Method for generating ground motion data matched with building damage based on physical guidance

Abstract

The application provides a physical guidance-based seismic motion data generation method matched with building earthquake damage, wherein a conditional generation adversarial network model comprises a generator and a discriminator, and a pre-trained building earthquake damage classification agent model is embedded as a physical performance evaluator, so that a physical guidance-based conditional generation adversarial network model is realized, the physical consistency problem in the seismic motion data generation process is solved, and the generation accuracy of the seismic motion data is improved. In addition, target seismic parameters and target earthquake damage label information are jointly used as conditional variable inputs of the conditional generation adversarial network model, so that the conditional generation adversarial network model can not only generate seismic waves conforming to a specific geological environment, but also generate target seismic acceleration time histories of a specified damage force, thereby solving the uncontrollability problem of disaster-causing in earthquake generation, further improving the applicability of the seismic motion data, and providing more reliable technical support for urban seismic resilience evaluation and post-earthquake rapid evaluation.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of earthquake engineering and artificial intelligence, and in particular to a method for generating ground motion data based on physics guidance and matching it with building earthquake damage. Background Technology

[0002] Seismic ground motion records are fundamental input data for structural seismic analysis, design, and earthquake risk assessment. With the increasing prevalence of performance-based seismic design, the engineering community's demand for seismic ground motion data is becoming increasingly sophisticated. It is not only necessary for seismic ground motions to conform to specific seismological parameters (such as magnitude, epicentral distance, and site conditions), but also urgently needed to reflect the destructive potential of seismic ground motions on specific structures. Currently, acquiring seismic ground motion data mainly faces the following challenges: ① Scarcity and imbalance of strong earthquake records; Despite the rapid development of global strong earthquake observation networks, records of strong earthquakes that cause severe structural damage or collapse under specific site conditions remain extremely scarce. Existing databases contain mostly small and medium-sized earthquake records, leading to severe "long-tail distribution" and sample imbalance problems when training deep learning earthquake damage prediction models or conducting structural vulnerability analysis. ② Limitations of traditional generation methods; To supplement data, traditional methods typically employ physical simulations (such as focal mechanism simulations) or stochastic process models (such as the Kanai-Tajimi spectrum). These methods are often computationally complex and struggle to capture the complex high-frequency components and non-stationary characteristics of real seismic ground motions, resulting in limitations in the generated seismic ground motion data. ③ The uncontrollability of deep learning generation; In recent years, Generative Adversarial Networks (GANs) have been widely used to generate artificial earthquakes. Although GANs can generate realistic waveforms, most current models are "data-driven," focusing only on the statistical similarity of waveforms or spectra. They typically cannot perceive the physical impact of earthquakes on structures, i.e., they cannot generate specific earthquakes that "collapse houses" or "leave houses intact." Therefore, how to generate artificial earthquakes that both conform to the statistical characteristics of real earthquakes and can be precisely controlled to determine their destructive impact on specific structures is a key problem that urgently needs to be solved in the current intersection of earthquake engineering and artificial intelligence. Summary of the Invention

[0003] In view of this, the object of the present invention is to provide a physical-guided method for generating seismic ground motion data that matches building earthquake damage, so as to alleviate at least some of the above-mentioned technical problems.

[0004] In a first aspect, embodiments of the present invention provide a method for generating ground motion data based on physical guidance and matching with building seismic damage. The method includes: acquiring target earthquake parameters and target seismic damage label information; wherein, the target earthquake parameters include the magnitude, epicentral distance, and site parameters of the target earthquake, and the target seismic damage label information is used to characterize the seismic damage level information of the target earthquake; inputting the target earthquake parameters and target seismic damage label information into a pre-trained conditional generative adversarial network model, so that the conditional generative adversarial network model outputs the target ground motion acceleration time history of the target earthquake; wherein, the conditional generative adversarial network model includes a generator, a discriminator, and a physical performance evaluator, and the physical performance evaluator is a pre-trained building seismic damage classification proxy model.

