GIS electromagnetic field simulation data enhancement method and system based on PU-GAN

Abstract

The present invention discloses a GIS electromagnetic field simulation data enhancement method and system based on PU-GAN, comprising obtaining real coarse grid data and real fine grid data of the GIS electromagnetic field; inputting the real coarse grid data and the real fine grid data into a generative adversarial network for training until the overall optimization loss function is minimized, and obtaining a trained generator as a data enhancement model for enhancing the real coarse grid data generated in real time; wherein the generative adversarial network comprises a generator and a discriminator, the generator is used for extracting features from the real coarse grid data, and cyclically enhancing the extracted features to obtain fine grid generated data; the present invention can quickly obtain high-precision GIS electromagnetic field simulation enhanced data by using the trained model.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment, and in particular to a GIS electromagnetic field simulation data enhancement method and system based on PU-GAN. Background Art

[0002] Gas-insulated switchgear (GIS) has a complex structure and operates in a volatile environment. GIS simulation research typically utilizes finite element simulation methods. For example, the paper "Simulation and Optimization of Switchgear Products Based on Finite Element Analysis, Master's Thesis by Ouyang Zhenguo" proposes using finite element analysis software to simulate and calculate the three-dimensional electric field strength of a new vacuum circuit breaker. However, traditional numerical methods face challenges in performing electromagnetic field simulations using non-uniform meshes, such as complex coupling between coarse and fine meshes and difficulty establishing boundary conditions. Furthermore, non-uniform meshing simulations require high computational resources. Simulating with a high number of meshes and a coarse mesh density results in long computation times and slow simulation speeds, significantly impacting detection time for high-frequency GIS partial discharge (PD) detection. While coarse mesh simulations can reduce time costs, they also suffer from low simulation accuracy, which can compromise PD detection quality. Therefore, a fast and high-precision grid-based electromagnetic field distribution simulation technology is urgently needed.

[0003] In related technology, patent application publication number CN111428407A proposes a deep learning-based electromagnetic scattering calculation method. This approach proposes using a deep learning network as a model to complete the training and prediction process, thereby characterizing the input information—the relationship between the scatterer, the incident field, and the induced current. However, this approach presupposes that the incident wave's angle of incidence and intensity are fixed and known. However, when enhancing GIS electromagnetic field data, only coarse-grid electromagnetic field information of the GIS cylinder can be obtained, and the location and intensity of the GIS local discharge are unknown. Therefore, this approach is not suitable for enhancing GIS electromagnetic field data. Summary of the Invention

[0004] The technical problem to be solved by the present invention is how to quickly obtain high-precision GIS grid electromagnetic field simulation results.

[0005] The present invention solves the above technical problems through the following technical means:

[0006] The present invention proposes a GIS electromagnetic field simulation data enhancement method based on PU-GAN, which includes:

[0007] Obtain real coarse grid data and real fine grid data of GIS electromagnetic field;

[0008] Inputting the real coarse grid data and the real fine grid data into a generative adversarial network for training until the overall optimization loss function is minimized, thereby obtaining a trained generator as a data augmentation model for augmenting the real coarse grid data generated in real time;

[0009] The generative adversarial network includes the generator and the discriminator. The generator is used to extract features from the real coarse grid data and perform cyclic multiple enhancement on the extracted features to obtain fine grid generated data.

[0010] Furthermore, the discriminator includes a structural similarity module, a field strength consistency module and an accuracy discrimination module, wherein:

[0011] The structural similarity module is used to calculate a first confidence value of structural information contained in the fine grid generation data and the real fine grid data;

[0012] The field intensity consistency module is used to calculate a second confidence value of the field intensity information included in the fine grid generation data and the real fine grid data;

[0013] The accuracy determination module is used to shorten the distance between the fine grid generated data and the true fine grid data distribution.

[0014] Furthermore, the structural similarity module includes a first structural branch network, a second structural branch network, and a first regression network. The enhanced structural information contained in the fine grid generation data and the real fine structural information contained in the real fine grid data are respectively used as inputs of the first structural branch network and the second structural branch network. The outputs of the first structural branch network and the second structural branch network are subjected to matrix multiplication operation to output a structural feature matrix.

[0015] The structural feature matrix is ​​used as an input of the first regression network to obtain a first confidence value between the enhanced structural information and the true fine structural information.

