CN113221614A - Power grid insulator damage image identification method based on hybrid neural network - Google Patents

Power grid insulator damage image identification method based on hybrid neural network Download PDF

Info

Publication number
CN113221614A
CN113221614A CN202011539694.8A CN202011539694A CN113221614A CN 113221614 A CN113221614 A CN 113221614A CN 202011539694 A CN202011539694 A CN 202011539694A CN 113221614 A CN113221614 A CN 113221614A
Authority
CN
China
Prior art keywords
neural network
image
gray
layer
hybrid neural
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202011539694.8A
Other languages
Chinese (zh)
Inventor
邓凯
张晓平
陈冠
吴嘉明
霍梓航
曾争
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
China Southern Power Grid Power Technology Co Ltd
Original Assignee
China Southern Power Grid Power Technology Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by China Southern Power Grid Power Technology Co Ltd filed Critical China Southern Power Grid Power Technology Co Ltd
Priority to CN202011539694.8A priority Critical patent/CN113221614A/en
Publication of CN113221614A publication Critical patent/CN113221614A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Multimedia (AREA)
  • Probability & Statistics with Applications (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Image Analysis (AREA)

Abstract

The invention provides a power grid insulator damage image identification method based on a hybrid neural network. Inputting a preprocessed image into a convolutional neural network, performing down-sampling on an original image sequentially through a convolutional layer and a pooling layer, extracting high-dimensional characteristics of the image, and performing down-sampling on a characteristic matrix through a full-connection layer; inputting image features extracted from a full connection layer of a convolutional neural network into a support vector machine, training the support vector machine, mapping the corresponding relation between independent variables and strain variables of an original low-dimensional vector space into a high-dimensional vector space to enable the high-dimensional vector space to be in a linearly separable state, finding a hyperplane in a feature space, and separating different types of data. The method combines the feature extraction capability of the convolutional neural network and the feature classification capability of the support vector machine, and can remarkably improve the image classification accuracy.