[0005] Optionally, the method further includes: constructing a building earthquake damage database; wherein the building earthquake damage database includes: multiple earthquake parameters, multiple ground motion acceleration time histories, and multiple earthquake damage label information; training a deep learning classification model based on the building earthquake damage database to obtain a building earthquake damage classification proxy model; and training an established conditional generative adversarial network model based on the building earthquake damage database to obtain a trained conditional generative adversarial network model.

[0006] Optionally, a deep learning classification model is trained based on a building earthquake damage database to obtain a building earthquake damage classification proxy model, including: converting multiple ground motion acceleration time histories into wavelet time-frequency matrices based on wavelet transform; using the wavelet time-frequency matrices as input to the deep learning classification model and multiple earthquake damage label information as output to the deep learning classification model, and training the deep learning classification model to obtain the building earthquake damage classification proxy model.

[0007] Optionally, the established conditional generative adversarial network (GAN) model is trained based on a building seismic damage database to obtain a trained GAN model. This includes: inputting multiple seismic parameters and multiple seismic damage label information into the generator to make the generator output ground motion acceleration time history; until the generator's total loss function converges to obtain the trained GAN model; wherein, the total loss function includes a generative adversarial loss function and a physical mechanism loss function, the physical mechanism loss function is generated by evaluating the wavelet time-frequency matrix of the ground motion acceleration time history output by the generator based on a physical performance evaluator, and the generative adversarial loss function is generated by discriminating the ground motion acceleration time history output by the generator based on a discriminator.

[0008] Optionally, a building seismic damage database is constructed, including: acquiring multiple strong ground motion record data; wherein the strong ground motion record data includes: earthquake parameters and ground acceleration data, and the earthquake parameters include: magnitude, epicentral distance, and site parameters of the strong earthquake; preprocessing the ground acceleration data to obtain the ground acceleration data time history; calculating the seismic response index of the target building under the action of each strong ground motion record data through nonlinear time history analysis; classifying the seismic response index into seismic damage levels and generating corresponding seismic damage label information; and constructing a building seismic damage database based on the earthquake parameters, ground acceleration time history, and corresponding seismic damage label information.

[0009] Optionally, the seismic response index includes the maximum inter-story drift angle. The seismic response index is classified into seismic damage levels, and corresponding seismic damage label information is generated. This includes: classifying the seismic damage level into multiple seismic damage levels based on the maximum inter-story drift angle and multiple preset intervals, and generating corresponding seismic damage label information for each seismic damage level.

[0010] Optionally, the ground motion acceleration data is preprocessed to generate a ground motion acceleration data time history, including: unifying the duration of all ground motion acceleration data corresponding to the same seismic parameter to a fixed length, and sampling the data using a fixed frequency to obtain the ground motion acceleration data time history corresponding to the seismic parameter.

[0011] Secondly, embodiments of the present invention also provide a physical-guided seismic ground motion data generation system for matching building earthquake damage, the system comprising: The acquisition module is used to acquire target earthquake parameters and target seismic damage label information. The target earthquake parameters include the magnitude, epicentral distance, and site parameters of the target earthquake, while the target seismic damage label information is used to characterize the seismic damage level of the target earthquake. The calculation module is used to input the target earthquake parameters and target earthquake damage label information into a pre-trained conditional generative adversarial network model, so that the conditional generative adversarial network model outputs the target ground motion acceleration time history of the target earthquake; wherein, the conditional generative adversarial network model includes a generator, a discriminator and a physical performance evaluator, and the physical performance evaluator is a pre-trained building earthquake damage classification proxy model.

[0012] Thirdly, embodiments of the present invention also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the physical-guided method for generating seismic ground motion data matching building earthquake damage described in the first aspect.

[0013] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the physical-guided seismic ground motion data generation method for matching building earthquake damage described in the first aspect.

