Dynamic updating method and device for finite element model of building structure based on additional stiffness

Abstract

The application relates to a dynamic updating method and device for a finite element model of a building structure based on additional stiffness, and is applied to the technical field of civil engineering structure analysis, and comprises the following steps: equivalent stiffness modules are arranged in an initial finite element model, the additional stiffness of non-load-bearing components is equivalently restored, the problem that a traditional high-rise initial finite element model ignores the stiffness of a filler wall and the initial modal deviation is too large is solved, and the workload of subsequent model correction is reduced from the source; a mass of physical parameters are screened and grouped by using a sensitivity matrix and a parameter correlation coefficient, invalid low-sensitivity parameters are removed, the output dimension of a deep learning network is significantly reduced, and the problem of prediction accuracy reduction caused by multi-parameter coupling is relieved; a fusion feature matrix is constructed by fusing modal frequencies and modes, parameter prediction is sufficiently constrained by the two types of modal information, and compared with a prediction scheme using only frequencies, the accuracy of the physical parameter correction result is higher.

Description

Technical Field

[0001] This invention relates to the field of civil engineering structural analysis technology, specifically to a method and apparatus for dynamic updating of finite element models of building structures based on additional stiffness. Background Technology

[0002] Finite element model updating technology serves as a bridge connecting structural numerical simulation analysis with the actual mechanical response of physical engineering structures. Its core objective is to adjust the physical parameters in the finite element model (such as material elastic modulus, member section moment of inertia, boundary condition stiffness, etc.) so that the modal characteristics output by the model (including natural frequencies, mode shapes, etc.) closely match the actual modal data of the building structure obtained through on-site measurements. For high-rise buildings, a precisely calibrated finite element model is the foundation and prerequisite for damage identification, condition assessment, digital twin creation, and seismic performance verification.

[0003] Currently, most mainstream finite element model update methods employ an iterative optimization-based approach. This method constructs a residual objective function between the computed and measured modes and uses optimization strategies such as gradient descent, genetic algorithms, and particle swarm optimization to iteratively adjust the physical parameters in the model until the objective function converges. While this approach is theoretically mature, it reveals significant shortcomings and deficiencies when applied to complex structural systems like high-rise buildings.

[0004] First, a significant "stiffness gap" exists between the initial model and the actual structure. In conventional high-rise building design or analysis modeling, numerous non-load-bearing components such as infill walls, enclosure walls, and curtain walls are often ignored or added only as mass sources because they do not bear the main structural loads. However, extensive engineering practice and research show that these non-load-bearing components interact complexly with the main frame structure under actual stress (such as the bracing effect of infill walls), providing a significant additional stiffness to the overall structure. If this stiffness contribution is completely ignored in the initial finite element model, the overall stiffness of the model will be much lower than that of the actual building, and the initial modal frequencies (especially low-order frequencies) will be systematically lower. This significant initial deviation forces the subsequent parameter correction process to bear enormous adjustment pressure, easily causing the optimization algorithm to get stuck in local optima, or even resulting in distorted parameter correction results.

[0005] Second, the sheer number of parameters to be corrected, coupled with strong coupling effects, leads to high iterative computation costs and slow convergence. A typical high-rise building contains tens of thousands of beams, slabs, and columns, each with its own elastic modulus, cross-sectional dimensions, etc., which can be considered physical parameters to be corrected. This high-dimensional parameter space, combined with the strong coupling between parameters (i.e., adjusting the stiffness of component A and adjusting the stiffness of component B may have similar effects on frequencies of the same order), makes traditional iterative optimization algorithms resemble blindly searching through a maze. To obtain a reasonable set of parameter solutions, hundreds of complete finite element modal calculations are often required. For high-rise building models, the computational cost of a single modal analysis is already considerable, and the high computational cost of the entire update process severely restricts its widespread application in practical engineering.

[0006] Third, existing deep learning-based model update schemes lack targeted and systematic optimization. In recent years, some research has attempted to introduce deep learning into the field of model updates, trying to use the nonlinear mapping capabilities of neural networks to replace expensive online iterative computations. However, these existing schemes generally suffer from two key flaws: First, they also fail to address the initial problem of "missing stiffness in non-load-bearing components," resulting in the network learning mapping relationships based on models with systematic biases, thus limiting prediction accuracy. Second, they directly use massive amounts of highly coupled physical parameters as the network output, leading to an excessively large output layer dimension and an exponential increase in network training difficulty. This not only requires massive amounts of training samples but also results in poor generalization ability of the trained network, making it difficult to guarantee prediction performance for new buildings. Therefore, existing deep learning schemes are difficult to truly implement in practical engineering applications of high-rise buildings.

[0007] In summary, how to achieve rapid and accurate updates of finite element models of high-rise buildings while taking into account the accuracy of the initial model, parameter dimensionality reduction, and efficient nonlinear mapping is a technical challenge that urgently needs to be solved in this field. Summary of the Invention

[0008] In view of this, the purpose of this invention is to provide a method and apparatus for dynamic updating of finite element models of building structures based on additional stiffness, which aims to solve the problems in the prior art, such as the lack of stiffness effect of non-load-bearing components, large computational load of multi-parameter coupling iteration, excessively high parameter dimensionality of deep learning schemes, and insufficient accuracy of model correction.

