A multi-fidelity bayesian based long-term service waterworks twin model updating method

Abstract

This invention discloses a method for updating a long-service hydraulic twin model based on multi-fidelity Bayesian methods. It involves collecting non-contact, full-field micro-vibration time-history data of hydraulic structures and extracting high- and low-fidelity data observations. Combining macroscopic physical aging equations, a time-varying degradation prior probability distribution with historical memory is constructed. Based on the high- and low-fidelity data observations, a spatially and mechanistically heterogeneous multi-fidelity twin architecture is built. A spatial covariance matrix of full-field measurement points is introduced, and the Bayesian heterogeneous MCMC method is used to infer parameters of the multi-fidelity twin architecture until convergence. The high-confidence posterior parameters of the structural physical parameters and boundary nonlinear stiffness are mapped to the long-service hydraulic twin model, enabling high-frequency adaptive evolution and synchronous updating of the full life-cycle state of the long-service hydraulic structure digital twin under complex uncertainties.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary technical field of digital twins of hydraulic structures and structural health monitoring, specifically involving a method for updating long-service hydraulic twin models based on multi-fidelity Bayes. Background Technology

[0002] Under long-term complex environmental excitation and multi-field coupling, the material properties (such as concrete carbonation) and boundary conditions (such as contact surface slippage and foundation constraint relaxation) of large hydraulic structures will progressively degrade, and structural damage will cause changes in local structural properties. Constructing a high-fidelity digital twin model is the key to realizing long-term structural safety assessment.

[0003] Existing finite element models and Bayesian update methods suffer from the following technical bottlenecks: (1) Traditional Bayesian inference often uses static empirical prior distribution. However, many of my country's water conservancy projects have been in service for a long time, and the static empirical prior distribution ignores the physical degradation of hydraulic structures over decades (i.e., lacks "historical memory" data). (2) When updating large finite element models containing complex nonlinear constitutive and contact, the Markov chain Monte Carlo (MCMC) algorithm needs to perform tens of thousands of repeated calculations, which often leads to a lot of time consumption or even computational crashes. (3) Most existing multi-fidelity models are limited to the coarse and fine division of the global grid scale, and fail to achieve heterogeneous fidelity alignment in terms of physical mechanisms and high and low frequency data characteristics, making it difficult to accurately locate local defects. Summary of the Invention

[0004] To address the aforementioned shortcomings in existing technologies, the present invention provides a method for updating long-service hydraulic twin models based on multi-fidelity Bayesian methods. By introducing a time-varying degradation prior mechanism and a heterogeneous multi-fidelity asynchronous proxy architecture, this method completely solves the engineering computing power problem of long-service complex hydraulic structure digital twin models struggling to achieve real-time, high-frequency adaptive evolution under massive uncertainties.

[0005] To achieve the aforementioned objectives, the present invention employs the following technical solution: a method for updating a long-service hydraulic twin model based on multi-fidelity Bayes, comprising the following steps: S100: Obtain the current non-contact full-field micro-vibration time history data of the hydraulic structure and decouple it into low-fidelity data observations characterizing the overall stiffness of the structure and high-fidelity data observations characterizing the local boundary and nonlinear states. S200, based on the macroscopic physical aging equation of hydraulic structure, constructs a time-varying degradation prior probability distribution with its own historical memory characteristics; S300, Construct a multi-fidelity twin architecture with spatial and mechanistic heterogeneity, including a global low-fidelity proxy model and a local high-fidelity three-dimensional finite element model; S400, introduce the spatial covariance matrix of the full field measurement points, construct a multivariate Gaussian likelihood function by combining low-fidelity data observations and high-fidelity data observations, and perform Bayesian heterogeneous MCMC parameter inference based on spatial correlation on the multifidelity twin architecture based on time-varying degenerate prior probability distribution until the posterior parameter sampling converges; S500: Extract the posterior probability distribution of the physical parameters and boundary nonlinear stiffness of the hydraulic structure to be corrected after convergence, and automatically map the parameters of its high confidence interval back into the long-term hydraulic twin model to complete the synchronous update of the hydraulic structure's full life cycle state.

[0006] Further, in S100, extracting low-fidelity data observations includes: Image domain measurement points in non-contact full-field micro-vibration time history data exist The analytical signal at time step is analyzed, its phase is extracted and the time-domain solution is obtained, the original cumulative displacement is calculated, and a low-pass digital filter is introduced to obtain the low-frequency steady-state displacement sequence. As a low-fidelity data observation, it is represented as: In the formula, This is the original cumulative displacement. This is the cutoff frequency of the low-pass digital filter. It is a low-pass digital filter. To analyze the signal, For time phase untangling operator, is the local spatial wavenumber.