Description

Power grid insulator damage image identification method based on hybrid neural network
Technical Field
The invention relates to the field of image recognition, in particular to a power grid insulator damage image recognition method based on a hybrid neural network.
Background
The insulator is large in use amount and various in variety in a power system, is one of key components of a high-voltage overhead power transmission line, and plays a very important role. On the one hand, the wire is responsible for the mechanical supporting function of the wire; on the other hand, the insulating layer plays a role of insulation and prevents current from forming a channel ground to the ground. As one of the important components of the transmission line, the insulator increases the creepage distance, insulates the charged part, and supports and positions it. Insulators operate in the field for a long time and are very prone to various mechanical and electrical faults. The insulator is often damaged due to problems of self-explosion, ice coating, foreign matters and the like. The insulator is an indispensable ring in the operation of the power system, and various electromechanical stress failures caused by the change of environmental and electrical load conditions should not occur, so that the service life and the operation life of the whole line are damaged. However, the insulator is a very vulnerable element in the power transmission line. Breakage of the insulator can lead to interruption of the power supply system and in severe cases can even lead to grid breakdown. The insulator is directly related to the safe and stable operation of the power transmission line, so that timely and accurate insulator damage detection is particularly important.
Before the deep learning method is applied to a large scale, the damage detection of the insulator mainly adopts an observation method, a laser Doppler vibration method, an ultrasonic detection method, an infrared temperature measurement method, an aerial photography method and the like. From the initial direct observation method to the manual analysis method based on aerial images and videos, although workers can remotely observe and judge whether the insulator is damaged through data, the workload of manual analysis is still too huge due to massive image and video data.
The traditional insulator breakage detection generally adopts a manual inspection method, which is simple, but has extremely low efficiency and certain danger. In the aspect of monitoring the damage condition of the insulator through a video, the traditional edge detection method has a good effect when simple insulator images are processed, but the insulator is exposed outdoors for a long time and can be influenced by the surrounding environment, the damage condition is complex, and the effect is poor when the edge detection technology is used for detecting the damage.
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides a power grid insulator damage image identification method based on a hybrid neural network. The technical scheme of the invention is as follows.
A power grid insulator damage image identification method based on a hybrid neural network comprises the following steps:
convolving a feature map by convolution layers of a hybrid neural network, the feature map being represented by equation one (1);
Figure BDA0002854196980000021
wherein,
Figure BDA0002854196980000022
is the kth characteristic diagram, W, of the ith layeri kIs a weight matrix Wi k，
Figure BDA0002854196980000023
Is a bias term;
(ii) a convolutional layer pooling profile through a hybrid neural network, the convolutional layer comprising neurons identified by formula two (2);
Figure BDA0002854196980000024
the function f () is an activation function and is used for realizing the nonlinear transformation of the characteristic value, and the function down () is a down-sampling function and is used for realizing the size transformation of the characteristic diagram by taking the maximum value or the average value;
converting the output values of the multiple classifications into probability distribution which ranges from [0,1] and is added to 1 by adopting a formula III (3);
Figure BDA0002854196980000025
wherein, σ (z)i) The number of output nodes is C;
calculating the Cross entropy loss by equation four (4)
Figure BDA0002854196980000026
Wherein N is the number of nodes, alphaiA is thatjIs given byiRecording the output of each node;
the decision classification is carried out by adopting a formula five (5),
Figure BDA0002854196980000027
wherein alpha isiFor penalty coefficients calculated by the Lagrangian transformation, xiAs a feature space of the input, yiB is the label of the class to which the input sample belongs, and b is the offset of the classification plane.
Preferably, the acquired damaged image of the power grid line insulator is preprocessed before the feature map is convolved.
Preferably, the preprocessing includes performing a gray scale transformation using equation six (6),
Gray＝0.299R(x,y)+0.587G(x,y)+0.114B(x,y)
R(x,y)＝G(x,y)＝B(x,y)＝Gray
where Gray represents the Gray value of the pixel, R (x, y) represents the red component, G (x, y) represents the green component, and B (x, y) represents the blue component.
Preferably, the contrast stretching transformation is performed by using the formula seven (7),
Figure BDA0002854196980000031
where r represents the gray level of the input image, s is the corresponding gray level value in the output image, and E is the slope.
Preferably, the median filtering is performed by using the formula eight (8),
g(x,y)＝median{f(x-k,y-l),(k,l)∈W} (8)
