Research And Implementation Of Automatic Recognition System For Substation Analog Instrument Based On Machine Vision

With the advancement of technology and the state electric power system reform in depth,automation and informatization of substation equipment have been developed rapidly.Most electrical signal data can be collected and recorded in real time by monitoring equipment.Part of the unmanned running state has been implemented.There is still a problem that the inspection data of instruments and meters monitoring equipment in a relatively special environment cannot be obtained immediately.However,using robots instead of manpower can greatly improve the production efficiency.It plays an important role in obtaining a large amount of information,executing equipment warning and defect detection,etc.In this paper,the detection and identification algorithms of instrument state or instrument reading are given respectively for switch state analog instruments and single pointer analog instruments with complex dial plate which are existing in substations.Because the images collected by the equipment mainly consists of environmental background in the actual scene,meter is not the main content of dial plate image.Therefore,it is necessary to study the meter positioning and extraction of the instrument before reading.The main research contents of this paper include the following:(1)Basic requirement analysis of substation instrument identification system,evaluation method of instrument positioning and reading accuracy,and system software and hardware design.Because the instrument is in a special environment,many areas cannot obtain a good lighting effect.Therefore,while increasing the illumination mode,the influence of different illumination modes on instrument is analyzed.Software system and algorithm design are the core issues of this paper.In the composition of the software system,the image acquisition module,the instrument positioning module,the instrument reading module and the display interface design are respectively built in a modular design mode.(2)Aiming at the problem of instrument positioning,images of the two types of instruments are collected for manual labeling and training sets and test sets are made.The Whitening operation is used to reduce the complexity of the image,and then the modified semantic segmentation network is used for training to obtain the network model.The model is analyzed and evaluated by TensorBoard,and instruments in various environments are detected and positioned.The experimental results show that the network can locate the instrument well and extract the complete meter area,with good anti-interference ability.(3)Identification algorithms are proposed for the two instruments respectively.Switch state analog instrument uses its structure and color information to extract color features and uses them as the basis for identification.However,pointer meters have special dial,Haar-Like feature is proposed to construct the filter,and then filter out instrument background.Afterwards carry out expansion projection correction to improve the identification accuracy of instrument scale.Finally,read result according to the distance ratio.Reading experiments show that this method can effectively identify the number of single pointer meter.