[0014] The embodiments of the present invention bring the following beneficial effects: This invention provides a physics-guided method for generating seismic ground motion data matched with building damage. The method inputs target earthquake parameters and target damage label information into a pre-trained conditional generative adversarial network (CGN) model, which outputs the target earthquake's ground motion acceleration time history. The CGN model includes a generator and a discriminator, and also embeds a pre-trained building damage classification proxy model as a physical performance evaluator. This achieves a physics-guided CGN model, solving the physical consistency problem in the ground motion data generation process and improving the accuracy of the generated data. Furthermore, by inputting the target earthquake parameters and target damage label information as conditional variables into the CGN model, the model can generate not only seismic waves consistent with specific geological environments but also target ground motion acceleration time histories with specified destructive forces. This addresses the uncontrollable disaster-causing nature of generated ground motions, further improving the applicability of the ground motion data and providing more reliable technical support for urban seismic resilience assessment and rapid post-earthquake assessment.

[0015] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a physical-guided method for generating seismic ground motion data that matches building earthquake damage, provided as an embodiment of the present invention; Figure 2A comparative schematic diagram of seismic acceleration time histories provided for an embodiment of the present invention; Figure 3 A schematic diagram of a set of normalized average FAS results provided for an embodiment of the present invention; Figure 4 This is a schematic diagram of another set of normalized average FAS results provided in an embodiment of the present invention; Figure 5 A schematic diagram of a set of maximum inter-story drift angle results provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of another set of maximum inter-story drift angle results provided in an embodiment of the present invention; Figure 7 A schematic diagram of a physical-guided seismic ground motion data generation system that matches building earthquake damage, provided as an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] To facilitate understanding of this embodiment, the embodiments of the present invention will be described in detail below.

[0021] This invention provides a method for generating seismic ground motion data based on physical guidance and matching it with building earthquake damage, such as... Figure 1 As shown, the method includes the following steps: Step S102: Obtain target earthquake parameters and target earthquake damage label information.

[0022] Specifically, for a target earthquake that is actually observed, the target earthquake parameters are obtained, and the corresponding target hazard label information is determined based on the target earthquake parameters. For example, the target hazard label information is obtained based on the target earthquake parameters and urban seismic elastoplastic analysis. The target earthquake parameters include the magnitude, epicentral distance, and site parameters of the target earthquake, while the target hazard label information is used to characterize the hazard level of the target earthquake. Therefore, by using both the target earthquake parameters and the target hazard label information as input conditional variables, not only can seismic waves conforming to a specific geological environment be generated, but also the target ground motion acceleration time history with a specified destructive force can be generated, thus solving the problem of uncontrollable disaster-causing effects of ground motion generation.

[0023] Step S104: Input the target earthquake parameters and target earthquake damage label information into the pre-trained conditional generative adversarial network model so that the conditional generative adversarial network model outputs the target ground motion acceleration time history of the target earthquake.

[0024] In practical applications, methods based on physical simulation and deep learning, such as standard GAN (Generative Adversarial Networks), WGAN (Wasserstein GAN), and cGAN (Conditional Generative Adversarial Network), exist in the field of artificial earthquake motion generation. For example, existing cGANs typically employ an architecture that includes a generator and a discriminator. These methods incorporate magnitude, epicentral distance, and site parameters (such as the average shear wave velocity of the soil layer within 30m below the surface). V S30 Seismic parameters such as seismic parameters are input into the generator as conditional variables, so that the generator attempts to generate ground motion acceleration time histories that conform to the characteristics of seismic parameters; at the same time, the discriminator judges whether the generated seismic waveform is consistent with the real seismic wave and whether it conforms to the input seismic parameters.