wherein, (x, y) is the coordinates of the pixel points, f (x, y) is the gray value of the pixel points, (k, l) is the set of the distances between all the pixel points in the pixel window and the original pixel points, and W is a constant.
Compared with the prior art, the invention has the beneficial technical effects that: inputting a preprocessed image into a convolutional neural network, performing down-sampling on an original image sequentially through a convolutional layer and a pooling layer, extracting high-dimensional characteristics of the image, and performing dimension reduction on a characteristic matrix through a full-connection layer; inputting image characteristics extracted by a convolution neural network full connection layer into a support vector machine, training the support vector machine, mapping the corresponding relation between independent variables and strain variables of an original low-dimensional vector space into a high-dimensional vector space to enable the high-dimensional vector space to be in a linearly separable state, finding a hyperplane in a characteristic space, and separating different types of data. The method combines the feature extraction capability of the convolutional neural network and the feature classification capability of the support vector machine, and can remarkably improve the image classification accuracy.
In order to make the aforementioned objects, features and advantages of the present application more comprehensible, preferred embodiments accompanied with figures are described in detail below.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings that are required to be used in the embodiments will be briefly described below, it should be understood that the following drawings only illustrate some embodiments of the present application and therefore should not be considered as limiting the scope, and that for a person skilled in the art, other related drawings can be obtained from these drawings without inventive effort.
FIG. 1 is a graph showing the accuracy rate variation trend of a non-hybrid model under different iteration times;
FIG. 2 is a graph showing the accuracy rate variation trend of the mixed model at different iteration times.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only a part of the embodiments of the present application, and not all the embodiments. The components of the embodiments of the present application, generally described and illustrated in the figures herein, can be arranged and designed in a wide variety of different configurations. Thus, the following detailed description of the embodiments of the present application, presented in the accompanying drawings, is not intended to limit the scope of the claimed application, but is merely representative of selected embodiments of the application. All other embodiments, which can be derived by a person skilled in the art from the embodiments of the present application without making any creative effort, shall fall within the protection scope of the present application.
The embodiment provides a grid line insulator damage image identification method based on a hybrid neural network, which comprises the following steps:
the acquired power grid line insulator damage image is preprocessed through three steps of gray-scale image conversion, image enhancement processing and image spatial filtering, so that noise in the image is suppressed, and errors in subsequent image feature extraction are reduced.
In computers, color images are represented with 3 bytes of data per pixel. It can be decomposed into three monochrome images of red (R), green (G), and blue (B), and the luminance of each pixel is identified by a gray value. The gray image is obtained by removing color information in the color image and only contains brightness information, so that a large amount of computer resources can be saved for storage and processing.
There are three types of processing methods for grayscale images: maximum, average, and weighted average. The expression is as follows:
Gray＝max(R(x,y),G(x,y),B(x,y))
Gray＝(R(x,y)+G(x,y)+B(x,y))/3
Gray＝WRR(x,y)+WGG(x,y)+WBB(x,y)
as can be seen from the above equation, the characteristics of the original image are changed or lost by any method. In view of practical applications, the following gray scale transformation formula is selected herein:
Gray＝0.299R(x,y)+0.587G(x,y)+0.114B(x,y)
R(x,y)＝G(x,y)＝B(x,y)＝Gray
wherein Gray represents the Gray level value of the pixel, R (x, y) represents the red component, G (x, y) represents the green component, B (x, y) represents the blue component, and the weight in the formula is the most reasonable Gray level image weight proved by the previous experiment.
Under the influence of complex environmental factors, such as illumination, mechanical vibration or shaking, the whole brightness of the actual image may be different, so that the gray value distribution of the image is relatively solidified, and the difficulty of subsequent treatment is increased. In the icing image of the power transmission line, the target icing conductor is not easy to distinguish from the background, and the gray value of the image is concentrated. In order to distinguish the icing conductor from the background ground, the range of the gray value in the image needs to be expanded, the brightness value of the image pixel is higher or lower, and the contrast of the icing conductor image is enhanced so as to be convenient to distinguish.
The method of the embodiment adopts a linear gray scale stretching algorithm to improve the dynamic range of the picture, and is used for emphasizing the frozen part in the image, and the following is a contrast stretching transformation function:
Figure BDA0002854196980000051
where r denotes the gray level of the input image, s is the corresponding gray level value in the output image, and E is used to control the slope of the function, which is implemented in MATLAB as follows:
Figure BDA0002854196980000052
wherein eps is a tiny number so as to avoid the problem that the eps cannot be calculated when the input is 0, and the value of eps in the method is 1 e-5.
After the image is enhanced, the details of the image are clearer, the contrast is enhanced, and meanwhile, the noise of the image is increased. In order to reduce the noise of the image, the image needs to be subjected to smooth filtering.
Spatial smoothing filtering typically includes mean filtering, gaussian filtering, median filtering, adaptive filtering, and the like. The first two methods belong to linear smoothing filtering methods, and the last two methods belong to nonlinear smoothing filtering methods. All the factors are comprehensively considered, and the method selects a median filtering method to process the acquired image.
The median filtering method is to replace the gray scale of a pixel f (x, y) with the median of the brightness of all pixels near the pixel. The specific method is to determine a window W, and then replace the original pixel f (x, y) with the median of the gray levels of all pixels in the window to obey the filtered image g (x, y). The function expression is as follows:
g(x,y)＝median{f(x-k,y-l),(k,l)∈W}
wherein, (x, y) is the coordinate of the pixel point, f (x, y) is the gray value of the pixel point, and (k, l) is the set of the distances between all the pixel points in the pixel window and the original pixel point.
Inputting the preprocessed image into a convolutional neural network, wherein the convolutional neural network does not comprise an output layer, performing down-sampling on the original image sequentially through convolution and pooling, extracting high-dimensional features of the image, and then performing down-dimension on a feature matrix through a full-connection layer.
A Convolutional Neural Network (CNN) is a multi-layer Neural Network that is built up from a number of Convolutional layers and pooling layers (lower sampling layers). Then, one or more fully connected layers are connected, and the image features generated by the previous layers are classified. Due to the adoption of the method of sharing weight by local connection and neurons, the number of free parameters of the CNN is greatly reduced, and the efficiency of the CNN is higher than that of a fully-connected network. In addition, due to the function of the pooling layer, the image features have better conversion, scalability and distortion invariance. The basic network working structure of a convolutional neural network can be divided into five parts: input layer, convolution layer, pooling layer, full-link layer and output layer.
The convolutional input layer may act directly on the original input data. For an input image, the input data is the pixel values of the image. For convolutional neural networks, the output of the convolutional layer is obtained by convolving the filter of the first layer with the input feature map, the convolution kernel is calculated by sliding the window over the feature map one by one, adding a bias term and then applying a non-linear activation function. The output value of a convolutional layer is the characteristic map of that layer. Each filter produces an output profile.
The input image is represented by X, and the kth feature map of the ith layer is represented by
Figure BDA0002854196980000061
The k filter of the ith layer is characterized by a weight matrix Wi kAnd bias term
Figure BDA0002854196980000062
And (6) determining. Then, the kth profile of the ith layer may be represented by:
Figure BDA0002854196980000063
the convolutional layer has the advantages of local connection and weight sharing, and through the local connection, the characteristics of an image can be better extracted, and the connection parameters of adjacent layers in the network are reduced so as to reduce the number of required parameters.
After the features of the image are extracted through the convolutional layer, the features can be theoretically directly input into the full-link layer and classified by the classifier. However, due to the large dimension of the features, the calculation amount is very large, and overfitting is easy to generate. Although the convolution layer adopts local connection to reduce the number of connections in the network structure, the increase of the number of feature maps increases the feature dimension, the network is still very complex, and the training difficulty is still very large. In order to further reduce the parameters of the network and reduce the complexity and overfitting degree of the model, a pooling layer is usually set after the convolutional layer. Pooling is a sampling process that integrates the outputs of neighboring neurons in the same profile. After a down-sampling mechanism is introduced, the characteristic dimensionality can be effectively reduced, the effective information of the image is retained, meanwhile, redundant data is removed, and the network training speed is accelerated. The principle of such aggregation is that the pixels of each neighboring region in the image have greater similarity. The region can be described by calculating the maximum value of the region as the sample value, or by averaging all the values of the region together and taking the average value as the sample value. The procedure of the pooling layer is represented by the following formula. The neurons of this layer employ downsampling functions for maximizing or averaging the pooled feature map.
Figure BDA0002854196980000071