[0016] Furthermore, the field strength consistency module includes a first enhancement branch network, a second enhancement branch network, and a second regression network. The enhanced field strength information contained in the fine grid generation data and the real fine field strength information contained in the real fine grid data are respectively used as inputs of the first enhancement branch network and the second enhancement branch network. The outputs of the first enhancement branch network and the second enhancement branch network are subjected to matrix multiplication operation to output a field strength feature matrix.

[0017] The field intensity characteristic matrix is ​​used as the input of the second regression network to obtain a second confidence value between the enhanced field intensity information and the true fine field intensity information.

[0018] Furthermore, the first structural branch network and the second structural branch network both include a multi-layer perceptron and a maximum pooling layer connected in sequence, and the outputs of the first structural branch network and the second structural branch network are subjected to matrix multiplication operations to output a structural feature matrix as follows:

[0019]

[0020] Where, is the structural feature matrix, Structural features of the data for the fine grid generation, is the structural feature of the real fine grid data, 、 respectively 、 The modulation weight of .

[0021] Furthermore, the first regression network includes a first self-attention unit, a multi-layer perceptron and a fully connected layer connected in sequence, the first self-attention unit is used to generate a structural attention weight based on the structural feature matrix, and the structural attention weight outputs a first confidence value after passing through the multi-layer perceptron and the fully connected layer.

[0022] Furthermore, the first enhanced branch network and the second enhanced branch network both include a multilayer perceptron and an activation function connected in sequence, and the outputs of the first enhanced branch network and the second enhanced branch network are subjected to a matrix multiplication operation to output a field intensity feature matrix as follows:

[0023]

[0024] Where, is the field strength characteristic matrix, The field strength characteristics of the data generated for the fine grid, is the field strength characteristic of the real fine grid data, '、 respectively 、 The modulation weight of .

[0025] Furthermore, the second regression network includes a second self-attention unit, a multi-layer perceptron and a fully connected layer connected in sequence, and the second self-attention unit is used to generate a field strength attention weight based on the field strength feature matrix, and the field strength attention weight outputs a second confidence value after passing through the multi-layer perceptron and the fully connected layer.

[0026] Furthermore, the formula of the overall optimization loss function is expressed as:

[0027]

[0028] Where, For the overall optimization loss function, a first confidence value for the structural information contained in the fine grid generation data and the true fine grid data, a second confidence value for the field strength information contained in the fine grid generation data and the true fine grid data, is the adversarial loss between the fine grid generated data and the true fine grid data, and is the weight parameter.

[0029] Furthermore, the generator includes a plurality of cascaded feature extraction and expansion networks, each of which is connected to a multi-layer perceptron;

[0030] The feature extraction and expansion network includes a feature extraction module and a feature expansion module connected in sequence, and the output of the feature extraction module in the feature extraction and expansion network of the previous level is connected to the output of the feature extraction module in the feature extraction and expansion network of the next level;

[0031] The feature extraction module includes a first MLPs layer, a KNN layer, a second MLPs layer, a maximum pooling layer, and a data compression layer connected in sequence. The real coarse grid data serves as the input of the first MLPs layer. The output of the first MLPs layer and the output of the second MLPs layer are spliced ​​and output to the maximum pooling layer. The real coarse grid data and the output of the maximum pooling layer are spliced ​​and used as the input of the data compression layer. The output of the data compression layer is connected to the feature expansion module.

[0032] The feature expansion module is used to add a vector coded as 1 or -1 to each feature output by the feature extraction module to generate position disturbance.

[0033] Furthermore, when the features output by the feature extraction module in the feature extraction and expansion network at the previous level are copied to the output of the feature extraction module in the feature extraction and expansion network at the next level, interpolation technology is used to match features at different levels.

[0034] Furthermore, the method further comprises: using bilateral interpolation to match features at different levels, The interpolation of the feature value output by the previous feature extraction module is:

[0035]

[0036] in, It is The coordinates and field strength of each point, It is The coordinates and field strength of each point, It is The eigenvalues ​​corresponding to the points are It is The eigenvalues ​​corresponding to the points are is the set of data points that need to be inserted, where the joint weighting function is expressed as follows:

[0037]

[0038]

[0039] Among them, the width parameter and is the average distance from a point to its nearest neighbor, Represents the magnitude of a vector.

