CN-121744943-B - Digital twin dynamic construction method based on multi-source data fusion and physical simulation

CN121744943BCN 121744943 BCN121744943 BCN 121744943BCN-121744943-B

Abstract

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

Inventors

JIANG JUNFENG
XIANG SIYU
LI WENBIN
KANG SHIJIE

Assignees

深圳市鼎粤科技有限公司
深圳华制智能制造技术有限公司

Dates

Publication Date: 20260508
Application Date: 20260224

Claims (9)

1. A digital twin dynamic construction method based on multi-source data fusion and physical simulation is characterized by comprising the following steps: acquiring a multi-source heterogeneous data stream from a preset sensor array, a numerical simulation result and a historical database, identifying key characteristic modes of a dominant physical process in the multi-source heterogeneous data stream, acquiring a key characteristic mode time-varying physical field evolution rule corresponding to the key characteristic modes by utilizing a time sliding window and a forgetting mechanism, and extracting a low-dimensional sparse characteristic parameter set reflecting dynamic behaviors from a high-dimensional observation space; The low-dimensional sparse characteristic parameter set is taken as input, and a Gaussian process regression, a neural network proxy model or an intrinsic orthogonal decomposition combined interpolation technology is adopted to construct a reduced order proxy model, wherein the extracted low-dimensional sparse characteristic parameter set is taken as the unique input of the reduced order proxy model, so that the reduced order proxy model only performs mapping operation in a physical main subspace, the complexity of the model is controlled from the root and the instantaneity is ensured; the method comprises the steps of constructing a unified reduced-order mapping relation, compressing sparse characteristic parameters to a fixed low-dimensional calculation space through a characteristic hash mapping matrix, reducing access memory and multiplication and addition costs of a reduced-order proxy model, selecting any one of Gaussian process regression, a lightweight neural network or interpolation operators based on intrinsic orthogonal decomposition as a reduced-order proxy model main body, inputting the compressed low-dimensional sparse characteristic parameters to the reduced-order proxy model main body, enabling the reduced-order proxy model main body to approach the system response of a high-fidelity physical solver under the same characteristic space at the lowest calculation cost through sample training, and completing construction of the reduced-order proxy model; Setting a trigger threshold based on sparse feature change rate or uncertainty estimation, if feature parameter mutation is detected according to the trigger threshold and exceeds the reliable prediction range of the reduced-order agent model, activating a high-fidelity physical solver of a local area, wherein the high-fidelity physical solver is used for calculating a physical subarea and a time segment with mutation, and feeding back the result as a true value to a corresponding model updating flow; receiving a corresponding real-time observation data stream, merging the observation information corresponding to the real-time observation data stream into the prediction output of the proxy model through a data assimilation technology, dynamically correcting the prediction deviation of the reduced-order proxy model, outputting a prediction result with uncertainty quantification, and establishing a full-closed loop feedback link from model prediction, high-fidelity solving and checking to observation data correction by combining the reduced-order proxy model to complete the construction of the digital twin body.
2. The method of claim 1, further comprising, after said completing the construction of the digital twins: Based on the prediction error, uncertainty index and frequency of triggering a high-fidelity physical solver generated by the digital twin in the system operation process, periodically or trigging a sparse feature extraction algorithm corresponding to a low-dimensional sparse feature parameter set, and retraining and optimizing the structure and super parameters of a reduced-order agent model so as to improve the capturing capacity of the sparse feature set on the dynamic evolution characteristic of the system and the prediction precision and robustness of the agent model.
3. The method of claim 1, wherein the identifying key feature modalities of a dominant physical process in the multi-source heterogeneous data stream comprises: Preprocessing the acquired multi-source heterogeneous data stream, unifying real-time data acquired by a sensor array, simulation data obtained by numerical simulation and archive data in a historical database to the same time reference, and introducing physical process consistency constraint in a data layer to eliminate interference caused by heterogeneous scale on feature identification; Constructing a state matrix which is updated in a rolling way along with time, performing modal decoupling processing on the state matrix in a preset time window, and restraining the dynamic evolution consistency of adjacent time slices; And analyzing a result after model decoupling by adopting any self-adaptive algorithm of dynamic mode decomposition, sparse coding or nonlinear independent component analysis to identify key characteristic modes which can reflect a system dominant physical process and have definite physical meanings, and eliminating redundant modes related to a non-dominant physical process.
4. The method of claim 3, wherein the obtaining, using the time sliding window and the forgetting mechanism, the evolution law of the time-varying physical field of the key feature mode corresponding to the key feature mode includes: Setting a time sliding window with fixed duration, enabling the time sliding window to continuously roll along with a time process, and capturing time sequence data of key characteristic modes in different time slices in real time; introducing an online updating mechanism with forgetting weight, carrying out exponential decay processing on key characteristic modal contributions at historical moments in a time sliding window, and reducing the influence of early state data on current characteristic estimation through a preset forgetting factor; Meanwhile, historical consistency constraint intensity parameters are set, continuous evolution trend of key characteristic modes along with time is controlled, interference caused by transient noise on evolution law tracking is avoided, and on-line tracking and accurate acquisition of time-varying physical field evolution laws corresponding to the key characteristic modes are realized.
5. The method of claim 4, wherein the extracting a low-dimensional sparse feature parameter set reflecting dynamic behavior from a high-dimensional observation space comprises: The method comprises the steps of performing sparsification treatment on key characteristic modal time sequence data acquired in a time sliding window by adopting an L1 regularization or structured sparsification constraint mode, strengthening sparsity of characteristic parameters, eliminating redundant characteristics and interference information corresponding to a non-dominant physical process, enabling only a small number of characteristic parameters to keep an activated state at any moment, and enabling part of activated characteristic parameters to be capable of displaying a dominant physical mechanism in a corresponding system; Introducing a characteristic screening criterion based on energy contribution and prediction sensitivity, and performing comprehensive importance evaluation on the characteristic parameters after the sparsification treatment, wherein the ratio of each characteristic parameter in the whole physical energy and the sensitivity of each characteristic parameter to the system output prediction are considered in the evaluation process; And sequencing and cutting the characteristic parameters according to the comprehensive importance evaluation result, reserving key characteristic parameters, removing irrelevant characteristic parameters, and forming a low-dimensional sparse characteristic parameter set which is controlled in dimension, consistent in time sequence, clear in physical meaning and reflects the intrinsic dynamics behavior of the system.
