Dynamic construction method of digital twin based on multi-source data fusion and physical simulation

Abstract

The application relates to the technical field of digital twinning, physical modeling and multi-source data fusion, and provides a dynamic construction method of a digital twinning body based on multi-source data fusion and physical simulation. The method comprises the following steps: acquiring multi-source heterogeneous data streams from a preset sensor array, numerical simulation results and a historical database, identifying key characteristic modes of a dominant physical process in the multi-source heterogeneous data streams, acquiring time-varying physical field evolution rules corresponding to the key characteristic modes by using a time sliding window and a forgetting mechanism, extracting a low-dimensional sparse characteristic parameter set reflecting dynamic behavior from a high-dimensional observation space, and constructing a reduced-order surrogate model by using a Gaussian process regression, a neural network surrogate model or intrinsic orthogonal decomposition combined with an interpolation technology; receiving corresponding real-time observation data streams, combining the reduced-order surrogate model to establish a full-closed-loop feedback link from model prediction, high-fidelity solution verification to observation data correction, and completing construction of the digital twinning body.

Description

Technical Field

[0001] This application relates to the fields of digital twin, physical modeling and multi-source data fusion technology, and in particular to a method for dynamically constructing a digital twin based on multi-source data fusion and physical simulation. Background Technology

[0002] In the fields of digital twin, physical modeling, and multi-source data fusion, the construction of digital twins for complex physical systems (such as industrial equipment and engineering systems) faces significant challenges. Traditional methods often rely on static models or single data sources, mapping the physical system to the virtual model through predefined rules. For example, while some solutions achieve the integration of geometric models, physical models, process models, and data models, their core flaw lies in:

[0003] 1. Insufficient static modeling and dynamic adaptability: Relying on multi-model fusion with a fixed structure, it lacks the ability to dynamically identify the dominant physical processes (such as strong nonlinearity and time-varying characteristics) in the real-time operation of the system, and cannot extract key features with physical significance from high-dimensional data, resulting in limited ability of the model to represent complex working conditions.

[0004] 2. Imbalance between computational efficiency and accuracy: Using high-fidelity physical solutions or synchronous updates of the complete model across the entire domain and all time periods, the computational burden increases sharply when the system state changes abruptly, making it difficult to achieve real-time response on edge devices or in environments with limited computing power. Furthermore, global solutions consume a large amount of resources and fail to achieve on-demand allocation of computing resources.

[0005] 3. Error accumulation and lack of self-optimization: Model errors rely on manual correction or offline updates. There is a lack of closed-loop feedback mechanism between the surrogate model, physical solution and observation data. Long-term operation is prone to accumulated bias due to environmental disturbances or simplified assumptions, and it cannot adaptively improve the reliability of predictions.

[0006] Therefore, a method is urgently needed to solve at least one of the above problems. Summary of the Invention

[0007] This application provides a dynamic construction method for digital twins based on multi-source data fusion and physical simulation. It aims to address the shortcomings of traditional methods that often rely on static models or single data sources, mapping physical systems to virtual models through predefined rules. For example, while some solutions integrate geometric, physical, process, and data models, their core shortcomings lie in insufficient static modeling and dynamic adaptability, an imbalance between computational efficiency and accuracy, and the accumulation of errors and a lack of self-optimization.

[0008] In a first aspect, embodiments of this application provide a method for dynamically constructing a digital twin based on multi-source data fusion and physical simulation, the method comprising:

[0009] Acquire multi-source heterogeneous data streams from preset sensor arrays, numerical simulation results, and historical databases; identify key feature modes that dominate physical processes in the multi-source heterogeneous data streams; utilize time sliding windows and forgetting mechanisms to obtain the time-varying physical field evolution laws of key feature modes corresponding to key feature modes; and extract low-dimensional sparse feature parameter sets reflecting dynamic behavior from high-dimensional observation space.

[0010] Using the low-dimensional sparse feature parameter set as input, a reduced-order surrogate model is constructed by employing Gaussian process regression, neural network surrogate model, or intrinsic orthogonal decomposition combined with interpolation techniques. The reduced-order surrogate model is then designed for lightweight performance and optimized for real-time performance.

[0011] Set a trigger threshold based on the rate of change of sparse features or uncertainty estimation. If a sudden change in feature parameters is detected according to the trigger threshold and exceeds the reliable prediction range of the reduced-order surrogate model, activate the high-fidelity physics solver for the local region. The high-fidelity physics solver is used to calculate for the physical sub-region and time segment where the sudden change occurred, and the result is fed back as the truth value to the corresponding model update process.

[0012] The system receives the corresponding real-time observation data stream, integrates the observation information corresponding to the real-time observation data stream into the prediction output of the surrogate model through data assimilation technology, dynamically corrects the prediction bias of the reduced-order surrogate model, and outputs prediction results with uncertainty quantification. This allows for the establishment of a fully closed-loop feedback link from model prediction, high-fidelity solution verification to observation data correction, in conjunction with the reduced-order surrogate model, thus completing the construction of the digital twin.

[0013] In some embodiments, after the construction of the digital twin is completed, the method further includes: periodically or triggerively optimizing the sparse feature extraction algorithm corresponding to the low-dimensional sparse feature parameter set based on the prediction error, uncertainty index and the frequency of triggering the high-fidelity physical solver generated by the digital twin during system operation, and retraining and optimizing the structure and hyperparameters of the reduced-order surrogate model, so as to improve the ability of the sparse feature set to capture the dynamic evolution characteristics of the system and the prediction accuracy and robustness of the surrogate model.

[0014] In some embodiments, identifying key feature modes that dominate physical processes in multi-source heterogeneous data streams includes: preprocessing the acquired multi-source heterogeneous data streams, unifying real-time data collected by sensor arrays, simulated data obtained from numerical simulations, and archived data in historical databases to the same time reference, and introducing physical process consistency constraints at the data layer to eliminate interference caused by heterogeneous scales to feature identification; constructing a state matrix that is updated over time, performing mode decoupling processing on the state matrix within a preset time window, and constraining the consistency of dynamic evolution between adjacent time slices; and using any one of the adaptive algorithms, such as dynamic mode decomposition, sparse coding, or nonlinear independent component analysis, to analyze the results after mode decoupling in order to identify key feature modes that can reflect the dominant physical processes of the system and have clear physical meanings, and to eliminate redundant modes related to non-dominant physical processes.

[0015] In some embodiments, the method of acquiring the time-varying physical field evolution law of key feature modes corresponding to key feature modes using a time sliding window and forgetting mechanism includes: setting a time sliding window of fixed duration, allowing the time sliding window to continuously scroll with the time process, and capturing the time-series data of key feature modes in different time segments in real time; introducing an online update mechanism with forgetting weights to exponentially decay the contribution of key feature modes at historical moments within the time sliding window, and reducing the impact of early state data on the current feature estimation through a preset forgetting factor; and simultaneously setting a historical consistency constraint strength parameter to control the continuous evolution trend of key feature modes over time, avoiding interference caused by transient noise to the tracking of evolution law, thereby realizing online tracking and accurate acquisition of the time-varying physical field evolution law corresponding to key feature modes.

[0016] In some embodiments, the extraction of a low-dimensional sparse feature parameter set reflecting dynamic behavior from a high-dimensional observation space includes: using L1 regularization or structured sparsity constraints to sparsify the time-series data of key feature modes acquired within a time sliding window, enhancing the sparsity of feature parameters, eliminating redundant features and interference information corresponding to non-dominant physical processes, ensuring that only a small number of feature parameters remain active at any given time, and that some of the activated feature parameters explicitly correspond to the dominant physical mechanism in the system; introducing a feature selection criterion based on energy contribution and prediction sensitivity to comprehensively evaluate the importance of the sparsified feature parameters, considering the proportion of each feature parameter in the overall physical energy and its sensitivity to system output prediction during the evaluation process; and sorting and pruning the feature parameters according to the comprehensive importance evaluation results, retaining key feature parameters and eliminating irrelevant feature parameters to form a low-dimensional sparse feature parameter set that is dimensionally controlled, time-consistent, physically meaningful, and reflects the essential dynamic behavior of the system.

[0017] In some embodiments, the step of constructing a reduced-order surrogate model using the low-dimensional sparse feature parameter set as input and employing Gaussian process regression, a neural network surrogate model, or intrinsic orthogonal decomposition combined with interpolation techniques includes: using the extracted low-dimensional sparse feature parameter set as the sole input to the reduced-order surrogate model, enabling the reduced-order surrogate model to perform mapping operations only within the physical dominant subspace, thereby controlling model complexity and ensuring real-time performance from the root; constructing a unified reduced-order mapping relationship, compressing sparse feature parameters to a fixed low-dimensional computational space through a feature hash mapping matrix, reducing the memory access and multiply-accumulate overhead of the reduced-order surrogate model; selecting any one of Gaussian process regression, lightweight neural networks, or interpolation operators based on intrinsic orthogonal decomposition as the main body of the reduced-order surrogate model, inputting the compressed low-dimensional sparse feature parameters into the main body of the reduced-order surrogate model, and training through samples to enable the main body of the reduced-order surrogate model to approximate the system response of a high-fidelity physics solver in the same feature space with the lowest computational cost, thus completing the construction of the reduced-order surrogate model.

[0018] In some embodiments, the lightweight design and real-time optimization of the downgraded proxy model includes: using at least one of feature hashing, quantization-aware training, or model pruning to perform lightweight design on the constructed downgraded proxy model; identifying and compressing the scale of non-critical parameters in the downgraded proxy model; eliminating redundant computational paths; enabling the downgraded proxy model to have deterministic computational latency during the deployment phase, and adapting to edge computing devices or environments with limited computing resources; configuring an online incremental learning function for the downgraded proxy model; setting a forgetting factor and learning rate consistent with the sparse feature update mechanism; enabling the downgraded proxy model to prioritize adapting to the latest physical state changes; and avoiding model structural instability caused by short-term fluctuations, thereby achieving real-time optimization of the downgraded proxy model.

[0019] In some embodiments, activating a high-fidelity physics solver for a local region if a feature parameter mutation is detected based on a trigger threshold and exceeds the reliable prediction range of the reduced-order surrogate model includes: constructing a unified trigger criterion function, which simultaneously characterizes the evolution rate of the dominant physical process in the feature space and the prediction uncertainty of the reduced-order surrogate model output, and balancing the impact of physical mutation intensity and insufficient model cognition on triggering behavior through preset weight coefficients; comparing the calculation result of the constructed trigger criterion function with a preset trigger threshold, and if the calculation result exceeds the preset trigger threshold, determining that the feature parameter has mutated and exceeds the reliable prediction range of the reduced-order surrogate model; locating the abnormal feature parameter to a specific physical sub-region through a feature-to-space mapping operator, the mapping process combining sparse feature components and modal basis functions of key feature modes, and ensuring that the high-fidelity physics solver is activated only in the spatial region where the dominant mode has a significant effect through a preset energy contribution threshold; and calling any one of the high-fidelity physics solvers among the finite element, finite volume, or spectral methods to perform high-precision calculations only for the located physical sub-region and the corresponding time segment.

[0020] In some embodiments, the process of integrating observation information corresponding to real-time observation data streams into the prediction output of a surrogate model using data assimilation techniques, dynamically correcting the prediction bias of the reduced-order surrogate model, and outputting prediction results with uncertainty quantification includes: continuously receiving real-time observation data streams during system operation; synchronously comparing the prediction results output by the reduced-order surrogate model based on sparse feature parameters with the real-time observation data streams; calculating the corresponding observation residuals, which are used to characterize the degree of prediction mismatch of the surrogate model under the current physical state; constructing a bias-aware data assimilation update operator, and updating the data based on the prediction uncertainty scalar output by the reduced-order surrogate model and... The observation system estimates the observation noise intensity online and calculates the time-dependent adaptive correction gain. This correction gain is used to adjust the correction intensity of the prediction results by the observation information, avoiding the amplification effect of observation noise on model stability. The observation residuals and the adaptive correction gain are substituted into the data assimilation update operator to dynamically correct the prediction results of the reduced-order surrogate model, resulting in a corrected system state estimate. The corrected system state estimate and the prediction uncertainty scalar output by the reduced-order surrogate model are output synchronously to form a prediction result with uncertainty quantification. The observation residuals are mapped back to the sparse feature space to form a bias-driven feature correction signal.

[0021] In some embodiments, the establishment of a fully closed-loop feedback chain from model prediction and high-fidelity solution verification to observation data correction, combined with a reduced-order surrogate model, to complete the construction of the digital twin, includes: unifying the prediction results of the reduced-order surrogate model, the reference truth value output by the high-fidelity physical solver, the correction state after data assimilation, and the observation residuals into the same closed-loop feedback framework; constructing a comprehensive feedback evaluation function, which comprehensively characterizes the true prediction error of the current model system, the prediction uncertainty of the surrogate model, and the triggering frequency of the high-fidelity solver; balancing prediction accuracy, model robustness, and computational cost through preset weight coefficients; and based on the calculation results and long-term evolution trend of the comprehensive feedback evaluation function, adjusting the sparseness corresponding to the low-dimensional sparse feature parameter set. The feature extraction algorithm performs parameter callbacks and adaptively adjusts the sparse constraint strength parameters, enabling the sparse feature parameter set to quickly respond to structural changes in the physical field while maintaining the ability to remember stable evolution patterns, thus conforming to the real dynamic evolution structure of the system. Using the calculation results of the comprehensive feedback evaluation function, the structural complexity of the reduced-order surrogate model is adaptively controlled. While ensuring that prediction errors and uncertainties meet the standards, the non-critical computational paths of the reduced-order surrogate model are continuously compressed, maintaining its lightweight characteristics. A fully closed-loop feedback chain covering sparse feature extraction, surrogate model prediction, high-fidelity solution verification, and observation data correction is established, enabling the corresponding digital twin to possess continuous self-optimization and adaptive capabilities, completing the dynamic construction of a high-fidelity digital twin.

