Pantograph Catenary Arcing Detection Algorithm Based On Attention Mechanism

Since the 21 st century,the construction of high-speed railroads in China has led the development of high-speed railroads in the world.As a heavy branch of electrified railroad,the traction power supply system of high-speed railway plays a crucial role in the development of the whole railroad.China’s vast territory and complex natural environment make high-speed railways encounter various climatic and geographical conditions during operation,and this complex and changing situation requires very high performance and maintenance of the traction power supply system.As one of the most important subsystems in the traction power supply system,the pantograph-catenary system works in a complex and harsh environment,and thus many types of faults can occur,among which the typical fault is arc-burning fault,i.e.high power discharge,which will produce Joule heat to melt or corrode the network components at the same time.In order to improve the effectiveness of the bow network system,it is necessary to explore the use of advanced science and technology to detect and identify the frequency,location and arc length of arc ignition in the bow network system during the operation of high-speed rail,and to control the arc ignition rate within a reasonable range.This has also become an important research topic for the study of high speed rail arch network system.Most of the existing arch network arc burning identification problems are based on traditional image processing methods,and few of them use machine vision algorithms.The existing machine vision based arch network arc burning recognition algorithms are mainly based on traditional target recognition algorithms(Faster R-CNN and YOLOv3).Due to the small image pixel size of the burning arc,such methods tend to ignore the features of small targets,resulting in poor generalization of the method,which makes it difficult to achieve a high effect of burning arc recognition.In order to solve the above problem of identifying arcburning faults in high-speed rail arch network,further improve the accuracy of arch network arc-burning fault detection,and reduce the probability of wrong and missed detection,this dissertation improves the classical target detection algorithms(Faster R-CNN and YOLOv5)based on attention mechanism,and builds four target detection algorithms incorporating attention mechanism for more accurate arc-burning fault detection problem(Faster RCNN+CBAM,Faster R-CNN+SE,YOLOv5+CBAM,YOLOv5+SE)for more accurate arc fault detection problems.In addition,given the advantage of fast multiscale feature fusion of weighted bidirectional feature pyramid network,which can better solve the difficulties of small target detection in combustion arc,a simplified model of BiFPN,BiFPN-mini,is introduced into the YOLOv5+SE model.The experimental results show that the YOLOv5+SE+BiFPNmini algorithm has the best experimental results in the arc-burning detection problem of the bow network compared with the benchmark model and four algorithms that fuse attention mechanisms.Finally,this dissertation further builds the basic YOLOv5+SE+BiFPN-mini algorithm for arch network arc-burning detection system.This study is of great significance to the improvement of intelligence and informationization in the high-speed railway industry,and has important application value to the intelligence of the bow network system.The main contributions of this dissertation are as follows:(1)This dissertation studies the arc-ignition fault identification problem of electric locomotive pantograph system,collects pantograph monitoring video,intercepts frames in separate scenes,and produces a VOC2007 format data set after data balancing and data enhancement expansion.(2)In this dissertation,we propose a series of attention fusion-based algorithms(Faster RCNN+CBAM,Faster R-CNN+SE,YOLOv5+CBAM,YOLOv5+SE)for the detection of arcburning faults in high-speed railway pantograph network with small target image pixel size and easy to ignore features,which can effectively enhance the existing target detection algorithms(Faster R-CNN+CBAM,Faster R-CNN+SE,YOLOv5+SE).detection algorithms(Faster RCNN and YOLOv5)to improve the accuracy and recall rate of the existing target detection algorithms(Faster R-CNN and YOLOv5)for arc ignition recognition.(3)In this dissertation,we propose the YOLOv5+SE+BiFPN-mini algorithm based on weighted bi-directional feature pyramid network for the recognition of arc-burning faults in the arch network of high-speed railways,which has further improved the recognition effect.(4)In this dissertation,the model with the best experimental results is packaged and applied to the bow network arc ignition detection system(6C system),which is practically applied to the field of high speed rail bow network monitoring.