Research On State Recognition Method Of Substation Equipment Based On Computer Vision

As an important part of power system,it is of great significance to accurately recognize the state of substation equipment for safe and economic operation of power grid.The traditional methods of on-site recognition or analysis of monitoring images have some problems such as high security risk and low efficiency.The image processing and analysis method based on computer vision technology provides a new auxiliary decision-making method for substation equipment state recognition.Aiming at the problem that the accuracy of substation equipment state recognition is limited by the quantity and quality of images,firstly,the general image augmentation method based on rotation and scaling is studied.Secondly,the image augmentation method based on style transfer algorithm is studied.By adjusting the proportion coefficient of style feature and content feature,the stylized image containing both content feature and style image feature of power equipment is generated,and the annotation file of original power equipment image corresponding to the generated image is reused to complete the annotation of position and category of power equipment in the generated image.In addition,an image enhancement method based on Poisson image editing algorithm is studied.This method generates a new image by embedding the target element into the specified position of the image of power equipment.Finally,the image quality evaluation and filtering method based on the deep learning classification model is studied.The method uses the classification algorithm to classify according to the image quality and clean the low-quality images in the dataset.In order to reduce the background complexity of power equipment and improve the accuracy of state recognition,firstly,the locating method of substation equipment based on YOLOv3 algorithm is studied.This method eliminates the background outside the area by detecting and locating the area where the power equipment is located.Secondly,the segmentation method of substation equipment based on semantic segmentation algorithm is studied.This method realizes the segmentation of power equipment pixel and background by classifying pixels according to the predefined semantic categories in the image.In this paper,a method to improve the robustness of semantic segmentation based on stylized images is studied.In this method,stylized images are introduced to enhance the learning of shape features of power equipment by semantic segmentation model and improve the accuracy of semantic segmentation in rain,snow and other corrupt images.Based on the location and segmentation of substation equipment,the method of substation equipment state recognition is studied.Firstly,the state recognition method of substation equipment based on object detection algorithm is studied.This method can realize the state recognition while positioning the power equipment area.Secondly,the method of substation equipment state recognition based on convolutional classification network is studied.In this method,the state categories of substation equipment area extracted by YOLOv3 algorithm is labeled manually to constitute a training set,and the classification algorithm is used to achieve state classification.In addition,a disconnector state recognition method based on semantic segmentation and connected component labeling method is studied.On the basis of semantic segmentation,the region growth algorithm is used to label connected component of the disconnector,and a component sorting statistical method is studied to optimize the component threshold and eliminate the non-switching arm components.Then,the state of the disconnector is judged according to the number of connected component of the switch arm.Finally,the algorithm of substation equipment state recognition based on heterogeneous model ensemble learning is studied.The recognition results of multiple single models are fused by the weighted voting method to improve the accuracy of the substation equipment state recognition.