[0022] This application extracts low-dimensional sparse feature parameters from high-dimensional data using algorithms such as dynamic mode decomposition and sparse coding, directly characterizing the evolutionary laws of the dominant physical processes, reducing redundant noise interference, and improving model stability and interpretability. A lightweight reduced-order surrogate model is constructed based on sparse features, and combined with model pruning and edge computing optimization, millisecond-level inference is achieved to meet real-time requirements. Local high-fidelity solving is triggered only when feature parameters undergo abrupt changes or model uncertainty exceeds limits, avoiding the waste of resources in global high-fidelity computation and balancing accuracy and efficiency. Real-time observation data is fused through data assimilation technology to dynamically correct model biases and output uncertainty quantification results, suppressing accumulated errors. Based on feedback information such as prediction error and trigger frequency, the feature set and model structure are collaboratively optimized, enabling the digital twin to have self-evolution capabilities and continuously improving robustness in complex scenarios.

[0023] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic flowchart illustrating the steps of a method for dynamically constructing a digital twin based on multi-source data fusion and physical simulation, provided in an embodiment of this application.

[0026] Figure 2 This is a flowchart illustrating the overall technical closed-loop process provided in one embodiment of this application;

[0027] Figure 3 This is a sparse feature adaptive extraction data flow graph provided in an embodiment of this application;

[0028] Figure 4 This is a data assimilation online correction control loop diagram provided in one embodiment of this application;

[0029] Figure 5 This is a diagram of a closed-loop feedback iterative optimization process provided in an embodiment of this application;

[0030] Figure 6 This is a schematic block diagram of a dynamic construction system for digital twins based on multi-source data fusion and physical simulation, provided in one embodiment of this application.

[0031] Figure 7 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application.

[0032] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Detailed Implementation

[0033] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the order described. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0035] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0036] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0037] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0038] In the fields of digital twin, physical modeling, and multi-source data fusion, the construction of digital twins for complex physical systems (such as industrial equipment and engineering systems) faces significant challenges. Traditional methods often rely on static models or single data sources, mapping the physical system to the virtual model through predefined rules. For example, while some solutions achieve the integration of geometric models, physical models, process models, and data models, their core flaw lies in:

[0039] 1. Insufficient static modeling and dynamic adaptability: Relying on multi-model fusion with a fixed structure, it lacks the ability to dynamically identify the dominant physical processes (such as strong nonlinearity and time-varying characteristics) in the real-time operation of the system, and cannot extract key features with physical significance from high-dimensional data, resulting in limited ability of the model to represent complex working conditions.

[0040] 2. Imbalance between computational efficiency and accuracy: Using high-fidelity physical solutions or synchronous updates of the complete model across the entire domain and all time periods, the computational burden increases sharply when the system state changes abruptly, making it difficult to achieve real-time response on edge devices or in environments with limited computing power. Furthermore, global solutions consume a large amount of resources and fail to achieve on-demand allocation of computing resources.

[0041] 3. Error accumulation and lack of self-optimization: Model errors rely on manual correction or offline updates. There is a lack of closed-loop feedback mechanism between the surrogate model, physical solution and observation data. Long-term operation is prone to accumulated bias due to environmental disturbances or simplified assumptions, and it cannot adaptively improve the reliability of predictions.

[0042] Therefore, a method is urgently needed to solve at least one of the above problems.

[0043] To solve the above problem, please refer to Figure 1 This application provides a method for dynamically constructing a digital twin based on multi-source data fusion and physical simulation, applicable to computer devices. The computer devices can be deployed on a single server or server cluster, or on handheld terminals, laptops, wearable devices, or robots. It should be noted that all information involved in the method provided in this application is extracted with the authorization of the relevant user and in accordance with relevant regulations, and will not infringe on user privacy.

[0044] The provided method for dynamically constructing digital twins based on multi-source data fusion and physical simulation includes steps S101 to S104. Details are as follows:

[0045] Step S101. Acquire multi-source heterogeneous data streams from preset sensor arrays, numerical simulation results and historical databases, identify key feature modes that dominate physical processes in the multi-source heterogeneous data streams, use time sliding windows and forgetting mechanisms to obtain the time-varying physical field evolution law of key feature modes corresponding to key feature modes, and extract low-dimensional sparse feature parameter sets reflecting dynamic behavior from high-dimensional observation space.

[0046] Specifically, the core of this step is to complete the fusion processing of multi-source heterogeneous data and extract key dynamic features, solving the problem that traditional methods cannot extract key physical features from high-dimensional data and are difficult to adapt to the time-varying characteristics of the system. By integrating multi-source data, identifying the core features that dominate the physical process, and capturing time-varying laws, the dimensionality is finally reduced to obtain a sparse feature set that reflects the dynamic behavior of the system, providing accurate input for subsequent model construction.

[0047] The multi-source heterogeneous data stream acquisition simultaneously collects multiple types and dimensions of data from preset sensor arrays (such as temperature, pressure, and vibration sensors for industrial equipment), numerical simulation results (such as finite element simulation and fluid simulation outputs), and historical databases (past system operation data and fault data), forming a unified data stream input that covers all dimensions of information such as system geometry, physics, and operating status.

[0048] Key feature mode identification is based on prior physical knowledge and data-driven methods (such as modal analysis and feature extraction algorithms). It identifies the feature modes corresponding to the core physical processes that dominate the operation of the system from high-dimensional data streams, eliminates redundant and noisy data, and focuses on features that play a decisive role in the dynamic behavior of the system (such as features corresponding to strongly nonlinear elements and time-varying parameters).

[0049] The extraction of time-varying physical field evolution laws adopts a time sliding window to dynamically extract feature data of continuous time periods. Combined with the forgetting mechanism, the weight of historical outdated data is weakened and the influence of recent real-time data is strengthened, so as to accurately capture the change law of key feature modes over time and clarify the dynamic evolution trend of physical field.

[0050] Low-dimensional sparse feature parameter set extraction uses dimensionality reduction algorithms (such as principal component analysis and sparse coding) to screen and retain the low-dimensional features that best reflect the dynamic essence of the system from the high-dimensional observation space, forming a sparse feature parameter set. This retains the core physical information while reducing the data dimensionality and the subsequent computational burden.

[0051] Step S102. Using the low-dimensional sparse feature parameter set as input, construct a reduced-order surrogate model using Gaussian process regression, neural network surrogate model, or intrinsic orthogonal decomposition combined with interpolation techniques, and perform lightweight design and real-time optimization on the reduced-order surrogate model.

[0052] Specifically, this step uses the low-dimensional sparse features output by S101 as a foundation to construct a lightweight, real-time-efficient reduced-order proxy model, addressing the issues of computational efficiency versus accuracy imbalance and inability to adapt to edge / limited computing environments in traditional methods. By selecting a suitable proxy model construction method and combining it with lightweight optimization, the model achieves both prediction accuracy and real-time response requirements.

[0053] The model input is determined by using the low-dimensional sparse feature parameter set obtained in step S101 as the sole input, ensuring that the model input focuses on the core dynamic information and avoids interference from redundant data.

[0054] The construction of the reduced-order surrogate model selects the appropriate method based on the system characteristics. If the system has weak nonlinearity and high accuracy requirements, Gaussian process regression is used to build the model. If the system has strong nonlinearity and large amount of data, a neural network surrogate model (such as lightweight CNN or LSTM) is used to capture complex mapping relationships. If both physical meaning and computational efficiency need to be considered, intrinsic orthogonal decomposition combined with interpolation technology is used to quickly build a reduced-order model based on the orthogonal basis functions of the physical field.

[0055] Lightweight design simplifies the structure of the constructed proxy model (such as neural network pruning and reducing hidden layer nodes), quantizes parameters (such as converting floating-point numbers to fixed-point numbers), and removes redundant modules, thereby reducing the number of model parameters and computational complexity.

[0056] Real-time optimization improves model inference speed through algorithm acceleration (such as parallel computing and inference engine optimization) and hardware adaptation (such as model compilation for edge devices), ensuring millisecond-level real-time response under limited computing power.

[0057] Step S103. Set a trigger threshold based on the sparse feature change rate or uncertainty estimation. If a feature parameter mutation is detected according to the trigger threshold and exceeds the reliable prediction range of the reduced-order surrogate model, activate the high-fidelity physics solver for the local region. The high-fidelity physics solver is used to calculate the physical sub-region and time segment where the mutation occurs, and feeds the result back to the corresponding model update process as the truth value.

[0058] Specifically, this step addresses the problems of heavy computational burden and unreasonable resource allocation in traditional methods by combining threshold triggering with local high-fidelity solving. Local high-fidelity solving is activated only when the system state abruptly exceeds the reliability range of the surrogate model, enabling on-demand allocation of computational resources while balancing accuracy and efficiency.

[0059] The trigger threshold is set based on the system's historical operating data and physical characteristics. Two types of trigger thresholds are set: one is the sparse feature change rate threshold (monitoring the abrupt change amplitude of feature parameters), and the other is the uncertainty estimation threshold (monitoring the confidence interval predicted by the proxy model). The two types of thresholds are cross-checked to ensure the accuracy of the trigger logic.

[0060] Mutation detection and range determination are achieved by monitoring changes in low-dimensional sparse feature parameters in real time. If the magnitude of a mutation in a feature parameter exceeds the rate of change threshold, or the uncertainty of the surrogate model prediction exceeds the estimation threshold, the system state is determined to have entered a range that the surrogate model cannot reliably predict.

[0061] Local high-fidelity solution activation is only applied to physical sub-regions where abrupt changes occur (such as equipment failure points or sudden changes in operating conditions) and time segments (the brief period during which the abrupt change occurs). It activates high-fidelity physical solvers (such as high-precision finite element solvers and multi-physics coupled solvers) to avoid the waste of resources caused by global and all-time-time solutions.

[0062] The truth feedback and model update process uses the calculation results of local high-fidelity solutions as "truth data" and feeds them back to the update process of the proxy model, providing accurate references for subsequent model corrections and ensuring the accuracy of the model under sudden changes.

[0063] Step S104. Receive the corresponding real-time observation data stream, integrate the observation information corresponding to the real-time observation data stream into the prediction output of the surrogate model through data assimilation technology, dynamically correct the prediction bias of the reduced-order surrogate model, and output the prediction result with uncertainty quantification. In order to combine the reduced-order surrogate model to establish a fully closed-loop feedback link from model prediction, high-fidelity solution verification to observation data correction, and complete the construction of the digital twin.

[0064] Specifically, this step addresses the issues of error accumulation and lack of self-optimization capability inherent in traditional methods through data assimilation and closed-loop feedback. It integrates real-time observation data to correct biases in the surrogate model, outputting prediction results with quantified uncertainty, forming a complete "prediction-verification-correction" closed loop to achieve adaptive model optimization and long-term reliable operation.

[0065] The real-time observation data stream receives real-time observation data from the sensor array through continuous acquisition, ensuring the real-time nature and accuracy of the data and providing a true basis for model bias correction.

[0066] Data assimilation and bias correction employ data assimilation techniques (such as Kalman filtering, ensemble Kalman filtering, and variational assimilation) to fuse real-time observation information with the prediction output of the reduced-order surrogate model, dynamically correcting the model's prediction bias and offsetting the cumulative errors caused by environmental disturbances and simplified assumptions.

[0067] Uncertainty quantification output is based on factors such as model prediction error, observation data noise, and parameter uncertainty. It quantifies the uncertainty of the prediction results (such as confidence interval and probability distribution) and outputs prediction results with uncertainty information, thereby improving the reliability and reference value of the prediction results.

[0068] The fully closed-loop feedback chain is constructed by integrating the entire process of "surrogate model prediction → local high-fidelity solution verification → observation data assimilation correction" to form a closed-loop feedback mechanism. Model prediction deviation triggers local solution, and the solution truth and observation data jointly correct the model. The corrected model further improves the subsequent prediction accuracy, realizing the dynamic self-optimization and continuous reliable construction of the digital twin.

[0069] In some embodiments, after the construction of the digital twin is completed, the method further includes: periodically or triggerively optimizing the sparse feature extraction algorithm corresponding to the low-dimensional sparse feature parameter set based on the prediction error, uncertainty index and the frequency of triggering the high-fidelity physical solver generated by the digital twin during system operation, and retraining and optimizing the structure and hyperparameters of the reduced-order surrogate model, so as to improve the ability of the sparse feature set to capture the dynamic evolution characteristics of the system and the prediction accuracy and robustness of the surrogate model.

