A multi-physics digital twin method and system based on spatially selective Bayesian inference and a medium

Abstract

The application discloses a multi-physical field digital twin method and system based on spatial selection Bayesian inference and a medium, mainly relates to the technical field of multi-physical field digital twin, and is used to solve the problem that the existing scheme conflicts between calculation efficiency and physical fidelity. It includes: determining the physical probe range, obtaining each node in the physical probe range; based on the predicted temperature field, extracting the node predicted temperature of each node in the physical probe range, and then calculating the mean value as the final prediction value of the sensor; obtaining the actual observation value of each sensor; based on the actual observation value and the final prediction value, calculating the log-likelihood function value and the joint log-posterior probability of the current global operation proposal parameter value; using a multi-chain parallel MCMC algorithm, generating a new global operation proposal parameter value based on the current global operation proposal parameter value; based on the joint log-posterior probability, judging whether to accept the updated global operation proposal parameter value.

Description

Technical Field

[0001] This application relates to the field of multiphysics digital twin technology, and in particular to a multiphysics digital twin method, system and medium based on spatially selective graph Bayesian inference. Background Technology

[0002] Currently, intelligent operation and maintenance of core power equipment (such as ultra-high voltage converter transformers) relies on high-fidelity heat flux multiphysics digital twin models to achieve real-time response to transient operating conditions. Due to limitations imposed by the high-voltage insulation structure and strong electromagnetic environment within the equipment, actual operation and maintenance can only obtain sparse boundary observation data such as top-level oil temperature and tank surface temperature. This data is then used to infer the internal temperature field distribution and heat source parameters through the inverse heat transfer problem (IHTP). To quantify uncertainty, research commonly employs Bayesian statistical frameworks (such as MCMC sampling) to fuse physical priors and observational information, constructing a parameter inversion system.

[0003] However, existing methods suffer from a fundamental conflict between computational efficiency and physical fidelity. Traditional forward simulations based on CFD or FEM take minutes to hours per run, and MCMC requires thousands to tens of thousands of simulations to ensure convergence, making the inversion process completely infeasible in engineering scenarios. Although data-driven surrogate models (such as CNN and DeepONet) attempt to improve inference speed, their globally coupled architecture forces each iteration to perform a full-field forward computation involving millions of nodes, just to obtain predictions for a few sensor locations. This results in significant computational redundancy and wasted GPU memory, severely limiting real-time performance and scalability. Summary of the Invention

[0004] This application provides a multiphysics digital twin method, system, and medium based on spatially selective graphical Bayesian inference to address the conflict between computational efficiency and physical fidelity in existing solutions.

[0005] Firstly, this application provides a multiphysics digital twin method based on spatially selective graph Bayesian inference, the method comprising: Obtain the global grid data and corresponding sensor spatial coordinate set of the target device; generate a local inference subgraph for each sensor position in the sensor spatial coordinate set; During the inversion iteration, the global operation proposal parameter value of the current inversion iteration is obtained, and the predicted temperature field of each node in the local inference subgraph under the current global operation proposal parameter value is output using a pre-trained graph neural network model. In the local inference subgraph, the area within a preset distance centered on the sensor location is defined as the physical probe range. Each node within the physical probe range is obtained. Based on the predicted temperature field, the node predicted temperature of each node within the physical probe range is extracted, and then the average value is calculated as the final predicted value of the sensor. Obtain the actual observation values ​​of each sensor; based on the actual observation values ​​and the final predicted values, calculate the log-likelihood function value and joint log-posterior probability of the current global operation proposal parameter values; A multi-chain parallel MCMC algorithm is adopted to generate new global operation proposal parameter values ​​based on the current global operation proposal parameter values; whether to accept the updated global operation proposal parameter values ​​is determined based on the joint logarithmic posterior probability; the global operation proposal parameter values ​​corresponding to each sensor are iteratively updated repeatedly until the MCMC algorithm reaches the posterior distribution state.

[0006] In one implementation of this application, the global grid data of the target device and the corresponding sensor spatial coordinate set are obtained; for each sensor position in the sensor spatial coordinate set, a local inference subgraph is generated, specifically including: Acquire global mesh data of the target device; wherein, the global mesh data includes node spatial coordinates, initial physical characteristics of nodes and edge connection topology; and the nodes contain sparsely distributed sensors corresponding to the target device; Determine the sensor spatial coordinate set based on node spatial coordinates. ;in, The m-th sensor represents its spatial coordinates; M represents the total number of sensors. Get the cutoff radius and the radius of the perceptual domain ,and ; Using the sensor coordinates as the center, a spatial indexing algorithm is used to calculate the Euclidean distance from global nodes to the center point, and nodes with distances less than or equal to... The set of nodes is used as the nodes for generating the corresponding local reasoning subgraph; The node features, edge indices, and edge features corresponding to the nodes in the set of nodes corresponding to the current sensor coordinates will be extracted from the global grid data and encapsulated into a local inference subgraph corresponding to the current sensor coordinates.

[0007] In one implementation of this application, during the inversion iteration process, the global operation proposal parameter value of the current inversion iteration is obtained, and a pre-trained graph neural network model is used to output the predicted temperature field of each node in the local inference subgraph under the current global operation proposal parameter value, specifically including: Obtain the global operation proposal parameter value in the current MCMC iteration; after normalizing the global operation proposal parameter value, use it as a global feature vector and explicitly concatenate it along the feature dimension to inject it into the feature vector of each node in the local subgraph to obtain the local inference subgraph after injecting the global operation proposal parameter value. After injecting the global operation proposal parameter values, the local inference subgraph is input into the pre-trained graph neural network model with frozen parameters, and the predicted temperature field of each node in the local inference subgraph is output under the current global operation proposal parameter values.

