Power grid partial discharge defect identification method and device, terminal equipment and storage medium

Abstract

The invention discloses a power grid partial discharge defect identification method and apparatus, a terminal device and a storage medium. The method comprises the steps of converting a partial discharge signal to obtain a partial discharge image data set; inputting the partial discharge image data set into a preset partial discharge defect identification model to obtain a partial discharge defect identification result; wherein the training of the partial discharge defect identification model comprises the following steps: carrying out dimension conversion on a historical partial discharge signal to obtain a historical partial discharge image data set; inputting the historical partial discharge image data set into a preset information generative adversarial network for image expansion to obtain a first image data set; and inputting a training data set generated by the first image data set into the to-be-trained partial discharge defect recognition model for iterative training until a preset model loss function converges, and obtaining a trained partial discharge defect recognition model. According to the invention, the partial discharge defect type identification accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid systems, and in particular, to a method, device, terminal device, and storage medium for identifying partial discharge defects in a power grid. Background Art

[0002] Building a stable and reliable large-scale power network is the power guarantee and solid foundation for ensuring the economic prosperity of each region. In recent years, with the continuous growth of electricity consumption load and the booming development of new electricity consumption forms such as electric vehicles and household photovoltaics, the scale of distribution network equipment has been continuously increasing. As important components in the power system, distribution equipment such as transformers, high-voltage switchgears, and power cables, their working status is directly related to the overall reliability of the power grid. Faults in distribution equipment will lead to the disconnection of faulty lines, power outages for some users, and power grid paralysis. Partial discharge is a manifestation of insulation damage after long-term operation of electrical equipment, and it is also an important cause of insulation equipment aging. When partial discharge occurs, it will cause damage and erosion to the surrounding insulating medium. Although weak partial discharge has little damage to insulating equipment and will not be immediately apparent, long-term or intense partial discharge will quickly weaken the insulation strength, ultimately leading to insulation breakdown, triggering equipment failures and power grid safety accidents.

[0003] Currently, the partial discharge defect states of electrical equipment do not occur frequently and the probabilities of different fault types are different, resulting in low accuracy of the model classifier in the partial discharge fault assessment of electrical equipment due to lack of data, leading to low accuracy in identifying partial discharge defects. Summary of the Invention

[0004] Embodiments of the present invention provide a method, device, terminal device, and storage medium for identifying partial discharge defects in a power grid, which can effectively solve the problem that the accuracy of the model classifier in the partial discharge fault assessment of the prior art is not high due to lack of data, resulting in low accuracy in identifying partial discharge defects.

[0005] An embodiment of the present invention provides a method for identifying partial discharge defects in a power grid, including:

[0006] Obtain a partial discharge signal;

[0007] Perform dimensionality conversion on the partial discharge signal to obtain a two-dimensional partial discharge image dataset;

[0008] Input the partial discharge image dataset into a preset partial discharge defect recognition model for defect type recognition to obtain a partial discharge defect recognition result;

[0009] Among them, the training of the partial discharge defect recognition model includes:

[0010] Obtain historical partial discharge signals;

[0011] Perform dimensionality conversion on the historical partial discharge signals to obtain a historical partial discharge image dataset;

[0012] Input the historical partial discharge image dataset into a preset information generation adversarial network for image dataset augmentation to obtain an augmented first image dataset;

[0013] Generate a training dataset based on the first image dataset;

[0014] Input the training dataset into the local discharge defect recognition model to be trained, and perform iterative training until the preset model loss function converges to obtain a trained local discharge defect recognition model.

[0015] Furthermore, it also includes: obtaining the corresponding historical partial discharge defect recognition results and initial model parameters;

[0016] The step of inputting the training dataset into the local discharge defect recognition model to be trained and performing iterative training until the preset model loss function converges to obtain a trained local discharge defect recognition model includes:

[0017] Perform class probability prediction based on the current model parameters and the training dataset to obtain a predicted local discharge defect probability; where the initial current model parameters are the initial model parameters;

[0018] Take the local discharge defect type corresponding to the maximum predicted local discharge defect probability as the predicted local discharge defect recognition result;

[0019] Calculate the loss function value of the model loss function based on the predicted local discharge defect recognition result and the historical local discharge defect recognition result;

[0020] In the case where it is determined that the loss function value has not converged, update the current model parameters according to the model loss function; and use the updated current model parameters as the current model parameters for the next training.

[0021] Furthermore, the local discharge defect recognition model to be trained includes: an input layer, a convolutional layer, an activation layer, a pooling layer, a fully connected layer, and an output layer:

[0022] The input layer is used to receive the training dataset and transmit the training dataset to the convolutional layer;

[0023] The convolutional layer is used to extract local features of the discharge image based on the training dataset to obtain a first image feature map, and transmit the first image feature map to the activation layer;

[0024] The activation layer is used to perform a non-linear transformation based on the first image feature map to obtain a second image feature map, and transmit the second image feature map to the pooling layer;

[0025] The pooling layer is used to reduce the feature dimension based on the second image feature map to obtain a third image feature map with reduced dimension, and transmit the third image feature map to the fully connected layer;

[0026] The fully connected layer is used to perform a tiled connection based on the third image feature map to obtain a one-dimensional feature vector, and transmit the one-dimensional feature vector to the output layer;

[0027] The output layer is used to perform class probability prediction based on the current model parameters and the one-dimensional feature vector to obtain the predicted probability of partial discharge defects.

