A structure vulnerability evaluation method based on a multi-fidelity neural network surrogate model
By combining low-fidelity and high-fidelity data with a multi-fidelity neural network surrogate model and utilizing a CNN-LSTM neural network model, the problems of high cost and low accuracy in the seismic vulnerability assessment of building structures are solved, and an efficient and economical structural vulnerability assessment is achieved.
Patent Information
- Application Number
- CN202510642044.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-19
- Publication Date
- 2026-02-24
- Estimated Expiration
- 2045-05-19
AI Technical Summary
In the assessment of seismic vulnerability of building structures, existing technologies rely on high-precision physical experiments, which are too costly, and pure numerical simulations, which lack sufficient accuracy, making it difficult to improve the assessment accuracy in an economical and practical way.
A multi-fidelity neural network surrogate model is adopted, which combines low-fidelity and high-fidelity data of ground motion and structural parameters. The data is fused through a CNN-LSTM neural network model to establish a multi-fidelity surrogate model, thereby reducing the experimental cost and improving the evaluation accuracy.
While controlling costs, it significantly improves the accuracy of structural seismic response prediction, solves the problems of high test costs and insufficient numerical simulation accuracy in traditional methods, and provides an efficient solution for seismic assessment of large structures.
Smart Images
Figure CN120562182B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of structural analysis technology, and specifically to a structural vulnerability assessment method based on a multi-fidelity neural network surrogate model. Background Technology
[0002] "Resilient cities" have become a core goal of sustainable development, with the core being the reduction of uncertainty and vulnerability in the development process. As the foundation of resilience assessment, seismic vulnerability assessment, based on the theory of total probability, is a key component of seismic probabilistic safety assessment and is of great significance for evaluating structural seismic safety, formulating disaster prevention measures, and improving seismic resistance. Compared with traditional deterministic seismic analysis methods, seismic vulnerability analysis can simultaneously consider the impact of uncertainties in ground motion and structural parameters on the seismic response of the structure, intuitively reflecting the probability of structural response and damage through a vulnerability function. However, the derivation of the vulnerability function relies on the actual structural system analysis model. If the hysteretic behavior of structural components is not accurately modeled, the accuracy of vulnerability analysis will be affected. Although the model accuracy can be improved by adding information such as nodes, elements, and complex hysteretic behaviors, the resulting huge computational cost makes it uneconomical to apply in practical engineering. Therefore, in the seismic vulnerability assessment of building structures, how to improve the accuracy of vulnerability assessment and obtain high-precision surrogate models through a small number of representative mixed tests to address the challenge of testing costs is an urgent problem to be solved.
[0003] Vulnerability assessment of complex structures typically relies on numerical simulations to obtain structural damage parameters, but their accuracy is limited, especially in high-dimensional complex structures and real-world earthquake scenarios. To address this challenge, multi-fidelity neural network surrogate models have emerged. Multi-fidelity neural networks combine numerical simulations of varying precision with limited experimental data, achieving more accurate predictions of structural behavior by integrating low-fidelity and high-fidelity data. This method can significantly improve the modeling accuracy of complex structures using a data-driven approach without requiring a large number of additional physical experiments. Summary of the Invention
[0004] To address the issues of excessively high costs of high-precision physical experiments and insufficient accuracy of pure numerical simulations in traditional structural vulnerability assessment, this invention provides a structural vulnerability assessment method based on a multi-fidelity neural network surrogate model, comprising:
[0005] S1. Collect seismic record data with uniformly distributed ground motion intensity parameters;
[0006] S2. Establish a three-story, double-order self-resetting buckling-restrained braced steel frame structural model and collect structural uncertainty parameters; construct a finite element model and data processing algorithm based on the interaction between OpenSees and MATLAB;
[0007] S3. Based on the seismic record data described in S1 and the structural uncertainty parameters described in S2, generate initial low-fidelity data through random sampling; and obtain initial high-fidelity data based on the aforementioned data and using the Latin hypercube-orthogonal design sampling method.