[0008] In one implementation of this application, before extracting the node-predicted temperatures of each node within the range of the physical probe based on the predicted temperature field, and then calculating the average value as the final predicted value of the sensor, the method further includes: In the node set corresponding to the current sensor local inference subgraph, find the nodes within the physical probe range and generate a Boolean type probe sensing mask; Based on the predicted temperature field, the predicted node temperature of each node within the range of the physical probe is extracted, and then the average value is calculated as the final predicted value of the sensor. Specifically, this includes: Extract the node-predicted temperature of the node corresponding to the probe sensing mask from the predicted temperature field; The spatial mean of the predicted temperature at the computing nodes is used as the final predicted value of the sensor.

[0009] In one implementation of this application, the log-likelihood function value and joint log-posterior probability of the current global operation proposal parameter value are calculated based on the actual observed value and the final predicted value, specifically including: Based on the actual observations and final predictions obtained from each sensor, a Gaussian noise observation model for the sensor is constructed, and the prediction residual for each sensor is calculated. Through the formula: Calculate the log-likelihood function value ; Where M represents the total number of sensors. Indicates the preset standard deviation. Indicates the predicted residual; Obtain the prior probability density of global operation parameters Using the formula: Obtain the joint logarithmic posterior probability .

[0010] In one implementation of this application, a multi-chain parallel MCMC algorithm is used to generate new global operation proposal parameter values ​​based on the current global operation proposal parameter values; and a decision on whether to accept the updated global operation proposal parameter values ​​is made based on the joint logarithmic posterior probability, specifically including: In multi-chain parallel MCMC inversion sampling, the current global operation proposal parameter state is maintained for each independent Markov chain. In each iteration, each chain will perform a random walk within the boundary of its prior parameter space based on its current parameter value, according to the preset proposal distribution and the specified step size, to generate a new global operation proposal parameter for the chain. Calculate the new joint log-posterior probability based on the new global operation proposal parameters; If the difference between the new joint log-posterior probability and the previous joint log-posterior probability is greater than or equal to 0, the global operation proposal parameter value is accepted; otherwise, the global operation proposal parameter value is not accepted.

[0011] In one implementation of this application, the global operation proposal parameter value corresponding to each sensor is iteratively updated repeatedly until the MCMC algorithm reaches the posterior distribution state, specifically including: After each parameter update, the global operation proposal parameter value of the current chain is stored in the history sequence, forming a trajectory of parameter changes with the number of iteration steps; Calculate the between-group and within-group variances of the mean parameters of all parallel chains in the most recent N steps, and determine the convergence of the posterior distribution based on the Gelman-Rubin statistic. When the between-group variance, within-group variance, and posterior distribution convergence meet the preset convergence conditions, the MCMC algorithm is considered to have reached the posterior distribution state.

[0012] Secondly, this application provides a multiphysics digital twin system based on spatially selective graph Bayesian inference, the system comprising: The generation module is used to acquire the global grid data of the target device and the corresponding sensor spatial coordinate set; for each sensor position in the sensor spatial coordinate set, a local inference subgraph is generated. The output module is used to obtain the global operation proposal parameter value of the current inversion iteration during the inversion iteration process, and use a pre-trained graph neural network model to output the predicted temperature field of each node in the local inference subgraph under the current global operation proposal parameter value. The calculation module is used to determine the physical probe range within a preset distance centered on the sensor location in the local inference subgraph, and to acquire the nodes within the physical probe range; based on the predicted temperature field, to extract the node predicted temperature of each node within the physical probe range, and then to calculate the mean as the final predicted value of the sensor; to acquire the actual observation value of each sensor; and based on the actual observation value and the final predicted value, to calculate the log-likelihood function value and joint log-posterior probability of the current global operation proposal parameter value. The iterative module is used to generate new global operation proposal parameter values ​​based on the current global operation proposal parameter values ​​using a multi-chain parallel MCMC algorithm; determine whether to accept the updated global operation proposal parameter values ​​based on the joint logarithmic posterior probability; and repeatedly iterate and update the global operation proposal parameter values ​​corresponding to each sensor until the MCMC algorithm reaches the posterior distribution state.

[0013] In one implementation of this application, the generation module includes a generation unit. Used to acquire global mesh data of the target device; wherein, the global mesh data includes node spatial coordinates, initial physical characteristics of nodes and edge connection topology; and the nodes contain sparsely distributed sensors corresponding to the target device; Determine the sensor spatial coordinate set based on node spatial coordinates. ;in, The m-th sensor represents its spatial coordinates; M represents the total number of sensors. Get the cutoff radius and the radius of the perceptual domain ,and ; Using the sensor coordinates as the center, a spatial indexing algorithm is used to calculate the Euclidean distance from global nodes to the center point, and nodes with distances less than or equal to... The set of nodes is used as the nodes for generating the corresponding local reasoning subgraph; The node features, edge indices, and edge features corresponding to the nodes in the set of nodes corresponding to the current sensor coordinates will be extracted from the global grid data and encapsulated into a local inference subgraph corresponding to the current sensor coordinates.