The function f () is an activation function and is used for realizing the nonlinear transformation of the characteristic value, and the function down () is a down-sampling function and is used for realizing the size transformation of the characteristic diagram by taking the maximum value or the average value. For a fully-connected layer, it may contain a plurality of fully-connected elements, which are in fact hidden layers of the multi-layer perceptron. The ganglion nodes of the later layer are connected with each nerve node of the previous layer, and the neuron nodes of the same layer are not connected. And each layer of neuron nodes are transmitted forwards through the weights on the connecting lines, and the weights are combined to obtain the input of the next layer of neuron nodes. The number of neural nodes in the output layer is set according to the specific application task. For the classification task, the output layer of the convolutional neural network is usually a classifier, and is composed of a fully-connected layer and an activation layer, and the embodiment adopts a softmax function as an activation function. The definition of the softmax function is as follows:
Figure BDA0002854196980000072
where σ (z)i) And C is the output value of the ith node, and the number of output nodes, namely the number of classified categories. The output value of the multi-classification can be converted into the range of [0,1] through the function]And add up to a probability distribution of 1.
Assuming that h is the activation value of the second last layer and W is the weight of the second last layer to the softmax layer, the input of the softmax layer can be expressed as:
Figure BDA0002854196980000073
suppose for the class N classification problem, the softmax layer has N nodes, and the output of each node is recorded as piWherein i is 1,2, …, N, because piIs a discrete probability distribution, satisfies the relationship
Figure BDA0002854196980000081
Wherein
Figure BDA0002854196980000082
The result of this equation is used to calculate the cross entropy loss for softmax.
In the method of the embodiment, the neural network is trained on a corresponding data set, and after the highest accuracy of the neural network is reached, the features acquired before the output layer are independently stored for subsequent classification research.
The image features extracted from the convolutional neural network full-link layer are input into a support vector machine, the support vector machine is trained, and original data are mapped to a high-dimensional feature space, so that the original data are divided into multiple classes, and the optimal classification effect is obtained.
The image features can be obtained by a convolutional neural network, but it cannot obtain the optimal classification accuracy. A Support Vector Machine (SVM) with fixed kernel functions cannot learn the complex features of the image. However, a "soft interval" approach may be used to maximize the interval and obtain the decision plane. Finally, an optimal solution to the classification problem can be obtained in the learning feature space.
The standard SVM is a non-probabilistic binary linear classifier, i.e., for each input, it predicts that the input will be one of two classes.
The principle of the support vector machine is as follows:
let the training set sample be { (x)i,yi)|xi∈Rd,yi∈{-1,1},i＝1,2,…,N}，yiThe labels of the classes to which the samples belong, N is the number of training samples, and d is the dimension of the samples. For linearly separable datasets, there is a generalized optimal classification hyperplane:
w·x+b＝0
where w is an n-dimensional vector, b is an offset, and is an inner product operation, so that the classification interval obtains the maximum value, i.e.
Figure BDA0002854196980000083
In order to be the maximum of the number,
Figure BDA0002854196980000084
is minimal. Thus, the classification of the optimization problem can be converted into the following form:
Figure BDA0002854196980000085
however, in actual work, many data are not completely linearly separable, so that the embodiment introduces a relaxation variable and a penalty coefficient on the basis of the above, and the relaxation variable and the penalty coefficient can be converted into the following optimization problems after lagrange transformation:
Figure BDA0002854196980000091
Figure BDA0002854196980000092
solving the above equation to obtain alphaiThen is further prepared by
Figure BDA0002854196980000093
To obtain w. In this problem, αiIs not 0, its corresponding training set sample is a support vector. We can use the following formula to judge the unknown class attribute vector.
Figure BDA0002854196980000094
The mapping function in the multidimensional space does not need to be explicitly calculated, and only the kernel function K (x)i,x)＝<Φ(xi),Φ(x)>Substituting into the above formula. The decision function is as follows:
Figure BDA0002854196980000095
wherein alpha isiFor penalty coefficients calculated by the Lagrangian transformation, xiAs a feature space of the input, yiB is the label of the class to which the input sample belongs, and b is the offset of the classification plane.
The support vector machine of this embodiment maps the corresponding relationship between the independent variables and the strain variables of the original low-dimensional vector space to a high-dimensional vector space (feature space) to make it a linearly separable state, so that the nonlinear feature vector can be linearly analyzed in the high-dimensional feature space by a linear algorithm. According to the structural risk minimization principle, an optimization tool is utilized to find a hyperplane in a feature space, and data and components thereof are divided into two classes so as to obtain the optimal classification effect.