[0025] However, this method cannot achieve targeted generation of seismic motion data for disaster induction on demand. Since most existing deep learning models use seismic parameters such as magnitude and epicentral distance as input, and there is a significant discrepancy between seismic parameters and structural damage, the same magnitude and epicentral distance can produce completely different structural damage consequences. Because existing schemes cannot directly accept damage levels such as "minor damage" or "collapse" as input instructions, they cannot meet the requirements for seismic waves specifically causing the "collapse" of certain types of structures. Therefore, they can only generate a large amount of data and then filter it, thus reducing the efficiency of seismic motion data generation.

[0026] Furthermore, existing deep learning models typically only focus on whether the mathematical and statistical characteristics of earthquake waveforms (such as amplitude distribution and spectral shape) are realistic. This results in deep learning models being completely unaware of the physical response (such as inter-story drift angle) that the generated earthquake waveforms will produce when they strike the structure during the training process. This "physical blind spot" means that the generated ground motion data may look very much like earthquake waves, but may not conform to the real disaster-causing laws in terms of dynamic characteristics. In other words, the lack of physical mechanism constraints and feedback leads to a lack of credibility of the generated samples in engineering analysis.

[0027] Based on this, this invention provides a novel conditional generative adversarial network (GAN) model. The GAN model includes a generator, a discriminator, and a physical performance evaluator. The physical performance evaluator is a pre-trained building seismic damage classification proxy model. In other words, besides including the generator and discriminator (i.e., the existing cGAN architecture), the GAN model also embeds a pre-trained building seismic damage classification proxy model as the physical performance evaluator, thus realizing a physics-guided GAN model. This solves the physical consistency problem in the generation of ground motion data and improves the accuracy of ground motion data generation.

[0028] Furthermore, in this embodiment of the invention, the target earthquake parameters and target earthquake damage label information are used together as conditional variables to input the conditional generative adversarial network model. This enables the conditional generative adversarial network model to not only generate seismic waves that conform to specific geological environments, but also to generate target ground motion acceleration time histories with specified destructive forces (such as explicitly specifying the generation of ground motion data that "leads to structural collapse" or "leads to structural integrity"). This achieves a leap from "generating earthquake-like waves" to "generating waves with specific disaster-causing capabilities," solving the problem of uncontrollable disaster-causing effects of ground motion generation, further improving the applicability of ground motion data, and providing more reliable technical support for urban seismic resilience assessment and rapid post-earthquake assessment.

[0029] In one embodiment, the method further includes: constructing a building earthquake damage database; wherein the building earthquake damage database includes: multiple earthquake parameters, multiple ground motion acceleration time histories, and multiple earthquake damage label information; training a deep learning classification model based on the building earthquake damage database to obtain a building earthquake damage classification proxy model; and training an established conditional generative adversarial network model based on the building earthquake damage database to obtain a trained conditional generative adversarial network model.

[0030] The building earthquake damage classification proxy model is trained based on a deep learning classification model. Specifically, the process of training the deep learning classification model based on the building earthquake damage database to obtain the building earthquake damage classification proxy model is as follows: multiple ground motion acceleration time histories are converted into wavelet time-frequency matrices using wavelet transform; the wavelet time-frequency matrices are used as input to the deep learning classification model, and multiple earthquake damage label information are used as output to the deep learning classification model. The deep learning classification model is then trained under supervision until convergence, resulting in a physical-guided building earthquake damage classification proxy model. It should be noted that the preferred wavelet transform here is Complex Gaussian Wavelets8 (cgau8) for video analysis of multiple ground motion acceleration time histories to obtain the corresponding wavelet time-frequency matrices. Specific wavelet transforms can be found in existing technologies, and will not be described in detail here.

[0031] Therefore, a building earthquake damage classification proxy model is obtained through wavelet transform and deep learning classification model. Here, the deep learning classification model includes, but is not limited to, convolutional neural networks (CNNs). This building earthquake damage classification proxy model establishes a fast mapping relationship between the time-frequency characteristics of ground motion and the building earthquake damage label information. It is also embedded into the conditional generative adversarial network model as a physical performance evaluator, which can play a physical guidance role. This not only alleviates the problems of high computational cost and inability to directly participate in the backpropagation of neural networks in traditional elastoplastic time history analysis, but also solves the physical consistency problem in the process of ground motion data generation, thus improving the accuracy of ground motion data generation.

