Research On Defects Detection Method Of High-speed Railway Catenary And Rail Based On Deep Learning

High-speed rail is a very important mode of transportation,which has the characteristics of strong carrying capacity,fast speed,low cost,and strong weather resistance.At present,our country’s investment in the development of high-speed rail continues to increase,the scale of high-speed rail is rapidly increasing,and related safety issues are increasingly worthy of attention.However,traditional defect detection methods have low efficiency and high error rates.Based on deep learning,this paper focuses on the identification methods of catenary components and the detection methods of catenary and rail defects.The main work and contributions are as follows:1.Data preprocessing.Perform image screening processing,data enhancement processing,and manual annotation processing on the collected data.Currently,most deep learning target detection methods are data-driven,and the quality of the data set is critical to the final result of the model.However,due to the special field,there is basically no public contact data and rail data.The data set in this thesis is the real catenary data and rail data collected from the Jiangguibei connecting line.Compared with the public training data sets such as coco,the single images of the collected data sets are more complicated,with lower purity and more invalid data.To apply some data to deep learning methods,it must be processed accordingly.This thesis experimentally analyzes the effects of using different data enhancement methods.The results show that the combination of geometric data enhancement and Mixup has achieved good results.2.Based on YOLO v3,the target detection method is improved.In the data set we actually collected,many objects to be detected occupies a very small proportion in the image,the categories are many and unbalanced,and the positive and negative examples between the categories are not balanced.Direct use of YOLO v3 for detection results is not good.In response to these problems of the data set,this thesis combines the multiscale detection in YOLO v3,and uses the organic combination of the Dice loss function and the Focal loss function to improve the original loss function in YOLO v3.After experimental comparison,the improved YOLO in this paper has achieved relatively good results on the data set.3.Improved the defect detection method based on YOLO v3.Due to security requirements,there are very few negative samples in the data we can collect,which is not conducive to the model learning the characteristics of the defect category,and even makes the learned model unable to be used to detect defects in practice.In response to this problem,this thesis uses the characteristics of the stacked denoising autoencoder and combines it with the multi-scale detection of YOLO v3 to identify the degree of similarity between the target to be detected and the real defect-free data,and finally the identification result is compared with the corresponding detection scale The test results are weighted and integrated.Experiments show that the defect detection method proposed in this thesis achieves better results than the original YOLO v3.4.The design and realization of the detection system for catenary and rail defects.Based on the improved target table detection method and defect detection method,a detection system for catenary and rail defects is designed and implemented.