[0040] In addition, the present invention also proposes a GIS electromagnetic field simulation data enhancement system based on PU-GAN, which includes:

[0041] A data acquisition unit is used to obtain real coarse grid data and real fine grid data of the GIS electromagnetic field;

[0042] A training unit is configured to input the real coarse grid data and the real fine grid data into a generative adversarial network for training until the overall optimization loss function is minimized, thereby obtaining a trained generator as a data augmentation model for augmenting the real coarse grid data generated in real time;

[0043] The generative adversarial network includes the generator and the discriminator. The generator is used to extract features from the real coarse grid data and perform cyclic multiple enhancement on the extracted features to obtain fine grid generated data.

[0044] The advantages of the present invention are:

[0045] (1) The present invention trains a generative adversarial network by using real coarse grid data and real fine grid data obtained by simulation. The trained generator can be used to enhance the real coarse grid data obtained by real-time simulation to generate GIS electromagnetic field fine data. Moreover, since GIS equipment has a certain physical structure, the data points of electromagnetic field simulation cannot exceed a certain range. The generator designed by the present invention adopts cyclic multiple enhancement to enhance the electromagnetic field between coarse grid data points, which ensures that it does not exceed the physical structure of GIS to a certain extent, thereby improving the accuracy of GIS electromagnetic field simulation data enhancement.

[0046] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 1 is a flow chart of a GIS electromagnetic field simulation data enhancement method based on PU-GAN proposed in one embodiment of the present invention;

[0048] Figure 2 This is a principle block diagram of GIS electromagnetic field simulation data enhancement based on PU-GAN proposed in one embodiment of the present invention;

[0049] Figure 3 is a schematic structural diagram of a generator in one embodiment of the present invention;

[0050] Figure 4 is a structural diagram of a feature extraction module in one embodiment of the present invention;

[0051] Figure 5 is a schematic diagram of a structurally similar module in one embodiment of the present invention;

[0052] Figure 6 is a structural diagram of a field strength consistency module in one embodiment of the present invention;

[0053] Figure 7 is a schematic diagram of the structure of a self-attention unit in one embodiment of the present invention;

[0054] Figure 8 This is a structural diagram of a GIS electromagnetic field simulation data enhancement system based on PU-GAN proposed in one embodiment of the present invention. DETAILED DESCRIPTION

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0056] like Figures 1 to 2 As shown, the first embodiment of the present invention proposes a GIS electromagnetic field simulation data enhancement method based on PU-GAN, which includes the following steps:

[0057] S10, obtaining real coarse grid data and real fine grid data of GIS electromagnetic field;

[0058] It should be noted that, in this embodiment, the electromagnetic field coarse grid data and fine grid data can be obtained by establishing a GIS spatial topology model and an electromagnetic field mathematical model, and solving the electromagnetic field model based on the finite difference method or the finite element method.

[0059] It should be noted that the acquired grid data includes the three-dimensional space coordinate data and field strength value of each grid vertex.

[0060] S20, inputting the real coarse grid data and the real fine grid data into a generative adversarial network for training until the overall optimization loss function is minimized, thereby obtaining a trained generator as a data enhancement model for enhancing the real coarse grid data generated in real time;

[0061] The generative adversarial network includes the generator and the discriminator. The generator is used to extract features from the real coarse grid data and perform cyclic multiple enhancement on the extracted features to obtain fine grid generated data.

[0062] It's important to note that the generator network is responsible for learning the underlying distribution characteristics of real samples and synthesizing new samples; the discriminator network is responsible for distinguishing between real samples and generated new samples. The generator and discriminator achieve a tactical effect through alternating training, continuously improving their respective generation and discrimination capabilities, ultimately leading to a Nash equilibrium between the generator and discriminator networks.

[0063] Because GIS equipment has a certain physical structure, the data points of electromagnetic field simulation cannot exceed a certain range. The PU-GAN (Projection Upsampling Generative Adversarial Network) designed in this embodiment of the present invention multiplies the coarse grid data points through upsampling. The generator uses cyclic multiplication enhancement to enhance the electromagnetic field between coarse grid data points, ensuring that it does not exceed the physical structure of the GIS to a certain extent, thereby improving the accuracy of GIS electromagnetic field simulation data enhancement.