[0070] This embodiment is a supplementary optimization to the core technical solution after the "digital twin is constructed". Its core purpose is to address the problems of decreased dynamic evolution capability of the sparse feature set capture system and insufficient prediction accuracy and robustness of the surrogate model during long-term operation of the digital twin. By introducing three key feedback information types—prediction error, uncertainty index, and high-fidelity solution trigger frequency—dynamic optimization of the sparse feature extraction algorithm parameters, surrogate model structure, and hyperparameters is achieved. This improves the closed-loop feedback chain, enabling the digital twin to possess continuous self-optimization and self-adaptation capabilities, maintaining high accuracy and robustness over the long term.

[0071] Feedback information collection involves collecting three types of core feedback information in real time after the digital twin is built and put into long-term operation: first, the deviation between the prediction results output by the digital twin and the actual physical system's operating state (i.e., prediction error); second, the prediction uncertainty index output synchronously by the downgraded surrogate model (reflecting the degree of model's insufficient understanding); and third, the triggering frequency of the high-fidelity physical solver (reflecting the surrogate model's sufficient coverage of the physical state).

[0072] The optimization triggering mechanism adopts a combination of "periodic triggering and trigger-based triggering". On the one hand, a fixed optimization cycle is set (such as daily or weekly, which can be flexibly set according to the system's operational complexity). On the other hand, when the prediction error exceeds the preset threshold, the uncertainty index rises abnormally, or the high-fidelity solution triggering frequency is too high / too low, the optimization process is immediately triggered.

[0073] Based on the three types of feedback information collected, the parameter tuning of the sparse feature extraction algorithm optimizes the key parameters in the sparse feature extraction process (such as the sparse constraint strength, forgetting factor, and comprehensive importance weight of feature selection). For example, when the prediction error is too large or the triggering frequency is too high, the sparse constraint strength is reduced to retain more features related to the dominant physical process and improve the feature set's ability to capture the dynamic evolution of the system.

[0074] Optimization of the reduced-order surrogate model involves retraining and adjusting the structure (such as the number of hidden layer nodes and pruning ratio in a neural network) and hyperparameters (such as the kernel function parameters of Gaussian process regression and the learning rate of incremental learning). For example, when the model uncertainty is too high, the parameter scale of the critical computation path of the model can be increased, or the forgetting factor of incremental learning can be adjusted to enable the model to adapt to the latest physical state more quickly while maintaining its lightweight characteristics, ultimately improving the prediction accuracy and robustness of the surrogate model.

[0075] In some embodiments, identifying key feature modes that dominate physical processes in multi-source heterogeneous data streams includes: preprocessing the acquired multi-source heterogeneous data streams, unifying real-time data collected by sensor arrays, simulated data obtained from numerical simulations, and archived data in historical databases to the same time reference, and introducing physical process consistency constraints at the data layer to eliminate interference caused by heterogeneous scales to feature identification; constructing a state matrix that is updated over time, performing mode decoupling processing on the state matrix within a preset time window, and constraining the consistency of dynamic evolution between adjacent time slices; and using any one of the adaptive algorithms, such as dynamic mode decomposition, sparse coding, or nonlinear independent component analysis, to analyze the results after mode decoupling in order to identify key feature modes that can reflect the dominant physical processes of the system and have clear physical meanings, and to eliminate redundant modes related to non-dominant physical processes.

[0076] This embodiment refines the core step S101, "identifying the key feature modes of the dominant physical process in multi-source heterogeneous data streams." It addresses the interference in feature recognition caused by scale differences and inconsistent time bases in multi-source heterogeneous data (sensor data, numerical simulation data, historical data), as well as the interference from redundant modes of non-dominant physical processes. Through data preprocessing, physical consistency constraints, and mode decoupling, it achieves accurate identification of key feature modes, ensuring that the identification results have clear physical meaning.

[0077] The multi-source heterogeneous data preprocessing first cleans and standardizes the three types of multi-source heterogeneous data streams acquired, removing outliers and missing values. Then, the real-time data collected by the sensor array (such as temperature and pressure data), the simulation data obtained by numerical simulation (such as finite element simulation results), and the archived data in the historical database (such as characteristic data and fault data of the system's past operation) are uniformly converted to the same time base to ensure that all types of data are synchronized in time and to eliminate interference caused by differences in time scales.

[0078] The physical consistency constraint at the data layer introduces physical process consistency constraints after data preprocessing. Combined with the inherent physical laws of complex physical systems (such as energy conservation and momentum conservation), it verifies multi-source data, eliminates data that contradicts physical laws, avoids interference from heterogeneous scale differences and data noise on subsequent feature recognition, and ensures the validity and consistency of input data.

[0079] The state matrix construction and modal decoupling are based on preprocessed multi-source data to construct a state matrix that is updated over time. This matrix contains multi-dimensional operating state information of the system under different time slices. A time window of preset duration is set, and modal decoupling is performed on the state matrix within the window. At the same time, the consistency of dynamic evolution between adjacent time slices is constrained to ensure that the decoupled modes can stably reflect the evolution trajectory of the system's physical processes and avoid modal distortion caused by time series fluctuations.

[0080] Key feature mode identification employs any one of the adaptive algorithms from dynamic mode decomposition, sparse coding, or nonlinear independent component analysis to conduct in-depth analysis of the decoupled mode results, and selects key feature modes that can reflect the dominant physical processes of the system and have clear physical meanings. At the same time, redundant modes related to non-dominant physical processes (such as small perturbations and irrelevant noise) are eliminated to ensure that the key feature modes identified in the end can accurately correspond to the core dynamic behavior of the system.

[0081] In some embodiments, the method of acquiring the time-varying physical field evolution law of key feature modes corresponding to key feature modes using a time sliding window and forgetting mechanism includes: setting a time sliding window of fixed duration, allowing the time sliding window to continuously scroll with the time process, and capturing the time-series data of key feature modes in different time segments in real time; introducing an online update mechanism with forgetting weights to exponentially decay the contribution of key feature modes at historical moments within the time sliding window, and reducing the impact of early state data on the current feature estimation through a preset forgetting factor; and simultaneously setting a historical consistency constraint strength parameter to control the continuous evolution trend of key feature modes over time, avoiding interference caused by transient noise to the tracking of evolution law, thereby realizing online tracking and accurate acquisition of the time-varying physical field evolution law corresponding to key feature modes.

[0082] This embodiment refines the core step S101, "using a time sliding window and a forgetting mechanism to obtain the time-varying physical field evolution law of the key feature mode corresponding to the key feature mode," and fundamentally solves the problems of difficulty in tracking the evolution law of the time-varying physical field online, interference of historical data with the current feature estimation, and distortion of the evolution law caused by transient noise. Through real-time capture of the time sliding window and dynamic adjustment of the forgetting mechanism, the accurate and stable acquisition of the evolution law of the time-varying physical field is achieved.

[0083] The time sliding window setting is based on the operating cycle and evolution rate of the complex physical system. A fixed time sliding window (such as 10s or 1min, which can be adaptively adjusted according to the system characteristics) is set so that the time sliding window continues to roll with the time process, capturing the time series data of key feature modes in different time segments in real time, ensuring that the key time series nodes of the physical field evolution can be covered, and realizing the real-time tracking of the evolution law.

[0084] An online update mechanism with forgetting weights is introduced on the basis of a time sliding window. This mechanism exponentially decays the contribution of key feature modes at historical moments within the window. A preset forgetting factor (with a value range of 0 < forgetting factor < 1) is used to reduce the impact of early state data on the current feature estimation, enabling feature parameters to respond quickly to structural changes in the physical field and avoiding feature lag caused by outdated historical data.

[0085] Historical consistency constraint control sets a historical consistency constraint strength parameter, which is used to control the continuous evolution trend of key feature modes over time, avoiding interference from transient noise (such as instantaneous sensor interference and small system fluctuations) on the tracking of evolution laws. Through this constraint, it is ensured that the acquired time-varying physical field evolution laws can respond to sudden structural changes and maintain the ability to remember stable evolution patterns, thus achieving stable and accurate acquisition of evolution laws.

[0086] The evolution law output, through the aforementioned window rolling, forgetting update and consistency constraint mechanism, outputs the time-varying physical field evolution trajectory corresponding to the key feature modes in real time, clarifies the changing trend of the dominant physical process over time, and provides a high-quality data foundation with temporal consistency and noise suppression for the subsequent extraction of low-dimensional sparse feature parameter sets.

[0087] In some embodiments, the extraction of a low-dimensional sparse feature parameter set reflecting dynamic behavior from a high-dimensional observation space includes: using L1 regularization or structured sparsity constraints to sparsify the time-series data of key feature modes acquired within a time sliding window, enhancing the sparsity of feature parameters, eliminating redundant features and interference information corresponding to non-dominant physical processes, ensuring that only a small number of feature parameters remain active at any given time, and that some of the activated feature parameters explicitly correspond to the dominant physical mechanism in the system; introducing a feature selection criterion based on energy contribution and prediction sensitivity to comprehensively evaluate the importance of the sparsified feature parameters, considering the proportion of each feature parameter in the overall physical energy and its sensitivity to system output prediction during the evaluation process; and sorting and pruning the feature parameters according to the comprehensive importance evaluation results, retaining key feature parameters and eliminating irrelevant feature parameters to form a low-dimensional sparse feature parameter set that is dimensionally controlled, time-consistent, physically meaningful, and reflects the essential dynamic behavior of the system.

[0088] This embodiment refines the core step S101, "extracting a set of low-dimensional sparse feature parameters reflecting dynamic behavior from a high-dimensional observation space." It addresses the issues of data redundancy in high-dimensional observation spaces, lack of physical interpretability of feature parameters, and heavy computational burden caused by excessive dimensionality. Through sparsification and physical meaning-oriented feature selection, a set of low-dimensional sparse feature parameters with controlled dimensionality, consistent temporal sequence, and clear physical meaning is formed.

[0089] Feature sparsity processing employs L1 regularization or structured sparsity constraints to sparsify the key feature modal time series data acquired within the time sliding window. Through regularization or constraints, the sparsity of feature parameters is enhanced, redundant features and noise interference information corresponding to non-dominant physical processes are eliminated, ensuring that only a small number of feature parameters remain active at any given time. Moreover, these activated feature parameters can explicitly correspond to the dominant physical mechanism in the system (such as energy transfer and momentum change), thus ensuring the physical interpretability of the feature parameters.

[0090] The comprehensive importance evaluation of features introduces a feature selection criterion based on energy contribution and prediction sensitivity to evaluate the comprehensive importance of feature parameters after sparsification. During the evaluation process, two core indicators are considered simultaneously: first, the proportion of each feature parameter in the overall physical energy (characterizing the degree to which the feature dominates the physical process); and second, the sensitivity of each feature parameter to the prediction of the system output (characterizing the degree to which the feature affects the prediction of the subsequent model). The comprehensive importance score of each feature parameter is obtained by weighted calculation of the two indicators.

[0091] Based on the comprehensive importance evaluation results, all sparsified feature parameters are sorted and pruned. Key feature parameters with high comprehensive importance scores are retained, while feature parameters with low scores, irrelevant to the dominant physical process, and with minimal impact on system prediction are removed. Through pruning, the dimensionality of the feature parameter set is controlled, ultimately forming a low-dimensional sparse feature parameter set that is dimensionally controlled, temporally consistent, physically meaningful, and reflects the essential dynamic behavior of the system. This provides low-redundancy, high-quality input for subsequent surrogate model construction.

[0092] In some embodiments, the step of constructing a reduced-order surrogate model using the low-dimensional sparse feature parameter set as input and employing Gaussian process regression, a neural network surrogate model, or intrinsic orthogonal decomposition combined with interpolation techniques includes: using the extracted low-dimensional sparse feature parameter set as the sole input to the reduced-order surrogate model, enabling the reduced-order surrogate model to perform mapping operations only within the physical dominant subspace, thereby controlling model complexity and ensuring real-time performance from the root; constructing a unified reduced-order mapping relationship, compressing sparse feature parameters to a fixed low-dimensional computational space through a feature hash mapping matrix, reducing the memory access and multiply-accumulate overhead of the reduced-order surrogate model; selecting any one of Gaussian process regression, lightweight neural networks, or interpolation operators based on intrinsic orthogonal decomposition as the main body of the reduced-order surrogate model, inputting the compressed low-dimensional sparse feature parameters into the main body of the reduced-order surrogate model, and training through samples to enable the main body of the reduced-order surrogate model to approximate the system response of a high-fidelity physics solver in the same feature space with the lowest computational cost, thus completing the construction of the reduced-order surrogate model.

[0093] This embodiment refines the core step S102, which involves "using the low-dimensional sparse feature parameter set as input and constructing a reduced-order surrogate model using Gaussian process regression, neural network surrogate model, or intrinsic orthogonal decomposition combined with interpolation techniques." It addresses the issues of high complexity, heavy computational burden, and difficulty in approximating high-fidelity physical solutions by unifying the reduced-order mapping and lightweight model design. This results in a reduced-order surrogate model that can approximate high-fidelity solution responses with low computational cost.

