Electromagnetic field simulation data enhancement generative adversarial network construction method and system

Abstract

This invention discloses a method and system for constructing a generative adversarial network (GAN) to enhance electromagnetic field simulation data. The method includes: establishing an attention mechanism to identify significant electric field distortion regions in coarse-grid electromagnetic field data and assigning processing weights; constructing a generative network to map the weighted coarse-grid electromagnetic field data to fine-grid electromagnetic field data; generating confidence scores and error distribution maps for the fine-grid electromagnetic field data to evaluate data quality; constructing a discriminative network to perform adversarial training on the coarse-grid and fine-grid electromagnetic field data; and introducing a priority-guided feedback mechanism to optimize the generation quality of significant electric field distortion regions and integrating it into the overall loss function for joint training. This invention prioritizes the generation quality of key regions, significantly improving the accuracy and reliability of electromagnetic field simulation data enhancement, and providing high-quality data support for the insulation status analysis of power equipment.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic data augmentation technology, specifically to a method and system for constructing a generative adversarial network for electromagnetic field simulation data augmentation. Background Technology

[0002] Currently, the main method for obtaining high-precision electric field distribution data is numerical simulation based on the finite element method (FEM) or boundary element method (BEM). Although these methods can provide accurate results, their calculations are usually extremely time-consuming, especially when a high-resolution, fine-mesh model is needed to capture local electric field details. To balance computational efficiency and resource consumption, engineering practice often adopts a strategy of performing coarse-mesh simulations first, followed by local refinement. However, this method heavily relies on human experience and is difficult to rapidly generate large-scale, comprehensive, high-density electric field data. This bottleneck greatly limits the development of data-driven intelligent diagnostic technology for gas-insulated switchgear (GIS), making the construction of a high-quality electric field sample library covering various typical defect conditions a significant challenge.

[0003] In recent years, deep learning techniques such as Generative Adversarial Networks (GANs) have been able to learn mapping relationships from low resolution to high resolution, providing a new solution for rapid super-resolution reconstruction of electric field data. However, directly applying standard GAN models to electromagnetic field simulation data augmentation still faces two key technical challenges: First, electromagnetic field data has high physical specificity, with huge differences in field strength distribution. Standard GANs often adopt an "average" learning strategy, resulting in insufficient feature preservation in key areas with significant electric field distortion, leading to blurring or distortion in the generated results at critical locations. Second, existing methods lack a quantitative quality assessment and feedback mechanism embedded within the generation process, making it impossible to dynamically evaluate and specifically optimize the reliability of the generated data, resulting in unstable generation quality.

[0004] Patent application CN115082318A discloses a method for super-resolution reconstruction of infrared images of electrical equipment. The main technical approach involves constructing low-resolution / high-resolution image pairs, utilizing generative networks and adversarial networks to reconstruct the infrared images, and evaluating the effect based on the quality of the reconstructed images. Its key aspects lie in the generator, discriminator, and image quality improvement during super-resolution reconstruction. Patent application CN121053514A discloses a method and system for multimodal data enhancement for vehicle recognition based on generative adversarial networks. This method collects multimodal data such as visible light, thermal infrared, and point cloud data, constructs a cross-modal generator, learns intermodal feature associations through a dynamic weight allocation module, and then uses dual discriminators for alternating adversarial training to generate enhanced data samples, ultimately serving vehicle recognition. It also involves confidence classification and error screening at the recognition result level. Its key aspects lie in multimodal enhancement, dual discriminator training, and reliability processing of the recognition results. These two patent documents are closer to the "generation + evaluation" technical path. The evaluation in CN115082318A mainly focuses on the external quality evaluation of the reconstruction effect, while the confidence processing in CN121053514A mainly involves the verification and screening of the recognition results. They are more inclined towards the judgment at the result level, rather than setting up a separate quality assessment / measurement module within the generation process to specifically output quality information for the generated results and continue to influence subsequent optimization. Judging from the publication context, their focus is still mainly on the evaluation of reconstruction effect or the confidence processing of recognition results.

[0005] Therefore, there is an urgent need in this field for an electromagnetic field data augmentation method that can automatically focus on key areas and has a built-in quality assessment and feedback loop, in order to overcome the technical limitations of existing numerical simulation and general generative models. Summary of the Invention

[0006] The technical problem to be solved by this invention is to address the issues of insufficient preservation of key regional features and unstable generation quality in generative adversarial networks during electromagnetic field data enhancement.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] A method for constructing a generative adversarial network to enhance electromagnetic field simulation data includes: An attention mechanism is established to identify regions of significant electric field distortion in coarse-grid electromagnetic field data and assign processing weights accordingly. A generative network is constructed to map the weighted coarse-grid electromagnetic field data into fine-grid electromagnetic field data; Generate confidence scores and error distribution plots for fine-grid electromagnetic field data to assess data quality; Construct a discriminative network and conduct adversarial training on coarse-grid and fine-grid electromagnetic field data; A priority-guided feedback mechanism is introduced to optimize the generation quality of regions with significant electric field distortion, and this optimization is integrated into the overall loss function for joint training.

