Catenary Key Components And Foreign Objects Detection Based On CNN And Sparse Coding

Catenary is an important equipment for electrified railway system.Due to its structural,mechanical characteristics and bad working environment,catenary can easily have failures such as loose components and fractures.Therefore,it is necessary to detect and monitor the suspension state of the catenary.Nowadays,monitoring the catenary support and suspension state is mainly by manual inspection or technical staff to view the image inspection,which both have problem of labor intensity,low work efficiency and low accuracy of detecting.Thus,an automatic detection system by image processing methods is in need.In this paper,we find some problems that need to be optimized by studying the existing algorithms for defects detection of catenary's components:(1)Since the algorithm for different components are parallel,it yields the problem of repeated calculation in the key regions localization step;(2)The precise localization and defect detection methods for the same type of components,which are located in different regions,are not different.That increases the complexity of the overall algorithm;(3)Some of precise localized images are blurred due to the failure of focus or low resolution,that makes it difficult for the following-up defect detection stage;(4)There is no algorithm of unknown foreign objects detection for catenary system.According to these four problems,this article mainly study on general algorithms for key regions/components localization,defect detection as well as the unknown foreign objects detection.The main work is as follows:Firstly,the paper compares the currently used object detection algorithms and feature extraction networks.According to the characteristics of region localization task for catenary images,the combination of SSD and MobileNet is adopted to achieve fast localization of seven key regions on catenary images.Secondly,the key components are precisely localized in different key regions by object segmentation algorithm: Mask R-CNN.At the same time,the finely localized images are aligned using the post-parting posture information,which is convenient for the application of subsequent algorithms.Then,for the images with low resolution and failure of focusing,the thesis proposes to use super-resolution algorithm to deblur the fuzzy images.And the object segmentation algorithm is used to segment the nuts and cotter pins in the fine localized image for state detection task.Finally,this thesis proposes to adopt sparse coding method,by using only normal catenary images,to achieve detection of a variety of unknown foreign objects.The methods in this paper have a positive meaning of improving the speed and efficiency,reducing the complexity of the current algorithm system for catenary automatic detection.Through the actual line data experiment results,it is proved that the methods in the article have strong feasibility and versatility.