Topological optimization method and system for aseismic structure based on meta-learning

Abstract

The invention discloses an aseismic structure topological optimization method and system based on meta-learning. The method comprises the steps of obtaining model data of a target structure; simulating structural response samples under different disturbance conditions from the model data through earthquake-like disturbance; calculating a response value of the structure response sample, and comparing the response value with a preset first threshold value; the structural response samples with the response values larger than the first threshold value serve as key samples, and the rest of the structural response samples are compressed to serve as non-key samples; complementing the missing part of the key sample through a deep learning model; if verification inconsistent samples exist in the complemented key samples, removing the part of samples, and then summarizing the part of samples with non-key samples to form a sample set; and extracting structure information of the model data as a conditional code, and inputting the conditional code and the sample set into a pre-constructed meta-learning model to obtain an anti-seismic structure topology scheme corresponding to the target structure. According to the invention, through simulation and deletion completion, the perception ability of a subsequent learning model to a structure change trend is improved.

Description

Technical Field

[0001] This invention relates to the field of structural topology optimization technology, specifically to a method and system for seismic-resistant structural topology optimization based on meta-learning. Background Technology

[0002] With the continuous increase in urban construction density and the number of high-rise buildings, the safety of building structures under extreme disturbances such as earthquakes has received increasing attention. To achieve efficient seismic design of building structures, the academic and engineering communities have widely adopted methods combining structural optimization and intelligent algorithms in recent years to explore optimal structural topology schemes that meet seismic performance requirements. Traditional seismic topology optimization techniques are mainly based on deterministic methods, such as topology optimization (TO), density method (SIMP), or evolutionary algorithm (GA). While these methods have certain engineering practicality, they have shortcomings in the following aspects:

[0003] Current methods often use static loads or linearly simplified seismic loads for structural response analysis, which frequently results in missing structural response data. Relying on interpolation and elimination strategies for missing data cannot restore the deep physical correlations and makes it difficult to comprehensively simulate the complex, nonlinear, and multi-disturbance scenarios in actual earthquakes, leading to poor adaptability of the generated structural topology. Summary of the Invention

[0004] To address the issues of generalization ability and deep physical correlation of existing sample data, this invention provides a method and system for topology optimization of seismic-resistant structures based on meta-learning.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] In the first aspect, this application discloses a topology optimization method for seismic-resistant structures based on meta-learning, which includes the following steps:

[0007] Obtain model data of the target structure;

[0008] The model data was used to simulate structural response samples under different perturbation conditions using earthquake-like perturbation.

[0009] The response value of the structural response sample is calculated by normalizing the response and the energy contribution ratio, and then the response value is compared with a preset first threshold.

[0010] Structural response samples with response values ​​greater than the first threshold are designated as key samples, while the remaining structural response samples are compressed and designated as non-key samples.

[0011] The missing parts of the key samples are filled in using a pre-built deep learning model, and the consistency of the filled key samples with non-key samples is verified.

[0012] If there are inconsistent samples in the completed key samples, remove those samples and then combine them with the non-key samples to form a sample set.

[0013] The structural information extracted from the model data is used as a conditional code. This code, along with the sample set, is input into a pre-constructed meta-learning model to obtain the seismic topology scheme corresponding to the target structure. The structural information includes material properties, number of stories, and construction form.

[0014] Secondly, this application discloses a topology optimization system for seismic-resistant structures based on meta-learning, which includes a data acquisition module, a simulation response module, a data comparison module, a data compression module, a consistency verification module, a sample aggregation module, and a result output module.

[0015] The data acquisition module is used to acquire model data of the target structure;

[0016] The simulation response module is used to simulate structural response samples under different disturbance conditions using model data through seismic-like disturbances;

[0017] The data comparison module is used to calculate the response value of the structural response sample by normalizing the response and the energy contribution ratio, and compare the response value with a preset first threshold.

[0018] The data compression module is used to designate structural response samples with response values ​​greater than a first threshold as key samples, and compress the remaining structural response samples as non-key samples.

[0019] The consistency verification module is used to fill in the missing parts of key samples using a pre-built deep learning model, and then verify the consistency between the filled key samples and non-key samples.

[0020] If there are inconsistent samples in the completed key samples, the sample aggregation module will remove those samples and then aggregate them with the non-key samples to form a sample set.

[0021] The output module is used to extract the structural information of the model data as conditional encoding, and input it along with the sample set into the pre-built meta-learning model to obtain the seismic topology scheme corresponding to the target structure; the structural information includes material properties, number of stories, and construction form.

[0022] Thirdly, this application discloses a computer-readable storage medium storing a program for implementing the aforementioned meta-learning-based seismic structure topology optimization method.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] 1. This application can generate structural response data covering a variety of complex disturbance scenarios by performing seismic-like disturbance simulation on the target structure. It can also use a deep learning model to complete the missing key response samples, restore response data with more physical logic and spatiotemporal correlation, and improve the ability of subsequent learning models to perceive the trend of structural changes.

[0025] 2. This application incorporates structural information such as material properties, number of stories, and construction form as conditional codes, which are then input into the meta-learning model along with response samples. This achieves deep coupling between structural parameters and performance output, enabling the generated seismic topology design to have stronger personalized adaptability. Attached Figure Description

[0026] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0027] Figure 1 This is a flowchart of the seismic structure topology optimization method based on meta-learning introduced in this invention;

[0028] Figure 2 Based on Figure 1 The flowchart shows the structural response samples under different perturbation conditions simulated using model data from earthquake-like perturbations;

[0029] Figure 3 Based on Figure 1 A flowchart showing how to adjust the compression ratio based on the response value during non-critical sample compression processing;

[0030] Figure 4 Based on Figure 1 A flowchart of the topology scheme output by the validation meta-learning model;

[0031] Figure 5 Based on Figure 4 The flowchart shows the process of stopping iteration and reporting an error when the meta-learning model optimization reaches the required number of iterations and still fails to meet the requirements for subsequent operations.

