Research On Image Generation And Defect Recognition Of Catenary Dropper Based On GAN

Catenary is the equipment that provides electric energy for electric locomotives in railway operation,and the dropper is its component,which is an important component connecting the bearing cable and the contact line,and plays the role of parallel diversion,carrying a certain force and adjusting the height and degree of the contact line.Due to the long-term service of the dropper and the influence of environmental factors,the number of defective dropper will increase day by day.The defective lifting string reduces the flow quality of the train,and the heavy causes pantograph accidents,so it needs to be repaired in time.In recent years,some intelligent algorithms have been gradually applied in the inspection and analysis of defects of catenary hanging parts,but there are problems such as low accuracy and high missed detection rate,and 4C analysts are still required to use manual interpretation and manual marking to review a large number of missed and erroneous images,and analysts have long-term and high-intensity repeated detection operations,which is easy to cause visual fatigue,resulting in recognition accuracy can not be guaranteed.Training and testing these intelligent methods at the same time requires a large number of defect samples,which the reality cannot meet.Therefore,in view of the above problems,this thesis conducts research in the field of image generation and defect recognition,generates dropper defect images and improves the accuracy of defect recognition algorithms,and provides certain guidance for catenary maintenance.The research work in this thesis is based on the high-definition images collected by the4 C device,and the dropper image is the research object.Since the dropper defect image is much smaller than the normal image,the research work is divided into two parts: the first part uses the improved Cycle GAN algorithm to complete the generation of the dropper defect image;The second part uses the improved GANomaly algorithm to complete the recognition of defect images.The main research contents are as follows:(1)Aiming at the current shortage of dropper defect samples,this thesis proposes an improved Cycle GAN defect sample generation model.By improving the network structure of the generator and replacing residual blocks with dense convolutional blocks in the generator,the problem of gradient disappearance and slow convergence in the process of generating suspender defect images is solved.At the same time,a coordinate attention mechanism is added to the generator convolution layer and dense convolutional blocks,enabling the generator to accurately locate the defect area,thereby making the characteristics of the final generated sample more obvious.Simulation experiments have compared the effects of DCGAN and Cycle GAN before and after the improvement in generating dropper defect images,verifying the effectiveness of the improved algorithm in generating defective dropper samples.The generated dropper defect image is applied to the test set of subsequent defect recognition research algorithms.(2)In view of the problems that most current algorithms use defect samples as training data,such as the need for manual labeling and multi-stage processing,and the inability to identify unexposed defects,this thesis proposes an improved GANomaly defect recognition model,which adds label smoothing technology after reconstructing the image to avoid model overfitting.At the same time,the original GANomaly discriminator is replaced with a Patch GAN discriminator to make the discriminator more accurate.The model only needs to be trained with normal dropper samples,and the difference between the reconstructed image and the input normal image can be determined whether it is a defective image.Simulation experiments compare the performance of Ano GAN,EGBAD and GANomaly models before and after improvement,and verify that the detection effect of the improved algorithm is better than that of other algorithms.