Research On Visual Detection Method Of Power Equipment By UAV Line Patrol

Power line patrol has an important role in maintaining the safe and stable operation of the power system,and can timely understand the operating status of power equipment on the transmission line.Traditional manual power line patrol has low efficiency,accuracy,comprehensiveness and safety.UAVs with safety and flexibility can replace manual power line patrol,which effectively makes up for the lack of manual line patrol.Utilizing UAVs to carry out power line patrol and realize intelligent detection and identification of transmission line power equipment and abnormal conditions is of great significance for the automatic detection,maintenance and management of power systems.In order to solve the problem of detection and identification of power equipment and abnormal conditions in the UAV line patrol,this article first analyzes the UAV line patrol scene and the key technology of power equipment visual detection.For the unmanned aerial vehicle patrol scene,an optimization scheme using a computer vision system is proposed,and a deep learning-based method is used for visual inspection of electric equipment in the unmanned aerial patrol scene.For the key technologies of visual detection,three key technologies about convolutional neural network are analyzed.The structure of the deep feedforward and deep convolutional neural networks is studied and analyzed respectively.Three object detection frameworks of the R-CNN series are described and their advantages and disadvantages are analyzed.Secondly,the sparse expression method in convolutional neural network is studied.An improved convolutional neural network is proposed by introducing the Dropout strategy in the pooling layer,and a dual probability weighted pooling model averaging method is proposed for model prediction.The derivation process of the improvement strategy and the proposed method are described in detail,and other improved methods of the pooling layer are introduced to conduct a comparative analysis of the proposed method and other methods.The results show that the dual probability weighted pooling model average method proposed in this paper has higher testing accuracy,which can effectively prevent overfitting and improve the generalization of convolutional neural network models.Then the classification model for the scene foreground object in the power equipment object detection and recognition task is pre-trained.Researching and constructing FCN and CNN model separately and add improved strategies for power equipment image foreground extraction and extracted foreground image classification.The overall framework of image foreground extraction classification model and transfer learning initialization scheme are described,and a migration step-by-step learning model initialization strategy is proposed.Data augmentation is used to expand the sample.Constructing a data set for training and compare and analyze the test results.The results show that migration step-by-step learning can effectively improve the accuracy of the model,promote better fusion and convergence of the model,and data enhancement can improve the performance of the model,and extracting the foreground can effectively improve the complexity Background power equipment image classification accuracy.This paper chooses the model with the best test result as the pre-trained model.Finally,the operating mechanism of Faster R-CNN is studied,witch focus on the research of the foreground target detection process.The various parameter factors affecting the model prediction are analyzed,and the shortcomings of the model are pointed out.An improved strategy is added to the shared feature extraction VGG16 network,and an optimization strategy is introduced to introduce the FPN structure to enhance the model detailed feature detection capability.This article constructs a data set containing power equipment and abnormal conditions images of UAV line patrol and uses pre-trained model initialization for training,comparing and analyzing the test results under different models and various parameters,which proves the effectiveness of the optimized model and setting parameters.It has a good detection effect on power equipment and abnormal conditions,can accurately locate the target power equipment and abnormal points,and accurately identify the target category.