Research On Insulator Identification And Damage Detection Of Lightweight Networks

With the construction of UHV in our country,the transmission mileage is also increasing.Insulators are important component of support and electrical insulation in transmission lines,but they work in harsh environments and are prone to various faults,this bring great hidden dangers to the safety and stability of transmission lines.Therefore,real-time detection of insulator status is especially necessary.In this paper,insulators and damage detection methods in transmission lines are taken as the research object.On the basis of the original YOLOv4 algorithm,lightweight compression and precision improvement are performed to achieve real-time detection of insulators and damaged targets at the edge.The main work of this paper is as follows:（1）Aiming at the large amount of parameters and computation of the YOLOv4,contrast and study two schemes of pruning compression and lightweight network replacement for lightweight improvement The experimental results show that the parameter compression of the two models reaches 90% and 73%.（2）Aiming at the problem of model accuracy degradation caused by model pruning,a self-attention mechanism that integrates feature information of different dimensions and a channel and spatial attention mechanism that amplifies target features are introduced.In the horizontal comparison,two improved methods for improving the insulator and the damage target were selected for fusion use,which improved the identification accuracy of the insulator and its damage.（3）For the lightweight improvement scheme of structural redesign,Dense Net is used to replace the backbone network to reduce the amount of model parameters,the lightweight attention model SA-Net is embedded in the feature extraction network to improve the target recognition ability,and shallow layers are added to the target detection module.Feature map to enhance the model’s ability to recognize small damaged targets.The experimental results show that this scheme can still improve the accuracy of the model while reducing the amount of model parameters.