走行部侧面高清线阵图像，研究了基于机器视觉的列车零部件缺陷检测流程的一系列算法，采用基于深度学习的数据扩充、零部件定位以及缺陷检测的方式，克服了采集的样本量少，正负样本极不均衡以及检测精度不高等难题，为地铁列车的自动化巡检提供可靠的解决方案。主要内容包括：1）分析了国内外列车关键部件缺陷检测、图像数据扩充以及一分类检测的研究现状，由此引出本文针对以上难点的技术方案。进一步介绍了卷积神经网络的原理、经典特征提取网络、一分类网络等，为后文提供了理论依据和技术途径。2）利用SRGAN网络，针对夹钳项点样本量较少和夹钳项点图片尺寸不一致且分布不均衡等特点，采取了超分辨率图像重构的方式实现了夹钳项点的数据扩充。3）利用YOLOv5网络，实现了对夹钳项点与电气箱盖项点的关键零件定位。为了保证技术方案在未来工作中能够更好地应用在嵌入式设备上，改进了YOLOv5网络的结构，将普通卷积替换为GhostConv，将CSP瓶颈层替换为GhostBottleneck，并添加了注意力机制，经过测试，改进的YOLOv5的mAP达到了0.991，同时在保持精度几乎不变的情况下，单张图片在CPU上的检测时间缩减了12.90%。4）针对夹钳项点中弹簧部件的负样本数目较少，正负样本数目不均衡的特点，采用了DifferNet算法实现了对夹钳项点中弹簧部件的缺陷检测，使得弹簧部件的AUC指标达到了0.9945。同时利用了迁移学习的方式，将DenseNet121网络作用于电气箱盖项点中的矩形、三角形标识和锁扣的缺陷检测步骤，保证了模型较小的情况下，准确率高达100%。在最终的联调实验中，两处项点的误报率均低于10%，漏报率为0。

In recent years, due to the further promotion of urbanization, urban rail transit has developed unprecedentedly in order to ensure the travel needs of urban residents. As a result, the subway train load is further increased, and the parts of the train running part are more prone to hidden dangers such as parts breakage and loss. Therefore, in order to ensure the travel safety of the people, it is necessary to put forward higher requirements for the safety inspection of trains. The traditional subway train inspection is mainly manual, which has the characteristics of high cost and low efficiency. Based on the high-definition linear array images of the side of the subway train running part collected at the inspection site, this paper studies a series of algorithms for the defect detection process of train parts based on machine vision, and adopts the methods of data expansion, part positioning and defect detection based on deep learning to overcome the problems of small sample size, unbalanced positive and negative samples and low detection accuracy, It provides a reliable solution for the automatic inspection of subway trains. The main contents include:1) This paper analyzes the research status of defect detection, image data expansion and one-class classification of key train components at home and abroad, and leads to the technical scheme for the above difficulties in this paper. It further introduces the principle of convolutional neural network, classical feature extraction network and one-class classification network, which provides a theoretical basis and technical way for the later paper.2) Using SRGAN network, aiming at the characteristics of small sample size of clamp item points and inconsistent picture size and uneven distribution of clamp item points, the data expansion of clamp item points is realized based on super-resolution image reconstruction.       3) Using YOLOv5 network, the key parts of clamp item and electric box cover item are located. In order to ensure that the technical scheme can be better applied to the embedded equipment in the future work, the structure of YOLOv5 network is improved, ordinary convolution is replaced by GhostConv, the CSPbottleneck layer is replaced by Ghostbottleneck, and an attention mechanism is added. After testing, the mAP of the improved YOLOv5 reaches 0.991, While keeping the accuracy unchanged, the detection time of a single picture on the CPU is reduced by 12.90%.4) In view of the characteristics of less negative samples and unbalanced number of positive and negative samples of spring parts in clamp item points, the DifferNet algorithm is used to detect the defects of spring parts in clamp item points, so that the AUC index of spring parts reaches 0.9945. At the same time, using the method of transfer learning, DenseNet121 network is applied to the defect detection steps of rectangle, triangle identification and lock in the item points of the electrical box cover, which ensures that the accuracy is close to 100% when the model is small. In the final joint commissioning experiment, the false alarm rate of the two items is less than 10%, and the omission rate is 0.