A structural damage identification method based on surrogate model and improved covariance matrix adaptation evolution strategy

By constructing a hybrid network model based on a surrogate model and an improved covariance matrix adaptive evolution strategy, and combining it with a variational Bayesian model correction method, the problem that existing structural damage identification methods are unable to accurately identify damage geometry and boundaries is solved, thus achieving efficient and accurate damage identification.

CN122470913APending Publication Date: 2026-07-28HUAZHONG UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610555477.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-04-24
Publication Date
2026-07-28

AI Technical Summary

Technical Problem

Existing structural damage identification methods struggle to efficiently and accurately identify the geometric dimensions and boundary information of damage, especially in complex structures where direct geometric interpretation and quantitative information are lacking, and are also subject to measurement noise interference and model uncertainty.

Method used

A structural damage identification method based on a surrogate model and an improved covariance matrix adaptive evolution strategy is adopted. By constructing a hybrid network model of convolutional neural network and long short-term memory network, and combining it with a variational Bayesian model correction method, the method can accurately identify the geometric parameters of the damage.

Benefits of technology

It achieves efficient and accurate identification of structural damage geometry and boundary information, reduces model prediction errors and uncertainties, and improves the efficiency and accuracy of the identification process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122470913A_ABST
    Figure CN122470913A_ABST
Patent Text Reader

Abstract

The application discloses a structural damage identification method based on a proxy model and an improved covariance matrix adaptive evolution strategy, and relates to the technical field of structural damage identification.The application comprises the following steps: obtaining a structure to be identified, constructing a finite element model of the structure to be identified, and respectively applying excitation to the structure to be identified and the finite element model to obtain a data set; a hybrid network model based on a convolutional neural network and a long short-term memory network is constructed, and the data set is used for training to obtain a proxy model; the proxy model is improved by using a covariance matrix adaptive evolution strategy based on a model gradient to obtain an optimal solution; a variational Bayesian model correction method is used to correct the optimal solution to obtain the most possible value of a structural damage geometric size parameter, so that damage geometric size and boundary information of the structure to be identified are determined.The application can efficiently and accurately identify structural damage geometric size information.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of structural damage identification technology, and in particular to a structural damage identification method based on a surrogate model and an improved covariance matrix adaptive evolution strategy. Background Technology

[0002] In recent years, people have increasingly emphasized the importance of structural health monitoring. Vibration-based structural health monitoring methods do not require prior knowledge of the approximate location of structural damage and offer advantages such as low detection costs and no disruption to the normal use of the structure, making them a research hotspot in the field of structural damage identification. Structural damage is mainly manifested in changes in structural physical parameters (such as stiffness, mass, and damping), and these physical parameters are directly related to their measured dynamic response. After damage occurs, changes in the structural physical parameters lead to certain alterations in its dynamic response. Vibration-based identification methods utilize this characteristic to determine the presence, location, and severity of structural damage by analyzing the structural dynamic response.

[0003] However, existing structural damage identification methods typically employ overall structural partitioning and material stiffness reduction to perform equivalent modeling and identification of damage. Discretizing the potential damage space of complex structures into a finite number of sub-regions reduces the dimensionality of the parameters to be identified, maintaining good numerical stability and computability even under conditions of limited measurement point information and high noise. Simultaneously, stiffness reduction, as a commonly used equivalent damage characterization method, can reflect the impact of local stiffness degradation caused by damage on the dynamic response, making it suitable for macroscopic localization of damage areas. This method's output usually corresponds to the degree of stiffness degradation in a certain region, but it does not uniquely correspond to a specific geometric defect. For example, in practical problems, response changes are often simultaneously influenced by factors such as defect location, area, thickness (height), contact state, and boundary conditions. For complex three-dimensional spatial structures, since damage may occur in different layers and regions within the structure, practical engineering focuses more on the geometric boundaries and scale information of the defects, such as the defect's center location, lateral and longitudinal extension range, and corresponding volume. These geometric quantities directly determine the spatial distribution of changes in the structural stress mechanism and are closely related to the operability of repair schemes. If we rely solely on the results of regional stiffness reduction, it is difficult to provide quantitative information such as the location, scale, and volume of damage that can be directly used in engineering, thus failing to meet the requirements for geometric interpretation of structural damage.

