Grounding grid topology identification and classification method based on a small number of measured samples

Abstract

The application provides a grounding grid topology identification and classification method based on a small amount of measured samples. The method is based on a measured sample set, uses a convolutional neural network for image recognition in a deep learning algorithm to realize the identification and classification of the grounding grid topology in the image. First, a sample library is established, a mature commercial simulation software is used to establish an electromagnetic response image sample library according to the parameter conditions of the measured environment, and through the way of transfer learning, the style of the measured noise image is transferred to the simulation sample library established according to the measured conditions, forming a measured image sample library. Then, part of the sample set is used as a training set, a neural network is established and trained, the grounding grid topology identification method is researched, and the extraction of the complex features in the grounding grid electromagnetic response image is realized, and the identification and classification are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of grounding grid topology identification, and particularly relates to a grounding grid topology identification classification method based on a small amount of measured samples. BACKGROUND

[0002] At present, the mainstream detection method of grounding grid topology structure is electromagnetic induction method, which can roughly judge the grounding grid topology structure and breakpoint by injecting current and analyzing the distribution characteristics of the ground electromagnetic field. However, due to the complex electromagnetic interference in the substation environment, the instrument detected grounding grid ground magnetic field intensity distribution chart contains high noise, and the topology structure is not clear. If the artificial identification method is used, the identification process will be limited by the subjective experience of the staff, and the identification result of the topology structure is not only low in accuracy but also not objective and real.

[0003] The convolutional neural network has the advantages of strong applicability, simultaneous feature extraction and classification, strong generalization ability, and less training parameters for global optimization. Due to the large area of the substation, the time for measuring the whole substation by using the instrument is long, the measured samples are few, and a large number of pictures cannot be obtained in a short time. However, by applying the advantages of the convolutional neural network and using the grounding grid electromagnetic response data as the driving, the identification of the topology structure can be made simple, accurate and objective, and the purpose of real-time intelligent identification of the grounding grid topology structure is achieved. Therefore, a method is needed to construct a sample library containing a large number of measured samples according to a small amount of measured samples to meet the application of the convolutional neural network, achieve the purpose of real-time intelligent identification of the grounding grid topology structure, and make full preparation for further identifying the position of the grounding grid conductor breakpoint and the corrosion degree through image recognition. SUMMARY

[0004] Based on the above problems, the present application provides a grounding grid topology structure identification classification method based on a small amount of measured samples, which comprises the following steps:

[0005] S100, based on a grounding grid simulation platform, a grounding grid physical model of different topology structures is established, each parameter is set according to the conditions at the time of measurement, grounding grid electromagnetic response data is obtained by simulation calculation, a magnetic field intensity distribution chart is drawn, and a basic sample library is established;

[0006] S200, the magnetic field intensity distribution chart is processed to further expand the sample size; the neural network is trained, and the method of transfer learning is used to perform style transfer of electromagnetic noise images based on a part of grounding grid measured response images, to obtain images containing electromagnetic noise measured in the actual substation environment of the grounding grid of different topology structures, and a measured image sample library is established;

[0007] S300, a neural network is established and trained according to the requirements of the measured image sample library and the grounding grid topology, so as to form a classification model; according to the requirements of image recognition, the input layer, convolution layer, full connection layer and output layer of the network are designed, and the related parameters are modified; a part of the images in the sample library are used to train the network, and the remaining part of the images are used as a test set to test the trained network and give result analysis.

[0008] Further, the step S200 specifically comprises:

[0009] S210, size normalization processing is performed on all images in the basic sample library, and the number of images in each category is increased to 6 times of the original number, so as to expand the basic sample library;

[0010] S220, a VGG-19 deep learning network is used as a neural network for image migration, part of the measured samples are used as a style source for style migration, a Gram matrix is used to model and extract the style in the image, and a slow image reconstruction method is used to update the pixel value of the reconstructed image in a gradient descent manner, so that the Gram matrix is close to the Gram matrix of the style image, and the Gram matrix is defined as:

[0011]

[0012] In the formula, i, j, and k are positions of the activation item, i represents the number of rows of the matrix, j represents the number of columns of the matrix, k represents different channels, l is the number of layers, F is a vectorized feature mapping function, represents the value of the vectorized feature mapping function of the lth layer in the ith row and the kth channel, represents the value of the vectorized feature mapping function of the lth layer in the jth column and the kth channel.

[0013] The high-level feature map extracted by the VGG network is used to represent the content information of the image.