[0009] According to a first aspect of the present invention, a method for dynamically updating a finite element model of a building structure based on additional stiffness is provided, the method comprising: Obtain the initial finite element model of the target high-rise building, and place equivalent stiffness modules at the positions corresponding to non-load-bearing components in the initial finite element model to complete the stiffness compensation of the initial finite element model. Modal analysis is performed on the initial finite element model after stiffness compensation, the natural frequency data and corresponding mode shape data of the initial finite element model are extracted, and the natural frequency data and mode shape data are fused to construct a frequency-mode shape fusion feature matrix. Calculate the sensitivity matrix of the frequency-mode fusion characteristic matrix to multiple candidate physical parameters in the initial finite element model, and solve the correlation coefficient between the sensitivity matrices of different candidate physical parameters; based on the sensitivity matrix and correlation coefficient, screen the multiple candidate physical parameters to obtain key physical parameters, and divide the screened key physical parameters into different parameter groups; An independent perturbation variable is set for each parameter group. Multiple physical parameter perturbation samples are generated by sampling the perturbation variable within a preset perturbation interval. Each physical parameter perturbation sample is substituted into the initial finite element model after stiffness compensation for modal calculation to obtain the fusion feature corresponding to each physical parameter perturbation sample. The fusion feature is used as the input label, and the physical parameter perturbation value corresponding to the fusion feature is used as the output label to construct a deep learning sample set. A deep learning network is constructed, with its input layer configured to receive the frequency-mode fusion feature matrix and its output layer configured to output the predicted physical parameter perturbation values ​​corresponding to each parameter group; the deep learning network is trained offline using the deep learning sample set until the network converges, and a trained deep learning prediction model is obtained. On-site modal testing was conducted on the target high-rise building, and the original vibration signals of each floor of the target high-rise building were collected. After preprocessing the original vibration signals, the measured natural frequencies and mode shapes of the structure were extracted, and the measured frequency-mode shape fusion characteristics were constructed based on the measured natural frequencies and mode shapes of the structure. The measured frequency-mode fusion features are input into the trained deep learning prediction model, which outputs a predicted perturbation correction. The original physical parameters in the initial finite element model are corrected based on the predicted perturbation correction to obtain the updated finite element model of the target high-rise building.

[0010] Preferably, The step of filtering the multiple candidate physical parameters based on the sensitivity matrix and correlation coefficient to obtain key physical parameters, and then dividing the filtered key physical parameters into different parameter groups, includes: Set sensitivity thresholds and correlation coefficient thresholds; Candidate physical parameters whose sensitivity matrix values ​​are lower than the sensitivity threshold are identified as low-sensitivity parameters and eliminated. For key physical parameters whose sensitivity matrix values ​​are higher than the sensitivity threshold, calculate the correlation coefficient between each pair of parameters, and group the key physical parameters whose correlation coefficients are higher than the correlation coefficient threshold into the same parameter group.

[0011] Preferably, The step of setting independent perturbation variables for each parameter group and generating multiple sets of physical parameter perturbation samples by sampling the perturbation variables within a preset perturbation interval includes: A spatial uniform sampling method is adopted to ensure that the generated perturbation samples are uniformly distributed within the preset perturbation interval.

[0012] Preferably, The modal analysis performed on the finite element model after stiffness compensation is performed to extract the natural frequency data and corresponding mode shape data of the initial finite element model, and the natural frequency data and mode shape data are fused to construct a frequency-mode shape fusion feature matrix, including: The frequency vector is formed by extracting the first N natural frequencies from the initial finite element model, and the corresponding N mode shape vectors are spliced ​​or weighted to construct the frequency-mode shape fusion feature matrix; where N is a positive integer greater than or equal to 3.

[0013] Preferably, The equivalent stiffness module includes linear spring units, nonlinear spring units, or combinations thereof; The equivalent stiffness value of the equivalent stiffness module is pre-calibrated based on the material properties, geometric dimensions, and connection method of the non-load-bearing component to the main structure.

[0014] Preferably, it further includes: The calculated modal features of the updated high-rise finite element model are extracted and compared with the measured frequency-mode fusion features. If the error between the two exceeds a preset threshold, the updated target high-rise building finite element model is used as the new initial finite element model. The new initial finite element model is iteratively updated until the error converges to within the preset threshold.

[0015] Preferably, The step of preprocessing the original vibration signal to extract the measured natural frequency and mode shape of the structure includes: The original vibration signal is bandpass filtered to remove environmental noise, and the natural frequency and mode shape of the measured structure are extracted using random subspace identification or frequency domain decomposition.

[0016] Preferably, The deep learning network employs a graph convolutional neural network, which mines the nonlinear correlations within the frequency-mode fusion features and the spatial topological relationships between nodes of the mode shape through its graph convolutional layers, thereby outputting predicted values ​​of physical parameter perturbations.