[0007] Furthermore, in step S100, extracting high-fidelity data observations includes: For non-contact full-field micro-vibration time history data, the instantaneous phase difference is calculated using the complex conjugate product of the analytical signals between adjacent video frames to obtain the instantaneous velocity response. Then, a bandpass filter or high-pass filter is used to extract specific high-frequency bands, resulting in a high-frequency transient velocity sequence. As a high-fidelity data observation, it is represented as In the formula, For instantaneous speed response, This is the cutoff frequency of the bandpass or highpass filter. It is a bandpass filter or a highpass filter. To analyze the signal, for The complex conjugate of a complex vector at time step.

[0008] Furthermore, in step S200, the time-varying degradation prior probability distribution constructed for the physical parameters to be corrected exhibiting decay characteristics is as follows: In the formula, Let be the mean evolution function. The prior equations are time-varying. For the physical parameters of the hydraulic structure to be corrected, These are the initial design values ​​for the physical parameters of the hydraulic structure to be corrected. As a comprehensive damage and degradation factor based on physical laws, For service time variable, It follows a Gaussian normal distribution.

[0009] Furthermore, step S300 includes the following sub-steps: S301. Based on low-fidelity data observations, a lightweight data-driven global proxy model is constructed as a low-fidelity model. S302. Based on high-fidelity data observations, a refined three-dimensional finite element local sub-model is constructed as a high-fidelity model; S303. Based on the multi-fidelity Gaussian process, the low-fidelity model and the high-fidelity model are bridged across scales to obtain a multi-fidelity twin architecture.

[0010] Furthermore, in S303, the bridging equation for cross-scale bridging of the low-fidelity model and the high-fidelity model is: In the formula, For high-fidelity model high-fidelity prediction of response, This represents the low-fidelity prediction vector for the low-fidelity model. This is the scale scaling factor. For high-fidelity residual processes, The probability distribution is a Gaussian process. It is a mean function. For the mean function of the high-fidelity residual process, For parameterized covariance kernel function, The covariance kernel function for a high-fidelity residual process. The input vector consists of the physical parameters of the hydraulic structure to be corrected. This is another set of reference input vectors in the residual space, used to... Jointly calculate spatial correlation. For hyperparameter set; The predicted mean of the high-fidelity predicted response output by the multi-fidelity twin architecture And the prediction variance of the cognitive uncertainty of the characterization model for: In the formula, For low-fidelity models at unknown prediction points The predicted mean at that location, For high-fidelity residual processes at unknown prediction points The predicted mean at that location, For low-fidelity subscripts, The subscript for the residual.

[0011] Furthermore, S400 includes the following sub-steps: S401. Introduce the spatial covariance matrix of all measurement points, combine it with the high-fidelity observation data of the whole field, use the multi-fidelity twin architecture to predict the mean to replace the real high-fidelity model, and construct a multivariate Gaussian likelihood function. S402, Configure independent dual-thread operation to run the asynchronous Markov chain Monte Carlo sampling algorithm; The main thread combines the time-varying degradation prior probability distribution with the multivariate Gaussian likelihood function to perform Metropolis-Hastings sampling. When the predicted variance of the parameter region exceeds the confidence tolerance threshold, the corresponding parameter point is pushed into the background task queue. The background daemon thread extracts parameter points from the task queue, automatically rewrites the input file of the local high-fidelity three-dimensional finite element model through the script, and performs parallel calculations. After the calculation is completed, it extracts the dynamic features from the result file. S403. Dynamically add the parameter points and dynamic features as a set of precise samples to the high-fidelity training set. S404. Repeat S402~S403 to incrementally retrain the hyperparameters of the multi-fidelity twin architecture based on the high-fidelity training set, update the high-fidelity residual process until the Markov chain converges, output the posterior probability distribution of the physical parameters to be corrected, and complete the Bayesian heterogeneous MCMC parameter inference.

[0012] Furthermore, in step S401, the constructed multivariate Gaussian likelihood function is: In the formula, For high-fidelity observation data across the entire field, The spatial covariance matrix of all measurement points is... The superscript indicates the candidate point for the new parameter, and the superscript indicates the transpose operator.

