A method and system for digital twin simulation of long-span arch bridges using sparse monitoring data
By using adaptive weighted sparse polynomial chaotic expansion and Sobol global sensitivity analysis to screen key parameters, and combining a surrogate model with dynamic and static feature fusion and an adaptive optimization algorithm, the problems of high computational cost, limited samples, and easy distortion of single features in the correction of finite element models of long-span arch bridges are solved, and efficient and accurate structural parameter identification is achieved.
Patent Information
- Application Number
- CN202611341854.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-01
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies face challenges in correcting finite element models of long-span arch bridges, including high computational costs, limited training samples, susceptibility to distortion of single features, and low efficiency of global optimization algorithms, making it difficult to achieve high-precision structural parameter identification and damage assessment.
We employ adaptive weighted sparse polynomial chaotic expansion and Sobol global sensitivity analysis to screen key parameters, construct a surrogate model oriented towards dynamic and static features, and identify key equivalent parameters through a multi-objective collaborative inversion strategy. We then combine this with an adaptive optimization algorithm to improve correction efficiency and accuracy.
It achieves efficient dimensionality reduction and high-precision finite element model correction under sparse monitoring data, improves the stability and engineering interpretability of parameter identification, and reduces computational costs and sample requirements.
Smart Images

Figure CN123116599A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of bridge engineering structural health monitoring technology, and relates to a digital twin simulation method and system for large-span arch bridges under sparse monitoring data. Background Technology
[0002] Long-span arch bridges, due to their strong span capacity, high stiffness, and beautiful design, occupy an important position in the construction of transportation infrastructure in mountainous areas. However, during long-term service, the actual physical state of the bridge structure deviates significantly from the initial finite element model established based on design drawings due to the coupled effects of factors such as solar temperature differences, random heavy traffic, wind-induced vibration, and material aging. Finite element model correction is a crucial link connecting discrete monitoring responses with the overall mechanical state of the bridge. Its core lies in using measured responses to inversely identify key physical parameters of the structure, thereby establishing a corrected analysis model that can characterize the current service state of the actual bridge, providing prerequisites for subsequent damage identification, load-bearing capacity assessment, and life prediction.
[0003] Existing digital twin simulation methods for long-span arch bridges mainly face the following technical challenges and limitations:
[0004] 1. Traditional proxy models rely on massive training samples, making them difficult to adapt to scenarios with high finite element computation costs and few samples.
[0005] Refined finite element models of long-span arch bridges typically have tens of thousands of degrees of freedom, making a single forward simulation (reanalysis) extremely time-consuming. In traditional model correction, while deep learning-based surrogate models (such as deep neural networks) can approximate complex nonlinear response surfaces, their gradient descent mechanisms heavily rely on massive training samples. For real-world bridge correction, the generation of finite element samples is limited by high computational costs, often resulting in only a few dozen to a hundred sets of samples. Under this "small sample" constraint, traditional deep learning methods are prone to overfitting, leading to a sharp decline in prediction stability and making it difficult to meet engineering accuracy requirements.
[0006] 2. Traditional statistical surrogate models have insufficient generalization ability in high-dimensional and strongly nonlinear mappings.
[0007] While traditional statistical surrogate models such as Kriging and response surface methodology are suitable for small to medium sample scenarios, their hyperparameter optimization and extrapolation prediction results are easily affected by sample distribution when facing the high-dimensional physical parameter space and strongly nonlinear mapping of long-span arch bridges, resulting in significant deficiencies in generalization stability. Furthermore, when the training sample size decreases, the prediction error degrades significantly, failing to provide a reliable mapping relationship for subsequent inversion optimization.
[0008] 3. Modifying a single feature can easily lead to the "different parameters, same effect" problem, resulting in distortion of the inversion results.
[0009] In existing model modification practices, using only a single physical field (such as dynamic or static characteristics alone) as the objective function easily leads to the ill-conditioned inverse problem of "different parameters having the same effect." Specifically: when relying solely on dynamic characteristics (such as the first-order fundamental frequency), since the frequency is mainly determined by the overall stiffness-mass ratio, the algorithm cannot effectively distinguish the independent effects of local stiffness changes and mass changes, resulting in mutual compensation of parameters; when relying solely on static characteristics (such as the temperature sensitivity coefficient), the objective function is insensitive to changes in mass parameters, and compensation effects easily occur between the stiffnesses of different components. This insufficient constraint from a single feature leads to severe distortion of the identified physical parameters, resulting in a loss of engineering interpretability.
[0010] 4. Global optimization algorithms lack the ability to adapt to parameter sensitivity, resulting in low optimization efficiency.
[0011] After obtaining the surrogate model, global optimization is required using an optimization algorithm. Traditional metaheuristic algorithms (such as genetic algorithms and particle swarm optimization) or the standard Crowned Porcupine Optimization (CPO) algorithm typically use an indiscriminate iterative step size for all parameters to be corrected, without considering the different sensitivity differences of different parameters to the structural response. This can easily lead to insufficient exploration of highly sensitive parameters and excessive divergence of low-sensitive parameters, which not only increases computational redundancy but may also miss the global optimal solution, affecting the convergence accuracy and efficiency of the correction results.
[0012] In summary, in response to the problems faced by long-term service long-span arch bridges, such as complex monitoring signals, high finite element calculation costs, limited training samples, and the susceptibility of single features to distortion, there is an urgent need for a finite element model correction method and system that can integrate multiple dynamic and static features, adapt to small sample conditions, and have efficient global optimization capabilities. Summary of the Invention
[0013] In view of this, in order to solve the problems of complex monitoring signals, high finite element calculation costs, limited training samples, and easy distortion of single features faced by long-term service long-span arch bridges, this invention provides a digital twin simulation and deduction method and system for long-span arch bridges under sparse monitoring data. The correction method first uses a global sensitivity analysis method to screen key parameters to be corrected, reducing the optimization dimension; secondly, it constructs a proxy model for static and dynamic features based on limited sample and limited data; finally, it combines a multi-objective collaborative inversion strategy of dynamic and static feature fusion to identify key equivalent parameters, and verifies the effectiveness of the proposed framework through numerical bridge examples.
[0014] To achieve the above objectives, the present invention provides the following technical solution:
[0015] A method for digital twin simulation of long-span arch bridges under sparse monitoring data includes the following steps:
[0016] S1. Parameter Selection and Dimensionality Reduction: By introducing sparse polynomial chaotic expansion with fusion of minimum angular regression and Sobol global sensitivity analysis, global sensitivity analysis is performed on multiple initial physical parameters in the finite element model of a long-span arch bridge. The sensitivity differences of each parameter to the dynamic and static response of the structure are quantified, and key parameters to be corrected are selected according to the set sensitivity threshold, thereby reducing the number of parameter types.
[0017] Specifically, in the process of revising the finite element model of a large arch bridge, the initial finite element model contains a large number of uncertain parameters. If all of these parameters are included in the revision, it will lead to an explosion in the dimensionality of the parameter space. Therefore, parameter sensitivity analysis (SA) is introduced as the core strategy for parameter pre-screening. By quantitatively assessing the impact of changes in physical parameters on the structural response, key parameters are identified and redundant variables are eliminated. To achieve evidence-based parameter dimensionality reduction, Sobol Global Sensitivity Analysis (GSA) based on variance decomposition is introduced. Unlike local partial derivatives, Sobol can comprehensively quantify the nonlinear interaction effects of parameters throughout their entire physical range. It does not rely on assumptions of linearity or monotonicity of the model and is widely used for parameter screening in complex engineering systems. Its core idea is to decompose the total output variance of the model into the sum of the partial variances generated by individual parameters and their interactions.