[0017] According to a second aspect of the present invention, a method for dynamically updating a finite element model of a building structure based on additional stiffness is provided, the apparatus being used to implement the method described in any one of the above embodiments, the apparatus comprising: Stiffness compensation module: used to obtain the initial finite element model of the target high-rise building, and to place equivalent stiffness modules at the positions corresponding to non-load-bearing components in the initial finite element model to complete the stiffness compensation of the initial finite element model. Feature fusion module: used to perform modal analysis on the initial finite element model after stiffness compensation, extract the natural frequency data and corresponding mode shape data of the initial finite element model, and fuse the natural frequency data and mode shape data to construct a frequency-mode fusion feature matrix; Parameter filtering module: used to calculate the sensitivity matrix of the frequency-mode fusion feature matrix to multiple candidate physical parameters in the initial finite element model, and solve the correlation coefficient between the sensitivity matrices of different candidate physical parameters; based on the sensitivity matrix and correlation coefficient, the multiple candidate physical parameters are filtered to obtain key physical parameters, and the filtered key physical parameters are divided into different parameter groups; The sample set construction module is used to set independent perturbation variables for each parameter group, and to generate multiple sets of physical parameter perturbation samples by sampling the perturbation variables within a preset perturbation interval; each set of physical parameter perturbation samples is substituted into the initial finite element model after stiffness compensation for modal calculation to obtain the fusion features corresponding to each set of physical parameter perturbation samples; the fusion features are used as input labels, and the physical parameter perturbation values ​​corresponding to the fusion features are used as output labels to construct a deep learning sample set; Model training module: used to construct a deep learning network. Its input layer is configured to receive the frequency-mode fusion feature matrix, and its output layer is configured to output the predicted values ​​of physical parameter perturbations corresponding to each parameter group. The deep learning network is trained offline using the deep learning sample set until the network converges, and a trained deep learning prediction model is obtained. The measured feature extraction module is used to conduct on-site modal testing on the target high-rise building, collect the original vibration signals of each floor of the target high-rise building, extract the natural frequency and mode shape of the measured structure after preprocessing the original vibration signals, and construct the measured frequency-mode shape fusion feature based on the natural frequency and mode shape of the measured structure. Model update module: used to input the measured frequency-mode fusion features into the trained deep learning prediction model, the deep learning prediction model outputs the predicted perturbation correction amount, and the original physical parameters in the initial finite element model are corrected according to the predicted perturbation correction amount to obtain the updated finite element model of the target high-rise building.

[0018] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This application addresses the problems of neglecting the stiffness of infill walls and excessive initial modal deviations in traditional high-rise initial finite element models by incorporating equivalent stiffness modules into the initial finite element model. This effectively restores the additional stiffness of non-load-bearing components, reducing the workload of subsequent model corrections from the outset. It utilizes sensitivity matrices and parameter correlation coefficients to filter and group massive amounts of physical parameters, eliminating invalid and low-sensitivity parameters, significantly reducing the output dimension of the deep learning network and mitigating the decline in prediction accuracy caused by multi-parameter coupling. Furthermore, it integrates modal frequencies and mode shapes to construct a fusion feature matrix, fully utilizing both types of modal information to constrain parameter predictions. Compared to prediction schemes using only frequencies or only mode shapes, the accuracy of physical parameter correction results is higher. It relies on offline simulation to generate deep learning sample sets in batches, requiring only a one-time completion of large-scale finite element modal calculations. After network training, the online model update process does not require repeated finite element iteration calculations; parameter corrections can be output simply through network forward propagation, significantly saving computation time for updating high-rise building models.

[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0021] Figure 1 This is a flowchart illustrating a method for dynamically updating a finite element model of a building structure based on additional stiffness, according to an exemplary embodiment. Figure 2 This is a schematic diagram of a system for dynamically updating a finite element model of a building structure based on additional stiffness, according to another exemplary embodiment. In the attached diagram: 1-Stiffness compensation module, 2-Feature fusion module, 3-Parameter selection module, 4-Sample set construction module, 5-Model training module, 6-Measured feature extraction module, 7-Model update module. Detailed Implementation

[0022] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0023] Example 1: Figure 1 This is a flowchart illustrating a method for dynamically updating a finite element model of a building structure based on additional stiffness, according to an exemplary embodiment. Figure 1 As shown, the method includes: S1. Obtain the initial finite element model of the target high-rise building, and place equivalent stiffness modules at the positions corresponding to non-load-bearing components in the initial finite element model to complete the stiffness compensation of the initial finite element model. S2, Perform modal analysis on the initial finite element model after stiffness compensation, extract the natural frequency data and corresponding mode shape data of the initial finite element model, and fuse the natural frequency data and mode shape data to construct a frequency-mode shape fusion feature matrix; S3, calculate the sensitivity matrix of the frequency-mode fusion feature matrix to multiple candidate physical parameters in the initial finite element model, and solve the correlation coefficient between the sensitivity matrices of different candidate physical parameters; based on the sensitivity matrix and correlation coefficient, screen the multiple candidate physical parameters to obtain key physical parameters, and divide the screened key physical parameters into different parameter groups; S4, set independent perturbation variables for each parameter group, and generate multiple sets of physical parameter perturbation samples by sampling the perturbation variables within a preset perturbation interval; substitute each set of physical parameter perturbation samples into the initial finite element model after stiffness compensation for modal calculation to obtain the fusion features corresponding to each set of physical parameter perturbation samples; use the fusion features as input labels and the physical parameter perturbation values ​​corresponding to the fusion features as output labels to construct a deep learning sample set; S5, construct a deep learning network, whose input layer is configured to receive the frequency-mode fusion feature matrix, and whose output layer is configured to output the predicted physical parameter perturbation values ​​corresponding to each parameter group; use the deep learning sample set to train the deep learning network offline until the network converges, and obtain the trained deep learning prediction model. S6, Conduct on-site modal testing on the target high-rise building, collect the original vibration signals of each floor of the target high-rise building, extract the measured natural frequency and mode shape of the structure after preprocessing the original vibration signals, and construct the measured frequency-mode shape fusion feature based on the measured natural frequency and mode shape of the structure. S7, The measured frequency-mode fusion features are input into the trained deep learning prediction model. The deep learning prediction model outputs the predicted perturbation correction amount. The original physical parameters in the initial finite element model are corrected according to the predicted perturbation correction amount to obtain the updated finite element model of the target high-rise building. It is understood that this embodiment specifically includes: First, in general finite element software (such as ABAQUS, ANSYS or SAP2000), an initial finite element model of the target high-rise building is established based on the architectural design drawings. This model accurately simulates all frame beams, frame columns, core tube shear walls and floor slabs. However, the model does not contain any infill wall elements, which is the main reason for the insufficient stiffness of the subsequent model.