[0009] Furthermore, regions of significant electric field distortion in the coarse-grid electromagnetic field data are identified and assigned processing weights, including: Based on coarse grid data of electromagnetic field and the electric field derived therefrom To obtain a set of complementary physical feature maps; The physical feature maps are concatenated along the channel dimension to obtain the fused features. ; Fusion features Input a lightweight convolutional network The network learns the spatial importance distribution through convolutional operations and finally normalizes it using the sigmoid function to obtain a preliminary attention weight map. ; Preliminary attention weight map Apply a joint filter based on multiple physical features to obtain a priority weight map. ; Based on priority weight graph coarse grid data of electromagnetic fields The weighted feature map is obtained by performing weighted processing. .

[0010] Furthermore, obtain the priority weight graph. include: An empirical threshold is set for each physical feature map, and a joint binary mask is generated. The mask is set to 1 in any region where any physical feature exceeds its threshold, and 0 otherwise.

[0011] Furthermore, the weighted coarse-grid electromagnetic field data is mapped to fine-grid electromagnetic field data, including: The generative network employs a physically guided encoder-decoder structure; it weights the feature maps... As input; wherein, in the encoder, not only are the spatial features of the data extracted, but physical information is also explicitly encoded, and the physical information is used as an additional channel with the input data. After being concatenated, they are input into the encoder together; in the decoder, a differentiable numerical differential operator is used to replace part of the standard convolutional layer, so that the decoder implicitly learns and satisfies the smoothness and continuity constraints of the field distribution in the feature space. Define the physical consistency loss function To ensure the fine-grid electromagnetic field data output by the decoder It conforms to the basic laws of electromagnetic fields and introduces physical regularization loss based on Maxwell's equations during the training process.

[0012] Furthermore, confidence scores and error distribution maps of the fine-grid electromagnetic field data are generated to assess data quality, including: Calculate fine-grid data for electromagnetic fields Physical rationality indicators; Based on electromagnetic field fine mesh data Coarse grid data of electromagnetic fields Physical rationality indicators extract quality characteristics from multiple dimensions and integrate them into comprehensive evaluation indicators; Based on comprehensive evaluation indicators, a confidence score for physical guidance is output. Error distribution plot ; And based on confidence scores Error distribution plot Define the measurement loss function.

[0013] Furthermore, output the confidence score of the physical guidance. Error distribution plot It can be expressed by the following formula: ; ; ; In the formula, For the Sigmoid function, It is a lightweight neural network. For comprehensive evaluation indicators, As a weight for physical rationality, For element-wise multiplication, It is an exponential function. This is an adjustment coefficient for the weight of physical rationality. For the physical residuals of the field distribution, This is a graph showing the error distribution predicted by the convolutional network. For fine-grid electromagnetic field data, For coarse-grid data of electromagnetic fields The result of upsampling to a fine grid resolution.

[0014] Furthermore, adversarial training is conducted using coarse-grid and fine-grid electromagnetic field data, including: The discriminator is defined as a fully convolutional binary classifier used to distinguish between generated electromagnetic field fine mesh data and real electromagnetic field fine mesh data. The least squares adversarial loss function is used to define the adversarial loss of the discriminator and to optimize the training.

[0015] Furthermore, a priority-guided feedback mechanism is introduced to optimize the generation quality of regions with significant electric field distortion, including: According to the priority weight diagram Confidence score Obtain the feedback weight graph ; Based on feedback weight graph Define a priority-guided feedback loss function. This measures the quality of generated fine-grid electromagnetic field data in regions with significant electric field distortion and low confidence levels. Compared with fine-grid data of real electromagnetic fields Weighted differences between them; Among them, the feedback loss function Expressed using the following formula: ; In the formula, This represents the total number of electromagnetic field mesh locations participating in the feedback loss statistics. For the first Feedback weight map at each location For the first Fine-grid data of the electromagnetic field at each location. For the first Fine-grid data of the actual electromagnetic field at each location. It is the square of the L2 norm.

[0016] Furthermore, a priority-guided feedback mechanism is introduced to optimize the generation quality of regions with significant electric field distortion, and this quality is integrated into the overall loss function for joint training, including: The overall loss function of a generative adversarial network (GAN), which consists of a generative network and a discriminative network, is mainly composed of the weighted sum of measurement loss, physical consistency loss, feedback loss, and adversarial loss during the joint training phase. Priority loss is defined as the reconstruction error of high-priority regions, which is used to guide regions with significant electric field distortion during the initial training phase. A two-stage training strategy is adopted: In the first stage, the priority loss is defined as the reconstruction error of the high-priority region, which is used to guide the training in the initial stage for regions with significant electric field distortion, and the priority loss and physical consistency loss are used to obtain the basic generation results; in the second stage, measurement loss, feedback loss and adversarial loss are introduced for end-to-end joint training to minimize the overall loss function. For the goal.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention achieves enhanced generation from coarse-grid data to fine-grid data at a lower simulation cost through the collaborative design of a priority module, enhancement module, measurement module, and discrimination module, thereby improving the resolution, fidelity of key areas, and quality assessability of electromagnetic field simulation data. Secondly, regarding the priority module, a weight allocation mechanism guided by multiple physical features prioritizes regions with high field strength, abrupt field strength changes, abnormal charge accumulation, and directional disorder during network training, thus improving the reconstruction accuracy of key defect areas. Thirdly, regarding physical consistency constraints, by introducing residual loss and boundary condition loss into the enhancement module training process, the generated results are made more consistent with the basic laws of electromagnetic fields, improving the physical rationality of the results. Fourthly, regarding the measurement module and feedback mechanism, by outputting confidence scores and error distribution maps, and jointly constructing feedback weights from key area information and low-confidence information, the optimization efficiency of difficult areas and the overall generation stability can be further improved.