[0018] Let the response function of the finite element model of the arch bridge be... , where the input parameter vector In a unit hypercube Independently distributed within the range. According to ANOVA decomposition theory, the function can be expressed as:
[0019]
[0020] In the formula: This is the mean.
[0021] Based on the above decomposition, the total variance D can be expressed as the sum of the partial variances of each order:
[0022]
[0023] Based on this, Sobol defined two types of core sensitivity indices for parameter evaluation:
[0024] Main effect index ( ): Measure a single parameter Direct contribution to output variance.
[0025]
[0026] Total effect index ( ): Measured parameters The total contribution of the variance to the output variance, including its interaction with all other parameters.
[0027]
[0028] In the formula: Indicates except All parameters except those mentioned above.
[0029] Traditional Sobol methods rely on massive Monte Carlo sampling, which is computationally too cumbersome for refined finite element models of large arch bridges. Conventional sparse PCE also suffers from drawbacks such as over-reliance on fixed criteria for basis function selection, orthogonality degradation under small samples, and uneven fitting biases between dynamic and static responses. To address these issues, an adaptive weighted sparse polynomial chaotic expansion method (AW-SparsePCE) is constructed. This method builds an algorithm through weighted orthogonal correction and adaptive sparse selection, and then uses this algorithm to perform analytical calculations of the Sobol exponent, resulting in an efficient global sensitivity analysis method suitable for large and complex bridge structures.
[0030] The PCE series approximation form of the standard polynomial chaotic expansion structural response is:
[0031]
[0032] In the formula: These are multivariate orthogonal basis functions; The expansion coefficients are to be determined. This is the set of basis function indices.
[0033] When the sample size is insufficient, PCE suffers from severe degradation of basis function orthogonality, leading to distorted sensitivity calculations. Furthermore, sparse PCE uses fixed L1 regularization, which cannot accommodate the nonlinear differences between the smooth static response and the sensitive dynamic response of bridges. By introducing a weighted orthogonality correction term, the orthogonality constraints of the basis functions are reconstructed:
[0034]
[0035] In the formula: The orthogonal correction weights are determined adaptively through iteration. The correlation coefficient is the basis function.
[0036] Furthermore, an adaptive weighted LARS algorithm is proposed to construct a differentiated sparse optimization objective function:
[0037]
[0038] In the formula: The weighting coefficients are for static and dynamic responses: static response has a lower weight, and dynamic response has a higher weight. The adaptive regularization parameter shrinks gradually with each iteration to reduce the risk of overfitting or underfitting. It is an L1 sparse constraint.
[0039] The improved AW-Sparse PCE final approximation is as follows:
[0040]
[0041] In the formula: The set of sparse basis functions after adaptive filtering; These are the optimal coefficients for the weighted least squares solution.
[0042] Based on the constructed adaptive weighted sparse PCE, the Sobol sensitivity exponent can be directly calculated analytically through coefficient expansion, thus improving efficiency.
[0043] Total variance analysis calculation:
[0044]
[0045] Main effect index (improved analytical formula):
[0046]
[0047] Total Effects Index (Improved Analytical Formula):
[0048]
[0049] like Significantly greater than This indicates that the parameter participates in a strong nonlinear interaction. The core objective of sensitivity analysis is to compress the parameter space. In practice, a threshold is usually set (e.g., ...). <0.05), parameters with extremely low sensitivity exponents are treated as insensitive parameters and set as constants, thus transforming the high-dimensional optimization problem into a low-dimensional parameter identification problem.
[0050] S2. Construction of a Small-Shot Learning Proxy Model: Utilizing the context learning mechanism of the TabPFN network fitted with prior tabular data, and combining it with the finite parameter-response sample set generated under the constrained finite element sample conditions of step S1, a small-shot proxy model for large-scale complex bridges oriented towards dynamic and static response characteristics is constructed. This establishes a nonlinear mapping relationship between the key parameters to be corrected and the dynamic and static response characteristics of the structure. The core of the small-shot proxy model for large-scale complex bridges is to construct a high-precision mapping relationship between structurally sensitive parameters and dynamic and static response characteristics. Specifically, it includes the following steps:
[0051] S21. Data Acquisition and Sample Set Construction: Select m core sensitive parameters for the whole-bridge health monitoring of long-span arch bridges, which are screened by global sensitivity analysis. Use orthogonal Latin hypercube sampling (OLHD) to generate n sets of samples. Extract static response characteristics and dynamic response characteristics through finite element model calculation and analysis. Then construct a parameter-response sample set and divide it into a training set with a small number of samples and a test set for verification.
[0052] Specifically, m core sensitive parameters As input vectors, each parameter follows a uniform distribution within its physically reasonable range; static response characteristics and dynamic response characteristics Synchronous fitting as the output vector: .
[0053] The Orthogonal Latin Hypercube Sampling (OLHD) algorithm is used to remove outliers and non-stationary noise from sensor monitoring data, ensuring that the sample data used for model correction has high quality and high signal-to-noise ratio.
[0054] S22. Construct an improved MFA-TabPFN dual-channel surrogate model. First, use a multi-feature alignment module to standardize the input multi-source heterogeneous data using the Z-score adaptive normalization method, and introduce a feature alignment attention mechanism to enhance the model's ability to fuse features of different physical dimensions. Simultaneously, construct an improved loss function that fuses mean squared error loss (MSE) and KL divergence loss to form a composite loss function. Finally, the weights of the TabPFN backbone network were frozen, and the multi-evidence alignment module and composite loss function were configured to complete the MFA-TabPFN initialization, providing an improved model foundation for subsequent few-shot learning.
[0055] When TabPFN is directly applied to the synchronous fitting of heterogeneous responses of dynamic and static types, it may be affected by differences in dimensions, sensitivity, and output scale, leading to uneven prediction accuracy of different response features. To address this, TabPFN is improved from two dimensions: the feature encoding layer and the loss constraint layer. A multi-feature alignment improved TabPFN surrogate model is proposed, denoted as MFA-TabPFN (Multi-Feature Alignment TabPFN). Here, MFA indicates that feature normalization and feature alignment mechanisms mitigate the differences in dimensions, scale, and sensitivity between dynamic and static response features.
[0056] Specifically, to address the differences in dimensions and sensitivity between static and dynamic features, an adaptive normalization encoding and feature alignment attention mechanism are introduced. This mechanism uniformly encodes the output features and performs Z-score adaptive normalization on the static and dynamic response features respectively, thus eliminating the differences in dimensions.
[0057]
[0058] In the formula: , These are the mean and standard deviation of the static characteristics, respectively. , These represent the mean and standard deviation of the dynamic characteristics, respectively.
[0059] By incorporating a feature-aligned attention branch, the association weights between static and dynamic features and input parameters are calculated separately, adaptively adjusting the contribution of the two types of features in the mapping to address the problem of uneven fitting accuracy.
[0060]
[0061] In the formula: The adaptive weights for dynamic and static features are learned through a multilayer perceptron (MLP) and are dynamically adjusted according to the input parameters.