[0024] Subsequently, a stiffness compensation operation is performed, traversing all planar regions enclosed by beam-column frames in the model (i.e., the expected locations of the infill walls). In each such region, an equivalent stiffness module is placed diagonally. This module consists of two intersecting oblique spring elements, each with an axial stiffness of... k eq Perform initial calibration using the following formula:

[0025] in, E w This is the equivalent elastic modulus of the infill wall masonry. t w For the thickness of the infill wall, L d The length of the frame diagonal. θ Let be the angle between the spring and the horizontal beam. Using this formula, the distributed in-plane stiffness of the wall is equivalent to a pair of concentrated rod element stiffnesses. To ensure simulation accuracy, the equivalent stiffness can be reduced according to the opening ratio of the infill wall. After the stiffness modules of all infill wall locations are set up, the initial stiffness of the model as a whole is significantly and reasonably compensated. It is worth emphasizing that the equivalent stiffness module can use linear spring elements, nonlinear spring elements, or a combination thereof; the equivalent stiffness value of the equivalent stiffness module is pre-calibrated based on the material properties, geometric dimensions, and connection method with the main structure of the non-load-bearing component. To more precisely simulate the mechanical behavior of different infill walls, the equivalent stiffness module can also be modified and combined in the following ways: Nonlinear spring elements: For infill walls with openings that may experience stiffness degradation or crushing under large deformations, the linear springs in the equivalent stiffness module can be replaced with nonlinear spring elements. For example, a bilinear kinematic reinforcement model or a spring with a damage factor can be used, and its force-displacement relationship can be defined as: F ( x )= k 0 x (1 d ( x )) In the formula, d ( x ) indicates that the displacement x The varying damage factor, with a value range of [0,1), is used to characterize stiffness degradation characteristics; F ( x ) represents the nonlinear spring element in displacement x The restoring force generated during time; k 0 represents the initial axial stiffness of the spring element; x This indicates the amount of deformation displacement of the spring element along the axial direction. This setup allows the calibrated model to not only reflect the stiffness in the elastic phase, but also to simulate the nonlinear response under extreme loads to some extent.

[0026] Multi-spring parallel / series system: To simulate the "combined arch" effect of an infill wall frame system, multiple sets of springs with different directions and stiffnesses can be arranged along the diagonal of a frame. For example, one set of compression springs simulates masonry under compression, and one set of tension springs simulates the interface bonding between the frame and the wall. These springs are connected in parallel to form an equivalent stiffness module. The stiffness coefficients of these springs are { k 1, k 2, ..., k m Together, they constitute the parameter vector of this module, which can also be used as candidate physical parameters in step S3 for screening.

[0027] Adaptive stiffness setting: In step S1, not all infill wall locations need to be equipped with modules of the same stiffness. An initial distribution function can be preset based on the structural height or floor location, such as:

[0028] In the formula, keq ( i ) indicates the first i The equivalent stiffness value of the equivalent stiffness module at the floor level; i This represents the floor number, with values ​​ranging from 1, 2, ... N ,i =1 indicates the lowest level; kbase Indicates the lowest level ( i =1) The reference stiffness value of the equivalent stiffness module is determined by the material and geometric parameters of the non-load-bearing component; This formula characterizes the linear decay distribution of equivalent stiffness along the building height, meaning that the bottom infill wall contributes the most to the overall structural stiffness, while the top wall contributes the least. Depending on the actual engineering situation, the linear decay can be replaced by exponential decay or piecewise function form. Its control parameters can also be used as physical parameters to be corrected in sensitivity screening and deep learning prediction, so that the deep learning model can automatically learn the optimal distribution law of structural stiffness along the height.

