Generative large model-based digital twin three-dimensional model construction method

Abstract

The invention provides a digital twin three-dimensional model construction method based on a generative large model, and the method comprises the steps: obtaining multi-source monitoring data of a power distribution network, and processing the multi-source monitoring data into a training data set; the method comprises the following steps: mapping multi-source monitoring data into a multi-scale tensor subspace through tensor wavelet structured transformation, adaptively extracting spatial features through a learnable wavelet kernel, and keeping the structural continuity of a physical field in combination with a geometric prior regular term; constructing and training a generative adversarial network through a training data set; inputting and analyzing the physical parameter vector of the target scene, and if the topological similarity score is lower than a preset threshold value, adjusting noise vector regeneration; and if yes, outputting a three-dimensional model tensor and importing the three-dimensional model tensor into a digital twin platform, and driving real-time physical field visualization. According to the method, characteristics of a multi-scale space structure and a nonlinear physical field can be reserved, the physical rationality and generalization ability of the generated model are remarkably improved, and depth identification of topological attributes (such as hole connectivity and surface defects) and local geometric defects of the three-dimensional model is realized.

Description

Technical Field

[0001] This invention relates to the field of power distribution networks, and in particular to a method for constructing a digital twin 3D model based on a generative large model. Background Technology

[0002] In the current context of digital twin technology development, 3D model construction has become a core component in fields such as industrial simulation, product design, and intelligent manufacturing. Currently, digital twin 3D models also play a crucial role in fault analysis and diagnosis of power distribution networks. However, existing 3D modeling methods generally rely on geometric reconstruction or low-dimensional feature extraction, such as principal component analysis (PCA), basic wavelet transform, or traditional deep convolutional networks. These methods have significant limitations when processing high-dimensional physical field data and complex topological structures. On the one hand, traditional methods cannot fully capture the spatial coupling relationships between multimodal physical quantities such as temperature, stress, and velocity fields, resulting in a lack of physical consistency in the modeling results. On the other hand, while existing generative adversarial networks (GANs) possess strong sample generation capabilities, they often neglect constraints on physical laws (such as mass conservation, energy balance, and stress continuity), leading to geometric defects in the generated models, such as cracks, missing holes, and non-manifold edges, which fail to meet the accuracy and reliability requirements of engineering applications. Furthermore, when faced with high-dimensional sparse data and complex spatial patterns, existing preprocessing methods often rely on linear or empirical compression techniques, which cannot preserve local geometric structures and non-Euclidean spatial characteristics. At the same time, during training, due to the coupling interference between physical gradients and adversarial gradients, traditional optimization strategies often lead to training instability, pattern collapse, or even convergence failure.

[0003] Chinese invention patent CN120317100A discloses a method, system, and equipment for automatically generating the spatial layout of a digital twin of a substation. The method includes: first, data acquisition and calibration to obtain accurate electrical parameters, capacity requirements, and operating status data; then, power flow calculation modeling to calculate the active and reactive power injection of nodes to ensure more accurate power flow calculation results; specifying that the layout scheme must meet various constraints when defining constraints; selecting appropriate optimization algorithms and setting objective functions to search for a global optimum or a series of Pareto optimum solutions; using a 3D model for display, constructing an interactive environment using virtual reality technology and the Three.js library, establishing a data transmission and communication network; employing an event-driven power system simulation method, and using filtering algorithms to fuse and estimate measurement data to obtain real-time electrical parameter estimates; and finally, utilizing the real-time simulation results.

[0004] The existing technology has the following shortcomings:

[0005] 1. Linear or fixed transformation methods such as PCA and conventional wavelets cannot accurately preserve topological connectivity and geometric complexity.

[0006] 2. Traditional generative adversarial networks do not employ physical control mechanisms, and the generative models often exhibit physical violations such as stress imbalance and displacement discontinuity.

[0007] 3. The mixed input of physical parameters and noise leads to unstable model learning, and the generated samples depend on a specific training data distribution.

[0008] 4. Standard Adam cannot handle the conflict between physical gradients and adversarial gradients, which can easily lead to mode collapse and gradient oscillations during training. Summary of the Invention

[0009] To address the problems existing in the background technology, this invention proposes a method for constructing a digital twin 3D model based on a generative large model.

[0010] A method for constructing a digital twin 3D model based on a generative large model, characterized by the following steps:

[0011] S1. Acquire multi-source monitoring data of the power distribution network and process it into a training dataset;

[0012] S2. Through tensor wavelet structured transformation, multi-source monitoring data is mapped into multi-scale tensor quantum space. Spatial features are adaptively extracted through learnable wavelet kernels, and the continuity of the physical field structure is maintained by combining geometric prior regularization terms.

[0013] S3. Construct and train a generative adversarial network using a training dataset, where the generator adopts a physically constrained decoupled generator and the discriminator adopts a multi-scale topology-aware discriminator.

[0014] S4. Input the physical parameter vector of the target scene into the Generative Adversarial Network and analyze it. If the topological similarity score is lower than the preset threshold, adjust the noise vector and regenerate. If the verification is successful, output the final three-dimensional model tensor. Directly import the three-dimensional model tensor into the digital twin platform to drive real-time physical field visualization.

[0015] Based on the above, in step S1, the multi-source monitoring data includes at least physical field distribution data, electromagnetic field intensity and phase information, material property distribution and internal microstructure geometric data, and high-dimensional physical field simulation data; the measured data and simulation data are spatiotemporally aligned and coordinate registered to construct a spatiotemporally synchronized multimodal dataset; after denoising, outlier removal and missing value imputation of the original multimodal dataset, the data is divided into channel dimensions according to physical field type to form a four-dimensional training tensor, and finally the physical parameter vector is labeled as conditional label to generate the training dataset required for supervised learning.