[0028] Further, the historical partial discharge signals are subjected to dimensional conversion to obtain a historical partial discharge image data set, including:

[0029] Perform normalization mapping according to the historical partial discharge signals and a preset mapping interval to obtain a first sampled value after mapping;

[0030] Map the first sampled value to a preset polar coordinate system, and calculate a first polar angle and a first polar radius;

[0031] Calculate based on the first polar angle and the first polar radius to obtain a first Gram angle and a field matrix;

[0032] Map the first Gram angle and the field matrix to the RGB channels to obtain first RGB channel values;

[0033] Generate a historical partial discharge image data set according to the first RGB channel values.

[0034] Further, the information generation adversarial network includes: a discriminator and a generator;

[0035] The training of the information generation adversarial network includes:

[0036] Obtain a noise vector of Gaussian distribution and an initial latent encoding;

[0037] Randomly select a number of the historical partial discharge image data sets as actual image samples;

[0038] Input the noise vector and the current latent encoding into the generator to generate a predicted image, obtaining a predicted image sample; where the current latent encoding at the beginning is the initial latent encoding;

[0039] Input the actual image sample and the predicted image sample into the discriminator to obtain a first probability for characterizing that the discriminator determines the actual image sample as an actual image sample, and a second probability for characterizing that the discriminator determines the predicted image sample as an actual image sample;

[0040] Calculate the mutual information between the current latent encoding and the predicted image sample according to the current latent encoding and the predicted image sample;

[0041] Calculate the loss function value of the cross-entropy loss function according to the first probability, the second probability, and the mutual information;

[0042] Determine whether the loss function value of the cross-entropy loss function converges;

[0043] If so, obtain the trained information generation adversarial network;

[0044] If not, update the current latent encoding according to the cross-entropy loss function, and use the updated current latent encoding as the current latent encoding for the next training.

[0045] Further, perform dimensional conversion on the partial discharge signal to obtain a two-dimensional partial discharge image dataset, including:

[0046] Perform normalization mapping according to the partial discharge signal and a preset mapping interval to obtain a second sampled value after mapping;

[0047] Map the second sampled value to a preset polar coordinate system, and calculate a second polar angle and a second polar radius;

[0048] Calculate according to the second polar angle and the second polar radius to obtain a second Gram angle and a field matrix;

[0049] Map the second Gram angle and the field matrix to the RGB channels to obtain second RGB channel values;

[0050] Generate a two-dimensional partial discharge image dataset according to the second RGB channel values.

[0051] Further, input the historical partial discharge image dataset into a preset information generation adversarial network for image dataset expansion to obtain an expanded first image dataset, including:

[0052] Input the historical partial discharge image dataset into a preset information generation adversarial network for image sample generation to obtain a new partial discharge image set;

[0053] Combine the new partial discharge image set and the historical partial discharge image dataset to obtain an expanded first image dataset.

[0054] As an improvement of the above solution, another embodiment of the present invention correspondingly provides a device for identifying local discharge defects in a power grid, including:

[0055] A signal acquisition module, configured to acquire local discharge signals;

[0056] An image conversion module, configured to perform dimensional conversion on the local discharge signals to obtain a two-dimensional local discharge image dataset;

[0057] A defect identification module, configured to input the local discharge image dataset into a preset local discharge defect identification model for defect type identification to obtain a local discharge defect identification result;

[0058] A model training module, configured to train the local discharge defect identification model;

[0059] Wherein, the model training module includes:

[0060] A historical signal acquisition unit, configured to acquire historical local discharge signals;

[0061] A historical image conversion unit, configured to perform dimensional conversion on the historical local discharge signals to obtain a historical local discharge image dataset;

[0062] A historical image augmentation unit, configured to input the historical local discharge image dataset into a preset information generation adversarial network for image dataset augmentation to obtain an augmented first image dataset;

[0063] A training dataset generation unit, configured to generate a training dataset according to the first image dataset;

[0064] A first training unit, configured to input the training dataset into the local discharge defect identification model to be trained for iterative training until a preset model loss function converges, so as to obtain a trained local discharge defect identification model.

[0065] Another embodiment of the present invention provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a method for identifying local discharge defects in a power grid as described in the above embodiment.

[0066] Another embodiment of the present invention provides a computer-readable storage medium, which includes a stored computer program. Wherein, when the computer program runs, it controls the device where the computer-readable storage medium is located to execute a method for identifying local discharge defects in a power grid as described in the above embodiment.

[0067] By implementing the present invention, at least the following beneficial effects are achieved:

[0068] The present invention provides a method, device, terminal device, and storage medium for identifying partial discharge defects in a power grid. The method can obtain partial discharge signals; convert the dimensions of the partial discharge signals to obtain a two-dimensional partial discharge image dataset; input the partial discharge image dataset into a preset partial discharge defect recognition model for defect type recognition to obtain a partial discharge defect recognition result. Among them, the training of the partial discharge defect recognition model includes: obtaining historical partial discharge signals; converting the dimensions of the historical partial discharge signals to obtain a historical partial discharge image dataset; inputting the historical partial discharge image dataset into a preset information generation adversarial network for image dataset expansion to obtain an expanded first image dataset; generating a training dataset according to the first image dataset; inputting the training dataset into the partial discharge defect recognition model to be trained for iterative training until a preset model loss function converges to obtain a trained partial discharge defect recognition model. By converting the dimensions of the partial discharge signals to obtain a partial discharge image dataset and converting one-dimensional signals into two-dimensional images, the features of the original signals are enhanced, and the accuracy of defect recognition is improved; at the same time, an expanded first image dataset is obtained through the information generation adversarial network, increasing the data diversity of model training, obtaining data with various fault types and various fault probabilities for training the partial discharge defect recognition model, making the data volume of model training more balanced, helping the model better learn the features of various partial discharge defect types, and improving the recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 is a flowchart of a method for identifying partial discharge defects in a power grid provided by an embodiment of the present invention;