[0004] Furthermore, due to challenges such as measurement noise interference and model uncertainties, vibration-based damage identification methods often produce results with unavoidable uncertainty. In addition, vibration-based overall damage identification methods often need to be combined with numerical analysis, identifying parameters through model correction. However, structural finite element models are typically large and complex, and traditional model correction methods require continuous model iteration, which contradicts the current demand for efficient and rapid damage identification. Overall, existing structural damage identification methods cannot achieve efficient and accurate identification of structural damage geometric boundary information.

[0005] Therefore, providing a structural damage identification method based on a surrogate model and an improved covariance matrix adaptive evolution strategy to address the difficulties in existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a structural damage identification method based on a surrogate model and an improved covariance matrix adaptive evolution strategy, which can achieve efficient and accurate identification of structural damage geometric dimensions.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A structural damage identification method based on a surrogate model and an improved covariance matrix adaptive evolution strategy includes the following steps: Obtain the structure to be identified, construct a finite element model of the structure to be identified, and apply excitation to the structure to be identified and the finite element model respectively to obtain the dataset; A hybrid network model based on convolutional neural networks and long short-term memory networks is constructed, and a proxy model is obtained by training it using a dataset. The surrogate model is improved by using an adaptive evolution strategy based on the model gradient covariance matrix. This includes obtaining parameter information, generating gradient information and candidate solutions using a pre-trained deep learning surrogate model, evaluating candidate solutions using the surrogate model, updating parameter information until the termination condition is met, and obtaining the optimal solution. The optimal solution is corrected using a variational Bayesian model correction method to obtain the most likely values ​​of the structural damage geometry parameters, thereby determining the damage geometry and boundary information of the structure to be identified.

[0008] Optionally, the hybrid network model includes a multilayer perceptron module, a multi-scale convolutional neural network module, a sequence modeling module, and a gated fusion module; The multilayer perceptron module extracts features from the geometric dimensions of structural damage in the dataset and performs copying and expansion processing to generate a static feature vector sequence. The multi-scale convolutional neural network module contains two parallel convolutional neural network branches. One branch uses a larger convolutional kernel and introduces dilated convolution, while the other branch uses a smaller convolutional kernel to process the static feature vector sequence to obtain multi-scale temporal feature vectors. The sequence modeling module uses an encoder-decoder structure consisting of two layers of long short-term memory networks for sequence modeling. Both the encoder and decoder ends consist of two stacked long short-term memory networks, and the temporal sequence features output by the decoder end are processed using a multi-head self-attention mechanism to obtain residual components. In the gated fusion module, the residual component and the trend component are weighted and fused step-by-step according to the gated sequence to form the acceleration response prediction sequence of the final model output. The trend component is obtained by directly mapping the multi-scale temporal feature vector to the baseline response through a multilayer perceptron.

[0009] Optionally, training the hybrid network model includes updating the learning rate according to a cosine annealing learning rate strategy, expressed as: , in, The maximum learning rate at the start of the cycle; T is the cycle length, obtained by dividing the total number of training rounds by the number of training cycles; the e-th training round number within a certain cycle, t is e divided by T and taking the remainder.

[0010] Optionally, a multi-weighted joint loss function can be used to simultaneously constrain amplitude error and waveform phase consistency, expressed as: , , , , in, and These are the actual value and the predicted value at time t, respectively; T is the total length of the time series. and For the true response sequence and the model-predicted response sequence; , >0 indicates a numerically stable term; , and These are the weighting coefficients.

