Research On Intelligent Identification Of Pins Defects In Positioning Tube Of High-speed Rail Catenary Support Device

As a key infrastructure to provide reliable and uninterrupted power to high-speed railway EMUs and electric locomotives,the safety and stability of the electrified railway catenary’s working state directly affects the safety of railway transportation.The split pin of the catenary positioning pipe is an important fastener in the catenary cantilever structure.The harsh environment and the vibration generated in the long-term operation of the train may cause it to fall off,resulting in the loosening of the cantilever structure and affecting the safe operation of the train.With the 4C inspection vehicle of high-speed railway catenary is put into use,the staff will judge the working status of parts one by one through images of catenary collected by the detection vehicle.This method is inefficient and cannot timely guide the maintenance.Therefore,it is necessary to study the intelligent defect identification method of catenary parts.In this paper,a three-level cascading architecture covering the locating and extraction of the split pins,the expansion of the defect sample and the training of the defect recognition model is designed.After the improved simulated defect samples generated by DCGAN are screened by lightweight CNN,they are used to optimize the training of the VGG16 defect recognition model,which overcomes the problem of insufficient defect samples of the split pins and realizes the defect recognition of the split pins based on the unbalanced samples of categories.Firstly,in the first level of the architecture,the pin location extraction is carried out.Since the image of the catenary cantilever structure collected by the 4C detection vehicle contains multiple parts and the image background gray level is close to the parts,Faster R-CNN was used to locate the split pins accurately,and the center point extraction method was designed to extract the split pins of the same scale according to the positioning coordinates to complete the establishment of the split pins sample.Then,the second level realizes the expansion of the split pins defect sample.To alleviate the problem of over fitting caused by insufficient sample size of defect recognition model training,compared to mainstream GANs generated result,choose DCGAN as generation model and adjust the structure of its model,with the improved DCGAN generate simulated defect samples,and building a lightweight CNN classification model to screene generate simulated defect samples,ensure the accuracy requirement of expansion of defect sample.Finally,the third level is the training of the defect identification model of the split pins.The expanded defect sample set in the second level and the split pins sample established in the first level are used as the training set and trained to convergence on the optimized VGG16 model of migrating the pre-training parameters of VGG16,so as to realize the defect identification of the split pins.Based on the Tensor Flow deep learning architecture,the experimental that the real images collected by 4C inspection vehicle are used for verification show three-level cascade architecture designed in this paper has a high accuracy rate of 99%in defect recognition of split pins,and has great robustness for images with different shooting angles and exposure intensities.