Research On Substation Image Analysis Method Based On Deep Learning

With the increasing demand for electricity in our country,the power grid is also facing more severe challenges,and the operating status of power equipment in the entire power system affects the safety and stability of the power grid.At present,the defect detection of power equipment is mainly realized manually,and the process of manual detection is usually by staff patrolling the substation or judging the defects through the data collected by the camera.The disadvantage of this detection method is that the detection efficiency is low,and the corresponding detection results may be interfered by external factors.In order to establish the application platform of electric power artificial intelligence technology,this paper applies deep learning algorithms to the field of electric power systems,using convolutional neural network models to analyze the images of electric power equipment to find out the defective areas,so as to solve the problems of traditional manual inspection And insufficient.The main tasks include:(1)According to the requirements of the actual engineering project,collect the defective pictures of the electric equipment of the substation at the power site,combine the defect assessment rules and existing related data sets,classify and sort the collected digital images of the electric equipment defects of the substation,and identify the defects Annotate the images and enhance the data according to the amount of defect data;(2)In-depth study of the structure and principle of several commonly used target detection and instance segmentation neural networks,these models are tested on the collected data sets,the performance between the models is compared,and the network's shortcomings in the power scene are studied.Aiming at these shortcomings,improve the network or design a new network to obtain higher detection accuracy;(3)Aiming at the problem that traditional methods are prone to deviations in the detection of power equipment defects in substations,a defect detection method based on sample migration network is proposed.By training the sample migration network,the model can spontaneously modify the candidate frame according to the needs of the training task,focus on the areas that are more important to the detection task,and reduce the deviation of the detection frame;a new loss function is designed,and the Io U-based Regression loss and classification loss based on hard case mining improve the detection ability of the model;the structure between modules is designed,and the common slibing head of the two-stage network is retained,which is complementary to the offset module.Through comparative experiments of power equipment defect detection methods,the proposed detection method based on sample migration network is evaluated.The experimental results show that the average accuracy rate(m AP)of the proposed method reaches 0.833,which is better than the detection effect of traditional algorithms;(4)Aiming at the situation that the irregular shape of individual categories of substations causes poor detection results,the method of instance segmentation is used to detect this category.By using the new feature extraction RPF and feature fusion method Res Ne Xt structure,the original backbone network is replaced,and the detection effect is improved without causing a lot of redundant calculations;using different methods of non-maximum suppression strategies,effectively reducing candidates The detection performance in the case of overlapping frames is significantly improved compared to the performance of the traditional instance segmentation method,and the detection AP can reach 0.821.This paper applies deep learning technology to the defect detection task of electric equipment in substations,and at the same time improves and optimizes the structure of the benchmark detection model,including application innovation and theoretical innovation,all of which are the application of deep learning technology in powerrelated fields Development provides a theoretical basis,and at the same time provides a certain reference value for follow-up research in related fields.