[0014] S230, the parameters of the neural network are set, and the content loss function is defined as:

[0015]

[0016] In the formula, l represents the layer used to extract the style feature, Q l and P l respectively represent the feature map of the generated picture and the content picture at the lth layer, M l represents the size of the feature map, N l is the number of different feature mappings in layer l, M l is the capacity of the feature mapping in layer l, and respectively represent the feature map of the generated picture and the content picture in the lth layer, the i th row and the j th column;

[0017] The style loss function is defined as follows:

[0018]

[0019] wherein,

[0020] In the formula, ω l is a correlation coefficient, that is, the weight of the layer; G l ij and A l ij respectively represent the value of the feature map Gram matrix of the generated picture and the style picture in the lth layer, the i th row and the j th column;

[0021] All images in the simulation sample library are input into the neural network as content images, each image makes the neural network generate a bottom picture composed of random noise, and the bottom picture is iteratively updated by calculating the style loss and the content loss to generate a measured picture simulating the field environment.

[0022] Further, the step S300 specifically comprises:

[0023] S310, determining that the neural network used for recognition and classification is an AlexNet network, adjusting the image size of the input layer to be consistent with the sample library; determining the types of different topological structures, adding a corresponding number of data labels to the output layer, and setting the output layer to be the same number of neurons as the data labels;

[0024] S320, modifying the middle layer, replacing the LRN layer with the BN layer, creating a competition mechanism for the activity of local neurons, and according to the structure of the convolutional neural network, pre-processing the output data of the previous layer: all normalized to data with a mean of 0 and a variance of 1 for subsequent input; if the input x is d-dimensional, then:

[0025] x = (x (1) x (2) ...x (d) )

[0026] After normalizing each dimension, we have

[0027]

[0028] Learnable parameters γ and β are introduced, and the values of the parameters γ and β are as shown in the following formula, and y is used to replace x to transform and reconstruct:

[0029]

[0030]

[0031] β (k) = E [x (k) ]

[0032] S330, reduce the number of convolution kernels of the first three layers, set the convolution step of the first three layers, the convolution kernel size remains unchanged, the noise and the features in the image which do not have representativeness in the training set are fitted together; reduce a full connection layer, and reduce the number of neurons in the remaining two full connection layers, add a dropout layer after each full connection layer;

[0033] S340, design the training process of the network, the network training is divided into four steps of loading data set, setting training parameters, network initialization and batch iteration operation;

[0034] S350, randomly take a part of images from the sample library as a test set, input into the trained network, and return the corresponding prediction data label; define the evaluation standard of the grounding net topology structure recognition accuracy as:

[0035]

[0036] In the formula, R rec is the success rate of recognition, Z a is the number of pictures of successful recognition, and Z t is the number of all pictures.

[0037] Further, the step S340 specifically comprises:

[0038] S341, the loading data set specifically comprises: randomly taking a part of images from the sample library as a test set, and dividing the remaining images into a training set and a validation set according to a certain proportion;

[0039] S342, the setting training parameters specifically comprises: using a random gradient descent method with momentum as a solver, and setting an initial learning rate as p;

[0040] S343, the network initialization specifically comprises: reasonable setting of initialization parameters;

[0041] S344, the batch iteration operation specifically comprises: updating the weight and bias, setting the verification frequency, the total number of training rounds, the number of iterations, outputting the training error of the network and the recognition accuracy of the network on the training set and the validation set, and saving the network after training.

[0042] Compared with the prior art, the grounding net topology structure recognition and classification method designed by the application can recognize and classify the topology structure of simulated or measured images, and facilitate the start of the next grounding net fault diagnosis work. BRIEF DESCRIPTION OF DRAWINGS

[0043] The application will be further described below in conjunction with the accompanying drawings and tables and examples, in which:

[0044] Figure 1 is a flow chart of a grounding grid topology identification and classification method of the application;

[0045] Figure 2 is a whole structure diagram of a neural network of an embodiment of the application. DETAILED DESCRIPTION

[0046] To make the object, technical scheme and advantages of the application clearer, the embodiments of the application will be further described below in conjunction with the accompanying drawings.

[0047] Reference Figure 1 and Figure 2 , Figure 1 is a flow chart of a grounding grid topology identification and classification method of the application, Figure 2 is a whole structure diagram of a neural network of an embodiment of the application.