[0070] Figure 2 is another flowchart of a method for identifying partial discharge defects in a power grid provided by an embodiment of the present invention;

[0071] Figure 3 is a schematic diagram of Gram angle field conversion provided by an embodiment of the present invention;

[0072] Figure 4 is a schematic diagram of a partial discharge signal waveform provided by an embodiment of the present invention;

[0073] Figure 5 is a schematic diagram of the structure of an infoGAN (information generation adversarial network) provided by an embodiment of the present invention;

[0074] Figure 6 is a schematic diagram of the structure of a partial discharge defect recognition model provided by an embodiment of the present invention;

[0075] Figure 7 It is a schematic diagram of the convolutional neural network structure provided by an embodiment of the present invention;

[0076] Figure 8 It is a schematic diagram of the structure of a device for identifying local discharge defects in a power grid provided by an embodiment of the present invention. Specific embodiments

[0077] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0078] See Figure 1 , which is a schematic flowchart of a method for identifying local discharge defects in a power grid provided by an embodiment of the present invention, including:

[0079] S1. Obtain local discharge signals;

[0080] Specifically, the local discharge signal is a set of voltage point signals that change with time.

[0081] S2. Perform dimensional conversion on the local discharge signal to obtain a two-dimensional local discharge image dataset;

[0082] Specifically, performing dimensional conversion on the local discharge signal to obtain a two-dimensional local discharge image dataset includes:

[0083] Perform normalization mapping according to the local discharge signal and a preset mapping interval to obtain a second sampled value after mapping;

[0084] Map the second sampled value to a preset polar coordinate system, and calculate to obtain a second polar angle and a second polar radius;

[0085] Perform calculations according to the second polar angle and the second polar radius to obtain a second Gram angle and a field matrix;

[0086] Map the second Gram angle and the field matrix to the RGB channels to obtain second RGB channel values;

[0087] Generate a two-dimensional local discharge image dataset according to the second RGB channel values.

[0088] In a preferred embodiment of the present invention, the second sampling value represents the sampling value after normalization mapping of the locally discharged signal collected in real time; the second polar angle and the second polar radius respectively represent the polar angle and the polar radius when the second sampling value is mapped to the polar coordinate system; the second Gram angle and the field matrix represent the Gram angle and the field matrix calculated from the second polar angle and the second polar radius; the second RGB channel value represents the RGB channel value when the second Gram angle and the field matrix are mapped to the RGB channel. The conversion process of converting the locally discharged signal into different dimensions is as follows Figure 3 shown. The locally discharged signal is used as the original time series signal X = {x1, x2, …, x i , …, x n}, which can be expressed as Figure 3 (a). First, according to the locally discharged signal X = {x1, x2, …, x i , …, x n} and the preset mapping interval, such as [-1, 1], normalization mapping is performed to obtain the second sampling value after mapping . The data after normalization mapping is as shown in Figure 3 (b). The normalization formula is as follows:

[0089]

[0090] where represents the i-th second sampling value, x max is the maximum value of the original time series signal X, and x min is the minimum value of the original time series signal X.

[0091] Then, the second sampling value is mapped to the preset polar coordinate system using inverse cosine coding to obtain the second polar angle φ i and the second polar radius r i of the i-th point, as shown in Figure 3 (c) for polar coordinate transformation:

[0092]

[0093] Next, by considering the angle sum and angle difference between each point in the polar coordinates to identify the time correlation within different time intervals, calculations are performed based on the second polar angle and the second polar radius, and the Gramian Angular Summation Field (GASF) is used to process the variables in the polar coordinates to obtain the second Gram angle and field matrix (G ASF matrix):

[0094] G ASF(i,j) = [cos(φ i + φ j )]

[0095]

[0096] Among them, φ i , φ j is the second sampling value and the polar angles of the i-th and j-th points in

[0097] Finally, map the second Gram angle and the field matrix to the RGB channels to obtain the second RGB channel values, that is, map the G ASF matrix to the pixel value range of 0 to 255, and copy this value to the three RGB channels. The second RGB channel values include: R (i,j) , G (i,j) , B (i,j) :

[0098]

[0099] Among them, G ASFmin is the minimum value of the G ASF matrix, and G ASFmax is the maximum value of the G ASF matrix. Then, according to the second RGB channel values, generate a partial discharge image dataset, such as Figure 3 (d).

[0100] During the feature extraction process of the partial discharge signal, it is easy to cause compression and loss of key fault information. By converting the one-dimensional partial discharge signal into a two-dimensional partial discharge image dataset with more complete information to retain the correlation in the time dimension of the data and "visually" identify and learn the data, the accuracy can be improved to a certain extent. After the above transformation, the time series signal with a length of n is converted into an n×n matrix symmetric along the main diagonal, and the integrity and time dependence of the original time series signal are retained. To reduce the complexity of the neural network, the partial discharge image dataset is uniformly scaled to an image with a size of 64×64×3.

[0101] S3. Input the partial discharge image dataset into a preset partial discharge defect recognition model for defect type recognition to obtain a partial discharge defect recognition result;

[0102] Preferably, input the converted partial discharge image dataset into a preset partial discharge defect recognition model for defect type recognition. The partial discharge defect types include floating electrode defect type, insulator air gap defect type, free metal particle defect type, and metal tip defect type. The partial discharge defect recognition results include: no partial discharge defect, floating electrode defect type, insulator air gap defect type, free metal particle defect type, and metal tip defect type.

[0103] Among them, the training of the partial discharge defect recognition model includes:

[0104] Obtain historical partial discharge signals;

[0105] Preferably, the historical partial discharge signals include four types of partial discharge defects, with 20 samples for each defect, a total of 80 samples. The specific information is shown in Table 1.

