Deep Learning For Risk Feature Intelligent Identification Of Distribution Network Facilities

The rapid development of China Power Grid is increasing the requirements of intelligentize in each section of safe operation of power grid as well as the reduction of labor cost.It is essential to proceed a timely and effective detection to avoid the potential risks of large-scale power supply failure,due to the complexity of equipment composition on grid lines and the close relation of equipment working.The manual testing in grid lines exists problems of panic detection,missed detection and false detection,therefore,decreasing labor cost could be an inexorable trend.Deep learning as a popular technology in recent years,is widely used in various fields.However,the existing intelligent detection method of deep learning mostly uses single-model to perform image target detection,which should be improved the accuracy of recognition and judgement and generalization ability.In addition,the recent intelligent detection method is mostly directed to a certain target device or a certain defect,but cannot detect the overall defect and fault of the power grid line.In this thesis,a deep vision-based intelligent identification technology for power distribution equipment risk characteristics is studied.The image data of the inspection is deeply studied and the classification of risk characteristics is automatically obtained.Firstly,this paper studies the electromagnetic compatibility problem of UAVs and the automatic obstacle avoidance strategy of UAVs in the process of obtaining the image data of the grid equipment in the intelligent inspection work of the power grid,and the inspection plan for image data capture.Designed to obtain high quality inspection image data.Secondly,the research will identify the image data obtained through the intelligent inspection,intelligently label the power equipment in the complex background,and specifically input the obtained power distribution equipment image data into the target recognition depth neural network.The network model obtained through the training can perform target recognition on the image data and crop out the subimage.Furthermore,this paper studies the technology of machine learning to classify the obtained sub-images of power equipment.This paper studies and uses a variety of machine learning models,including deep convolutional neural networks,to obtain different models to classify distribution network equipment through training.As a result,the differences in the results of several models under different iterations of different layers are compared.Finally,this paper studies the use of integrated learning technology to combine the classification results of distribution network equipment features in multiple models,and conducts a risk prediction method for secondary learning,and compares the differences with the classification results of single classifier models.