[0029] Next, modal analysis is performed on the finite element model after stiffness compensation, extracting the natural frequency data and corresponding mode shape data of the initial finite element model. The natural frequency data and mode shape data are then fused to construct a frequency-mode shape fusion feature matrix, including: The first N natural frequencies extracted from the initial finite element model are used to form a frequency vector, and the corresponding N mode shape vectors are concatenated or weighted to construct the frequency-mode shape fusion feature matrix; where N is a positive integer greater than or equal to 3; specifically including: Modal analysis was performed on the finite element model after stiffness compensation, and the Lanczos algorithm was used to extract the first 6 natural frequencies of the model. f 1 to f 6), these 6 frequencies typically correspond to the translational and torsional modes of the structure along its two horizontal principal axes. Simultaneously, 6 mode shape vectors corresponding to the first 6 frequencies are extracted. Each mode shape vector contains the displacement values ​​of all free nodes in the model. Considering that high-rise buildings primarily focus on floor response, only the horizontal displacement components of the core tube or frame column nodes on each floor are extracted here. U x , U y ) and torsional component ( R z This forms the mode shape vector for each floor; the six frequencies are combined into a 6×1 angular frequency vector. F Then, the normalized mode shape vectors of each order (assuming there are 30 floors, each floor outputs 3 degrees of freedom, then the dimension of a single mode shape vector is 90) are concatenated row by row to form a 6×90 normalized mode shape matrix. Φ Finally, the mode shape matrix Φ With the circular frequency vector F The squared values ​​of the elements are divided accordingly to form a 6×90 "frequency-mode shape fusion characteristic matrix". M The matrix M This serves as an input sample for subsequent deep learning networks; Based on the sensitivity matrix and correlation coefficient, the multiple candidate physical parameters are screened to obtain key physical parameters, which are then divided into different parameter groups, including: Set sensitivity thresholds and correlation coefficient thresholds; Candidate physical parameters whose sensitivity matrix values ​​are lower than the sensitivity threshold are identified as low-sensitivity parameters and eliminated. For key physical parameters whose sensitivity matrix values ​​are higher than the sensitivity threshold, calculate the correlation coefficient between each pair of parameters, and group key physical parameters with correlation coefficients higher than the correlation coefficient threshold into the same parameter group; specifically including: In this embodiment, it is assumed that the candidate physical parameter set includes the following categories: Correction factors for the elastic modulus of concrete in the core tube shear wall at different floors (3 groups in total: low zone, medium zone, and high zone). Stiffness reduction factors of frame beams at different floors (considering the interaction of floor slabs, a total of 3 groups); Global scaling factor of the equivalent stiffness module of the infill wall deployed in step S1 (1 group in total). Rotational spring stiffness in three directions under fixed boundary conditions at the bottom of the structure (3 groups in total); A total of 10 candidate physical parameters; calculate M For the sensitivity matrix of these 10 parameters, the finite difference method was used to apply a small perturbation of ±5% to each parameter, and the modes were recalculated to obtain the results. M Change Δ M Sensitivity S ≈ Δ M / 5%. After the calculation is completed, a 6×90 sensitivity vector is obtained for each parameter (which is flattened into one dimension); The sensitivity threshold is set to 0.05. The L2 norm of the sensitivity vector of each parameter is calculated. Parameters with a sensitivity threshold lower than 0.05 (e.g., the reduction factor of some frame beams far from the core tube) are judged as low-sensitivity redundant parameters and are directly removed. For the remaining highly sensitive parameters (e.g., core tube low-zone elastic modulus, infill wall global factor, boundary rotational stiffness), the Pearson correlation coefficients of the sensitivity vectors between each pair were calculated. With a correlation coefficient threshold of 0.7, the calculations showed that the correlation coefficient between "core tube low-zone elastic modulus" and "infill wall global factor" was as high as 0.85, indicating that their influence on modal characteristics was highly similar. Therefore, they were assigned to the same parameter group (parameter group A). ​​The correlation coefficients of "boundary rotational stiffness" with other parameters were all below 0.3, so they formed their own group (parameter group B). Through this process, the number of parameters to be corrected was reduced from 10 to 2 groups (parameter group A and parameter group B), and the output dimension of the deep learning network was also set to 2 accordingly. It is worth noting that the sensitivity threshold and correlation coefficient threshold directly determine the aggressiveness of parameter selection. Setting the threshold too high may lead to the accidental deletion of key parameters; setting it too low will result in insignificant dimensionality reduction. In this embodiment, it is recommended to initially set the sensitivity threshold to 0.05 and the correlation coefficient threshold to 0.7. Then, observe the grouping results. If the deep learning network is difficult to train after grouping, the threshold can be appropriately increased to further reduce dimensionality. If the prediction results deviate greatly from the true values, the threshold can be appropriately decreased to retain more parameter details. This is an optimization process based on engineering experience. Setting independent perturbation variables for each parameter group, and generating multiple sets of physical parameter perturbation samples by sampling the perturbation variables within a preset perturbation interval, includes: employing a spatially uniform sampling method to ensure that the generated perturbation samples are uniformly distributed within the preset perturbation interval; specifically including: As in the example above, set the perturbation variable for parameter group A. α Its physical meaning is the synchronous scaling factor of the parameters within this group (elastic modulus of the core tube in the lower region and stiffness of the infill wall). α The perturbation interval is set to [0.7, 1.3], and the perturbation variable is set for parameter group B. β Its physical meaning is the multiplier factor of the rotational stiffness of the bottom boundary, and the perturbation interval of β is set to [0.5, 5.0]; the Latin hypercube sampling method is used in... α and β Within the constructed two-dimensional rectangular perturbation space [0.7, 1.3] × [0.5, 5.0], 5000 uniformly distributed perturbation sample points are generated. α i , β i ), i =1,2,...,5000; It should be noted that the disturbance range should cover the possible variation range of parameters throughout the entire life cycle of the building. For material parameters such as the elastic modulus of concrete, ±30% of the recommended value in the code can be taken. For empirical parameters such as the equivalent stiffness module, the range can be appropriately widened to ±50% or even larger to fully accommodate modeling uncertainties. If the range is set too narrowly, the sample cannot cover the true solution; if it is too wide, it will dilute the density of the sample in the effective region. For each perturbation sample ( α i , β i Substitute these values ​​into the finite element model: multiply all the original core tube low-zone elastic modulus and infill wall stiffness by . α i Multiply the original bottom rotational stiffness by β iKeeping other parameters in the model constant, perform a complete modal analysis to calculate the fusion feature matrix corresponding to the sample. M i This results in a dataset containing 4000 sample pairs. M i ->( α i , β i The dataset is divided into two parts: 80% is randomly allocated as the training set and 20% as the validation set. A graph convolutional network model is constructed with 30 nodes (corresponding to 30 floors). The node features are fused feature vectors from each layer. Then, two graph convolutional layers are stacked, each followed by a ReLU activation function. The graph convolutional layers are used to capture the spatial dependencies between different floor vibration modes. Finally, two fully connected layers map the features to a 2D output, corresponding to... α and β The predicted value; The mean squared error (MSE) is used as the loss function, the Adam optimizer is used, the initial learning rate is set to 0.001, the batch size is set to 64, and the network is trained using the training set generated above. After each epoch, the loss is evaluated using the validation set. The training continues for about 200 epochs until the loss on the validation set no longer decreases, and the network model weights at this point are saved. At the site, high-sensitivity wireless accelerometers were deployed on the 1st, 5th, 10th, 15th, 20th, 25th, and 30th floors of the target high-rise building, as well as on the roof, to continuously collect 5 minutes of horizontal acceleration time history data in the X and Y directions. The raw data is preprocessed as follows: First, trend terms are removed from the data; second, a bandpass filter of 0.1Hz to 10Hz is designed to filter out high-frequency structural noise and low-frequency temperature drift; finally, the data is downsampled to improve computational efficiency. The preprocessed multi-channel data was imported into ARTEMIS Modal Pro or similar modal identification software. The covariance-driven stochastic subspace identification method (SSI-COV) was used to identify modal parameters. By plotting stability diagrams, the measured frequencies corresponding to the first 6 orders of the finite element model were identified. f 1' to f 6', and the corresponding measured vibration mode. Φ Following the exact same data format as in step S2, construct the measured frequency-mode shape fusion feature matrix. M It should be noted that the modal order NThe value of N is not necessarily better the larger it is. For high-rise buildings, the first 3 to 6 modes (usually translational and first torsional in two horizontal directions) already contain the most important macroscopic stiffness information of the structure. Although higher-order modes are richer in information, they are more affected by field test noise and modeling errors (such as local vibration of the floor slab) are more likely to be mixed in, which may introduce interference. It is generally recommended that N be 3 to 8, and fine-tuned according to the accuracy of the validation set. The constructed measured fusion feature matrix M The input is fed into the trained graph convolutional network, which performs forward computation and outputs the predicted perturbation value in a very short time (less than 0.1 seconds): α =0.92, β =1.75; Substitute the predicted values ​​into the initial finite element model: multiply all parameters in the original parameter group A (core tube low zone elastic modulus and infill wall stiffness) by 0.92; multiply the original bottom rotational stiffness by 1.75. After performing this update, a calibrated high-rise building finite element model is immediately obtained. Calculate the first 6 frequencies of the updated model f 1' ' to f 6' ', compared to the measured frequency f 1' to f By comparison, the frequency mean relative error decreased from 15.2% before the update (after only stiffness compensation) to 2.5% after the update. Since this error is still slightly higher than the set engineering accuracy threshold of 2.0%, an iterative calibration can be performed. The updated model is used as the new initial model, and steps S6 and S7 are repeated (in this example, since the model changes little, there is no need to retrain the network; the prediction is directly repeated and the parameters are fine-tuned). After the second iteration, the frequency mean relative error further decreased to 1.8%, meeting the accuracy requirements, and the iteration ends. The final finite element model has a modal response that is highly consistent with the field measured data and can be delivered for subsequent seismic performance assessment or digital twin modeling.

