Defect Detection Method For Catenary Parts Based On Few-shot Learning

Catenary is the main equipment of electrified railway,which is used for the power supply.It transmits power to the locomotive through sliding contact with pantograph to ensure the normal operation of electric locomotives.Nevertheless,the long-term complex interaction between catenary and pantograph is easy to lead to various defects of catenary parts,such as loose bolts,damage of parts,etc.In serious cases,it will lead to the interruption of catenary power supply and affect the safety of train operation.At present,the defect detection of catenary equipment mainly depends on the manual judgment of professionals.The workload is large and the labor intensity is high,which is easy to lead to long cycle of detection and possible omission.With the development of image processing and computer vision technology,many scholars have applied it to catenary image defect detection and achieved many research results.Nevertheless,due to the issue of many types of catenary parts,many types of defects and few defective images,the relevant research results are still far from practical application.To address this problem,this thesis proposes a defect detection method for catenary parts based on Few-shot Learning(FSL).Firstly,the target detection algorithm is used to extract the catenary parts,and then the FSL is used to detect the parts defects,so as to improve the intelligence of the detection system.The research of this thesis mainly includes the following aspects:(1)A target detection algorithm based on improved Yolo v4 is proposed.The "coarse to fine" strategy is adopted to realize the accurate positioning of catenary parts in two stages,and the accuracy of the model is improved by combining the Squeeze-andExcition Networks(SENet),so as to make precise detection of small target in catenary images and alleviate the interference caused by complex environment.According to the positional relationship among the parts in the insulator area,the redundant insulator marking frame in the scene of double support is removed,so as to support the defect detection in the next step.In stage of coarse positioning,the average accuracy of this algorithm is 96.22%,which is 2.75% higher than that of Yolo v4.In the comprehensive positioning experiment,the average accuracy of 10 parts reaches 94.91%.(2)A multi-modal fusion method based on visual information and semantic information is proposed.Firstly,Wide Residual Network(WRN)and Word2 Vec network are used to extract visual information and semantic information of catenary images respectively.Subsequently,Variational Auto-Encoder(VAE)is used to project the two kinds of information into the same common space for alignment.Finally,the decoded two kinds of information are fused in proportion to enhance the granularity difference between state change of characteristics in catenary parts defective image.(3)A few-shot defect detection method based on multimodal fusion and Graph Neural Network(GNN)is proposed.The GNN is used to capture the relationship between the fused features to construct an undirected graph,and the defect detection of the input samples is realized by embedding propagation and label propagation.Combined with transfer learning,FSL is extended to the defective classification of catenary parts,which can accelerate the training and improve the generalization performance of the model.At the same time,semi supervised learning based on pseudo tags expands the support set to achieve accurate classification in scenarios of small samples.The average recognition accuracy of this method is 94.6% in 13 defect states of 10 parts.