[0016] Based on the above, in step S2, an adaptive sparse mask matrix is ​​generated based on the original input four-dimensional training tensor:

[0017] M spa =sig(X-τ) sp )

[0018] In the formula, M spa τ represents the adaptive sparse mask matrix used to filter noise; sig(·) is the Sigmoid activation function, which compresses the input to the [0,1] interval; X is the four-dimensional training tensor of the original input; τ sp For dynamic threshold parameters, the subscript sp indicates sparsity control; X-τ sp Implement thresholding;

[0019] A learnable 3D convolution kernel is used to perform multi-scale convolution operations on the masked original input 4D training tensor. Then, features from different scales are concatenated and fused with a geometric curvature regularization term to output multi-scale structured features, represented as follows:

[0020]

[0021] In the formula, J wav (·) denotes the tensor wavelet transform operator, J wav (X) represents multi-scale structured features; This indicates the implementation of multi-scale feature stitching operation; K is the preset total number of wavelet decomposition scales; W k The kernel is a learnable 3D convolution kernel, with the subscript k representing the scale index, which is adaptively optimized through training. Represents a 3D convolution operation to extract local spatial patterns; ⊙ represents element-wise multiplication, applying a sparse mask to filter noise; P geo (X) is the geometric curvature regularization term, calculated as follows: It is a second-order partial derivative tensor that describes the physical field in x. i ,x j Spatial curvature of direction; g m is the curvature basis vector, and the subscript m represents the index of the basis, used to fit the local surface geometry.

[0022] Based on the above, step S3 includes step S301: The physical constraint decoupling generator constrains the physical sensing features through orthogonal encoding and combines the physical decoupling loss to forcibly separate physical parameters from random noise. The specific steps are as follows:

[0023] (1) Construct a physical sensing encoder, represented as:

[0024] Q phys (z)=MLP(z)·A orth

[0025] In the formula, Q phys(·) represents the physical sensing encoder, which outputs physical constraint features; MLP(z) is a multilayer perceptron that maps the noise vector z to the latent space; z is the noise vector; A orth This is the orthogonality constraint matrix, where the subscript "orth" indicates orthogonality.

[0026] (2) Design a physical condition injection module to output physical condition characteristics, represented as follows:

[0027] F con (C phy ) = tanh(W c C phy +b c )

[0028] In the formula, F con (·) indicates the physical condition injection module, F con (C phy ) represents physical condition characteristics; C phy This is a vector of physical parameters, where the subscript phy indicates a physical property; W c The subscript c represents the learnable weight matrix; b c is the bias vector; tanh(·) is the hyperbolic tangent activation function, which restricts the output to the interval [-1, 1];

[0029] (3) Generate a decoupled 3D model and output the generated 3D model tensor, represented as:

[0030] Y gen =D dec (Q phys (z)||F con (C phy ))

[0031] In the formula, Y gen D is the tensor of the generated 3D model. dec (·) represents a deconvolutional decoder; || represents a feature concatenation operation, which connects the input along the channel dimension;

[0032] (4) The physical decoupling loss constraint is expressed as:

[0033]

[0034] In the formula, L dis This represents the physical decoupling loss function, where the subscript dis indicates decoupling; ||·|| F Given the Frobenius norm, calculate the square root of the sum of squares of all elements in the matrix; Y gen J is the tensor of the generated 3D model; phyγ is the Jacobian matrix of the physical field, with the subscript phy indicating the physical field; γ is the decoupling strength coefficient, which controls the weight of the physical law constraints; E[·] represents the expectation operator, which calculates the mean on the batch data.

[0035] Based on the above, step S3 includes step S302: The multi-scale topology-aware discriminator enhances its ability to identify the global topological structure and local geometric defects of the 3D model by fusing continuous coherence features and multi-scale convolutional features. The specific steps are as follows:

[0036] (1) Calculate the persistent homology feature, expressed as:

[0037] S topo =PH(Y);

[0038] Among them, S topo Represents the set of persistent homology features; PH(·) is the persistent homology computation operator; Y is the input 3D model tensor;

[0039] (2) Construct a multi-scale convolutional feature extraction network, represented as:

[0040]

[0041] This represents a 3D convolutional subnetwork at scale s, which outputs the feature vector at that scale; s is the scale index, s∈{1,2,…,S}, and S is the preset total number of scales; the superscript (s) indicates that the network parameters of different scale branches are independent;

[0042] (3) Calculate the adaptive scaling weights, expressed as:

[0043]

[0044] In the formula, α s f represents the adaptive weights at the s-th scale; s w is the output feature vector of the s-th scale convolutional network; s The subscript s represents the scale index; exp(·) is the exponential function that performs softmax normalization. f is the transpose of the learnable weight vector at the j-th scale; j This represents the output feature vector of the j-th scale convolutional network. This represents the vector transpose operation;

[0045] (4) Evaluate topological similarity, expressed as:

[0046]

[0047] In the formula, N sim B represents the topological similarity score, with a value range of [0,1].k (·) represents the set of k-dimensional Betti numbers, describing the number of k-dimensional topological holes; S topo S is the set of persistently homologous features of the input 3D model; ref The reference topological features are usually derived from real sample datasets; Δ represents the symmetric difference operator; |·| represents the cardinality of the set;

[0048] (5) The output of the integrated discriminator is represented as:

[0049]

[0050] In the formula, D φ (·) represents the final output of the discriminator; β is the topological similarity weight coefficient, which controls the contribution of topological features in the discrimination. Achieve multi-scale feature weighted fusion.

