Detection Of Bolts On Apron Boards Of EMU Based On Convolutional Neural Networks

Trouble of moving EMU Detection System(TEDS)is a system that guarantees the safety of high-speed running EMU,but the state of the key components in the EMU images collected by TEDS is mainly judged manually.With the rapid development of deep learning,object detection algorithms based on convolutional neural network are widely used.Based on YOLOv2,this paper proposes the YOLOv2-Plus with improved SPP layer,combined with multiple receptive fields,YOLOv2-Plus improves the performance of detecting small targets.Then it is applied to the detection of bolts on EMU apron boards,improving the accuracy and effectiveness.In the dataset of bolts on EMU apron boads,fault samples with missing bolts are scarce.First,according to the characteristics of normal bolts and bolt-missing samples,this paper uses image augmentation technology to augment the dataset of bolts on EMU apron boards.Then the K-Means algorithm is used to cluster the bounding box size of the bolt area to be detected.The target areas share similar sizes and occupy small areas relative to the whole image,then an SPP layer that combines multiple receptive fields is proposed.According to the different sizes of small targets in multi-resolution images,the kernel sizes of the pooling layers in the SPP layer are designed to enhance the model’s ability of perceiving small targets of different sizes.On this basis,the low-level feature map of Darknet-19 is selected,and the features are further extracted by the convolutional layer after the SPP layer,then the Passthrough layer is used for feature rearrangement,the feature map is concatenated with the high-level feature map and fed into the classifier,higher network resolution and Mish activation function are selected to improve the performance of the model,the YOLOv2-Plus which improves the accuracy of the detection of small targets is proposed.In order to validate the improvement of the algorithm,YOLOv2,YOLOv3,YOLOv4,Faster R-CNN and YOLOv2-Plus are trained and tested by the augmented dataset of bolts on EMU apron boards.With the input resolution of 608×608,the mean average accuracy(m AP)of YOLOv2-Plus is 2.12% higher than that of YOLOv2,reaching 95.66%,and the detection speed reaches 67.1FPS.The m AP and detection speed of YOLOv2-Plus respectively exceed that of YOLOv3 and YOLOv4 in the testing dataset.Although the m AP is slightly lower than that of Faster R-CNN,the detection speed of YOLOv2-Plus is 3 times as that of Faster R-CNN.YOLOv2-Plus achieves a good balance between detection accuracy and detection speed,can quickly and accurately detect the bolts and bolt-missing areas on apron boards of running EMU.Realizing YOLOv2-Plus with Darknet framework,this paper designs a web-based bolt detection system for EMU’s apron boards.In the task of manual troubleshooting,this system has certain practicability and can reduce the working pressure of analysts.