Defect Identification Of Catenary Dropper Based On Image Recognition And Deep Learning

The contact network is an important energy source for the electrified railway and is an important part of the supply of traction for high-speed railways,ensuring the safety of the railway’s energy supply.In the high-speed railroad power supply system,the dropper connects the load-bearing cable and the contact line,which is responsible for conducting the current in the system and improving the electrical quality of the contact line.It also improves the elasticity of the contact line in the horizontal direction and reduces the chord in the middle of the span in the vertical direction.Therefore,it is of great significance to ensure the stability of the railway power supply system and the safe operation of high-speed railway to ensure that the dropper is in normal working condition.The working environment of the dropper is bad,it is not only eroded by the natural environment,but also impacted by the pantograph during the operation of the train,resulting in the dropper line relaxation,loose strands,fracture,current-carrying ring fracture and other faults,endangering the power supply quality and safe operation of the high-speed railway.Although the introduction of 4C devices and computer image processing technology has improved the efficiency of contact network condition detection,the existing algorithms are not perfect and there are problems like missing detection and false detections when used for on-site inspection,so manual methods of re-checking still need to be used.In this thesis,using the catenary dropper as the target,the object recognition algorithm was improved to enhance the efficiency of contact network condition detection and to guarantee power supply reliability.First,optimized the images taken by the 4C device and created a data set,restored and enhanced the image with poor effect,and removed the motion blur for the image by least square filter method;use histogram equalization for the image with too low brightness to enhance the contrast of the dropper in the image;the dataset was also expanded using rotation and the addition of pretzel noise.Then,based on the Faster R-CNN algorithm to position the dropper images from all images taken by the 4C device.The proportion and area of the anchor box in the RPN layer were modified by K-means clustering algorithm to make it more suitable for the size of the dropper,and the original VGG16 network was substituted with Res Net feature extraction network to improve the feature extraction capability of the network by increasing the network depth.Through comparative tests,it is verified that when the feature extraction network is the Res Net101 network,the best effect is achieved.After image preprocessing and network improvement,the proposed dropper positioning method can achieve accurate positioning,with positioning accuracy of 96.59% and recall rate of 97.19%.Finally,the FCOS network was used to identify the status of the dropper images.Because there were fewer images with defects in the dropper data set and VOVNet can make full use of the extracted features,so the VOVNet39 was used in the feature extraction part in place of the Res Net network in order to better facilitate the identification of the dropper status.In the regression calculation of the anchor box,The IOU intersection ratio calculation was substituted by the Generalized intersection over union(GIOU)calculation to enhance the ability to distinguish the relative spatial position between the detection box and the ground-truth box in the training process;In the focus loss function,the positive and negative samples imbalance during the training process was eliminated by adjusting the α parameter;Through the above improvements to the FCOS network,the recognition ability of the network for the dropper status is enhanced.In summary,with the dropper images taken by 4C device as the data set,the catenary droppers are localized by using the improved Faster R-CNN network,and the improved FCOS network is used to identify the status of the dropper,and the simulation results are compared with other target detection networks.The simulation results shows that the proposed network can accurately locate droppers and accurately identify the relaxation,breakage of the dropper string and breakage of the current-carrying ring,the validity and feasibility of the proposed algorithm are proved.