Research On Transmission Line Equipment Identification System Based On Deep Learning

With the construction of UHV and Strong Smart Grids,the difficulty of operation and maintenance of transmission lines has gradually increased.Due to multiple conditions such as geographic location and light intensity,traditional manual inspection methods lack the ability of guaranteeing the identification of foreign matters in transmission lines.Moreover,it has been unable to accurately detect of power components.Now depth study and artificial intelligence are on the rise.It has become the main trend of smart grid power inspection to use UAV in realizing the identification of foreign matters in transmission lines and automatic detection of important components.In this thesis,the related research is based on convolution neural network.It also combines with deep learning for UAV aerial image multi-scale small target detection algorithm.The main tasks are as following:Firstly,this thesis compares and analyzes VGG16 and Res Net basic networks.And the Res Net101 basic network is selected as the best model.On this basis,the article focuses on two algorithms: Faster R-CNN and Mask R-CNN.At the same time,VGG16,Res Net50,and Res Net101 feature extraction networks are compared.Aiming at the problems of low recognition accuracy and poor robustness in the power grid inspection,this method is used to identify and detect five different targets in the transmission line,including insulators,anti-vibration hammers,and bird nests.The method is based on the Faster R-CNN algorithm.In deep learning,the richer the sample set is,the better the detection effect is.Therefore,this thesis studies an enlarged sample algorithm to achieve expansion of the hidden danger sample set in batch.The test results prove that through sample augment,the detection network can accurately realize the detection of multiple targets and small targets in the background of complex aerial photography.It concurrently improves the accuracy of the detection of power components of aerial transmission lines.In this thesis,by modifying the Mask R-CNN anchor point frame structure,the accuracy of mask prediction is further improved.And the semantic segmentation of insulator strings under complex background is realized.The identification of three types of aerial insulators under complex backgrounds is achieved by constructing deep learning environment.The extraction network is to compare and analyze the two features of Res Net101 and Res Net50 throughexperiments.The test results show that the method proposed in this thesis can accurately identify and locate insulator strings in complex environments.