Hydrophobicity Classification Of Composite Insulators Based On Continuous Representation Of Image And Deep Learning

Hydrophobicity is an important index to evaluate the aging state of composite insulators.At present,the most widely used hydrophobicity detection method is spray method,but the accuracy of this method depends on the experience of the staff,and the detection efficiency is low.Image analysis and deep learning are introduced into the hydrophobicity detection to achieve intelligent detection and improve the detection efficiency.However,due to the complexity of the environment and the transparency of bead images,the detection results of such methods are not satisfactory.In order to detect the hydrophobicity classes,the main works of this thesis are as follows:(1)Aiming at the problems of low contrast and difficult segmentation of bead images,firstly,the images are preprocessed by grayscale,image enhancement and other technical means to highlight the edges of beads and eliminate the occlusion of adjacent beads due to uneven illumination and shadows of water droplets.Secondly,the segmentation rules of bead images are formulated,and the segmentation of bead image is realized automatically by using threshold segmentation method based on entropy and deep learning model,and the binary image of bead or water trace is obtained.The segmentation accuracy of this method reaches 93.94%,and it only takes0.11 s to segment a bead image.(2)The segmentation task based on deep learning can be regarded as a mapping from matrix to matrix,which is essentially a function estimation problem.In this thesis,the representation of this function is discussed by the continuous representation of digital image.Firstly,the two-dimensional image segmentation is transformed into the optimization problem of continuous functional,then the continuous representation of digital image is used to establish the constraint conditions of continuous functional,the optimization problem is transformed into the solution problem of overdetermined equations,then the unknown coefficients of equations are solved by BP neural network,and finally the segmentation of bead images is realized by using the optimization model.(3)Most of the hydrophobicity classification based on machine learning needs to extract the characteristics of bead images,such as shape factor of beads and maximum area ratio.However,it is time-consuming to study the representative characteristics of beads,and there is no direct correlation between the characteristics of beads and hydrophobicity classes.After completing the segmentation of bead images,this thesis introduces GoogLeNet network,takes the obtained binary images of bead as the input of GoogLeNet and the corresponding hydrophobicity classes as the output,sets reasonable parameters,and automatically learns the shallow or deep features of the binary images through supervised learning.The results show that the classification accuracy of the method reaches 89.1%,and the detection of an image takes only0.03 s.