[0106] Table 1

[0107]

[0108]

[0109] Select a sample from the four types of partial discharge defects for visualization, and obtain the partial discharge signal as Figure 4 shown, Figure 4 where (a)-(d) are respectively: Figure 4 (a) Suspended discharge; Figure 4 (b) Insulator air gap; Figure 4 (c) Free metal particles; Figure 4 (d) Metal tip.

[0110] Perform dimensionality conversion on the historical partial discharge signals to obtain a historical partial discharge image dataset;

[0111] Preferably, performing dimensionality conversion on the historical partial discharge signals to obtain a historical partial discharge image dataset includes:

[0112] Perform normalization mapping according to the historical partial discharge signals and a preset mapping interval to obtain a first sampled value after mapping;

[0113] Map the first sampled value to a preset polar coordinate system, and calculate a first polar angle and a first polar radius;

[0114] Calculate according to the first polar angle and the first polar radius to obtain a first Gram angle and a field matrix;

[0115] Map the first Gram angle and the field matrix to the RGB channels to obtain first RGB channel values;

[0116] Generate a historical partial discharge image dataset according to the first RGB channel values.

[0117] In a preferred embodiment of the present invention, the first sampling value represents the sampling value of the historical partial discharge signal after normalized mapping; the first polar angle and the first polar radius respectively represent the polar angle and the polar radius after the first sampling value is mapped to a preset polar coordinate system; the first Gram angle and field matrix represent the Gram angle and field matrix calculated according to the first polar angle and the first polar radius; the first RGB channel value represents the RGB channel value after the first Gram angle and field matrix are mapped to the RGB channel. The process of converting the dimension of the historical partial discharge signal to obtain the historical partial discharge image dataset is similar to the process of converting the dimension of the partial discharge signal to obtain the partial discharge image dataset.

[0118] Input the historical partial discharge image dataset into a preset information generation adversarial network for image dataset augmentation to obtain the augmented first image dataset;

[0119] Specifically, the information generation adversarial network includes: a discriminator and a generator;

[0120] The training of the information generation adversarial network includes:

[0121] Obtain a noise vector of Gaussian distribution and an initial latent encoding;

[0122] Randomly select several of the historical partial discharge image datasets as actual image samples;

[0123] Input the noise vector and the current latent encoding into the generator to generate a predicted image to obtain a predicted image sample; where the current latent encoding at the beginning is the initial latent encoding;

[0124] Input the actual image sample and the predicted image sample into the discriminator to obtain a first probability for characterizing that the discriminator determines the actual image sample as an actual image sample, and a second probability for characterizing that the discriminator determines the predicted image sample as an actual image sample;

[0125] Calculate the mutual information between the current latent encoding and the predicted image sample according to the current latent encoding and the predicted image sample;

[0126] Calculate the loss function value of the cross-entropy loss function according to the first probability, the second probability, and the mutual information;

[0127] Determine whether the loss function value of the cross-entropy loss function converges;

[0128] If so, obtain the trained information generation adversarial network;

[0129] If not, update the current latent encoding according to the cross-entropy loss function, and use the updated current latent encoding as the current latent encoding for the next training.

[0130] Preferably, the Information Generative Adversarial Network (InfoGAN) includes a discriminator and a generator, as Figure 5 shown. Its principle is to constrain the relationship between the input noise vector z and the interpretable latent code c, so that c can contain interpretable information of the predicted image samples. The generator G and discriminator D of InfoGAN are both composed of convolutional neural networks. The input of the generator G combines the noise vector z of the Gaussian distribution and the latent code c, and the output is the predicted image sample. z can be understood as incompressible noise, while c can be understood as the interpretable latent code. By changing the latent code c, the change of the predicted image sample is changed. At the same time, there is a constraint relationship between c and the generator, that is, the mutual information I(c; G(z,c)) between c and the predicted image sample G(z,c). Therefore, InfoGAN can control image generation. The input of the discriminator D is the actual image sample and the predicted image sample. Its tail is connected to the auxiliary network C. C is essentially a classifier used to predict the latent code of the image sample. D and C use the same convolutional neural network. D outputs a judgment on the authenticity of the input image sample, and C is used to approximate the posterior probability P(c|x) and output the code c ′ . If the latent code c input by the generator can have a clear impact on the predicted image sample, that is, the correlation between the two is high, the output c of C ′ should be as consistent with c as possible. To improve the correlation between the latent code c and the generated sample G(z,c) and make a certain dimension of the latent code correspond to some semantic information of the predicted image sample, InfoGAN needs to maximize the mutual information I(c; G(z,c)) between the latent code c and the generated sample G(z,c).

[0131] The traditional Generative Adversarial Network (GAN) mainly consists of two modules, including the generator G and the discriminator D. The main task of the generator is to receive the noise z of the random distribution and make the sample G(z) output by itself consistent with the real sample distribution; the main task of the discriminator D is to receive the generated data G(z) of the generator and the real sample data x, and distinguish the authenticity of the received data. Ideally, the discriminator cannot judge whether the received data comes from the generated data G(z) or the real data x. When the model reaches the optimal state, the probability value output by the discriminator each time is 1 / 2, that is, the discriminator and the generator are alternately trained to reach the Nash equilibrium, and its loss is:

[0132]

[0133] In the formula, V(D,G) represents the cross-entropy loss, and P data(x) represents the distribution of real samples, and P z(z)It represents the noise distribution. E represents the expectation, D(x) is the probability that the discriminator D determines that x is a real sample, and D(G(z)) is the probability that the discriminator determines that the generated sample G(z) is a real sample.

[0134] For traditional GANs, completely random and continuous noise vectors are usually used as inputs. However, an obvious problem is that the generator has little constraint, making the generated data too free and unable to generate images of specified targets. To solve this problem, the Information maximizing generative adversarial nets (InfoGAN) introduces mutual information and combines GAN with feature learning, enabling the description and control of features without relying on external information.