[0032] Figure 6 Based on Figure 1 A flowchart for filtering based on the number of target structures;

[0033] Figure 7 Based on Figure 1 The flowchart of meta-learning.

[0034] Figure 8 This is a block diagram of a topology optimization system for earthquake-resistant structures based on meta-learning. Detailed Implementation

[0035] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0036] Application Overview

[0037] In existing technologies, traditional seismic topology optimization methods typically employ a deterministic analysis framework, relying on static loads or simplified seismic loads to calculate structural response. These methods struggle to capture the true response characteristics of structures under nonlinear multi-disturbance conditions when dealing with complex seismic disturbances. In practical engineering, limitations in sensor deployment or computational resources often lead to localized gaps in structural response data. While linear interpolation or direct data removal are commonly used to address missing data, they fail to restore the physical relationships between data points, resulting in insufficient adaptability of the optimized structural topology scheme to real-world seismic scenarios.

[0038] To address the aforementioned issues, research revealed that the core flaw of traditional methods lies in the disconnect between the data recovery mechanism and physical laws. To address the problem of missing structural response data, attempts were made to introduce deep learning models for data completion, but it was found that the data after simple completion deviated from the true perturbation response. Further investigation revealed that structural responses under different perturbation conditions exhibit differences in energy distribution, and high-energy response regions have a decisive impact on topology optimization. Based on this, a proposal was made to select key samples by energy contribution ratio, combining deep learning completion with a meta-learning framework to achieve dual constraints of data-driven approaches and physical laws.

[0039] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0040] Example 1

[0041] like Figure 1 As shown, a topology optimization method for seismic-resistant structures based on meta-learning is introduced, including the following steps:

[0042] S1. Obtain model data for the target structure;

[0043] S2. Simulate structural response samples under different perturbation conditions using model data through seismic-like perturbations;

[0044] S3. Calculate the response value of the structural response sample by normalizing the response and energy contribution ratio, and compare the response value with the preset first threshold.

[0045] S4. Structural response samples with response values ​​greater than the first threshold are designated as key samples, and the remaining structural response samples are compressed and designated as non-key samples.

[0046] S5. Use a pre-built deep learning model to fill in the missing parts of the key samples, and verify the consistency between the filled key samples and non-key samples.

[0047] S6. If there are inconsistent samples in the completed key samples, remove those samples and then combine them with the non-key samples to form a sample set;

[0048] S7. Extract the structural information from the model data as conditional encoding, and input it along with the sample set into the pre-built meta-learning model to obtain the seismic topology scheme corresponding to the target structure; the structural information includes material properties, number of stories, and construction form.

[0049] Regarding step S1, the model data is a set of parameters used to characterize the static and dynamic response characteristics of the target structure. It includes structural geometric information, such as the plan dimensions of each floor, floor height and total number of floors, nodes (node ​​number, spatial coordinates), component layout (topological connection relationship of beams, columns, walls, floor slabs, etc.), and component dimensions (beam cross-sectional dimensions, column height-to-width ratio, shear wall thickness, etc.).

[0050] Material properties: such as the elastic modulus, yield strength, Poisson's ratio, and density of the material used in the component.

[0051] Construction form and connection method: support type, structural redundancy, connection node type;

[0052] It can also include boundary and load conditions: foundation stiffness model, fixed constraint conditions.

[0053] It can also include mesh / discrete information: finite element type (beam element, shell element, solid element), mesh partitioning information (partition density, number of elements), element number and its material / component mapping relationship.

[0054] Apart from material properties, number of layers, and structural form, the remaining model information is processed and categorized according to actual needs for subsequent steps. Specifically, mesh / discrete information can be used for subsequent in-depth simulation, and boundary and load conditions can be used for disturbance simulation.

[0055] Regarding step S2, firstly, define a simulation model for seismic disturbances. These disturbances may include:

[0056] Seismic waves of different frequencies and amplitudes.

[0057] Different propagation modes of seismic waves (P-waves, S-waves, etc.).

[0058] Different modes of action of seismic waves (base vibration, wind load, etc.).

[0059] Choose a seismic excitation model suitable for the target structure, such as using real seismic records (e.g., NPRD, El Centro seismic waves) or synthetic seismic waves.

[0060] Disturbance conditions settings:

[0061] Different input conditions for seismic waves can be set based on the type, material properties, size, and shape of the target structure. These conditions can include the amplitude, frequency, and duration of the seismic waves. By selecting different disturbance input conditions such as acceleration, displacement, and frequency, seismic waves of varying intensities and properties can be simulated.

[0062] Using selected seismic waves and structural models, multiple seismic disturbance samples are designed to simulate structural responses under different intensities, frequencies, and time characteristics. The comprehensiveness of the simulation results can be improved by increasing the number of disturbance samples.

[0063] Structural analysis and response acquisition:

[0064] Finite element analysis (FEA) or other structural dynamic analysis methods are used to simulate the response of the target structure under different seismic disturbances. For example, the seismic wave as the input disturbance is usually represented by acceleration time history data, which shows the time variation of the seismic wave.

[0065] Extract structural response data samples under various seismic disturbances. Each sample should include structural response data at different time points and under different disturbance conditions.