[0032] Furthermore, the established conditional generative adversarial network (GAN) model is trained based on the building seismic damage database to obtain a trained GAN model. Specifically, the process includes: inputting multiple seismic parameters and multiple seismic damage label information into the generator to output ground motion acceleration time histories; until the generator's total loss function converges, the trained GAN model is obtained. The total loss function includes a generative adversarial loss function and a physical mechanism loss function. The physical mechanism loss function is generated by evaluating the wavelet time-frequency matrix of the ground motion acceleration time histories output by the generator using a physical performance evaluator. The generative adversarial loss function is generated by discriminating the ground motion acceleration time histories output by the generator using a discriminator.

[0033] Specifically, for the established conditional generative adversarial network (GAN) model, a framework such as WCGAN (Wasserstein Conditional GAN) can be adopted. Based on the generator and discriminator, a pre-trained building seismic damage classification proxy model is embedded as a physical performance evaluator. During training, multiple seismic parameters and multiple seismic damage label information are input into the generator as conditional variables. In some scenarios, random noise vectors are also input into the generator. Simultaneously, the physical performance evaluator is used to evaluate the seismic acceleration time history output by the generator, calculating the physical mechanism loss. Furthermore, the generator parameters are jointly optimized by combining the generative adversarial loss of the discriminator until the generator's total loss function converges, resulting in a well-trained generator.

[0034] It should be noted that the generator includes a fully connected network module, a two-dimensional convolutional neural network generation module, and subsequent convolutional modules. The generator first combines the input conditional variables with a random noise vector and increases the dimension through the fully connected network module. Then, it generates the seismic acceleration time history and the absolute peak acceleration scaling factor through convolution and pooling operations, and outputs the seismic acceleration time history.

[0035] The total loss function is obtained by weighted summation of the generative adversarial loss function and the physical mechanism loss function; its expression is as follows: (1) in, L G Represents the total loss function. L G-GM This represents the generative adversarial loss function. L G-PHY Represents the physical mechanism loss function. This indicates the generation of adversarial loss weights. This represents the weight of the physical mechanism loss.

[0036] In addition, the generative adversarial loss function L G-GM The expression is as follows: (2) in, Represents the mathematical expectation. This indicates that the discriminator is responsible for the generated data. G (Z) score.

[0037] And, for the physical mechanism loss function L G-PHY The calculation process is as follows: First, the seismic acceleration time history output by the generator is subjected to wavelet transform to obtain the corresponding wavelet time-frequency matrix. Then, the wavelet time-frequency matrix is ​​input into the physical performance evaluator so that the physical performance evaluator outputs the prediction result, which is the predicted seismic damage label probability distribution information or feature information. Finally, the difference between the prediction result and the seismic damage label information input to the generator is used as the physical mechanism loss function. L G-PHY .

[0038] Furthermore, the discriminator's input includes: seismic ground acceleration time histories, seismic parameters, and seismic damage label information from the building damage database, and / or, seismic ground acceleration time histories generated by the generator. The discriminator's loss function uses the Wasserstein distance and its gradient penalty term, specifically expressed as follows: (3) in, This represents the loss function of the discriminator. Represents the mathematical expectation. This indicates that the discriminator is responsible for the generated data. G The score of (Z), D(x) represents the score of the discriminator on the real data x, and λ represents the regularization coefficient of the gradient penalty term. Represents the gradient of the discriminator L 2-norm, To pass through the real data distribution With the generated distribution P g The distribution is obtained by uniform sampling along the straight line.