[0135] In a preferred embodiment of the present invention, first obtain a noise vector of Gaussian distribution and an initial latent encoding; then randomly select several of the historical partial discharge image datasets as actual image samples; according to the noise vector, the current latent encoding, and the actual image samples, repeatedly execute the training operation until the preset cross-entropy loss function converges to obtain a trained information generative adversarial network; the training operation includes: first input the noise vector and the current latent encoding into the generator to generate a predicted image to obtain a predicted image sample; then input the actual image sample and the predicted image sample into the discriminator to obtain a first probability D(x) representing that the discriminator determines the actual image sample as an actual image sample, and a second probability D(G(z,c)) representing that the discriminator determines the predicted image sample as an actual image sample; calculate the mutual information I(c; G(z,c)) between the current latent encoding and the predicted image sample according to the current latent encoding and the predicted image sample; calculate the cross-entropy loss function according to the first probability, the second probability, and the mutual information:

[0136]

[0137] where λ is a hyperparameter, and V1(D,G) represents the cross-entropy loss function; P data(x) represents the distribution of the actual image samples; P z(x) represents the noise distribution of the noise vector; E represents the expectation; D(x) is the first probability determined by the discriminator D; D(G(z,c)) is the second probability determined by the discriminator D;

[0138] Update the current latent encoding according to the cross-entropy loss function, and use the updated current latent encoding as the current latent encoding for the next training.

[0139] In another preferred embodiment of the present invention, train the discriminator and update the gradient θ using the gradient ascent methodd : Randomly sample a set of noise vectors \(z = \{z (1) , z (2) , \ldots, z (m) \}\) from a Gaussian distribution, and input it together with a set of historical partial discharge image datasets \(x = \{x (1) , x (2) , \ldots, x (m) \}\) into the discriminator, and calculate the formula:

[0140] In the formula, \(D(x (i) )\) is the probability that the discriminator \(D\) determines \(x (i) \) to be an actual sample, that is, the first probability, and \(G(z (i) )\) represents the sample generated by the generator \(G\) when the input is the random vector \(z (i) \), that is, the predicted image sample. \(D(G(z i ))\) is the probability that the discriminator \(D\) determines the generated sample \(G(z (i) )\) to be a real sample, that is, the second probability. Train the generator, and use the gradient descent method to update the gradient \(\theta g : Randomly sample a set of noise vectors \(z = \{z (1) , z (2) , \ldots, z (m) \}\) from a Gaussian distribution, and then randomly sample a one-hot vector from a discrete random distribution and several random vectors from a continuous random distribution. Both are denoted as the latent encoding \(c\). After concatenating the above random vectors, input them into the generator \(G\), and calculate the formula:

[0141] In the formula, \(\lambda\) is a hyperparameter, generally set to a small number, and is set to \(0.1\) in this embodiment. \(G(z, c)\) represents the sample generated by the generator \(G\) when the input is the random noise vector \(z\) and the latent encoding \(c\). Repeat the training of the generator and the discriminator until the Nash equilibrium is reached, that is, the cross-entropy loss function converges, and obtain the trained information generation adversarial network.

[0142] Generate a training dataset according to the first image dataset

[0143] In a preferred embodiment of the present invention, according to the first image dataset and a preset division ratio, divide the first image dataset into a training dataset and a test dataset; divide it into a training dataset and a test dataset according to a preset division ratio of 4:1.

[0144] Input the training dataset into the local discharge defect recognition model to be trained, and perform iterative training until the preset model loss function converges, and obtain the trained local discharge defect recognition model;

[0145] Specifically, obtain the corresponding historical partial discharge defect recognition results and initial model parameters;

[0146] The step of inputting the training data set into the partial discharge defect recognition model to be trained for iterative training until the preset model loss function converges to obtain the trained partial discharge defect recognition model includes:

[0147] Predict the class probability according to the current model parameters and the training data set to obtain the predicted partial discharge defect probability; initially, the current model parameters are the initial model parameters;

[0148] Take the partial discharge defect type corresponding to the maximum predicted partial discharge defect probability as the predicted partial discharge defect recognition result;

[0149] Calculate the loss function value of the model loss function according to the predicted partial discharge defect recognition result and the historical partial discharge defect recognition result;

[0150] In the case where it is determined that the loss function value has not converged, update the current model parameters according to the model loss function; and use the updated current model parameters as the current model parameters for the next training.

[0151] In a preferred embodiment of the present invention, the partial discharge defect recognition model is as Figure 2 shown in the CNN model in, and training the partial discharge defect recognition model is like Figure 2 training the CNN model in. Although the convolutional neural network can automatically and effectively extract key features, the high performance of the convolutional neural network depends on a sufficient amount of samples with a uniform distribution among classes. However, the probability of insulation defects in actual electrical equipment is relatively low, resulting in difficulty in obtaining a large number of partial discharge defect samples with a uniform quantity distribution. To address this problem, in this embodiment, the Gramian angular field is used to convert the one-dimensional time series signal into an image, effectively enhancing the features of the original signal; the InfoGAN model is used to augment the original image data, which not only ensures that there are enough samples input for the convolutional neural network but also increases the diversity of the image sample set. At the same time, a multi-convolutional neural network is used to identify partial discharges, which not only avoids the subjectivity of manual feature extraction but also can automatically extract deep features to achieve accurate identification of partial discharge defect types.