[0048] A grounding grid topology identification and classification method comprises the following steps:

[0049] S100, based on a grounding grid simulation platform, performing a grounding grid electromagnetic response simulation experiment, establishing a grounding grid physical model of different topologies, setting parameters such as soil type, current injection frequency, size and mode according to the conditions at the time of measurement, simulating and calculating grounding grid electromagnetic response data, and drawing a magnetic field intensity distribution map to establish a basic sample library;

[0050] S200, processing the magnetic field intensity distribution map to further expand the sample size; training a neural network, using a transfer learning method, based on part of the grounding grid measured response images, performing style transfer of electromagnetic noise images to obtain images containing electromagnetic noise measured by the grounding grid of different topologies in the actual substation environment, and establishing a measured image sample library;

[0051] S210, performing size normalization processing on all images in the basic sample library, first rotating each image in the basic sample library by 90°, 180° and 270° respectively, then using horizontal flipping, vertical flipping and random shifting methods on the original image to expand the sample library and increase the number of images of each category to 6 times the original, and expanding the basic sample library;

[0052] S220, using VGG-19 deep learning network as a neural network used for image migration, using part of the measured sample as the style source of style migration, using Gram matrix to model and extract the style in the image, and then using slow image reconstruction method to update the pixel value of the reconstructed image in the form of gradient descent, so that the Gram matrix is close to the Gram matrix of the style image, and the Gram matrix is defined as:

[0053]

[0054] Where i, j, k are the positions of the activation term, i represents the number of rows of the matrix, j represents the number of columns of the matrix, k represents different channels, l is the number of layers, F is the vectorized feature mapping function, represents the value of the vectorized feature mapping function in the i-th row and the k-th channel of the l-th layer, represents the value of the vectorized feature mapping function in the j-th column and the k-th channel of the l-th layer.

[0055] The high-level feature map extracted by the VGG network represents the content information of the image;

[0056] S230, setting the parameters of the neural network, and defining the content loss function as:

[0057]

[0058] Where l represents the layer used to extract the style feature, Q l and P l represent the feature map of the generated image and the content image at layer l, M l represents the size of the feature map, N l is the number of different feature mappings in layer l, M l is the capacity of the feature mapping in layer l, and represent the feature map of the generated image and the content image at layer l, i-th row and j-th column;

[0059] The style loss function is defined as follows:

[0060]

[0061] Where,

[0062] Where ω l is the correlation coefficient, i.e. the weight of the layer; G l ij and A l ij represent the value of the feature map Gram matrix of the generated image and the style image at layer l, i-th row and j-th column;

[0063] The high-level feature map extracted by the VGG network represents the content information of the image, and the extracted high-level feature map of the VGG network for the base map is close to the high-level feature map of the target image to achieve content similarity, so that the final reconstructed result image has both the style of the style image and the content of the content image;

[0064] S230, setting the parameters of the neural network, in order to quickly calculate the loss and achieve the required style transfer effect, the epoch round is set to 20, the learning rate is set to 0.03, and the content loss function is defined as:

[0065]

[0066] In the formula, l represents the layer used to extract the style feature, Q l and P l respectively represent the feature map of the generated image and the content image at layer l, M l represents the size of the feature map, N l is the number of different feature mappings in layer l, M l is the capacity of the feature mapping in layer l, and respectively represent the feature map of the generated image and the content image at layer l, i row and j column.

[0067] The style loss function is defined as follows:

[0068]

[0069] Wherein,

[0070] In the formula, ω l is the correlation coefficient, that is, the weight of the layer; G l ij and A l ij respectively represent the value of the Gram matrix of the generated image and the style image at layer l, i row and j column.

[0071] All images in the simulation sample library are input into the neural network as content images, and each image makes the neural network generate a base image composed of random noise. By calculating the style loss and content loss, the base image is iteratively updated to make it similar in style and texture to the style image and similar in content to the content image, generating a measured image simulating the field environment. A measured image sample library is established based on a small number of measured samples.

[0072] S300, a neural network is established and trained according to the requirements of the measured image sample library and the grounding grid topology, and a classification model is formed; according to the requirements of image recognition, the input layer, convolution layer, full connection layer and output layer of the network are designed, and the related parameters are modified; a part of the images in the sample library are used to train the network, and the remaining part of the images are used as a test set to test the trained network and give result analysis.