[0013] Furthermore, S500 includes the following sub-steps: S501. Extract the optimal mapping parameters required for the long-term hydraulic twin model, construct the posterior covariance matrix to characterize the uncertainty of identification, and then determine the high-confidence evolution interval of the physical parameters of each hydraulic structure; whereby the optimal mapping parameters are the posterior mean estimates of the physical parameters of the hydraulic structure to be corrected after convergence. S502. The extracted optimal mapping parameters are automatically mapped to the underlying physical equations of the long-term hydraulic twin model. Combined with the material stiffness degradation factor and nonlinear boundary contact stiffness, the global evolution stiffness matrix of the long-term hydraulic twin model is reconstructed. S503. Based on the global evolution stiffness matrix, apply actual working conditions to the long-term hydraulic twin model, solve the current safe bearing capacity state of the structure in a forward manner, and output the full life-cycle health index of the long-term hydraulic structure to complete the synchronous update of the full life-cycle state of the hydraulic structure.

[0014] Furthermore, in S502, the global evolution stiffness matrix for: In the formula, For optimal mapping parameters The first mapping Regional material stiffness degradation factor The nonlinear boundary contact stiffness matrix of the structure is obtained by inversion from high-frequency micro-vibration data. For the first time in a healthy state The element stiffness matrix of each region This represents the number of physical sub-regions into which the hydraulic structure is divided.

[0015] The beneficial effects of this invention are as follows: (1) The prior characteristics of long service life are redefined: a dynamic prior distribution based on time-varying degradation mechanism is creatively proposed, which gives the digital twin model the ability of "historical memory", effectively narrows the search space of Bayesian inversion, and improves the physical accuracy of identifying the true degradation state of materials under complex working conditions.

[0016] (2) It breaks through the dimensional limitations of traditional multi-fidelity: It constructs a heterogeneous multi-fidelity update architecture that covers the spatial scale of "global data proxy-local finite element sub-model" and the data characteristics of "low-frequency steady state-high-frequency transient state", making full use of the high spatial resolution advantage of visual non-contact measurement.

[0017] (3) Improved cross-platform computing efficiency: By combining the asynchronous MCMC algorithm with MFGP, the extremely time-consuming complex finite element trial calculation is transformed into on-demand calculation triggered by background adaptive triggering, which makes the update time of the huge hydraulic digital twin model decrease exponentially, meeting the near real-time requirements of structural inspection and online safety assessment. Attached Figure Description

[0018] Figure 1 The flowchart of the long-term hydraulic twin model update method based on multi-fidelity Bayes provided by the present invention is shown. Detailed Implementation

[0019] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0020] This invention provides a method for updating long-service hydraulic twin models based on multi-fidelity Bayesian methods, referencing... Figure 1 This includes the following steps: S100: Obtain the current non-contact full-field micro-vibration time history data of the hydraulic structure and decouple it into low-fidelity data observations characterizing the overall stiffness of the structure and high-fidelity data observations characterizing the local boundary and nonlinear states. S200, based on the macroscopic physical aging equation of hydraulic structure, constructs a time-varying degradation prior probability distribution with its own historical memory characteristics; S300, Construct a multi-fidelity twin architecture with spatial and mechanistic heterogeneity, including a global low-fidelity proxy model and a local high-fidelity three-dimensional finite element model; S400, introduce the spatial covariance matrix of the full field measurement points, construct a multivariate Gaussian likelihood function by combining low-fidelity data observations and high-fidelity data observations, and perform Bayesian heterogeneous MCMC parameter inference based on spatial correlation on the multifidelity twin architecture based on time-varying degenerate prior probability distribution until the posterior parameter sampling converges; S500: Extract the posterior probability distribution of the physical parameters and boundary nonlinear stiffness of the hydraulic structure to be corrected after convergence, and automatically map the parameters of its high confidence interval back into the long-term hydraulic twin model to complete the synchronous update of the hydraulic structure's full life cycle state.

[0021] In S100 of this embodiment of the invention, the current non-contact full-field micro-vibration time history data of the hydraulic structure is acquired and decomposed into low-frequency steady-state displacement data (low-fidelity data observations) characterizing the overall stiffness of the structure and high-frequency transient velocity response values ​​(high-fidelity data observations) characterizing local nonlinearity and boundary state.