[0152] Preferably, the partial discharge defect recognition model to be trained includes: an input layer, a convolutional layer, an activation layer, a pooling layer, a fully connected layer, and an output layer:

[0153] The input layer is used to receive the training data set and transmit the training data set to the convolutional layer;

[0154] The convolutional layer is used to extract local features of the discharge images according to the training data set, obtain a first image feature map, and transmit the first image feature map to the activation layer;

[0155] The activation layer is used to perform a non-linear transformation according to the first image feature map, obtain a second image feature map, and transmit the second image feature map to the pooling layer;

[0156] The pooling layer is used to reduce the feature dimension according to the second image feature map, obtain a third image feature map with reduced dimension, and transmit the third image feature map to the fully connected layer;

[0157] The fully connected layer is used to perform a flattened connection according to the third image feature map, obtain a one-dimensional feature vector, and transmit the one-dimensional feature vector to the output layer;

[0158] The output layer is used to predict the class probability according to the current model parameters and the one-dimensional feature vector, and obtain the predicted probability of local discharge defects.

[0159] In a preferred embodiment of the present invention, as Figure 6 shown, the input layer, as the starting position of the convolutional neural network, receives the original data and then sends the data to the convolutional layer and the pooling layer for feature extraction. In the convolutional neural network, the role of the convolutional layer is to extract features from the local area of the input data, and through the convolutional operation and the non-linear transformation of the activation function, generate the feature map output of this layer. The pooling layer is located after the convolutional layer, and its purpose is to reduce the dimension of the data and avoid the overfitting phenomenon. The fully connected layer is a transitional structure between the previous pooling layer and the Softmax classifier. The output features of the final pooling layer are flattened into a one-dimensional feature vector through the fully connected layer. The Softmax classifier can convert the neurons of the last fully connected layer into a probability distribution with a sum of 1 to predict the class label of the input data. The Softmax function is used in the multi-classification process, and its output represents the relative probabilities between different classes. It maps the outputs of multiple neurons to the interval (0,1) to perform multi-classification. It transforms the output values into a probability distribution that is positive and has a sum of 1 through the following formula:

[0160]

[0161] where V i is the output of the output unit in front of the classifier. i represents the class index, and the total number of classes is C. S i represents the ratio of the exponent of the current element to the sum of the exponents of all elements.

[0162] In another preferred embodiment of the present invention, the common partial discharge types of power distribution equipment include floating electrodes, insulator air gaps, free metal particles, and metal tips. Different from machine learning that requires manual extraction of features such as maximum values and kurtosis, a convolutional neural network can automatically extract shallow features and deep features of partial discharge defect types, thus accurately identifying partial discharge defect types and achieving end-to-end learning. That is, when inputting a partial discharge signal, after automatically completing feature extraction and classification inside the model, the category probability of the signal is directly given. The constructed convolutional neural network structure is as shown in Figure 7 Figure Figure 7 . The network consists of two convolutional layers, two pooling layers, and two fully connected layers. The optimizer is Adam, the learning rate is 0.001, and the number of iterations is 200 times. The first convolutional layer uses a 3×3 convolutional kernel, the stride is 1, and the padding is same. The first pooling layer uses a 2×2 pooling kernel, the stride is 2, and the padding is same. The kernel sizes and strides of the remaining convolutional layers and pooling layers remain unchanged. Finally, the fully connected layer stretches the obtained feature subgraph into a one-dimensional feature, and performs recognition and classification through a Softmax classifier to obtain the category probability. For example, the input layer receives an image with a size of (64, 64, 3), where 64×64 is the width and height of the image, and 3 represents the three RGB channels. The first convolutional module Conv2d, 32: A two-dimensional convolutional layer that uses 32 convolutional kernels to extract basic features of the image, such as edges and textures; BN: Batch normalization, which normalizes the input data to accelerate training and prevent gradient disappearance; ReLU: Activation function, introducing non-linearity; Maxpooling: Max pooling, which reduces parameters through downsampling (such as a 2×2 window) and retains the main features, and the output size becomes (32, 32, 32). In the second convolutional module, Conv2d, 64: Two-dimensional convolution, using 64 convolutional kernels to extract more complex features; BN: Batch normalization; ReLU: Activation function; Maxpooling: Downsampling, with an output size of (16, 16, 64); Dropout: Randomly discarding some neurons to prevent overfitting. In the fully connected layer module, Flatten: Flattens the multi-dimensional features into a one-dimensional vector (16×16×64 = 16384); Linear, 128: A fully connected layer that maps the 16384-dimensional input to 128 dimensions to extract high-level abstract features; BN: Batch normalization; ReLU: Activation function; Dropout: Prevent overfitting. In the output layer, Linear, 4: A fully connected layer that maps the 128 dimensions to 4 dimensions, corresponding to 4 categories; Softmax: Converts the output into category probabilities and outputs 4-dimensional category probabilities.

[0163] By implementing this embodiment, a partial discharge signal is obtained; the partial discharge signal is subjected to dimensionality conversion to obtain a two-dimensional partial discharge image dataset; the partial discharge image dataset is input into a preset partial discharge defect recognition model for defect type recognition to obtain a partial discharge defect recognition result; wherein, the training of the partial discharge defect recognition model includes: obtaining historical partial discharge signals; performing dimensionality conversion on the historical partial discharge signals to obtain a historical partial discharge image dataset; inputting the historical partial discharge image dataset into a preset information generation adversarial network for image dataset augmentation to obtain an augmented first image dataset; generating a training dataset according to the first image dataset; inputting the training dataset into the partial discharge defect recognition model to be trained for iterative training until a preset model loss function converges to obtain a trained partial discharge defect recognition model. By performing dimensionality conversion on the partial discharge signal to obtain a partial discharge image dataset and converting the one-dimensional signal into a two-dimensional image, the features of the original signal are enhanced, and the accuracy of defect recognition is improved; at the same time, an augmented first image dataset is obtained through the information generation adversarial network, increasing the data diversity of model training, obtaining data with various fault types and various fault probabilities for training the partial discharge defect recognition model, making the data volume of model training more balanced, helping the model better learn the features of various partial discharge defect types, and improving the recognition accuracy.