[0018] The invention prioritizes the generation quality of key areas, significantly improves the accuracy and reliability of electromagnetic field simulation data enhancement, and provides high-quality data support for the insulation status analysis of power equipment. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a method for constructing a generative adversarial network to enhance electromagnetic field simulation data, according to an embodiment of the present invention.

[0020] Figures 2 to 5 The embodiments of the present invention provide simulation results of coarse-grid electromagnetic field data under operating conditions of 220kV, 500kV, 800kV, and 1000kV. Figures 6 to 8 The simulation results of the electromagnetic field under the conditions of 220kV, 500kV and 800kV are shown in the figure. Detailed Implementation

[0021] To facilitate understanding of the technical solution of the present invention by those skilled in the art, the technical solution of the present invention will now be further described in conjunction with the accompanying drawings.

[0022] 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 technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0023] Please see Figure 1 As shown, this invention provides a method for constructing a generative adversarial network to enhance electromagnetic field simulation data, comprising: S10: Establish an attention mechanism to identify regions with significant electric field distortion in coarse-grid electromagnetic field data and assign processing weights.

[0024] In one embodiment of the present invention, it specifically includes: S11, based on coarse-grid electromagnetic field data and the electric field derived therefrom To obtain a set of complementary physical feature maps.

[0025] In this embodiment, multi-physics feature extraction is performed: receiving input coarse-grid simulation data of the electromagnetic field. . This refers to a coarse-grid potential distribution or a coarse-grid field distribution directly output by simulation software. When the input is a potential distribution, the electric field E is calculated from this potential distribution. To comprehensively identify the electric field distortion regions in GIS equipment caused by typical defects such as fixed protrusions and suspended metal particles, the first step is to... And from the electric field E obtained therefrom, a set of complementary physical characteristic maps are calculated, including: Field strength amplitude diagram: It is used to identify areas with high absolute field strength, which are high-risk areas for insulation breakdown. E is the electric field vector used to calculate physical characteristics; when the input is a potential distribution, E is calculated from that potential distribution; when the input is directly a field distribution, E is that field distribution.

[0026] Field strength gradient magnitude map: It is used to identify regions with drastic changes in field strength, such as the edges of metal spikes or the boundaries of local field distortion. The gradient of the electric field intensity scalar field |E| is used to characterize the rate of change of the field intensity with spatial location.

[0027] Electric field divergence plot: This is used to identify regions with abnormal charge accumulation. For passive regions ( This feature is simplified to .in, Represents the gradient operator, Represents free charge density, This represents the vacuum permittivity. This characteristic is extremely sensitive to charge concentration regions caused by fixed metal spike defects and local charge accumulation caused by suspended particles, even if the electric field gradient in the region does not increase significantly.

[0028] Electric field direction consistency diagram: .in, A measure of the uniformity of electric field direction within a local neighborhood. This represents the set of electric field vectors within the local neighborhood of the current location. This feature identifies regions with disordered electric field orientations by calculating the consistency of the electric field vector directions within the local neighborhood (such as mean cosine similarity). When suspended particles reciprocate or rotate in an alternating electric field, it leads to a high degree of disorder in the electric field orientation around them, while the field strength amplitude and gradient may not change much. Regions with low directional consistency are an important indicator of dynamic defects.

[0029] S12, concatenate the physical feature maps along the channel dimension to obtain the fused features. .

[0030] In this embodiment, the original input The calculated multi-physical feature map is concatenated with the fused feature map along the channel dimension to obtain the fused feature. : ; S13 will integrate features Input a lightweight convolutional network The network learns the spatial importance distribution through convolutional operations and finally normalizes it using the sigmoid function to obtain a preliminary attention weight map. .

[0031] In this embodiment, attention weight calculation: Input a lightweight convolutional network The network learns the spatial importance distribution through convolutional operations and finally normalizes it to the [0,1] interval using the Sigmoid function to obtain a preliminary attention weight map. : ; In the formula, This is the Sigmoid function.

[0032] S14, Preliminary Attention Weight Map Apply a joint filter based on multiple physical features to obtain a priority weight map. .