[0066] The simulated response data is normalized to ensure comparability under different perturbation conditions. Preprocessing steps such as filtering and denoising can be performed to remove irrelevant noise data.

[0067] The processed data is decomposed into response samples to ensure that each sample represents the structural response under a specific disturbance condition. The structural response of each sample may include acceleration response, displacement response, vibration mode, etc., to ensure that the effects of seismic waves on various parts of the structure are covered.

[0068] Based on the simulation results, the response samples under different perturbation conditions are calibrated to ensure that the perturbation conditions, frequency, and intensity of each sample can accurately correspond one-to-one with the sample data.

[0069] Finally, the structural and response data of all response samples are saved, providing input for subsequent steps such as normalization and energy contribution calculation.

[0070] Regarding step S3, in order to make the structural response samples under different perturbation conditions comparable, it is first necessary to normalize the structural response samples. That is:

[0071] ;

[0072] in: It is the structural response at time t; It is the maximum value of the response sample.

[0073] Normalized structural response It can eliminate the scaling effect, allowing different response samples to be compared under the same standard.

[0074] In the dynamic analysis of multi-degree-of-freedom systems, each degree of freedom of the structure may contribute different amounts of energy to the overall response. Therefore, it is necessary to calculate the energy contribution ratio of each degree of freedom in order to weigh the impact of each degree of freedom on the overall structural response.

[0075] Energy contribution ratio It is usually calculated by the ratio of the response energy of each degree of freedom to the total response energy:

[0076] ;in, Let be the response energy of the i-th degree of freedom; This represents the total response energy of the entire structure.

[0077] Assumption Let be the displacement response of the i-th degree of freedom, then the response energy of that degree of freedom is... It can be calculated as follows: ;

[0078] in, It is the mass of the i-th degree of freedom; It is the natural frequency corresponding to the i-th degree of freedom; and These are the velocity and displacement of the i-th degree of freedom, respectively.

[0079] Total response energy : .

[0080] Therefore, the final response value of the structural response sample is:

[0081] .

[0082] Regarding the first threshold The determination is based on one of the following methods, or a combination thereof: structural design standards, structural performance requirements, statistical analysis, risk assessment, numerical simulation, and expert experience.

[0083] Regarding step S4, The structural response sample is used as the key sample. The structural response sample is a non-critical sample and needs to be compressed (e.g., through data compression algorithms, denoising, etc.) to reduce storage and computation costs and facilitate further analysis.

[0084] Regarding step S5, firstly, each key sample needs to be analyzed to identify the missing data. Typically, missing data can be labeled based on the following methods:

[0085] Missing time steps: Some time step data may be missing in the structural response data.

[0086] Missing degrees of freedom response: Some degrees of freedom (such as displacement or acceleration in a certain direction) may not have response data.

[0087] Incomplete amplitude: The amplitude of the partial response may be missing data.

[0088] The method for marking missing parts can be based on the actual missing data situation, and can be achieved by supplementing the data with markings.

[0089] Then, a deep learning model was used to complete the data.

[0090] The deep learning model is trained using a sample set containing complete data. The training objective is to enable the model to predict missing parts from known responses. The input sample data consists of known data (e.g., known structural responses, displacements, accelerations, etc.), and the output is the predicted missing parts.

[0091] Regarding the selection of model architecture, it is based on the missing data type, as detailed below:

[0092] Autoencoders are suitable for dimensionality reduction and feature learning of data. They generate a representation of the latent space from the missing parts through an encoder and then restore it through a decoder.

[0093] Variational autoencoder: It adds a probabilistic model to the autoencoder and is suitable for handling missing data with high uncertainty.

[0094] Generative Adversarial Networks (GANs): These networks use generators to generate data with missing parts, and discriminators to evaluate the authenticity of the generated data. They are suitable for handling complex and non-linear missing patterns.

[0095] The following explanation uses the example of selecting a generative adversarial network.

[0096] A GAN model consists of a generator, a discriminator, and a loss function.

[0097] The generator's task is to generate samples that closely approximate real-world data. For missing data completion, the generator's goal is to generate the missing parts based on the known parts.

[0098] The input is a known portion of the data (samples with missing values). The generator network typically contains multiple fully connected layers (or convolutional layers) that use non-linear activation functions (such as ReLU and Tanh) to generate complete data. The output is the generated complete sample data. .

[0099] The discriminator's task is to distinguish whether the input sample comes from real data or generated data. The discriminator outputs a probability value, representing the probability that the input data is real data. Input: Samples generated by the generator. and real data samples Output: Probability value or , which represents the probability that the input data is the real data.

[0100] The training objective of GANs is to enable the generator to produce data that is as realistic as possible, while the discriminator can effectively distinguish between generated and real data. The loss function of a GAN consists of two parts:

[0101] Generator loss function: .

[0102] Discriminator loss function: .

[0103] During training, the generator and discriminator constantly compete against each other, eventually enabling the generator to produce high-quality completion data.

[0104] Consistency verification can employ methods such as Euclidean distance and cosine similarity. The following example uses Euclidean distance.

[0105] Euclidean distance is used to calculate the key samples after completion. Non-critical samples Differences between them:

[0106] .

[0107] Regarding step S6, a threshold ϵ is set to judge the consistency between samples. If the verification result is less than the threshold, the two samples are considered to have high consistency at the feature level. When the consistency verification finds that the difference between the completed key sample and the non-key sample is greater than the set threshold ϵ, inconsistent samples need to be removed.

[0108] Finally, the completed key and non-key samples, after consistency verification and retention, will be aggregated to form a new sample set:

[0109] .

