Abnormality Detection Of Catenaty Suspender Of High-speed Railway Based On Deep Learning

The detection and monitoring of high-speed railway catenary is an important means to ensure the safe operation of railway traction power supply.The current image detection methods for catenary components are not efficient and the detection accuracy is low,so it is necessary to study an efficient deep learning model and recognition algorithm for anomaly recognition.There are many kinds of catenary components,and the suspender is taken as a typical representative to identify and improve abnormalities.This article introduces the situation of the catenary inspection system of the railway bureau,considering the identification of all components,and designing an overall plan for its initial positioning and classification,and then identifying abnormal states separately.Due to the phenomenon of motion blur and underexposure in railway bureau component images,before anomaly recognition,Wiener filtering method is used to restore the motion blurred images and global histogram equalization method is used to brighten the underexposed images.Through experimental comparison,both methods have a certain enhancement effect on the image of the contact network.In addition,data expansion methods are used to increase data set samples,including Mosaic,Copy paste,affine transformation,Mix Up and other methods.In order to improve the detection speed during the initial positioning recognition of contact network components,the lightweight PP-LCNet model backbone network was replaced with the original YOLOv5 model backbone network to become the P-YOLOv5 model.Through training and testing of the P-YOLOv5 model,the experimental results showed that the recognition speed was as high as 42.6 frames,an increase of 3.1 frames,and the accuracy,recall rate,and m AP value slightly decreased.The experimental results show that the algorithm effectively improves recognition speed and ensures accuracy with minimal impact.In addition,the "two-step" method is used to identify the abnormal suspension wire.The Res Net50 model is used to identify the status of the suspension wire strand,and the loss value of this model is 0.0012 through experiments,with a verification accuracy of 98%.The experiment shows that the method has good recognition effect on suspension wires;Then,when it is difficult to identify the current carrying ring and nut small target of the dropper,the original YOLOv5 is improved,including CSPPF module,adding small target detection head,introducing centralized feature pyramid to display the visual center structure,and modifying the EIOU loss function.In the experimental results m AP@0.5 The value has reached 93.9%,an increase of 4.3 percentage points,and there has been a good improvement in accuracy and recall,but the detection speed has decreased.The experiment shows that although the improved accuracy of the model slightly decreases in detection,there is a significant improvement in average recognition accuracy,which effectively solves the problem of low recognition of small targets in nut loosening faults and achieves the expected recognition effect.