[0073] S310, the neural network used for recognition and classification is determined as an AlexNet network, the image size at the input layer is adjusted to be consistent with the sample library; the types of different topologies are determined, data labels corresponding to the output layer are added, and the output layer is correspondingly set to the same number of neurons as the data labels;

[0074] In this embodiment, the image size at the input layer is adjusted to 227x227x3;

[0075] S320, the middle layer is modified, the LRN layer is replaced by the BN layer, the LRN is a non-trainable layer, which is used for square normalization of pixel values in the feature map in the local field. A competition mechanism is created for the activity of local neurons, so that the value with relatively large response becomes relatively large, and other neurons with small feedback are suppressed, thereby enhancing the generalization ability of the model. The BN is a trainable layer, which is usually located after the convolution layer. According to the structure of the convolutional neural network, the output data of the previous layer are preprocessed, all normalized to data with mean value of 0 and variance of 1 for subsequent input; if the input x is d-dimensional, then

[0076] x=(x (1) x (2) ...x (d) )

[0077] Each dimension is normalized, then

[0078]

[0079] If only the above two formulas are used to normalize the feature map to output data, the feature map extracted by the convolution kernel will be changed, and the learnable parameters γ and β are introduced, and the values of the parameters γ and β are as shown in the following formula, and y is used to replace x to transform and reconstruct:

[0080]

[0081]

[0082] β (k) =E[x (k) ]

[0083] The parameters γ and β can be learned, so that the batch normalization layer can also be deep learning through back propagation, by updating the parameters, to restore the extracted feature map; after introducing the batch normalization layer, the dependence on initialization is greatly reduced, a larger initial learning rate can be used, and the network has the characteristics of fast convergence training.

[0084] S330, for the requirement of grounding grid topology structure recognition classification with less image categories and less sample library content, the middle layer is modified. Two convolution layers are reduced to reduce network parameters and improve calculation speed without affecting training accuracy; the number of convolution kernels in the first three layers is reduced and set to 16, 32 and 16, the convolution step of the first three layers is set to 4, 1 and 1, and the convolution kernel size remains unchanged, so as to avoid fitting the noise and features in the training set and images that do not have representative features together, and reduce the generalization ability of the network; one full connection layer is reduced, and the number of neurons in the remaining two full connection layers is reduced, which are respectively set to 2048 and 512 neurons, and in order to further prevent network overfitting, a dropout layer is added after each full connection layer;

[0085] S340, the training process of the network is designed, and the network training is divided into four steps of loading data set, setting training parameters, network initialization and batch iteration operation;

[0086] S341, loading data set is specifically: 20% of the images in the sample library are randomly taken as the test set, and the remaining images are divided into training set and validation set according to the ratio of 9:1;

[0087] S342, setting training parameters is specifically: the invention selects the random gradient descent method with momentum as the solver, the initial learning rate represents the amplitude of weight update at each iteration, and the initial learning rate is set to 0.001 through experimental test;

[0088] S343, network initialization is specifically the reasonable setting of initialization parameters, which can improve the convergence speed of the network, and the invention selects random initialization;

[0089] S344, batch iteration operation is specifically: updating the weight and bias, setting the verification frequency to 50, the total number of iterations of 5 training rounds is 4320, the training error of the network and the recognition accuracy of the network on the training set and the validation set are output once every 50 iterations, and the network is saved after training.

[0090] S350, 20% of the images in the sample library are randomly taken as the test set, input into the trained network, and return the corresponding predicted data label; the evaluation standard of grounding grid topology structure recognition accuracy is defined as:

[0091]

[0092] wherein R rec is the success rate of recognition, Z a is the number of pictures of successful recognition, Z t is the number of all pictures. The recognition accuracy of classification is defined by the criterion.

[0093] The embodiment is simulated by software simulation, sample library expansion and image migration. The simulation sample library is established according to the parameter conditions of the measured environment, and the images detected in the field are migrated to the images in the sample library. The images containing electromagnetic noise measured by the grounding net in the actual substation environment are simulated, the field environment is simulated, and the measured image sample library is established. The improved neural network is used to carry out the research on the grounding net topology recognition method, and the extraction of the complex features in the electromagnetic response image of the grounding net is realized, and the classification is recognized.