[0110] in, This is to complete the key samples after the consistency verification is passed.

[0111] Regarding step S7, assume that the model data of the target structure consists of the following parts:

[0112] Material properties: .

[0113] Number of floors: L, which is the number of floors in the building, usually an integer.

[0114] Construction form: F represents the construction form of the structure, such as frame structure, shear wall structure, etc. It can be represented by numerical code, such as F∈{0,1,2,…} corresponding to different construction forms.

[0115] For material properties and the number of layers, this information is processed through normalization and standardization to obtain a unified encoding form. The construction form F can be represented by one-hot encoding or integer encoding.

[0116] The normalized material properties, number of layers, and structural form are combined to obtain the final conditional code. . It is a vector that contains material properties, number of layers, and construction form.

[0117] like Figure 7 As shown, with For sample set Taking the feature vector of the i-th sample as an example, suppose the sample This is a multi-dimensional feature vector. The input to the meta-learning model is... and The two are combined to form a joint feature vector: . This represents the fusion function.

[0118] Meta-learning layer based on joint features The model generates seismic-resistant structural topology schemes through a task generation strategy. The goal of this part is to learn corresponding seismic design schemes based on different structural conditions.

[0119] Output layer: Outputs the seismic-resistant structural topology scheme corresponding to the target structure. The topology scheme can be a set of structural parameters, including the dimensions and layout of columns, beams, and walls.

[0120] For training the meta-learning model, a training objective is first defined: to enable the model to generate suitable seismic-resistant structural topology schemes based on different structural information (such as material properties, number of stories, and construction form) and seismic response samples. To achieve this objective, mean squared error (MSE) can be used as a loss function to measure the predicted topology scheme. Differences between the actual topology scheme y and the actual topology scheme y:

[0121] .

[0122] in, Let i be the predicted topology scheme for the i-th sample; This is a realistic topology scheme.

[0123] By continuously updating the model parameters through backpropagation and optimization algorithms (such as Adam or SGD) to minimize the loss function, a meta-learning model capable of effectively generating seismic-resistant structural topology schemes is trained.

[0124] Therefore, this application ensures the physical rationality of the sample data through deep learning completion and consistency verification. The generated topology scheme can adapt to different seismic disturbance modes, avoiding the local failure phenomenon that occurs in traditional schemes in complex seismic scenarios. The introduction of structural attribute encoding enables the optimization process to fully integrate engineering constraints, generating seismic-resistant structural schemes that meet actual construction requirements.

[0125] The main scheme of this application has been introduced above. The following section details step S2, simulating structural response samples under different perturbation conditions using seismic-like perturbations based on model data. Figure 2 As shown, the specific steps are as follows:

[0126] S201. The model data is superimposed through a pre-established stochastic process to obtain seismic-like disturbance inputs simulating different seismic scenarios;

[0127] S202. The structural response under different disturbance conditions is obtained by multi-degree-of-freedom linear calculation of the seismic disturbance input, and the structural response samples under different disturbance conditions are summarized.

[0128] Modeling of stochastic processes with seismic-like disturbance inputs:

[0129] The input of seismic disturbances typically uses stochastic processes to represent the stochastic characteristics of seismic waves. In this embodiment, the seismic wave model describes the distribution of the disturbance signal in the frequency domain using the power spectral density (PSD) function.

[0130] Assuming the seismic disturbance a(t) is a Gaussian white noise process, with power spectral density... for:

[0131] Where C is a constant that controls the intensity of the disturbance; is the frequency; n is the spectral index, usually 1 or 2.

[0132] For seismic input, a time-domain seismic wave signal can be generated using the Discrete Fourier Transform (DFT) after a frequency-domain to time-domain transformation. The formula for generating the seismic wave input a(t) is as follows:

[0133] ;in, For frequency components; The phase is random; N is the number of frequency components.

[0134] This method generates It is a time series of earthquake-like disturbances that can simulate dynamic disturbances under different earthquake conditions.

[0135] To obtain more complex seismic wave inputs, disturbance signals with multiple frequency components can be combined by superposition.

[0136] In the finite element method, the response calculation under the action of seismic waves usually involves the strain, stress and other states of each element.

[0137] Typical dynamic response representation of a structural element: ;in, Represents the mass matrix, , These are acceleration and velocity, respectively. Indicates the displacement response of the structure; It is the damping matrix; It is a quality matrix. It is an external force excited by the acceleration field of seismic waves.

[0138] Numerical methods (such as the Newmark method or the Runge-Kutta method) can be used to numerically solve the above differential equations, and obtain the displacement, velocity and acceleration responses of the structure at each time step.

[0139] Taking the Newmark method as an example, the update formula is:

[0140]

[0141]

[0142] in, It is the time step; It is the displacement of the structure at time t; by using this method, the structural response at each time step can be obtained, forming a complete response sample set of the structure under seismic disturbance.

[0143] Structural response calculated under each disturbance condition (Including displacement, velocity, and acceleration) can be used as structural response samples. By summarizing the response samples under different disturbance conditions, a final structural response sample containing response data under all disturbance conditions is obtained.

[0144] Therefore, this application generates a sample set covering multidimensional perturbation conditions, providing more complete data support for subsequent topology optimization. Simultaneously, the response data obtained based on multi-degree-of-freedom calculations can accurately reflect the actual mechanical behavior of the structure under dynamic loads, avoiding the loss of physical correlation caused by single-degree-of-freedom simplification, thereby improving the adaptability of seismic-resistant structural topology schemes to complex actual earthquake conditions.