[0164] See Figure 8 , which is a schematic structural diagram of a device for identifying partial discharge defects in a power grid provided by an embodiment of the present invention, includes:

[0165] A signal acquisition module for acquiring partial discharge signals;

[0166] An image conversion module for performing dimensionality conversion on the partial discharge signal to obtain a two-dimensional partial discharge image dataset;

[0167] A defect recognition module for inputting the partial discharge image dataset into a preset partial discharge defect recognition model for defect type recognition to obtain a partial discharge defect recognition result;

[0168] A model training module for training the partial discharge defect recognition model;

[0169] Wherein, the model training module includes:

[0170] A historical signal acquisition unit for acquiring historical partial discharge signals;

[0171] A historical image conversion unit for performing dimensionality conversion on the historical partial discharge signals to obtain a historical partial discharge image dataset;

[0172] A historical image augmentation unit for inputting the historical partial discharge image dataset into a preset information generation adversarial network for image dataset augmentation to obtain an augmented first image dataset;

[0173] A training dataset generation unit for generating a training dataset according to the first image dataset;

[0174] A first training unit for inputting the training dataset into a to-be-trained partial discharge defect recognition model for iterative training until a preset model loss function converges to obtain a trained partial discharge defect recognition model.

[0175] The present invention provides a power grid partial discharge defect recognition device. According to a signal acquisition module, a partial discharge signal is acquired; according to an image conversion module, the partial discharge signal is subjected to dimensionality conversion to obtain a two-dimensional partial discharge image dataset; according to a defect recognition module, the partial discharge image dataset is input into a preset partial discharge defect recognition model for defect type recognition to obtain a partial discharge defect recognition result; according to a model training module, training of the partial discharge defect recognition model is performed; which includes: according to a historical signal acquisition unit, a historical partial discharge signal is acquired; through a historical image conversion unit, the historical partial discharge signal is subjected to dimensionality conversion to obtain a historical partial discharge image dataset; through a historical image augmentation unit, the historical partial discharge image dataset is input into a preset information generation adversarial network for image dataset augmentation to obtain an augmented first image dataset; in a training dataset generation unit, a training dataset is generated according to the first image dataset; according to a first training unit, the training dataset is input into a to-be-trained partial discharge defect recognition model for iterative training until a preset model loss function converges to obtain a trained partial discharge defect recognition model. By performing dimensionality conversion on the partial discharge signal to obtain a partial discharge image dataset, converting a one-dimensional signal into an image, the features of the original signal are enhanced, and the accuracy of defect recognition is improved; at the same time, through the information generation adversarial network, an augmented first image dataset is obtained, the data diversity for model training is increased, and data with various fault types and various fault probabilities are obtained for training the partial discharge defect recognition model, making the data volume for model training more balanced, helping the model to better learn the features of various partial discharge defect types, and improving the recognition accuracy.

[0176] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in the attached drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that they have a communication connection, which can be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art can understand and implement it without creative work.

[0177] Those skilled in the art can clearly understand that for the sake of convenience and brevity, the specific working process of the device described above can refer to the corresponding process in the foregoing method embodiments, and will not be described herein again.

[0178] Another embodiment of the present invention also provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a method for identifying local discharge defects in a power grid as described in the above embodiments. The terminal device can be a computing device such as a desktop computer, a notebook, a palm computer, and a cloud server. The terminal device may include, but is not limited to, a processor and a memory.

[0179] The processor may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The processor is the control center of the terminal device, and connects various parts of the entire terminal device through various interfaces and lines.

[0180] The memory can be used to store the computer program. By running or executing the computer program stored in the memory and invoking the data stored in the memory, the processor realizes various functions of the terminal device. The memory mainly includes a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for at least one function, etc.; the data storage area can store data created according to the use of the mobile phone, etc. In addition, the memory can include high-speed random access memory and can also include non-volatile memory, such as a hard disk, memory, plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, at least one magnetic disk storage device, flash device or other volatile solid-state storage devices.

[0181] Another embodiment of the present invention provides a computer-readable storage medium. The computer-readable storage medium includes a stored computer program. When the computer program runs, it controls the device where the computer-readable storage medium is located to execute a method for identifying grid partial discharge defects described in the above embodiment.

[0182] The storage medium is a computer-readable storage medium, and the computer program is stored in the computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above various method embodiments. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0183] The above is the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.

Claims

1. A method for identifying partial discharge defects in a power grid, characterized in that: include: Acquire partial discharge signals; Performing dimension conversion on the partial discharge signal to obtain a two-dimensional partial discharge image data set; Inputting the partial discharge image data set into a preset partial discharge defect recognition model to perform defect type recognition to obtain a partial discharge defect recognition result; The training of the partial discharge defect recognition model includes: Obtain historical partial discharge signals; Performing dimension conversion on the historical partial discharge signal to obtain a historical partial discharge image data set; Inputting the historical partial discharge image data set into a preset information generation adversarial network to expand the image data set to obtain an expanded first image data set; Generate a training data set according to the first image data set; The training data set is input into the partial discharge defect recognition model to be trained, and iterative training is performed until a preset model loss function converges to obtain a trained partial discharge defect recognition model.