6. The method of claim 1, wherein the lightweight design and real-time optimization of the reduced-order proxy model comprises: Adopting at least one technology of characteristic hashing, quantitative perception training or model pruning to carry out lightweight design on the constructed reduced-order agent model, identifying and compressing non-key parameter scale in the reduced-order agent model, eliminating redundant calculation paths, enabling the reduced-order agent model to have deterministic calculation delay in a deployment stage, and adapting to edge calculation equipment or limited calculation resource environments; The online incremental learning function is configured for the reduced-order agent model, forgetting factors and learning rates consistent with a sparse feature updating mechanism are set, so that the reduced-order agent model can be preferentially adapted to the latest physical state change, meanwhile, the instability of a model structure caused by short-time fluctuation is avoided, and the real-time optimization of the reduced-order agent model is realized.
7. The method according to claim 1, wherein activating the high-fidelity physical solver of the local region if the characteristic parameter mutation is detected according to the trigger threshold and exceeds the reliable prediction range of the reduced-order proxy model comprises: Constructing a unified trigger criterion function, and simultaneously describing the evolution rate of a dominant physical process in a feature space and the prediction uncertainty output by a reduced-order agent model by the trigger criterion function, and balancing the influence of physical mutation strength and model cognition deficiency on the trigger behavior through a preset weight coefficient; Comparing the calculated result of the constructed trigger criterion function with a preset trigger threshold, and if the calculated result exceeds the preset trigger threshold, judging that the characteristic parameters are suddenly changed and exceed the reliable prediction range of the reduced-order agent model; Positioning abnormal characteristic parameters to specific physical subareas through a characteristic-to-space mapping operator, wherein the mapping process combines a sparse characteristic component and a modal basis function of a key characteristic mode, and a preset energy contribution threshold value is used for ensuring that a high-fidelity physical solver is started only in a space area with a remarkable effect of a dominant mode; And calling any one high-fidelity physical solver in finite element, finite volume or spectrum methods, and performing high-precision calculation only on the positioned physical subarea and the corresponding time segment.
8. The method according to claim 1, wherein the merging, by the data assimilation technology, the observation information corresponding to the real-time observation data flow into the proxy model prediction output, dynamically correcting the prediction bias of the reduced-order proxy model, and outputting the prediction result with uncertainty quantization, includes: Continuously receiving a real-time observation data stream in the running process of the system, synchronously comparing a prediction result output by the reduced-order agent model based on sparse characteristic parameters with the real-time observation data stream, and calculating a corresponding observation residual, wherein the observation residual is used for describing the prediction mismatch degree of the agent model in the current physical state; Constructing a deviation-aware data assimilation update operator, calculating time-dependent self-adaptive correction gain according to a prediction uncertainty scalar output by a reduced order agent model and the on-line estimated observation noise intensity of an observation system, and adjusting the correction intensity of observation information to a prediction result through the correction gain to avoid the amplification effect of observation noise on model stability; Substituting the observation residual and the self-adaptive correction gain into a data assimilation update operator, and dynamically correcting the prediction result of the reduced-order agent model to obtain corrected system state estimation; synchronously outputting the corrected system state estimation and the prediction uncertainty scalar output by the reduced order agent model to form a prediction result with uncertainty quantification; and mapping the observation residual back to the sparse feature space to form a bias-driven feature correction signal.
9. The method of claim 1, wherein the establishing a full closed loop feedback link from model prediction, high fidelity solution verification to observed data correction in combination with the reduced order proxy model, completes the construction of the digital twin, comprises: uniformly integrating a prediction result of the reduced-order agent model, a reference true value output by the high-fidelity physical solver, a correction state after data assimilation and an observation residual into the same closed-loop feedback frame to construct a comprehensive feedback evaluation function, wherein the comprehensive feedback evaluation function is used for comprehensively describing a real prediction error of a current model system, a prediction uncertainty of the agent model and a trigger frequency of the high-fidelity solver, and balancing prediction precision, model robustness and calculation cost through preset weight coefficients; based on the calculation result of the comprehensive feedback evaluation function and the long-term evolution trend, parameter callback is carried out on a sparse feature extraction algorithm corresponding to the low-dimensional sparse feature parameter set, sparse constraint intensity parameters are adaptively adjusted, so that the sparse feature parameter set can rapidly respond to structural changes of a physical field, memory capacity of a stable evolution mode is maintained, and a real dynamic evolution structure of a system is attached; The method comprises the steps of performing self-adaptive control on the structural complexity of a reduced-order agent model by utilizing a calculation result of a comprehensive feedback evaluation function, continuously compressing a non-key calculation path of the reduced-order agent model on the premise of ensuring that the prediction error and uncertainty reach standards, and maintaining the lightweight characteristic of the reduced-order agent model; And establishing a full-closed loop feedback link for covering sparse feature extraction, agent model prediction, high-fidelity solving verification and observation data correction, so that the corresponding digital twin body has continuous self-optimization and self-adaptation capability, and the dynamic construction of the high-fidelity digital twin body is completed.