[0022] In this embodiment, extracting low-fidelity data observations includes: Image domain measurement points in non-contact full-field micro-vibration time history data exist The analytical signal at time step is analyzed, its phase is extracted and the time-domain solution is obtained, the original cumulative displacement is calculated, and a low-pass digital filter is introduced to obtain the low-frequency steady-state displacement sequence. As a low-fidelity data observation, it is represented as: In the formula, This is the original cumulative displacement. This is the cutoff frequency of the low-pass digital filter. It is a low-pass digital filter. To analyze the signal, For time phase untangling operator, For the local spatial wavenumber. Among them, the time-phase unwrapping operator... Used to eliminate phase truncation The transition low-pass digital filter (which can be a zero-phase-shift Butterworth low-pass filter) is used to filter out high-frequency environmental noise and local oscillations, and to obtain low-frequency steady-state displacement data sequences.

[0023] In this embodiment, the low-frequency steady-state displacement data sequence dominates the first few macroscopic modes of the hydraulic structure. Due to its long wavelength and insensitivity to local details, it is defined as low-fidelity data observations and is specifically used in subsequent steps for Bayesian likelihood matching with the output of the coarse-grid finite element or lightweight surrogate model (low-fidelity model).

[0024] In this embodiment, extracting high-fidelity data observations includes: For non-contact full-field micro-vibration time history data, the instantaneous phase difference is calculated using the complex conjugate product of the analytical signals between adjacent video frames to obtain the instantaneous velocity response. Then, a bandpass filter or high-pass filter is used to extract specific high-frequency bands, resulting in a high-frequency transient velocity sequence. As a high-fidelity data observation, it is represented as: In the formula, For instantaneous speed response, This is the cutoff frequency of the bandpass or highpass filter. It is a bandpass filter or a highpass filter. To analyze the signal, for The complex conjugate of a complex vector at time step.

[0025] Specifically, in high-fidelity data observations, to avoid the masking of weak high-frequency signals by low-frequency, long-period environmental drift (such as baseline drift caused by wind load) during integration, this invention switches to a conjugate multiplication differential mode. By utilizing the complex conjugate product of the analytical signals between adjacent video frames, the instantaneous phase difference is directly calculated to obtain the instantaneous velocity response. This complex-domain differential operation has inherent high-pass filtering characteristics, completely eliminating low-frequency trend terms. Subsequently, specific high-frequency bands (such as bands containing local impulse responses or higher-order modes) are extracted using bandpass or high-pass filters to obtain the high-frequency transient velocity response sequence.

[0026] High-frequency stress waves exhibit significant distortion and scattering when encountering internal micro-damage or boundary nonlinear relaxation. Therefore, this high-frequency transient velocity response sequence is extremely sensitive to local defects and is defined as high-fidelity data observations. In subsequent Bayesian asynchronous sampling, it is specifically used to trigger and verify a three-dimensional Abaqus refined finite element model (high-fidelity model) containing complex nonlinear contacts.

[0027] In S200 of this embodiment of the invention, a time-varying degradation prior probability distribution with historical memory characteristics is constructed based on the macroscopic physical aging equation of the hydraulic structure (such as fatigue damage accumulation or carbonization depth model); so that the prior probability density function of the parameter to be corrected evolves dynamically with the service time, rather than being statically assumed.

[0028] Unlike the traditional assumption of a Gaussian constant, this invention uses the service time variable... Explicitly introduced into Bayesian prior inference. Combined with the long-term damage accumulation rule, this is applied to the physical parameters to be corrected, which exhibit decay characteristics. (e.g., the elastic modulus of a material), its constructed time-varying degradation prior probability distribution is: In the formula, Let be the mean evolution function. The prior equations are time-varying. For the physical parameters of the hydraulic structure to be corrected, These are the initial design values ​​for the physical parameters of the hydraulic structure to be corrected. As a comprehensive damage and degradation factor based on physical laws, For service time variable, It follows a Gaussian normal distribution.

[0029] The above formula enables the prior search probability density of the parameters to automatically undergo a reasonable baseline shift as the service life of the structure increases, thus avoiding locally optimal solutions with physical distortion.

[0030] In S300 of this embodiment, in order to overcome the problem of huge computation time of a single high-fidelity finite element model, this invention constructs a lightweight data-driven global proxy model as a low-fidelity model to quickly capture the macroscopic self-vibration frequency response; constructs a refined three-dimensional finite element local sub-model as a high-fidelity model; and based on the multi-fidelity Gaussian process (MFGP), the global lightweight proxy and the local three-dimensional finite element sub-model are bridged across scales to obtain a multi-fidelity twin architecture, namely the MFGP model.