[0014] Thirdly, this application provides a non-volatile computer storage medium storing computer instructions, which, when executed, implement a multiphysics digital twin method based on spatially selective graphical Bayesian inference as described above.

[0015] As can be seen from the above technical solutions, this application has the following advantages: By constructing a local inference subgraph independently for each sensor location, the redundant mode of traditional data-driven models, which require full forward computation across millions of global grid nodes, is eliminated. During the inversion iteration process, the graph neural network only needs to predict the temperature field within a preset range of local subgraphs, rather than traversing the entire physical domain, thus compressing the computational node scale of a single inference from millions to less than a thousand. This design directly eliminates the memory explosion and wasted computing power caused by being forced to perform full-field simulations due to "only a few sensor values ​​needed," making the computational load linearly related to the number of sensors, rather than exponentially related to the global grid size, fundamentally achieving precise allocation and efficient utilization of computing resources.

[0016] A spatially selective graphical Bayesian inference framework is adopted, decomposing global parameter updates into local posterior probability calculation tasks corresponding to multiple independent sensors, and combining this with a multi-chain parallel MCMC algorithm for synchronous iteration. Each MCMC chain calculates the local log-likelihood and joint posterior probability based only on the predicted values ​​and actual observation values ​​of nodes within the physical probe range of its corresponding sensor, avoiding parameter coupling oscillations and convergence delays caused by global coupling in traditional methods. This mechanism reduces the computational complexity of each iteration from O(N²) to O(M), where M is the number of local probe nodes and N is the number of global nodes, shortening the time consumed in a single iteration.

[0017] In the local inference subgraph, a physical probe range centered on the sensor and bounded by a preset distance is defined, ensuring that predicted values ​​originate only from neighboring nodes closely coupled to the sensor's physical location, thus preserving the local physical consistency of multi-physics coupling. Simultaneously, the local subgraphs of each sensor are independent of the MCMC chain, and parameter updates do not interfere with each other, achieving distributed decoupling of the global parameter space. This structure allows the system to flexibly add or remove sensors without retraining or reconstructing the global model, and does not cause a non-linear increase in computational load due to the addition of new sensors. Attached Figure Description

[0018] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a multiphysics digital twin method based on spatially selective graph Bayesian inference provided in an embodiment of this application.

[0020] Figure 2 This is a schematic diagram of a multi-scale spatial truncation and asymmetric observation model provided in an embodiment of this application.

[0021] Figure 3 This is a schematic diagram of the internal structure of a multiphysics digital twin system based on spatially selective graph Bayesian inference, provided in an embodiment of this application. Detailed Implementation

[0022] 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.

[0023] Those skilled in the art should understand that the embodiments described below are merely preferred embodiments of this disclosure and do not imply that this disclosure can only be implemented through these preferred embodiments. These preferred embodiments are merely used to explain the technical principles of this disclosure and are not intended to limit the scope of protection of this disclosure. Based on the preferred embodiments provided by this disclosure, all other embodiments obtained by those skilled in the art without creative effort should still fall within the scope of protection of this disclosure.

[0024] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0025] To address the issues of global coupling rigidity and computational resource waste inherent in existing surrogate models when handling sparse observation inverse problems, this application proposes a multiphysics digital twin method, system, and medium based on spatially selective graph Bayesian inference. This method uses a pre-trained graph neural network (GNN) as the physics surrogate solver, leveraging its "topological independence" during the inference phase to dynamically construct a lightweight local subgraph focused on the sensor's sensing domain. By performing high-frequency MCMC sampling optimization on the local subgraph and injecting the parameters into the global graph after convergence, efficient inversion and full-field twin reconstruction of multiphysics with millions of nodes are achieved.

[0026] The core technical solution of this application includes: A spatially selective subgraph inference mechanism that breaks global coupling: This mechanism fully leverages the topological independence and inductive bias characteristics of graph neural networks during the inference phase. After the model parameters are frozen, information transmission is strictly limited to local neighborhoods, and the physical states of spatially non-overlapping sub-regions are conditionally independent under given global parameters. Based on this, this application abandons the traditional "full-field forward computation—sparse point sampling" model and instead adopts a "computation follows observation" paradigm, extracting local subgraphs for inference only within the effective sensing domain of the sensor.

[0027] Global parameter injection and MCMC sampling based on asymmetric observation scale: In each MCMC iteration, the global operating parameters to be inverted (such as heat source power, ambient temperature, etc.) are explicitly injected as feature vectors into each node of the local subgraph, ensuring the global physical consistency of local inference. Simultaneously, when calculating the likelihood function, the predicted mean of nodes within the probe's sensing radius within the subgraph is extracted and compared with noisy observation data, effectively smoothing the numerical distortion of single-point grids. This strategy significantly reduces the computational complexity of a single iteration from the global scale to the local scale, completely eliminating the computational barrier to high-frequency MCMC iterations.

[0028] Low-cost, million-node full-field digital twin reconstruction: After MCMC sampling converges and the posterior statistical features of the parameters are obtained, the optimal parameter mean is extracted and re-injected into a global graph structure covering the entire computational domain. By performing full-field parallel inference through a single call to the graph neural network, a high-fidelity distribution of temperature and pressure containing all nodes (on a million-node scale) is instantly generated, achieving a leap from "sparse perception" to "full-field twin delivery".

