Intelligent Natural Scene Substation Instrument Reading System

At present,with the rapid development of China’s power grid automation,the status detection of substation distribution equipment has gradually changed from manual inspection and manual recording to intelligent and automatic monitoring,so as to realize the intelligent identification and real-time warning of distribution equipment.In substation power distribution equipment,pointer instrument is widely used because of its simple structure,good electromagnetic resistance and low price.Therefore,many domestic and foreign research institutions have studied the reading recognition of pointer type instruments.However,in the main theoretical research stage of its recognition,only some or no consideration has been given to the various factors existing in the actual application scene,such as the low area of the instrument panel in the image,the deformation of the instrument panel,the exposure,the dim brightness,the complex background of the instrument panel and so on.Based on the above considerations,this paper develops an intelligent reading recognition system for substation instruments,in order to realize the clear,accurate and comprehensive reading of hidden dangers.In this paper,the reading recognition model of the system firstly uses the instrument panel positioning module to locate the instrument panel,then uses the distortion correction module to deal with the deformation of the instrument panel,and finally uses the pointer feature extraction module to obtain the pointer feature,and uses the angle method to convert it into the final reading.The instrument panel positioning module mainly solves the problem of low proportion instrument panel positioning.The module is based on the two-stage Faster R-CNN target detection algorithm,and improves the VGG16 convolutional neural network used in the feature extraction module to ShufflenetV2 convolutional neural network,so as to reduce the hardware cost of storage and display card,and accelerate the reasoning speed.In addition,in order to simplify the iterative optimization process after the system landing deployment,TensorFlow-Serving technology is introduced to realize the deployment of dashboard positioning model.Distortion correction uses traditional image processing,innovatively obtains the calibration points needed for perspective transformation through the C-shaped black ring in the instrument panel,and according to the acquisition feedback,applies image enhancement technologies such as brightening and reducing exposure.The pointer feature extraction module abandons the traditional Hough line method to obtain the pointer feature,and improves it based on the silhouette template method.It constructs a digital matrix as the template,and finally obtains the pointer feature through the matrix point multiplication operation.In addition,the software part of the system is developed based on C/S architecture.The client can collect real-time data or offline data,and interact with the server through the interface version or API version.The server interacts with the interface based on Flash,uses MySQL database for historical analysis and data persistence,and joins the log system to record the system running log.Finally,in order to verify the effectiveness of the identification model,we construct a pointer instrument data set in the real substation environment as the main data set,and use a public data set as the auxiliary data set for the pre training of the instrument panel positioning module model.On the main data set,the effectiveness of the proposed model is verified by comparing the instrument panel positioning module with the instrument reading recognition module,and the effectiveness of each component module of the instrument reading recognition module is verified by ablation experiments.