| In recent years,high-voltage transmission inspection has begun to be completed by taking aerial photos of power system through UAV and manually identifying the abnormalities or obstacles of key components.At present,researchers are studying the automatic object detection of UAV aerial photography,that is,the detection is realized at the same time of aerial photographing.Image and video object recognition is the key technology,so automatic recognition has become the main direction of the development of power automatic inspection system,and it is the research focus and difficulty of smart grid inspection.As a new technology in the field of artificial intelligence,deep learning has many advantages.It has been applied to practical work fields,including automatic detection of power system,and has obtained some results.Based on the deep learning method,this thesis proposes a detection model for electronic components such as insulators based on YOLOv2 network,which uses prior information and experiments to determine the size of image blocks,reduces the number of convolution layers,and reduces operation time;Using multiple approximate objects,such as the correlation of multiple insulators,the design formula comprehensively calculates the confidence of the insulators,and optimizes the model for occlusion and light and shade changes.When the number of samples in the training set is 50%of the total number of samples in the database,the highest recognition accuracy of insulators can reach 99%.After further improving the network structure and adjusting the network parameters,it has a high recognition rate for all kinds of objects.In terms of the pin identification of tower nut,A pin fault recognition model based on YOLO network and dictionary learning is proposed.Image blocks containing nuts and pins are extracted from the original image through YOLO network,and then the image blocks are identified through dictionary learning.sort.The initial data set is annotated by labeling software and manual selection methods,and the effectiveness of the improved YOLO network and dictionary learning method in the identification of missing pins is verified.Through the comparative analysis of the improved YOLO network,the three-scale output YOLO v3 network,and the Fast R-CNN network,the influence of the number of training samples on the recognition accuracy of untrained samples is studied.Experimental results show that various metrics in this thesis outperform other methods. |