[0064] As a further preferred technical solution, Figure 3 As shown, the generator includes several cascaded feature extraction and expansion networks, each of which is connected to a multi-layer perceptron;

[0065] The feature extraction and expansion network includes a feature extraction module Featuer Extraction and a feature expansion module connected in sequence, and the output of the feature extraction module Featuer Extraction in the feature extraction and expansion network of the previous level is connected to the output of the feature extraction module Featuer Extraction in the feature extraction and expansion network of the next level;

[0066] like Figure 4As shown, the feature extraction module Featuer Extraction includes a first MLPs layer, a KNN layer, a second MLPs layer, a maximum pooling layer maxpooling and a data compression layer compression connected in sequence. The real coarse grid data serves as the input of the first MLPs layer. The output of the first MLPs layer and the output of the second MLPs layer are spliced ​​and output to the maximum pooling layer maxpooling. The real coarse grid data and the output of the maximum pooling layer maxpooling are spliced ​​and used as the input of the data compression layer compression. The output of the data compression layer compression is connected to the feature expansion module.

[0067] The feature expansion module is used to add a vector coded as 1 or -1 to each feature output by the feature extraction module to generate position disturbance.

[0068] It should be noted that the function of the generator designed in this embodiment is to enhance the coarse grid data of GIS and generate the required GIS fine grid data; the designed generator adopts a cyclic multiple enhancement design, and can autonomously adjust the coarse grid enhancement multiple according to demand, and the multiple enhancement can better retain the characteristics of the coarse grid data.

[0069] Specifically, the feature extraction module in this embodiment operates as follows: First, the N×4-dimensional information consisting of the 3D spatial coordinates (x, y, z) and field strength value v of each input sampling point is extracted as fixed-scale features C' through the first MLPs layer. The features are then grouped into N×K groups using the KNN layer. The group features are further optimized using the second MLPs layer to obtain G-channel features. The features output by the second MLPs layer are combined with the features output by the first MLPs layer to obtain a new G-channel feature (for a total of N×K×(2G+C')). Finally, the grouped information is condensed through maxpooling to obtain the final N×(2G+C') feature, which is then added to the N×4-dimensional feature of the initial sampling point. This feature, aggregated as a multi-scale feature of the module, is fed into subsequent modules for reuse, improving reconstruction accuracy while also increasing parameter efficiency and reducing model size. Finally, the N×C features are obtained through compression in the final data compression layer. After copying the obtained N×C features, a vector encoded as 1 or -1 is added after each feature to generate position perturbations, and then the subsequent multi-layer perceptron MLP is used to regress the 2N×(C+1) features into 2N×4 data.

[0070] It should be noted that after multiple cycles of feature extraction and expansion, a 2 r The N×C features are finally regressed to four-dimensional data by a multi-layer perceptron MLP2r N×4.

[0071] It should be noted that the value of r depends on the accuracy difference between the input simulation coarse grid and fine grid, and is not specifically limited in this embodiment.

[0072] Specifically, if the previous feature needs to be passed to the next layer, a good interpolation technology is required to match features at different levels. In this embodiment, when copying the features output by the feature extraction module in the feature extraction and expansion network of the previous level to the output of the feature extraction module in the feature extraction and expansion network of the next level, the interpolation technology is used to match features at different levels.

[0073] As a further preferred technical solution, this embodiment uses the bilateral interpolation technology to interpolate the points using the features around the target point. The interpolation of the feature value output by the previous feature extraction module is:

[0074]

[0075] in, It is The coordinates and field strength of each point, It is The coordinates and field strength of each point, It is The eigenvalues ​​corresponding to the points are It is The eigenvalues ​​corresponding to the points are is the set of data points that need to be inserted, where the joint weighting function is expressed as follows:

[0076]

[0077]

[0078] Among them, the width parameter and is the average distance from a point to its nearest neighbor, Represents the magnitude of a vector.

[0079] This embodiment is for coarse grid data enhancement. Using bilateral interpolation can effectively avoid the generated grid data from overlapping with the source grid data, thus avoiding waste.

[0080] As a further preferred technical solution, Figure 2 As shown, the discriminator includes a structural similarity module, a field strength consistency module and an accuracy discrimination module, wherein:

[0081] The structural similarity module is used to calculate a first confidence value of structural information contained in the fine grid generation data and the real fine grid data;

[0082] The field intensity consistency module is used to calculate a second confidence value of the field intensity information included in the fine grid generation data and the real fine grid data;

[0083] The accuracy determination module is used to shorten the distance between the fine grid generated data and the true fine grid data distribution.