[0011] Optionally, the specific method to obtain the optimal solution is as follows: S1. Initialize the population distribution and path variables, including the initial population mean, initial global step size, covariance matrix, and evolutionary path of step size and covariance; S2. Before sampling to generate the g-th generation population, load the pre-trained deep learning agent model and calculate the gradient; S3, Generation A number of candidate solutions are identified, and gradient-guided sampling bias is introduced during the sampling process. S4. Use a surrogate model to evaluate the objective function values ​​of all candidate solutions and sort them by size, starting from... The top m individuals from the candidate population are selected for recombination to update the mean; S5. Update gradients and related parameters, including related adaptive parameters, step size path, covariance path, covariance matrix, and global step size; S6. Repeat steps S2-S5 until the termination condition is met, and output the optimal solution.

[0012] Optionally, the most likely values ​​to be obtained include: Using the acceleration time history data of the structure to be identified in the experiment as the test value, and the acceleration response of the finite element model predicted by the surrogate model as the predicted value, the objective function is constructed. By combining an improved covariance matrix adaptive evolution strategy, a variational Bayesian learning method is used to obtain the most probable values ​​of the geometric dimension parameters of the damage to the structure to be identified. Determine the damage geometry and boundary information of the structure to be identified.

[0013] Optional, objective function The expression is: , in, Let n be the geometric dimension parameter of the i-th structural damage, and n be the parameter dimension. This represents the total number of measurement points actually measured on the structure to be identified. This is the total response time step; and , respectively, are the measured acceleration response and the surrogate model predicted acceleration response of the j-th measurement point of the structure to be identified at time step s; η is a hyperparameter representing the accuracy of the measurement data, and is the reciprocal of the variance of the square of the acceleration response; To represent the hyperparameters of the model parameters, for The reciprocal of the variance.

[0014] As can be seen from the above technical solution, compared with the prior art, the present invention provides a structural damage identification method based on a surrogate model and an improved covariance matrix adaptive evolution strategy, which has the following beneficial effects: 1. This invention uses the setting of geometric dimension parameters for structural damage to replace the traditional approach of equivalent modeling and identification of damage by dividing the overall structure and reducing material stiffness, thereby realizing the identification of the geometric dimension boundary information of structural damage.

[0015] 2. This invention proposes a hybrid network model based on convolutional neural networks and long short-term memory neural networks to achieve accurate prediction of the dynamic response of complex structures. Since the model input consists of static parameters characterizing the geometric dimensions of structural damage, it is difficult to establish a time-step-wise mapping relationship between the input and the output temporal acceleration response. To address the problem that such time-invariant model inputs cannot support long-sequence output modeling, the surrogate model designed in this invention includes a multilayer perceptron module, a multi-scale convolutional neural network module, a sequence modeling module, and a gated fusion module. Simultaneously, a cosine annealing learning rate strategy and a weighted joint loss function are employed to enhance the model's temporal feature modeling capability, achieving high-precision and highly robust structural dynamic response surrogate model prediction, thereby better reducing the uncertainty in the damage identification process caused by model prediction errors.

[0016] 3. This invention improves the adaptive evolution strategy of the covariance matrix by introducing differentiable gradient information from a pre-trained surrogate model, guiding the algorithm's sampling distribution and update process, thereby reducing the algorithm's dependence on a large number of expensive function evaluations and improving the convergence speed.