[0029] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0030] The embodiment provides a multiphysics digital twin method based on spatially selective graphical Bayesian inference, such as Figure 1 As shown in the embodiments of this application, the method mainly includes the following steps: Step 110: Obtain the global grid data of the target device and the corresponding sensor spatial coordinate set; generate a local inference subgraph for each sensor position in the sensor spatial coordinate set.

[0031] In some embodiments, this step may specifically be: Acquire global mesh data of the target device; wherein, the global mesh data includes node spatial coordinates, initial physical characteristics of nodes and edge connection topology; and the nodes contain sparsely distributed sensors corresponding to the target device; Determine the sensor spatial coordinate set based on node spatial coordinates. ;in, The m-th sensor represents its spatial coordinates; M represents the total number of sensors. Get the cutoff radius and the radius of the perceptual domain ,and ; Using the sensor coordinates as the center, a spatial indexing algorithm is used to calculate the Euclidean distance from global nodes to the center point, and nodes with distances less than or equal to... The set of nodes is used to extract the internal edge connections of the set in the global grid. If there are isolated nodes that are not connected, only the largest connected subgraph is retained to ensure the continuity of physical heat transfer, and these nodes are used to generate the corresponding local inference subgraph. The node features, edge indices, and edge features corresponding to the nodes in the set of nodes corresponding to the current sensor coordinates will be extracted from the global grid data and encapsulated into a local inference subgraph corresponding to the current sensor coordinates.

[0032] It should be noted that, in the extracted node set, nodes with a distance less than or equal to the center of the sensor should be identified. The nodes generate a Boolean-type sensor mask for subsequent region mean aggregation.

[0033] Step 120: During the inversion iteration, obtain the global operation proposal parameter value of the current inversion iteration, and use the pre-trained graph neural network model to output the predicted temperature field of each node in the local inference subgraph under the current global operation proposal parameter value.

[0034] It should be noted that during the inversion iteration process, a pre-trained graph neural network model (as a black-box physics solver) is used to perform efficient reasoning on the local subgraph.

[0035] In some embodiments, this step may specifically be: Obtain the global operation proposal parameter value in the current MCMC iteration; after normalizing the global operation proposal parameter value, use it as a global feature vector and explicitly concatenate it along the feature dimension to inject it into the feature vector of each node in the local subgraph to obtain the local inference subgraph after injecting the global operation proposal parameter value. After injecting the global operation proposal parameter values, the local inference subgraph is input into the pre-trained graph neural network model with frozen parameters, and the predicted temperature field of each node in the local inference subgraph is output under the current global operation proposal parameter values.

[0036] It should be noted that in the first In the next MCMC iteration, the proposed values ​​of the global operation parameters to be inverted are received. (Such as grid-side heat source power, valve-side heat source power, ambient temperature, etc.). For After normalization, it is used as a global feature vector and explicitly concatenated along the feature dimension and injected into the local subgraph. In the feature vector of each node, enhanced node features are formed. This step ensures that spatially disconnected local subgraphs can share the same set of macroscopic physical boundary conditions. The local subgraphs injected with global parameters are then... The input is fed into a pre-trained graph neural network model whose parameters are frozen. Since message passing during the inference phase relies only on the local neighborhood topology, the model can directly output the values ​​of each node within the subgraph at the current parameters. The predicted temperature field.

[0037] To further clarify, the pre-trained graph neural network model's pre-training process constructs a graph structure based on a physical domain discretized from a finite element mesh. Nodes correspond to mesh vertices, and edges are defined by spatial adjacency relationships. Input features include node coordinates, initial temperature field, and boundary conditions. Training data consists of labeled global temperature sequences generated from CFD simulations. The graph neural network employs a multi-layer graph convolutional architecture, aggregating heat conduction information from neighboring nodes through message passing, and outputting predicted temperature values ​​for each node within a local subgraph. The loss function uses node-level L2 error, and end-to-end training is performed on a GPU using the Adam optimizer.

[0038] Physical constraints based on the heat conduction equation are embedded during pre-training. The spatiotemporal partial derivatives of the network output are calculated through automatic differentiation, and the PDE residuals are added as regularization terms to the loss function, forcing the predicted field to satisfy energy conservation and heat diffusion laws. This mechanism enables the model to maintain physical consistency in sensorless data regions, improves its generalization ability to unseen parameter combinations, and allows for unsupervised pre-training using massive amounts of simulation data without additional annotations.

[0039] Step 130: In the local inference subgraph, the area within a preset distance centered on the sensor location is the physical probe range. Obtain the nodes within the physical probe range. Based on the predicted temperature field, extract the node predicted temperature of each node within the physical probe range, and then calculate the average value as the final predicted value of the sensor.

[0040] Before extracting the node-predicted temperature of each node within the physical probe range based on the predicted temperature field, and then calculating the average value as the final predicted value of the sensor, the method also includes: In the node set corresponding to the current sensor local inference subgraph, find the nodes within the physical probe range and generate a Boolean type probe sensing mask; Specifically, based on the predicted temperature field, the predicted node temperature of each node within the range of the physical probe is extracted, and then the average value is calculated as the final predicted value of the sensor. This can be done as follows: Extract the node-predicted temperature of the node corresponding to the probe sensing mask from the predicted temperature field; The spatial mean of the predicted temperature at the computing nodes is used as the final predicted value of the sensor.