[0084] It should be noted that the discriminator designed in this embodiment is used to determine whether the input data meets the requirements. The discriminator consists of a structural similarity module, a field strength consistency module, and an accuracy discrimination module. The grid data is divided into structural data and field strength data for comparative analysis, which can effectively improve the discrimination accuracy.

[0085] As a further preferred technical solution, Figure 5 As shown, the structural similarity module includes a first structural branch network, a second structural branch network and a first regression network. The enhanced structural information contained in the fine grid generation data and the real fine structural information contained in the real fine grid data are respectively used as inputs of the first structural branch network and the second structural branch network. The outputs of the first structural branch network and the second structural branch network are subjected to matrix multiplication operation to output a structural feature matrix.

[0086] The structural feature matrix is ​​used as an input of the first regression network to obtain a first confidence value between the enhanced structural information and the true fine structural information.

[0087] Specifically, if Figure 5 As shown, the first structure branch network and the second structure branch network both include a multi-layer perceptron and a maximum pooling layer connected in sequence. The output of the first structure branch network and the second structure branch network is subjected to a matrix multiplication operation to output a structural feature matrix:

[0088]

[0089]

[0090]

[0091] Where, is the structural feature matrix, Structural features of the data for the fine grid generation, is the structural feature of the real fine grid data, 、 respectively 、 The modulation weight, represents the multilayer perceptron operation, represents the maximum pooling operation, Generate data for the fine grid The structural information of each point, is the real fine grid data The structural information of each point.

[0092] Specifically, if Figure 5 As shown, the first regression network includes a first self-attention unit, a multi-layer perceptron and a fully connected layer connected in sequence, the first self-attention unit is used to generate a structural attention weight based on the structural feature matrix, and the structural attention weight outputs a first confidence value after passing through the multi-layer perceptron and the fully connected layer.

[0093] As a further preferred technical solution, Figure 6 As shown, the field strength consistency module includes a first enhancement branch network, a second enhancement branch network, and a second regression network. The enhanced field strength information contained in the fine grid generation data and the real fine field strength information contained in the real fine grid data are respectively used as inputs of the first enhancement branch network and the second enhancement branch network. The outputs of the first enhancement branch network and the second enhancement branch network are subjected to matrix multiplication operation to output a field strength feature matrix.

[0094] The field intensity characteristic matrix is ​​used as the input of the second regression network to obtain a second confidence value between the enhanced field intensity information and the true fine field intensity information.

[0095] Specifically, the first enhanced branch network and the second enhanced branch network both include a multilayer perceptron and an activation function connected in sequence. After the outputs of the first enhanced branch network and the second enhanced branch network are subjected to matrix multiplication operations, the output field intensity feature matrix is:

[0096]

[0097]

[0098]

[0099] Where, is the field strength characteristic matrix, The field strength characteristics of the data generated for the fine grid, is the field strength characteristic of the real fine grid data, '、 respectively 、 The modulation weight, Generate data for the fine grid Field strength information of each point, is the real fine grid data Field strength information of each point.

[0100] Specifically, the second regression network includes a second self-attention unit, a multi-layer perceptron and a fully connected layer connected in sequence. The second self-attention unit is used to generate a field intensity attention weight based on the field intensity feature matrix. The field intensity attention weight outputs a second confidence value after passing through the multi-layer perceptron and the fully connected layer.

[0101] As a further preferred technical solution, Figure 7 As shown, the self-attention unit used in this embodiment is used to perform enhanced feature integration. Specifically, the input features are converted into G, H and F through three independent MLPs, and then the attention weights are generated by G and H:

[0102] M

[0103] in, Represents the softmax function.

[0104] Then the weighted feature M is obtained, and finally the output feature is generated, which is the sum of the input feature and the weighted feature.

[0105] As a further preferred technical solution, in this embodiment, the outputs of the structural similarity module and the field strength consistency module are both connected to the accuracy discrimination module. The accuracy discrimination module is used to control the generated grid structure and field strength data and distinguish the accuracy of the generated results. The accuracy discrimination module can be defined as a regional-level fully convolutional binary classifier, which can narrow the distance between the distribution of generated fine grid simulation data and the fine grid simulation data of the numerical method, and at the same time use the least squares adversarial loss to ensure the stability of the training process.