[0094] The model input setting uses the extracted low-dimensional sparse feature parameter set as the sole input to the reduced-order proxy model, ensuring that the reduced-order proxy model performs mapping operations only within the physical dominant subspace. This avoids redundant features and irrelevant information from entering the model, thereby controlling the model complexity from the root and providing a foundation for the model's real-time performance.

[0095] The unified reduction mapping relationship is constructed by introducing a feature hash mapping matrix to compress low-dimensional sparse feature parameters into a fixed low-dimensional computation space. Through feature hash mapping, the memory access overhead and multiply-accumulate operation overhead of the reduction proxy model are reduced, further simplifying the model's computational complexity and adapting it to environments with limited computing power.

[0096] The selection and construction of the surrogate model are based on the characteristics of the complex physical system (such as the intensity of nonlinearity and the amount of data). A suitable surrogate model is selected: if the system has weak nonlinearity and high prediction accuracy is required, Gaussian process regression is selected as the model; if the system has strong nonlinearity and a large amount of data is processed, a lightweight neural network (such as pruned CNN or LSTM) is selected as the model; if both physical meaning and computational efficiency need to be considered, an interpolation operator based on intrinsic orthogonal decomposition is selected as the model.

[0097] Model training and optimization involves inputting compressed low-dimensional sparse feature parameters into the selected surrogate model and using high-fidelity physics solutions or real system observation data as training samples to train the surrogate model. During training, the model parameters are adjusted with the goal of "approaching the system response of the high-fidelity physics solver in the same feature space with the lowest computational cost" to ensure that the surrogate model has high prediction accuracy while simplifying complexity, thus completing the construction of the reduced-order surrogate model.

[0098] In some embodiments, the lightweight design and real-time optimization of the downgraded proxy model includes: using at least one of feature hashing, quantization-aware training, or model pruning to perform lightweight design on the constructed downgraded proxy model; identifying and compressing the scale of non-critical parameters in the downgraded proxy model; eliminating redundant computational paths; enabling the downgraded proxy model to have deterministic computational latency during the deployment phase, and adapting to edge computing devices or environments with limited computing resources; configuring an online incremental learning function for the downgraded proxy model; setting a forgetting factor and learning rate consistent with the sparse feature update mechanism; enabling the downgraded proxy model to prioritize adapting to the latest physical state changes; and avoiding model structural instability caused by short-term fluctuations, thereby achieving real-time optimization of the downgraded proxy model.

[0099] This embodiment elaborates on the core step S102, "lightweight design and real-time optimization of the downgraded proxy model," and addresses the problems of the downgraded proxy model being difficult to deploy on edge computing devices, in environments with limited computing power, and the instability of the model caused by data distribution drift during long-term operation. Through lightweight technology and online incremental learning, the model achieves real-time optimization and long-term stable operation.

[0100] The lightweight design of the proxy model employs at least one technique among feature hashing, quantization-aware training, or model pruning to lightweight the constructed reduced-order proxy model. For example, model pruning identifies and removes non-critical parameters and redundant computational paths in the proxy model, compressing the model parameter scale; quantization-aware training quantizes floating-point parameters into fixed-point parameters, reducing model storage and computational overhead; and lightweight design ensures that the reduced-order proxy model has deterministic computational latency during deployment, making it adaptable to environments with limited computing resources such as edge computing devices and handheld terminals.

[0101] The online incremental learning function is configured by setting the forgetting factor and learning rate consistent with the sparse feature update mechanism in step S101. The forgetting factor is used to reduce the impact of early training samples on the current model update, so that the model can adapt to the latest physical state changes first. The learning rate is used to control the update magnitude of the model parameters, avoiding model oscillation caused by too fast updates and model lag caused by too slow updates. Through online incremental learning, the surrogate model can adaptively cope with the slow drift of data distribution in long-term operation and maintain the prediction accuracy of the model.

[0102] Real-time verification and adjustment: After the lightweight design and incremental learning configuration are completed, the real-time performance of the proxy model is verified. The inference speed of the model is tested on edge devices and in environments with limited computing power to ensure that millisecond-level inference can be achieved. If the inference speed does not meet the preset requirements, the lightweight parameters are further optimized (such as increasing the pruning ratio and optimizing the quantization accuracy) until the real-time requirements are met, thus completing the real-time optimization of the downgraded proxy model.

[0103] In some embodiments, activating a high-fidelity physics solver for a local region if a feature parameter mutation is detected based on a trigger threshold and exceeds the reliable prediction range of the reduced-order surrogate model includes: constructing a unified trigger criterion function, which simultaneously characterizes the evolution rate of the dominant physical process in the feature space and the prediction uncertainty of the reduced-order surrogate model output, and balancing the impact of physical mutation intensity and insufficient model cognition on triggering behavior through preset weight coefficients; comparing the calculation result of the constructed trigger criterion function with a preset trigger threshold, and if the calculation result exceeds the preset trigger threshold, determining that the feature parameter has mutated and exceeds the reliable prediction range of the reduced-order surrogate model; locating the abnormal feature parameter to a specific physical sub-region through a feature-to-space mapping operator, the mapping process combining sparse feature components and modal basis functions of key feature modes, and ensuring that the high-fidelity physics solver is activated only in the spatial region where the dominant mode has a significant effect through a preset energy contribution threshold; and calling any one of the high-fidelity physics solvers among the finite element, finite volume, or spectral methods to perform high-precision calculations only for the located physical sub-region and the corresponding time segment.

[0104] This embodiment refines the core step S103, "If a sudden change in the feature parameter is detected according to the trigger threshold and exceeds the reliable prediction range of the reduced-order surrogate model, activate the high-fidelity physics solver in the local region." It addresses the problems of high computational overhead in high-fidelity solving, the triggering mechanism being prone to false / missed triggering, and inaccurate physical sub-region positioning. By unifying the triggering criteria and precise spatial positioning, it achieves on-demand and precise activation of the high-fidelity solver.

[0105] The unified triggering criterion function is constructed by simultaneously characterizing two types of core information: first, the evolution rate of the dominant physical process in the feature space (reflecting the intensity of abrupt changes in feature parameters); and second, the prediction uncertainty of the output of the reduced-order surrogate model (reflecting the degree of model cognitive insufficiency). By using preset weight coefficients, the influence of the intensity of physical abrupt changes and the model cognitive insufficiency on the triggering behavior is balanced, enabling the criterion function to measure whether the current state exceeds the stable approximation capability of the surrogate model with a unified dimension. This avoids false triggering (such as triggering the solution due to slight fluctuations) or missed triggering (such as undetected abrupt changes) caused by relying on a single indicator.

[0106] The mutation detection and range determination process calculates the result of the trigger criterion function in real time and compares it with the system's preset trigger threshold. If the result of the criterion function exceeds the preset trigger threshold, it is determined that the feature parameter has mutated and the current state has exceeded the reliable prediction range of the reduced-order surrogate model, triggering the activation process of the high-fidelity physics solver. If the threshold is not exceeded, the real-time operation of the surrogate model is maintained, and the high-fidelity solver is not activated to reduce computational overhead.

[0107] Precise physical sub-region localization uses a feature-to-space mapping operator to locate anomalous feature parameters to specific physical sub-regions. During the mapping process, the energy contribution threshold is set by combining sparse feature components with the physically meaningful modal basis functions obtained in step S101. This ensures that the high-fidelity physical solver is activated only in spatial regions where the dominant mode has a significant effect and where feature mutations are obvious, thus avoiding irrelevant regions from being included in the solution range and further reducing computational overhead.

[0108] Local high-fidelity solution execution utilizes any high-fidelity physics solver from the finite element, finite volume, or spectral methods to perform high-precision calculations only on the located physical sub-region and the time segment corresponding to the abrupt change in characteristic parameters. During the solution process, it focuses on the core physical processes in the abrupt change region to obtain high-precision physical response results, providing a reliable truth reference for subsequent surrogate model updates.

[0109] In some embodiments, the process of integrating observation information corresponding to real-time observation data streams into the prediction output of a surrogate model using data assimilation techniques, dynamically correcting the prediction bias of the reduced-order surrogate model, and outputting prediction results with uncertainty quantification includes: continuously receiving real-time observation data streams during system operation; synchronously comparing the prediction results output by the reduced-order surrogate model based on sparse feature parameters with the real-time observation data streams; calculating the corresponding observation residuals, which are used to characterize the degree of prediction mismatch of the surrogate model under the current physical state; constructing a bias-aware data assimilation update operator, and updating the data based on the prediction uncertainty scalar output by the reduced-order surrogate model and... The observation system estimates the observation noise intensity online and calculates the time-dependent adaptive correction gain. This correction gain is used to adjust the correction intensity of the prediction results by the observation information, avoiding the amplification effect of observation noise on model stability. The observation residuals and the adaptive correction gain are substituted into the data assimilation update operator to dynamically correct the prediction results of the reduced-order surrogate model, resulting in a corrected system state estimate. The corrected system state estimate and the prediction uncertainty scalar output by the reduced-order surrogate model are output synchronously to form a prediction result with uncertainty quantification. The observation residuals are mapped back to the sparse feature space to form a bias-driven feature correction signal.

[0110] This embodiment refines the core step S104, which involves "integrating the observation information corresponding to the real-time observation data stream into the surrogate model's prediction output through data assimilation technology, dynamically correcting the prediction bias of the reduced-order surrogate model, and outputting prediction results with uncertainty quantification." It addresses the issues of accumulated bias, amplified observation noise, and unreliable prediction results during the long-term operation of the surrogate model. Through bias-aware data assimilation and adaptive correction, it achieves dynamic correction of prediction bias and reliable output.

[0111] Real-time observation and residual calculation continuously receive real-time observation data streams collected by a multi-source sensor array during system operation, ensuring that the observation data and the prediction results of the surrogate model are consistent in terms of time series and physical quantities. The prediction results output by the reduced-order surrogate model based on sparse feature parameters are compared with the synchronously acquired real-time observation data, and the observation residual between the two is calculated. This residual directly characterizes the degree of prediction mismatch of the surrogate model under the current physical state, providing a basis for bias correction.

[0112] The adaptive correction gain calculation constructs a bias-aware data assimilation update operator and calculates the time-related adaptive correction gain based on two types of core information: one is the prediction uncertainty scalar output of the reduced-order surrogate model (reflecting the cognitive risk brought about by model simplification and incomplete parameters), and the other is the observation noise intensity estimated online by the observation system (reflecting the reliability of real-time observation data). The magnitude of the correction gain is positively correlated with the prediction uncertainty and negatively correlated with the observation noise intensity. It is used to adjust the correction strength of the observation information on the prediction results and avoid the amplification effect of observation noise on model stability.

[0113] The prediction bias dynamic correction is achieved by substituting the calculated observation residuals and adaptive correction gain into the data assimilation update operator to dynamically correct the prediction results of the reduced-order surrogate model. The correction process is performed online in real time to ensure that the prediction results of the surrogate model can quickly fit the real physical state, eliminate the cumulative bias caused by model simplification assumptions and environmental disturbances, and obtain the corrected system state estimate.

[0114] Uncertainty quantification output and residual feedback are achieved by synchronously outputting the corrected system state estimate and the prediction uncertainty scalar output by the reduced-order surrogate model, forming a prediction result with uncertainty quantification, which provides a reliable basis for subsequent decision-making and control. At the same time, the observation residual is mapped back to the sparse feature space through the feature sensitivity matrix to form a bias-driven feature correction signal, which is used for parameter tuning of the subsequent sparse feature extraction algorithm and adaptive adjustment of the trigger threshold.

[0115] In some embodiments, the establishment of a fully closed-loop feedback chain from model prediction and high-fidelity solution verification to observation data correction, combined with a reduced-order surrogate model, to complete the construction of the digital twin, includes: unifying the prediction results of the reduced-order surrogate model, the reference truth value output by the high-fidelity physical solver, the correction state after data assimilation, and the observation residuals into the same closed-loop feedback framework; constructing a comprehensive feedback evaluation function, which comprehensively characterizes the true prediction error of the current model system, the prediction uncertainty of the surrogate model, and the triggering frequency of the high-fidelity solver; balancing prediction accuracy, model robustness, and computational cost through preset weight coefficients; and based on the calculation results and long-term evolution trend of the comprehensive feedback evaluation function, adjusting the sparseness corresponding to the low-dimensional sparse feature parameter set. The feature extraction algorithm performs parameter callbacks and adaptively adjusts the sparse constraint strength parameters, enabling the sparse feature parameter set to quickly respond to structural changes in the physical field while maintaining the ability to remember stable evolution patterns, thus conforming to the real dynamic evolution structure of the system. Using the calculation results of the comprehensive feedback evaluation function, the structural complexity of the reduced-order surrogate model is adaptively controlled. While ensuring that prediction errors and uncertainties meet the standards, the non-critical computational paths of the reduced-order surrogate model are continuously compressed, maintaining its lightweight characteristics. A fully closed-loop feedback chain covering sparse feature extraction, surrogate model prediction, high-fidelity solution verification, and observation data correction is established, enabling the corresponding digital twin to possess continuous self-optimization and adaptive capabilities, completing the dynamic construction of a high-fidelity digital twin.