[0062] To address the overfitting problem that may occur in surrogate models under limited sample scenarios, at the method design level, a composite loss function containing a mean squared error term and a prediction uncertainty constraint term can be introduced to describe the constraint idea of model output stability:
[0063]
[0064] In the formula: To compensate for the mean squared error loss between the predicted and actual values, and to ensure fitting accuracy; To constrain the prediction uncertainty of the model and avoid overfitting, the KL divergence loss between the predicted distribution and the Bayesian posterior distribution is calculated. The regularization weight coefficients are adaptively determined through cross-validation. The improved MFA-TabPFN dual-channel surrogate model, while retaining the advantages of TabPFN's small-sample context learning, further alleviates the fitting bias problem of heterogeneous features with dynamic and static multi-output features, and better meets the engineering needs of correcting large and complex bridge models.
[0065] S23. Model Initialization and Few-Shot Learning: Load the pre-trained TabPFN backbone network weights, and then configure the multi-feature alignment module and composite loss function designed in step S22 to complete the initialization of the MFA-TabPFN dual-channel surrogate model. Then, using the In-Context Learning (ICL) mechanism, the initialized MFA-TabPFN dual-channel surrogate model is iteratively fine-tuned under the condition of a small number of training samples in step S21 to achieve high-precision inversion of key parameters of the finite element model, providing a reliable learning model foundation for subsequent model prediction and validation.
[0066] S24. Model Prediction: Input the test parameter samples from the test set in step S21 into the MFA-TabPFN dual-channel surrogate model trained in step S23, perform a single forward propagation calculation, output the predicted values of the bridge's dynamic and static responses and compare them with the true values, then determine whether the prediction accuracy meets the engineering requirements; if it does, a high-precision multi-feature aligned MFA-TabPFN surrogate model is constructed; if it does not, a feedback mechanism is triggered, the sample size is replenished and the previous steps S22 and S23 are iterated repeatedly until the accuracy meets the engineering requirements, and finally a high-precision surrogate model that can be used for the correction of the finite element model of a long-span arch bridge is obtained.
[0067] S3. Fusion of dynamic and static features and global optimization: Construct a multi-objective function for fusion of dynamic and static features, perform residual matching between the structural response predicted by the small sample surrogate model in step S2 and the dynamic and static target features extracted from the actual measurements; rely on the CPO algorithm to perform global optimization of the multi-objective function, invert and identify the optimal key physical parameter correction coefficients, and complete the update of the finite element model.
[0068] The MFA-TabPFN dual-channel surrogate model constructed in step S2, and the essence of the finite element model correction, can be reduced to a constrained nonlinear inversion optimization problem. The core objective is to find a set of optimal physical parameter vectors within the physically reasonable range of the parameters. This minimizes the residual between the structural response predicted by the surrogate model and the target response. However, for scenarios such as large and complex bridges with sparse measurement points and high parameter dimensionality, using only a single dynamic or static feature as the correction target will face serious ill-posedness and heterogeneous parameter isomorphism problems. That is, multiple combinations of stiffness and mass parameters may output similar structural response characteristics, resulting in non-uniqueness of the correction results and difficulty in parameter interpretation.
[0069] To address the aforementioned core challenges, a multi-objective collaborative inversion framework integrating dynamic and static features is constructed. By compressing the feasible solution space through complementary constraints of different response features, the problem of different parameters having the same effect in single feature correction is alleviated. At the same time, to address the shortcomings of the CPO algorithm, such as indiscriminate optimization step size and iterative redundancy, an adaptive step size improvement strategy based on Sobol sensitivity prior and an adaptive early stopping mechanism are proposed to achieve efficient and high-precision global optimization of the correction parameters.
[0070] Because static and dynamic characteristics differ significantly in dimensions and numerical magnitudes, directly using absolute residuals to construct the objective function would cause the numerically dominant static term to dominate the descent direction of the optimization gradient, rendering the constraints of the dynamic characteristics ineffective and ultimately reducing the physical interpretability of the corrected results. Therefore, a dimensionless objective function based on the relative residual sum of squares is constructed. :
[0071]
[0072] In the formula: The vector of physical parameters to be corrected; , For the TabPFN proxy model in parameters The predicted value at that location; , The true value of the target feature identified based on measured data; , The weighting systems for static and dynamic characteristics respectively satisfy... The allocation of weight coefficients is determined based on the physical properties, statistical robustness, and parameter identification ability of the two types of features.
[0073] For the constructed multi-objective optimization function, the CPO optimization algorithm is used to perform global optimization. Compared with traditional genetic algorithms and particle swarm optimization algorithms, the CPO algorithm has strong global search capabilities and is applicable to complex nonlinear optimization problems. However, the traditional CPO algorithm uses an indiscriminate iteration step size for all parameters to be corrected, without considering the different sensitivity differences of different parameters to the structural response. This easily leads to insufficient exploration of highly sensitive parameters and excessive divergence of low-sensitive parameters. At the same time, the termination criterion of a fixed number of iterations can easily cause computational redundancy or insufficient convergence. Therefore, the traditional CPO algorithm is improved from two dimensions: adaptive adjustment of the optimization step size and early stopping mechanism for iteration termination.
[0074] Based on the Sobol sensitivity index calculated in step S1, an adaptive adjustment mechanism for the parameter optimization step size is constructed, making the algorithm's exploration intensity positively correlated with the parameter sensitivity. The expression is:
[0075]
[0076] In the formula: Let be the exploration step size of the i-th parameter in the t-th iteration; The initial reference step size for the i-th parameter is determined by the physical value range of the parameter. Let be the Sobol sensitivity index of the i-th parameter, representing the total contribution of this parameter to the structural response.
[0077] By assigning larger exploration step sizes to highly sensitive parameters, the global optimization capability is enhanced, and the omission of optimal solutions is avoided. On the other hand, compressing the exploration step size for low-sensitive parameters and constraining their fluctuations within a physically reasonable range helps to improve the stability of the search process and reduce the computational cost caused by invalid fluctuations of low-sensitive parameters.
[0078] To balance optimization accuracy and computational efficiency, and to avoid unnecessary redundant iterations, an adaptive early stopping mechanism is embedded in the algorithm framework. This mechanism monitors the changing trend of the optimal fitness value of the population in real time, and the convergence criterion is set as follows:
[0079]
[0080] In the formula: Let be the optimal objective function value of the t-th generation population; To continuously monitor the number of iteration steps, the baseline value is taken as follows: =5; The benchmark value is taken as the convergence threshold. = It can be flexibly adjusted according to accuracy requirements.
[0081] When the convergence criterion in the above formula is met, it is determined that the algorithm has converged to a stable optimal solution, and the iteration process is automatically terminated without having to complete the preset maximum number of iterations, thus improving computational efficiency while ensuring accuracy.
[0082] The beneficial effects of this invention are as follows:
[0083] 1. The digital twin simulation method for long-span arch bridges under sparse monitoring data disclosed in this invention, by introducing adaptive weighted sparse polynomial chaotic expansion (AW-Sparse PCE) and Sobol global sensitivity analysis, can quantitatively identify insensitive parameters that contribute little to the dynamic and static responses of the structure from a large number of initial physical parameters and fix them as constants. This method achieves reasonable dimensionality reduction of the high-dimensional parameter space, avoids the "dimensional explosion" problem caused by overall correction, significantly narrows the search range of subsequent optimization algorithms, and thus improves the overall computational efficiency of finite element model correction.
[0084] 2. The digital twin simulation method for long-span arch bridges under sparse monitoring data disclosed in this invention utilizes the pre-training prior and context learning (ICL) mechanism of the TabPFN (Tab Prior Data Fitting Network) to construct a small-sample surrogate model for finite element model correction. Unlike traditional deep learning, this method does not require time-consuming iterative training; it only needs a small number of finite element samples as contextual cues to establish the parameter-response mapping relationship in a single forward propagation. This effectively alleviates the bottleneck of high single-cycle computational cost and limited training sample acquisition for refined finite element models of large bridges, providing a feasible approach for high-precision surrogate modeling under small-sample conditions.