[0039] It should be noted that, for mathematical expectation ,in, This represents a vector randomly sampled from the noise distribution. This represents the prior distribution of noise; similarly, for , x This means randomly sampling a sample from the real dataset. Represents the distribution of real data; for , This indicates a mixed distribution (in the distribution of the real data). And generating data distribution, i.e. generating distribution P g The sample obtained by interpolation between the samples is the sample obtained from sampling.

[0040] Therefore, by embedding a physical performance evaluator as a supervision module in the conditional generative adversarial network (GAN) model, the lack of physical feedback in traditional GAN ​​models is addressed. This allows the physical performance evaluator to provide physical constraint gradients to the generator during the training loop, feeding back errors generated by structural response (maximum inter-story drift angle) to the generator, forcing it to learn the intrinsic mapping relationship between waveform features and structural damage. Simultaneously, the generator's total loss function is obtained by weighting the discriminator's generative adversarial loss function (data distribution consistency) and the physical mechanism loss function (earthquake damage label information consistency) of the physical performance evaluator. Optimal weighting ratios are determined through parameter analysis, such as a physical mechanism loss weight of 1.4 and a generative adversarial loss weight of 1. This ensures that the generated seismic acceleration time histories are not only realistic and consistent with seismic waves in specific geological environments, but also accurately induce disasters, i.e., generating target seismic acceleration time histories with specified destructive forces. This solves the problem of uncontrollable disaster-causing effects in seismic motion generation, further improving seismic data and providing more reliable technical support for urban seismic resilience assessment and rapid post-earthquake assessment.

[0041] In one implementation, constructing a building seismic damage database includes: acquiring multiple strong ground motion record data; wherein the strong ground motion record data includes: earthquake parameters and ground acceleration data, the earthquake parameters including: magnitude, epicentral distance, and site parameters of the strong earthquake; preprocessing the ground acceleration data to obtain the ground acceleration data time history; calculating the seismic response index of the target building under the action of each strong ground motion record data through nonlinear time history analysis; classifying the seismic response index into seismic damage levels and generating corresponding seismic damage label information; and constructing a building seismic damage database based on the earthquake parameters, ground acceleration time history, and corresponding seismic damage label information.

[0042] The seismic response index includes the Maximum Inter-story Drift Ratio (MIDR). The seismic response index is used to classify seismic damage levels and generate corresponding damage labels. This includes dividing the damage severity into multiple damage levels based on the MIDR and multiple preset intervals, and generating corresponding damage labels for each damage level (i.e., using numerical labels to identify each damage level). For example, the preset intervals include first interval, second interval, third interval, etc., and the damage levels include first level, second level, third level, etc. If the MIDR is in the first interval, the damage level is first level, indicating severe building damage; similarly, if the MIDR is in the second interval, the damage level is second level, indicating building collapse; if the MIDR is in the third interval, the damage level is third level, indicating minor building damage. The specific number of preset intervals and damage levels can be adaptively adjusted according to actual conditions.

[0043] Furthermore, the seismic acceleration data is preprocessed to generate a seismic acceleration data time history. This includes: firstly, unifying the duration of all seismic acceleration data to a fixed length, such as by truncating or padding with zeros to a fixed length; for example, seismic acceleration data is extracted from segments 30 seconds before and after the PGA time. For segments shorter than 30 seconds, zeros are added to increase the data points, resulting in a 60-second seismic acceleration data duration. Then, sampling is performed at a fixed frequency to obtain the seismic acceleration data time history. This preprocessing ensures the accuracy of the seismic acceleration data time history, further guaranteeing the training accuracy of the subsequent building seismic damage classification proxy model and conditional generative adversarial network model, thereby improving the accuracy of seismic data generation.

