Research On Railway Fastener Detection And Recognition Based On Deep Learning

Taking high-speed rail and railway transportation has become an important part of people’s lives and social production.As the basic equipment for railway transportation,the performance of track directly affects the service life of railway lines and vehicles,so there is a significant investment in the research on automatic detection of rail at home and abroad.Generally,automatic detection of rail fasteners requires both lightweight real-time detection capabilities and high accuracy detection capabilities for abnormal state fasteners.However,due to the high cost of manual labeling,it is extremely difficult to produce large-scale rail fastener datasets.Therefore,this paper studies rail fastener detection and recognition based on deep learning from the perspectives of lightweight detection,semi-supervised learning,and long-tailed recognition of abnormal fastener.The main research content and innovation points are as follows:1.Improved YOLOv5 is proposed to achieve real-time detection of fastener status.By proposing improvements based on Mosaic data augmentation and Non Maximum Suppression,and adding offline Hard Example Learning module,the network improves network detect speed and GPU utilization,and increases the ability of network learning for hard samples.Experiments were conducted on railway track data sets.Compared with the original YOLOv5 model and other commonly used target detection algorithms,it was found that the improved network significantly improved detection accuracy and convergence speed.2.A semi-supervised fastener detection algorithm based on Fully Convolutional One-Stage Object Detection is proposed,which uses Exponential Moving Average to build a teacher-student framework.The strong data augmentation method of Position Encode/Decode is introduced to improve the robustness of the student model,and a dual channel feature fusion strategy is adopted to improve the student model.In addition,the predictions generated by the teacher model are finely divided through adaptive thresholds through a Fuzzy Region Filtering mechanism to obtain more accurate pseudo labels.Experiments were conducted on rail dataset by setting different degrees of labeled data ratios,and it was found that the Average Precision of the proposed improved algorithm was improved at different IOU thresholds.3.A long-tailed recognition model based on Dual Branch Learning Networks is proposed.In order to enhance the network’s ability to characterize tail abnormal fastener data,dynamic weighting parameters are used to organically combine the re-weighting learning branch using a class-balanced loss function with the re-sampling learning branch based on the improved Mixup algorithm through dynamic parameters.Dual Branch Learning Network achieves both improvements in the recognition accuracy of full category data and tail data in both the rail fastener dataset and the public dataset,thereby balancing the tendency of deep learning algorithms to favor the head category and improving the recognition ability of tail data.