[0145] Step S2 has been described above. The following section details the adjustment of the compression ratio based on the response value during the compression processing of non-critical samples in step S4. Figure 3 As shown, the details are as follows:

[0146] S401. Calculate the response difference by comparing the response value of the non-critical sample with the first threshold.

[0147] S402. Calculate the mean difference of the response differences of several non-critical samples;

[0148] S403. The corresponding compression ratio is obtained by matching a preset linear curve representing the mapping relationship between compression ratio and response difference.

[0149] During the compression process, the difference between the response value of each non-critical sample and a preset threshold is first calculated to obtain a response difference index reflecting the importance of the sample. Then, the mean of the response differences for non-critical samples of the same structure (a portion of the target structure) processed in the same batch is calculated to eliminate compression ratio setting deviations caused by local data fluctuations. Based on a preset linear mapping relationship, the compression ratio parameter corresponding to the mean difference is output. For example, when the mean difference is in a high range, it indicates that the batch of samples has high potential value; in this case, a lower compression ratio is matched to retain more detailed features. When the mean difference is in a low range, it indicates that the sample redundancy is high; in this case, a higher compression ratio is used to improve storage efficiency.

[0150] The normalized response difference is obtained by using a preset linear function. Mapped to compression ratio Define the linear compression mapping function: ;

[0151] in, This represents the minimum compression ratio, corresponding to the sample with the smallest response difference, and is usually set to 1.0 (i.e., no compression). This represents the maximum compression ratio, corresponding to the sample with the largest response difference, and is usually taken as a small value such as 0.1.

[0152] This application addresses the problem of wasted storage resources or loss of critical information caused by using a fixed compression ratio when compressing non-critical samples in traditional seismic topology optimization. By dynamically adjusting the compression ratio, it avoids both low-value samples occupying excessive storage space and high-potential-value samples suffering feature distortion due to over-compression, thus providing a more accurate data foundation for subsequent meta-learning model training.

[0153] The above details how to adjust the compression ratio based on the response value when compressing non-critical samples. The following describes how, before the conditional codes and sample sets are input into the pre-built meta-learning model, the significance of the conditional codes is determined by a normalized weighted combination method, and the processing order is set accordingly.

[0154] Due to conditional coding It contains multiple different structural information parts, such as material properties, number of layers, construction form, etc., in order to enable conditional coding. Each condition component Normalization can be performed under the same units of measurement. Assume the maximum value of each feature is... The minimum value is Then the normalized conditional coding The calculation is as follows: . The normalized feature typically takes values ​​in the range [0,1].

[0155] To determine the importance or significance of each feature, it is necessary to consider each conditional feature. Assign a weight This indicates the degree to which the feature influences the model. Weights It can be determined through domain knowledge, prior experience, or data-based training.

[0156] The significance of each conditional feature is obtained by weighted combination. This value represents the importance of each conditional feature to the overall structural design. Significance The calculation formula is as follows:

[0157] .

[0158] To determine the processing order of each feature based on its significance, the overall significance can be calculated. It represents the sum of the importance of all features. The formula for calculating the overall significance is:

[0159] .

[0160] Based on significance The size allows you to set the processing order for each feature. This is based on the size of each structure. It allows you to set the processing order for multiple structures.

[0161] This application addresses the problem of low model efficiency caused by unreasonable weighting of structural information features. For example, in the seismic optimization of high-rise buildings, prioritizing material property features enables the model to quickly establish a relationship between strength and deformation; in the optimization of irregular structures, weighted processing of structural form features helps the model accurately identify the stress characteristics of key connection nodes. This saliency-based processing order allows the meta-learning model to complete higher-quality feature extraction within the same training period, improving the generation accuracy and convergence speed of topology schemes.

[0162] The above details how the processing order of the meta-learning model is determined based on the saliency of the conditional encoding. The following section, before obtaining the seismic-resistant topology scheme corresponding to the target structure, also provides a detailed explanation of how to validate the topology scheme output by the meta-learning model. Figure 4 As shown, the details are as follows:

[0163] S701. The seismic performance index of the topology scheme output by the meta-learning model is calculated by finite element simulation;

[0164] S702. Determine whether the seismic performance index is greater than or equal to the preset seismic standard. If so, output the topology scheme as the seismic structure topology scheme.

[0165] S703. Otherwise, trigger the meta-learning model to re-encode the topology scheme based on the conditions and the sample set until the seismic resistance standard is met, or stop iterating and report an error if the number of iterations is reached but the standard is still not met.

[0166] The seismic performance indicators include the maximum inter-story drift angle, the equivalent base shear coefficient, the equivalent damping ratio, and the periodic elongation ratio. One or more of these indicators can be selected as metrics to evaluate the feasibility of the topology scheme.

[0167] The following are examples of the calculation formulas for several indicators.

[0168] Maximum inter-story drift angle: ;

[0169] in, , These are the horizontal displacements of the i-th layer and the (i-1)-th layer, respectively; Let the height be the height of the i-th layer; This represents the total number of structural layers.

[0170] Equivalent base shear coefficient: ; This represents the total shear force at the bottom of the structure. This represents the total weight of the structure.

[0171] Equivalent damping ratio: ; Energy dissipation for each cycle, This represents the maximum elastic potential energy.

[0172] Structural periodicity ratio: ; The effective period of the structure under disturbance. This is the initial structural period.

[0173] The preset seismic performance standard is , indicating the permissible range of multiple indicators.

[0174] Define the performance vector corresponding to the structural topology scheme as: .