[0085] 3. The digital twin simulation method for long-span arch bridges under sparse monitoring data disclosed in this invention proposes a multi-feature alignment improved TabPFN (MFA-TabPFN) model and introduces... The adaptive normalization and feature alignment attention mechanism eliminates output bias caused by differences in physical dimensions, numerical magnitudes, and sensitivity between static and dynamic features. This design enables the surrogate model to learn the two types of heterogeneous response features in a balanced manner, avoiding the negative transfer phenomenon caused by feature scale differences in single multi-output regression, and improving the stability and reliability of parameter-response mapping.
[0086] 4. The digital twin simulation method for long-span arch bridges under sparse monitoring data disclosed in this invention constructs a dimensionless multi-objective function that fuses static and dynamic features, and weightedly and collaboratively constrains the temperature sensitivity coefficient (static feature) and the structural fundamental frequency (dynamic feature). Static features are sensitive to stiffness changes, while dynamic features are sensitive to changes in the mass-stiffness ratio. These two features form complementary orthogonal constraints in the parameter space, effectively compressing the feasible solution space and reducing the risk of mutual compensation between different parameters, thereby improving the physical interpretability and engineering credibility of the inversion results.
[0087] 5. The digital twin simulation method for long-span arch bridges under sparse monitoring data disclosed in this invention, based on the Sobol exponent obtained from sensitivity analysis, adaptively improves the iteration step size of the traditional CPO algorithm, increasing the exploration step size for highly sensitive parameters and compressing the step size for low-sensitive parameters. This differentiated optimization strategy enhances the global exploration capability of key parameters while suppressing the ineffective divergence of redundant parameters. Combined with an adaptive early stopping mechanism, it reduces unnecessary computational redundancy while ensuring convergence accuracy, further improving the engineering practicality of the correction framework.
[0088] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0089] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:
[0090] Figure 1 This is a flowchart illustrating the construction of the few-shot learning agent model of the present invention;
[0091] Figure 2 This is a graph showing the global sensitivity analysis results of Sobol in an example.
[0092] Figure 3 This is a schematic diagram illustrating the physical and mechanical interpretation of the physical parameter identification results in the example.
[0093] Figure 4 This is a spatial distribution diagram of the sample in the example.
[0094] Figure 5 The example is a CPO convergence curve.
[0095] Figure 6 The following is a comparison of modal shapes and MAC values before and after the modification in the example;
[0096] Figure 7 This is a verification diagram of the correlation coefficient of the spatial morphology of the entire bridge in the example embodiment;
[0097] Figure 8 (a) is a comparison diagram of the spatial morphology on the left side of the embodiment. Figure 8 (b) is a comparison diagram of the spatial morphology on the right side of the embodiment. Detailed Implementation
[0098] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention.
[0099] Example
[0100] This study focuses on the correction and verification of a long-serving, large-span, upper-bearing steel truss arch bridge using the finite element method (FEM) model. First, based on the bridge's current service status and the initial FEM model, a set of parameters to be corrected is constructed, and dimensionality reduction of high-dimensional parameters is achieved through global sensitivity analysis. Second, a surrogate model is trained based on FEM calculation samples, and a small-sample learning surrogate modeling method is introduced. A multi-objective collaborative inversion strategy integrating dynamic and static features is employed to identify key physical parameters of the bridge. Finally, the corrected model is verified in multiple dimensions from the perspectives of dynamic characteristics and quasi-static spatial deformation. The aim is to establish a corrected FEM model of the bridge that shows good consistency with the current monitoring response, providing a foundation for corrected analysis models for subsequent structural state assessment, response analysis, and engineering applications.
[0101] S1. Parameter Definition of Initial Finite Element Model of Real Bridge
[0102] Based on the three-dimensional finite element baseline model of the entire bridge, 14 key physical parameters that have potential influence on the dynamic and static responses of the structure were selected and parameterized. Regarding material parameters, considering the potential reduction in overall stiffness due to semi-rigid connections at the nodes and the uncertainty of the secondary dead load distribution, elastic modulus correction coefficients were defined for the main arch ring and the main beam, respectively. , ) and density correction factor ( , In addition, the coefficient of thermal expansion of the steel used in the entire bridge was specifically introduced. The parameters to be corrected are used to characterize the overall temperature deformation response of the structure. Considering the aging and wear of the support system under long-term service, the combined spring element COMBIN14 is used to simulate the support boundary at the junction pier and column. The equivalent spring stiffness in the longitudinal, transverse, and vertical directions is defined respectively. , , This allows for limited elastic deformation and rotation at the boundary. All parameters to be analyzed and their initial value ranges are detailed in Table 1. These parameters collectively constitute the high-dimensional input space for subsequent global sensitivity analysis.
[0103] Table 1 Definitions and value ranges of physical parameters to be corrected
[0104]
[0105] S2. Implementation of Correction of Real Bridge Finite Element Model Based on Dynamic and Static Feature Fusion and Small Sample Learning
[0106] S21. Screening of parameters to be corrected and sample generation
[0107] Based on the constructed adaptive weighted sparse PCE and Sobol global sensitivity analysis parameter selection framework, efficient dimensionality reduction of the high-dimensional parameter space was achieved for 14 defined initial parameters to be corrected, identifying the core sensitive parameters that have a significant impact on the dynamic and static responses of the structure, and constructing a sample dataset. The temperature sensitivity coefficient k = 6.1778 mm / ℃ was used as the static feature, and the first-order vertical fundamental frequency f = 0.6226 Hz was used as the dynamic feature.
[0108] Sample generation is a core prerequisite for response solving and sensitivity analysis. Employing OLHD technology, compared to ordinary Latin hypercube sampling, OLHD relies on algorithms to iteratively optimize the spatial arrangement of sample points within a multidimensional hypercube, adhering to the principle of maximizing the minimum distance to ensure excellent spatial filling and uniformity of the sample points. After sample generation, ANSYS APDL parametric batch processing technology is used to batch call finite element models to complete calculations, accurately determining the core structural responses corresponding to the samples. This constructs a sample-response-based foundation dataset, achieving precise matching between input parameters and the dynamic and static responses of the structure.
[0109] Based on AW-Sparse PCE, a 14-dimensional mapping relationship between input parameters and dynamic and static responses was constructed. Using the Sobol global sensitivity analysis method, the contribution of each parameter to the temperature sensitivity slope and the first-order vertical fundamental frequency was quantified. The main effect index and total effect index of each parameter were calculated. The calculation results are as follows: Figure 2 As shown, for the temperature sensitivity slope k characterizing the static properties of the structure, the coefficient of thermal expansion of the steel in the entire bridge is... It exhibits absolute control, with its Sobol global index approaching 1.0. This result indicates that for long-span steel arch bridges, the quasi-static temperature deformation mainly originates from the thermal expansion and contraction effect of the material itself. Although the stiffness distribution of a structure affects the redistribution of internal forces, from the perspective of macroscopic deformation trends, the coefficient of thermal expansion is the fundamental factor determining the slope of the temperature-displacement linear relationship. Furthermore, the figure shows the stiffness parameters of various supports. , The sensitivity to slope is extremely low. This indicates that under normal service conditions, the elastic constraint of the support is approximately an ideal boundary relative to the main arch ring with a huge span, and its stiffness fluctuation is insufficient to change the quasi-static temperature deformation of the entire bridge.
