Research On Fastener Identification And Positioning Algorithm Applied To Rail Bolt Maintaining Machine

Railway fastener is an important part of connecting sleeper and railway,which plays an important role in ensuring the safety of train,and it is urgent to improve the manual maintenance of fasteners with the increasing of railway in China.Thus,it is particularly important to develop a railway equipment which can replace the railway workers for fastener maintenance,and one of the keys to develop such equipment is the identification of fasteners,including fastener nut positioning and fastener status classification.At present,the image recognition technology based on deep learning develops rapidly compared with the traditional image recognition method,which has many advantages and more application prospects in the future.Besides,deep learning has great advantages in how to locate the fastener nut and how to detect and judge the integrity of the fastener.Therefore,the research of fastener positioning detection technology based on deep learning is of great significance to the development of automatic fastener maintenance equipment.Based on deep learning technology,this paper first introduces the basic knowledge of convolutional neural network,compares and analyzes the commonly used activation functions in detail,and clarifies the most essential characteristics of the two convolutional neural networks: local connection and weight sharing.Then it introduces three classical network structures and explains the development trend of neural network models in recent years.The RF8000 was established through the data collection,enhancement,and labeling because there is no public fastener data set.Then,Based on three algorithms,Faster-RCNN,YOLOv3,and MobileNetv1-SSD,three fastener detection models were constructed.Use RF8000 as the data sample to train the model.Completed the analysis of the accuracy,speed and memory usage of the three algorithms.A lightweight fastener detection algorithm Mobilnetv3-TL was constructed because of insufficient computing power of the hardware platform.A feature pyramid structure is introduced to fuse multi-scale feature maps in order to improve the positioning accuracy of the nut.Mobilnetv3-TL widely uses the inception structure to replace the 3×3 convolution kernel to achieve the purpose of reducing the amount of calculation.It is found that the spe ed of Mobilnetv3-TL fastener detection model is optimal,and the accuracy is equivalent to Faster-RCNN through experimental comparison.Finally,the hardware environment of the maintenance equipment is analyzed,the advantages and disadvantages of the industrial control computer and the Jetson TX2 are compared.Jetson TX2 is selected as the hardware platform of the maintenance equipment considering the requirements of computing power.Mobilnetv3-TL was deployed in the development board after the Jetson TX2 environment built,and it was verified that Mobilnetv3-TL's inference speed reached 13.2 FPS/s,which met the requirements of the fastener detection task of maintenance equipment.