[0051] Based on the above, step S3 includes step S303: using physical residual adversarial loss, embedding partial differential equation constraints in standard adversarial training, and forcing the generator to output a three-dimensional model that satisfies the physical conservation law. The specific steps are as follows:

[0052] (1) Define the physical residual operator as follows:

[0053] R(Y) = L c Yf

[0054] In the formula, R(·) represents the physical residual operator, which outputs the residual tensor; L c The differential operator describes the governing equations of the physical field, expressed as: Let be the Laplace operator, representing the second derivative in space; k is the physical coefficient; Here, f is the time differential operator, characterizing transient physical processes; f is the physical field source term.

[0055] (2) Construct physical residual adversarial loss, expressed as:

[0056]

[0057] In the formula, This represents the adversarial loss function for physical constraints, with the superscript phy indicating physical embedding. p is the expected value of the real data. data It is a true 3D model distribution; G is the expectation of the noise vector; θ (z) is the generator output; λ is the residual weight coefficient, which balances adversarial loss and physical constraint strength; ||·||2 is the L2 norm, which is used to calculate the Euclidean distance of the residual tensor.

[0058] Based on the above, step S3 includes step S304: using projective adaptive moment estimation, through parameter projection constraints and gradient orthogonalization decomposition, to enhance training stability while maintaining the advantages of Adam's adaptive learning rate. The specific steps are as follows:

[0059] (1) Calculate the adaptive moment estimate, expressed as:

[0060]

[0061] In the formula, m t This represents the first-order moment estimate at the current time, with the subscript t being the time step index; v t β1 represents the estimated second moment at the current time; β2 represents the exponential decay rate of the first moment; m represents the exponential decay rate of the second moment; m represents the second moment. t-1 This is an estimate of the first moment from the previous time step; v t-1 This is an estimate of the second moment from the previous moment; Let G be the gradient of the loss function with respect to the parameter θ, where θ is the generator G. θ Or discriminator D φ The parameters; (·) °2 This represents the element-wise squaring operation;

[0062] (2) The deviation correction moment estimate is expressed as:

[0063]

[0064] In the formula, This represents the first moment after deviation correction; This represents the second moment after deviation correction; It is β1 raised to the power of t; β² raised to the power of t; t is the current training iteration number;

[0065] (3) Projection parameter update, expressed as:

[0066]

[0067] In the formula, θ t Represents the updated parameters; η is the base learning rate; ∈ is the numerical stability constant; Π C (·) is the projection operator, denoted as C represents the stable constraint region;

[0068] (4) Perform gradient orthogonalization, expressed as:

[0069]

[0070] In the formula, d phy This indicates the direction of the physical residual gradient, and the subscript phy indicates the physical constraint. The gradient operator with respect to parameter θ is represented by ‖·‖; L represents the L2 norm; L represents the loss function; <·, ·> represent the vector inner product operation; ← represents the gradient overwrite operation, subtracting the physical gradient component from the original gradient; orthogonalization ensures... Eliminate optimization path conflicts.

[0071] This invention has outstanding substantive features and significant progress compared to the prior art, specifically:

[0072] 1. Tensor structure transformation combining learnable wavelet kernels and geometric regularization terms is adopted to preserve multi-scale spatial structure and nonlinear physical field characteristics;

[0073] 2. By separating physical laws and randomness through orthogonal constraints and Jacobi decoupling loss function, the physical rationality and generalization ability of the generative model are significantly improved.

[0074] 3. By integrating continuous cohomology features with multi-scale convolutional networks, we can achieve deep identification of topological properties (such as hole connectivity and surface genus) and local geometric defects of 3D models.

[0075] 4. The partial differential physical control equations are explicitly embedded to counteract the loss, and physical residual regularization is used to force the generated results to satisfy the physical conservation law. Attached Figure Description

[0076] Figure 1 This is a flowchart illustrating the process of this invention.

[0077] Figure 2 This is a comparative kernel density curve of the topological similarity distribution of the present invention.

[0078] Figure 3 This is a box plot comparing the physical residual norm of this invention.

[0079] Figure 4 This is a comparison curve of the structural similarity index of the present invention. Detailed Implementation

[0080] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0081] like Figure 1 As shown, a method for constructing a digital twin 3D model based on a generative large model is presented, and the main steps are as follows:

[0082] S1. Data Acquisition and Training Dataset Construction

[0083] To address the need for constructing digital twin 3D models of power distribution networks, this embodiment collects raw data of the power distribution network through a multi-source heterogeneous sensor network and a high-fidelity physical simulation system. Specifically, this includes: deploying temperature sensor arrays, strain sensor arrays, displacement sensor arrays, etc., at key points on the surface or inside the target entity (such as mechanical structures, fluid systems, etc.) to acquire real-time physical field distribution data of the target, including at least temperature field, stress field, and displacement field; acquiring electromagnetic field intensity and phase information through an electromagnetic probe array; and combining laser scanning and industrial CT equipment to acquire material property distribution and internal microstructure geometric data.

[0084] Simultaneously running physical simulations based on finite element analysis and computational fluid dynamics generates high-dimensional physical field data (velocity field, energy density field) covering different operating conditions (such as extreme loads and transient heat conduction). Sensor measured data and simulation data are spatiotemporally aligned and coordinate registered to construct a spatiotemporally synchronized multimodal dataset. After denoising, outlier removal, and missing value imputation of the original data, channel dimensions are divided according to physical field type, forming a four-dimensional training tensor (number of samples × spatial dimension × number of physical field channels).

[0085] Finally, the physical parameter vectors (such as Young's modulus of the material, boundary conditions, external load values, etc.) are labeled as conditional labels to generate the training dataset required for supervised learning.

[0086] S2. Tensor wavelet structured preprocessing of digital twin big data

[0087] Digital twin big data, such as physical field distributions and structural stress fields, exhibits high-dimensional sparsity, multi-scale spatial correlation, and non-Euclidean structural characteristics. Conventional processing methods, such as Min-Max normalization, destroy spatial topological relationships, and principal component analysis, due to the limitations of linear transformation, cannot capture local geometric features and nonlinear spatial patterns, resulting in processing results that cannot effectively preserve the structural characteristics of the physical field.

