Research On Insulator Target Recognition Method Combined With Artificial Samples

Insulator is an important electrical insulation and support device of the transmission line,but it is prone to fault,which has an important impact on the safety of the power grid.Therefore,insulator is the key recognition object of the transmission line.The method of transmission line inspection combined with UAV acquisition and manual recognition can’t meet the current task requirements,so the deep learning method becomes the key research content.In the application of the deep learning to the insulator recognition,a problem of the insufficiency of both public dataset and quality for insulator image emerges.In the light of this situation,this thesis proposes two methods of insulator artificial sample generation and studies how to build a "real-artificial" parallel training dataset,which provides a theoretical basis for alleviating the problem of sample shortage.First of all,this thesis not only describes the status of insulator recognition and sample problems,but also analyzes the advantages and disadvantages of the existing artificial sample amplification technologies.To solve the problem of insufficient samples,this thesis proposes a framework of "real-artificial" parallel training dataset creation.Two methods of insulator artificial sample generation based on 3D Max software platform and generation adversarial nets are proposed respectively,which could provide artificial samples for the framework.In view of the negative influence of the samples generated by generation adversarial nets,a sample selection scheme based on Ada Boost algorithm is realized.Finally,this paper uses Res Net50,Inception＿v3 and Dense Net121 to identify insulator respectively.The results show that based on the framework of “real-artificial”parallel training dataset creation,the recognition accuracy of Res Net50 increased by10.1% to 99.04%,that of Inception＿v3 by 3.98%,to 97.02%,and that of Dense Net121 by 7.52%,to 96.10%,which alleviates the problem of lacking enough samples and effectively improves the performance of insulator identification.