[0031] Specifically, S300 includes the following steps: S301. Based on low-fidelity data observations, a lightweight data-driven global proxy model is constructed as a low-fidelity model. S302. Based on high-fidelity data observations, a refined three-dimensional finite element local sub-model is constructed as a high-fidelity model; S303. Based on the multi-fidelity Gaussian process, the low-fidelity model and the high-fidelity model are bridged across scales to obtain a multi-fidelity twin architecture.

[0032] In S301, for low-fidelity data observations, the input vector is composed of the physical parameters of the hydraulic structure to be corrected (such as local elastic modulus and boundary spring stiffness). Utilizing low-fidelity models with extremely low computational overhead (such as global coarse-mesh finite element methods or reduced-order analytical models) to obtain massive training sample sets. Define the low-fidelity predicted response of the low-fidelity model. For a Gaussian process: In S302, for high-fidelity data observations, a small number of high-fidelity three-dimensional finite element (such as Abaqus refined mesh) sample sets are obtained, which is extremely time-consuming. ,and .

[0033] In S303, the Kennedy-O'Hagan autoregressive framework is introduced to establish low-fidelity and high-fidelity responses. The bridging equation between them is: In the formula, For high-fidelity model high-fidelity prediction of response, This represents the low-fidelity prediction vector for the low-fidelity model. This is the scale scaling factor. For high-fidelity residual processes, The probability distribution is a Gaussian process. It is the mean function (usually a multinomial regression model). For the mean function of the high-fidelity residual process, The parameterized covariance kernel function is selected (preferably the Matern 5 / 2 kernel to ensure the smoothness of the physical response). The covariance kernel function for a high-fidelity residual process. The input vector consists of the physical parameters of the hydraulic structure to be corrected. This is another set of reference input vectors in the residual space, used to... Jointly calculate spatial correlation. Let be the set of hyperparameters; where is the scale scaling factor. Used to correct global linearity bias in low-fidelity models; high-fidelity residual process Used to fit complex local nonlinear features and contact state boundary responses.

[0034] Furthermore, hyperparameters were analyzed using the co-kriging method. Perform maximum likelihood estimation. After training, for any unknown combination of parameters... This multifidelity twin architecture (MFGP model) can output high-fidelity response prediction mean at millisecond speeds. and the prediction variance of the cognitive uncertainty of the characterization model They are represented as follows: In the formula, For low-fidelity models at unknown prediction points The predicted mean at that location, For high-fidelity residual processes at unknown prediction points The predicted mean at that location, For low-fidelity subscripts, The subscript for the residual.

[0035] In S400 of this embodiment, a spatial covariance matrix of full-field measurement points is introduced to quantify the spatial redundancy and measurement noise of full-field mode measurement points, and a multivariate Gaussian likelihood function is constructed; an asynchronous Markov chain Monte Carlo sampling algorithm is run, the main thread uses MFGP to perform millisecond-level global likelihood estimation, the background worker thread filters high uncertainty samples through an adaptive learning function, silently drives the three-dimensional finite element software to perform high-fidelity calculation of local sub-models, and dynamically enriches the calculation results into the training set of the MFGP model.

[0036] S400 of this embodiment of the invention includes the following sub-steps: S401. Introduce the spatial covariance matrix of all measurement points, combine it with the high-fidelity observation data of the whole field, use the multi-fidelity twin architecture to predict the mean to replace the real high-fidelity model, and construct a multivariate Gaussian likelihood function. S402, Configure independent dual-thread operation to run the asynchronous Markov chain Monte Carlo sampling algorithm; The main thread combines the time-varying degradation prior probability distribution with the multivariate Gaussian likelihood function to perform Metropolis-Hastings sampling. When the predicted variance of the parameter region exceeds the confidence tolerance threshold, the corresponding parameter point is pushed into the background task queue. The background daemon thread extracts parameter points from the task queue, automatically rewrites the input file of the local high-fidelity three-dimensional finite element model through the script, and performs parallel calculations. After the calculation is completed, it extracts the dynamic features from the result file. S403. Dynamically add the parameter points and dynamic features as a set of precise samples to the high-fidelity training set. S404. Repeat S402~S403 to incrementally retrain the hyperparameters of the multi-fidelity twin architecture based on the high-fidelity training set, update the high-fidelity residual process (i.e., residual surface) until the Markov chain converges, output the posterior probability distribution of the physical parameters to be corrected, and complete the Bayesian heterogeneous MCMC parameter inference.

