| Smart grid is the inevitable trend of the development of power grid.It can improve the security and reliability of power system effectively and realize the sustainable development of power grid.Intelligent inspection robots have been gradually applied to the maintenance and management of intelligent substations,replacing manual table reading in the detection and identification link of pointer meters,ensuring the safety of inspection personnel and reducing the total cost of data governance.At present,when the inspection robot processes the instrument image collected,it still needs to manually preset the template or manually adjust the parameters to complete the inspection,which is difficult to achieve the goal of intelligent substation.At the same time,due to the complex environment of pointer meters,various types of meters,and the uneven illumination,tilt,occlusion and other problems in the meter image taken by the inspection robot,the meter reading is greatly affected.In view of various problems in the identification of pointer meters,this thesis proposes an automatic reading algorithm of pointer meters in substation based on deep learning,which reduces the reading errors caused by uneven illumination,tilt,occlusion and other problems.The main research work of this thesis is as follows:(1)The technical scheme of instrument panel detection and positioning,instrument digital detection and recognition,instrument information recovery and indicator reading is put forward.A practical and stable automatic reading method of pointer type instrument is realized,which improves the universality of the algorithm.Without manual intervention and prior information,the meter reading can be read directly by inputting the image to be recognized,which solves the problems of complex environment,uneven illumination,meter tilt,meter occlusion and multiple types of meters in the automatic reading of the pointer meter in substation,thus effectively improving the accuracy and reliability of the automatic reading of the pointer meter in substation.(2)The instrument panel detection and recognition algorithm based on YOLOv5 model is studied,which can accurately detect the target instrument in the substation background.This algorithm processes the detected pointer meter according to the position of the detection box in the current field of vision,so as to get the pointer meter image which removes the complex background and makes its features more obvious.Then these images are used as a new data set to improve the identification accuracy of the instrument digital detection.(3)The digital detection and recognition algorithm of the dashboard,which combines the YOLOv5 target detection model with the improved CRNN text recognition algorithm,is studied and designed to obtain the digital information on the dashboard.To meet the needs of portability,a lightweight network was used to replace the backbone network,and then the network was optimized by modifying the feature map resolution and online quantization operation of parametric shear activation.The model size is reduced by 58.27%,the accuracy is increased by 8.82%,and the reasoning time is accelerated by 2.78%.(4)The instrument information recovery method under complex substation environment is studied.Image enhancement processing is carried out for instrument images with uneven illumination.In order to solve the problem of tilt caused by poor shooting Angle,perspective transformation method based on the number of scale line is adopted.To solve the occlusion problem,a circle fitting method based on least square method is proposed to determine the center position of the instrument panel.By processing the instrument image,the information of pointer position and center position on the instrument panel is recovered,and the reading accuracy is improved.(5)A pointer indicator reading algorithm is designed.The reading method of pointer instrument is studied,the Hough Lines P method in Hough linear detection method is used to locate the pointer position,and finally the Angle method is used to read the pointer indicator number.After the implementation of the algorithm,the method in this thesis has been verified that it is not only suitable for various types of instruments,but also has a very stable performance while ensuring the identification accuracy.The maximum relative error of reading is less than 1.5%,which meets the requirements of high identification accuracy.The average time of the whole process of reading the instrument information is 716 ms,and the speed can also meet the needs. |
