The Detection Of Small Sample Of High-speed Railway Catenary Support Devices

With the rapid development of high-speed railways,the safety inspection of railway catenary support devices is more stringent.However,the manual analysis is still used in railway system to find faults in the captured images,which is inefficient and extremely easy to miss faults.Moreover,the sizes of catenary components are different,and the fault are various.The fault sample amounts are smaller compared with the normal ones.The imbalance of positive and negative sample quantities leads to difficulty in constructing the automatic fault recognition model.Therefore,the multi-classification analysis system for the detected images of catenary support device and the state discrimination of isoelectric line in the case of small failure samples are studied in this paper.Firstly,the location and identification of 12 key components are realized based on the Faster R-CNN framework in this paper.Considering the wide variety of catenary network components and the clear difference in size,the database containing 12 types of key components is established.The basic structure of network layer,the loss function and the non-maximum suppression are optimized.The positioning of components is extracted finally.Secondly,the image processing technology,the deep concatenated convolutional network structure and the fault diagnosis based on generative adversarial network are adopted in order to train the detect model under the less samples limitation.After a series of experimental comparisons,it is shown that the fault diagnosis based on generative adversarial network is the most effective one.The improved Deep Convolution Generative Adversarial Network(DCGAN)is used to realize the fault diagnosis of isoelectric line.The isoelectric line image under normal working condition is used to train and generate the model.Then detected images are filled into the model.Finally,the generated image is compared with the input image and the working state is judged.The fault diagnosis under the condition of small samples can be realized and the applicability is better.Finally,the images in a section of high-speed railway are used as training data and the images of other sections are selected as test data for experimental comparison.The experimental results show that the fault diagnosis based on generative adversarial network can report the working status of isoelectric line quickly and accurately in small sample faults.This method has strong robustness.