[0110] For the first-order vertical fundamental frequency, which reflects the overall stiffness and mass distribution of the structure, the sensitivity exhibits characteristics driven by multiple parameters. (Main arch ring elastic modulus) The sensitivity index is the highest, at approximately 0.63. This is because the main arch ring, as the primary load-bearing component, provides the majority of the bridge's vertical bending stiffness, directly determining the upper limit of the fundamental frequency. Main beam density. Density of the main arch Following closely behind, the sensitivity indices were approximately 0.20 and 0.09, respectively. Based on dynamic principles, modal frequencies are highly sensitive to changes in mass. In particular, the significant impact of the main girder density suggests that secondary dead loads such as bridge deck pavement, guardrails, and pipelines account for a large proportion of the overall bridge modal mass, and that there may be estimation biases in the initial model.
[0111] Based on engineering experience, a sensitivity threshold of 0.05 was set. Parameters whose contribution to both types of target characteristics was below the threshold were removed, while parameters covering the three major physical properties of stiffness, mass, and thermodynamics were retained to ensure the physical completeness of the correction results. Finally, four core parameters to be corrected were selected from the 14 initial parameters, namely: the correction coefficient of the equivalent elastic modulus of the main arch ring. The overall stiffness matrix that controls the structure macroscopically. The equivalent density correction factor for the main beam. Macroscopically, this controls the modal quality of the structure. The equivalent density correction factor for the main arch ring. : Auxiliary correction for the mass distribution of the structure, especially for mass compensation of the ancillary facilities in the arch rib section. Equivalent thermal expansion correction coefficient for the steel of the entire bridge. : The reference physical quantity for controlling the thermal response of a structure.
[0112] Due to the aforementioned physical mechanism, the range of values for each parameter is not a simple symmetrical interval, but rather takes into account significant bias. Considering the semi-rigidity and damage effects, its true value is likely lower than or close to the theoretical design value, and may only be slightly higher than the design value under extreme working conditions of material property dispersion. Accordingly, the value range of the elastic modulus correction coefficient of the main arch ring is set to [0.80, 1.05], and the interval of [0.80, 1.00] corresponding to the stiffness reduction is set as the priority search area for parameter inversion; , Among these factors, the mass of the main girder is significantly affected by deviations in the second-phase dead load construction, bridge deck paving updates during operation, and damage from the addition of ancillary facilities, making it prone to large positive deviations. The mass of the main arch ring, however, is only affected by the inherent material dispersion and local defects, exhibiting a smaller fluctuation range. Therefore, the range for the density correction coefficient of the main girder is set to [0.80, 1.30], and the range for the density correction coefficient of the main arch ring is set to [0.90, 1.10]. The range [1.00, 1.30] corresponding to an increase in the mass of the main girder is then designated as the priority search area. Considering the material's discreteness and constraint effects, the interval is set to [0.85, 1.15] to cover possible bidirectional deviations. A detailed schematic diagram is shown below. Figure 3 As shown.
[0113] Again, OLHD technology was used to generate 100 sets of finite element samples in 4D parameter space, such as... Figure 4 As shown, this ensures that the sample points maintain good spatial filling within a physically reasonable range, reducing the possibility of clustering or gaps in random sampling.
[0114] Based on the generated sampling matrix A parameterized batch processing program for ANSYS APDL was developed. The program automatically reads each line of parameter samples, updates the material properties and real constants of the finite element model, and performs static thermal analysis and modal analysis to extract the corresponding temperature sensitivity slope and first-order vertical fundamental frequency, forming a response matrix. Finally, the finite element sample dataset D is constructed to drive the TabPFN few-shot learning agent model:
[0115]
[0116] Wherein, input vector It contains 4 key physical parameters and output vectors. It contains two dynamic and static feature indices. Although the dataset is limited in size, after sensitivity dimensionality reduction and OLHD sampling design, it can provide relatively uniform sample coverage within a given parameter space, providing a data foundation for subsequent surrogate modeling and parameter inversion.
[0117] S22, Training and Modification of Real-Bridge Proxy Model and Fusion of Dynamic and Static Features
[0118] Before correcting the global bounding box of the finite element model constructed using the improved TabPFN and CPO algorithms, data preprocessing is a crucial step in ensuring the numerical stability and prediction accuracy of the model. Although the physical parameter correction coefficients at the input ( , , All are dimensionless quantities distributed within the interval [0.85, 1.15], but the structural response characteristics at the output end show significant differences in magnitude. To address the dimensional differences and numerical magnitude gap between the dynamic and static responses of the actual bridge, a Min-Max normalization strategy is adopted to linearly map the input parameter matrix and the output response matrix to a [0,1] normalized hypercube. The normalization formula is:
[0119]
[0120] In the formula: , These are the upper and lower limits of the physical range for the corresponding parameters; , These are the minimum and maximum values of the corresponding response samples, respectively.
[0121] To address the physical heterogeneity of the static and dynamic characteristics of real bridges, the MFA-TabPFN surrogate model adapted to real bridges constructed in this paper is adopted. To prevent the negative transfer phenomenon caused by forcibly using a single multi-output regression model to fit these static and dynamic characteristics simultaneously, a dual-channel independent modeling strategy is adopted in the algorithm architecture.
[0122] The specific implementation logic is as follows:
[0123] Static prediction channel: Instantiate the first TabPFN regressor, load only the normalized context samples of physical parameters → temperature slope, and focus on learning the mapping of static features.
[0124] Dynamics Prediction Channel: Instantiate a second TabPFN regressor, load only the normalized context samples of physical parameters → modal fundamental frequencies, and focus on learning the mapping of dynamic features.
[0125] Based on the trained dual-channel TabPFN proxy model, the CPO algorithm is improved by relying on sensitivity priors to carry out global inversion of the physical parameters of the real bridge, and the optimal combination of physical parameters is locked by orthogonal constraints of dynamic and static features.
[0126] For the dimensionless objective function based on the relative residual sum of squares mentioned above:
[0127]
[0128] To reduce the impact of weight values on the inversion results, a weight sensitivity analysis was conducted, taking into account the physical properties and statistical stability of the static and dynamic target characteristics of the actual bridge. Four typical weight combinations were set up for comparison, mainly examining the stability of target matching error, convergence iteration steps, and parameter change trends under different weight conditions. It should be noted that Table 2 is mainly used for trend analysis in the weight selection stage, and the final corrected results are based on Tables 3 and 4.
[0129] Table 2 Comparison of target matching and parameter trends under different weight combinations
[0130]
[0131] The comparative results show that when the weight combination is (0.7, 0.3), both static and dynamic characteristics maintain a low error level, the number of convergence iterations is small, and the corresponding parameter change trend is well consistent with the engineering experience under long-term service conditions of actual bridges. Therefore, this weight combination is selected as the objective function weight for subsequent actual bridge parameter inversion. The results also indicate that the temperature sensitivity slope, as a quasi-static feature obtained based on long-term monitoring data, has relatively good statistical stability; appropriately increasing its weight helps improve the convergence stability of the inversion process. However, considering the important constraint effect of dynamic characteristics on the structural mass-stiffness relationship, a certain frequency weight still needs to be retained to avoid the correction results from over-reliance on a single static feature.