[0094] The above merely describes preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for identifying and classifying grounding grid topology based on a small number of measured samples, characterized in that, The method comprises the following steps: S100, based on a grounding grid simulation platform, a grounding grid physical model of different topological structures is established, each parameter is set according to the condition at the time of measurement, electromagnetic response data of the grounding grid are obtained through simulation calculation, a magnetic field intensity distribution map is drawn, and a basic sample library is established; S200, the magnetic field intensity distribution map is processed to further expand the sample size; A neural network is trained, a method of transfer learning is used, a part of the measured response images of the grounding grid are used as the basis, the style of the electromagnetic noise images is transferred, images containing electromagnetic noise measured in the actual substation environment of the grounding grid of different topological structures are obtained, and a measured image sample library is established; S300, a neural network is established and trained according to the requirements of the measured image sample library and grounding grid topological structure identification, a classification model is formed, the input layer, convolution layer, full connection layer and output layer of the network are designed according to the requirements of image recognition, and the related parameters are modified; Part of the images in the sample library are used to train the network, the remaining part of the images are used as a test set, the trained network is tested, and a result analysis is given; Step S200 specifically comprises: S210, all images in the basic sample library are subjected to size normalization processing, and the number of images in each category is increased to 6 times the original number, and the basic sample library is expanded; S220, a VGG-19 deep learning network is used as the neural network for image transfer, part of the measured samples are used as the style source for style transfer, a Gram matrix is used to model and extract the style in the image, and a slow image reconstruction method is used to update the pixel value of the reconstructed image in a gradient descent manner, so that the Gram matrix is close to the Gram matrix of the style image, and the Gram matrix is defined as: where i, j, k are the position of the activation term, i represents the row number of the matrix, j represents the column number of the matrix, k represents different channels, l is the layer number, F is a vectorized feature mapping function, represents the value of the vectorized feature mapping function of the lth layer at the i-th row and the k-th channel, represents the value of the vectorized feature mapping function of the lth layer at the j-th column and the k-th channel; The high-level feature map extracted by the VGG network is used to represent the content information of the image; S230, the parameters of the neural network are set, and the content loss function is defined as: where l represents a layer used to extract style features, Ql and Pl represent feature maps of the generated picture and the content picture at layer l, respectively, Ml represents the size of the feature maps, Nl is the number of different feature maps in layer l, and Ml is the capacity of the feature maps in layer l, and represent the feature maps of the generated picture and the content picture at layer l, respectively, and represent the feature maps of the generated picture and the content picture at the i-th row and the j-th column of layer l, respectively. The style loss function is defined as follows: wherein In the formula, ωl is a correlation coefficient, that is, the weight of the layer; Glij and Alij represent the values of the feature map Gram matrix of the generated picture and the style picture in the l layer, respectively; All images in the simulation sample library are input into the neural network as content images, each image makes the neural network generate a bottom picture composed of random noise, the bottom picture is iteratively updated by calculating the style loss and the content loss, and a measured picture simulating the on-site environment is generated.

2. The method of claim 1, wherein, The step S300 specifically comprises: S310, the neural network used for identification and classification is determined as an AlexNet network, the image size of the input layer is adjusted to be consistent with the sample library; the categories of different topological structures are determined, a corresponding number of data labels are added to the output layer, and the output layer is correspondingly set to the same number of neurons as the data labels; S320, modify the intermediate layer, replace the LRN layer with the BN layer, create a competition mechanism for the activity of local neurons, and preprocess the output data of the previous layer according to the structure of the convolutional neural network: all normalized to mean 0, variance 1 for subsequent input; input x is d dimension, then: Each dimension is normalized, that is, Introducing learnable parameters ​ and β , the values of the parameters ​ and β are given by y instead of x transformation reconstruction: S330, reduce the number of convolution kernels of the first three layers, set the convolution step of the first three layers, keep the convolution kernel size unchanged, fit the noise and the features in the image which are not representative in the training set; reduce one fully connected layer, and reduce the number of neurons in the remaining two fully connected layers, add a dropout layer after each fully connected layer; S340, design the training process of the network, the network training includes four steps of loading data set, setting training parameters, network initialization and batch iteration operation; S350, randomly take a part of images from the sample library as the test set, input them into the trained network, and return the corresponding prediction data label; define the evaluation standard of the grounding net topology structure recognition accuracy as: wherein R rec the success rate of recognition, Z a the number of pictures recognized successfully, Z t the number of all pictures.

3. The method of claim 2, wherein, The step S340 specifically comprises: S341, the loading data set specifically comprises: randomly taking a part of images from the sample library as the test set, and dividing the remaining images into the training set and the validation set according to a certain proportion; S342, the setting training parameter specifically uses a random gradient descent method with momentum as a solver, and sets an initial learning rate as p ; S343, the network initialization specifically comprises the reasonable setting of the initialization parameters; S344, the batch iteration operation specifically comprises updating the weight and the bias, setting the verification frequency, the total number of training rounds, the number of iteration intervals, outputting the training error of the network and the recognition accuracy of the network to the training set and the validation set, and saving the network after the training is completed.

Images