Research On The Application Of Lightweight Optimization Network In Substation Equipment Defect Detection

The stable operation of substation equipment is an important guarantee for the safety of the power grid,so the regular inspection of equipment is an essential task.Due to the disadvantages of poor efficiency and high risk of manual inspection,with the continuous promotion of smart grid,the combination of robot inspection and computer vision technology,the intelligent inspection method of using inspection equipment equipped with target detection model can effectively realize the defect detection of substation equipment,making the inspection more efficient and safe.The generalized target detection model not only lacks accuracy,but also has a large number of parameters,which is difficult to deploy on the substation equipment inspection equipment with low arithmetic power.Therefore,this paper is oriented to design a lightweight and efficient detection network and combine the characteristics of defect datasets.This paper first analyzes the common defects of substation equipment and creates the infrared dataset IFDS and the visible dataset VFDS according to the characteristics of defects,and then proposes an improved network MSD-YOLOv3 for the problems of high occlusion in IFDS images,large size of targets to be detected,and difficulty of detection.The network uses Mobile Net V3 to replace the original YOLOv3 backbone network to complete the lightweight improvement;uses the K-means algorithm to re-cluster the anchor frames to ensure that the candidate frames are close to the substation equipment to be identified;introduces the SPP module to improve the perceptual field of the model to achieve the detection of multi-size targets;and adds the Drop Block module to optimize the model fitting situation.Compared with the preoptimization,MSD-YOLOv3 not only improves the m AP by 2.18% on the IFDS dataset,but also decreases the model size by 84.45% to a minimum of 36.5 M.In addition,this paper proposes a method to determine the thermal defect level of substation equipment based on the infrared MSD-YOLOv3 model.The thermal defect areas detected by using the MSD-YOLOv3 network are extracted with grayscale values,the temperature values of the thermal defect areas are fitted using the grayscale-temperature formula,and specific diagnostic results are given according to the thermal defect criterion.In this paper,the thermal defect level diagnosis of disconnect switches is carried out by this method,and the results verify the feasibility of the method.Different from the optimization strategy of MSD-YOLOv3,SG-YOLOv3 performs the following optimization for VFDS image features: based on Shuffle Net V2 network,incorporating SE attention module in its basic unit to enhance the extraction of effective features,combining with deep convolutional network to complete the lightweight design of Shuffle Net V2 network,using GIo U as border regression loss function to improve the detection accuracy.Experiments on the VFDS dataset show that the SG-YOLOv3 model m AP can reach 89.42% and the model size is maintained at 15.5M.In this paper,a cell phone is used to design a substation equipment defect detection APP in order to verify the detection effect of the improved lightweight detection network on low arithmetic devices.This allows the lightweight network to have a practical application in the work of substation equipment inspection.Figure [52],Table [12],Reference [73]...