The structure of the hybrid model is the combination of a convolutional neural network and a support vector machine, namely, in the last step of classification, the traditional softmax layer is replaced by the support vector machine. The fully-connected layer of CNN can be viewed as a set of features of the original image, and therefore it is practical to train and classify using classifiers with these features. And after the original CNN is trained by a reverse propagation algorithm, the output of the full-connection layer is taken as the newly extracted feature. They are then sent to train the SVM classifier. After the SVM is trained, a task of recognizing test data is started.
Since the hybrid model combines the advantages of the convolutional neural network and the support vector machine and compensates for their limitations, the performance of the hybrid model will be superior to each of the independent models. The learning algorithm of CNN is based on empirical risk minimization, minimizing the error of the training set. When the first classification hyperplane is found by the back-propagation algorithm, the training process will be terminated whether it is locally optimal or globally optimal. The classification hyperplane of the SVM is globally optimal using the structured risk minimization principle. Therefore, the generalization capability of the multilayer neural network is stronger. It can be seen that the generalization capability of the multi-layer neural network is lower than that of the SVM. Therefore, the SVM is used for replacing the softmax layer of the convolutional neural network, and the classification accuracy is improved.
Based on the acquired data, simulation evaluation is performed on the method of the embodiment on MATLAB to determine the influence of the iteration steps of the algorithm on the method. According to the method, a plurality of monitoring points of typical insulator damage conditions are obtained from a power grid line insulator detection system in a certain area, and the damage conditions are divided into five grades according to the insulator damage area: normal, slight damage, moderate damage, severe damage, alarm damage.
In order to verify the classification effect of the hybrid model, the non-hybrid model and the hybrid model are subjected to an insulator damage condition image classification experiment, and the hybrid model provided by the method is verified to have a good classification effect. Both models used the same CNN structure. First, the non-hybrid models are classified with a convolutional neural network. Finally, the non-hybrid models are classified with a softmax layer. From the experimental results in table 1, it can be found that the accuracy of the training set and the test set increases with the number of iterations from the first iteration to the tenth iteration, but the accuracy of the test set is much lower than that of the training set, and is substantially less than 10%. After 10 iterations, the classification accuracy on the training set can reach 95.1%, and the classification accuracy on the test set can only reach 84.1%. The mixed models are then classified by experiment. The last layer of classifier of the hybrid model is SVM, after 10 iterations, the classification accuracy of the training set can reach 98.8%, and the classification accuracy of the test set can reach 93.2%.
Fig. 1 and fig. 2 respectively plot accuracy rate variation trends of the non-hybrid model and the hybrid model at different iteration times, and it is obvious that as the number of iteration steps increases, the accuracy difference between the training set and the test set gradually decreases, and both approaches convergence at the tenth iteration. Due to the type difference between the training set and the testing set, the testing condition and the training condition of the training set have certain difference, but the difference of the mixed model is smaller, and the generalization capability to different samples is stronger. Analysis in conjunction with the following table can lead to the conclusion: compared with simple convolutional neural network classification, the classification precision of the mixed model training set is improved by 3.7%, and the classification precision of the test set is improved by 9.1%.
Figure BDA0002854196980000101
Figure BDA0002854196980000111
In the method for identifying the damaged image of the insulator of the power grid line based on the hybrid neural network, firstly, gray processing is carried out on an image acquired by a power grid junction insulator damage detection system, so that the image is reduced from three dimensions to two dimensions; secondly, through image enhancement, the image quality is improved for subsequent processing. Because image noise is increased due to image enhancement, the method adopts median filtering for denoising, and then extracts image characteristic information through a convolutional neural network. In order to improve the classification effect, the patent provides a hybrid model combining a convolutional neural network and an SVM. The method mainly adopts an SVM classifier to replace the last softmax layer of a trained convolutional neural network, and effectively combines the advancement of the convolutional neural network and the SVM to classify the icing image.
The embodiment of the invention provides a power grid line insulator damage image identification method based on a hybrid neural network. The above description of the embodiments is only for the purpose of helping understanding the method of the present patent and its core ideas; while the specification should not be considered as limiting the present patent, it will be appreciated by those skilled in the art that various modifications, equivalent alterations, improvements and the like can be made without inventive faculty.

