Research On Intelligent Inspection System For Substation Based On Machine Vision

With the increasing scale of China’s power grid,the number of substations is also gradually increasing.Substation is an important component of the power system,and its safe operation will have a direct impact on the safety and stability of the entire system.At present,substation video surveillance and robot and drone inspections based on safety,reliability,and low cost are gradually replacing traditional manual inspections.With the widespread application of unmanned inspection in power grids,how to use computer vision technology to intelligently,efficiently,and accurately detect it has become an important development direction of intelligent inspection in power grids.In this paper,the main research objects are post insulators and their defects in substations,foreign objects in substations,and workers wearing safety helmets.A substation inspection dataset based on PASCAL VOC2007 format is constructed,and histogram equalization is used to make the dataset more evenly distribute gray and brightness.At the same time,geometric transformation and noise transformation are used to expand the dataset,and Labelimg is used to label and format images.In order to solve the problems of low recognition accuracy,missed detection,and false detection during the recognition process using YOLOv5 s,a method is proposed to replace the loss function of the original network with SIOU loss function,and to optimize the non maximum suppression of the original network using Soft-NMS.The accuracy of the improved method was improved by 5.9% by comparing the performance of the model with ablation experiments on the substation patrol data set,mAP@0.5 Improved by 2.1%,and visual testing was performed on the improved model,resulting in improved confidence levels.This makes it possible to apply the improved YOLOv5 in the intelligent patrol inspection of video surveillance in substations.Aiming at the real-time and accuracy requirements of intelligent patrol inspection,a lightweight YOLOv5 workpiece detection model integrating attention mechanism is proposed.Based on the theory of deep separable convolution,the original Bottlencek module of YOLOv5 was replaced with a MobileNET-v3 module,which reduced model parameters and incorporated the GAM attention mechanism to optimize the detection accuracy of the lightweight model.The characteristics of the dataset were analyzed,and the large target detection header was removed,ensuring detection accuracy while increasing detection speed.The FPS of the lightweight model is 24% higher than the original model.The goal of lightening the model is to make it possible to apply the model to the patrol inspection of small unmanned aerial vehicles.