Research On Anti-vibration Hammer Detection Algorithm Of Overhead Transmission Line Based On Deep Learning

With the increasing demand of social development for power resources,the strategic planning of the power industry tends to the development of intelligent power grid.Overhead transmission lines play an important role of the power grid,and the intelligent daily inspection of them is also a key link in the intelligentization of the power grid.The anti-vibration hammer is an electrical equipment on the overhead transmission line.Its main function is to prevent the vibration caused by the strong wind pulling the transmission line in the natural environment,and prevent the transmission line from being damaged due to periodic vibration.In recent years,the commonly used daily inspection method of transmission lines is drone aerial photography.Accurately detecting and locating the anti-vibration hammer on the transmission line through the aerial image of the UAV is conducive to grasping its normal operation status,and can also find its various defects,respond to the defects in time and take measures,which is significance for the normal operation of the power grid.However,the image collected by UAV aerial photography has complex background,small object and wide field of vision,which is a major difficulty in anti-vibration hammer detection.Based on this background,this paper investigates related research at home and abroad,compares the traditional object detection algorithm and the object detection algorithm based on deep learning,selects the FCOS algorithm as the basic framework,and proposes two improved methods:1.A method of anti-vibration hammer detection based on moment pooling spatial attention is proposed.Aiming at the problems of small target and complex background of anti-vibration hammer in aerial images,a space-based moment pooling attention mechanism is proposed.Each feature point is regarded as an independent random variable,and its random distribution form is uniquely described based on the combination of the central moments of each order.By improving the weight distribution of the spatial domain,the feature expression can be more accurately reflected in the feature extraction of the convolutional neural network and the detection accuracy can be improved.Experimental results on VOC2007 data set constructed from aerial anti-vibration hammer images collected under natural conditions show that compared with the mainstream attention mechanism,the proposed method greatly improves the detection accuracy of anti-vibration hammer without increasing the number of parameters and operations.2.An anti-vibration hammer detection method based on dynamic adaptive weight adjustment of loss function is proposed.Different from most methods to improve the feature expression of convolutional neural networks,this method focuses on adjusting the weight relation among tasks of the loss function.In this method,vibration hammer detection is regarded as a special case of multi-task learning,and the training process of FCOS detection network is controlled by dynamically adjusting the weight of classification loss and regression loss,so as to improve the detection accuracy of anti-vibration hammer.The loss function is adjusted by adding dynamic weight generation structure after feature extraction network and generating weight based on input information.The experimental results show that the detection accuracy of the anti-vibration hammer is higher than that of other weight adjustment methods without changing the test model.The thesis includes 31 figures,6 tables and 71 references.