[0044] In practical applications, records of strong earthquakes that cause severe structural damage or collapse are extremely rare (i.e., the long-tail distribution problem of data). Existing generation techniques, if relying solely on learning the distribution of real data, tend to generate a large number of common small and medium-sized earthquake records (head data), while struggling to generate rare, highly destructive extreme samples. This severe scarcity of extreme earthquake damage samples leads to undertraining of deep learning-based earthquake damage prediction models when faced with large earthquakes, resulting in a significant drop in prediction accuracy. Therefore, this invention generates a large amount of virtual artificial ground motion data with accurate high-damage labels in a building earthquake damage database. This expands the "tail data" in the training set (i.e., the building earthquake damage database), alleviating the problems of data imbalance and scarcity of strong earthquake records in earthquake damage prediction model training, thereby improving the generalization ability and accuracy of conditional generative adversarial network models under extreme conditions.

[0045] Furthermore, for the constructed building seismic damage database, the generated high-damage samples were used to specifically enhance the training of the conditional generative adversarial network (GAN) model. This significantly improved the model's accuracy in identifying severe damage and collapse states, enhancing the application effectiveness of artificial intelligence in disaster prevention and providing more reliable technical support for urban seismic resilience assessment and rapid post-earthquake assessment. Additionally, for the generated GAN model, the acquired target earthquake parameters and target seismic damage label information were input into the generator, causing the generator to output ground motion data (i.e., target ground motion acceleration time history) that matches the seismic damage level corresponding to the target seismic damage label information. This allows for the evaluation of structural seismic performance and post-earthquake damage analysis based on the target ground motion acceleration time history. The target ground motion acceleration time history is also stored in the building seismic damage database, expanding the database used as training data and further improving the generation accuracy of the GAN model.

[0046] For example, taking a four-story reinforced concrete frame structure as an example, the beneficial effects of the above-mentioned physical-guided method for generating seismic ground motion data matching building seismic damage are illustrated. Specifically: (1) The accuracy of earthquake damage matching is significantly improved. Compared with the traditional generative model without a physical performance evaluator, the conditional generative adversarial network model with a physical performance evaluator in this invention has significantly improved the consistency ratio between the generated target ground motion acceleration time history and the target earthquake damage label information from 55.7% to 81.5%, thus proving that the physical guidance mechanism can effectively control the destructive force of the generated ground motion data on specific structures.

[0047] (2) The physical authenticity of the generated ground motion data is relatively high.

[0048] like Figure 2 As shown, the horizontal axis of each graph represents time t, in seconds; the vertical axis represents the time history of ground motion acceleration, in m / s².2 For (2A), the magnitude is 6.0. M w =6.0, epicenter distance EpiD=12.24km, site parameters V S30 =314.7 m / s, label=4. For the same earthquake, the results include four graphs. The top graph shows the observed ground acceleration time history, while the bottom three graphs show the ground acceleration time histories generated by the conditional generative adversarial network model provided in this embodiment of the invention. Specifically, building damage is classified according to the maximum inter-story drift angle, and the classification result is used as the label. For example, label=4 indicates that the maximum inter-story drift angle corresponds to level four building damage, where the maximum inter-story drift angle is greater than 0.03345°.

[0049] Similarly, for (2B), the magnitude is 6.5. M w =6.5, epicenter distance EpiD=57.79km, site parameters V S30 =213.44 m / s, label=4. For the same earthquake, the results include four graphs. The top graph shows the observed ground motion acceleration time history, while the bottom three graphs show the ground motion acceleration time histories generated by the conditional generative adversarial network model provided in this embodiment of the invention. For (2C), the magnitude is 7.0. M w =7.0, epicenter distance EpiD=55.51km, site parameters V S30 =211m / s, label=4. At this time, for the same earthquake, the results include four graphs. The top graph is the observed ground motion acceleration time history, and the bottom three graphs are the ground motion acceleration time histories generated by the conditional generative adversarial network model provided in the embodiment of the present invention.

[0050] Therefore, according to such Figure 2 The results show that the waveform of the ground motion acceleration time history generated by the conditional generative adversarial network model provided in this embodiment of the invention is highly consistent with the waveform of the observed ground motion acceleration time history, thereby improving the accuracy of ground motion data generation.