Description

Digital twin dynamic construction method based on multi-source data fusion and physical simulation Technical Field The application relates to the technical field of digital twin, physical modeling and multi-source data fusion, in particular to a digital twin dynamic construction method based on multi-source data fusion and physical simulation. Background In the technical fields of digital twinning, physical modeling and multi-source data fusion, the digital twinning construction of complex physical systems (such as industrial equipment and engineering systems) faces significant challenges. The traditional method relies on a static model or a single data source, and the mapping between the physical system and the virtual model is realized through a predefined rule. For example, while some schemes implement the integration of geometric models, physical models, flow models, and data models, their core drawbacks are: 1. Static modeling and dynamic adaptability are insufficient, namely multimode fusion depending on a fixed structure is insufficient, dynamic recognition capability of dominant physical processes (such as strong nonlinearity and time-varying characteristics) in real-time operation of a system is insufficient, key features with physical significance cannot be extracted from high-dimensional data, and therefore the representation capability of a model on complex working conditions is limited. 2. The computing efficiency and the precision are unbalanced, namely, global and full-time high-fidelity physical solution or complete model synchronous update is adopted, the computing burden is increased sharply when the system state is suddenly changed, real-time response is difficult to realize in an edge device or limited computing force environment, a large amount of resources are consumed by global solution, and the on-demand distribution of computing resources is not realized. 3. Error accumulation and self-optimization are lacking, namely model errors depend on manual correction or off-line updating, closed loop feedback mechanisms are absent among the proxy model, physical solution and observation data, accumulated deviation is easily generated due to environmental disturbance or simplifying assumptions during long-term operation, and prediction reliability cannot be adaptively improved. Accordingly, a need exists for a method that addresses at least one of the problems described above. Disclosure of Invention The application provides a digital twin dynamic construction method based on multi-source data fusion and physical simulation, which aims to solve the problem that the traditional method relies on a static model or a single data source, and the mapping between a physical system and a virtual model is realized through a predefined rule. For example, although the partial scheme realizes the integration of a geometric model, a physical model, a flow model and a data model, the core defects of the partial scheme are problems of insufficient static modeling and dynamic adaptability, unbalanced calculation efficiency and precision, error accumulation, self-optimization deficiency and the like. In a first aspect, an embodiment of the present application provides a method for dynamically constructing a digital twin body based on multi-source data fusion and physical simulation, where the method includes: acquiring a multi-source heterogeneous data stream from a preset sensor array, a numerical simulation result and a historical database, identifying key characteristic modes of a dominant physical process in the multi-source heterogeneous data stream, acquiring a key characteristic mode time-varying physical field evolution rule corresponding to the key characteristic modes by utilizing a time sliding window and a forgetting mechanism, and extracting a low-dimensional sparse characteristic parameter set reflecting dynamic behaviors from a high-dimensional observation space; taking the low-dimensional sparse characteristic parameter set as input, constructing a reduced-order proxy model by adopting Gaussian process regression, a neural network proxy model or an intrinsic orthogonal decomposition combined interpolation technology, and carrying out lightweight design and real-time optimization on the reduced-order proxy model; Setting a trigger threshold based on sparse feature change rate or uncertainty estimation, if feature parameter mutation is detected according to the trigger threshold and exceeds the reliable prediction range of the reduced-order agent model, activating a high-fidelity physical solver of a local area, wherein the high-fidelity physical solver is used for calculating a physical subarea and a time segment with mutation, and feeding back the result as a true value to a corresponding model updating flow; receiving a corresponding real-time observation data stream, merging the observation information corresponding to the real-time observation data stream into the prediction output of the p