Research On Deep Learning Based Power Insulator Identification Method

As a way of overhead transmission line insulator used for electrical insulation and mechanical fixation important insulation controls,the number of large and long exposure to the natural environment,climate,temperature,its durability are prone to the influence of such factors as explosive,damaged,missing,such as fault problem,has a great influence to the stability of the power transmission,Therefore,it is indispensable for the inspection of insulators.With the continuous development of UAV technology,personnel take aerial photos of the insulators tested by UAV and judge the insulator images collected by naked eyes.This method seems to be a reliable method,but in fact,due to the large amount of insulator data captured by the UAV and the complex shooting background,the staff often rely on experience to make judgment,which will inevitably lead to errors.In recent years,with the rapid development of machine vision and computer performance,it has become a reliable and popular method to use computer to assist staff in insulator and its fault detection.In this paper,a method of "first location,then identification" is proposed for the detection of insulators and their defects.By improving the existing target detection framework Faster R-CNN,and based on deep learning technology,a new cascade target detection model is proposed.Aerial image of insulator is used as the sample to locate and identify the defects,so the automatic detection of insulators and their defects on overhead transmission lines is realized.The research work is as follows:(1)for the existing insulator datasets,especially fault insulator data set sample shortage,poor quality problems,using background radiation transformation,the fusion technology of insulator positive and negative samples augmentation operation,and the amplification of insulator image after image enhancement,mark and insulator data set processing,such as: The problem of over-fitting and low recognition success rate in network training and testing due to the small sample size is solved.(2)According to the shape and size of insulators and defects,the anchor style and related hyperparameters in the Faster R-CNN detection framework were optimized;The number of candidate regions(NAP)was added to realize the dynamic adjustment of the number of candidate regions and improve the speed of training and detection.(3)By inserting saliency learning module(ALM)into the insulator positioning network,the image saliency technology is introduced into the insulator positioning task to highlight the importance of insulators in the image and assist the insulator positioning network to locate insulators more quickly.(4)A cascading target detection model of "insulator location-image clipping-defect detection" is built.(5)by designing various types of contrast experiment,using the amplification of insulator for model training and testing data sets,insulator localization rate of 91.18% and defect detection rate is 90.06%,which verifies the paper put forward the parameters optimization and improvement of the method and the feasibility and effectiveness of the cascading detection structure and identify performance boost.