Based On Image Processing Of The Catenary Pipe Cap And Steady Ear Nut Defect Recognition

The catenary support and positioning device is an important part of the catenary,and its normal operation is directly related to the normal operation of the high-speed railway,which is of great significance in the safe operation of the railway.The pipe cap and the steady ear are important components of the catenary support and positioning device,and the nut is one of the important parts of the steady ear.The pipe cap plays a role in protecting the insulation between the cantilever of the support device and the positioning pipe.The steady ear plays an important role in fixing the contact wire of the catenary.As the train runs at high speed for a long time,it will inevitably have an impact on the support and positioning device,which may cause the failure of the parts of the support and positioning device,and then cause the train to run abnormally.Therefore,it is particularly important to detect and eliminate the defects of catenary parts in time.At present,the main detection methods for catenary parts failure are mainly manual line inspection,manual inspection of high-definition pictures taken by 4C devices,and traditional algorithms for defect identification.These identification methods not only cannot guarantee the accuracy,but also the detection time is too long to meet the real-time requirements.The traditional algorithm also has the disadvantage of poor robustness.In recent years deep learning methods have appeared in the catenary parts defect recognition,deep learning has the characteristics of fast,high accuracy,good robustness,etc,however,in the existing research,the fault identification of small targets such as catenary pipe cap and catenary steady ear nut is less,and the research is still difficult.In this paper,a deep learning method is used to locate the pipe caps and locating steady ear,and the nuts and caps in the area obtained by locating them are identified as defects.The research content of this paper is built on the high-definition pictures taken by 4C system,taking the pictures of pipe cap and steady ear nut as the research object,and using the improved Faster R-CNN algorithm to realize the defect recognition of pipe cap and steady ear nut,and the main work within is as follows.（1）Generating adversarial networks to expand samples using DCGAN in the process of making pipe caps and steady ear to locate sample sets,and increasing sample diversity using rotation and adding pretzel noise;Subsequently,the Labelimg labeling tool is used to label the pipe caps and steady ear and train the target localization model;the localized pipe caps and steady ear areas are cropped and the defective sample set is created using the same method used to create the locating sample set;then the defective sample recognition model is trained.（2）In the target localization model,the proportion of anchor boxes and the area in the Faster R-CNN algorithm are improved by the K-means clustering algorithm,and the VGG16feature extraction network in the original Faster R-CNN algorithm is replaced with Res Net50,Res Net101,and Res Net152,respectively.By comparing the recognition accuracy,recall,F₁value,single detection time and other indexes,the optimal feature extraction network for pipe cap localization is selected as Res Net50,and the optimal feature extraction network for steady ear localization is Res Net101.Finally,the recognition effect of the improved algorithm in this paper is compared with SSD and YOLOv3,and the effectiveness of the improved localization model in this paper is verified.（3）In the defect recognition model,an improved Faster R-CNN algorithm is proposed to add the feature pyramid network FPN layer to the Faster R-CNN backbone network with Res Net101 as the feature extraction network as a way to improve the feature extraction ability of the feature extraction layer for the target,and the K-means algorithm is used to cluster the proportion of the labeled boxes of the defect samples to derive a suitable.The proportion of anchor boxes for defect recognition in this paper and replace the proportion of anchor boxes in the RPN layer of the original algorithm,and increase the area of two kinds of anchor boxes,and change the ROIpooling algorithm to ROIAlign algorithm.Simulation experiments are conducted on the improved defect identification algorithm of this paper,and the simulation experiments prove that the improved algorithm of this paper can effectively and quickly identify the pipe cap and steady ear nut defects.Subsequently,the improved defect recognition algorithm of this paper is compared with SSD and YOLOv3 algorithms to verify the effectiveness of this paper’s algorithm.