[0037] In S401, the full-field high-fidelity observation data obtained from the previous decoupling is combined. With spatial covariance matrix By using a multi-fidelity twin architecture to predict the mean instead of the true high-fidelity model, a multivariate Gaussian likelihood function is constructed as follows: In the formula, For high-fidelity observation data across the entire field, The spatial covariance matrix of all measurement points is... The superscript indicates the candidate point for the new parameter, and the superscript indicates the transpose operator.

[0038] In S402, two threads are allocated within the main control algorithm system (such as the MATLAB platform) to run independently: In the main thread, the sampler generates new parameter candidate points based on the proposal distribution. The main thread rapidly calls the MFGP model to obtain... and And calculate the Metropolis-Hastings acceptance rate. : in, Using a time-varying degenerate prior probability distribution, the main thread continuously advances the evolution of the Markov chain with this acceptance rate, enabling rapid exploration of a parameter space on the order of millions.

[0039] While traversing the main thread, the prediction variance is monitored in real time. If a candidate point is found to meet the active learning trigger condition, the system will take action. (That is, if the surrogate model is highly "unconfident" in the current parameter range), then the parameter point will be... Push it into the background task queue.

[0040] Background daemon process extracts queue The script automatically rewrites the Abaqus input file (.inp) and submits it for multi-core parallel computation. After the computation is complete, the script silently extracts the dynamic features from the result file (.odb). .

[0041] In S403, the newly obtained accurate sample Dynamically add high-fidelity training sets In the background, the system incrementally retrains the hyperparameters of the MFGP model, updating the high-fidelity residuals. Complete the Bayesian heterogeneous MCMC parameter inference.

[0042] Through the above asynchronous parallel logic, the Bayesian sampling of the main thread does not need to wait for the long computation of Abaqus. This not only ensures the rapid convergence of the Markov chain, but also, as the sampling proceeds, the physical fidelity of the MFGP model in the high-probability region of the target becomes higher and higher, completely solving the computational barrier of high-frequency evolution of long-term digital twin models.

[0043] In S500 of this embodiment, after the MCMC sampling algorithm reaches a stationary distribution, the system obtains a joint posterior Markov chain sample set of the physical parameters to be corrected (including the elastic modulus degradation factor and boundary stiffness of each aging region). To map the statistically significant probability distribution to a uniquely determined long-service hydraulic twin model in the physical world, a parameter evolution and physical matrix reconstruction mechanism is employed; specifically, S500 includes the following sub-steps: S501. Extract the optimal mapping parameters required for the long-term hydraulic twin model, construct the posterior covariance matrix to characterize the uncertainty of identification, and then determine the high-confidence evolution interval of the physical parameters of each hydraulic structure; whereby the optimal mapping parameters are the posterior mean estimates of the physical parameters of the hydraulic structure to be corrected after convergence. S502. The extracted optimal mapping parameters are automatically mapped to the underlying physical equations of the long-term hydraulic twin model. Combined with the material stiffness degradation factor and nonlinear boundary contact stiffness, the global evolution stiffness matrix of the long-term hydraulic twin model is reconstructed. S503. Based on the global evolution stiffness matrix, apply actual working conditions to the long-term hydraulic twin model, solve the current safe bearing capacity state of the structure in a forward manner, and output the full life-cycle health index of the long-term hydraulic structure to complete the synchronous update of the full life-cycle state of the hydraulic structure.

[0044] In S501, the effective MCMC sample chain length after removing the burn-in period is set as follows: , No. The parameter vector for the second sampling is Extracting the optimal mapping parameters (i.e., posterior mean estimates) required for a long-term hydraulic twin model. ) and the posterior covariance matrix representing the uncertainty of identification : In the formula, Let be the posterior covariance matrix of the physical parameters to be corrected.

[0045] By extracting the square roots of the diagonal elements of the covariance matrix, the high-confidence evolution intervals (such as the 95% confidence interval) of each physical parameter can be determined. ).

[0046] In S502, the extracted optimal mapping parameters Automatically mapped to the underlying physical equations of a high-fidelity 3D finite element model (such as Abaqus). Assume the structure is divided into... The first physical sub-region or unit, in the initial healthy state, is the... The element stiffness matrix of each region is: Based on damage mechanics and boundary evolution theory, a global evolutionary stiffness matrix for a long-service hydraulic twin model is constructed. for: In the formula, For optimal mapping parameters The first mapping Regional material stiffness degradation factor The nonlinear boundary contact stiffness matrix of the structure is obtained by inversion from high-frequency micro-vibration data. For the first time in a healthy state The element stiffness matrix of each region This represents the number of physical sub-regions into which the hydraulic structure is divided. Wherein, when This indicates no damage. This indicates varying degrees of aging or microcrack damage. It reflects the evolution of topological boundaries caused by foundation voids or contact surface slippage.

