Research On Defect Detection Method Of Insulator In Power Line Based On Lightweight Networks

With the rapid development of the industrial economy,the supply demand of electricity consumption is becoming more and more prominent,and the safety and sustainability of power transmission in the power industry has become increasingly focused.As an important insulation part in transmission equipment,insulators in power line are prone to defects in complex environments,which poses great challenges to safe transmission.Existing insulator defect detection methods based on deep learning often rely on drones,but it is difficult to achieve the balance between detection accuracy and model lightweight.In the strategic context of smart grids,it is an urgent problem to be solved.In this paper,the application of lightweight networks to the detection of insulator defects in power lines is carried out in depth.The main research contents are as follows:The insulator data is confidential,the amount of public data is small,and the pixel area occupied by the defective part is too small.Through data enhancement,the insulator defect dataset was amplified,resulting in a total of 2158 images.Insulator defect data is annotated with the help of specialized software.In order to improve the detection effect,a two-stage cascade deep learning inspection framework is used that quickly locates insulators and then accurately detects defects.Under this framework,YOLOv5 n,a lightweight model with fast positioning speed,and Mobile Vi T,a lightweight model with accurate detection,are selected as the basic networks.Aiming at the slow speed of the YOLOv5 n algorithm in the insulator localization stage,a fast insulator localization algorithm of Fast-YOLOv5 n is proposed.Using Shuffle Netv2 network and deep separable convolution,the C3 module of YOLOv5 n backbone network and its feature fusion layer were transformed respectively,which reduced the amount of model calculation.By using the Re LU activation function,the overfitting and slow inference caused by the Si LU activation function of the original network are avoided.Finally,SIo U is used to transform the positioning loss function,which enhances the accuracy of the algorithm.Experiments show that under the premise of model lightweight,the algorithm obtains a good combination of accuracy and speed.For the task of accurate defect detection,the accurate defect detection algorithm of TR-Mobile Vi T is proposed.In the feature fusion stage,it is proposed to replace the traditional two-feature fusion with three-scale feature fusion.In the stage of three-scale feature fusion,based on the semantic information that is easy to lose in small-scale feature upsampling,the traditional sampling method is replaced by transpose convolution and the modified attention mechanism module EDCA is added to enhance the learning ability of the network.Experiments show that the algorithm has good accuracy and achieves the balance between accuracy and lightweight.