[0175] The following Boolean conditional expression is used for evaluation:

[0176]

[0177]

[0178] like If the topology is correct, output the solution. Otherwise, proceed to the next iteration.

[0179] The iteration will re-invoke the meta-learning model based on the same conditional encoding and sample set to output a new topology scheme, and repeat the above process until the criteria are met or the maximum number of iterations is reached. If no effective scheme is found after exceeding the maximum number of iterations, an error flag will be output or the process will proceed to the manual verification stage (depending on the actual deployment strategy).

[0180] This application combines finite element simulation with iterative optimization to ensure that the final output scheme meets both the efficiency requirements of data-driven optimization and the seismic performance standards under the constraints of physical laws, thereby improving the engineering reliability of the structural topology scheme.

[0181] The topology scheme for validating the meta-learning model output has been described in detail above. The following section provides a detailed explanation of the subsequent operations after the meta-learning model optimization reaches the required number of iterations but still fails to meet the requirements, resulting in a stop iteration and an error message. Figure 5 As shown, the details are as follows:

[0182] S711. Obtain the vibration isolation elements associated with the target structure;

[0183] S712. Simulate the response samples of the seismic isolation elements under different disturbance conditions through seismic-like disturbances;

[0184] S713. Input the component response samples, sample set, and conditional codes back into the meta-learning model to obtain the seismic-resistant structural topology scheme corresponding to the target structure.

[0185] The model parameters for each seismic isolation element may include one or more of the following: horizontal stiffness, hysteresis characteristics, energy dissipation coefficient, yield displacement, ultimate displacement, and installation location. Other parameters may also be included. These parameters are then processed to form a parameter vector for the seismic isolation element.

[0186] Each isolation element can generate multiple sets of disturbance samples, and all element samples are integrated to form an element response sample.

[0187] The component response samples are fused with the original input sample set and conditional codes to form an enhanced input, which is then fed into the meta-learning model. If the output scheme meets the seismic resistance criteria, it is adopted as the seismic-resistant structural topology scheme; otherwise, iteration can continue or the design can be terminated.

[0188] When the meta-learning model fails to optimize, this application supplements the response data of the seismic isolation elements, enabling the model to generate a topology scheme that meets seismic requirements based on the structure-device collaborative working mechanism, thus avoiding the waste of computational resources caused by repeatedly adjusting invalid parameters.

[0189] The above details the subsequent operations after stopping the iteration and reporting an error. Below, before obtaining the seismic isolation elements associated with the target structure, further filtering is performed based on the number of target structures, such as... Figure 6 As shown, the details are as follows:

[0190] S721. Count the number of target structures and determine whether it is greater than the preset second threshold;

[0191] S722. If so, match the corresponding associated seismic isolation element according to the key sample of the target structure.

[0192] The number of target structures refers to the total number of building structures to be optimized. This can be calculated using database queries or file parsing. This parameter is used to determine the scale of the optimization scenario.

[0193] The associated isolation elements of the target structure are screened by a second threshold. If the number of target structures is greater than the second threshold, it indicates that there are too many target structures. When the number of structures is large, the isolation element screening strategy based on the correlation of key samples is introduced to reduce unnecessary computational redundancy and improve response speed and model inference efficiency.

[0194] Regarding the determination of the second threshold, an adaptive mechanism can be constructed based on historical data to dynamically adjust the threshold according to factors such as the average time consumption, failure rate, and resource consumption of historical tasks.

[0195] ;

[0196] in, This is the update threshold after the (n+1)th iteration; This is the second threshold during the nth iteration; For adjustment coefficients; The target performance index value; These are the actual measured performance index values.

[0197] Matching the corresponding isolation elements to key samples of the target structure can be achieved using Euclidean distance to measure the degree of matching between the sample response and the element parameters.

[0198] Weighted Euclidean distance: ;

[0199] in, Assign importance coefficients to the response parameters as weight vectors; The key sample response vector; This represents the performance vector of the seismic isolation element.

[0200] For each structure, select the K best-matching isolation elements to obtain the recommended isolation elements for each structure.

[0201] This application enables rapid resource selection in large-scale building complex scenarios. Simultaneously, through a physical correlation mechanism between key samples and the performance of seismic isolation components, it ensures that the selected isolation components can specifically improve the weak response areas of the structure, avoiding the correlation loss problem caused by full data processing in traditional methods.

[0202] After obtaining the seismic topology scheme corresponding to the target structure, the non-critical samples will be deleted, and only the critical samples will be stored for further explanation.

[0203] After confirming the effectiveness of the seismic-resistant structural topology scheme, structural response samples with response values ​​below a first threshold are removed from the complete sample set; these are termed "non-critical samples." Removal methods include, but are not limited to:

[0204] Clear from memory;

[0205] Delete from the temporary disk directory;

[0206] Mark it as "processed" for subsequent periodic cleanup.

[0207] The retained key samples will undergo data structuring processing, including:

[0208] Add index information (such as structure number, disturbance number, disturbance magnitude, etc.); organize according to structure ID.

[0209] Update the current key samples to the sample database or model input cache, and record in the system log that the structure has completed topology generation and sample simplification, providing traceable analysis; if there is a structure for multi-structure parallel optimization task, mark it as "archived" to avoid duplicate calculations.

[0210] This application optimizes storage space utilization through an intelligent filtering mechanism, while improving the efficiency of data retrieval during subsequent model iterations. Retaining key samples provides high-value data support for model updates, while deleting non-key samples avoids the negative impact of invalid data on system performance.