[0033] In this embodiment, a priority weight map is generated through a multi-physical feature-guided attention network. This network integrates multiple physical features mentioned above to more comprehensively identify key regions. Multi-feature constraint filtering: This involves adjusting the initial attention weight map... Apply a joint filter based on multiple physical features. Set an empirical threshold for each physical feature. ), and generate a joint binary mask. The mask is set to 1 in regions where any physical feature exceeds its threshold, and 0 otherwise. This ensures that any physical feature indicating a potential defect in the region is prioritized, as shown in the following formula: ; In the formula, , , , These are a set of complementary physical feature maps, namely, the field strength amplitude map, the field strength gradient amplitude map, the electric field divergence map, and the electric field direction consistency map; ( ) are the empirical thresholds corresponding to a set of complementary physical feature maps. It is a union. Among them, This refers to directional consistency; a lower value indicates more disordered directions, hence the use of values ​​below the threshold. The judgment conditions. Finally, the priority weight graph. The calculation is shown in the following formula: ; Here, ⊙ represents element-wise multiplication. The final priority weights are assigned only to candidate key regions (regions with significant electric field distortion) indicated by one or more physical features, and then the importance of the candidate regions is distinguished by continuous weights, thereby suppressing noise interference.

[0034] S15, based on priority weight graph coarse grid data of electromagnetic fields The weighted feature map is obtained by performing weighted processing. .

[0035] Among them, weighted feature map Obtain it using the following formula: ; In the formula, This is the weighting enhancement coefficient.

[0036] S20, construct a generative network to map the weighted coarse-grid electromagnetic field data into fine-grid electromagnetic field data.

[0037] In one embodiment of the present invention, the weighted feature map The input to the generator network is used as input. The generator network employs a physically guided encoder-decoder structure, which, based on standard convolution and upsampling operations, incorporates the following targeted design features to ensure the physical validity of the generated data.

[0038] In this embodiment, physical feature encoding involves not only extracting the spatial features of the data in the encoder but also explicitly encoding physical information, such as the position coordinates of grid nodes and material property boundaries (e.g., the interface between conductors and insulating media). This physical information is then used as an additional channel alongside the input data. After splicing, they are input into the encoder together to achieve this. In this embodiment, the differential operator is embedded: in the decoder, a portion of the standard convolutional layer is replaced by a differentiable numerical differential operator (such as a gradient operator based on convolution), so that the decoder implicitly learns and satisfies the smoothness and continuity constraints of the field distribution in the feature space. In this embodiment, a physical consistency loss function is defined. To ensure the fine-grid electromagnetic field data output by the decoder To conform to the fundamental laws of electromagnetic fields, a physical regularization loss based on Maxwell's equations is introduced during training. This loss function measures the degree to which the decoder output deviates from physical laws, and is defined as follows: Residual loss: For a passive region in an electrostatic field, it is equivalent to requiring the fine-grid predicted potential distribution output by the decoder to satisfy the Laplace equation. Let the predicted potential distribution output by the decoder be... enh From this, we obtain the electric field E = - enh The residual loss of PDEs (partial differential equations) is defined as the squared L2 norm of the residuals of the equations within the computational domain: ; In the formula, For PDE residual loss, The Laplace operator can be computed using the automatic differentiation function of a convolutional network. For the first The generated potential value at each sampling point To calculate the number of sampling points within the domain, The value is the squared L2 norm. This loss is used to constrain the generated potential distribution to approximately satisfy the Laplace equation.

[0039] Boundary condition loss: Simultaneously, the generated data must satisfy known boundary conditions (such as potential values ​​on specific boundaries). The boundary loss is defined as follows: ; In the formula, For boundary condition loss, The number of boundary points. For the first The predicted potential values ​​of the data generated at each boundary point For the first The known boundary potential values ​​at each boundary point.

[0040] Total Physical Loss: Combining PDE residual loss and boundary condition loss, the total physical consistency loss is obtained: ; In the formula, For the total physical consistency loss, and These are weighting coefficients, used to balance the residual loss of PDE. L pde and boundary condition loss L bc . contributions.

[0041] S30 generates confidence scores and error distribution maps for fine-grid electromagnetic field data to assess data quality.

[0042] In one embodiment of the present invention, physical plausibility calculation is performed by receiving fine-grid electromagnetic field data during runtime. And the original coarse-grid electromagnetic field data The system outputs confidence assessment results based on the generated results and their related features; during the training phase, it can also utilize real electromagnetic field fine-grid data. A monitoring signal is constructed to calibrate the evaluation results. Among these, I real This indicates real, fine-grid supervised data used only during the training phase.

[0043] S31, Calculate fine-grid electromagnetic field data Physical rationality indicators.

[0044] In this embodiment, to ensure that the confidence assessment has physical meaning, the physical reasonableness index of the generated data is first calculated: Physical residuals of the field distribution: These are the physical residuals of the generated data calculated based on Maxwell's equations. For electrostatic problems, the Laplace equation residuals are primarily calculated. ; In the formula, To generate the physical residuals of the field distribution corresponding to the data, For electromagnetic field fine mesh data The obtained potential distribution, the residual, measures the degree to which the generated data conforms to the basic laws of electromagnetic fields.

[0045] Boundary condition compliance: Calculate the degree to which the generated data conforms to the known boundary conditions, as shown in the following formula: ; In the formula, For boundary condition residuals, To generate the potential values ​​of the data at the boundary, The boundary potential value is known.

[0046] Medium interface constraints: For medium interfaces and conductor boundaries in GIS, the interface constraint residuals are calculated based on the boundary / interface conditions preset in the simulation model. ; The larger the value, the weaker the degree to which the generated result satisfies the preset physical constraints of the region. When constructing the physical rationality features mentioned above, the following will be considered: Unified integration with boundary condition residuals It is used in the project and reflected in subsequent measurement loss and error assessment.