[0030] Deep extensions of deep learning network architectures: This embodiment also extends the possible architectures of deep learning networks to adapt to different types of architectural and modal data, specifically including: Multi-channel convolutional neural network (1D-CNN) architecture: When mode shape data is viewed as a one-dimensional sequence (by floor order), using a one-dimensional convolutional network is an efficient choice to fuse the frequency-mode shape feature matrix. MTreating signals as multiple channels (one channel per frequency order, or one channel per mode shape), the network can automatically extract multi-scale features from local to global by stacking multiple convolutional and pooling layers. For example, shallow convolutional kernels can capture the relative deformation relationships between adjacent floors, while deep convolutions can integrate the overall bending or shear deformation patterns of the entire building. This architecture has fast training speed, few parameters, and is suitable for high-rise buildings with regular structures and smooth mode shapes.

[0031] Multi-head attention mechanism (Transformer) architecture: For complex high-rise buildings with irregular structural plans or significant torsional effects, the relationships between different vibration modes and between different floors are more complex. In such cases, introducing the multi-head attention mechanism from the Transformer architecture offers significant advantages. Treating the frequency and mode shape vectors as a sequence, the network can dynamically learn which frequency components have the strongest correlation with which floor's mode shape components through self-attention modules. The attention score matrix can be visualized, providing engineers with "interpretability" of the model's decisions—that is, which modal features the network primarily relies on when predicting parameter perturbations, thereby enhancing the reliability of the prediction results.

[0032] Physical Information Neural Network (PINN) Constraints: To further improve the physical plausibility of predictions, physical constraint terms can be added to the loss function. For example, we can utilize the Rayleigh quotient formula for structures. Theoretically, for a predicted stiffness correction parameter, the corrected model's frequency should be closer to the measured value. An auxiliary loss term can be designed. L phys = |(Predicted frequency of the perturbated model) - (Measured frequency)|, although the predicted frequency of the perturbated model cannot be obtained directly analytically, it can be approximated by constructing a lightweight, differentiable reduced-order model (such as a condensed model based on the first few modes). This physical loss is then weighted and added to the data-driven MSE loss to form a hybrid loss function. L total = w 1 L MSE + w 2 L phys ,in, w 1 and w 2 represents the weight; this PINN framework can guide the solution space of the network to fall on the manifold constrained by physical laws, effectively avoiding the "physically unrealizable" prediction results that may occur in pure data-driven methods.

[0033] One of the greatest advantages of this invention lies in its efficient online prediction capability, making it an ideal tool for iterative updates of digital twin models in structural health monitoring (SHM) systems. Suppose a high-rise building has deployed a long-term health monitoring system with wind vibration accelerometers and GPS installed at the top, and strain gauges deployed at key locations on each floor. This system generates massive amounts of data in real time. Traditional finite element model update methods cannot achieve real-time or near-real-time updates due to computational time consumption. However, by applying this invention, the following workflow can be constructed: Offline Phase (Construction / Early Opening): An initial model containing equivalent stiffness modules is built based on the design drawings. Benchmark measured data is obtained by conducting one or two high-quality detailed modal tests (e.g., using artificial blasting or environmental excitation) during the design phase or early opening of the building. Steps S2 to S5 of this invention are executed to train a "benchmark deep learning prediction model," a process that takes several days but is only performed once.

