| The safety of the power grid contributes to the steady development of the social economy.With the continuous improvement of the construction of ultra-high voltage,long-distance and large-capacity transmission lines in China,how to ensure the safe operation of transmission lines in complex environments and timely eliminate safety hazards is the line.A major problem facing inspections.Traditional manual inspections are gradually being replaced by aircraft inspection methods represented by drones because of their low efficiency and heavy workload.With the large-scale application of the aircraft in the power grid,a large number of aerial images will be generated.How to identify and analyze the defects of these aerial images with intelligence,high efficiency and high reliability is a problem to be solved by the transmission line machine.At present,the intelligent identification and analysis method of the aircraft patrol image mainly extracts the defect characteristics of the line image through computer vision technology,and then judges whether there is a fault or not.However,the existing methods have the following problems:(1)Manual adjustment of parameters is required,and any image needs The parameters of the artificial selection feature cannot be adapted to the processing of massive images.(2)The detection accuracy is low,and it is prone to misjudgment and missed judgment,which cannot meet the requirements of the refined inspection of the transmission line.In view of the above problems,this paper combines computer vision and deep learning technology to propose a defect detection model specifically for power line patrol-Power Net,and does the following work:(1)Analyze two classic methods in the current object detection field: Faster RCNN model and YOLO model,analyze and reconstruct the model from the perspectives of model architecture,feature extraction mechanism and object detection implementation steps,and use actual inspection image training.Compared with the test model,the performance of digital image processing method,machine learning method and deep learning method are compared.The experimental results show that the image processing method and the machine learning method have the problem of model generalization,while the Faster RCNN and YOLO have the detection speed and detection respectively.The problem of poor precision is difficult to meet the quasi-real-time and refined inspection requirements of the grid.(2)Optimize the accuracy and speed of detection to build the Power Net model.At the accuracy level,Power Net uses the context information fusion architecture to fuse the shallow feature map and the depth feature map to reduce information loss and improve small-scale object detection.Accuracy;At the speed level,Power Net uses volume integral solution techniques to decompose standard convolution calculations into deep convolution and point-by-point convolution,reducing the computational complexity of the model.(3)Re-establish RCNN,Fast RCNN,Faster RCNN,YOLO,SSD and Power Net models,and use the same test data to compare and analyze the detection performance of several types of models.The results verify the full-time of the proposed transmission line for the transmission line.Type object detection model-Power Net is superior to other models in terms of speed and accuracy of detection,and is more suitable for the identification of defect image of transmission line machine.The object detection model dedicated to the transmission line patrol,Power Net,can quickly and accurately detect the defects of the aerial image of the transmission line,help to improve the efficiency of the transmission line patrol,eliminate safety hazards in time,and ensure the transmission line safe. |
