| The health status of track fasteners plays an extremely important role in ensuring the safe operation of trains.In recent years,with the rapid development of computer vision technology and deep learning,track inspection trains based on deep learning have replaced manual inspection methods to achieve automatic inspection of tracks.However,in terms of track fastener defect detection,the target detection algorithm used has many shortcomings such as difficulty in identifying overlapping defects,insufficient feature extraction in complex backgrounds,and difficulty in achieving real-time detection,which can easily lead to missed and false detections.Therefore,studying high-precision and high-efficiency rail fastener defect detection methods and lightweight rail fastener detection systems has important engineering practical significance for improving the real-time inspection efficiency of China’s rail status.In response to the above issues,this article designs a track fastener defect detection algorithm based on lightweight convolutional neural network,which is deployed on the Jetson AGX Xaiver deep learning development platform,applied to fastener defect detection tasks,and constructs a self collected high-definition fastener dataset to train,test,and evaluate the designed network.The main research work and results are as follows:(1)Build a large-scale dataset on rail fastener defects.The data is mainly collected by the intelligent inspection car equipped with linear array cameras in railway sections and manually designed by the research group.Mixup,Mosaic and other data augmentation technologies are used to expand the sample and balance the number of categories.(2)Propose a weighted fusion algorithm based on Softer NMS to optimize the regression box screening stage,thereby improving the problem of difficult identification of overlapping defects and target box offset in fastener detection tasks;Based on the CBAM attention mechanism,an EAM module that combines spatial attention and efficient channel attention is proposed by integrating ECA,and embedded into the Neck network.This strategy can effectively improve the model’s ability to extract key effective feature information,thereby improving the problem of insufficient defect feature extraction in complex backgrounds in the target detection network,and reducing missed and false detections during the detection process.(3)Propose a lightweight rail fastener defect detection network architecture ELYOLOv5.First,the improved ECA attention is introduced into Ghost Bottleneck,and the EG module is formed by combining depth separable convolution,and then the DBF module is formed by combining depth separable convolution with funnel activation function,which is used to reduce redundant operations in the backbone network and build a new lightweight backbone network.This strategy significantly improves the detection accuracy of the model and reduces the number of network parameters and computational complexity,enabling detection algorithms to be deployed on embedded platforms with limited computing resources for real-time detection.(4)Design a rail fastener defect detection system based on lightweight convolutional neural network,deploy the improved lightweight detection algorithm framework to the Jetson AGX Xaiver deep learning development platform,and conduct Tensor RT acceleration experiments.Build a human-machine interaction interface on the development platform to process video data in real-time.The experimental results show that compared with the improved main body method YOLOv5,the EL-YOLOv5 lightweight convolutional neural network model proposed in this paper has higher detection accuracy and less parameter calculation while ensuring a certain detection speed.Its detection accuracy has been improved by 4.7%,and the parameter quantity has been reduced by 48.7%.This algorithm can meet the industrial demand for real-time detection of rail fastener defects,Applied to large-scale rail fastener detection equipment. |