[0116] This embodiment refines the core step S104, which involves "establishing a fully closed-loop feedback chain from model prediction and high-fidelity solution verification to observation data correction by combining a reduced-order proxy model to complete the construction of the digital twin." It addresses the core issues of the digital twin's lack of self-optimization capabilities and the inability of the feature set and model structure to adapt to dynamic changes in the system. Through a fully closed-loop feedback framework and collaborative optimization, it achieves the self-evolution and high-fidelity construction of the digital twin.

[0117] The fully closed-loop feedback framework is constructed by unifying the prediction results of the reduced-order surrogate model, the reference truth output of the high-fidelity physics solver, the corrected state after data assimilation, and the observation residuals into a single closed-loop feedback framework. A comprehensive feedback evaluation function is constructed, which comprehensively characterizes three core indicators: the true prediction error of the current model system (the deviation between the corrected state and the high-fidelity reference truth), the prediction uncertainty of the surrogate model (reflecting insufficient model cognition), and the triggering frequency of the high-fidelity solver (reflecting the sufficiency of the surrogate model in covering the physical state). Through preset weight coefficients, prediction accuracy, model robustness, and computational cost are balanced, and the overall performance of the current features and model configuration is quantified in a single scalar form.

[0118] The parameter callback of the sparse feature extraction algorithm is performed on the sparse feature extraction algorithm corresponding to the low-dimensional sparse feature parameter set by calculating the results of the comprehensive feedback evaluation function and the long-term evolution trend. For example, when the comprehensive evaluation function value is too high (large prediction error, high uncertainty, abnormal triggering frequency), the sparse constraint strength parameter is adaptively adjusted to reduce the sparse constraint, so that the feature parameter set can quickly respond to the structural changes of the physical field. At the same time, by adjusting the forgetting factor and feature selection weight, the ability to remember the stable evolution mode is maintained, ensuring that the feature parameter set fits the real dynamic evolution structure of the system.

[0119] The adaptive control of the reduced-order surrogate model structure adaptively controls the structural complexity of the reduced-order surrogate model by utilizing the calculation results of the comprehensive feedback evaluation function. Under the premise of ensuring that the prediction error and uncertainty meet the standards, the non-critical calculation paths and redundant parameters of the surrogate model are continuously compressed to maintain the lightweight characteristics of the surrogate model. If the prediction error is too large, the parameter scale of the critical calculation path is appropriately increased to improve the model prediction accuracy and achieve a dynamic balance between model accuracy and lightweight.

[0120] The formation of the fully closed-loop link and the construction of the digital twin are achieved by integrating the above-mentioned parameter callback and model optimization processes to establish a fully closed-loop feedback link covering sparse feature extraction, surrogate model prediction, high-fidelity solution verification, and observation data correction. Through this link, the sparse feature set and the surrogate model structure can evolve co-evolved. High-fidelity solution and real-time observation are no longer isolated verification tools, but key signal sources driving model and feature optimization. After multiple rounds of iterative optimization, the digital twin has continuous self-optimization and self-adaptation capabilities, and can map the operating state of complex physical systems in real time and accurately, thus completing the dynamic construction of a high-fidelity digital twin.

[0121] In some embodiments, to address problems existing in the prior art, such as Figure 2 , Figure 3 , Figure 4 and Figure 5 As shown, this scheme proposes a method for dynamically constructing a high-fidelity digital twin based on multi-source data fusion and physical simulation, including the following steps:

[0122] Step 1: Adaptively extract sparse feature parameters of the dominant physical processes from massive multi-source data.

[0123] For heterogeneous data streams from multiple sources such as sensor arrays, numerical simulations, and historical databases, adaptive algorithms such as dynamic mode decomposition, sparse coding, or nonlinear independent component analysis are used to identify key feature modes that dominate the physical processes in the data in real time. By introducing a time sliding window and a forgetting mechanism, the evolution of time-varying physical fields is tracked online. Using L1 regularization or structured sparse constraints, a low-dimensional sparse feature parameter set that reflects the essential dynamic behavior of the system is extracted from the high-dimensional observation space. This parameter set has physical interpretability and can significantly reduce the dimensionality of subsequent processing.

[0124] Step 2: Construct a reduced-order proxy model optimized for real-time computing based on sparse features.

[0125] Using the aforementioned sparse feature parameters as input, a reduced-order surrogate model of the system response is constructed using Gaussian process regression, neural network surrogate model, or intrinsic orthogonal decomposition combined with interpolation techniques. Through feature hashing, quantization-aware training, or model pruning techniques, the surrogate model is designed to be lightweight and optimized for real-time performance, enabling it to achieve millisecond-level inference on edge computing devices or in environments with limited computing resources. At the same time, the model structure supports online incremental learning to adapt to the slow drift of data distribution.

[0126] Step 3: Trigger the local high-fidelity physics solver only when a feature parameter undergoes a sudden change.

[0127] A trigger threshold based on the rate of change of sparse features or uncertainty estimation is set. When a sudden change in feature parameters is detected and exceeds the reliable prediction range of the surrogate model, a high-fidelity physics solver for the local region is automatically activated. This solver performs high-precision calculations only for the physical sub-regions and time segments where the sudden change occurs, and feeds the results back to the model update process as true values. This significantly reduces the computational burden of global high-fidelity solving while ensuring the computational accuracy of key regions.

[0128] Step 4: Dynamically correct the prediction bias of the surrogate model using real-time observation data.

[0129] During system operation, real-time observation data streams are continuously received. Through data assimilation technology, the observation information is integrated into the prediction output of the surrogate model, dynamically correcting the prediction bias caused by simplification assumptions or environmental disturbances. This correction process is performed online to ensure that the surrogate model always remains consistent with the current actual physical state and outputs prediction results with uncertainty quantification.

[0130] Step 5: Iteratively optimize the sparse feature set and surrogate model structure through closed-loop feedback.

[0131] A closed-loop feedback chain is established, from model prediction and high-fidelity solution verification to observation data correction. Based on information such as prediction error, uncertainty index and trigger frequency, the parameters of the sparse feature extraction algorithm are tuned periodically or triggered, and the structure and hyperparameters of the surrogate model are retrained and optimized. Through continuous iteration, the sparse feature set can better capture the dynamic evolution characteristics of the system, and the surrogate model can further improve prediction accuracy and robustness while maintaining lightweightness, forming a self-evolving modeling and computing system.

[0132] Through the above steps, a complete technical loop is achieved, from massive data to sparse features, from reduced-order models to local high-fidelity solutions, and from bias correction to model self-optimization, which significantly improves the efficiency and reliability of real-time simulation and prediction of complex physical processes.

[0133] In step one, multi-source data is continuously fed in with a unified time reference, and physical process consistency constraints are introduced at the data layer to avoid interference from heterogeneous scales in feature extraction. For the joint observation sequence formed by the sensor array, numerical simulation, and historical database, a state matrix that is updated over time is first constructed, and modal decoupling is performed on it within a sliding time window. By constraining the consistency of dynamic evolution between adjacent time slices, it is ensured that the extraction results stably reflect the evolution trajectory of the dominant physical process. Based on this, the joint sparse mode decomposition objective function is defined:

[0134] ;

[0135] in This represents the fused multi-source observation matrix within the current time window. This represents a shared modal basis with a definite physical meaning. This represents the feature parameter vector corresponding to the time slice. Used to enhance the sparsity of feature parameters to eliminate non-dominant physical processes. This objective function constrains the continuous evolution of features over time to avoid transient noise interference. While ensuring reconstruction accuracy, it ensures that only a small number of feature parameters remain active at any given time, thus explicitly corresponding to the dominant physical mechanism in the system. To adapt to the slow changes in the distribution of physical states during long-term operation, an online update mechanism with forgetting weights is introduced, which exponentially decays the contribution of historical features. The update rule is expressed as follows:

[0136] ;

[0137] in This is a forgetting factor used to reduce the influence of early states on the current feature estimation. The strength of the historical consistency constraint is controlled to enable feature parameters to respond quickly to structural changes in the physical field while maintaining their ability to remember stable evolution patterns. Furthermore, to ensure the extracted features possess physical interpretability and usability for subsequent modeling, a feature selection criterion based on energy contribution and prediction sensitivity is introduced to rank and prune sparse feature parameters. The evaluation index is defined as follows:

[0138] ;

[0139] in Indicates the first The overall importance of each feature parameter is determined by two terms: the first characterizes its proportion in the overall physical energy, and the second characterizes its sensitivity to system output prediction. This metric ensures that the retained features both dominate the physical process and directly influence the system response. Through this mechanism, a sparse feature parameter set with controlled dimensions, consistent temporality, and clear physical meaning is ultimately formed, providing a stable and low-redundancy input foundation for subsequently constructing a reduced-order surrogate model capable of real-time inference and supporting incremental updates.

[0140] In step two, the sparse feature parameter set obtained in the previous step is used as the sole input, ensuring that the surrogate model maps only within the physically dominant subspace, thereby controlling model complexity and guaranteeing real-time performance from the root. Let's assume at time... The obtained sparse feature vector is Its dimensions are determined by the comprehensive importance index in step one. After screening, it was determined that the goal of the surrogate model is to approximate the system response of the high-fidelity physics solver in the same feature space with the lowest computational cost. To this end, a unified reduced-order mapping form is constructed:

[0141] ;

[0142] in Indicates the proxy model at time... The prediction results of the system response, This represents a feature hash mapping matrix, used to convert sparse features... Compressing to a fixed low-dimensional computational space reduces memory access and multiply-accumulate overhead. The parameter is The proxy model body supports Gaussian process regression, lightweight neural networks, or interpolation operators based on intrinsic orthogonal decomposition. The purpose of this expression is to unify feature dimensionality reduction and response prediction into a single mapping, ensuring that all subsequent calculations revolve around a small number of effective degrees of freedom, thereby meeting the real-time inference requirements of edge devices. To further reduce computational burden and suppress redundant parameters, a real-time-oriented structural sparsity training objective is introduced:

[0143] ;

[0144] in This represents the true value of the system response obtained from high-fidelity solutions or actual observations. Used to enhance the sparsity of model parameters to support subsequent pruning. Used to constrain parameter stability on critical computational paths This is a structure mask matrix, whose elements are determined by feature importance. The sign is determined together with historical prediction errors. This represents element-wise multiplication. The objective function ensures that the surrogate model automatically compresses the scale of non-critical parameters while maintaining the required approximation accuracy, thus providing deterministic computational latency during deployment. To address the slow drift of the physical state distribution over long-term operation, the surrogate model introduces an incremental update rule consistent with the time-forgetting mechanism in step one:

[0145] ;

[0146] in Indicates the model parameters at the current time. Indicates the learning rate. Forgetting factor consistent with sparse feature updates This represents the time interval between the current sample and the previous update. This update method allows the model to prioritize adapting to the latest physical state while avoiding structural instability caused by short-term fluctuations. Through the above modeling and training mechanism, the surrogate model forms a fast approximate mapping of the system response in the sparse feature space and simultaneously provides the prediction residual and parameter uncertainty assessment at the output. This uncertainty will serve as the core criterion for whether to trigger the local high-fidelity physics solver in the next step, thereby achieving a dynamic balance between computational accuracy and resource consumption.

[0147] In step three, the system no longer performs high-fidelity numerical solutions on the global physical field, but instead uses the sparse characteristic parameters obtained in the previous steps. The uncertainty information output by the proxy model is used as the sole triggering criterion to achieve precise allocation of computing resources. To simultaneously characterize both abrupt changes in physical state and a decrease in model credibility, a unified triggering criterion function is constructed:

[0148] ;

[0149] in and These represent the sparse feature vectors at adjacent time points, Representing the corresponding time interval, the first term characterizes the evolution rate of the dominant physical process in the feature space. This represents the scalar of prediction uncertainty output synchronously by the surrogate model in step two. , representing a weighting coefficient, is used to balance the impact of physical mutation intensity and model cognitive limitations on triggering behavior. The purpose of this function is to measure, using a unified scale, whether the current state exceeds the stable approximation capability of the surrogate model within the sparse feature subspace, thereby avoiding false triggers or missed triggers caused by relying on a single metric. When Exceeding the system's set threshold At this point, the triggering mechanism is activated, and the anomaly is further located to a specific physical sub-region through the feature-to-space mapping operator. This mapping relationship is defined as:

[0150] ;

[0151] in This indicates the physical subregion for which a high-fidelity solution needs to be performed. Represents the position coordinates in physical space. For the first sparse feature components These are the physically meaningful modal basis functions obtained in step one. An energy contribution threshold is defined to ensure that high-precision computations are initiated only within spatial regions where the dominant mode has a significant effect. This definition directly couples the computational domain of the high-fidelity solver with sparse features, preventing irrelevant regions from being included in the solution. Subsequently, Within the corresponding time segment, the finite element method, finite volume method, or spectral method solver is invoked to obtain a high-precision physical response. This is fed back to the surrogate model update channel in the form of residuals, and the feedback amount is defined as:

[0152] ;

[0153] in This refers to the prediction results of the surrogate model in step two under the same feature input. This directly characterizes the degree of mismatch in the surrogate model under abrupt changes. The residual not only serves as a truth constraint for updating the surrogate model parameters but is also used to correct feature importance assessments and adaptively adjust subsequent trigger thresholds, thus achieving a dynamic balance between computational accuracy and trigger frequency. Through this on-demand triggering and local solution mechanism, the system relies on a lightweight surrogate model during most stable evolutionary phases, while automatically switching to a high-fidelity computation mode at critical physical process abrupt changes. This provides a highly reliable reference state for data assimilation and prediction bias correction based on real-time observations in the next step.