[0088] This invention employs tensor wavelet structured transform to map the original data into a multi-scale tensor quantum space. Spatial features are adaptively extracted using a learnable wavelet kernel, and geometric prior regularization terms are combined to maintain the continuity of the physical field structure. The specific steps are as follows:

[0089] 1) Constructing an adaptive sparse mask

[0090] Based on the original input four-dimensional training tensor, thresholding is performed using dynamic threshold parameters to highlight significant physical field features. The processed result is then input into the Sigmoid activation function to generate an adaptive sparse mask matrix, used to filter noise interference while preserving key physical field abrupt change regions, as shown below:

[0091] M spa =sig(X-τ)sp )

[0092] In the formula, M spa τ represents the adaptive sparse mask matrix used to filter noise; sig(·) is the Sigmoid activation function, which compresses the input to the [0,1] interval; X is the original input four-dimensional training tensor, which is the original physical field data in the digital twin scene. In this embodiment, it includes at least: physical field distribution data (such as temperature field, flow velocity field, electromagnetic field, etc.), structural mechanics data (such as stress field, strain field, displacement field, etc.), and other high-dimensional physical quantities (such as material property distribution, energy density field, etc.); sp This is a dynamic threshold parameter, adaptively calculated based on the data distribution; the subscript sp indicates sparsity control; X-τ sp Thresholding is implemented to highlight significant physical field characteristics.

[0093] 2) Perform multi-scale tensor wavelet transform

[0094] A learnable 3D convolution kernel is used to perform multi-scale convolution operations on the masked original input 4D training tensor to extract local spatial patterns. Then, features at different scales are concatenated and fused with a geometric curvature regularization term to output multi-scale structured features, enhancing spatial continuity. This is represented as follows:

[0095]

[0096] In the formula, J wav (·) denotes the tensor wavelet transform operator, J wav (X) represents multi-scale structured features; This indicates the implementation of multi-scale feature stitching operation; K is the preset total number of wavelet decomposition scales; W k The kernel is a learnable 3D convolution kernel, with the subscript k representing the scale index, which is adaptively optimized through training. Represents a 3D convolution operation to extract local spatial patterns; ⊙ represents element-wise multiplication, applying a sparse mask to filter noise; P geo (X) is the geometric curvature regularization term, calculated as follows:

[0097]

[0098] It is a second-order partial derivative tensor that describes the physical field in x. i ,x j Spatial curvature of direction; g m is the curvature basis vector, and the subscript m represents the index of the basis, used to fit the local surface geometry.

[0099] In one embodiment, topological similarity distribution analysis was performed to verify the ability of different methods to preserve the topological structure of 3D models (such as hole connectivity and surface genus). The topological similarity score was calculated by comparing the difference in Betti numbers between the generated model and the real model; a higher score indicates more complete preservation of topological attributes. The experiment collected the score distribution of 500 generated samples. The results are as follows: Figure 2 As shown, the kernel density curves indicate that: 1) the technique of this invention (red peak) exhibits a significant right-skewed distribution, with the main peak concentrated above 0.9; 2) it has the narrowest distribution width, with 90% of samples having scores greater than 0.85; and 3) there is no low-score tailing phenomenon. Other methods, however, all exhibit varying degrees of distribution defects: traditional principal component analysis (yellow peak) has the lowest main peak position and the most dispersed distribution; basic wavelet transform (green peak) has a secondary peak at 0.6-0.7; and physical constraint generative adversarial network (blue peak) shows a plateau region at 0.8. This demonstrates that the multi-scale topology-aware discriminator (S302) of this invention, through continuous homology feature analysis, can effectively identify and avoid geometric defects such as missing holes and non-manifold edges.

[0100] S3. Improved Generative Adversarial Network Construction and Training

[0101] S301, Generator Physical Constraint Embedding and Feature Decoupling

[0102] Conventional generative adversarial networks (GANs) directly learn the mapping from noise vectors to 3D models, but ignore physical constraints. This leads to generated models that do not conform to actual working conditions, such as stress singularities and material discontinuities. Traditional methods cannot separate the coupling effect between physical laws and random noise, resulting in generated results lacking physical plausibility. This invention designs a physically constrained decoupling generator. It constrains physically perceived features through orthogonal encoding and combines physical decoupling loss to forcibly separate physical parameters from random noise. The specific steps are as follows:

[0103] 1) Constructing a physical sensing encoder

[0104] The noise vector is input into the multilayer perceptron and mapped to the latent space. Then, it is transformed using an orthogonal constraint matrix to output physical constraint features, which force the preservation of the independence of physical laws. This is represented as:

[0105] Q phys (z)=MLP(z)·A orth

[0106] In the formula, Q phys (·) represents the physical sensing encoder, which outputs physical constraint features; MLP(z) is a multilayer perceptron that maps the noise vector z to the latent space; z is the noise vector; A orth This is an orthogonal constraint matrix, where the subscript "orth" indicates orthogonality and forces feature decoupling.

[0107] 2) Design of physical condition injection module

[0108] The physical parameter vectors are input into the learnable weight matrix and bias vector, and after linear transformation, processed by the hyperbolic tangent activation function. The output physical condition features are expressed as follows:

[0109] F con (C phy ) = tanh(W c C phy +b c )

[0110] In the formula, F con (·) indicates the physical condition injection module, F con (C phy ) represents physical condition characteristics; C phy This is a vector of physical parameters, where the subscript phy indicates a physical property; W c The subscript c represents the learnable weight matrix; b c is the bias vector; tanh(·) is the hyperbolic tangent activation function, which restricts the output to the interval [-1,1].