[0106] The adversarial loss between the fine mesh generated data and the true fine mesh data is defined as follows:

[0107]

[0108] Where, and They are the fine grid generated data output by the discriminator D from the generator Compared with the real fine grid data The confidence value of the prediction.

[0109] As a further preferred technical solution, the formula of the overall optimization loss function is expressed as:

[0110]

[0111] Where, For the overall optimization loss function, a first confidence value for the structural information contained in the fine grid generation data and the true fine grid data, a second confidence value for the field strength information contained in the fine grid generation data and the true fine grid data, is the adversarial loss between the fine grid generated data and the true fine grid data, and is the weight parameter.

[0112] The overall optimization loss function set in this embodiment comprehensively considers the structure and field strength confidence of the generated fine grid and the real fine grid, and has the following advantages: (1) Promoting the generator to generate realistic data: By optimizing the loss function, the generator is forced to generate samples that are as close to the real data distribution as possible. (2) Improving the accuracy of the discriminator: The discriminator improves its ability to distinguish between real data and generated data by optimizing the loss function. This includes enabling the discriminator to more accurately identify the data generated by the generator and better distinguish between generated data and real data. (3) Achieving a balance between the generator and the discriminator: Optimizing the loss function ensures that there is a balance between the generator and the discriminator, avoiding over-optimization of one of them and causing unstable training. This balance helps the generator generate more realistic data while ensuring that the discriminator can effectively distinguish between real data and generated data.

[0113] In addition, if Figure 8 As shown, the second embodiment of the present invention proposes a GIS electromagnetic field simulation data enhancement system based on PU-GAN, the system comprising:

[0114] A data acquisition unit 10 is used to acquire real coarse grid data and real fine grid data of the GIS electromagnetic field;

[0115] A training unit 20 is configured to input the real coarse grid data and the real fine grid data into a generative adversarial network for training until the overall optimization loss function is minimized, thereby obtaining a trained generator as a data augmentation model for augmenting the real coarse grid data generated in real time;

[0116] The generative adversarial network includes the generator and the discriminator. The generator is used to extract features from the real coarse grid data and perform cyclic multiple enhancement on the extracted features to obtain fine grid generated data.

[0117] It should be noted that other embodiments or implementation methods of the generative adversarial network in the GIS electromagnetic field simulation data enhancement system based on PU-GAN described in the present invention can refer to the above-mentioned method embodiments and will not be repeated here.

[0118] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0119] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0120] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A GIS electromagnetic field simulation data enhancement method based on PU-GAN, characterized in that: The method comprises: Obtain the real coarse grid data and real fine grid data of GIS electromagnetic field; Inputting the real coarse grid data and the real fine grid data into a generative adversarial network for training until the overall optimization loss function is minimized, and obtaining a trained generator as a data enhancement model for enhancing the real coarse grid data generated in real time; The generative adversarial network includes the generator and the discriminator, the generator is used to extract features from the real coarse grid data, and perform cyclic multiple enhancement on the extracted features to obtain fine grid generated data; The discriminator includes a structural similarity module, a field strength consistency module and an accuracy discrimination module, wherein: the structural similarity module is used to calculate a first confidence value of structural information contained in the fine grid generation data and the real fine grid data; the field strength consistency module is used to calculate a second confidence value of field strength information contained in the fine grid generation data and the real fine grid data; the accuracy discrimination module is used to narrow the distance between the distribution of the fine grid generation data and the real fine grid data; The structural similarity module includes a first structural branch network, a second structural branch network and a first regression network, the enhanced structural information contained in the fine grid generation data and the real fine structural information contained in the real fine grid data are respectively used as inputs of the first structural branch network and the second structural branch network, and the outputs of the first structural branch network and the second structural branch network are output as a structural feature matrix after matrix multiplication operation; the structural feature matrix is ​​used as an input of the first regression network to obtain a first confidence value between the enhanced structural information and the real fine structural information; The field intensity consistency module includes a first enhanced branch network, a second enhanced branch network and a second regression network. The enhanced field intensity information contained in the fine grid generation data and the real fine field intensity information contained in the real fine grid data are respectively used as inputs of the first enhanced branch network and the second enhanced branch network. The outputs of the first enhanced branch network and the second enhanced branch network are subjected to matrix multiplication operation to output a field intensity feature matrix. The field intensity feature matrix is ​​used as an input of the second regression network to obtain a second confidence value between the enhanced field intensity information and the real fine field intensity information. The generator includes a plurality of cascaded feature extraction and expansion networks, each of which is connected to a multi-layer perceptron; The feature extraction and expansion network includes a feature extraction module and a feature expansion module connected in sequence, and the output of the feature extraction module in the feature extraction and expansion network of the previous level is connected to the output of the feature extraction module in the feature extraction and expansion network of the next level; The feature extraction module includes a first MLPs layer, a KNN layer, a second MLPs layer, a maximum pooling layer maxpooling and a data compression layer connected in sequence, the real coarse grid data is used as the input of the first MLPs layer, the output of the first MLPs layer and the output of the second MLPs layer are spliced ​​and output to the maximum pooling layer maxpooling, the real coarse grid data and the output of the maximum pooling layer maxpooling are spliced ​​and used as the input of the data compression layer, and the output of the data compression layer is connected to the feature expansion module; the feature expansion module is used to add a vector encoded as 1 or -1 to each feature output by the feature extraction module to generate position disturbance; When copying the features output by the feature extraction module in the feature extraction and expansion network of the previous level to the output of the feature extraction module in the feature extraction and expansion network of the next level, an interpolation technique is used to match features at different levels.