[0154] In step four, the prediction results of the surrogate model are no longer directly used as the system state output, but are instead integrated online with the real-time observation data, thereby continuously suppressing the cumulative prediction bias without compromising the real-time performance of step two. Let's assume that at time... Based on sparse features by proxy model The predicted results are The synchronous observation data obtained from multiple source sensors is Both maintain a consistent dimension in the physical quantity space. By constructing a bias-aware data assimilation update operator, the predicted state is dynamically corrected. The unified correction expression is defined as follows:

[0155] ;

[0156] in This represents the corrected system state estimate. Representing the observation residuals, it directly characterizes the degree of prediction mismatch of the surrogate model under the current physical conditions. This is the time-dependent adaptive correction gain, used to adjust the strength of the correction to the prediction results based on observational information. The purpose of this formula is to establish a controlled feedback channel between the surrogate model output and the actual observations, enabling the model to quickly conform to the real physical state while maintaining smooth evolution. To avoid the amplifying effect of observational noise on model stability, the correction gain... It is not a fixed setting, but is jointly determined by the uncertainty of the surrogate model and the reliability of the observation. Its calculation method is defined as follows:

[0157] ;

[0158] in This indicates that the proxy model in step two is for... The given scalar of prediction uncertainty reflects the cognitive risks arising from model simplification and incomplete parameters. This represents the observation noise intensity estimated online by the observation system, used to characterize the reliability of real-time data. This design ensures that when the surrogate model is under high uncertainty or has just undergone the transition phase triggered by the high-fidelity solution in step three, the observation data dominates the correction process. Conversely, when the model is stable and the observation noise is high, the correction process automatically becomes conservative, thus guaranteeing the numerical stability of the overall estimation process. Furthermore, to use the correction results to support the adaptive evolution of the model, the observation residuals are explicitly mapped back to the sparse feature space, forming a bias-driven feature correction signal. The mapping relationship is defined as follows:

[0159] ;

[0160] in This represents the correction amount at the feature level. This indicates that the proxy model is in the current parameters The sensitivity matrix from features to response obtained through implicit learning is transposed to feed the bias information in the output space back to the sparse feature space. This correction amount does not directly change the current feature value, but rather participates as a weight signal in subsequent feature importance updates, uncertainty assessments, and trigger threshold adjustments. This allows the data assimilation process to not only correct the current prediction results but also continuously influence the long-term evolution of the model and feature system. Through this online correction mechanism, the surrogate model output remains consistent with the real physical state at every moment and simultaneously carries a reliable uncertainty measure, providing a stable and quantifiable source of error information for the next step of sparse feature set optimization based on closed-loop feedback and the self-evolution of the surrogate model structure.

[0161] In step five, the system will use the prediction results generated in steps two through four. High-fidelity solution of residuals and the correction status after data assimilation A unified feedback framework is adopted to co-evolve the sparse feature set and the surrogate model structure, rather than optimizing them in isolation. To quantify the current modeling system's ability to depict the true dynamics of the system, a comprehensive feedback evaluation function is first constructed:

[0162] ;

[0163] in This represents the cost of closed-loop feedback. This is the system state estimate after correction in step four. The first term, representing the reference true value obtained when triggering the high-fidelity physics solution in step three, characterizes the actual prediction error of the current model system. The uncertainty in the surrogate model's output at this moment reflects the lack of understanding at the structural and parameter levels. This indicates the triggering frequency of the high-fidelity solver within the statistical window, used to measure the adequacy of the surrogate model's coverage of the physical state space. , and These are the weighting coefficients used to balance accuracy, robustness, and computational cost. The purpose of this evaluation function is to characterize the overall performance of the current features and model configuration in terms of accuracy and resource consumption as a single scalar, providing a clear direction for subsequent adaptive tuning. Based on Based on the long-term evolution trend, the sparse feature extraction mechanism in step one is subject to parameter callback, and its core update rule is defined as:

[0164] ;

[0165] in This represents the current sparse constraint strength parameter. This represents the feature importance vector constructed from the comprehensive importance index in step one. To adjust the rate coefficient, this update mechanism automatically adjusts the sparsity constraint strength as the prediction error and trigger frequency continue to increase, thereby introducing new physical dominant features or suppressing redundant modes, ensuring that the feature set always closely matches the system's true dynamic evolution structure. At the surrogate model level, the same feedback cost is used to adaptively control the model's structural complexity, and its structural update criterion is defined as:

[0166] ;

[0167] in This represents the set of parameters for the proxy model. This indicates the number of effective parameters in the model, used to directly constrain the structural size. As a complexity penalty coefficient, this optimization objective continuously compresses unnecessary computational paths while reducing prediction errors and uncertainties, thus maintaining lightweight characteristics. Through the aforementioned two-layer feedback mechanism, the sparse feature set and the surrogate model structure are jointly driven to evolve in the same closed loop. High-fidelity solving and real-time observation are no longer just passive verification tools, but become important signal sources shaping the model structure and feature representation. After multiple iterations, the system gradually forms a feature representation and computational structure that highly matches complex physical processes, laying the foundation for further expansion to cross-scene transfer, multi-twin collaboration, or group-level digital twin management.

[0168] Please see Figure 6 As shown, Figure 6 This is a schematic diagram of the structure of a dynamic digital twin construction system 200 based on multi-source data fusion and physical simulation provided in this application embodiment. This dynamic digital twin construction system 200 is used to execute the steps of the dynamic digital twin construction method based on multi-source data fusion and physical simulation shown in the above embodiments. The dynamic digital twin construction system 200 can be a single server or a server cluster, or it can be a terminal, such as a handheld terminal, laptop computer, wearable device, or robot.

[0169] like Figure 6 As shown, the digital twin dynamic construction system 200 based on multi-source data fusion and physical simulation includes:

[0170] The data acquisition unit 201 is used to acquire multi-source heterogeneous data streams from preset sensor arrays, numerical simulation results and historical databases, identify key feature modes that dominate physical processes in multi-source heterogeneous data streams, use time sliding windows and forgetting mechanisms to obtain the time-varying physical field evolution law of key feature modes corresponding to key feature modes, and extract low-dimensional sparse feature parameter sets reflecting dynamic behavior from high-dimensional observation space.

[0171] The model building unit 202 is used to construct a reduced-order surrogate model by taking the low-dimensional sparse feature parameter set as input and using Gaussian process regression, neural network surrogate model or intrinsic orthogonal decomposition combined with interpolation technology, and to perform lightweight design and real-time optimization of the reduced-order surrogate model.

[0172] The threshold setting unit 203 is used to set a trigger threshold based on the sparse feature change rate or uncertainty estimation. If a feature parameter mutation is detected according to the trigger threshold and exceeds the reliable prediction range of the reduced-order surrogate model, the high-fidelity physics solver of the local region is activated. The high-fidelity physics solver is used to calculate for the physical sub-region and time segment where the mutation occurred, and the result is fed back to the corresponding model update process as the truth value.

[0173] The data receiving unit 204 is used to receive the corresponding real-time observation data stream. Through data assimilation technology, the observation information corresponding to the real-time observation data stream is integrated into the prediction output of the surrogate model. The prediction bias of the reduced-order surrogate model is dynamically corrected, and the prediction result with uncertainty quantification is output. In order to combine the reduced-order surrogate model, a closed-loop feedback link is established from model prediction, high-fidelity solution verification to observation data correction, so as to complete the construction of the digital twin.

[0174] In some embodiments, after the construction of the digital twin is completed, the method further includes: periodically or triggerively optimizing the sparse feature extraction algorithm corresponding to the low-dimensional sparse feature parameter set based on the prediction error, uncertainty index and the frequency of triggering the high-fidelity physical solver generated by the digital twin during system operation, and retraining and optimizing the structure and hyperparameters of the reduced-order surrogate model, so as to improve the ability of the sparse feature set to capture the dynamic evolution characteristics of the system and the prediction accuracy and robustness of the surrogate model.

[0175] In some embodiments, identifying key feature modes that dominate physical processes in multi-source heterogeneous data streams includes: preprocessing the acquired multi-source heterogeneous data streams, unifying real-time data collected by sensor arrays, simulated data obtained from numerical simulations, and archived data in historical databases to the same time reference, and introducing physical process consistency constraints at the data layer to eliminate interference caused by heterogeneous scales to feature identification; constructing a state matrix that is updated over time, performing mode decoupling processing on the state matrix within a preset time window, and constraining the consistency of dynamic evolution between adjacent time slices; and using any one of the adaptive algorithms, such as dynamic mode decomposition, sparse coding, or nonlinear independent component analysis, to analyze the results after mode decoupling in order to identify key feature modes that can reflect the dominant physical processes of the system and have clear physical meanings, and to eliminate redundant modes related to non-dominant physical processes.

[0176] In some embodiments, the method of acquiring the time-varying physical field evolution law of key feature modes corresponding to key feature modes using a time sliding window and forgetting mechanism includes: setting a time sliding window of fixed duration, allowing the time sliding window to continuously scroll with the time process, and capturing the time-series data of key feature modes in different time segments in real time; introducing an online update mechanism with forgetting weights to exponentially decay the contribution of key feature modes at historical moments within the time sliding window, and reducing the impact of early state data on the current feature estimation through a preset forgetting factor; and simultaneously setting a historical consistency constraint strength parameter to control the continuous evolution trend of key feature modes over time, avoiding interference caused by transient noise to the tracking of evolution law, thereby realizing online tracking and accurate acquisition of the time-varying physical field evolution law corresponding to key feature modes.

[0177] In some embodiments, the extraction of a low-dimensional sparse feature parameter set reflecting dynamic behavior from a high-dimensional observation space includes: using L1 regularization or structured sparsity constraints to sparsify the time-series data of key feature modes acquired within a time sliding window, enhancing the sparsity of feature parameters, eliminating redundant features and interference information corresponding to non-dominant physical processes, ensuring that only a small number of feature parameters remain active at any given time, and that some of the activated feature parameters explicitly correspond to the dominant physical mechanism in the system; introducing a feature selection criterion based on energy contribution and prediction sensitivity to comprehensively evaluate the importance of the sparsified feature parameters, considering the proportion of each feature parameter in the overall physical energy and its sensitivity to system output prediction during the evaluation process; and sorting and pruning the feature parameters according to the comprehensive importance evaluation results, retaining key feature parameters and eliminating irrelevant feature parameters to form a low-dimensional sparse feature parameter set that is dimensionally controlled, time-consistent, physically meaningful, and reflects the essential dynamic behavior of the system.

[0178] In some embodiments, the step of constructing a reduced-order surrogate model using the low-dimensional sparse feature parameter set as input and employing Gaussian process regression, a neural network surrogate model, or intrinsic orthogonal decomposition combined with interpolation techniques includes: using the extracted low-dimensional sparse feature parameter set as the sole input to the reduced-order surrogate model, enabling the reduced-order surrogate model to perform mapping operations only within the physical dominant subspace, thereby controlling model complexity and ensuring real-time performance from the root; constructing a unified reduced-order mapping relationship, compressing sparse feature parameters to a fixed low-dimensional computational space through a feature hash mapping matrix, reducing the memory access and multiply-accumulate overhead of the reduced-order surrogate model; selecting any one of Gaussian process regression, lightweight neural networks, or interpolation operators based on intrinsic orthogonal decomposition as the main body of the reduced-order surrogate model, inputting the compressed low-dimensional sparse feature parameters into the main body of the reduced-order surrogate model, and training through samples to enable the main body of the reduced-order surrogate model to approximate the system response of a high-fidelity physics solver in the same feature space with the lowest computational cost, thus completing the construction of the reduced-order surrogate model.

[0179] In some embodiments, the lightweight design and real-time optimization of the downgraded proxy model includes: using at least one of feature hashing, quantization-aware training, or model pruning to perform lightweight design on the constructed downgraded proxy model; identifying and compressing the scale of non-critical parameters in the downgraded proxy model; eliminating redundant computational paths; enabling the downgraded proxy model to have deterministic computational latency during the deployment phase, and adapting to edge computing devices or environments with limited computing resources; configuring an online incremental learning function for the downgraded proxy model; setting a forgetting factor and learning rate consistent with the sparse feature update mechanism; enabling the downgraded proxy model to prioritize adapting to the latest physical state changes; and avoiding model structural instability caused by short-term fluctuations, thereby achieving real-time optimization of the downgraded proxy model.