[0111] 3) Generate a decoupled 3D model

[0112] The physical constraint features and physical condition features are concatenated along the channel dimension, input into the deconvolution decoder for 3D deconvolution operation, and the generated 3D model tensor is output as follows:

[0113] Y gen =D dec (Q phys (z)||F con (C phy ))

[0114] In the formula, Y gen D is the tensor of the generated 3D model. dec (·) represents a deconvolutional decoder, which consists of three-dimensional deconvolutional layers; || represents a feature concatenation operation, which connects the input along the channel dimension.

[0115] 4) Physical decoupling loss constraint

[0116] Calculate the Jacobian matrix of the generated 3D model tensor with respect to the physical parameter vector, obtain its Frobenius norm, and combine it with the expected value of the physical field Jacobian matrix using a weighted combination. The weights are controlled by a decoupling strength coefficient to forcibly separate the coupling effect between physical parameters and noise, expressed as:

[0117]

[0118] In the formula, L disThis represents the physical decoupling loss function, where the subscript dis indicates decoupling; ||·|| F Given the Frobenius norm, calculate the square root of the sum of squares of all elements in the matrix; Y gen J is the tensor of the generated 3D model; phy γ is the Jacobian matrix of the physical field, with the subscript phy indicating the physical field; γ is the decoupling strength coefficient, which controls the weight of the physical law constraints; E[·] represents the expectation operator, which calculates the mean on the batch data.

[0119] S302, Define the discriminator and multi-scale topological sensing structure

[0120] Traditional discriminators using convolutional layers struggle to model the global topological structure of 3D models, such as hole connectivity and surface genus, and are insensitive to local geometric defects such as micro-cracks and non-manifold edges. Conventional convolutional neural networks primarily focus on local texture features and cannot effectively capture the overall topological characteristics of 3D models, which may result in geometric defects in the generated models.

[0121] This invention employs a multi-scale topology-aware discriminator, which enhances the ability to identify the global topological structure and local geometric defects of 3D models by fusing persistent coherence features and multi-scale convolutional features. The specific steps are as follows:

[0122] 1) Calculate the persistent homology feature

[0123] Perform persistent homology computation on the input 3D model tensor, analyze its topological invariant properties, and output a persistent homology feature set describing the geometric structure, denoted as S. topo =PH(Y);

[0124] S topo PH(·) represents the set of persistent homology features, describing the topological invariant properties of the model; PH(·) is the persistent homology computation operator, based on algebraic topology theory; Y is the input three-dimensional model tensor.

[0125] Continuous cohomology quantifies the topological structure of a model by analyzing the "birth and death process" of n-dimensional topological voids (such as 0-dimensional connected components, 1-dimensional holes, and 2-dimensional cavities).

[0126] 2) Construct a multi-scale convolutional feature extraction network

[0127] By employing multiple independent 3D convolutional sub-networks, features are extracted from the input 3D model tensor at different scales. Each sub-network outputs a feature vector corresponding to its scale.

[0128] This represents a 3D convolutional subnetwork at scale s, which outputs the feature vector at that scale; s is the scale index, s∈{1,2,…,S}, and S is the preset total number of scales; the superscript (s) indicates that the network parameters of different scale branches are independent.

[0129] 3) Calculate the adaptive scaling weights

[0130] The transpose of the learnable weight vector is applied to the feature vectors output by convolutional networks at each scale. After exponential function and normalization, adaptive weight coefficients at different scales are generated, as follows:

[0131]

[0132] In the formula, α s f represents the adaptive weights at the s-th scale; s w is the output feature vector of the s-th scale convolutional network; s The subscript s represents the scale index; exp(·) is the exponential function that performs softmax normalization. f is the transpose of the learnable weight vector at the j-th scale; j This represents the output feature vector of the j-th scale convolutional network. This represents the vector transpose operation.

[0133] 4) Evaluate topological similarity

[0134] Compare the persistently cohomological feature sets of the input model and the reference model, calculate the symmetric difference cardinality of their k-dimensional Betti number sets, and output the topological similarity score by the ratio of this cardinality to the cardinality of the reference model, expressed as:

[0135]

[0136] In the formula, N sim B represents the topological similarity score, with a value range of [0,1]. k (·) represents the set of k-dimensional Betti numbers, describing the number of k-dimensional topological holes; S topo S is the set of persistently homologous features of the input 3D model; ref The reference topological features are usually derived from real sample datasets; Δ represents the symmetric difference operator; |·| represents the cardinality of the set.

[0137] 5) Integrated discriminator output

[0138] The multi-scale convolutional features are summed using adaptive weight coefficients, and then combined with the topological similarity score in a weighted manner to output the discrimination result. The larger the value, the more likely the input is to be a real sample, as shown below:

[0139]

[0140] In the formula, D φ (·) represents the final output of the discriminator; the larger the value, the more likely it is a real sample. β is the topological similarity weight coefficient, which controls the contribution of topological features in the discrimination. Achieve multi-scale feature weighted fusion.

[0141] S303, Consistency Adversarial Training Based on Physical Residues

[0142] Conventional adversarial loss functions focus on matching data distributions, but cannot ensure that the generated model conforms to physical conservation laws such as mass conservation and energy balance. Therefore, the generated 3D model may violate basic physical laws in engineering applications, such as exhibiting non-equilibrium stress distributions and discontinuous displacement fields.

[0143] This invention employs physical residual adversarial loss and embeds partial differential equation constraints into standard adversarial training, forcing the generator to output a 3D model that satisfies physical conservation laws. The specific steps are as follows:

[0144] 1) Define the physical residual operator

[0145] Based on differential operators and physical field source terms, physical conservation law operations are performed on the input 3D model tensor to output a residual tensor that quantifies the deviation of physical laws, expressed as:

[0146] R(Y) = L c Yf

[0147] In the formula, R(·) represents the physical residual operator, which outputs the residual tensor; L c The differential operator describes the governing equations of the physical field, expressed as: Let be the Laplace operator, representing the second derivative in space; k is the physical coefficient; is the time differential operator, characterizing transient physical processes; f is the physical field source term.