[0008] S4. Input the initial low-fidelity data and the initial high-fidelity data from S3 into the steel frame model, obtain expanded low-fidelity data through incremental dynamic analysis, and obtain expanded high-fidelity data by combining mixed experiments to construct a neural network training sample database.
[0009] S5. Construct a CNN-LSTM neural network model based on the neural network training sample database described in S4, and optimize the network structure and hyperparameters;
[0010] S6. Calculate the correlation coefficient between the expanded low-fidelity data and the expanded high-fidelity data mentioned in S4, introduce uncertainty parameters to optimize the robustness of the model, and establish a multi-fidelity proxy model through hybrid training;
[0011] S7. Based on the hybrid training described in S6, establish a multi-fidelity proxy model to obtain seismic response data. Calculate seismic response parameters under different ground motion intensity parameters using the seismic response data. Construct a seismic demand model based on the regression analysis of the ground motion intensity parameters and the seismic response parameters.
[0012] S8. Calculate the structural damage probability density function and failure probability based on the earthquake demand model described in S7, and plot the vulnerability curve.
[0013] Furthermore, in S1, the earthquake record data adopts the 22 far-field ground motion sets recommended by ATC-63, including 14 earthquake records with magnitudes of 6.5-7.6 and epicentral distances of 8.7-98.22 km, with site shear wave velocities of 192-724 m / s, conforming to the US standard C / D site category, with a PGA range of 0.2-2.1g and evenly divided into four intervals.
[0014] Furthermore, in S2, the three-layer double-stage self-resetting buckling-restrained braced steel frame structure model includes: a reset system, an energy dissipation system, and external connectors;
[0015] The reset system includes: an inner tube, an outer tube, a combined disc spring, inner and outer tube baffles, and a disc spring;
[0016] The energy-consuming system includes: a two-stage energy-consuming inner core and a constraining channel steel;
[0017] The external connectors include: inner core connectors, inner tube connectors, outer tube connectors, and structural connectors.
[0018] Furthermore, in S7, the earthquake demand model is obtained through the following steps:
[0019] S7.1 Set the ground motion intensity parameters IM, including: peak ground acceleration (PGA), peak ground velocity (PGV), overall duration intensity (Ia), and average spectral acceleration (AvgS). a Earthquake intensity information; seismic response parameters DM include: top displacement and maximum inter-story drift angle θ. max ;
[0020] S7.2. Based on the fact that the seismic ground motion intensity parameter IM and the seismic response parameter DM follow a normal distribution, their mathematical relationship is expressed as follows:
[0021] DM = α(IM) β
[0022] Where α and β are regression coefficients;
[0023] S7.3. According to step S6.1, select the ground motion intensity parameter IM as the peak ground acceleration (PGA) and the seismic response parameter DM as the maximum inter-story drift angle θ. max The formula is then obtained as follows:
[0024] θ max =α(PGA) β
[0025] Taking the logarithm of both sides yields the following equation:
[0026] lnθ max = lnα + βln(PGA)
[0027] By performing linear regression on the above formula, the values of α and β are obtained, thus yielding the earthquake demand model.
[0028] Furthermore, S8 specifically includes:
[0029] S8.1 Define the failure probability of a structure exceeding its structural demand capacity parameter C under different seismic intensities, wherein the failure probability is determined by:
[0030] P f =P(θ) c / θ max <1)=P(lnθ) c -lnθ max <0)
[0031] We obtain, where Pf is the failure probability, and θ c For structural requirement capacity parameters;
[0032] S8.2, Set Z = lnθ c -lnθ max And it follows a normal distribution, with the mean expressed as: Standard deviation is The formula for deriving the failure probability is:
[0033]
[0034] Where Z is the representative value;
[0035] S8.3, N(μ) z ,σ z Converting the distribution to a standard normal distribution N(0,1), the failure probability is obtained as follows:
[0036]
[0037] S8.4 According to HAZUS 99, when IM is PGA, Use a value of 0.5 to plot the fragility curve.