[0017] 4. This invention establishes a variational Bayesian damage identification method that uses acceleration response as the target data, and uses a surrogate model to predict the structural dynamic response to replace the finite element model for evaluation and calculation. It also uses an improved covariance matrix adaptive evolution strategy to help minimize the objective function, thereby achieving efficient and accurate identification of the geometric boundary information of structural damage. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 This is a flowchart of a structural damage identification method based on a surrogate model and an improved covariance matrix adaptive evolution strategy disclosed in this invention. Figure 2 This is a schematic diagram of the proxy model structure disclosed in this invention; Figure 3 This is a flowchart of the variational Bayesian model correction process disclosed in this invention; Figure 4 This is a schematic diagram of the finite element model of the structure disclosed in an embodiment of the present invention; Figure 5a Structural damage geometric parameters disclosed in embodiments of the present invention The probability density function graph; Figure 5bStructural damage geometric parameters disclosed in embodiments of the present invention The probability density function graph; Figure 5c Structural damage geometric parameters disclosed in embodiments of the present invention The probability density function graph; Figure 5d Structural damage geometric parameters disclosed in embodiments of the present invention The probability density function graph; Figure 5e Structural damage geometric parameters disclosed in embodiments of the present invention The probability density function graph. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figure 1 As shown, this invention discloses a structural damage identification method based on a surrogate model and an improved covariance matrix adaptive evolution strategy, comprising the following steps: Obtain the structure to be identified, construct a finite element model of the structure to be identified, and apply excitation to the structure to be identified and the finite element model respectively to obtain the dataset; A hybrid network model based on convolutional neural networks and long short-term memory networks is constructed, and a proxy model is obtained by training it using a dataset. The surrogate model is improved by using an adaptive evolution strategy based on the model gradient covariance matrix. This includes obtaining parameter information, generating gradient information and candidate solutions using a pre-trained deep learning surrogate model, evaluating candidate solutions using the surrogate model, updating parameter information until the termination condition is met, and obtaining the optimal solution. The optimal solution is corrected using a variational Bayesian model correction method to obtain the most likely values ​​of the structural damage geometry parameters, thereby determining the damage geometry and boundary information of the structure to be identified.

[0022] Furthermore, obtaining the dataset includes: applying excitation to the structure to be identified and measuring the acceleration time history data of each data acquisition point; By setting structural damage with different geometric parameters for the finite element model and obtaining the corresponding acceleration response, a dataset of structural damage geometric parameters and acceleration response is constructed.

[0023] Specifically, structural damage can be described as defects located at different locations in the three-dimensional spatial structure to be identified, with different geometric dimensional information.

[0024] Furthermore, such as Figure 2 As shown, the hybrid network model includes a multilayer perceptron module, a multi-scale convolutional neural network module, a sequence modeling module, and a gated fusion module; The multilayer perceptron module extracts features from the geometric dimensions of structural damage in the dataset and performs copying and expansion processing to generate a static feature vector sequence. The multi-scale convolutional neural network module contains two parallel convolutional neural network branches. One branch uses a larger convolutional kernel and introduces dilated convolution, while the other branch uses a smaller convolutional kernel to process the static feature vector sequence to obtain multi-scale temporal feature vectors. The sequence modeling module uses an encoder-decoder structure consisting of two layers of long short-term memory networks for sequence modeling. Both the encoder and decoder ends consist of two stacked long short-term memory networks, and the temporal sequence features output by the decoder end are processed using a multi-head self-attention mechanism to obtain residual components. In the gated fusion module, the residual component and the trend component are weighted and fused step-by-step according to the gated sequence to form the acceleration response prediction sequence of the final model output. The trend component is obtained by directly mapping the multi-scale temporal feature vector to the baseline response through a multilayer perceptron.

[0025] Furthermore, training the hybrid network model includes updating the learning rate according to the cosine annealing learning rate strategy, expressed as: , in, The maximum learning rate at the start of the cycle; T is the cycle length, obtained by dividing the total number of training rounds by the number of training cycles; the e-th training round number within a certain cycle, t is e divided by T and taking the remainder.

[0026] Furthermore, a multi-weighted joint loss function is used to simultaneously constrain amplitude error and waveform phase consistency, expressed as: , , , , in, and These are the actual value and the predicted value at time t, respectively; T is the total length of the time series. and For the true response sequence and the model-predicted response sequence; , >0 indicates a numerically stable term; , and These are the weighting coefficients.