[0047] S32, based on fine-grid electromagnetic field data Coarse grid data of electromagnetic fields Physical rationality indicators extract quality characteristics from multiple dimensions and integrate them into comprehensive evaluation indicators.

[0048] In this embodiment, quality features are extracted from multiple dimensions and fused into a comprehensive evaluation index: Data consistency feature: Calculate the difference between the generated data and the coarse grid data in the overlapping area to ensure the consistency between the generated result and the input data, as shown in the following formula: ; In the formula, For data consistency characteristics, This is the result of downsampling the generated data to a coarse grid resolution.

[0049] Physical rationality characteristics: Based on the calculated field distribution residuals and the unified fusion of boundary / interface constraint residuals, physical rationality characteristics are constructed as shown in the following formula; where, the above steps yield... In construction When merged This item will be used in subsequent confidence assessments and loss calculations. ; formula, As a physical rationality feature, This is a normalization operation to make the characteristics of different physical quantities comparable. (In constructing...) At that time, the interface constraint residuals obtained from the above steps Unified integration with boundary condition residuals This item is used to supplement constraints on physical rationality characteristics.

[0050] Structural integrity features: Multi-scale structural features of the generated data are extracted using a pre-trained feature extractor (such as a VGG network). During the training phase, the structural feature differences between the generated data and the real data can be used as supervisory information to calibrate the measurement module. During the application phase, the multi-scale structural response of the generated data itself is used for quality assessment. This feature is combined with data consistency features and physical plausibility features in S33 to form a comprehensive quality feature.

[0051] S33, based on comprehensive evaluation indicators, outputs a confidence score for physical guidance. Error distribution plot .

[0052] In this embodiment, feature fusion involves concatenating multi-dimensional features along the channel dimension to form a comprehensive quality feature, as shown in the following formula: ; In the formula, For comprehensive evaluation indicators, This refers to structural integrity characteristics.

[0053] In this embodiment, the physical weighted confidence score is calculated by inputting the comprehensive evaluation index into a lightweight neural network (such as an MLP) and weighting it according to physical plausibility. Specifically, a weight for physical rationality is introduced. This weight is inversely proportional to the physical residual, ensuring that regions with poor physical plausibility receive a lower confidence weight, as shown in the following formula: ; ; In the formula, For the Sigmoid function, It is a lightweight neural network. For comprehensive evaluation indicators, As a weight for physical rationality, For element-wise multiplication, It is an exponential function. This is an adjustment coefficient for the weight of physical rationality. This represents the physical residual of the field distribution.

[0054] Error distribution plot generation: Generating error distribution plots based on physical residuals and data differences Highlighting physically illogical areas, as shown in the following formula: ; In the formula, This is a graph showing the error distribution predicted by the convolutional network. For fine-grid electromagnetic field data, For coarse-grid data of electromagnetic fields The result of upsampling to a fine grid resolution.

[0055] S34, and based on confidence scores Error distribution plot Define the measurement loss function.

[0056] In the embodiment, a measurement loss function is defined. It can be expressed by the following formula: ; In the formula, represents the total number of fine grid locations involved in the measurement loss statistics; For location index; Let be the confidence score at the i-th position; For the first Error values ​​at each location; bolded letters For the first The comprehensive physical residual at each location can be obtained from the field distribution residual and the unified fusion of the boundary / interface constraint residual, which is used to directly constrain the physical rationality of the generated data. and is the weighting coefficient; ||·|1 is the L1 norm. When it is a scalar, It degenerates into its absolute value.

[0057] S40 constructs a discriminative network and performs adversarial training on coarse-grid and fine-grid electromagnetic field data.

[0058] In one embodiment of the present invention, the generator network and the discriminator network constitute a standard adversarial learning framework, specifically: S41 defines the discriminator as a fully convolutional binary classifier used to distinguish between generated electromagnetic field fine-grid data and real electromagnetic field fine-grid data.

[0059] S42 uses the least squares adversarial loss function to define the adversarial loss of the discriminator and performs training optimization, specifically expressed by the following formula: ; In the formula, For counter-loss, For mathematical expectation operators, This is the output function of the discriminator.

[0060] S50 introduces a priority-guided feedback mechanism to optimize the generation quality of regions with significant electric field distortion, and integrates it into the overall loss function for joint training.

[0061] In one embodiment of the present invention, a feedback loss mechanism based on priority weights and confidence evaluation is constructed. This mechanism updates the network parameters through backpropagation of the loss function, rather than directly modifying the generated data, thereby ensuring consistency in processing logic between the training and testing phases. Specifically, it includes the following steps: S511, according to the priority weight diagram Confidence score Obtain the feedback weight graph Specifically, it is obtained through the following formula: ; in, This indicates element-wise multiplication. The weight map has larger values ​​in key regions and where the generation confidence is low, thus guiding the training process to focus on these challenging areas.

