Research On The Identification Method Of High Voltage Transmission Line Tower Bird’s Nests

With the development of my country’s long-distance power transmission technology,the scale of the power transmission network is getting larger and larger,and there are more and more overhead line towers.Some large birds take advantage of the location of the tower to avoid predators and build their nests on the tower,which brings safety hazards such as short circuit and damage to the power system.Therefore,it is very necessary to detect the bird’s nest of the transmission line tower.In order to explore a target detection algorithm based on deep learning that can quickly and accurately detect and identify a large number of transmission line inspection pictures,this topic mainly studies the transmission line tower bird nest location model based on YOLOv4.The false detection phenomenon proposes an improvement to YOLOv4: replace the 9*9maximum pooling in the SPP module with 9*9 average pooling,which reduces the sensitivity of the algorithm to small targets and makes the algorithm prediction results more stable.In order to improve the recognition accuracy m AP value and recognition speed FPS value of the experiment in this paper,a second improvement to YOLOv4 is proposed: at the end of the YOLOv4 backbone feature extraction network,three large and small feature layers are output,and three attention mechanisms SENet,CBAM and ECA are added respectively.In the end,the attention mechanism with the best improvement effect is CBAM.The main research contributions of this subject are as follows:(1)Combined with the development status of foreign object inspection of UAV transmission lines in foreign countries and the current situation of foreign object identification in transmission lines based on deep learning in China,the background and significance of this research are summarized.The collected high-voltage transmission line tower bird’s nest images are augmented by data enhancement technology,and Label Img is used to mark the bird’s nest to construct a high-voltage transmission line tower bird’s nest dataset.(2)The constructed high-voltage transmission line tower bird’s nest dataset is used for training and detection of the three detection model frameworks Faster RCNN,YOLOv3 and YOLOv4 respectively.Comparing the test results,it is finally obtained that YOLOv4 has the best recognition effect among the above three detection algorithms,with a m AP value of 82.72%and an FPS value of 35 frames/s.Using VGG,Resnet,Mobilenet and CSPdarknet as the backbone feature extraction network of the YOLOv4 detection algorithm respectively,the network with the best test effect is CSPdarknet,the m AP value of the test is 89.34%,and the FPS value is 32 frames/s.(3)The YOLOv4 detection model of the improved SPP module is used to detect the highvoltage transmission line tower bird’s nest dataset of this subject.On the original basis,the m AP is increased by 0.47%,and the FPS is increased by 1 frame/s.The three attention mechanisms of SENet,CBAM and ECA are added to the YOLOv4 structure respectively.The test results show that adding the CBAM attention mechanism has the best effect,with the m AP value reaching 91.02% and the FPS reaching 44 frames/s.In order to verify that the improved YOLOv4 algorithm in this paper has certain robustness and generalization,the influence of different lighting conditions and different shooting distances on the test results is discussed.The test results show that the improved YOLOv4 algorithm in this paper can accurately locate and identify different lighting conditions and different shooting distance conditions.A bird’s nest image of a transmission line tower below.(4)On the basis of the improved model,the COCO dataset is used as the source domain,and the tower bird’s nest image in this paper is used as the target domain for model transfer.The test results show that the FPS value is increased by 12 frames/s and the m AP value is increased after the introduction of transfer learning 0.76%.