Research On Automatic Reading Technology Of Substation Instrumentation Based On Machine Vision

Instrument is an important device in a transformer substation.It is a source of various parameter data of the substation and an important tool for data measurement and monitoring.However,current transformer substations still require a large amount of manual instrument data collection.This collection method consumes a lot of manpower and material resources,and due to the harsh working environment of some substation instruments,manual reading cannot be performed.In order to solve the above problems,the paper studies the automatic recognition technology of substation instruments based on machine vision.At present,related research mainly focuses on instrument contour extraction,pointer detection,and extraction of pointer line segment features,mainly based on traditional image processing methods,and dial scale recognition depends on calibration parameters on the template in advance.When the inspection robot travels to When there is a deviation in the position of the observation point or the deflection angle of the gimbal,the perspective distortion may cause a large error.The automatic reading system of the substation instrument that has been put into use still depends on manual labeling to a certain extent,and still requires a lot of manpower.In order to more accurately and automatically recognize the readings of the pointer meter,the paper conducts in-depth research on the automatic meter reading technology based on deep learning technology.In view of the problems of angle tilt and blur of the meter in the scene,the paper divides the automatic reading process of the meter into three steps: meter dial detection,dial registration,and pointer detection.The calibration of the front template image assists the pointer and scale positioning.In order to solve the problem of meter tilt and blur,this paper analyzes the design ideas of the current mainstream image matching methods,including traditional SIFT feature matching methods,image matching methods Hard Net and D2-Net based on convolutional neural networks,and analyzes these in combination with experiments.The performance of the image matching method in dial registration tasks.Experiments show that D2-Net performs best in actual dial registration tasks.The thesis improves the key point detection algorithm of the D2-Net network,using the sum of the local proportions of the characteristic response as the key points score,and uses the window method to detect sparse key points,making it more suitable for application in dial registration tasks.Increasing the registration success rate while reducing the calculation time.The paper analyzes the dial detection algorithm based on Faster R-CNN network.In the self-made dial detection data set,m AP reaches 92.7%.The paper designs a network DM-Net combining dial detection and dial registration.The feature extraction part of dial registration network and dial detection network is used to share parameters,so that a single network model can complete the two tasks of dial detection and dial registration.In this way,the network reduces computational redundancy and further reduces time consumption.Based on the calibration of the frontal template image and a robust dial detection and registration algorithm,this paper proposes a pointer detection algorithm that combines polar coordinate transformation with a small target detection network YOLOv3,eliminating interference in most areas of the dial and solving The problem that the traditional pointer detection algorithm has high requirements on the dial image quality is improved,and the stability of the pointer detection algorithm is improved.In the end,the paper based on the video surveillance system realized the automatic reading of substation meters.