[0038] The beneficial effects of this invention are:
[0039] This invention reduces the sample size of mixed experiments through a Latin hypercube-orthogonal design sampling strategy, acquiring key high-fidelity data while controlling costs. Simultaneously, it utilizes a CNN-LSTM neural network to construct a low-fidelity numerical simulation model, which is then corrected using high-fidelity experimental data, ultimately establishing a multi-fidelity surrogate model. This method overcomes the dual bottlenecks of excessively high costs in traditional mixed experiments and insufficient accuracy in single numerical simulations. By organically integrating the three core components of low-fidelity simulation, high-fidelity experimentation, and model correction through a data fusion algorithm, it significantly improves the accuracy of structural seismic response prediction. Compared to traditional methods, this approach effectively reduces experimental costs while ensuring assessment accuracy, providing an efficient solution for seismic assessment of large structures. Attached Figure Description
[0040] Figure 1 This is a flowchart of the structural vulnerability assessment process based on a multi-fidelity neural network surrogate model.
[0041] Figure 2 The flowchart shows the experimental method based on the CNN-LSTM neural network surrogate model.
[0042] Figure 3 This is a schematic diagram of a vulnerability assessment method based on a multi-fidelity neural network surrogate model. Detailed Implementation
[0043] The technical solution of the present invention will be further described below with reference to embodiments, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention. In the following embodiments, the process equipment or apparatus not specifically specified are all conventional equipment or apparatus in the art. Unless otherwise specified, the raw materials used in the embodiments of the present invention are all commercially available; unless otherwise specified, the technical means used in the embodiments of the present invention are all conventional means well known to those skilled in the art.
[0044] Example 1, combined with Figure 1 and Figure 2 This embodiment describes a structural vulnerability assessment method based on a multi-fidelity neural network surrogate model, comprising:
[0045] S1. Collect seismic record data with uniformly distributed ground motion intensity parameters;
[0046] S2. Establish a three-story, double-order self-resetting buckling-restrained braced steel frame structural model and collect structural uncertainty parameters; construct a finite element model and data processing algorithm based on the interaction between OpenSees and MATLAB;
[0047] S3. Based on the seismic record data described in S1 and the structural uncertainty parameters described in S2, generate initial low-fidelity data through random sampling; and obtain initial high-fidelity data based on the aforementioned data and using the Latin hypercube-orthogonal design sampling method.
[0048] S4. Input the initial low-fidelity data and the initial high-fidelity data from S3 into the steel frame model, obtain expanded low-fidelity data through incremental dynamic analysis, and obtain expanded high-fidelity data by combining mixed experiments to construct a neural network training sample database.
[0049] S5. Construct a CNN-LSTM neural network model based on the neural network training sample database described in S4, and optimize the network structure and hyperparameters;
[0050] S6. Calculate the correlation coefficient between the expanded low-fidelity data and the expanded high-fidelity data mentioned in S4, introduce uncertainty parameters to optimize the robustness of the model, and establish a multi-fidelity proxy model through hybrid training;
[0051] S7. Based on the hybrid training described in S6, establish a multi-fidelity proxy model to obtain seismic response data. Calculate seismic response parameters under different ground motion intensity parameters using the seismic response data. Construct a seismic demand model based on the regression analysis of the ground motion intensity parameters and the seismic response parameters.
[0052] S8. Calculate the structural damage probability density function and failure probability based on the earthquake demand model described in S7, and plot the vulnerability curve.
[0053] Specifically, this invention first generates random samples of seismic motion and structural parameters. Through structural numerical simulation, low-fidelity data of the structural response is obtained, and a CNN-LSTM neural network low-fidelity surrogate model is established. Then, the Latin hypercube-orthogonal design method is used to sample the random samples of seismic motion and structural parameters, and high-fidelity data of the structural response is obtained through superposition experiments. Next, the high-fidelity data is used to further train the CNN-LSTM neural network low-fidelity surrogate model, thereby establishing a CNN-LSTM neural network multi-fidelity surrogate model. Finally, the neural network multi-fidelity surrogate model is used to conduct seismic vulnerability assessment on a self-setting buckling-resisting braced structure, obtaining structural vulnerability curves, evaluating the structure's seismic performance, and analyzing the influence of the dual uncertainties of seismic motion and structural parameters on the structure's seismic performance.
