Research On Transmission Line Defect Identification Based On Deep Learning Technolog

At any time,the power grid transmission lines continue to grow.Taking Yunnan power grid as an example,the length of transmission lines has reached 89576 kilometers in 2022.With the increase of transmission lines,the operation and maintenance workload of power grid companies is increasing.At present,the operation and maintenance work of transmission lines is still mainly manual.In recent years,with the continuous development of UAV and AI technology,a large number of photos are taken during UAV patrol and manual patrol,and photos are viewed manually,Find line defects and eliminate them in advance to ensure the safe and stable operation of the transmission line.Using deep learning technology to identify defects in the photos taken can not only improve the efficiency of defect detection and reduce personnel costs,but also have great prospects for the deep application of deep learning technology in all aspects of power grid transmission line defect identification.However,the following problems exist in the actual application process:1.In the establishment of the sample library of power grid transmission line defect identification,most of them are normal samples,and the number of defect samples is relatively small,so it is necessary to build the defect sample library.By building the defect sample library,the accuracy of transmission line defect identification can be improved.2.Due to the fact that Yunnan is located in many mountainous areas,there are more fog in the photos taken.In order to improve the accuracy of defect recognition in the later stage,it is necessary to preprocess the pictures before defect recognition.It is necessary to study the image preprocessing algorithm with good defogging effect.In order to better improve the defogging effect,we use the defogging method of teacher guidance and dual antagonistic learning to defog the pictures.Through data set construction,model selection Parameter adjustment,ablation experiment and other methods were used to train the model.The model improved the PSNR and SSIM indexes of defogging.In the construction of three test data sets a,B and C,the PSNR and SSIM indexes in test data set a increased by 45.89% and 7.02% on average,respectively.The PSNR and SSIM indexes in test data set B increased by 35.89% and 7.58% on average,respectively.The PSNR and SSIM indexes in test set C increased by 36.83% and 13.47% on average,respectively,It lays the foundation for the improvement of the recognition rate of component defects and the deployment of the algorithm in practical application.3.Due to the limited conditions of manual patrol shooting and the distance from the transmission line,the small target objects of the transmission line,such as insulators,cannot be effectively identified,and different equipment takes photos at different times,places,angles,and weather conditions,which will cause large differences in brightness,contrast,background,etc.between images,that is,there is a problem of domain offset between data,In addition,the algorithm for small target object recognition cannot be controlled independently,and the algorithm model cannot be continuously optimized after the sample is increased.Therefore,it is necessary to establish a picture sample library.In order to solve the problem of domain offset for a single small target object,research the target recognition method based on multiple unsupervised domain counterwork intelligent recognition and the domain adaptive target recognition method based on counterwork training,and improve the cross domain recognition ability for small target objects,Through the research of data set construction,pre training,self training network,ablation experiment and comparison,the experiment shows that the average prediction accuracy(map)value of glass span composite material in insulator is 263.64%,87.23% and 32% higher than that of yolov3,yolov5 and fast RCNN based on multiple unsupervised domain confrontation intelligent target recognition method.Yolov3,yolov5 and faster RCNN of composite cross glass materials were increased by 235.64%,13.9% and 27.18%,respectively.The domain adaptive target recognition method based on countermeasure training improved the insulator target detection algorithm.The average prediction accuracy(map)value of composite cross glass materials in insulators was increased by 56.47%,53.46% and 51.1%compared with faster RCNN,Aug FPN and Pisa RCNN.The fast RCNN,Aug FPN and Pisa RCNN of composite cross glass materials increased by 40.72%,71.28% and 64.53%,respectively.The research results have been verified in the specific paper,which can improve the efficiency of power grid transmission line maintenance work,provide intelligent support for transmission line maintenance work,and provide technical support for ensuring the normal operation of transmission lines,so as to ensure the normal operation of national economy and society.