[0047] In S503, based on the reconstructed evolutionary stiffness matrix By applying actual working conditions such as the current reservoir water level, temperature field, and environmental wind load to the long-term hydraulic twin model, the current safe bearing capacity state of the structure is solved in a forward approach, and the full life cycle health index of the long-term hydraulic structure is output. .

[0048] This completes the digital twin closed loop from "non-contact micro-vibration wave field capture" to "synchronous evolution of entity-virtual state".

[0049] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0050] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for updating a long-service hydraulic twin model based on multi-fidelity Bayesian methods, characterized in that, Includes the following steps: S100: Obtain the current non-contact full-field micro-vibration time history data of the hydraulic structure and decouple it into low-fidelity data observations characterizing the overall stiffness of the structure and high-fidelity data observations characterizing the local boundary and nonlinear states. S200, based on the macroscopic physical aging equation of hydraulic structure, constructs a time-varying degradation prior probability distribution with its own historical memory characteristics; S300, Construct a multi-fidelity twin architecture with spatial and mechanistic heterogeneity, including a global low-fidelity proxy model and a local high-fidelity three-dimensional finite element model; S400, introduce the spatial covariance matrix of the full field measurement points, construct a multivariate Gaussian likelihood function by combining low-fidelity data observations and high-fidelity data observations, and perform Bayesian heterogeneous MCMC parameter inference based on spatial correlation on the multifidelity twin architecture based on time-varying degenerate prior probability distribution until the posterior parameter sampling converges; S500: Extract the posterior probability distribution of the physical parameters and boundary nonlinear stiffness of the hydraulic structure to be corrected after convergence, and automatically map the parameters of its high confidence interval back into the long-term hydraulic twin model to complete the synchronous update of the hydraulic structure's full life cycle state.

2. The method for updating a long-service hydraulic twin model based on multi-fidelity Bayes as described in claim 1, characterized in that, In step S100, the extraction of low-fidelity data observations includes: Image domain measurement points in non-contact full-field micro-vibration time history data exist The analytical signal at time step is analyzed, its phase is extracted and the time-domain solution is obtained, the original cumulative displacement is calculated, and a low-pass digital filter is introduced to obtain the low-frequency steady-state displacement sequence. As a low-fidelity data observation, it is represented as: In the formula, This is the original cumulative displacement. This is the cutoff frequency of the low-pass digital filter. It is a low-pass digital filter. To analyze the signal, For time phase untangling operator, is the local spatial wavenumber.

3. The method for updating a long-service hydraulic twin model based on multi-fidelity Bayes as described in claim 1, characterized in that, In step S100, the extraction of high-fidelity data observations includes: For non-contact full-field micro-vibration time history data, the instantaneous phase difference is calculated using the complex conjugate product of the analytical signals between adjacent video frames to obtain the instantaneous velocity response. Then, a bandpass filter or high-pass filter is used to extract specific high-frequency bands, resulting in a high-frequency transient velocity sequence. As a high-fidelity data observation, it is represented as In the formula, For instantaneous speed response, This is the cutoff frequency of the bandpass or highpass filter. It is a bandpass filter or a highpass filter. To analyze the signal, for The complex conjugate of a complex vector at time step.

4. The method for updating a long-service hydraulic twin model based on multi-fidelity Bayes as described in claim 1, characterized in that, In step S200, the time-varying degradation prior probability distribution constructed for the physical parameters to be corrected exhibiting decay characteristics is as follows: In the formula, Let be the mean evolution function. The prior equations are time-varying. For the physical parameters of the hydraulic structure to be corrected, These are the initial design values ​​for the physical parameters of the hydraulic structure to be corrected. As a comprehensive damage and degradation factor based on physical laws, For service time variable, It follows a Gaussian normal distribution.

5. The method for updating a long-service hydraulic twin model based on multi-fidelity Bayes as described in claim 1, characterized in that, S300 includes the following steps: S301. Based on low-fidelity data observations, a lightweight data-driven global proxy model is constructed as a low-fidelity model. S302. Based on high-fidelity data observations, a refined three-dimensional finite element local sub-model is constructed as a high-fidelity model; S303. Based on the multi-fidelity Gaussian process, the low-fidelity model and the high-fidelity model are bridged across scales to obtain a multi-fidelity twin architecture.

