| As an important part of the electric catenary,the long-term operation of the overhead contact system(OCS)is directly affected by the high frequency of electric contact system(OCS),and these parts will inevitably fall off,fracture and other faults.These hidden dangers will seriously affect the stable operation of high-speed trains,so it is necessary to find and replace these defective parts in time.At present,the catenary suspension state detection and monitoring device is mainly used in China to inspect these devices by means of high-definition camera,and manual analysis is used to process the image data generated in this process.Due to the large amount of image data,manual analysis has the problems of low efficiency and low accuracy.In view of the above problems,this paper takes catenary suspension device parts as the research object,based on deep learning and image processing technology.In this paper,the accurate location of components in OCS and the automatic fault identification method are studied and improved.The main contents of this paper are as follows:(1)Firstly,the detection system of catenary suspension device and its characteristics are studied.It is clear that the image has the characteristics of low illumination,high noise and some unavailability.For the overexposed images and background images in the sample set,a data cleaning method using ResNet50 deep network is proposed.Experiments show that the method can effectively eliminate the images that do not meet the detection requirements.After the comparison of different schemes,the data preprocessing process of data cleaning,image noise reduction and contrast enhancement is designed,which is prepared for the data set production and the establishment of defect automatic detection model.(2)This paper compares the most commonly used target detection algorithms,analyzes the characteristics of the target in the scene image,improves on the basis of yolov5 algorithm,and puts forward a model S-YOLOv5 which is more suitable for small target detection.Firstly,based on the idea of feature fusion,this paper proposes a model S-YOLOv5 which is more suitable for small target detection,in the neck of the network structure,bi-directional feature pyramid(BiFPN)structure is used to replace the original feature pyramid(FPN)structure and path aggregation network(PANet)structure to improve the detection ability of small targets;Secondly,we try to introduce different attention mechanisms into the model and compare them.Finally,we choose to add the convolutional attention module(CBAM)to the backbone network of YOLOv5.Through the test and comparison of Faster R-CNN,YOLOv5 and S-YOLOv5 models on actual data sets,it is proved that the improved target detection algorithm has advantages in average detection accuracy and small target detection ability.(3)In this paper,a cascaded network of catenary parts defect detection is proposed,which cascades the S-YOLOv5 network with ResNet50 network.After locating the catenary components,the ResNet50 model with L2 regularization term is used to identify the working state.The experimental results show that the improved cascaded network can quickly and accurately judge the various states of components in the catenary.Figures 47;Tables 4;References 60... |