Claims (4)

1. A power grid insulator damage image identification method based on a hybrid neural network is characterized by comprising the following steps:
preprocessing the obtained power grid line insulator damage image to extract a characteristic diagram;
convolving a feature map by convolution layers of a hybrid neural network, the feature map being represented by equation one (1);
Figure FDA0002854196970000011
wherein,
Figure FDA0002854196970000012
is the kth characteristic diagram, W, of the ith layeri kIs a weight matrix Wi k，
Figure FDA0002854196970000013
Is a bias term;
a convolutional layer pooled feature map through a hybrid neural network, the convolutional layer comprising neurons identified by equation two (2),
Figure FDA0002854196970000014
the function f () is an activation function and is used for realizing the nonlinear transformation of the characteristic value, and the function down () is a down-sampling function and is used for realizing the size transformation of the characteristic diagram by taking the maximum value or the average value;
converting the output values of the multi-classification into probability distribution which ranges from [0,1] and is added to 1 by adopting a formula III (3);
Figure FDA0002854196970000015
wherein, σ (z)i) The number of output nodes is C;
the cross entropy loss of the network is solved through the calculation result of the formula four (4),
Figure FDA0002854196970000016
wherein N is the number of nodes, alphaiTo specify the output value of the node, ajFor output values of different classes of tags, piAn output probability for each node;
the decision classification is carried out by adopting a formula five (5),
Figure FDA0002854196970000021
wherein alpha isiFor penalty coefficients calculated by the Lagrangian transformation, xiAs a feature space of the input, yiB is the label of the class to which the input sample belongs, and b is the offset of the classification plane.
2. The power grid insulator breakage image identification method based on the hybrid neural network as claimed in claim 1, wherein the preprocessing includes performing gray scale transformation using formula six (6),
Gray＝0.299R(x,y)+0.587G(x,y)+0.114B(x,y)
R(x,y)＝G(x,y)＝B(x,y)＝Gray
where Gray represents the Gray value of the pixel, R (x, y) represents the red component, G (x, y) represents the green component, and B (x, y) represents the blue component.
3. The grid insulator breakage image recognition method based on the hybrid neural network as claimed in claim 2, wherein a contrast ratio stretching transformation is performed by using a formula seven (7),
Figure FDA0002854196970000022
where r represents the gray level of the input image, s is the corresponding gray level value in the output image, and E is the slope.
4. The grid insulator damage image identification method based on the hybrid neural network as claimed in any one of claims 1-3, wherein the median filtering is performed by using equation eight (8),
g(x,y)＝median{f(x-k,y-l),(k,l)∈W} (8)
wherein, (x, y) is the coordinates of the pixel points, f (x, y) is the gray value of the pixel points, (k, l) is the set of the distances between all the pixel points in the pixel window and the original pixel points, and W is a constant.
CN202011539694.8A 2020-12-23 2020-12-23 Power grid insulator damage image identification method based on hybrid neural network Pending CN113221614A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011539694.8A CN113221614A (en) 2020-12-23 2020-12-23 Power grid insulator damage image identification method based on hybrid neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011539694.8A CN113221614A (en) 2020-12-23 2020-12-23 Power grid insulator damage image identification method based on hybrid neural network