[0062] S512, based on feedback weight graph Define a priority-guided feedback loss function. This measures the quality of generated fine-grid electromagnetic field data in regions with significant electric field distortion and low confidence levels. Compared with fine-grid data of real electromagnetic fields Weighted differences between them; Among them, the feedback loss function Expressed using the following formula: ; In the formula, This represents the total number of electromagnetic field mesh locations participating in the feedback loss statistics. For the first Feedback weight map at each location For the first Fine-grid data of the electromagnetic field at each location. For the first Fine-grid data of the actual electromagnetic field at each location. The L2 norm squared. The loss term incentivizes the augmentation module to prioritize improving the generation accuracy of high-weight regions by imposing a larger penalty for reconstruction error.

[0063] In one embodiment of the present invention, the feedback loss is integrated into the overall optimization objective and participates in the joint training of the network. This loss term, as an important supervisory signal, together with the measurement loss, physical consistency loss, and adversarial loss, guides the optimization direction of the network and enhances the network parameters through backpropagation.

[0064] In this embodiment, joint optimization includes: S521, a generative adversarial network consisting of a generator network and a discriminator network, has an overall loss function during the joint training phase that is mainly composed of the weighted sum of measurement loss, physical consistency loss, feedback loss, and adversarial loss.

[0065] Wherein, the overall loss function Expressed using the following formula: ; In the formula, λ 1. λ 2 and λ 3 represents the weighting parameters, used to balance the importance of measurement loss, physical consistency loss, and feedback loss. To measure the loss, For the total physical consistency loss, For feedback loss, in this formula, adversarial loss L adv The coefficient is fixed at 1.

[0066] S522 defines the priority loss as the reconstruction error of the high-priority region, which is used to guide the training of regions with significant electric field distortion during the initial training phase. ; In the formula, For priority loss, The total number of fine grid locations participating in the priority loss statistics. For fine grid position indexing, For the first Priority weights at each fine grid location For the first Generated electromagnetic field fine-grid data at each fine-grid location. For the first Real electromagnetic field fine-grid data at each fine-grid location.

[0067] S523 employs a two-stage training strategy: The first stage defines the priority loss as the reconstruction error in high-priority regions, used to guide the initial training of regions with significant electric field distortion, and utilizes priority loss and physical consistency loss to obtain the basic generated results; the second stage introduces measurement loss, feedback loss, and adversarial loss for end-to-end joint training to minimize the overall loss function. For the goal.

[0068] This invention also provides a system for constructing an electromagnetic field simulation data-enhanced generative adversarial network, which applies the above-described method for constructing an electromagnetic field simulation data-enhanced generative adversarial network, including: The priority module is used to establish an attention mechanism to identify regions of significant electric field distortion in coarse-grid electromagnetic field data and assign processing weights.

[0069] The enhancement module is used to build a generative network that maps the weighted coarse-grid electromagnetic field data to fine-grid electromagnetic field data.

[0070] The measurement module is used to generate confidence scores and error distribution maps for fine-grid electromagnetic field data to assess data quality.

[0071] The discriminant module is used to construct a discriminant network, performing adversarial training on coarse-grid and fine-grid electromagnetic field data. It also introduces a priority-guided feedback mechanism to optimize the generation quality of significant electric field distortion regions, integrating this optimization into the overall loss function for joint training.

[0072] Experimental testing; Please see Figures 2 to 8 As shown, to make the technical solution, training data source, and testing process of this invention clearer, the experimental testing scheme of this invention is described below, combining existing electromagnetic field coarse mesh simulation results, electromagnetic field fine mesh simulation results, and electric field data files exported from regular volumetric meshes. The following experimental content only focuses on experimental designs supported by existing simulation data and does not introduce additional verification steps that do not correspond to existing data.

[0073] I. Sources of Experimental Data like Figures 2 to 5 As shown. The experimental data comes from coarse-grid and fine-grid simulation results of the same GIS-type electric field simulation model under different voltage conditions. The coarse-grid simulation results cover four voltage conditions: 220kV, 500kV, 800kV, and 1000kV; the fine-grid simulation results cover three voltage conditions: 220kV, 500kV, and 800kV. For the 220kV, 500kV, and 800kV conditions, coarse-grid simulation results and corresponding fine-grid simulation results are constructed into one-to-one supervised sample pairs for training and testing the network of this invention. The 1000kV condition provides coarse-grid simulation results and their derived regular volume grid electric field data, which can be used as input samples for higher voltage conditions.

[0074] Figures 2 to 5 The coarse-grid simulation results for 220kV, 500kV, 800kV, and 1000kV operating conditions are presented. This data reflects the coarse-grid field distribution of the same structural model under different voltage conditions and can serve as input data for the priority module, enhancement module, and measurement module of this invention. The fine-grid simulation results for 220kV, 500kV, and 800kV operating conditions can be used as monitoring targets, while the 1000kV result can be used as input samples for higher voltage conditions.

[0075] Figures 6 to 8The simulation results of fine mesh under 220kV, 500kV, and 800kV operating conditions are presented. This set of data is consistent with... Figures 2 to 4 The coarse mesh results correspond one-to-one and can be used as a supervision target for the output results of the enhancement module of this invention to evaluate the difference between the generated fine mesh data and the actual fine mesh simulation results.