[0041] It should be noted that the generated sensor mask is used to extract the predicted temperature of nodes within the range of the physical probe, and its spatial mean is calculated as the final predicted value Tpred(m) of the sensor. This region aggregation operation effectively smooths out the discretization numerical error of the single-point grid.

[0042] Step 140: Obtain the actual observation values ​​of each sensor; based on the actual observation values ​​and the final predicted values, calculate the log-likelihood function value and the joint log-posterior probability of the current global operation proposal parameter values.

[0043] In some embodiments, the actual observation data vector Yobs of each sensor is obtained. Assuming that the observation data is disturbed by Gaussian white noise with a standard deviation of σ, the log-likelihood function and joint log-posterior probability of the current proposed parameter θ(k) are calculated based on the predicted value Tpred output by the local subgraph.

[0044] This step can be specifically described as follows: Based on the actual observations and final predictions obtained from each sensor, a Gaussian noise observation model for the sensor is constructed, and the prediction residual for each sensor is calculated. Through the formula: Calculate the log-likelihood function value ; Where M represents the total number of sensors. Indicates the preset standard deviation. Indicates the predicted residual; Obtain the prior probability density of global operation parameters Using the formula: Obtain the joint logarithmic posterior probability .

[0045] Step 150: Using the multi-chain parallel MCMC algorithm, generate new global operation proposal parameter values ​​based on the current global operation proposal parameter values; determine whether to accept the updated global operation proposal parameter values ​​based on the joint logarithmic posterior probability; iteratively update the global operation proposal parameter values ​​corresponding to each sensor until the MCMC algorithm reaches the posterior distribution state.

[0046] Specifically, a multi-chain parallel MCMC algorithm is employed to generate new global operation proposal parameter values ​​based on the current global operation proposal parameter values. The decision to accept the updated global operation proposal parameter values ​​is based on the joint logarithmic posterior probability, which can be achieved as follows: In multi-chain parallel MCMC inversion sampling, the current global operation proposal parameter state is maintained for each independent Markov chain. In each iteration, each chain will perform a random walk within the boundary of its prior parameter space based on its current parameter value, according to the preset proposal distribution and the specified step size, to generate a new global operation proposal parameter for the chain. Calculate the new joint log-posterior probability based on the new global operation proposal parameters; If the difference between the new joint log-posterior probability and the previous joint log-posterior probability is greater than or equal to 0, the global operation proposal parameter value is accepted; otherwise, the global operation proposal parameter value is not accepted.

[0047] Specifically, the global operation proposal parameter value for each sensor is iteratively updated repeatedly until the MCMC algorithm reaches the posterior distribution state. This can be achieved as follows: After each parameter update, the global operation proposal parameter value of the current chain is stored in the history sequence, forming a trajectory of parameter changes with the number of iteration steps; Calculate the between-group and within-group variances of the mean parameters of all parallel chains in the most recent N steps, and determine the convergence of the posterior distribution based on the Gelman-Rubin statistic. When the between-group variance, within-group variance, and posterior distribution convergence meet the preset convergence conditions, the MCMC algorithm is considered to have reached the posterior distribution state.

[0048] After MCMC sampling converges, the parameters obtained from local inversion are used to drive high-fidelity physical reconstruction of the entire macroscopic field: Posterior statistical feature extraction: Samples from the initial warm-up period of MCMC sampling are removed. Statistics are then performed on the stationary sampling chain to extract the posterior mean of the parameters to be inverted. And the 95% confidence interval (95% CI) that reflects uncertainty.

[0049] Full-field twin reconstruction: using the optimal posterior mean It is re-injected into a global graph structure containing all nodes of the device (such as a scale of millions of nodes). A single call to the pre-trained graph neural network model performs parallel inference across the entire field, instantly outputting the temperature field, fluid pressure field, and velocity field distribution of all nodes inside the device, completing the reconstruction and delivery of a high-fidelity full-field digital twin from sparse local perception.

[0050] Based on the above description, such as Figure 2 As shown, this application can be specifically described as follows: Phase 1 involves spatially selective subgraph decoupling. It inputs a global grid of millions of target devices and sparse sensor coordinates, sets multi-scale cutoff radii, extracts local inference subgraphs corresponding to each sensor using spatial indexing, generates a receptive field mask, and transfers the subgraphs to GPU memory, thus decoupling global computation from local computation. Phase 2 is high-frequency, low-cost MCMC iteration. It first proposes global operating parameters and injects them into the local subgraph, calls a pre-trained GNN black-box model to complete local inference, aggregates the receptive field mean and compares it with actual observation data to complete likelihood calculation, and uses the MH criterion to determine whether to accept parameter updates. This iteration is repeated until the iteration limit is reached. Phase 3 is million-level full-field digital twin reconstruction. After MCMC convergence, it extracts the optimal posterior mean of the parameters, injects it into the million-level global grid, calls the GNN to perform a single global inference, and finally outputs a high-fidelity multiphysics digital twin result.

[0051] Example 2: Multiphysics Digital Twin Parameter Inversion and Full-Field Reconstruction of ±800kV / 400MVA Converter Transformer: This embodiment describes the specific execution process of transforming sparse sensor observation data of a converter transformer into a multi-physics digital twin of global thermal flux at the million-node level using the spatially selective graph Bayesian inference method of the present invention: Step 1: Data preparation and configuration of asymmetric observation model: The global mesh dataset of the target transformer (including coordinates of approximately 915,000 nodes, material property parameters, and adjacency indexes) and the weights of a pre-trained graph neural network black-box surrogate model are loaded.