[0054] pass Figure 1 The construction logic of the multi-fidelity surrogate model can be intuitively grasped. Based on the earthquake motion and structural uncertainty parameters, the CNN-LSTM model is trained and iteratively corrected through random sampling (numerical simulation to generate low-fidelity data) and Latin hypercube orthogonal design (mixed experiment to generate high-fidelity data) respectively, until the accuracy reaches the standard and the multi-fidelity surrogate model is output, which ultimately supports the assessment of structural vulnerability.
[0055] Figure 2 This paper demonstrates the experimental methodology based on a CNN-LSTM neural network surrogate model. The diagram illustrates the overall technical process for predicting structural seismic response, from identifying key structural parameters and seismic motion parameters, using Latin hypercube-orthogonal design to generate input samples through hybrid sampling, to constructing a neural network model that integrates CNN (convolutional layer feature extraction) and LSTM (temporal modeling), and finally outputting the structural response prediction results through iterative training (conditional on the convergence of the loss function), providing data support for subsequent vulnerability analysis.
[0056] In S3, the steps of the Latin hypercube-orthogonal design sampling method are as follows:
[0057] S3.1 Input Parameter Definition
[0058] Given k input variables, each variable has a range of values. Therefore, n sampling points need to be generated;
[0059] S3.2 Initialize Latin hypercube samples
[0060] The initial samples are generated using the standard Latin hypercube method. The interval of each variable is divided into n sub-intervals. A point is randomly selected in each interval. Finally, the k-dimensional data are combined to form the initial sample matrix S.
[0061] S3.3 Orthogonal Optimization Design
[0062] The initial sample is optimized to ensure that the distribution of sample points satisfies orthogonality constraints. The correlation matrix C between sample variables is calculated, and the sample order is adjusted using a sorting and exchange method to reduce the correlation between variables. Optimization algorithms (such as simulated annealing and genetic algorithms) are then applied to further optimize the sample distribution.
[0063] Define the correlation matrix C between sample points, where C ij This indicates the correlation between the i-th and j-th variables;
[0064] Optimization goal: Minimize That is, to reduce the correlation between variables;
[0065] S3.4, Multidimensional Spatial Uniformity Optimization
[0066] The minimum distance φ between sample points is calculated, and the value of φ is improved through iterative optimization to enhance the uniformity of the samples in multidimensional space. A φ-criterion is introduced to define the minimum distance φ between sample points and optimize its value.
[0067]
[0068] S3.5 Validate the sample distribution
[0069] Generate the optimized sample matrix S * To verify its uniformity and orthogonality, calculate whether the distribution of sample points in each dimension satisfies the uniformity requirement, and whether the sample correlation matrix C satisfies the orthogonality requirement.
[0070] S5. Using the sample database from step S4, construct a CNN-LSTM neural network model. Determine the neural network's inputs (such as structural parameters, dynamic loads, and seismic wave characteristics) and outputs (such as the maximum inter-story drift angle), and optimize the neural network structure, including hyperparameters such as the number of network layers, number of nodes, and activation functions.
[0071] In S4, incremental dynamic analysis is applied to the steel frame model to obtain sample data of structural deformation response in batches; important characteristic variables in the input, including seismic motion and structural parameters, are determined, and the maximum inter-story drift angle of the first and second floors is used as the key seismic response index to establish a sample database of neural network model parameters.
[0072] In S5, the basic structural components of the CNN-LSTM neural network model are as follows:
[0073] S5.1 Input Layer
[0074] It receives time-series input data, typically in the form of a three-dimensional tensor (number of samples × time step × feature dimension), such as earthquake time history records or structural response time history data.
[0075] S5.2 Convolutional Layers (CNN)
[0076] One-dimensional convolutional layer (Conv1D): used to extract local temporal features, usually using multiple convolutional kernels for sliding computation.
[0077] The first convolutional layer consists of 32 convolutional kernels of size 2 (stride 1) and the ReLU activation function.
[0078] The second convolutional layer consists of 64 kernels of size 2 (stride 1) and the ReLU activation function.