2. A method for identifying partial discharge defects in a power grid as claimed in claim 1, characterized in that: Also includes: Obtain the corresponding historical partial discharge defect recognition results and initial model parameters; The step of inputting the training data set into the partial discharge defect recognition model to be trained, performing iterative training until a preset model loss function converges, and obtaining a trained partial discharge defect recognition model comprises: Performing a category probability prediction based on the current model parameters and the training data set to obtain a predicted partial discharge defect probability; wherein the initial current model parameters are initial model parameters; The partial discharge defect type corresponding to the maximum predicted partial discharge defect probability is used as the predicted partial discharge defect identification result; Calculate the loss function value of the model loss function according to the predicted partial discharge defect recognition result and the historical partial discharge defect recognition result; When it is determined that the loss function value has not converged, the current model parameters are updated according to the model loss function; and the updated current model parameters are used as the current model parameters for the next training.

3. A method for identifying partial discharge defects in a power grid as claimed in claim 2, characterized in that: The partial discharge defect recognition model to be trained includes: an input layer, a convolution layer, an activation layer, a pooling layer, a fully connected layer and an output layer: The input layer is used to receive the training data set and transmit the training data set to the convolution layer; The convolution layer is used to extract local features of the discharge image according to the training data set to obtain a first image feature map, and transmit the first image feature map to the activation layer; The activation layer is used to perform a nonlinear transformation according to the first image feature map to obtain a second image feature map, and transmit the second image feature map to the pooling layer; The pooling layer is used to reduce the feature dimension according to the second image feature map to obtain a third image feature map after the dimension is reduced, and transmit the third image feature map to the fully connected layer; The fully connected layer is used to perform tile connection according to the third image feature map to obtain a one-dimensional feature vector, and transmit the one-dimensional feature vector to the output layer; The output layer is used to perform category probability prediction based on the current model parameters and the one-dimensional feature vector to obtain a predicted partial discharge defect probability.

4. A method for identifying partial discharge defects in a power grid as claimed in claim 1, characterized in that: The historical partial discharge signal is dimensionally transformed to obtain a historical partial discharge image data set, including: Performing normalized mapping according to the historical partial discharge signal and a preset mapping interval to obtain a mapped first sampling value; Mapping the first sampling value to a preset polar coordinate system to calculate a first polar angle and a first polar diameter; Calculating according to the first polar angle and the first polar diameter to obtain a first Gram angle and a field matrix; Mapping the first Gram angle and field matrix to RGB channels to obtain first RGB channel values; A historical partial discharge image data set is generated according to the first RGB channel value.

5. A method for identifying partial discharge defects in a power grid as claimed in claim 1, characterized in that: The information generation adversarial network includes: a discriminator and a generator; The training of the information generation adversarial network includes: Obtain a Gaussian distributed noise vector and an initial latent code; Randomly selecting a number of the historical partial discharge image data sets as actual image samples; Inputting the noise vector and the current implicit code into the generator to generate a predicted image, and obtaining a predicted image sample; wherein the initial current implicit code is the initial implicit code; Inputting the actual image sample and the predicted image sample into the discriminator to obtain a first probability for characterizing that the discriminator determines that the actual image sample is the actual image sample, and a second probability for characterizing that the discriminator determines that the predicted image sample is the actual image sample; Calculating the mutual information between the current implicit code and the predicted image sample according to the current implicit code and the predicted image sample; Calculate a loss function value of a cross entropy loss function according to the first probability, the second probability, and the mutual information; Determine whether the loss function value of the cross entropy loss function converges; If so, the trained information generation adversarial network is obtained; If not, the current implicit code is updated according to the cross entropy loss function, and the updated current implicit code is used as the current implicit code for the next training.

6. A method for identifying partial discharge defects in a power grid as claimed in claim 1, characterized in that: The local discharge signal is dimensionally converted to obtain a two-dimensional local discharge image data set, including: Performing normalized mapping according to the partial discharge signal and a preset mapping interval to obtain a mapped second sampling value; Mapping the second sampling value to a preset polar coordinate system to calculate a second polar angle and a second polar diameter; Calculating according to the second polar angle and the second polar diameter to obtain a second Gram angle and a field matrix; Mapping the second Gram angle and field matrix to RGB channels to obtain second RGB channel values; A two-dimensional partial discharge image data set is generated according to the second RGB channel value.

7. A method for identifying partial discharge defects in a power grid as claimed in claim 1, characterized in that: The historical partial discharge image data set is input into a preset information generation adversarial network to expand the image data set to obtain an expanded first image data set, including: Inputting the historical partial discharge image data set into a preset information generation adversarial network to generate image samples to obtain a new partial discharge image set; The new partial discharge image set and the historical partial discharge image data set are combined to obtain an expanded first image data set.

8. A power grid partial discharge defect identification device, characterized in that: include: A signal acquisition module, used for acquiring partial discharge signals; An image conversion module, used for performing dimension conversion on the partial discharge signal to obtain a two-dimensional partial discharge image data set; A defect recognition module, used for inputting the partial discharge image data set into a preset partial discharge defect recognition model to perform defect type recognition and obtain a partial discharge defect recognition result; A model training module, used for training the partial discharge defect recognition model; Wherein, the model training module includes: A historical signal acquisition unit, used for acquiring historical partial discharge signals; A historical image conversion unit, used for performing dimension conversion on the historical partial discharge signal to obtain a historical partial discharge image data set; A historical image expansion unit, used for inputting the historical partial discharge image data set into a preset information generation adversarial network to expand the image data set, so as to obtain an expanded first image data set; A training data set generating unit, configured to generate a training data set according to the first image data set; The first training unit is used to input the training data set into the partial discharge defect recognition model to be trained, and perform iterative training until a preset model loss function converges to obtain a trained partial discharge defect recognition model.

9. A terminal device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, a method for identifying partial discharge defects in a power grid as claimed in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute a method for identifying partial discharge defects in a power grid as claimed in any one of claims 1 to 7.

Images