| With the rapid development of technology in the field of high-speed trains in China,the demand for online intelligent maintenance technology for power system and other infrastructure is more urgent.The abnormal stress and cracked dropper of the high-speed rail catenary will cause the poor energy absorption of locomotive,resulting in large energy loss and damage to the pantograph and catenary,thus posing a threat to the safety of the high-speed railway.The intelligent detection technology of machine vision simulates human visual functions,receives and processes image data,which is suitable for the research on the detection of the dropper defects.This thesis aims at the high-speed rail catenary dropper structure,starting from the existing methods of each function,the framework of detection model algorithm is defined,and the improved LeNet network for the catenary fault diagnosis is proposed.Also,the effect of several positioning and fault diagnosis methods is evaluated through experimental comparison.The specific contents are as follows:(1)The research status of catenary detection system in fault diagnosis technology and convolution neural network in the direction of target detection at home and abroad is reviewed in this thesis,and the prospect and direction of research in the field of dropper wire defect detection are discussed.(2)Aiming at the structure of the dropper,the existing methods of the three detection functions of dropper image preprocessing,dropper positioning and dropper wire defect diagnosis are analyzed.Aiming at deep learning methods,several mainstream target detection and several image classification network structures,working principles are proposed to provide a theoretical basis for the thesis research and related algorithms.(3)The Fire module is used to replace the C3 convolutional layer in the LeNet network for improvement,then the detection of the dropper wire defects is made via this network.Also,the optimization of the Fire module is proceed in the improved LeNet network,the impact of the 3?3 convolutional layer and the proportion of the Squeeze layer in the module on the network performance is explored;the influence of variation in the number of convolutional layers on the network performance is also demonstrated,finally,the effect of LeNet network before and after improvement is contrasted for highlighting the progressiveness of the improvement LeNet network.The results show that for the data set in this paper,when the number of convolutions in the extruded layer,the extended 1?1convolution layer and the 3?3 convolution layer is 24,16 and 16 respectively,the Fire module has the best effect in the LeNet network.The improved LeNet network compresses the model to 1 / 10 of the original LeNet network,and the accuracy is 5.7 percent higher than that as well.(4)In view of the specific function of dropper defect detection,the basic steps and specific framework of the dropper positioning and wire fault diagnosis model are summarized.Combined with the preliminary processing effects of typical dopper images on each function,the details of the dropper positioning and wire fault diagnosis model are discussed.The specific contents are as follows: presenting the pre-processing effect of dropper images based on the pre-processing theory;selecting the depth learning object detection method based on the preliminary application effect of several positioning methods;selecting the depth learning image classification method based on the preliminary application effect of dropper’s wire detection.(5)The experimental comparison is made on the dropper positioning effects by using five mainstream models of Faster R-CNN,YOLO v3,R-FCN,Corner Net-Squeeze and Corner Net-Saccade.The results show that Corner Net-Saccade has the highest positioning accuracy(99.1%).Corner Net-Squeeze maintains good detection accuracy(94.7%)at a high calculation speed.At the same time,the performance of improved LeNet is explored,the Mobile Net v2 and Res Net-34 are introduced as a comparison to realize the defects detection and fault classification of the dropper wire.The results show that the improved LeNet has most significant balance between the accuracy and speed in the detection of droppers.Therefore,the combined detection method of Corner Net-Saccade and improved LeNet has a significant effect,with an accuracy of 95.0%;Corner Net-Squeeze and improved LeNet combined detection methods in the test set can realize real-time detection,which an accuracy rate of 90.7%. |