[0148] 2) Construct physical residuals to counteract losses

[0149] Based on the standard adversarial loss, an L2 norm regularization term is added to the residual tensor of the generative model. This term balances the adversarial loss and the strength of physical constraints through residual weight coefficients, and is expressed as follows:

[0150]

[0151] In the formula, The adversarial loss function represents the physical constraints, and the superscript phy indicates the physical embedding. p is the expected value of the real data. data It is a true 3D model distribution; G is the expected value of the noise vector, and G is the noise distribution. θ (z) is the generator output; λ is the residual weight coefficient, which balances adversarial loss and physical constraint strength; ||·||2 is the L2 norm, which is used to calculate the Euclidean distance of the residual tensor.

[0152] S304, Parameter update strategy for dynamic stability optimization

[0153] Traditional Adam optimizers are prone to getting trapped in local optima in the parameter space of high-dimensional generative adversarial networks, and gradient oscillations during training can lead to pattern collapse, resulting in decreased generative diversity and increased violation of physical laws. Conventional optimization methods lack explicit constraints on the parameter stability domain, and there is coupling interference between the physical constraint gradient and the adversarial gradient.

[0154] This invention employs projective adaptive moment estimation, which enhances training stability while maintaining the advantages of Adam's adaptive learning rate through parameter projection constraints and gradient orthogonalization decomposition. The specific steps are as follows:

[0155] 1) Calculate the adaptive moment estimate

[0156] The historical first and second moments of the loss function gradient are used to perform an exponential moving average with respect to the exponential decay rate, and the current first and second moment estimates are updated as follows:

[0157]

[0158] In the formula, m t This represents the first-order moment estimate at the current time, with the subscript t being the time step index; v t This represents the estimate of the second moment at the current time; β1 is the exponential decay rate of the first moment, e.g., β1 = 0.9; β2 is the exponential decay rate of the second moment, e.g., β2 = 0.999; m t-1 This is an estimate of the first moment from the previous time step; v t-1 This is an estimate of the second moment from the previous moment; Let G be the gradient of the loss function with respect to the parameter θ, where θ is the generator G. θ Or discriminator D φ The parameters; (·) °2 This represents the element-wise squaring operation.

[0159] 2) Deviation correction moment estimation

[0160] By correcting the first-order moment estimate and second-order moment estimate at the current moment using a power-law correction of the exponential decay rate, the bias in the early stages of iteration is eliminated, as expressed in:

[0161]

[0162] In the formula, This represents the first moment after deviation correction; This represents the second moment after deviation correction; It is β1 raised to the power of t; β is 2 raised to the power of t; t is the current training iteration number.

[0163] 3) Projection parameter update

[0164] The parameter update is calculated based on the bias-corrected moment estimate and the basic learning rate. The updated parameters are then constrained to the Lipschitz stability region using the projection operator, as follows:

[0165]

[0166] In the formula, θ t Represents the updated parameters; η is the base learning rate; ∈ is the numerical stability constant, e.g., ∈ = 10 -8 ; Π C (·) is the projection operator, denoted as C represents the stable constraint region.

[0167] 4) Perform gradient orthogonalization

[0168] Calculate the L2 norm direction of the physical residual gradient, subtract its projected component in that direction from the original loss function gradient, and eliminate the directional conflict between the physical constraint gradient and the adversarial gradient, expressed as:

[0169]

[0170]

[0171] In the formula, d phy This indicates the direction of the physical residual gradient, and the subscript phy indicates the physical constraint. The gradient operator with respect to parameter θ is represented by ‖·‖; L represents the L2 norm; L represents the loss function; <·, ·> represent the vector inner product operation; ← represents the gradient overwrite operation, subtracting the physical gradient component from the original gradient; orthogonalization ensures... Eliminate optimization path conflicts.

[0172] In one embodiment, the physical residual norm is compared to evaluate the performance of different 3D modeling methods in satisfying physical conservation laws. The physical residual norm is an index obtained by calculating the deviation between the generated model and the physical governing equations (such as mass conservation and energy balance); the smaller the value, the more the model conforms to actual physical laws. Four methods were compared in the experiment: traditional principal component analysis, basic wavelet transform, physically constrained generative adversarial networks, and the technique of this invention. Figure 3As shown in the box plot results, the overall distribution of the physical residual norm of the present invention is significantly lower than that of other methods. Specifically: 1) the box position of the present invention is the lowest, and the median is close to the bottom of the coordinate axis; 2) the interquartile range (box height) is the smallest, indicating the best stability of the results; 3) the data points are concentrated in the range of 0.1-0.3, with no abnormally high values. In contrast, the traditional principal component analysis method has the widest residual distribution and the highest median. Although the basic wavelet transform and physical constraint generative adversarial network methods have improved this, there are still obvious outliers. This verifies the effectiveness of the present invention in the physical constraint embedding (S301) and physical residual adversarial training (S303) modules, ensuring that the generated model meets engineering requirements such as stress distribution balance and displacement field continuity.

[0173] S4, Model Reasoning Application

[0174] After the generator and discriminator complete adversarial training, 3D model inference is achieved through the following process:

[0175] First, input the physical parameter vector of the target scene (such as material properties and load conditions) and a random noise vector that conforms to a Gaussian distribution to the physical constraint decoupling generator.

[0176] The generator's physical perception encoder performs orthogonal constraint mapping on the noise vector and outputs decoupled physical constraint features. At the same time, the physical condition injection module generates condition features by nonlinearly transforming the physical parameter vector.