[0027] Furthermore, refer to Figure 3 As shown, the specific method for obtaining the optimal solution is as follows: S1. Initialize population distribution and path variables, including initial population mean. Initial global step size Covariance matrix And the evolutionary path of step size and covariance , ; S2. Before sampling to generate the g-th generation population, load the pre-trained deep learning proxy model. , Let be the geometrical parameters of the structural damage. Utilizing its differentiability, the mean value in the g-th generation... Calculate gradient ; S3, Generation A number of candidate solutions are identified, and gradient-guided sampling bias is introduced during the sampling process. ,Right now , , in, To guide the direction, For g, the covariance matrix Cholesky factor ; The gradient bias coefficient is applied to only a subset of candidate solutions in each generation. The proportion of guided samples is determined by an adaptive coefficient. control; S4. Use a surrogate model to evaluate the objective function values ​​of all candidate solutions and sort them by size, starting from... The top m individuals from the candidate population are selected for recombination to update the mean, expressed as: , In the formula, express The h-th optimal individual among the candidate solutions, sorted by objective function value; Let h be the recombination weight of the h-th sample; The mean learning coefficient; The value of m is determined based on the dimension n of the structural damage geometric parameter, i.e. and ; S5, Update Gradient With related parameters, including related adaptive parameters , and Step Path Covariance path and covariance matrix With global step size The corresponding expression is: , , , , , in, , , , and Both are the state variables and learning rate coefficients of the algorithm; The expectation of the Euclidean norm of the standard normal vector; S6. Repeat steps S2-S5 until the termination condition is met, and output the optimal solution.

[0028] Furthermore, obtaining the most likely value includes: Using the acceleration time history data of the structure to be identified in the experiment as the test value, and the acceleration response of the finite element model predicted by the surrogate model as the predicted value, the objective function is constructed. By combining an improved covariance matrix adaptive evolution strategy, variational Bayesian learning is used to obtain the most likely values ​​of the geometric dimension parameters of the damage to the structure to be identified. Determine the damage geometry and boundary information of the structure to be identified.

[0029] Furthermore, the objective function The expression is: , in, Let n be the geometric dimension parameter of the i-th structural damage, and n be the parameter dimension. This represents the total number of measurement points actually measured on the structure to be identified. This is the total response time step; and , respectively, are the measured acceleration response and the surrogate model predicted acceleration response of the j-th measurement point of the structure to be identified at time step s; η is a hyperparameter representing the accuracy of the measurement data, and is the reciprocal of the variance of the square of the acceleration response; To represent the hyperparameters of the model parameters, for The reciprocal of the variance.

[0030] Furthermore, combining an improved covariance matrix adaptive evolution strategy, the most probable values ​​of the geometric dimension parameters of the damage to the structure to be identified are obtained using a variational Bayesian learning method, specifically as follows: First, the objective function is minimized using an improved covariance matrix adaptive evolution strategy. Then take this time As the distribution mean, the sampling method is used to... Sampling is performed to update hyperparameters and Most likely values, hyperparameter calculation formulas, and known hyperparameters time The calculation formula, derived from variational Bayesian inference, is shown below: , , .

[0031] According to the iterative update The most probable value determines the damage geometry and boundary information of the structure to be identified, and describes the probability distribution of the damage geometry parameters of the structure based on the distribution of the sampled samples, thereby describing the uncertainty of the damage identification results and making the damage identification results more consistent with the actual situation.

[0032] In one specific embodiment, this method is used to identify the geometric dimensions of structural damage caused by voids in the intermediate CA mortar layer of a slab track structure model. Specifically: A finite element model of the slab track structure was established using the finite element software ABAQUS, such as... Figure 4 As shown, the model comprises three structural layers: a precast track slab, a CA mortar layer, and a base plate. The track slab measures 1.6 m × 1.4 m × 0.2 m. The CA mortar layer is designed with a T-shaped cross-section along its width, measuring 1.6 m × 1.4 m × (0.05-0.15) m. The base plate measures 1.6 m × 1.4 m × 0.25 m. All components are modeled using solid elements. The elastic foundation employs the Winkler foundation assumption and is modeled using an elastic foundation. Specific material parameters for each component include Young's modulus, density, and Poisson's ratio for the concrete track slab, Young's modulus, density, and Poisson's ratio for the concrete base plate, and Young's modulus, density, Poisson's ratio, and elastic foundation coefficient for the CA mortar. Boundary conditions are set between the structural layer components through bonded contact relationships.