[0079] Dropout layer: A random deactivation is introduced after the convolutional layer with a dropout rate of 0.3 to prevent overfitting;
[0080] S5.3, Reshape Layer
[0081] The features extracted by the CNN are resized to fit the input format of the LSTM layer (time step × feature dimension).
[0082] S5.4, LSTM layer (Long Short-TermMemory Layers)
[0083] Stacked LSTM layers: A 3-layer LSTM structure with 128 hidden units per layer is used to enhance the ability to model time series.
[0084] Dropout layer: Add dropout between LSTM layers or after the output. The dropout rate is 0.3 to improve generalization performance.
[0085] S5.5, Dense Layer
[0086] It is used to map the temporal features of the LSTM output to the target prediction value. It contains 64 neurons, uses the ReLU activation function, and includes a Dropout mechanism with a Dropout rate of 0.3.
[0087] S5.6, Output Layer
[0088] Predictive regression tasks: such as predicting inter-layer displacement angles, set ReLU activation.
[0089] S5.7, Training Optimization Component
[0090] Loss functions: Mean squared error (MSE) is commonly used for regression tasks, while cross-entropy loss is commonly used for classification tasks.
[0091] Optimizer: Typically, the Adam optimizer is used, combined with a learning rate scheduler to dynamically adjust the learning rate.
[0092] Training strategy:
[0093] Batch Training: Set the batch size to 16.
[0094] Early Stopping: Set the patience value to 200 rounds to prevent overfitting.
[0095] This structure extracts local features through CNN, captures long-term temporal dependencies through LSTM, and combines Dropout and optimization strategies to effectively improve the model's prediction accuracy and generalization ability.
[0096] S5.8 Neural Network Preprocessing and Evaluation
[0097] Data normalization: Normalize the input data (such as structural parameters and seismic wave characteristics) to improve training results. Training and test set split: The dataset is usually split into 85% training set and 15% test set.
[0098] Evaluation metrics: The mean squared error (MSE) or coefficient of determination (R²) is used to evaluate model performance.
[0099] By appropriately selecting input / output design, network architecture, and hyperparameters, efficient CNN-LSTM neural network models for predicting structural dynamic responses can be constructed.
[0100] In S1, the earthquake record data adopts the 22 far-field ground motion sets recommended by ATC-63, including 14 earthquake records with magnitudes of 6.5-7.6 and epicentral distances of 8.7-98.22 km, with site shear wave velocities of 192-724 m / s, conforming to the US standard C / D category site, with a PGA range of 0.2-2.1g and uniformly divided into four intervals.
[0101] In S2, the three-layer double-stage self-resetting buckling-restrained braced steel frame structure model includes: a reset system, an energy dissipation system, and external connectors;
[0102] The reset system includes: an inner tube, an outer tube, a combined disc spring, inner and outer tube baffles, and a disc spring;
[0103] The energy-consuming system includes: a two-stage energy-consuming inner core and a constraining channel steel;
[0104] The external connectors include: inner core connectors, inner tube connectors, outer tube connectors, and structural connectors.
[0105] In S7, the earthquake demand model is obtained through the following steps:
[0106] S7.1 Set the ground motion intensity parameters IM, including: peak ground acceleration (PGA), peak ground velocity (PGV), overall duration intensity (Ia), and average spectral acceleration (AvgS). a Earthquake intensity information; seismic response parameters DM include: top displacement and maximum inter-story drift angle θ. max ;
[0107] S7.2. Based on the fact that the seismic ground motion intensity parameter IM and the seismic response parameter DM follow a normal distribution, their mathematical relationship is expressed as follows:
[0108] DM = α(IM) β
[0109] Where α and β are regression coefficients;
[0110] S7.3. According to step S6.1, select the ground motion intensity parameter IM as the peak ground acceleration (PGA) and the seismic response parameter DM as the maximum inter-story drift angle θ. max The formula is then obtained as follows:
[0111] θ max =α(PGA) β
[0112] Taking the logarithm of both sides yields the following equation:
[0113] lnθ max = lnα + βln(PGA)
[0114] By performing linear regression on the above formula, the values of α and β are obtained, thus yielding the earthquake demand model.