[0052] The core physical operation parameter to be inverted is set as follows: grid-side heat source power ( ), valve-side heat source power ( ) and ambient temperature ( ).

[0053] To simulate the highly challenging blind search environment with no prior information, a broad uniform prior boundary is set: the power parameter search boundary is 5.0 to 150.0 kW / m, and the ambient temperature search boundary is 253.0 to 323.0 K.

[0054] Set the actual physical sensor locations (e.g., three top-level oil field sensor measurement points with coordinates (334, -607), (107, -534), and (226, 302)). Set the large cutoff radius to simulate real-world full-field physical heat transfer. To generate a pure observational baseline value without truncation error; and to inject the standard deviation into this pure value. (i.e., 1.0 K) Gaussian white noise is used to generate the final virtual sensor observation vector. .

[0055] Step 2: Dynamic extraction of spatially selective subgraphs: Set the Pareto optimal cutoff radius for the MCMC inversion inference period. And set the sensing range radius of the physical probe. .

[0056] Regarding the coordinates of the aforementioned three sensors, Extract the corresponding node and its edge connection from the global grid for the radius, and generate three independent local inference subgraph objects, each of which is reduced to the level of thousands of nodes.

[0057] In each subgraph, according to Generate a sensor mask for subsequent extraction of the region mean.

[0058] Step 3: Parallel inversion sampling of high-frequency MCMC based on local subgraphs: Initialize MCMC algorithm parameters: Set up 6 parallel Markov chains, each chain performs 1500 high-frequency iterations, and set the first 20% (i.e., 300 steps) to be discarded as a warm-up period.

[0059] After the proposal parameters are generated in each iteration: The proposed three-dimensional parameter vector It is explicitly spliced ​​into the feature tensor of each node in the three local inference subgraphs.

[0060] A pre-trained graphical neural network is invoked in parallel for a single forward inference. The prediction results are averaged and aggregated using a sensor mask to obtain the predicted temperatures from the three sensors. .

[0061] Calculate and predict temperature With noisy observations The Gaussian log-likelihood error between the two probabilities is used to calculate the posterior probability of the current proposal by superimposing the prior probability.

[0062] The Metropolis-Hastings criterion is used to determine whether to accept the proposed parameters until 1500 iterations are completed. Since only a lightweight subgraph needs to be evaluated each time, the entire process of thousands of MCMC samplings takes only about 48 seconds.

[0063] Step 4: Posterior statistical analysis and full-scale twin delivery (worth millions of devices): After removing the burn-in preheating data from the MCMC chain, the posterior statistics of the stationary distribution are extracted. For example, under the operating condition of (126 kW / m, 126 kW / m, 296 K), the algorithm output includes a report containing the posterior mean (such as 125.04 kW / m, 124.31 kW / m, etc.) and the corresponding 95% confidence interval.

[0064] The mean of the obtained posterior optimal parameters is extracted, remapped, and spliced ​​into the global transformer graph containing a complete set of 915,000 nodes.

[0065] The pre-trained graph neural network is invoked for a final, complete full-field forward inference. Within an extremely short period of approximately 2.52 seconds, high-resolution temperature and fluid pressure distribution cloud maps of all nodes within the equipment can be reconstructed. Even under harsh conditions relying on top-layer oil engineering sensors with information dissipation, the mean absolute error (MAE) of the full-field reconstructed temperature can still be strictly controlled within 0.47 K, enabling the system to deliver a high-fidelity digital twin of the entire field.

[0066] As described above, this embodiment drastically reduces the computational cost of inversion and overcomes the bottleneck of real-time twin computing power. By introducing a spatially selective inference mechanism of "computation following observation," it successfully breaks the limitation of "global coupling" in solving inverse problems by the data-driven model. In a single Markov chain Monte Carlo (MCMC) iteration, the computational complexity is significantly reduced from O(Ntotal) for million-level full-field inference to O(Nsub) for local subgraph inference. Under the Pareto optimal truncation radius, the time required for a complete Bayesian inversion involving thousands of iterations is rapidly compressed from hours or even days in traditional methods to tens of seconds (e.g., around 48 seconds), clearing away the computational power obstacle for real-time high-frequency uncertainty inference of the equipment.

[0067] Using a pre-trained graph neural network as a black-box physics solver, the system maintains extremely high parameter point estimation accuracy even under large fluctuations in the full-condition envelope (e.g., from 33% light load to 145% extreme overload). When faced with extreme industrial measurement noise (e.g., Gaussian white noise interference up to 3.0σ), the system not only maintains high robustness against deviations in the inverted mean but also ensures complete coverage of the true parameters by adaptively widening the 95% confidence interval (95% CI), effectively avoiding the fatal weakness of deterministic surrogate models that are "blindly confident" under harsh data.

[0068] In addition, this application Figure 3 This application provides an embodiment of a multiphysics digital twin system based on spatially selective graph Bayesian inference. For example... Figure 3 As shown in the embodiments of this application, the system mainly includes: The generation module 210 is used to acquire the global grid data of the target device and the corresponding sensor spatial coordinate set; and to generate a local inference subgraph for each sensor position in the sensor spatial coordinate set.