[0033] A finite element model is constructed by setting up structural damage with different geometric parameters. Excitation is applied to the structural surface, and several measuring points are set up. Finite element analysis is then performed to obtain the corresponding acceleration response, thus constructing a dataset of structural damage geometric parameters and acceleration response. In this specific example, such as... Figure 4 As shown, eight measuring points were set up to extract the acceleration response, and a cylindrical void constructed from five parameters was used to simulate the geometric boundary of void damage within the CA mortar layer of a slab track. The structural damage geometric dimension parameter vector is then characterized as follows: ,in and The three parameters represent the spatial coordinates of the center of the top surface circle. Represents the radius of the cylinder. This represents the tensile height of the cavity along its thickness.

[0034] A hybrid network model based on convolutional neural network and long short-term memory network was constructed. The surrogate model that can predict the acceleration response of the structure to be identified was obtained by training with the dataset. The predictive performance of the surrogate model was evaluated using the coefficient of determination (R²), mean square error (MSE), mean absolute error (MAE), and root mean square error (RMSE) as evaluation indicators, as shown in Table 1.

[0035] Table 1 Calculation Results of Predictive Performance Evaluation Indicators

[0036] The results show that the surrogate model has high accuracy in predicting the dynamic response of slab track structures.

[0037] The surrogate model is improved using an adaptive evolution strategy based on the model gradient and covariance matrix. This includes acquiring parameter information and generating gradient information and candidate solutions using a pre-trained deep learning surrogate model. The candidate solutions are evaluated using a surrogate model. The top m populations are selected and their parameter information is updated until the termination condition is met to obtain the optimal solution. The objective function is constructed using the model response under the preset damage condition as the target value and the finite element model acceleration response predicted by the surrogate model as the predicted value. Combined with an improved covariance matrix adaptive evolution strategy, the most probable values ​​of the geometric dimensions of the damage to the structure to be identified are obtained using a variational Bayesian learning method. The objective function expression is: , Among them, settings The initial values ​​of the hyperparameters are defined as follows: as well as The superscripts of α and η indicate the number of iterations. The damage conditions and identification results are shown in Table 2.

[0038] Table 2 Damage Condition Settings and Identification Results

[0039] The identification results show that the geometric boundaries and scale information of the CA mortar layer void damage were accurately identified, with low coefficients of variation. Using the degree of overlap between the parameter settings and the overall spatial area corresponding to the identification results as an error index, the relative error was 5.36%. The probability density distribution function of the final structural damage geometric dimensions was calculated based on the sampled data, and the posterior probability distribution is as follows: Figures 5a-5e As shown.

[0040] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A structural damage identification method based on a surrogate model and an improved covariance matrix adaptive evolution strategy, characterized in that, Includes the following steps: Obtain the structure to be identified, construct a finite element model of the structure to be identified, and apply excitation to the structure to be identified and the finite element model respectively to obtain the dataset; A hybrid network model based on convolutional neural networks and long short-term memory networks is constructed, and a proxy model is obtained by training it using a dataset; The surrogate model is improved by using an adaptive evolution strategy based on the model gradient covariance matrix. This includes obtaining parameter information, generating gradient information and candidate solutions using a pre-trained deep learning surrogate model, evaluating candidate solutions using the surrogate model, updating parameter information until the termination condition is met, and obtaining the optimal solution. The optimal solution is corrected using a variational Bayesian model correction method to obtain the most likely values ​​of the structural damage geometry parameters, thereby determining the damage geometry and boundary information of the structure to be identified.