[0076] II. Experimental Data Preprocessing and Sample Construction In the experiment, the coarse-grid simulation results and fine-grid simulation results were first unified in coordinate alignment and represented by a regular grid. For the directly exported vector field data file, the file must contain at least the spatial coordinates X, Y, Z and the corresponding electric field components Ex, Ey, Ez. Taking the cu1000.csv exported under the 1000kV condition as an example, its sampling space range is [-420mm, -420mm, 0mm] to [420mm, 420mm, 1300mm], the sampling step size is 15mm, and the corresponding regular volumetric grid size is 57×57×87.

[0077] For regular volumetric grid electric field data, the electric field amplitude |E| is first calculated from Ex, Ey, and Ez. Based on this, a field strength amplitude map, a field strength gradient amplitude map, an electric field divergence map, and a direction consistency map are constructed, serving as multi-physical feature inputs for the priority module to identify key regions. For paired coarse and fine grid samples, the coarse grid field distribution is used as the input data Ilow in a unified coordinate system, and the corresponding fine grid field distribution is used as the supervised target Ireal, to train the enhancement module's output generation result Ienh.

[0078] III. Experimental Training and Test Design The experimental training was based on paired coarse and fine grid samples under three operating conditions: 220kV, 500kV, and 800kV. For the field distribution results under each operating condition, training samples could be further extracted according to the central axis section, typical radial section, and local high-field regions to increase the number of effective samples and retain information on the overall field distribution and key local areas. During training, coarse grid data... Its multi-physical feature input priority module, the priority module outputs a priority weight map and weights the input; the enhancement module receives the weighted input and outputs the generated fine mesh result. During the training phase, the corresponding real fine-mesh results are then utilized. Calculate the reconstruction error, physical consistency loss, and feedback loss, and update the network parameters accordingly.

[0079] The testing phase still primarily uses paired samples under 220kV, 500kV, and 800kV operating conditions, and results are generated through comparison. Compared with the actual fine mesh results The differences were evaluated to assess the ability of the proposed method to reconstruct fine-grid field distributions. Since only coarse-grid simulation results and regular-grid field data were provided for the 1000kV operating condition, these were used as input samples for higher voltage levels in the experimental design and input into the trained model to obtain the corresponding refined field distribution results.

[0080] IV. Evaluation Methods The experimental evaluation was set up around indicators that could be directly achieved under the existing data conditions of this invention. First, the results were generated. Compared with the actual fine mesh results The mean square error, relative L2 error, and structural similarity indices are used to evaluate the overall reconstruction effect. Secondly, regions with higher field strength in the actual fine mesh results are designated as key regions, and the average and peak errors within these key regions are statistically analyzed to evaluate the improvement effect of the priority guidance mechanism of this invention on the fidelity of key regions. Thirdly, the physical consistency loss defined in this invention and the output of the measurement module are used to check the physical rationality of the generated results in the known boundary neighborhood and regions with drastic local field changes.

[0081] The above experimental test content corresponds to the existing coarse grid data, fine grid data and regular volume grid electric field data of this invention, and can characterize the training data organization method, test process and evaluation method of the method of this invention in the task of enhancing electromagnetic field data from coarse grid to fine grid.

[0082] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0083] The above embodiments are merely examples of implementation methods of the invention. The scope of protection of the present invention is not limited to the above embodiments. For those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for constructing a generative adversarial network to enhance electromagnetic field simulation data, characterized in that, include: An attention mechanism is established to identify regions of significant electric field distortion in coarse-grid electromagnetic field data and assign processing weights accordingly. A generative network is constructed to map the weighted coarse-grid electromagnetic field data into fine-grid electromagnetic field data; Generate confidence scores and error distribution plots for fine-grid electromagnetic field data to assess data quality; Construct a discriminative network and conduct adversarial training on coarse-grid and fine-grid electromagnetic field data; A priority-guided feedback mechanism is introduced to optimize the generation quality of regions with significant electric field distortion, and this optimization is integrated into the overall loss function for joint training.

2. The method for constructing a generative adversarial network to enhance electromagnetic field simulation data according to claim 1, characterized in that, Identify regions of significant electric field distortion in coarse-grid electromagnetic field data and assign processing weights, including: Based on coarse grid data of electromagnetic field and the electric field derived therefrom To obtain a set of complementary physical feature maps; The physical feature maps are concatenated along the channel dimension to obtain the fused features. ; Fusion features Input a lightweight convolutional network The network learns the spatial importance distribution through convolutional operations and finally normalizes it using the sigmoid function to obtain a preliminary attention weight map. ; Preliminary attention weight map Apply a joint filter based on multiple physical features to obtain a priority weight map. ; Based on priority weight graph coarse grid data of electromagnetic fields The weighted feature map is obtained by performing weighted processing. .

3. The method for constructing a generative adversarial network to enhance electromagnetic field simulation data according to claim 2, characterized in that, Obtain the priority weight graph include: An empirical threshold is set for each physical feature map, and a joint binary mask is generated. The mask is set to 1 in any region where any physical feature exceeds its threshold, and 0 otherwise.