2. The GIS electromagnetic field simulation data enhancement method based on PU-GAN as claimed in claim 1, characterized in that: The first structure branch network and the second structure branch network both include a multi-layer perceptron and a maximum pooling layer connected in sequence. The outputs of the first structure branch network and the second structure branch network are subjected to matrix multiplication operations to output a structural feature matrix: In the formula, is the structural feature matrix, The structural characteristics of the data for the fine grid generation, is the structural feature of the real fine grid data, , respectively , The modulation weight.

3. The GIS electromagnetic field simulation data enhancement method based on PU-GAN as claimed in claim 1, characterized in that: The first regression network includes a first self-attention unit, a multi-layer perceptron and a fully connected layer connected in sequence, the first self-attention unit is used to generate a structural attention weight based on the structural feature matrix, and the structural attention weight outputs a first confidence value after passing through the multi-layer perceptron and the fully connected layer.

4. The GIS electromagnetic field simulation data enhancement method based on PU-GAN as claimed in claim 1, characterized in that: The first enhanced branch network and the second enhanced branch network both include multi-layer perceptrons and activation functions connected in sequence. The outputs of the first enhanced branch network and the second enhanced branch network are subjected to matrix multiplication operations to output a field intensity feature matrix as follows: In the formula, is the field strength characteristic matrix, The field strength characteristics of the data generated for the fine grid, is the field intensity characteristic of the real fine grid data, '、 respectively , The modulation weight.

5. The GIS electromagnetic field simulation data enhancement method based on PU-GAN as claimed in claim 1, characterized in that: The second regression network includes a second self-attention unit, a multi-layer perceptron and a fully connected layer connected in sequence. The second self-attention unit is used to generate field intensity attention weights based on the field intensity feature matrix. The field intensity attention weights output a second confidence value after passing through the multi-layer perceptron and the fully connected layer.

6. The GIS electromagnetic field simulation data enhancement method based on PU-GAN according to any one of claims 1 to 5, characterized in that: The formula of the overall optimization loss function is expressed as: In the formula, For the overall optimization loss function, A first confidence value of the structural information contained in the fine grid generation data and the true fine grid data, a second confidence value for the field strength information contained in the fine grid generation data and the real fine grid data, is the adversarial loss between the fine grid generated data and the true fine grid data, and is the weight parameter.

7. The GIS electromagnetic field simulation data enhancement method based on PU-GAN as claimed in claim 1, characterized in that: The method further includes: using bilateral interpolation to match features at different levels, The interpolation of the feature values ​​output by the previous feature extraction module is: in, It is The coordinates and field strength of the points, It is The coordinates and field strength of the points, It is The eigenvalues ​​corresponding to the points are It is The eigenvalues ​​corresponding to the points are , where the joint weighting function is expressed as follows: Among them, the width parameter and is the average distance from a point to its nearest neighbor, Represents the magnitude of a vector.