[0069] The generation module 210 includes a generation unit. Used to acquire global mesh data of the target device; wherein, the global mesh data includes node spatial coordinates, initial physical characteristics of nodes and edge connection topology; and the nodes contain sparsely distributed sensors corresponding to the target device; Determine the sensor spatial coordinate set based on node spatial coordinates. ;in, The m-th sensor represents its spatial coordinates; M represents the total number of sensors. Get the cutoff radius and the radius of the perceptual domain ,and ; Using the sensor coordinates as the center, a spatial indexing algorithm is used to calculate the Euclidean distance from global nodes to the center point, and nodes with distances less than or equal to... The set of nodes is used as the nodes for generating the corresponding local reasoning subgraph; The node features, edge indices, and edge features corresponding to the nodes in the set of nodes corresponding to the current sensor coordinates will be extracted from the global grid data and encapsulated into a local inference subgraph corresponding to the current sensor coordinates.

[0070] The output module 220 is used to obtain the global operation proposal parameter value of the current inversion iteration during the inversion iteration process, and output the predicted temperature field of each node in the local inference subgraph under the current global operation proposal parameter value using a pre-trained graph neural network model.

[0071] The calculation module 230 is used to determine the physical probe range within a preset distance centered on the sensor location in the local inference subgraph, and to obtain the nodes within the physical probe range; based on the predicted temperature field, to extract the node predicted temperature of each node within the physical probe range, and then to calculate the mean as the final predicted value of the sensor; to obtain the actual observation value of each sensor; and based on the actual observation value and the final predicted value, to calculate the log-likelihood function value and the joint log-posterior probability of the current global operation proposal parameter value.

[0072] The iteration module 240 is used to generate new global operation proposal parameter values ​​based on the current global operation proposal parameter values ​​using a multi-chain parallel MCMC algorithm; determine whether to accept the updated global operation proposal parameter values ​​based on the joint logarithmic posterior probability; and repeatedly iterate and update the global operation proposal parameter values ​​corresponding to each sensor until the MCMC algorithm reaches the posterior distribution state.

[0073] In addition, embodiments of this application also provide a non-volatile computer storage medium storing executable instructions, which, when executed, implement a multiphysics digital twin method based on spatially selective graph Bayesian inference as described above.

[0074] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multiphysics digital twin method based on spatially selective graphical Bayesian inference, characterized in that, The method includes: Obtain the global grid data and corresponding sensor spatial coordinate set of the target device; for each sensor location in the sensor spatial coordinate set, generate a local inference subgraph, specifically including: Obtain the global mesh data of the target device; the global mesh data includes node spatial coordinates, initial physical characteristics of nodes, and edge connection topology; and the nodes contain sparsely distributed sensors corresponding to the target device; based on the node spatial coordinates, determine the sensor spatial coordinate set. ;in, The m-th sensor's spatial coordinates are represented; M represents the total number of sensors; the cutoff radius is obtained. and the radius of the perceptual domain ,and Using the sensor coordinates as the center, a spatial indexing algorithm is used to calculate the Euclidean distance from global nodes to the center point, and nodes with distances less than or equal to... The set of nodes is used as the nodes for generating the corresponding local inference subgraph; the node features, edge indices and edge features corresponding to the nodes in the set of nodes corresponding to the current sensor coordinates are extracted from the global grid data and encapsulated into the local inference subgraph corresponding to the current sensor coordinates; During the inversion iteration, the global operation proposal parameter values ​​for the current inversion iteration are obtained. Using a pre-trained graph neural network model, the predicted temperature field for each node within the local inference subgraph under the current global operation proposal parameter values ​​is output, specifically including: Obtain the global operation proposal parameter values ​​in the current MCMC iteration; after normalizing the global operation proposal parameter values, explicitly concatenate them along the feature dimension as global feature vectors and inject them into the feature vectors of each node in the local subgraph to obtain the local inference subgraph after injecting the global operation proposal parameter values; input the local inference subgraph after injecting the global operation proposal parameter values ​​into the pre-trained graph neural network model with frozen parameters to output the predicted temperature field of each node in the local inference subgraph under the current global operation proposal parameter values; In the local inference subgraph, the area within a preset distance centered on the sensor location is defined as the physical probe range. Each node within the physical probe range is obtained. Based on the predicted temperature field, the node predicted temperature of each node within the physical probe range is extracted, and then the average value is calculated as the final predicted value of the sensor. Obtain the actual observation values ​​of each sensor; based on the actual observation values ​​and the final predicted values, calculate the log-likelihood function value and joint log-posterior probability of the current global operation proposal parameter values; A multi-chain parallel MCMC algorithm is adopted to generate new global operation proposal parameter values ​​based on the current global operation proposal parameter values; whether to accept the updated global operation proposal parameter values ​​is determined based on the joint logarithmic posterior probability; the global operation proposal parameter values ​​corresponding to each sensor are iteratively updated repeatedly until the MCMC algorithm reaches the posterior distribution state.

2. The multiphysics digital twin method based on spatially selective graphical Bayesian inference according to claim 1, characterized in that, Before extracting the node-predicted temperature of each node within the physical probe range based on the predicted temperature field, and then calculating the average value as the final predicted value of the sensor, the method further includes: In the node set corresponding to the current sensor local inference subgraph, find the nodes within the physical probe range and generate a Boolean type probe sensing mask; Based on the predicted temperature field, the predicted node temperature of each node within the range of the physical probe is extracted, and then the average value is calculated as the final predicted value of the sensor. Specifically, this includes: Extract the node-predicted temperature of the node corresponding to the probe sensing mask from the predicted temperature field; The spatial mean of the predicted temperature at the computing nodes is used as the final predicted value of the sensor.