[0115] S8 specifically includes:
[0116] S8.1 Define the failure probability of a structure exceeding its structural demand capacity parameter C under different seismic intensities, wherein the failure probability is determined by:
[0117] P f =P(θ) c / θ max <1)=P(lnθ) c -lnθ max <0)
[0118] We obtain, where Pf is the failure probability, and θ c For structural requirement capacity parameters;
[0119] S8.2, Set Z = lnθ c -lnθ max And it follows a normal distribution, with the mean expressed as: Standard deviation is The formula for deriving the failure probability is:
[0120]
[0121] Where Z is the representative value;
[0122] S8.3, N(μ) z ,σ z Converting the distribution to a standard normal distribution N(0,1), the failure probability is obtained as follows:
[0123]
[0124] S8.4 According to HAZUS 99, when IM is PGA, Use a value of 0.5 to plot the fragility curve.
[0125] Example 2. Taking a three-story, three-span steel frame structure with double-stage self-resetting buckling-restrained bracing as an example, the basic principles and usage steps of the present invention are illustrated.
[0126] To prepare building structures for earthquake disasters, it is essential to reasonably estimate, predict, and mitigate losses. Seismic vulnerability curves are a crucial component in assessing earthquake damage and losses. Double-stage self-resetting buckling-restrained braces (BRS), as energy-dissipating components, offer advantages such as good energy dissipation and control of residual structural displacement. However, for a three-story, three-span steel frame structure with BRS, the accuracy of the overall structural seismic response obtained solely through nonlinear time history analysis is insufficient. Therefore, this implementation method employs a multi-fidelity neural network surrogate model to efficiently obtain the structural seismic response, proposing a structural vulnerability assessment method based on this model.
[0127] according to Figure 3 As can be seen, firstly, seismic response data was obtained based on the incremental dynamic analysis method, and a seismic demand model was derived. Finally, seismic vulnerability curves were plotted for a three-story, three-span steel frame structure with double-stage self-resetting buckling-restrained braces under minor, moderate, severe, and collapse conditions. The structure was essentially undamaged when the PGA reached a frequent earthquake of intensity 8. When the PGA reached the design intensity of intensity 8, there was a very small probability of minor damage. When the PGA reached a rare earthquake of intensity 8, there was an 82.36% probability of minor damage. This differs from the situation of a frame structure with double-stage self-resetting buckling-restrained braces under different conditions.
[0128] The working principle is consistent with that under strong earthquake action. Under minor earthquake action, the structure is basically in an elastic working state. As the earthquake intensity reaches the level of a rare earthquake, the structure begins to enter an elastoplastic state to dissipate seismic energy, and damage gradually accumulates. The collapse probability of the structure under an 8-degree rare earthquake action is 0. When the PGA reaches 0.8g, the structure has only a very small probability of collapse of 4.3%. When the PGA reaches 2g, the collapse probability of the structure is 87.16%, which can fully meet the design requirement of "no collapse under major earthquakes".
[0129] The structural vulnerability assessment method based on a multi-fidelity neural network surrogate model described in this embodiment combines multi-fidelity modeling, structural hybrid testing technology, neural networks, and vulnerability calculation methods. It uses a multi-fidelity neural network surrogate model to obtain seismic structural response data of a three-story, three-span steel frame structure with double-order self-resetting buckling-resisting braces, and plots vulnerability curves based on the data, thereby improving the vulnerability assessment accuracy of large and complex civil engineering structures.