[0211] To facilitate understanding of the above embodiments, a specific application scenario of the above embodiments will be used as an example for illustration below:

[0212] A city located in a region of medium to high seismic intensity plans to build a new medical center complex, comprising 15 buildings of varying heights, uses, and structural forms (including inpatient buildings, outpatient buildings, emergency centers, and research buildings). Given the crucial role of medical buildings in social rescue efforts during earthquakes, the designers require each building to not only meet current seismic design standards but also further optimize its seismic performance and structural topology. This aims to achieve high seismic toughness, cost control and efficient material allocation, unified optimization of diverse building types, and support for future modular construction and upgrades.

[0213] A total of 15 BIM model files were acquired. Multiple domestic seismic record samples were used, and 30 sets of artificial seismic records (peak acceleration 0.15–0.5g) were generated. These were superimposed to form a disturbance input. 30 multi-degree-of-freedom dynamic analyses were performed on each structure, resulting in a total of 450 response samples.

[0214] By comparing the response values ​​of the structural response samples with the first threshold, 142 key samples and 308 non-key samples were selected.

[0215] A trained deep learning model is used to complete the missing data of key samples. Some key samples have local missing data due to complex structure or severe perturbation. The system completes the samples using a pre-trained deep learning model and verifies the consistency of response trends with non-key samples, removing abnormal samples. The final sample set is constructed as follows: 125 key samples + 308 compressed samples (vector dimensionality reduced to 64 dimensions).

[0216] The structural information of each building (such as steel-concrete composite, frame-shear wall structure, structural height, etc.) is converted into conditional codes and input into the meta-learning model in combination with the sample set.

[0217] After the preliminary topology design is output, the seismic performance of the structure is calculated, with the main indicators being:

[0218] Maximum inter-story drift angle Requirement: <1 / 550;

[0219] Base shear ratio: Meets the design limit without exceeding the limit;

[0220] Modal periodicity comparison: dominant mode frequency > 0.5Hz;

[0221] The first batch of buildings that met the standards: 12 buildings;

[0222] Not up to standard: 3 buildings (high number of floors).

[0223] After two rounds of iteration, 14 buildings met the standards, while 1 building still failed.

[0224] Failed building information:

[0225] The building is 18 stories high and has a frame-shear wall structure. The original foundation design was a rigid raft foundation. The system recommended the introduction of 12 high-damping rubber seismic isolation bearings (horizontal stiffness: 3000kN / m; damping ratio: 15%; vertical bearing capacity: 3500kN). After seismic isolation simulation, all structural indicators met the standards.

[0226] Only the final key samples are stored: 125 groups + 8 groups of seismic isolation enhancement samples; all non-key samples are deleted. A sample index database is built for the stored samples to support rapid topology reasoning for similar future buildings.

[0227] This embodiment combines a meta-learning model with a key sample screening mechanism to distinguish between key and non-key samples when simulating earthquake disturbances. It uses deep learning to fill in missing data and performs consistency verification, ultimately generating a topology scheme that meets seismic resistance standards, thereby improving optimization efficiency and enhancing the accuracy and reliability of the scheme.

[0228] Example 2

[0229] like Figure 8 As shown in the figure, this embodiment introduces a seismic structure topology optimization system based on meta-learning, including a data acquisition module 801, a simulation response module 802, a data comparison module 803, a data compression module 804, a consistency verification module 805, a sample aggregation module 806, and a result output module 807.

[0230] The data acquisition module 801 is used to acquire model data of the target structure;

[0231] The simulation response module 802 is used to simulate structural response samples under different disturbance conditions by using model data through seismic-like disturbances;

[0232] The data comparison module 803 is used to calculate the response value of the structural response sample by normalizing the response and the energy contribution ratio, and compare the response value with a preset first threshold.

[0233] The data compression module 804 is used to take the structural response samples with response values greater than the first threshold as key samples, and compress the remaining structural response samples as non-key samples;

[0234] The consistency verification module 805 is used to complete the missing parts of the key samples through a pre-built deep learning model, and verify the consistency between the completed key samples and the non-key samples;

[0235] If there are inconsistent samples in the completed key samples, the sample summary module 806 removes these samples, and then summarizes them with the non-key samples into a sample set;

[0236] The result output module 807 is used to extract the structural information of the model data as conditional encoding, and input it and the sample set into a pre-built meta-learning model to obtain the seismic structural topology scheme corresponding to the target structure; the structural information includes material properties, number of floors, and construction form.

[0237] In some specific embodiments, the data compression module can adopt a compression algorithm based on wavelet transform to perform dimensionality reduction processing on the low-frequency vibration data in the non-key samples. Specifically, each non-key sample contains a multi-channel time series response, such as acceleration / velocity / displacement response, and each channel time series signal is processed separately.

[0238] Using discrete wavelet transform (DWT), common wavelet bases include Daubechies (db4), Symlets (sym5), etc., to decompose the signal into L levels: ; where is the low-frequency approximation coefficient of the Lth layer, is the high-frequency detail coefficient of the first layer.

[0239] Retain the low-frequency principal component , discard or quantize the high-frequency details , i < L / 2. Taking the 4-level decomposition as an example, is completely retained, 、 is quantized to 4~8bit representation, 、 is set to zero or discarded, and the compressed structure is stored.

[0240] The simulation response module can be configured with multiple groups of parallel computing units to process seismic wave inputs in different directions respectively. During the conditional encoding process, the construction form parameters can be transformed into spatial topology feature vectors through a graph neural network.

[0241] This embodiment reduces computational resource consumption while ensuring data integrity through collaborative processing of key sample screening and dynamic compression of non-key samples. By combining conditional coding and a meta-learning model for joint optimization, the generated topology scheme can adapt to different material properties and building construction forms, improving the adaptability of seismic-resistant structures in complex earthquake scenarios.