[0177] After being concatenated along the channel dimension, the two are input into the deconvolution decoder for multi-level three-dimensional deconvolution operations, which are then progressively upsampled to generate a high-resolution three-dimensional model tensor (containing geometric structure and embedded physical field distribution).

[0178] Subsequently, the discriminator's multi-scale topology-aware module performs continuous cohomology analysis on the generated model, calculates its Betti number to verify topological properties such as hole connectivity, and combines it with a multi-scale convolutional network to evaluate local geometric continuity (such as crack detection).

[0179] If the topological similarity score is lower than the preset threshold, the noise vector is adjusted and regenerated; if the verification is successful, the final 3D model tensor is output.

[0180] This tensor can be directly imported into the digital twin platform to drive real-time physical field visualization (such as stress cloud maps and temperature distribution) and provide high-fidelity input for subsequent working condition simulation prediction.

[0181] In one embodiment, the structural similarity index is analyzed over training epochs to examine the evolution trend of model accuracy generated by different methods during training. The structural similarity index is a key parameter measuring the geometric consistency between the generated model and the real 3D model, with a value range of 0-1 (1 indicating complete similarity). The experiment recorded the index changes of four methods over 200 training epochs. Figure 4 As shown in the curves, the results indicate that: 1) the initial value of the proposed technique (red curve) is higher than other methods, and it rapidly increases after 50 rounds; 2) it first enters a stable plateau period around 100 rounds, with the final value approaching 0.95; 3) it exhibits the smallest fluctuation range throughout the process. In contrast, the traditional principal component analysis method (yellow curve) shows slow growth and the lowest final value. Although the basic wavelet transform (green curve) and physically constrained generative adversarial networks (blue curve) show some improvement in the later stages, they still exhibit significant oscillations. Experimental results demonstrate that the tensor wavelet structured preprocessing (S2) and feature decoupling mechanism (S301) of this invention can efficiently extract multi-scale spatial features while maintaining geometric continuity, significantly accelerating model convergence.

[0182] 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 in all respects as exemplary and non-limiting, 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. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for constructing a digital-twin three-dimensional model based on a generative large model, characterized in that, The method comprises the steps of: S1, obtaining multi-source monitoring data of the power distribution network, and processing the multi-source monitoring data into a training data set; S2, mapping the multi-source monitoring data into a multi-scale tensor subspace through a tensor wavelet structured transformation, adaptively extracting spatial features through a learnable wavelet kernel, and combining a geometric prior regular term to maintain the continuity of the physical field structure; S3, constructing and training a generative adversarial network through the training data set, wherein the generator adopts a physical constraint decoupling generator, and the discriminator adopts a multi-scale topological perception discriminator; S4, inputting a physical parameter vector of a target scene into the generative adversarial network and performing analysis, if a topological similarity score is lower than a preset threshold, adjusting a noise vector to regenerate, if verified, outputting a final three-dimensional model tensor, and directly importing the three-dimensional model tensor into a digital twin platform to drive real-time physical field visualization.

2. The generative large model-based digital twin three-dimensional model construction method according to claim 1, characterized in that: In step S1, the multi-source monitoring data at least includes physical field distribution data, electromagnetic field intensity and phase information, material attribute distribution and internal microstructure geometry data, and high-dimensional physical field simulation data; the measured data and the simulation data are time-space aligned and coordinate registered to construct a multi-modal data set synchronized in time and space; after denoising, outlier elimination and missing value interpolation of the original obtained multi-modal data set, the data is divided into channel dimensions according to the physical field type to form a four-dimensional training tensor, and finally the physical parameter vector is labeled as a conditional label to generate a training data set required for supervised learning.

3. The generative large model-based digital twin three-dimensional model construction method of claim 1, wherein: In step S2, based on the original input four-dimensional training tensor, an adaptive sparse mask matrix is generated: M spa = sig(X-τ sp ) In the formula, M spa denotes an adaptive sparse mask matrix for filtering noise; sig(·) is a Sigmoid activation function that compresses the input to the interval [0, 1]; X is a four-dimensional training tensor of the original input; τ sp is a dynamic threshold parameter, and the subscript sp represents sparsity control; X-τ sp achieves thresholding processing; The original input four-dimensional training tensor after mask processing is subjected to multi-scale convolution operation by using a learnable three-dimensional convolution kernel, then the features of different scales are spliced and fused with a geometric curvature regular term, and a multi-scale structured feature is output, which is represented as: where J wav (·) denotes the tensor wavelet transform operator, J wav (X) denotes the multi-scale structured features; denotes the multi-scale feature concatenation operation; K is the total number of preset wavelet decomposition scales; W k is the learnable 3D convolution kernel, subscript k denotes the scale index, which is adaptively optimized by training; denotes the 3D convolution operation, which extracts local spatial patterns; ⊙ denotes element-wise multiplication, which applies sparse mask to filter noise; P geo (X) is the geometric curvature regularization term, and the calculation method is represented as is the second-order partial derivative tensor, which describes the spatial curvature of the physical field in the x i ,x j direction; g m is the curvature basis vector, subscript m denotes the index of the basis, which is used to fit the local surface geometric characteristics.