Publications (1)

Publication Number Publication Date
CN113221614A true CN113221614A (en) 2021-08-06

Family

ID=77085860

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011539694.8A Pending CN113221614A (en) 2020-12-23 2020-12-23 Power grid insulator damage image identification method based on hybrid neural network

Country Status (1)

Country Link
CN (1) CN113221614A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113449769A (en) * 2021-05-18 2021-09-28 内蒙古工业大学 Power transmission line icing identification model training method, identification method and storage medium
CN114111647A (en) * 2021-12-06 2022-03-01 国网山东省电力公司电力科学研究院 Artificial intelligence-based method and system for measuring damaged area of insulator umbrella skirt
CN114266720A (en) * 2021-11-08 2022-04-01 国网河南省电力公司电力科学研究院 Power wire fault detection method and system based on hybrid neural network

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108734691A (en) * 2018-03-22 2018-11-02 广东电网有限责任公司清远供电局 A kind of transmission line of electricity defect image recognition methods
CN110288571A (en) * 2019-06-06 2019-09-27 中南大学 A kind of high iron catenary insulator method for detecting abnormality based on image procossing
CN111507972A (en) * 2020-04-20 2020-08-07 南京航空航天大学 Tunnel surface defect detection method combining convolutional neural network and support vector machine

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108734691A (en) * 2018-03-22 2018-11-02 广东电网有限责任公司清远供电局 A kind of transmission line of electricity defect image recognition methods
CN110288571A (en) * 2019-06-06 2019-09-27 中南大学 A kind of high iron catenary insulator method for detecting abnormality based on image procossing
CN111507972A (en) * 2020-04-20 2020-08-07 南京航空航天大学 Tunnel surface defect detection method combining convolutional neural network and support vector machine

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
周筑博 等: "基于深度卷积神经网络的输电线路可见光图像目标检测", 《液晶与显示》, vol. 33, no. 4, 30 April 2018 (2018-04-30), pages 317 - 325 *
张华 等: "《遥感数据可靠性分类方法研究》", 北京航空航天大学出版社, pages: 161 - 163 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113449769A (en) * 2021-05-18 2021-09-28 内蒙古工业大学 Power transmission line icing identification model training method, identification method and storage medium
CN114266720A (en) * 2021-11-08 2022-04-01 国网河南省电力公司电力科学研究院 Power wire fault detection method and system based on hybrid neural network
CN114266720B (en) * 2021-11-08 2024-04-23 国网河南省电力公司电力科学研究院 Power wire fault detection method and system based on hybrid neural network
CN114111647A (en) * 2021-12-06 2022-03-01 国网山东省电力公司电力科学研究院 Artificial intelligence-based method and system for measuring damaged area of insulator umbrella skirt

Similar Documents

Publication Publication Date Title
CN108961235B (en) Defective insulator identification method based on YOLOv3 network and particle filter algorithm
CN108960140B (en) Pedestrian re-identification method based on multi-region feature extraction and fusion
CN109118479B (en) Capsule network-based insulator defect identification and positioning device and method
CN113221614A (en) Power grid insulator damage image identification method based on hybrid neural network
CN112233073A (en) Real-time detection method for infrared thermal imaging abnormity of power transformation equipment
CN109635636B (en) Pedestrian re-identification method based on fusion of attribute characteristics and weighted blocking characteristics
CN110728694B (en) Long-time visual target tracking method based on continuous learning
Mussina et al. Multi-modal data fusion using deep neural network for condition monitoring of high voltage insulator
CN108108772A (en) A kind of insulator contamination condition detection method based on distribution line Aerial Images
She et al. Insulator breakage detection utilizing a convolutional neural network ensemble implemented with small sample data augmentation and transfer learning
CN110414616B (en) Remote sensing image dictionary learning and classifying method utilizing spatial relationship
CN110837865A (en) Domain adaptation method based on representation learning and transfer learning
Pei et al. UV‐flashover evaluation of porcelain insulators based on deep learning
CN110211127A (en) Image partition method based on bicoherence network
Li et al. Stochastic configuration networks for self-blast state recognition of glass insulators with adaptive depth and multi-scale representation
CN112749633A (en) Separate and reconstructed individual radiation source identification method
CN110929611A (en) Modeling method of power transmission line icing thickness prediction model based on PR-KELM
CN114359167A (en) Insulator defect detection method based on lightweight YOLOv4 in complex scene
CN116485802B (en) Insulator flashover defect detection method, device, equipment and storage medium
CN117853596A (en) Unmanned aerial vehicle remote sensing mapping method and system
CN113283371A (en) Local discharge feature extraction and classification method based on BRISK features
CN116385950A (en) Electric power line hidden danger target detection method under small sample condition
CN117197517A (en) Insulator pollution classification method and device, electronic equipment and medium
CN113538382B (en) Insulator detection algorithm based on non-deep network semantic segmentation
Qiu et al. Hydrophobicity Classification of Composite Insulators Based on Light‐Weight Convolutional Neural Networks

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20210806