At present,the defect detection of key components of electric multiple units (EMUs) has the problems of complex model, high missed detection rate of small targets and low detection efficiency.To solve the existing problems,a defect detection method based on improved YOLOv5 is proposed.On the basis of using generative adversarial network (GAN) to expand the dataset,the YOLOv5m backbone extraction network is changed to the MobileNetV3-large network structure,and the neck 3×3 convolution layer is optimized by using depthwise separable convolution to further reduce the model complexity.Then,the coordinate attention (CA) is introduced into the improved backbone network to capture the location information and channel information of small targets, thereby enhancing the feature expression ability of the network.The non-max suppression (NMS) algorithm is optimized by integrating the position information of the center point of the overlapping detection box to improve the accuracy of the prediction box location.The experimental results on the EMUs defect dataset show that,compared with YOLOv5m,the improved model reduces the amount of parameters by 77%,the amount of computation by 80.9%,the detection time of a single image by 31.7%, and the mean average precision (mAP) can reach 0.804.In addition,the experimental results on the NEU-DET dataset show that the improved model also has a strong generalization ability.