2. The structural damage identification method based on a surrogate model and an improved covariance matrix adaptive evolution strategy according to claim 1, characterized in that, The hybrid network model includes a multilayer perceptron module, a multi-scale convolutional neural network module, a sequence modeling module, and a gated fusion module; The multilayer perceptron module extracts features from the geometric dimensions of structural damage in the dataset and performs copying and expansion processing to generate a static feature vector sequence. The multi-scale convolutional neural network module contains two parallel convolutional neural network branches. One branch uses a larger convolutional kernel and introduces dilated convolution, while the other branch uses a smaller convolutional kernel to process the static feature vector sequence to obtain multi-scale temporal feature vectors. The sequence modeling module uses an encoder-decoder structure consisting of two layers of long short-term memory networks for sequence modeling. Both the encoder and decoder ends consist of two stacked long short-term memory networks, and the temporal sequence features output by the decoder end are processed using a multi-head self-attention mechanism to obtain residual components. In the gated fusion module, the residual component and the trend component are weighted and fused step-by-step according to the gated sequence to form the acceleration response prediction sequence of the final model output. The trend component is obtained by directly mapping the multi-scale temporal feature vector to the baseline response through a multilayer perceptron.

3. The structural damage identification method based on a surrogate model and an improved covariance matrix adaptive evolution strategy according to claim 1, characterized in that, Training the hybrid network model involves updating the learning rate according to the cosine annealing learning rate strategy, expressed as: , in, The maximum learning rate at the start of the cycle; T is the cycle length, obtained by dividing the total number of training rounds by the number of training cycles; the e-th training round number within a certain cycle, t is e divided by T and taking the remainder.

4. The structural damage identification method based on a surrogate model and an improved covariance matrix adaptive evolution strategy according to claim 1, characterized in that, A multi-weighted joint loss function is used to simultaneously constrain amplitude error and waveform phase consistency, and its expression is: , , , , in, and These are the actual value and the predicted value at time t, respectively; T is the total length of the time series. and For the true response sequence and the model-predicted response sequence; , >0 indicates a numerically stable term; , and These are the weighting coefficients.

5. The structural damage identification method based on a surrogate model and an improved covariance matrix adaptive evolution strategy according to claim 1, characterized in that, The specific method for obtaining the optimal solution is as follows: S1. Initialize the population distribution and path variables, including the initial population mean, initial global step size, covariance matrix, and evolutionary path of step size and covariance; S2. Before sampling to generate the g-th generation population, load the pre-trained deep learning agent model and calculate the gradient; S3, Generation A number of candidate solutions are identified, and gradient-guided sampling bias is introduced during the sampling process. S4. Use a surrogate model to evaluate the objective function values ​​of all candidate solutions and sort them by size, starting from... The top m individuals from the candidate population are selected for recombination to update the mean; S5. Update gradients and related parameters, including related adaptive parameters, step size path, covariance path, covariance matrix, and global step size; S6. Repeat steps S2-S5 until the termination condition is met, and output the optimal solution.

6. The structural damage identification method based on a surrogate model and an improved covariance matrix adaptive evolution strategy according to claim 1, characterized in that, The most likely values ​​include: Using the acceleration time history data of the structure to be identified in the experiment as the test value, and the acceleration response of the finite element model predicted by the surrogate model as the predicted value, the objective function is constructed. By combining an improved covariance matrix adaptive evolution strategy, a variational Bayesian learning method is used to obtain the most probable values ​​of the geometric dimension parameters of the damage to the structure to be identified. Determine the damage geometry and boundary information of the structure to be identified.

7. The structural damage identification method based on a surrogate model and an improved covariance matrix adaptive evolution strategy according to claim 6, characterized in that, objective function The expression is: , in, Let n be the geometric dimension parameter of the i-th structural damage, and n be the parameter dimension. This represents the total number of measurement points actually measured on the structure to be identified. This is the total response time step; and , respectively, are the measured acceleration response and the surrogate model predicted acceleration response of the j-th measurement point of the structure to be identified at time step s; η is a hyperparameter representing the accuracy of the measurement data, and is the reciprocal of the variance of the square of the acceleration response; To represent the hyperparameters of the model parameters, for The reciprocal of the variance.