[0034] Online Phase (Operational Period): Whenever the SHM system collects new vibration data (e.g., after a typhoon or during quarterly routine maintenance), the following process is automatically triggered: The data preprocessing module automatically extracts the measured modal features at the current moment from stable environmental vibration data that lasts for several hours using robust online identification algorithms (such as recursive random subspace identification method).

[0035] The measured features are fed into a pre-trained benchmark deep learning prediction model, which outputs the predicted perturbation values ​​of the current structural physical parameters within seconds (e.g., whether the core tube stiffness has decreased, or whether the stiffness of the infill wall in a certain area has decreased due to cracking). Automatically update the underlying finite element model parameters of the digital twin; The updated model will be used to assess whether the current structural stiffness has deteriorated abnormally and to provide risk warnings.

[0036] This process automates the entire chain from "real-time data acquisition -> second-level model update -> instant performance evaluation." Whenever a minor physical change occurs in the structure (such as material properties creeping over time or damage accumulation), it can be quickly captured and reflected in the digital twin model, greatly improving the timeliness and accuracy of the structural health monitoring system and achieving a leap from "passive monitoring" to "active perception and mirror synchronization."

[0037] Example 2: Figure 2 This is a schematic diagram of a system for dynamically updating a finite element model of a building structure based on additional stiffness, according to another exemplary embodiment. The device includes: Stiffness compensation module 1: used to obtain the initial finite element model of the target high-rise building, and to place equivalent stiffness modules at the positions corresponding to non-load-bearing components in the initial finite element model to complete the stiffness compensation of the initial finite element model. Feature fusion module 2: is used to perform modal analysis on the initial finite element model after stiffness compensation, extract the natural frequency data and corresponding mode shape data of the initial finite element model, and fuse the natural frequency data and mode shape data to construct a frequency-mode shape fusion feature matrix; Parameter filtering module 3: is used to calculate the sensitivity matrix of the frequency-mode fusion feature matrix to multiple candidate physical parameters in the initial finite element model, and solve the correlation coefficient between the sensitivity matrices of different candidate physical parameters; based on the sensitivity matrix and correlation coefficient, the multiple candidate physical parameters are filtered to obtain key physical parameters, and the filtered key physical parameters are divided into different parameter groups; Sample set construction module 4: used to set independent perturbation variables for each parameter group, and generate multiple sets of physical parameter perturbation samples by sampling the perturbation variables within a preset perturbation interval; substitute each set of physical parameter perturbation samples into the initial finite element model after stiffness compensation for modal calculation to obtain the fusion features corresponding to each set of physical parameter perturbation samples; use the fusion features as input labels and the physical parameter perturbation values ​​corresponding to the fusion features as output labels to construct a deep learning sample set; Model training module 5: used to construct a deep learning network, its input layer is configured to receive the frequency-mode fusion feature matrix, and its output layer is configured to output the predicted values ​​of physical parameter perturbations corresponding to each parameter group; the deep learning network is trained offline using the deep learning sample set until the network converges, and a trained deep learning prediction model is obtained. Measured feature extraction module 6: used to conduct on-site modal testing on the target high-rise building, collect the original vibration signals of each floor of the target high-rise building, extract the natural frequency and mode shape of the measured structure after preprocessing the original vibration signals, and construct the measured frequency-mode shape fusion feature based on the natural frequency and mode shape of the measured structure. Model update module 7: is used to input the measured frequency-mode fusion features into the trained deep learning prediction model, the deep learning prediction model outputs the predicted perturbation correction amount, and the original physical parameters in the initial finite element model are corrected according to the predicted perturbation correction amount to obtain the updated finite element model of the target high-rise building.

[0038] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0039] It should be noted that in the description of this invention, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this invention, unless otherwise stated, "a plurality of" means at least two.

[0040] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0041] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0042] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it includes one or a combination of the steps of the method embodiments.

[0043] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0044] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0045] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0046] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A dynamic update method for finite element models of building structures based on additional stiffness, characterized in that, The method includes: Obtain the initial finite element model of the target high-rise building, and place equivalent stiffness modules at the positions corresponding to non-load-bearing components in the initial finite element model to complete the stiffness compensation of the initial finite element model. Modal analysis is performed on the initial finite element model after stiffness compensation, the natural frequency data and corresponding mode shape data of the initial finite element model are extracted, and the natural frequency data and mode shape data are fused to construct a frequency-mode shape fusion feature matrix. Calculate the sensitivity matrix of the frequency-mode fusion characteristic matrix to multiple candidate physical parameters in the initial finite element model, and solve the correlation coefficient between the sensitivity matrices of different candidate physical parameters; based on the sensitivity matrix and correlation coefficient, screen the multiple candidate physical parameters to obtain key physical parameters, and divide the screened key physical parameters into different parameter groups; An independent perturbation variable is set for each parameter group. Multiple physical parameter perturbation samples are generated by sampling the perturbation variable within a preset perturbation interval. Each physical parameter perturbation sample is substituted into the initial finite element model after stiffness compensation for modal calculation to obtain the fusion feature corresponding to each physical parameter perturbation sample. The fusion feature is used as the input label, and the physical parameter perturbation value corresponding to the fusion feature is used as the output label to construct a deep learning sample set. A deep learning network is constructed, with its input layer configured to receive the frequency-mode fusion feature matrix and its output layer configured to output the predicted physical parameter perturbation values ​​corresponding to each parameter group; the deep learning network is trained offline using the deep learning sample set until the network converges, and a trained deep learning prediction model is obtained. On-site modal testing was conducted on the target high-rise building, and the original vibration signals of each floor of the target high-rise building were collected. After preprocessing the original vibration signals, the measured natural frequencies and mode shapes of the structure were extracted, and the measured frequency-mode shape fusion characteristics were constructed based on the measured natural frequencies and mode shapes of the structure. The measured frequency-mode fusion features are input into the trained deep learning prediction model, which outputs a predicted perturbation correction. The original physical parameters in the initial finite element model are corrected based on the predicted perturbation correction to obtain the updated finite element model of the target high-rise building.