4. The generative large model-based digital twin three-dimensional model construction method of claim 1, wherein, In step S3, step S301 is included: the physical constraint decoupling generator constrains the physical perception features through orthogonal encoding, and combines a physical decoupling loss to force the separation of physical parameters and random noise, and the specific steps are as follows: (1) constructing a physical perception encoder, represented as: Q phys (z) = MLP(z) · A orth wherein Q phys (·) denotes a physical perceptual encoder, outputting physical constraint features; MLP(z) is a multi-layer perceptron mapping a noise vector z to a latent space; z is a noise vector; A orth is an orthogonal constraint matrix, with the subscript orth denoting orthogonality; (2) designing a physical condition injection module to output physical condition features, represented as: F con (C phy )=tanh(W c C phy +b c ) where F con (·) denotes the physical condition injection module, F con (C phy ) denotes the physical condition feature; C phy is the physical parameter vector, subscript phy denotes the physical property; W c is the learnable weight matrix, subscript c denotes the condition; b c is the bias vector; tanh(·) is the hyperbolic tangent activation function, which limits the output to the interval [-1, 1]; (3) generating a decoupled three-dimensional model to output a generated three-dimensional model tensor, represented as: Y gen = D dec (Q phys (z) || F con (C phy )) In the formula, Y gen is a generated three-dimensional model tensor; D dec (·) represents a deconvolution decoder; || represents a feature concatenation operation, connecting the input along the channel dimension; (4) physical decoupling loss constraint, represented as: where L dis represents the physical decoupling loss function, subscript dis represents decoupling; ||·||F F is the Frobenius norm, which calculates the square root of the sum of squares of all elements of a matrix; Y gen is the generated three-dimensional model tensor; J phy is the physical field Jacobian matrix, subscript phy represents the physical field; γ is a decoupling strength coefficient, which controls the weight of the physical law constraint; E[·] represents an expectation operator, which calculates the mean on the batch data.

5. The generative large model-based digital twin three-dimensional model construction method of claim 1, wherein, In step S3, step S302 is included: the multi-scale topological perception discriminator enhances the recognition ability of the global topological structure and local geometric defects of the three-dimensional model by fusing persistent homology features and multi-scale convolution features, and the specific steps are as follows: (1) calculating persistent homology features, represented as: S topo = PH(Y); where S topo denotes the persistent homology feature set; PH(·) is the persistent homology computation operator; Y is the input three-dimensional model tensor; (2) constructing a multi-scale convolution feature extraction network, represented as: denotes the s-th scale three-dimensional convolution subnetwork, and outputs a feature vector at this scale; s is a scale index, s e {1, 2, …, S}, S is a preset total scale number; the superscript (s) indicates that the network parameters of different scale branches are independent; (3) calculating adaptive scale weights, represented as: where α s denotes the adaptive weight of the s-th scale; f s is the output feature vector of the s-th scale convolutional network; w s is the learnable weight vector, and subscript s denotes the scale index; exp(·) is the exponential function, which realizes the softmax normalization; is the transpose of the learnable weight vector of the j-th scale; f j is the output feature vector of the j-th scale convolutional network; denotes the vector transpose operation; (4) evaluating topological similarity, represented as: where N sim denotes the topological similarity score, with value range [0, 1]; B k is the k-dimensional Betti number set, describing the number of k-dimensional topological holes; S topo is the persistent homology feature set of the input 3D model; S ref is the reference topological feature, usually from a real sample dataset; Δ denotes the symmetric difference operator; |·| denotes the cardinality of the set; (5) integrating discriminator output, represented as: In the formula, D φ (·) represents the final output of the discriminator; β is a topological similarity weight coefficient, which controls the contribution of topological features in the judgment; Realize multi-scale feature weighted fusion.

6. The generative large model-based digital twin three-dimensional model building method according to claim 1, characterized in that, In step S3, step S303 is included: a physical residual adversarial loss is used to embed a partial differential equation constraint in standard adversarial training to force the generator to output a three-dimensional model that satisfies the physical conservation law, and the specific steps are as follows: (1) defining a physical residual operator, represented as: R(Y) = L c Y-f where R(·) represents a physical residual operator, outputting a residual tensor; L c is a differential operator, describing a physical field governing equation, denoted as is a Laplacian operator, representing spatial second-order derivative; k is a physical coefficient; is a time differential operator, representing a transient physical process; f is a physical field source term; (2) constructing a physical residual adversarial loss, represented as: where, is the adversarial loss function representing physical constraints, the superscript phy denotes the physical embedding; is the real data expectation, p data is the real three-dimensional model distribution; is the noise vector expectation; G θ (z) is the generator output; λ is the residual weight coefficient balancing the adversarial loss and the strength of the physical constraints; ||·||2 is the L2 norm, calculating the Euclidean distance of the residual tensor.

7. The generative large model-based digital twin three-dimensional model building method according to claim 1, characterized in that, The step S3 comprises a step S304: using projection adaptive moment estimation, through parameter projection constraint and gradient orthogonalization decomposition, the training stability is enhanced while the advantages of Adam adaptive learning rate are maintained, and the specific steps are as follows: (1) calculate the adaptive moment estimation, denoted as: where m t denotes the first moment estimate at the current time instant, and subscript t is the time step index; v t denotes the second moment estimate at the current time instant; β1 is the exponential decay rate of the first moment; β2 is the exponential decay rate of the second moment; m t-1 is the first moment estimate at the previous time instant; v t-1 is the second moment estimate at the previous time instant; is the gradient of the loss function with respect to the parameters θ of the generator G θ or the discriminator D φ ; (·) °2 denotes the element-wise square operation; (2) bias correction moment estimation, denoted as: wherein denotes the first moment after bias correction; denotes the second moment after bias correction; is the t-th power of β1; is the t-th power of β2; t is the current number of training iterations; (3) projection parameter update, denoted as: where θ t denotes the updated parameters; η is the base learning rate; ∈ is a numerical stability constant; Π C (·) is the projection operator, denoted as C denotes the stable constraint domain; (4) perform gradient orthogonalization, denoted as: where d phy denotes the physical residual gradient direction, and the subscript phy denotes the physical constraint; denotes the gradient operator with respect to the parameter θ; || · || denotes the L2 norm; L denotes the loss function; <·, ·> denotes the vector inner product operation; ← denotes the gradient overwrite operation, subtracting the physical gradient component from the original gradient; and the orthogonalization ensures eliminates the optimization path conflict.

Images