[0132] To address the optimization requirements in the 4D parameter space of real bridges, the improved CPO algorithm was engineered and adapted. The core parameters were set as follows: population size 30, maximum number of iterations 50, and the adaptive early stopping criterion was that the rate of change of the optimal objective function value over five consecutive generations was less than [a certain value]. The parameter search boundary remains consistent with the physical range set in step S1. Simultaneously, a sensitivity prior adaptive step-size mechanism is introduced, dynamically adjusting the exploration step size for each parameter based on the Sobol total effect sensitivity index calculated in step S1: a larger exploration step size is assigned to the elastic modulus and thermal expansion coefficient of the high-sensitivity main arch ring to enhance global optimization capabilities; the exploration step size is compressed for the density coefficient of the low-sensitivity main arch ring to constrain its fluctuation within a physically reasonable range, avoiding ineffective searches and parameter divergence. The convergence curve is shown below. Figure 5 As shown.
[0133] Based on the above optimization process, the correction coefficients for the four key physical parameters finally locked are shown in Table 3. The results show that the correction coefficient for the equivalent thermal expansion coefficient of the steel of the entire bridge is 0.9527, indicating that the correction model needs to use an equivalent thermal expansion parameter slightly lower than the nominal value to match the actual bridge temperature response; the elastic modulus coefficient of the main arch ring is 0.9231, indicating that the model needs to use a relatively low overall equivalent stiffness to characterize the current actual bridge response; the density coefficients of the main beam and the main arch are corrected to 1.0819 and 0.9918 respectively, which can be understood as an equivalent characterization of the differences in the dead load of the bridge deck system and the mass distribution of the main arch.
[0134] Table 3. Results of Physical Parameter Correction for Finite Element Model
[0135]
[0136] The material parameters, density parameters, and thermal expansion parameters in the table are dimensionless correction coefficients relative to the initial finite element model values. The initial value of 1.0 indicates that the initial model parameters remain unchanged. The actual parameter values used in the calculation are the products of the initial model parameter values and the corresponding correction coefficients. Substituting the corrected parameters into the initial finite element model of the actual bridge completes the parameter update of the full bridge model, resulting in a corrected finite element model that has good consistency with the current monitoring response.
[0137] S23, Comparative Analysis of Real Bridge Correction Agent Model
[0138] To further illustrate the applicability of our proposed method in the scenario of correcting real-bridge finite element models using few-shot learning methods, this section further selects ordinary TabPFN, XGBoost, BP neural network, and Kriging surrogate models as comparative methods, while maintaining consistency in the set of parameters to be corrected, training samples, and target features. Ordinary TabPFN is used to analyze the correction effect when only the TabPFN surrogate model is introduced, without employing multi-feature alignment, sensitivity prior weights, and improved CPO strategies; XGBoost represents a relatively new ensemble learning surrogate model; and BP neural network and Kriging are compared as traditional neural network surrogate models and classical statistical surrogate models, respectively.
[0139] It should be noted that the standard TabPFN, XGBoost, BP neural network, and Kriging surrogate model are all used as baseline methods in this section, solely for fitting the parameter-response mapping relationship. They employ an equal-weight objective function and random search for parameter inversion, without introducing additional strategies such as parameter importance weights, sensitivity prior adaptive step sizes, or improved CPO. Our proposed method, however, builds upon the improved TabPFN surrogate modeling by further incorporating dynamic and static feature fusion, parameter sensitivity priors, and an improved CPO optimization strategy. This comparison allows us to analyze the effectiveness of our improved strategy in a finite-sample correction scenario for real bridges.
[0140] The corresponding comparison results are shown in Tables 4 and 5.
[0141] Table 4 Comparison of Real Bridge Target Feature Matching Results Using Different Methods
[0142]
[0143] Table 5 Comparison of Real Bridge Correction Parameter Identification Results Using Different Methods
[0144]
[0145] As shown in Table 4, different surrogate models can all achieve target feature matching of the real bridge to a certain extent, but their error distribution and parameter identification results differ significantly. The method proposed in this paper achieves the highest matching accuracy on the static target of temperature sensitivity coefficient, with a slope relative error of only 0.0065%, which is significantly lower than that of ordinary TabPFN, XGBoost, BP neural network and Kriging method. Since the temperature sensitivity coefficient is derived from the long-term temperature-displacement statistical relationship, it is a quasi-static target feature with high weight in the real bridge correction in this paper. Therefore, this result shows that the method proposed in this paper has a better fitting ability for the main control feature of the temperature response of the real bridge.
[0146] From the perspective of dynamic targets, the first-order vertical frequency errors of ordinary TabPFN and XGBoost are 0.0163% and 0.0102%, respectively, both lower than the 0.1446% of the proposed method. This indicates that some ordinary surrogate models can also achieve smaller residuals for single-frequency targets. However, it should be noted that finite element model correction does not only pursue the minimum of a single response error, but should also consider the comprehensive matching of dynamic and static targets, parameter identification stability, and engineering interpretability. Although the proposed method does not minimize the single-term frequency error, its frequency error is still controlled at a low level, while significantly improving the matching accuracy of static targets and demonstrating better coordination between dynamic and static targets.
[0147] Table 5 further shows that there are significant differences in the parameter combinations obtained by different methods. The equivalent elastic modulus coefficient of the main arch ring of the ordinary TabPFN is 0.8981, which is lower than the result of the method in this paper, indicating that it may compensate for the target response error by adjusting the stiffness parameter significantly under the current comparison criteria. Although XGBoost achieves a smaller error in the frequency target, its equivalent density coefficient of the main beam reaches 1.1959 and the equivalent density coefficient of the main arch ring reaches 1.0747, showing a certain tendency to compensate for mass parameters. The equivalent density coefficient of the main beam of the BP neural network reaches 1.2889, which is close to the upper limit of the search interval, indicating that it is easy to achieve target fitting by amplifying local parameter changes under limited sample conditions, and its parameter stability and engineering interpretability are relatively weak. The parameters obtained by the Kriging method are generally close to the results of this paper, but its static and dynamic target errors are higher than those of the method in this paper, indicating that its comprehensive fitting ability is still insufficient under the current limited sample correction scenario of real bridges.
[0148] Based on the combined target feature matching and parameter identification results, it can be seen that the ordinary TabPFN and XGBoost can achieve smaller errors in some individual responses, but their parameter identification results exhibit a certain degree of compensation. While the BP neural network and Kriging method can achieve basic target matching, they are inferior to the proposed method in terms of error control and parameter stability. In contrast, the improved TabPFN+CPO method in this paper demonstrates greater stability in static master target matching, dynamic-static target integration coordination, and engineering interpretation of parameter results. Therefore, the proposed method has better applicability for the current real bridge correction problem, which faces limitations in finite element samples, coupling between dynamic and static targets, and the risk of parameter compensation.
[0149] S3. Physical verification of the real-bridge response of the modified model.
[0150] S31. Error Analysis of Modal Frequency and Mode Shape (MAC)
[0151] To comprehensively evaluate the physical fidelity of the modified model at the dynamic level, this section selects the experimentally identified first-order vertical bending mode as the benchmark and compares and analyzes the differences in dynamic characteristics between the initial finite element model and the modified finite element model. The evaluation metrics include relative frequency error and modal confidence criterion (MAC). The former quantifies the numerical accuracy of eigenvalues, while the latter measures the geometric consistency of the mode shapes corresponding to the eigenvectors from the perspective of vector space.
[0152] In performing modal data matching, this section selects the acceleration response of the main girder as the identification basis, rather than the arch rib measuring points. This is mainly based on two considerations: First, for a deck-type steel truss arch bridge, the mass distribution of the pavement layer, guardrails, and vehicle loads is concentrated in the bridge deck system, and the main girder is the mass control body for dynamic characteristics; second, under environmental excitation, the signal-to-noise ratio of the vibration response of the directly stressed bridge deck system is significantly better than that of the arch rib, and the modal parameters identified accordingly can more accurately reflect the overall dynamic behavior of the structure.
