Research On Image Detection Method For Surface Defects Of Catenary Insulators Based On Deep Learning

The maintenance of the catenary is related to the safety of railway operations.In order to improve the reliability of railway operations,the China Railway Corporation has developed a high-speed railway power supply safety inspection and monitoring system(6C system).In the6 C system,the catenary suspension state detection and monitoring device(4C device)that has been put into operation uses manual inspection and traditional template matching methods to detect the state of the catenary components,which is difficult to meet the demand for real-time detection of the catenary.In recent years,target recognition methods based on deep learning have developed rapidly.This thesis uses the Faster Region Convolutional Neural Networks(Faster R-CNN)algorithm under the deep learning framework to complete the precise and rapid positioning of the insulators.Due to the timely maintenance of insulator faults and the scarcity of defective samples,an improved Deep Convolutional Generative Adversarial Networks(DCGAN)is used to expand the defective samples of insulators.In the insulator defect detection,the improved deep learning target detection algorithm Mask Region with Convolutional Neural Networks(Mask R-CNN)is used to identify the insulator defects,and realize the multi-target detection of the insulator surface defects.First,a series of deep learning target detection algorithms for YOLO,SSD,R-CNN are introduced respectively,the advantages and disadvantages of each algorithm are analyzed and summarized.The Faster R-CNN algorithm is selected as the network structure for insulator positioning in this article,and the Mask R-CNN algorithm is selected as the network structure of insulator defect detection in this thesis.Secondly,according to the relatively single feature of the catenary insulator,a simple and efficient VGG16 network is selected to extract the insulator features,and the insulator positioning is completed under the Faster R-CNN framework.By comparing and analyzing the experimental results of current mainstream deep learning algorithms,the feasibility and effectiveness of the positioning algorithm in this thesis are verified.Then,in view of the problem of the shortage of defect samples of catenary insulators,a deep convolution generation confrontation network is selected to expand the defect samples of insulators.Combined with the shortcomings of the DCGAN algorithm used in the expansion of insulator defect samples,the problem of unstable model training,disappearance of gradients and low resolution of insulators generated under unsupervised conditions is solved by improving the generation network and the discriminant network.Finally,combined with the characteristics of insulator defects,the Mask R-CNN algorithm with mask function is used to identify insulator defects.In this thesis,the Mask R-CNN algorithm uses an improved Feature Pyramid Networks(FPN)to make full use of all scale feature information and improve the accuracy of insulator defect detection.Furthermore,the non-maximum value soft suppression algorithm(Soft-NMS)is used to replace the non-maximum value suppression algorithm(NMS)to improve the detection performance of insulator defects.Finally,through comparative experiments,the results are as follows: the m AP value of this algorithm when the Io U threshold is 0.5 is increased by 6.2%,and the m AP value when the Io U threshold is 0.7 is increased by 7.2%,and the detection accuracy has been greatly improved.The experimental results show that the image-based detection method for surface defects of catenary insulators based on deep learning in this thesis is feasible and effective,and can be extended to other parts of the catenary,thereby improving the accuracy and efficiency of catenary detection.