Claims
1. A structural vulnerability assessment method based on a multi-fidelity neural network surrogate model, characterized in that, include: S1. Collect seismic record data with uniformly distributed ground motion intensity parameters; S2. Establish a three-story, double-stage self-resetting buckling-restrained braced steel frame structure model and collect structural uncertainty parameters; base Using OpenSees and MATLAB to interactively construct finite element models and data processing algorithms; S3. Based on the seismic record data described in S1 and the structural uncertainty parameters described in S2, generate through random sampling. Initial low-fidelity data, and a CNN-LSTM neural network low-fidelity proxy model is built using the initial low-fidelity data; Based on the seismic record data described in S1 and the structural uncertainty parameters described in S2, and using Latin hypercubic-orthogonal design... This method is used to obtain initial high-fidelity data; S4. Input the initial low-fidelity data and the initial high-fidelity data described in S3 into the steel frame model, and then perform incremental dynamic... Force analysis method is used to obtain extended low-fidelity data, and combined with mixed test to obtain extended high-fidelity data. By applying incremental dynamic analysis method to steel frame model, sample data of structural deformation response are obtained in batches. The input includes important characteristic variables in seismic motion and structural parameters. The output is the maximum inter-story drift angle of the first and second floors as the key seismic response index. A neural network training sample database is constructed. S5. Based on the neural network training sample database described in S4, determine the input and output of the CNN-LSTM neural network low-fidelity proxy model, and optimize the neural network structure and hyperparameters; S6. Calculate the correlation coefficient between the augmented low-fidelity data and the augmented high-fidelity data described in S4, introducing uncertainty. Parameter optimization improves model robustness; a multi-fidelity agent model is established through hybrid training. S7. Based on the hybrid training described in S6, a multi-fidelity proxy model is established to obtain seismic response data. The seismic response parameters should be calculated based on the data under different ground motion intensity parameters, and the seismic response parameters should be calculated based on the ground motion intensity parameters and the earthquake. Regression analysis of response parameters is used to construct an earthquake demand model; S8. Calculate the structural damage probability density function and failure probability based on the earthquake demand model described in S7, and plot the vulnerability profile. curve.
2. The structural vulnerability assessment method based on a multi-fidelity neural network surrogate model according to claim 1, characterized in that, In S1, the earthquake record data adopts the 22 far-field ground motion sets recommended by ATC-63, which includes 14 earthquake records with magnitudes of 6.5-7.6 and epicentral distances of 8.7-98.22 km, with a field shear wave velocity of 192-724 m / s and a PGA range of 0.2-2.1g, and is evenly divided into four intervals.
3. The structural vulnerability assessment method based on a multi-fidelity neural network surrogate model according to claim 1, characterized in that, In S2, the three-layer double-stage self-resetting buckling-restrained braced steel frame structure model includes: a reset system, an energy dissipation system, and external connectors; The reset system includes: an inner tube, an outer tube, a combined disc spring, inner and outer tube baffles, and a disc spring; The energy-consuming system includes: a two-stage energy-consuming inner core and a constraining channel steel; The external connectors include: inner core connectors, inner tube connectors, outer tube connectors, and structural connectors.
4. The structural vulnerability assessment method based on a multi-fidelity neural network surrogate model according to claim 1, characterized in that, In S7, the earthquake demand model is obtained through the following steps: S7.1 Set the ground motion intensity parameters IM, including: peak ground acceleration (PGA), peak ground velocity (PGV), overall duration intensity (Ia), and average spectral acceleration. Ground motion intensity information; seismic response parameters DM include: top displacement and maximum inter-story drift angle. ; S7.
2. Based on the fact that the seismic ground motion intensity parameter IM and the seismic response parameter DM follow a normal distribution, their mathematical relationship is expressed as follows: ； Where α and β are regression coefficients; S7.
3. According to step S6.1, select the ground motion intensity parameter IM as the peak ground acceleration (PGA) and the seismic response parameter DM as the maximum inter-story drift angle. The formula is then obtained as follows: ； Taking the logarithm of both sides yields the following equation: ； By performing linear regression on the above formula, the values of α and β are obtained, thus yielding the earthquake demand model.
5. The structural vulnerability assessment method based on a multi-fidelity neural network surrogate model according to claim 4, characterized in that, S8 specifically includes: S8.1 Define the failure probability of a structure exceeding its structural demand capacity parameter C under different seismic intensities, wherein the failure probability is determined by: ； The value is obtained, where Pf is the failure probability. For structural requirement capacity parameters; S8.2, Settings And it follows a normal distribution, with the mean expressed as: Standard deviation is The formula for deriving the failure probability is: ； Where Z is the representative value; S8.3, will Convert to standard normal distribution The failure probability is obtained as follows: ； S8.4 When IM is PGA, Use a value of 0.5 to plot the fragility curve.