[0153] Table 6 shows the comparison results of the first-order vertical fundamental frequency of the structure before and after the correction. Analysis shows that the initial model, because it was built solely based on design drawings, failed to accurately account for the secondary dead load mass of the pavement layer, crash barriers, and various ancillary facilities, resulting in a calculated fundamental frequency higher than the measured and identified fundamental frequency, with a relative error of +3.81%. This positive deviation indicates a stiffness-mass ratio mismatch in the initial model. In contrast, after TabPFN-CPO collaborative inversion, the equivalent density coefficient of the main beam in the corrected model was adjusted to 1.0819, successfully compensating for the missing secondary dead load mass; simultaneously, the elastic modulus coefficient of the main arch ring was adjusted to 0.9231, correcting the overall stiffness reduction caused by the semi-rigid effect of the nodes. Driven by this reconstruction of physical parameters, the fundamental frequency of the corrected model decreased to 0.6253Hz, with an absolute error of 0.0027Hz and a relative error convergence to +0.43%. This result indicates that, through calibration of physical parameters, the corrected model closely approximates the identified first-order vertical fundamental frequency of the actual bridge.
[0154] Table 6 Comparison of Frequency Errors in Dynamic Characteristics Before and After Correction
[0155]
[0156] However, frequency matching alone is insufficient to fully demonstrate the rationality of the dynamic response of the modified model. Therefore, a mode shape consistency index is further introduced to assist in verifying the model's ability to represent dynamic response. Figure 6 The results show the comparison between the measured first-order vertical bending mode shape and the corrected finite element model, and the MAC value between the two is calculated.
[0157] The calculation results show that the modal confidence score (MAC) of the corrected model is 0.77, indicating that the measured mode shapes and the corrected model mode shapes have a certain consistency in overall morphology, but there are still local differences. Based on the theoretical characteristics of the first-order antisymmetric vertical bending mode, this difference may be related to factors such as asymmetric traffic flow on the operating bridge, uneven mass distribution of the bridge deck, eccentric accelerometer installation, and drift of the mid-span stationary point. Therefore, the mode shape verification results can serve as supplementary evidence for the rationality of the dynamic response of the corrected model. Combining the frequency error and mode shape comparison results, it can be seen that the corrected model can reflect the low-order vertical dynamic characteristics of the structure well, but its comprehensive characterization ability still needs further verification by combining subsequent quasi-static deformation and local stress response.
[0158] S32, Quasi-static temperature deformation verification
[0159] To further verify the deformation compatibility of the modified finite element model at the full-bridge scale, this section extracts measured deflection data from 14 measuring points at key sections of the entire bridge, specifically from 1 / 8 to 7 / 8 spans on both the left and right sides, and constructs the measured deformation vector of the structure. This vector is then compared with the deformation vector calculated by the modified model under the same temperature difference condition, aiming to verify the modified model's ability to represent the quasi-static temperature deformation spatial morphology of the entire bridge.
[0160] like Figure 7 The figure shows a comparison of the spatial deformation patterns of the entire bridge. As can be seen from the figure, due to the different definitions of the initial zero point of the monitoring system and the origin of the finite element model coordinates, there are certain systematic differences between the measured displacement increment and the simulated absolute displacement on the numerical baseline. However, the two curves exhibit a relatively consistent spatial variation trend at the peaks, troughs, and inflection points. The calculation results show that the correlation coefficient of the entire bridge deformation pattern is 0.9760, indicating that the corrected model can better characterize the spatial deformation pattern of the actual bridge under temperature effects. It should be noted that the correlation coefficient mainly reflects the consistency of the deformation curve shape and cannot alone prove the uniqueness of the spatial distribution of structural stiffness, but it can serve as an effective verification index for the rationality of the corrected model's static response.
[0161] To further analyze the characterization effect of the modified model on the temperature deformation response of the left and right axes, the data from the left and right axes were validated separately, and the comparison results are as follows: Figure 8 (a) and Figure 8 As shown in (b), the analysis reveals that the correlation coefficient of the deformation vector of the left main girder is 0.9700, and the correlation coefficient of the deformation vector of the right main girder is 0.9880, indicating that the modified model has a good characterization ability for the temperature deformation patterns of both the left and right main girders. The high correlation coefficients of both the left and right girders suggest that the bridge's lateral temperature deformation response is relatively coordinated within the selected monitoring window. In summary, the modified model shows good consistency with the measured results at the quasi-static temperature deformation level, providing a basis for a modified analysis model for subsequent structural response analysis.
[0162] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for digital twin simulation and deduction of long-span arch bridges under sparse monitoring data, characterized in that, Includes the following steps: S1. Parameter Selection and Dimensionality Reduction: By introducing sparse polynomial chaotic expansion with fusion of minimum angular regression and Sobol global sensitivity analysis, global sensitivity analysis is performed on multiple initial physical parameters in the finite element model of a long-span arch bridge. The sensitivity differences of each parameter to the dynamic and static response of the structure are quantified, and key parameters to be corrected are selected according to the set sensitivity threshold, thereby reducing the number of parameter types. S2. Construction of small sample learning surrogate model: Using the context learning mechanism of the TabPFN network to fit the prior data of the table, combined with the finite parameter-response sample set generated under the limited conditions of the finite element samples in step S1, a small sample surrogate model for large and complex bridges oriented towards dynamic and static response characteristics is constructed. The small sample surrogate model establishes the nonlinear mapping relationship between the key parameters to be corrected and the dynamic and static response characteristics of the structure. S3. Fusion of dynamic and static features and global optimization: Construct a multi-objective function for fusion of dynamic and static features, perform residual matching between the structural response predicted by the small sample surrogate model in step S2 and the dynamic and static target features extracted from the actual measurements; rely on the CPO algorithm to perform global optimization of the multi-objective function, invert and identify the optimal key physical parameter correction coefficients, and complete the update of the finite element model.
2. The long-span arch bridge digital twin simulation deduction method of claim 1, wherein, Step S1 specifically involves: constructing an adaptive weighted sparse polynomial chaotic expansion model; by introducing a weighted orthogonal correction term and an adaptive weighted LARS algorithm, analytically calculating the Sobol main effect index and total effect index of each parameter; and determining parameters with total effect indices below a preset threshold as insensitive parameters and fixing them as constants, thereby achieving low-dimensional optimization of high-dimensional parameters.
3. The long-span arch bridge digital twin simulation deduction method of claim 2, wherein, The PCE series approximation form of the standard polynomial chaotic expansion structural response in step S1 is as follows: In the formula: These are multivariate orthogonal basis functions; The expansion coefficients are to be determined. A set of basis function indices; Sparse PCE, employing fixed L1 regularization, cannot accommodate the nonlinear differences between the smooth static response and the sensitive dynamic response of bridges. Therefore, by introducing a weighted orthogonal correction term, the orthogonality constraints of the basis functions are reconstructed. In the formula: The orthogonal correction weights are determined adaptively through iteration. The correlation coefficient is the basis function. Furthermore, an adaptive weighted LARS algorithm is proposed to construct a differentiated sparse optimization objective function: In the formula: The weighting coefficients are for static and dynamic responses: static response has a lower weight, and dynamic response has a higher weight. To adapt the regularization parameter, it gradually shrinks with each iteration to reduce the risk of overfitting or underfitting; For L1 sparse constraints; The improved AW-Sparse PCE final approximation is as follows: In the formula: The set of sparse basis functions after adaptive filtering; These are the optimal coefficients for the weighted least squares solution; Based on the constructed adaptive weighted sparse PCE, parameter space compression is achieved by directly calculating the Sobol sensitivity index through coefficient expansion; the total variance is calculated analytically as follows: The formula for calculating the main effect index is: The formula for calculating the total effect index is: like Significantly greater than If the sensitivity index is below the threshold, it indicates that the parameter is involved in a strong nonlinear interaction. Parameters with a sensitivity index below the threshold are considered insensitive parameters and set as constants, thus transforming the high-dimensional optimization problem into a low-dimensional parameter identification problem.