6. The method for updating a long-service hydraulic twin model based on multi-fidelity Bayes as described in claim 5, characterized in that, In S303, the bridging equation for cross-scale bridging of the low-fidelity model and the high-fidelity model is: In the formula, For high-fidelity model high-fidelity prediction of response, This represents the low-fidelity prediction vector for the low-fidelity model. This is the scale scaling factor. For high-fidelity residual processes, The probability distribution is a Gaussian process. It is a mean function. For the mean function of the high-fidelity residual process, For parameterized covariance kernel function, The covariance kernel function for a high-fidelity residual process. The input vector consists of the physical parameters of the hydraulic structure to be corrected. This is another set of reference input vectors in the residual space, used to... Jointly calculate spatial correlation. For hyperparameter set; The predicted mean of the high-fidelity predicted response output by the multi-fidelity twin architecture And the prediction variance of the cognitive uncertainty of the characterization model for: In the formula, For low-fidelity models at unknown prediction points The predicted mean at that location, For high-fidelity residual processes at unknown prediction points The predicted mean at that location, For low-fidelity subscripts, The subscript for the residual.

7. The method for updating a long-service hydraulic twin model based on multi-fidelity Bayes as described in claim 1, characterized in that, S400 includes the following steps: S401. Introduce the spatial covariance matrix of all measurement points, combine it with the high-fidelity observation data of the whole field, use the multi-fidelity twin architecture to predict the mean to replace the real high-fidelity model, and construct a multivariate Gaussian likelihood function. S402, Configure independent dual-thread operation to run the asynchronous Markov chain Monte Carlo sampling algorithm; The main thread combines the time-varying degradation prior probability distribution with the multivariate Gaussian likelihood function to perform Metropolis-Hastings sampling. When the predicted variance of the parameter region exceeds the confidence tolerance threshold, the corresponding parameter point is pushed into the background task queue. The background daemon thread extracts parameter points from the task queue, automatically rewrites the input file of the local high-fidelity three-dimensional finite element model through the script, and performs parallel calculations. After the calculation is completed, it extracts the dynamic features from the result file. S403. Dynamically add the parameter points and dynamic features as a set of precise samples to the high-fidelity training set. S404. Repeat S402~S403 to incrementally retrain the hyperparameters of the multi-fidelity twin architecture based on the high-fidelity training set, update the high-fidelity residual process until the Markov chain converges, output the posterior probability distribution of the physical parameters to be corrected, and complete the Bayesian heterogeneous MCMC parameter inference.

8. The method for updating a long-service hydraulic twin model based on multi-fidelity Bayes as described in claim 7, characterized in that, In S401, the constructed multivariate Gaussian likelihood function is: In the formula, For high-fidelity observation data across the entire field, The spatial covariance matrix of all measurement points is... The superscript indicates the candidate point for the new parameter, and the superscript indicates the transpose operator.

9. The method for updating a long-service hydraulic twin model based on multi-fidelity Bayes as described in claim 1, characterized in that, The S500 includes the following steps: S501. Extract the optimal mapping parameters required for the long-term hydraulic twin model, construct the posterior covariance matrix to characterize the uncertainty of identification, and then determine the high-confidence evolution interval of the physical parameters of each hydraulic structure; whereby the optimal mapping parameters are the posterior mean estimates of the physical parameters of the hydraulic structure to be corrected after convergence. S502. The extracted optimal mapping parameters are automatically mapped to the underlying physical equations of the long-term hydraulic twin model. Combined with the material stiffness degradation factor and nonlinear boundary contact stiffness, the global evolution stiffness matrix of the long-term hydraulic twin model is reconstructed. S503. Based on the global evolution stiffness matrix, apply actual working conditions to the long-term hydraulic twin model, solve the current safe bearing capacity state of the structure in a forward manner, and output the full life-cycle health index of the long-term hydraulic structure to complete the synchronous update of the full life-cycle state of the hydraulic structure.

10. The method for updating a long-service hydraulic twin model based on multi-fidelity Bayes as described in claim 1, characterized in that, In S502, the global evolution stiffness matrix for: In the formula, For optimal mapping parameters The first mapping Regional material stiffness degradation factor The nonlinear boundary contact stiffness matrix of the structure is obtained by inversion from high-frequency micro-vibration data. For the first time in a healthy state The element stiffness matrix of each region This represents the number of physical sub-regions into which the hydraulic structure is divided.

Images