Research On Detection And Recognition Of High-voltage Switchgear Equipment Based On Machine Vision

The power system plays an indispensable role in industrial producing and our daily life,for its importance in power generating and power transmission.With the development of modern society,the demand of energy is increasing as well.Recently,"smart grid" has been proposed to ensure the normal manipulation of the power system and improve the efficiency of power distribution.In order to build the strong smart grid,the operating information and user demand should be gathered in the first place.As an important hub of the power system,substation is a basic unit for collecting the information of electric power and controlling the system with a large number of power equipment used for monitoring as well as protecting,hence the substation need regular inspections for daily maintenance.Most of the inspection tasks are usually done manually by the staff,which is not only inefficient and easy to make error,but it is also very dangerous to work in such high-voltage places for a long time.Nowadays,a variety of solutions for power system inspection are proposed in the industry,which are mainly realized by automatic inspection robots.However,those systems still have some limitations.Firstly,most of these robots are equipped with complex control systems and sensors,which greatly increases the cost of the robot.Secondly,state recognition methods are usually implemented by some traditional image processing and computer vision algorithms,which may not be suitable to deal with the complex environments and equipments with rich types and shapes.Aiming at the inspection task of high-voltage switchgear equipment in substations,this paper uses low-quality images obtained from the surveillance camera in the high-voltage room to try to complete detection tasks and recognition tasks using deep learning methods.The main work of this paper are as follows:(1)Inspired by the task of moving object detection,We propose a foreground segmentation method based on deep neural networks for high-voltage switch cabinet foreign object detection.In view of the difficulty of obtaining the image data of foreign body occlusion,the high-voltage switch cabinet foreign object detection datasets are constructed through the synthesis algorithm for training and validating of the model,with the switch cabinet images captured in real scene used as the background images and the object images intercepted from the open-source segmentation dataset used as the foreground images.The proposed method is proven to be promising to solve the foreground object segmentation in the substation,by testing the method on several experiments and applying it to real scenes.(2)Aiming at the task of identifying the status of the protection plates on the switchgear,and in view of the low poor quality of the plate images obtained from the surveillance camera,a lite image classification model based on convolutional neural network is constructed to identify the plate state.In order to solve the problem that it is difficult to obtain sufficient plate images with rich states from online switchgear frames,the disassembled plate images are used as the material to build artificial plate datasets for model training.In addition,to improve the accuracy of model transferred to the real scenes,an unsupervised domain adaptation method called Maximum Classifier Discrepancy is used to perform adversarial learning for the model.Experiment results show that this method can effectively identify the states of plates in the real scenes with high accuracy.(3)An intelligent high-voltage switchgear inspection system is constructed depending on the actual demand of substation.The system can automatically perform inspection tasks and push the results to servers,which has been deployed and applied in State Grid Sichuan Electric Power Company Chengdu power supply company 220 k V Sansheng substation.