Obstacle Recognition Technology For Deicing Robots On Power Lines Based On Modified-ELM

Electricity supply is an indispensable element of people's lives and national operations,which requires transmission systems to work reliably under all conditions.As a front-end solution for solving the problem of ice coating on transmission lines,de-icing robots,such as suspension clamps,spacers,insulators and other accessories installed on the transmission line,can become obstacles to their deicing operations.In order to enable the de-icing robot to perform operations stably and efficiently,accurately identifying the location of the obstacle on the transmission line and classifying it becomes the primary problem to be solved.This paper investigates and analyzes the environmental characteristics of transmission lines,the requirements of deicing operations and the types of obstacles.It analyzes the installation methods and structural characteristics of common transmission line devices,and summarizes the difficulties in the identification of transmission line obstacles.The algorithm training efficiency is low,the classification accuracy is not high,and the over-fitting is serious.A set of Modified-ELM-based obstacle recognition scheme for the de-icing robots of transmission lines is designed.The scheme is mainly composed of two parts: image preprocessing and neural network based obstacle target localization classification.Image preprocessing includes noise reduction and edge detection.Firstly,the original image captured by the deicing robot imaging system is denoised using BM3 D algorithm.Then,based on the adaptive threshold wavelet transform algorithm,the object edge information in the image is obtained.The location classification is completed by neural network and deep learning algorithm,and a neural network structure with convolution and fully connected parts is constructed.The Dropout layer is introduced,and the layers of neurons are customized according to the requirements of transmission line obstacle recognition.The number and connection relationship.On the problem of image preprocessing,this paper discusses the influence of the similarity threshold and the truncation threshold in the DCT transform domain onthe image denoising effect in the BM3 D algorithm,and analyzes the performance of different wavelet functions in the wavelet edge detection algorithm under different thresholds.On the problem of obstacle target location classification,the genetic algorithm is used to train the convolution kernel of neural network.The limit learning machine algorithm(ELM)of Hinton is analyzed,the mathematical model is improved,and the Modified-ELM algorithm is proposed to train the network.The full connection part improves the convergence speed of the whole training process.Finally,experiments are carried out on the simulated transmission line photo dataset.The results show that the scheme can solve the obstacle identification problem on the transmission line well.The cutting edge method is more accurate and more stable,and meets the real-time requirements of robot deicing operations.