3. The multiphysics digital twin method based on spatially selective graphical Bayesian inference according to claim 1, characterized in that, Based on actual observations and final predictions, the log-likelihood function and joint log-posterior probability of the current global operation proposal parameters are calculated, specifically including: Based on the actual observations and final predictions obtained from each sensor, a Gaussian noise observation model for the sensor is constructed, and the prediction residual for each sensor is calculated. Through the formula: Calculate the log-likelihood function value ; Where M represents the total number of sensors. Indicates the preset standard deviation. Indicates the predicted residual; Obtain the prior probability density of global operation parameters Using the formula: Obtain the joint logarithmic posterior probability .

4. The multiphysics digital twin method based on spatially selective graphical Bayesian inference according to claim 1, characterized in that, A multi-chain parallel MCMC algorithm is employed to generate new global operation proposal parameter values ​​based on the current global operation proposal parameter values. The decision to accept the updated global operation proposal parameter values ​​is based on the joint logarithmic posterior probability, specifically including: In multi-chain parallel MCMC inversion sampling, the current global operation proposal parameter state is maintained for each independent Markov chain. In each iteration, each chain will perform a random walk within the boundary of its prior parameter space based on its current parameter value, according to the preset proposal distribution and the specified step size, to generate a new global operation proposal parameter for the chain. Calculate the new joint log-posterior probability based on the new global operation proposal parameters; If the difference between the new joint log-posterior probability and the previous joint log-posterior probability is greater than or equal to 0, the global operation proposal parameter value is accepted; otherwise, the global operation proposal parameter value is not accepted.

5. The multiphysics digital twin method based on spatially selective graphical Bayesian inference according to claim 1, characterized in that, Repeatedly iterate and update the global operation proposal parameter values ​​for each sensor until the MCMC algorithm reaches the posterior distribution state, specifically including: After each parameter update, the global operation proposal parameter value of the current chain is stored in the history sequence, forming a trajectory of parameter changes with the number of iteration steps; Calculate the between-group and within-group variances of the mean parameters of all parallel chains in the most recent N steps, and determine the convergence of the posterior distribution based on the Gelman-Rubin statistic. When the between-group variance, within-group variance, and posterior distribution convergence meet the preset convergence conditions, the MCMC algorithm is considered to have reached the posterior distribution state.

6. A multiphysics digital twin system based on spatially selective graph Bayesian inference, characterized in that, The system includes: The generation module is used to acquire the global grid data of the target device and the corresponding sensor spatial coordinate set; for each sensor position in the sensor spatial coordinate set, a local inference subgraph is generated. The generation module includes generation units. Used to acquire global mesh data of the target device; wherein, the global mesh data includes node spatial coordinates, initial physical characteristics of nodes and edge connection topology; and the nodes contain sparsely distributed sensors corresponding to the target device; Determine the sensor spatial coordinate set based on node spatial coordinates. ;in, The m-th sensor represents its spatial coordinates; M represents the total number of sensors. Get the cutoff radius and the radius of the perceptual domain ,and ; Using the sensor coordinates as the center, a spatial indexing algorithm is used to calculate the Euclidean distance from global nodes to the center point, and nodes with distances less than or equal to... The set of nodes is used as the nodes for generating the corresponding local reasoning subgraph; The node features, edge indices, and edge features corresponding to the nodes in the set of nodes corresponding to the current sensor coordinates will be extracted from the global grid data and encapsulated into a local inference subgraph corresponding to the current sensor coordinates. The output module is used to obtain the global operation proposal parameter values ​​for the current inversion iteration during the inversion iteration process. Utilizing a pre-trained graph neural network model, it outputs the predicted temperature field for each node within the local inference subgraph under the current global operation proposal parameter values. Specifically, it includes: Obtain the global operation proposal parameter values ​​in the current MCMC iteration; after normalizing the global operation proposal parameter values, explicitly concatenate them along the feature dimension as global feature vectors and inject them into the feature vectors of each node in the local subgraph to obtain the local inference subgraph after injecting the global operation proposal parameter values; input the local inference subgraph after injecting the global operation proposal parameter values ​​into the pre-trained graph neural network model with frozen parameters to output the predicted temperature field of each node in the local inference subgraph under the current global operation proposal parameter values; The calculation module is used to determine the physical probe range within a preset distance centered on the sensor location in the local inference subgraph, and to acquire the nodes within the physical probe range; based on the predicted temperature field, to extract the node predicted temperature of each node within the physical probe range, and then to calculate the mean as the final predicted value of the sensor; to acquire the actual observation value of each sensor; and based on the actual observation value and the final predicted value, to calculate the log-likelihood function value and joint log-posterior probability of the current global operation proposal parameter value. The iterative module is used to generate new global operation proposal parameter values ​​based on the current global operation proposal parameter values ​​using a multi-chain parallel MCMC algorithm; determine whether to accept the updated global operation proposal parameter values ​​based on the joint logarithmic posterior probability; and repeatedly iterate and update the global operation proposal parameter values ​​corresponding to each sensor until the MCMC algorithm reaches the posterior distribution state.

7. A non-volatile computer storage medium, characterized in that, It stores computer instructions that, when executed, implement a multiphysics digital twin method based on spatially selective graphical Bayesian inference as described in any one of claims 1-5.

Images