8. A GIS electromagnetic field simulation data enhancement system based on PU-GAN, characterized in that: The system comprises: A data acquisition unit, used to acquire real coarse grid data and real fine grid data of GIS electromagnetic field; A training unit, used for inputting the real coarse grid data and the real fine grid data into a generative adversarial network for training until the overall optimization loss function is minimized, and obtaining a trained generator as a data enhancement model for enhancing the real coarse grid data generated in real time; The generative adversarial network includes the generator and the discriminator, the generator is used to extract features from the real coarse grid data, and perform cyclic multiple enhancement on the extracted features to obtain fine grid generated data; The discriminator includes a structural similarity module, a field strength consistency module and an accuracy discrimination module, wherein: the structural similarity module is used to calculate a first confidence value of structural information contained in the fine grid generation data and the real fine grid data; the field strength consistency module is used to calculate a second confidence value of field strength information contained in the fine grid generation data and the real fine grid data; the accuracy discrimination module is used to narrow the distance between the distribution of the fine grid generation data and the real fine grid data; The structural similarity module includes a first structural branch network, a second structural branch network and a first regression network, the enhanced structural information contained in the fine grid generation data and the real fine structural information contained in the real fine grid data are respectively used as inputs of the first structural branch network and the second structural branch network, and the outputs of the first structural branch network and the second structural branch network are output as a structural feature matrix after matrix multiplication operation; the structural feature matrix is ​​used as an input of the first regression network to obtain a first confidence value between the enhanced structural information and the real fine structural information; The field intensity consistency module includes a first enhanced branch network, a second enhanced branch network and a second regression network. The enhanced field intensity information contained in the fine grid generation data and the real fine field intensity information contained in the real fine grid data are respectively used as inputs of the first enhanced branch network and the second enhanced branch network. The outputs of the first enhanced branch network and the second enhanced branch network are subjected to matrix multiplication operation to output a field intensity feature matrix. The field intensity feature matrix is ​​used as an input of the second regression network to obtain a second confidence value between the enhanced field intensity information and the real fine field intensity information. The generator includes a plurality of cascaded feature extraction and expansion networks, each of which is connected to a multi-layer perceptron; The feature extraction and expansion network includes a feature extraction module and a feature expansion module connected in sequence, and the output of the feature extraction module in the feature extraction and expansion network of the previous level is connected to the output of the feature extraction module in the feature extraction and expansion network of the next level; The feature extraction module includes a first MLPs layer, a KNN layer, a second MLPs layer, a maximum pooling layer maxpooling and a data compression layer connected in sequence, the real coarse grid data is used as the input of the first MLPs layer, the output of the first MLPs layer and the output of the second MLPs layer are spliced ​​and output to the maximum pooling layer maxpooling, the real coarse grid data and the output of the maximum pooling layer maxpooling are spliced ​​and used as the input of the data compression layer, and the output of the data compression layer is connected to the feature expansion module; the feature expansion module is used to add a vector encoded as 1 or -1 to each feature output by the feature extraction module to generate position disturbance; When copying the features output by the feature extraction module in the feature extraction and expansion network of the previous level to the output of the feature extraction module in the feature extraction and expansion network of the next level, an interpolation technique is used to match features at different levels.

9. The GIS electromagnetic field simulation data enhancement system based on PU-GAN as claimed in claim 8, characterized in that: The first structure branch network and the second structure branch network both include a multi-layer perceptron and a maximum pooling layer connected in sequence. The outputs of the first structure branch network and the second structure branch network are subjected to matrix multiplication operations to output a structural feature matrix: In the formula, is the structural feature matrix, The structural characteristics of the data for the fine grid generation, is the structural feature of the real fine grid data, , respectively , The modulation weight.

10. The GIS electromagnetic field simulation data enhancement system based on PU-GAN as claimed in claim 8, characterized in that: The first enhanced branch network and the second enhanced branch network both include multi-layer perceptrons and activation functions connected in sequence. The outputs of the first enhanced branch network and the second enhanced branch network are subjected to matrix multiplication operations to output a field intensity feature matrix as follows: In the formula, is the field strength characteristic matrix, The field strength characteristics of the data generated for the fine grid, is the field intensity characteristic of the real fine grid data, '、 respectively , The modulation weight.

Images