Defect Detection Of Catenary Components With Large-scale Difference In High-speed Railways Based On Structural Inference

The catenary support components(CSCs)are the overhead devices along the railway.Because of the bad working environment and frequent violent vibration,the CSCs have become one of the most vulnerable parts in traction power system.They are important to the operatetion and maintenance in high-speed electrified railways.With the rapid development of deep learning,the positioning and detecting efficiency of the CSCs are gradually improved.However,it’s still difficult to accurately locate small components because of the large-scale difference.The other existing intelligent detection methods only focus on some serious faults.Therefore,it is necessary to develop a real-time and intelligent detection method to locate and detect faults of the large-scale CSCs.This thesis presents a vision-based method related to deep learning technology to detect CSCs conditions of high-speed railway,which is based on multiple image processing methods and convolutional neural network models.First,the datasets of CSCs image are established to locate the CSCs of high-speed railway images.The catenary image clustering algorithm is proposed for the coarse positioning network.Then the coarse positioning network and the fine positioning network are cascaded to form a large-scale difference positioning network,which is CSCNET.Meanwhile,this thesis introduces the structural inference network(SIN)based on the gated recurrent units(GRU).GRU groupes nodes,boundaries and scene features to detect CSCs.Finally,various network models are tested to locate 12 categories of CSCs at the same time.The advantages of the two structure inference network models are compared and analyzed.Second,the fault detection of some CSCs and lines are presented.For the messenger wire supporters,adaptive RANSAC algorithm is proposed to detect the front side,the back side and the balance line.In the suspected crack region,the short line is detected by Beamlet transformation and the local chain searching algorithm.For the various poor line judgments,the different types of hanging lines are devided to apply the projection sampling difference algorithm and projection average gradient algorithm.Finally,the extraction and fault detection of the CSCs is realized under tensorflow and Matlab environment.CSCs in the different angles and backgrounds are analyzed and tested.The experiment results show that the proposed methods for the extraction and detection(CSCNET,SIN,Beamlet algorithm,etc.)have a high precision,efficiency and robustness.