4. The digital twin simulation method for long-span arch bridges as described in claim 3, characterized in that, Step S2 specifically includes: S21. Data Acquisition and Sample Set Construction: Select m core sensitive parameters from the whole-bridge health monitoring of the long-span arch bridge in step S1, which were screened by global sensitivity analysis. Use OLHD technology to generate n sets of samples. Extract static response characteristics and dynamic response characteristics through finite element model calculation and analysis. Then construct a parameter-response sample set and divide it into a training set with a small number of samples and a test set for verification. S22, Improvements to the Core Mechanism of MFA-TabPFN: An improved MFA-TabPFN dual-channel surrogate model is constructed. First, a multi-feature alignment module is used to standardize the input multi-source heterogeneous data using the Z-score adaptive normalization method. A feature alignment attention mechanism is introduced to enhance the model's ability to fuse features of different physical dimensions. Simultaneously, an improved loss function is constructed, fusing MSE difference loss and KL divergence loss to form a composite loss function. ; S23. Model Initialization and Few-Shot Learning: Load the pre-trained TabPFN backbone network weights, and then configure the multi-feature alignment module and composite loss function designed in step S22 to complete the initialization of the MFA-TabPFN dual-channel surrogate model. Then, using the ICL mechanism, iteratively fine-tune the initialized MFA-TabPFN dual-channel surrogate model under the condition of a small number of training samples in step S21 to achieve high-precision inversion of key parameters of the finite element model, providing a reliable learning model foundation for subsequent model prediction and validation. S24. Model Prediction: Input the test parameter samples from the test set in step S21 into the MFA-TabPFN dual-channel surrogate model trained in step S23, perform a single forward propagation calculation, output the predicted values of the bridge's dynamic and static responses and compare them with the true values, then determine whether the prediction accuracy meets the engineering requirements; if it does, a high-precision multi-feature aligned MFA-TabPFN surrogate model is constructed; if it does not, a feedback mechanism is triggered, the sample size is replenished and the previous steps S22 and S23 are iterated repeatedly until the accuracy meets the engineering requirements, and finally a high-precision surrogate model that can be used for the correction of the finite element model of a long-span arch bridge is obtained.
5. The digital twin simulation method for long-span arch bridges as described in claim 4, characterized in that, m core sensitive parameters in step S21 As an input vector, each parameter follows a uniform distribution within its physically reasonable range; static response characteristics and dynamic response characteristics Synchronous fitting as output vector: .
6. The digital twin simulation method for long-span arch bridges as described in claim 5, characterized in that, In step S22, to address the differences in dimensions and sensitivity between static and dynamic features, an adaptive normalization encoding and feature alignment attention mechanism are introduced. This unifies the encoding of output features and performs Z-score adaptive normalization on static and dynamic response features respectively, thus eliminating the dimensional differences. In the formula: , These are the mean and standard deviation of the static characteristics, respectively. , These are the mean and standard deviation of the dynamic characteristics, respectively. By incorporating a feature alignment attention branch, the association weights between static and dynamic features and input parameters are calculated separately, adaptively adjusting the contribution of the two types of features in the mapping to address the problem of uneven fitting accuracy. In the formula: The adaptive weights for dynamic and static features are learned through MLP and dynamically adjusted according to the input parameters.
7. The digital twin simulation method for long-span arch bridges as described in claim 6, characterized in that, In step S22, to address the potential overfitting problem of the surrogate model in scenarios with limited samples, a composite loss function containing a mean squared error term and a prediction uncertainty constraint term is introduced to describe the constraint concept for the stability of the model output. In the formula: To compensate for the mean squared error loss between the predicted and actual values, and to ensure fitting accuracy; To constrain the prediction uncertainty of the model and avoid overfitting, the KL divergence loss between the predicted distribution and the Bayesian posterior distribution is calculated. The regularization weight coefficients are determined adaptively through cross-validation.
8. The digital twin simulation method for long-span arch bridges as described in claim 7, characterized in that, Step S3 involves constructing a multi-objective function that fuses static and dynamic features. Specifically, it selects the temperature sensitivity coefficient as a static feature and the first-order vertical fundamental frequency as a dynamic feature; and constructs a dimensionless objective function based on the sum of squared relative residuals. : In the formula: The vector of physical parameters to be corrected; , For the TabPFN proxy model in parameters The predicted value at that location; , The true value of the target feature identified based on measured data; , The weighting systems for static and dynamic characteristics respectively satisfy... The weighting coefficients are determined based on the physical properties, statistical robustness, and parameter identification ability of the two types of features.
9. The digital twin simulation method for long-span arch bridges as described in claim 8, characterized in that, Step S3 relies on the CPO algorithm for global optimization, specifically including: introducing an adaptive step size mechanism based on Sobol sensitivity prior, so that the exploration step size of each parameter in the CPO algorithm is proportional to its global sensitivity exponent; the expression is: In the formula: Let be the exploration step size of the i-th parameter in the t-th iteration; The initial reference step size for the i-th parameter is determined by the physical value range of the parameter. Let be the Sobol sensitivity index of the i-th parameter, which represents the total contribution of the parameter to the structural response; An adaptive early stopping mechanism is also set up. When the rate of change of the optimal objective function value of the population over several consecutive generations is lower than the convergence threshold, the iteration is automatically terminated, and the optimal parameter combination is output. The convergence criterion is set as follows: In the formula: Let be the optimal objective function value of the t-th generation population; To continuously monitor the number of iteration steps, the baseline value is taken as follows: =5; The benchmark value is taken as the convergence threshold. = Adjust flexibly according to accuracy requirements.
10. A simulation system based on the digital twin simulation method for long-span arch bridges as described in claim 9, characterized in that, include: The parameter selection and dimensionality reduction module is used to introduce sparse polynomial chaotic expansion with fusion of minimum angular regression and Sobol global sensitivity analysis to perform global sensitivity analysis on multiple initial physical parameters in the finite element model of a long-span arch bridge. It quantifies the sensitivity differences of each parameter to the dynamic and static response of the structure, and selects key parameters to be corrected according to the set sensitivity threshold, thereby reducing the types of parameters. The few-shot learning proxy model construction module is used to fit the context learning mechanism of the TabPFN network using tabular prior data, and combine it with the finite parameter-response sample set generated under the constraint of finite element samples to construct a few-shot proxy model oriented towards dynamic and static response characteristics, and establish the nonlinear mapping relationship between the key parameters to be corrected and the dynamic and static response characteristics of the structure. The dynamic and static feature fusion and global optimization module is used to construct a multi-objective function for dynamic and static feature fusion. It performs residual matching between the structural response predicted by the small sample surrogate model and the dynamic and static target features extracted from the actual measurements, and uses the CPO algorithm to perform global optimization of the multi-objective function. It then inverts and identifies the optimal key physical parameter correction coefficients to complete the update of the finite element model.