[0180] In some embodiments, activating a high-fidelity physics solver for a local region if a feature parameter mutation is detected based on a trigger threshold and exceeds the reliable prediction range of the reduced-order surrogate model includes: constructing a unified trigger criterion function, which simultaneously characterizes the evolution rate of the dominant physical process in the feature space and the prediction uncertainty of the reduced-order surrogate model output, and balancing the impact of physical mutation intensity and insufficient model cognition on triggering behavior through preset weight coefficients; comparing the calculation result of the constructed trigger criterion function with a preset trigger threshold, and if the calculation result exceeds the preset trigger threshold, determining that the feature parameter has mutated and exceeds the reliable prediction range of the reduced-order surrogate model; locating the abnormal feature parameter to a specific physical sub-region through a feature-to-space mapping operator, the mapping process combining sparse feature components and modal basis functions of key feature modes, and ensuring that the high-fidelity physics solver is activated only in the spatial region where the dominant mode has a significant effect through a preset energy contribution threshold; and calling any one of the high-fidelity physics solvers among the finite element, finite volume, or spectral methods to perform high-precision calculations only for the located physical sub-region and the corresponding time segment.

[0181] In some embodiments, the process of integrating observation information corresponding to real-time observation data streams into the prediction output of a surrogate model using data assimilation techniques, dynamically correcting the prediction bias of the reduced-order surrogate model, and outputting prediction results with uncertainty quantification includes: continuously receiving real-time observation data streams during system operation; synchronously comparing the prediction results output by the reduced-order surrogate model based on sparse feature parameters with the real-time observation data streams; calculating the corresponding observation residuals, which are used to characterize the degree of prediction mismatch of the surrogate model under the current physical state; constructing a bias-aware data assimilation update operator, and updating the data based on the prediction uncertainty scalar output by the reduced-order surrogate model and... The observation system estimates the observation noise intensity online and calculates the time-dependent adaptive correction gain. This correction gain is used to adjust the correction intensity of the prediction results by the observation information, avoiding the amplification effect of observation noise on model stability. The observation residuals and the adaptive correction gain are substituted into the data assimilation update operator to dynamically correct the prediction results of the reduced-order surrogate model, resulting in a corrected system state estimate. The corrected system state estimate and the prediction uncertainty scalar output by the reduced-order surrogate model are output synchronously to form a prediction result with uncertainty quantification. The observation residuals are mapped back to the sparse feature space to form a bias-driven feature correction signal.

[0182] In some embodiments, the establishment of a fully closed-loop feedback chain from model prediction and high-fidelity solution verification to observation data correction, combined with a reduced-order surrogate model, to complete the construction of the digital twin, includes: unifying the prediction results of the reduced-order surrogate model, the reference truth value output by the high-fidelity physical solver, the correction state after data assimilation, and the observation residuals into the same closed-loop feedback framework; constructing a comprehensive feedback evaluation function, which comprehensively characterizes the true prediction error of the current model system, the prediction uncertainty of the surrogate model, and the triggering frequency of the high-fidelity solver; balancing prediction accuracy, model robustness, and computational cost through preset weight coefficients; and based on the calculation results and long-term evolution trend of the comprehensive feedback evaluation function, adjusting the sparseness corresponding to the low-dimensional sparse feature parameter set. The feature extraction algorithm performs parameter callbacks and adaptively adjusts the sparse constraint strength parameters, enabling the sparse feature parameter set to quickly respond to structural changes in the physical field while maintaining the ability to remember stable evolution patterns, thus conforming to the real dynamic evolution structure of the system. Using the calculation results of the comprehensive feedback evaluation function, the structural complexity of the reduced-order surrogate model is adaptively controlled. While ensuring that prediction errors and uncertainties meet the standards, the non-critical computational paths of the reduced-order surrogate model are continuously compressed, maintaining its lightweight characteristics. A fully closed-loop feedback chain covering sparse feature extraction, surrogate model prediction, high-fidelity solution verification, and observation data correction is established, enabling the corresponding digital twin to possess continuous self-optimization and adaptive capabilities, completing the dynamic construction of a high-fidelity digital twin.

[0183] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the digital twin dynamic construction system and its modules based on multi-source data fusion and physical simulation described above can be found in the corresponding contents of the various embodiments of the digital twin dynamic construction method based on multi-source data fusion and physical simulation, and will not be repeated here.

[0184] The aforementioned method for dynamically constructing digital twins based on multi-source data fusion and physical simulation can be implemented as a computer program, which can be used in various ways, such as... Figure 6 It runs on the device shown.

[0185] Please see Figure 7 , Figure 7 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application. The computer device includes a processor, a memory, and a network interface connected via a device bus, wherein the memory may include a storage medium and internal memory.

[0186] The storage medium can store operating devices and computer programs. The computer program includes program instructions that, when executed, cause the processor to perform any method for dynamically constructing a digital twin based on multi-source data fusion and physical simulation.

[0187] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0188] Internal memory provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to execute any dynamic construction method of digital twins based on multi-source data fusion and physical simulation.

[0189] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the terminal to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0190] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0191] In one embodiment, the processor is configured to run a computer program stored in memory to perform the following steps:

[0192] Acquire multi-source heterogeneous data streams from preset sensor arrays, numerical simulation results, and historical databases; identify key feature modes that dominate physical processes in the multi-source heterogeneous data streams; utilize time sliding windows and forgetting mechanisms to obtain the time-varying physical field evolution laws of key feature modes corresponding to key feature modes; and extract low-dimensional sparse feature parameter sets reflecting dynamic behavior from high-dimensional observation space.

[0193] Using the low-dimensional sparse feature parameter set as input, a reduced-order surrogate model is constructed by employing Gaussian process regression, neural network surrogate model, or intrinsic orthogonal decomposition combined with interpolation techniques. The reduced-order surrogate model is then designed for lightweight performance and optimized for real-time performance.

[0194] Set a trigger threshold based on the rate of change of sparse features or uncertainty estimation. If a sudden change in feature parameters is detected according to the trigger threshold and exceeds the reliable prediction range of the reduced-order surrogate model, activate the high-fidelity physics solver for the local region. The high-fidelity physics solver is used to calculate for the physical sub-region and time segment where the sudden change occurred, and the result is fed back as the truth value to the corresponding model update process.

[0195] The system receives the corresponding real-time observation data stream, integrates the observation information corresponding to the real-time observation data stream into the prediction output of the surrogate model through data assimilation technology, dynamically corrects the prediction bias of the reduced-order surrogate model, and outputs prediction results with uncertainty quantification. This allows for the establishment of a fully closed-loop feedback link from model prediction, high-fidelity solution verification to observation data correction, in conjunction with the reduced-order surrogate model, thus completing the construction of the digital twin.

[0196] In some embodiments, after the construction of the digital twin is completed, the method further includes: periodically or triggerively optimizing the sparse feature extraction algorithm corresponding to the low-dimensional sparse feature parameter set based on the prediction error, uncertainty index and the frequency of triggering the high-fidelity physical solver generated by the digital twin during system operation, and retraining and optimizing the structure and hyperparameters of the reduced-order surrogate model, so as to improve the ability of the sparse feature set to capture the dynamic evolution characteristics of the system and the prediction accuracy and robustness of the surrogate model.

[0197] In some embodiments, identifying key feature modes that dominate physical processes in multi-source heterogeneous data streams includes: preprocessing the acquired multi-source heterogeneous data streams, unifying real-time data collected by sensor arrays, simulated data obtained from numerical simulations, and archived data in historical databases to the same time reference, and introducing physical process consistency constraints at the data layer to eliminate interference caused by heterogeneous scales to feature identification; constructing a state matrix that is updated over time, performing mode decoupling processing on the state matrix within a preset time window, and constraining the consistency of dynamic evolution between adjacent time slices; and using any one of the adaptive algorithms, such as dynamic mode decomposition, sparse coding, or nonlinear independent component analysis, to analyze the results after mode decoupling in order to identify key feature modes that can reflect the dominant physical processes of the system and have clear physical meanings, and to eliminate redundant modes related to non-dominant physical processes.

[0198] In some embodiments, the method of acquiring the time-varying physical field evolution law of key feature modes corresponding to key feature modes using a time sliding window and forgetting mechanism includes: setting a time sliding window of fixed duration, allowing the time sliding window to continuously scroll with the time process, and capturing the time-series data of key feature modes in different time segments in real time; introducing an online update mechanism with forgetting weights to exponentially decay the contribution of key feature modes at historical moments within the time sliding window, and reducing the impact of early state data on the current feature estimation through a preset forgetting factor; and simultaneously setting a historical consistency constraint strength parameter to control the continuous evolution trend of key feature modes over time, avoiding interference caused by transient noise to the tracking of evolution law, thereby realizing online tracking and accurate acquisition of the time-varying physical field evolution law corresponding to key feature modes.

[0199] In some embodiments, the extraction of a low-dimensional sparse feature parameter set reflecting dynamic behavior from a high-dimensional observation space includes: using L1 regularization or structured sparsity constraints to sparsify the time-series data of key feature modes acquired within a time sliding window, enhancing the sparsity of feature parameters, eliminating redundant features and interference information corresponding to non-dominant physical processes, ensuring that only a small number of feature parameters remain active at any given time, and that some of the activated feature parameters explicitly correspond to the dominant physical mechanism in the system; introducing a feature selection criterion based on energy contribution and prediction sensitivity to comprehensively evaluate the importance of the sparsified feature parameters, considering the proportion of each feature parameter in the overall physical energy and its sensitivity to system output prediction during the evaluation process; and sorting and pruning the feature parameters according to the comprehensive importance evaluation results, retaining key feature parameters and eliminating irrelevant feature parameters to form a low-dimensional sparse feature parameter set that is dimensionally controlled, time-consistent, physically meaningful, and reflects the essential dynamic behavior of the system.

[0200] In some embodiments, the step of constructing a reduced-order surrogate model using the low-dimensional sparse feature parameter set as input and employing Gaussian process regression, a neural network surrogate model, or intrinsic orthogonal decomposition combined with interpolation techniques includes: using the extracted low-dimensional sparse feature parameter set as the sole input to the reduced-order surrogate model, enabling the reduced-order surrogate model to perform mapping operations only within the physical dominant subspace, thereby controlling model complexity and ensuring real-time performance from the root; constructing a unified reduced-order mapping relationship, compressing sparse feature parameters to a fixed low-dimensional computational space through a feature hash mapping matrix, reducing the memory access and multiply-accumulate overhead of the reduced-order surrogate model; selecting any one of Gaussian process regression, lightweight neural networks, or interpolation operators based on intrinsic orthogonal decomposition as the main body of the reduced-order surrogate model, inputting the compressed low-dimensional sparse feature parameters into the main body of the reduced-order surrogate model, and training through samples to enable the main body of the reduced-order surrogate model to approximate the system response of a high-fidelity physics solver in the same feature space with the lowest computational cost, thus completing the construction of the reduced-order surrogate model.

[0201] In some embodiments, the lightweight design and real-time optimization of the downgraded proxy model includes: using at least one of feature hashing, quantization-aware training, or model pruning to perform lightweight design on the constructed downgraded proxy model; identifying and compressing the scale of non-critical parameters in the downgraded proxy model; eliminating redundant computational paths; enabling the downgraded proxy model to have deterministic computational latency during the deployment phase, and adapting to edge computing devices or environments with limited computing resources; configuring an online incremental learning function for the downgraded proxy model; setting a forgetting factor and learning rate consistent with the sparse feature update mechanism; enabling the downgraded proxy model to prioritize adapting to the latest physical state changes; and avoiding model structural instability caused by short-term fluctuations, thereby achieving real-time optimization of the downgraded proxy model.

[0202] In some embodiments, activating a high-fidelity physics solver for a local region if a feature parameter mutation is detected based on a trigger threshold and exceeds the reliable prediction range of the reduced-order surrogate model includes: constructing a unified trigger criterion function, which simultaneously characterizes the evolution rate of the dominant physical process in the feature space and the prediction uncertainty of the reduced-order surrogate model output, and balancing the impact of physical mutation intensity and insufficient model cognition on triggering behavior through preset weight coefficients; comparing the calculation result of the constructed trigger criterion function with a preset trigger threshold, and if the calculation result exceeds the preset trigger threshold, determining that the feature parameter has mutated and exceeds the reliable prediction range of the reduced-order surrogate model; locating the abnormal feature parameter to a specific physical sub-region through a feature-to-space mapping operator, the mapping process combining sparse feature components and modal basis functions of key feature modes, and ensuring that the high-fidelity physics solver is activated only in the spatial region where the dominant mode has a significant effect through a preset energy contribution threshold; and calling any one of the high-fidelity physics solvers among the finite element, finite volume, or spectral methods to perform high-precision calculations only for the located physical sub-region and the corresponding time segment.