4. The method for constructing a generative adversarial network to enhance electromagnetic field simulation data according to claim 1, characterized in that, The weighted coarse-grid electromagnetic field data is mapped to fine-grid electromagnetic field data, including: The generative network employs a physically guided encoder-decoder structure; it weights the feature maps... As input; in the encoder, not only are the spatial features of the data extracted, but physical information is also explicitly encoded and used as an additional channel with the input data. After being concatenated, they are input into the encoder together; in the decoder, a differentiable numerical differential operator is used to replace part of the standard convolutional layer, so that the decoder implicitly learns and satisfies the smoothness and continuity constraints of the field distribution in the feature space. Define the physical consistency loss function To ensure the fine-grid electromagnetic field data output by the decoder It conforms to the basic laws of electromagnetic fields and introduces physical regularization loss based on Maxwell's equations during the training process.

5. The method for constructing a generative adversarial network to enhance electromagnetic field simulation data according to claim 1, characterized in that, Generate confidence scores and error distribution plots for fine-grid electromagnetic field data to assess data quality, including: Calculate fine-grid data for electromagnetic fields Physical rationality indicators; Based on electromagnetic field fine mesh data Coarse grid data of electromagnetic fields Physical rationality indicators extract quality characteristics from multiple dimensions and integrate them into comprehensive evaluation indicators; Based on comprehensive evaluation indicators, a confidence score for physical guidance is output. Error distribution plot ; And based on confidence scores Error distribution plot Define the measurement loss function.

6. The method for constructing a generative adversarial network to enhance electromagnetic field simulation data according to claim 5, characterized in that, Output the confidence score of the physical guidance Error distribution plot It can be expressed by the following formula: ； ； ； In the formula, For the Sigmoid function, It is a lightweight neural network. For comprehensive evaluation indicators, As a weight for physical rationality, For element-wise multiplication, It is an exponential function. This is an adjustment coefficient for the weight of physical rationality. For the physical residuals of the field distribution, This is a graph showing the error distribution predicted by the convolutional network. For fine-grid electromagnetic field data, For coarse-grid data of electromagnetic fields The result of upsampling to a fine grid resolution.

7. The method for constructing a generative adversarial network to enhance electromagnetic field simulation data according to claim 1, characterized in that, Adversarial training was conducted using coarse-grid and fine-grid electromagnetic field data, including: The discriminator is defined as a fully convolutional binary classifier used to distinguish between generated electromagnetic field fine-grid data and real electromagnetic field fine-grid data. The least squares adversarial loss function is used to define the adversarial loss of the discriminator and to optimize the training.

8. The method for constructing a generative adversarial network to enhance electromagnetic field simulation data according to claim 1, characterized in that, A priority-guided feedback mechanism is introduced to optimize the generation quality of regions with significant electric field distortion, including: According to the priority weight diagram Confidence score Obtain the feedback weight graph ; Based on feedback weight graph Define a priority-guided feedback loss function. This measures the quality of generated fine-grid electromagnetic field data in regions with significant electric field distortion and low confidence levels. Compared with fine-grid data of real electromagnetic fields Weighted differences between them; Among them, the feedback loss function Expressed using the following formula: ； In the formula, This represents the total number of electromagnetic field mesh locations participating in the feedback loss statistics. For the first Feedback weight map at each location For the first Fine-grid data of the electromagnetic field at each location. For the first Fine-grid data of the actual electromagnetic field at each location. It is the square of the L2 norm.

9. The method for constructing a generative adversarial network to enhance electromagnetic field simulation data according to claim 8, characterized in that, A priority-guided feedback mechanism is introduced to optimize the generation quality of regions with significant electric field distortion, and this optimization is integrated into the overall loss function for joint training, including: The overall loss function of a generative adversarial network (GAN), which consists of a generative network and a discriminative network, is mainly composed of the weighted sum of measurement loss, physical consistency loss, feedback loss, and adversarial loss during the joint training phase. Priority loss is defined as the reconstruction error of high-priority regions, which is used to guide regions with significant electric field distortion during the initial training phase. A two-stage training strategy is adopted: In the first stage, the priority loss is defined as the reconstruction error of the high-priority region, which is used to guide the training in the initial stage for regions with significant electric field distortion, and the priority loss and physical consistency loss are used to obtain the basic generation results; in the second stage, measurement loss, feedback loss and adversarial loss are introduced for end-to-end joint training to minimize the overall loss function. For the goal.

10. A system for constructing an electromagnetic field simulation data-enhanced generative adversarial network, employing the electromagnetic field simulation data-enhanced generative adversarial network construction method according to any one of claims 1-9, characterized in that, include: The priority module is used to establish an attention mechanism, identify regions of significant electric field distortion in coarse-grid electromagnetic field data, and assign processing weights accordingly. The enhancement module is used to build the generative network, which maps the weighted coarse electromagnetic field data to fine electromagnetic field data. The measurement module is used to generate confidence scores and error distribution maps for fine-grid electromagnetic field data to assess data quality. The discrimination module is used to construct a discrimination network and perform adversarial training on coarse-grid and fine-grid electromagnetic field data. Simultaneously, a priority-guided feedback mechanism is introduced to optimize the generation quality of significant electric field distortion regions, and this is integrated into the overall loss function for joint training.

Images