Fault Detection Of Catenary Components Based On Generative Adversarial Networks

High-speed railway catenary is a key power supply facility in traction power supply system,which is mainly responsible for the power supply function of electric multiple units.Catenary support components(CSCs)are important parts of the catenary.Due to the high-density operation of the train in the long term,a series of faults are easily to happen,which seriously threaten the safe operation of trains.However,traditional image processing methods need to construct special detection algorithm for different fault characteristics because of the variety of defects.These algorithms are not universal and cannot complete the real-time fault identification.At the same time,the proportion of defective samples in all samples is very low,and the number of defective samples is insufficient,which leads to the over-fitting problem for supervised learning methods such as deep convolution neural networks(DCNNs)classifier.So it is difficult to effectively monitor abnormal samples,and the accuracy is low.Therefore,it is of great significance to develop a general and efficient fault detection system.In this dissertation,an end-to-end fault detection system for CSCs is designed.The unsupervised learning technology of emerging generative adversarial networks(GANs)in the field of computer vision is adopted,as well as DCNNs which has achieved excellent achievements in target detection.Firstly,various DCNNs for target detection such as SSD(Single Shot Multi Box Detector),Faster-RCNN(Faster Region Convolutional Neural Networks),YOLO(You Only Look Once),FPN(Feature pyramid networks),etc.are analyzed and compared in depth.In order to obtain accurate positioning results for subsequent fault detection,Faster-RCNN(Res Net-101)is finally selected to improve the accuracy of positioning targets.At the same time,the network parameters are adjusted,and some preprocessing methods(data augmentation,image denoising,etc.)are adopted to realize the positioning and extraction of CSCs which can obtain many CSCs samples.To some extent,it alleviates the heavy workload caused by manual labeling.Then the difficulty of limited fault samples is considered,generative models based on GANs are constructed,and various generation network frameworks,such as DCGAN(Deep Convolutional Generative Adversarial Networks),EGBAD(Efficient GAN Based Anomaly Detection)and GANomaly,etc.are analyzed and compared to find a good mapping from image space to high-dimensional feature spaces implicitly.Finally,according to the differences of the image distribution or potential space between the abnormal samples and the normal samples,a standard model for defect detection and evaluation can be established which synthesizes fault types and features of CSCs.In this dissertation,insulator and isoelectric line,as the most characteristic parts of CSCs,are selected to carry out relevant experimental tests.The experimental results show that the fault detection system can pre-screen faults of CSCs quickly and accurately and has good general fault detection capabilities.However,there are also some problems such as the fault area of the components is not located and the robustness of the generated network model is weak.In the end of this paper,an improved scheme is proposed to solve these two problems.The texture information is adopted instead of the overall components information to construct the generated network model for higher performance of the general anomaly detector.In addition,the memory unit can be adopted for avoiding the false positive problems caused by too strong performance of the generative models or encoders.