2. The method according to claim 1, characterized in that, The step of filtering the multiple candidate physical parameters based on the sensitivity matrix and correlation coefficient to obtain key physical parameters, and then dividing the filtered key physical parameters into different parameter groups, includes: Set sensitivity thresholds and correlation coefficient thresholds; Candidate physical parameters whose sensitivity matrix values ​​are lower than the sensitivity threshold are identified as low-sensitivity parameters and eliminated. For key physical parameters whose sensitivity matrix values ​​are higher than the sensitivity threshold, calculate the correlation coefficient between each pair of parameters, and group the key physical parameters whose correlation coefficients are higher than the correlation coefficient threshold into the same parameter group.

3. The method according to claim 2, characterized in that, The step of setting independent perturbation variables for each parameter group and generating multiple sets of physical parameter perturbation samples by sampling the perturbation variables within a preset perturbation interval includes: A spatial uniform sampling method is adopted to ensure that the generated perturbation samples are uniformly distributed within the preset perturbation interval.

4. The method according to claim 3, characterized in that, The modal analysis performed on the finite element model after stiffness compensation is performed to extract the natural frequency data and corresponding mode shape data of the initial finite element model, and the natural frequency data and mode shape data are fused to construct a frequency-mode shape fusion feature matrix, including: The frequency vector is formed by extracting the first N natural frequencies from the initial finite element model, and the corresponding N mode shape vectors are spliced ​​or weighted to construct the frequency-mode shape fusion feature matrix; where N is a positive integer greater than or equal to 3.

5. The method according to claim 1, characterized in that, The equivalent stiffness module includes linear spring units, nonlinear spring units, or combinations thereof; The equivalent stiffness value of the equivalent stiffness module is pre-calibrated based on the material properties, geometric dimensions, and connection method of the non-load-bearing component to the main structure.

6. The method according to claim 1, characterized in that, Also includes: The calculated modal features of the updated high-rise finite element model are extracted and compared with the measured frequency-mode fusion features. If the error between the two exceeds a preset threshold, the updated target high-rise building finite element model is used as the new initial finite element model. The new initial finite element model is iteratively updated until the error converges to within the preset threshold.

7. The method according to claim 6, characterized in that, The step of preprocessing the original vibration signal to extract the measured natural frequency and mode shape of the structure includes: The original vibration signal is bandpass filtered to remove environmental noise, and the natural frequency and mode shape of the measured structure are extracted using random subspace identification or frequency domain decomposition.

8. The method according to claim 7, characterized in that, The deep learning network employs a graph convolutional neural network, which mines the nonlinear correlations within the frequency-mode fusion features and the spatial topological relationships between nodes of the mode shape through its graph convolutional layers, thereby outputting predicted values ​​of physical parameter perturbations.

9. A dynamic update method for finite element models of building structures based on additional stiffness, characterized in that, The apparatus is used to implement the method according to any one of claims 1-8, and the apparatus comprises: Stiffness compensation module: used to obtain the initial finite element model of the target high-rise building, and to place equivalent stiffness modules at the positions corresponding to non-load-bearing components in the initial finite element model to complete the stiffness compensation of the initial finite element model. Feature fusion module: used to perform modal analysis on the initial finite element model after stiffness compensation, extract the natural frequency data and corresponding mode shape data of the initial finite element model, and fuse the natural frequency data and mode shape data to construct a frequency-mode fusion feature matrix; Parameter filtering module: used to calculate the sensitivity matrix of the frequency-mode fusion feature matrix to multiple candidate physical parameters in the initial finite element model, and solve the correlation coefficient between the sensitivity matrices of different candidate physical parameters; based on the sensitivity matrix and correlation coefficient, the multiple candidate physical parameters are filtered to obtain key physical parameters, and the filtered key physical parameters are divided into different parameter groups; The sample set construction module is used to set independent perturbation variables for each parameter group, and to generate multiple sets of physical parameter perturbation samples by sampling the perturbation variables within a preset perturbation interval; each set of physical parameter perturbation samples is substituted into the initial finite element model after stiffness compensation for modal calculation to obtain the fusion features corresponding to each set of physical parameter perturbation samples; the fusion features are used as input labels, and the physical parameter perturbation values ​​corresponding to the fusion features are used as output labels to construct a deep learning sample set; Model training module: used to construct a deep learning network. Its input layer is configured to receive the frequency-mode fusion feature matrix, and its output layer is configured to output the predicted values ​​of physical parameter perturbations corresponding to each parameter group. The deep learning network is trained offline using the deep learning sample set until the network converges, and a trained deep learning prediction model is obtained. The measured feature extraction module is used to conduct on-site modal testing on the target high-rise building, collect the original vibration signals of each floor of the target high-rise building, extract the natural frequency and mode shape of the measured structure after preprocessing the original vibration signals, and construct the measured frequency-mode shape fusion feature based on the natural frequency and mode shape of the measured structure. Model update module: used to input the measured frequency-mode fusion features into the trained deep learning prediction model, the deep learning prediction model outputs the predicted perturbation correction amount, and the original physical parameters in the initial finite element model are corrected according to the predicted perturbation correction amount to obtain the updated finite element model of the target high-rise building.

Images