Research On Grading Ring Detection Technique Of Power Line Based On Deep Learning

The grading ring is an important component of the power line.The periodic inspections of the grading ring is an important measure to ensure the safe operation of the power system.At present,the inspections of the grading ring in our country are mainly conducted by manual methods,and supplemented by aerial image analysis of Unmanned Aerial Vehicles.Grading ring edge detection and local shape contour matching using traditional digital image processing technique require manual design features and setting of algorithm parameters for the different environments,resulting in low detection accuracy,poor applicability of the algorithm,and inability to achieve positioning of the grading ring.With the maturity of deep learning theory,convolutional neural network has the capabilities of strong feature learning and complex scene processing,which has been widely used in the field of target detection.This paper conducts the research of grading ring detection technique based on deep learning,and realizes the grading ring recognition and position in aerial images.Feature extraction is an important step of achieving target detection.First of all,by extracting the Histogram of Oriented Gradient(HOG)features of grading ring,the detection approach of grading ring based on convolution neural network is proposed aiming at expression insufficiency of manual design shallow features.Compared with various CNN networks and according to the characteristics of grading ring with different sizes and shapes in aerial images,Inception V2 network is selected as the feature extraction network and the experiment of feature map visualization is carried out.Through the comparison of different convolution layers of feature maps,it is found that as the number of network layers increases,the abstract and deep features of the grading ring can be learned which verifies the rationality of the research approach.Secondly,compared with various deep learning frameworks and R-CNN series algorithms,TensorFlow and Faster R-CNN are selected to train the grading ring detection model based on actual aerial image data set.The accuracy of the training model is verified by analyzing the loss curves,weight parameter distribution of the training process and test results of eval program by using Tensorboard visualization output.Finally,after conducting the actual test experiments of grading ring detection model based on deep learning,the average accuracy rate is 87.8%.By comparing the effect of different iterations on the accuracy of model actual test,the relationship between iterations and accuracy is determined.Robustness of grading ring detection model is verified by designing light intensity and motion blur experiments.And tilt malfunction detection of grading ring under certain conditions is implemented.The experimental results show that the grading ring detection model in this paper has certain accuracy and practical value.