[0242] Example 3

[0243] This embodiment introduces a computer-readable storage medium storing a program for implementing the aforementioned meta-learning-based seismic structure topology optimization method.

[0244] The meta-learning-based seismic structure topology optimization method in Example 1 can be applied in software form, such as by designing it as a program that can run independently on a computer-readable storage medium, which can be a USB flash drive or a USB shield. The program is designed to start the entire method via an external trigger.

[0245] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A topology optimization method for seismic-resistant structures based on meta-learning, characterized in that, It includes the following steps: Obtain model data of the target structure; The model data was used to simulate structural response samples under different perturbation conditions using earthquake-like perturbation. The response value of the structural response sample is calculated by normalizing the response and the energy contribution ratio, and the response value is compared with a preset first threshold. Structural response samples with response values ​​greater than the first threshold are designated as key samples, while the remaining structural response samples are compressed and designated as non-key samples. The missing parts of the key samples are filled in using a pre-built deep learning model, and the consistency of the filled key samples with the non-key samples is verified. If there are inconsistent samples in the completed key samples, remove those samples and then combine them with the non-key samples to form a sample set. The structural information extracted from the model data is used as a conditional code, and it is input into the pre-constructed meta-learning model along with the sample set to obtain the seismic topology scheme corresponding to the target structure; the structural information includes material properties, number of layers, and construction form.

2. The seismic-resistant structural topology optimization method based on meta-learning according to claim 1, characterized in that, The specific steps for simulating structural response samples under different perturbation conditions using the model data are as follows: The model data is superimposed through a pre-established stochastic process to obtain seismic-like disturbance inputs simulating different seismic scenarios; The structural responses under different perturbation conditions are obtained by multi-degree-of-freedom linear calculation using the aforementioned seismic disturbance input, and the structural response samples under different perturbation conditions are summarized.

3. The seismic-resistant structural topology optimization method based on meta-learning according to claim 1, characterized in that, When compressing non-critical samples, the compression ratio is adjusted based on the response value, as follows: The response difference is calculated by comparing the response value of the non-critical sample with the first threshold. The mean difference is obtained by calculating the average response difference of several non-critical samples. The corresponding compression ratio is obtained by matching a pre-defined linear curve representing the mapping relationship between compression ratio and response difference.

4. The seismic-resistant structural topology optimization method based on meta-learning according to claim 1, characterized in that, Before inputting the conditional codes and sample set into the constructed meta-learning model, the significance of the conditional codes is determined by normalized weighted combination, and the processing order is set accordingly.

5. The seismic-resistant structural topology optimization method based on meta-learning according to claim 1, characterized in that, Before obtaining the seismic-resistant structural topology scheme corresponding to the target structure, the method further includes validating the topology scheme output by the meta-learning model, as follows: The seismic performance index of the topology scheme output by the meta-learning model is calculated by finite element simulation. Determine whether the seismic performance index is greater than or equal to the preset seismic standard; if so, output the topology scheme as the seismic structure topology scheme. Otherwise, the meta-learning model will be triggered to re-encode the topology scheme based on the conditions and the sample set until the seismic resistance standard is met, or if the number of iterations is reached but the standard is still not met, the iteration will stop and an error will be reported.

6. The seismic-resistant structural topology optimization method based on meta-learning according to claim 5, characterized in that, After the meta-learning model optimization reaches the required number of iterations but still fails to meet the stop iteration and reports an error, the following steps are also included: Obtain the vibration isolation elements associated with the target structure; The seismic isolation element was simulated under different disturbance conditions using seismic-like disturbances; The component response samples, along with the sample set and conditional codes, are input again into the meta-learning model to obtain the seismic-resistant structural topology scheme corresponding to the target structure.

7. The seismic-resistant structural topology optimization method based on meta-learning according to claim 6, characterized in that, Before obtaining the seismic isolation elements associated with the target structure, a screening process based on the number of target structures is also included, as follows: The number of target structures is counted, and it is determined whether the number exceeds a preset second threshold. If so, the corresponding isolation elements are matched based on the key samples of the target structure.

8. The seismic-resistant structural topology optimization method based on meta-learning according to claim 1, characterized in that, After obtaining the seismic-resistant structural topology scheme corresponding to the target structure, non-critical samples are deleted, and only critical samples are stored.

9. A topology optimization system for seismic-resistant structures based on meta-learning, characterized in that, It includes: The data acquisition module is used to acquire model data of the target structure. The simulation response module is used to simulate structural response samples under different disturbance conditions using the model data through seismic-like disturbances. The data comparison module is used to calculate the response value of the structural response sample by normalizing the response and the energy contribution ratio, and compare the response value with a preset first threshold. The data compression module is used to treat structural response samples with response values ​​greater than the first threshold as key samples, and compress the remaining structural response samples as non-key samples. The consistency verification module is used to fill in the missing parts of the key samples using a pre-built deep learning model, and then perform consistency verification between the filled key samples and the non-key samples. The sample aggregation module is used to remove samples with inconsistent verification if there are such samples in the completed key samples, and then aggregate them with non-key samples to form a sample set. The result output module is used to extract the structural information of the model data as conditional encoding, and input it along with the sample set into the pre-constructed meta-learning model to obtain the seismic structural topology scheme corresponding to the target structure; the structural information includes material properties, number of layers, and construction form.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program for implementing the topology optimization method for seismic-resistant structures based on meta-learning as described in any one of claims 1-8.

Images