[0203] In some embodiments, the process of integrating observation information corresponding to real-time observation data streams into the prediction output of a surrogate model using data assimilation techniques, dynamically correcting the prediction bias of the reduced-order surrogate model, and outputting prediction results with uncertainty quantification includes: continuously receiving real-time observation data streams during system operation; synchronously comparing the prediction results output by the reduced-order surrogate model based on sparse feature parameters with the real-time observation data streams; calculating the corresponding observation residuals, which are used to characterize the degree of prediction mismatch of the surrogate model under the current physical state; constructing a bias-aware data assimilation update operator, and updating the data based on the prediction uncertainty scalar output by the reduced-order surrogate model and... The observation system estimates the observation noise intensity online and calculates the time-dependent adaptive correction gain. This correction gain is used to adjust the correction intensity of the prediction results by the observation information, avoiding the amplification effect of observation noise on model stability. The observation residuals and the adaptive correction gain are substituted into the data assimilation update operator to dynamically correct the prediction results of the reduced-order surrogate model, resulting in a corrected system state estimate. The corrected system state estimate and the prediction uncertainty scalar output by the reduced-order surrogate model are output synchronously to form a prediction result with uncertainty quantification. The observation residuals are mapped back to the sparse feature space to form a bias-driven feature correction signal.

[0204] In some embodiments, the establishment of a fully closed-loop feedback chain from model prediction and high-fidelity solution verification to observation data correction, combined with a reduced-order surrogate model, to complete the construction of the digital twin, includes: unifying the prediction results of the reduced-order surrogate model, the reference truth value output by the high-fidelity physical solver, the correction state after data assimilation, and the observation residuals into the same closed-loop feedback framework; constructing a comprehensive feedback evaluation function, which comprehensively characterizes the true prediction error of the current model system, the prediction uncertainty of the surrogate model, and the triggering frequency of the high-fidelity solver; balancing prediction accuracy, model robustness, and computational cost through preset weight coefficients; and based on the calculation results and long-term evolution trend of the comprehensive feedback evaluation function, adjusting the sparseness corresponding to the low-dimensional sparse feature parameter set. The feature extraction algorithm performs parameter callbacks and adaptively adjusts the sparse constraint strength parameters, enabling the sparse feature parameter set to quickly respond to structural changes in the physical field while maintaining the ability to remember stable evolution patterns, thus conforming to the real dynamic evolution structure of the system. Using the calculation results of the comprehensive feedback evaluation function, the structural complexity of the reduced-order surrogate model is adaptively controlled. While ensuring that prediction errors and uncertainties meet the standards, the non-critical computational paths of the reduced-order surrogate model are continuously compressed, maintaining its lightweight characteristics. A fully closed-loop feedback chain covering sparse feature extraction, surrogate model prediction, high-fidelity solution verification, and observation data correction is established, enabling the corresponding digital twin to possess continuous self-optimization and adaptive capabilities, completing the dynamic construction of a high-fidelity digital twin.

[0205] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to implement the steps of the dynamic construction method for a digital twin based on multi-source data fusion and physical simulation provided in any embodiment of this application.

[0206] The computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiments, such as the hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the computer device.

[0207] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for dynamically constructing a digital twin based on multi-source data fusion and physical simulation, characterized in that, include: Acquire multi-source heterogeneous data streams from preset sensor arrays, numerical simulation results, and historical databases; identify key feature modes that dominate physical processes in the multi-source heterogeneous data streams; utilize time sliding windows and forgetting mechanisms to obtain the time-varying physical field evolution laws of key feature modes corresponding to key feature modes; and extract low-dimensional sparse feature parameter sets reflecting dynamic behavior from high-dimensional observation space. Using the aforementioned low-dimensional sparse feature parameter set as input, a reduced-order surrogate model is constructed using Gaussian process regression, a neural network surrogate model, or intrinsic orthogonal decomposition combined with interpolation techniques. This includes: using the extracted low-dimensional sparse feature parameter set as the sole input to the reduced-order surrogate model, ensuring that the model performs mapping operations only within the physical dominant subspace, thereby controlling model complexity and guaranteeing real-time performance; constructing a unified reduced-order mapping relationship, compressing sparse feature parameters to a fixed low-dimensional computational space through a feature hash mapping matrix, reducing memory access and multiply-accumulate overhead of the reduced-order surrogate model; selecting any one of Gaussian process regression, lightweight neural networks, or interpolation operators based on intrinsic orthogonal decomposition as the main body of the reduced-order surrogate model, inputting the compressed low-dimensional sparse feature parameters into the main body of the reduced-order surrogate model, and training with samples to enable the main body of the reduced-order surrogate model to approximate the system response of a high-fidelity physics solver in the same feature space with minimal computational cost, thus completing the construction of the reduced-order surrogate model; and performing lightweight design and real-time optimization on the reduced-order surrogate model. Set a trigger threshold based on the rate of change of sparse features or uncertainty estimation. If a sudden change in feature parameters is detected according to the trigger threshold and exceeds the reliable prediction range of the reduced-order surrogate model, activate the high-fidelity physics solver for the local region. The high-fidelity physics solver is used to calculate for the physical sub-region and time segment where the sudden change occurred, and the result is fed back as the truth value to the corresponding model update process. The system receives the corresponding real-time observation data stream, integrates the observation information corresponding to the real-time observation data stream into the prediction output of the surrogate model through data assimilation technology, dynamically corrects the prediction bias of the reduced-order surrogate model, and outputs prediction results with uncertainty quantification. This allows for the establishment of a fully closed-loop feedback link from model prediction, high-fidelity solution verification to observation data correction, in conjunction with the reduced-order surrogate model, thus completing the construction of the digital twin.

2. The method according to claim 1, characterized in that, After the construction of the digital twin is completed, the following is also included: Based on the prediction errors, uncertainty indicators, and frequency of triggering the high-fidelity physical solver generated by the digital twin during system operation, the parameters of the sparse feature extraction algorithm corresponding to the low-dimensional sparse feature parameter set are periodically or triggered, and the structure and hyperparameters of the reduced-order surrogate model are retrained and optimized to improve the ability of the sparse feature set to capture the dynamic evolution characteristics of the system and the prediction accuracy and robustness of the surrogate model.

3. The method according to claim 1, characterized in that, The identification of key characteristic modes that dominate physical processes in multi-source heterogeneous data streams includes: The acquired multi-source heterogeneous data streams are preprocessed to unify the real-time data collected by the sensor array, the simulated data obtained by numerical simulation, and the archived data in the historical database to the same time base. Physical process consistency constraints are introduced at the data layer to eliminate the interference of heterogeneous scale on feature recognition. A state matrix is ​​constructed and updated over time. Modal decoupling is performed on the state matrix within a preset time window, and the consistency of dynamic evolution between adjacent time slices is constrained. The results of mode decoupling are analyzed using any one of the adaptive algorithms, such as dynamic mode decomposition, sparse coding, or nonlinear independent component analysis, to identify key feature modes that can reflect the dominant physical process of the system and have clear physical meaning, and to eliminate redundant modes related to non-dominant physical processes.

4. The method according to claim 3, characterized in that, The method of obtaining the time-varying physical field evolution law of key feature modes corresponding to key feature modes using a time sliding window and forgetting mechanism includes: Set a fixed-duration time sliding window so that the time sliding window continues to scroll with the time process, and capture the time series data of key feature modalities in different time segments in real time; An online update mechanism with forgetting weights is introduced to exponentially decay the contributions of key feature modes at historical moments within a time sliding window, and to reduce the impact of early state data on current feature estimation by using a preset forgetting factor. Simultaneously, a historical consistency constraint strength parameter is set to control the continuous evolution trend of key feature modes over time, avoid interference from transient noise on the tracking of evolution laws, and realize online tracking and accurate acquisition of the evolution laws of time-varying physical fields corresponding to key feature modes.

5. The method according to claim 4, characterized in that, The extraction of a low-dimensional sparse feature parameter set reflecting dynamic behavior from a high-dimensional observation space includes: L1 regularization or structured sparsity constraints are used to sparsify the key feature modal time series data obtained within the time sliding window, enhance the sparsity of feature parameters, eliminate redundant features and interference information corresponding to non-dominant physical processes, so that only a small number of feature parameters remain active at any time, and some of the activated feature parameters can explicitly correspond to the dominant physical mechanism in the system. A feature selection criterion based on energy contribution and prediction sensitivity is introduced to comprehensively evaluate the importance of the feature parameters after sparsification. During the evaluation process, the proportion of each feature parameter in the overall physical energy and its sensitivity to the prediction of the system output are considered at the same time. Based on the comprehensive importance evaluation results, the feature parameters are sorted and pruned, retaining key feature parameters and eliminating irrelevant feature parameters, forming a low-dimensional sparse feature parameter set that is dimensionally controlled, temporally consistent, physically meaningful, and reflects the essential dynamic behavior of the system.

6. The method according to claim 1, characterized in that, The lightweight design and real-time optimization of the downgraded proxy model include: By employing at least one of the techniques of feature hashing, quantization-aware training, or model pruning, the constructed reduced-order proxy model is designed to be lightweight. The scale of non-critical parameters in the reduced-order proxy model is identified and compressed, and redundant computation paths are eliminated. This enables the reduced-order proxy model to have deterministic computation latency during the deployment phase and to adapt to edge computing devices or environments with limited computing resources. Configure online incremental learning function for the reduced-order surrogate model, set forgetting factor and learning rate consistent with sparse feature update mechanism, so that the reduced-order surrogate model can adapt to the latest physical state changes first, while avoiding model structure instability caused by short-term fluctuations, and realize real-time optimization of the reduced-order surrogate model.

7. The method according to claim 1, characterized in that, If a sudden change in a feature parameter is detected based on a trigger threshold and exceeds the reliable prediction range of the reduced-order surrogate model, the high-fidelity physics solver for the local region is activated, including: A unified trigger criterion function is constructed, which simultaneously characterizes the evolution rate of the dominant physical process in the feature space and the prediction uncertainty of the output of the reduced-order surrogate model. The influence of the intensity of physical mutation and the model's lack of cognition on the triggering behavior is balanced by preset weight coefficients. The calculation result of the constructed trigger criterion function is compared with the preset trigger threshold. If the calculation result exceeds the preset trigger threshold, it is determined that the feature parameter has abruptly changed and exceeds the reliable prediction range of the reduced-order surrogate model. By using a feature-to-space mapping operator, anomalous feature parameters are located to specific physical sub-regions. The mapping process combines sparse feature components with modal basis functions of key feature modes. By using a preset energy contribution threshold, it ensures that the high-fidelity physical solver is activated only in the spatial region where the dominant mode has a significant effect. Call any high-fidelity physics solver from the finite element, finite volume, or spectral methods to perform high-precision calculations only for the located physical sub-region and its corresponding time segment.

8. The method according to claim 1, characterized in that, The process involves integrating observation information from real-time observation data streams into the surrogate model's prediction output using data assimilation technology. This dynamically corrects the prediction bias of the reduced-order surrogate model, outputting prediction results with quantified uncertainty, including: During system operation, real-time observation data streams are continuously received. The prediction results output by the reduced-order proxy model based on sparse feature parameters are synchronously compared with the real-time observation data streams, and the corresponding observation residuals are calculated. The observation residuals are used to characterize the degree of prediction mismatch of the proxy model under the current physical state. A bias-aware data assimilation update operator is constructed. Based on the prediction uncertainty scalar output by the reduced-order surrogate model and the observation noise intensity estimated online by the observation system, the time-dependent adaptive correction gain is calculated. This correction gain is used to adjust the correction intensity of the prediction results by the observation information, so as to avoid the amplification effect of observation noise on model stability. Substituting the observation residuals and adaptive correction gain into the data assimilation update operator, the prediction results of the reduced-order surrogate model are dynamically corrected to obtain the corrected system state estimate. The corrected system state estimate is output synchronously with the prediction uncertainty scalar output by the reduced-order surrogate model to form a prediction result with uncertainty quantification; and the observation residuals are mapped back to the sparse feature space to form a bias-driven feature correction signal.

9. The method according to claim 1, characterized in that, The method combines a reduced-order surrogate model to establish a fully closed-loop feedback chain from model prediction and high-fidelity solution verification to observation data correction, thereby completing the construction of the digital twin, including: The prediction results of the reduced-order surrogate model, the reference true value output by the high-fidelity physics solver, the correction state after data assimilation, and the observation residuals are all incorporated into the same closed-loop feedback framework to construct a comprehensive feedback evaluation function. The comprehensive feedback evaluation function is used to comprehensively characterize the true prediction error of the current model system, the prediction uncertainty of the surrogate model, and the triggering frequency of the high-fidelity solver. The prediction accuracy, model robustness, and computational cost are balanced by preset weight coefficients. Based on the calculation results and long-term evolution trend of the comprehensive feedback evaluation function, the sparse feature extraction algorithm corresponding to the low-dimensional sparse feature parameter set is parameter-reverted and the sparse constraint strength parameter is adaptively adjusted so that the sparse feature parameter set can quickly respond to the structural changes of the physical field, while maintaining the ability to remember the stable evolution mode, thus conforming to the real dynamic evolution structure of the system. By utilizing the calculation results of the comprehensive feedback evaluation function, the structural complexity of the reduced-order surrogate model is adaptively controlled. Under the premise of ensuring that the prediction error and uncertainty meet the standards, the non-critical calculation paths of the reduced-order surrogate model are continuously compressed to maintain the lightweight characteristics of the reduced-order surrogate model. Establish a closed-loop feedback chain covering sparse feature extraction, surrogate model prediction, high-fidelity solution verification, and observation data correction, so that the corresponding digital twin has continuous self-optimization and self-adaptation capabilities, and completes the dynamic construction of high-fidelity digital twin.

Images