[0051] In addition, when the epicenter is 20km to 60km from EpiD, the site parameters... V S30 The normalized average FAS (Fourier Amplitude Spectra) results of the ground motion data at speeds between 270 m / s and 360 m / s and label=4 are as follows: Figure 3As shown in the figure. In the figure, the horizontal axis of (3A) to (3F) represents the frequency (in Hz), the vertical axis represents the normalized average FAS (in m / s), L1 represents the change of the normalized average FAS of the observed ground motion data with frequency, and L2 represents the change of the normalized average FAS of the ground motion data generated by the conditional generative adversarial network model with frequency.

[0052] For (3A), the magnitude is 5.3~5.5. M w =[5.3, 5.5], Number of ground motion data N obs =19; For (3B), the magnitude is 5.7~5.9. M w =[5.7, 5.9], Number of ground motion data N obs =72; For (3C), the magnitude is 6.1~6.3. M w =[6.1, 6.3], Number of ground motion data N obs =44; For (3D), the magnitude is 6.5~6.7. M w =[6.5, 6.7], Number of ground motion data N obs =52; for (3E), the magnitude is 6.9~7.1. M w =[6.9, 7.1], Number of ground motion data N obs =40; for (3F), the magnitude is 7.1~7.3. M w =[7.1, 7.3], Number of ground motion data N obs =10.

[0053] And, when the epicenter is 60km~100km from EpiD, site parameters V S30 The normalized average FAS results of the ground motion data at speeds between 270 m / s and 360 m / s and label=4 are as follows: Figure 4 As shown in the figure. In the figure, the horizontal axis of (4A) to (4F) represents the frequency (in Hz), the vertical axis represents the normalized average FAS (in m / s), L1 represents the change of the normalized average FAS of the observed ground motion data with frequency, and L2 represents the change of the normalized average FAS of the ground motion data generated by the conditional generative adversarial network model with frequency.

[0054] For (4A), the magnitude is 5.9~6.1. M w =[5.9, 6.1], Number of ground motion data N obs =16; For (4B), the magnitude is 6.1~6.3. M w =[6.1, 6.3], Number of ground motion data N obs =26; For (4C), the magnitude is 6.7~6.9. M w =[6.7, 6.9], Number of ground motion data N obs =19; For (4D), the magnitude is 6.9~7.1. M w =[6.9, 7.1], Number of ground motion data N obs =18; For (4E), the magnitude is 7.3~7.5. M w =[7.3, 7.5], Number of ground motion data N obs =36; For (4F), the magnitude is 7.5~7.7. M w =[7.5, 7.7], Number of ground motion data N obs =17.

[0055] Therefore, according to such Figure 3-4 The results show that the normalized average FAS of the ground motion data generated by the conditional generative adversarial network model provided in this embodiment of the invention is highly consistent with the waveform of the normalized average FAS of the observed ground motion data. Thus, the conditional generative adversarial network model can capture the frequency domain energy distribution of real ground motion better, further improving the accuracy of ground motion data generation.

[0056] Furthermore, for key seismic intensity indicators (IMs) such as Peak Ground Acceleration (PGA), Sa (0.4 s), Cumulative Absolute Velocity (CAV), and Effective Peak Acceleration (EPA), the correlation coefficient (R) between the generated seismic ground motion data and the actual seismic ground motion data is crucial.2 The values ​​all exceeded 0.81, thus proving that the generated seismic motion data possessed real physical statistical characteristics.

[0057] (3) Good structural response reproduction: The generated ground motion data is input into the structural model for calculation. The resulting MIDR is very close to the response under the actual ground motion. Here, the structural model is the seismic response calculation model of the target building. The model uses the urban seismic elastoplastic analysis method to calculate the maximum inter-story drift angle MIDR of the corresponding building.

[0058] When the earthquake damage label information is label 3 (i.e., earthquake damage level is 3), the generated maximum inter-story drift angle (MIDR) is as follows: Figure 5 As shown; where (5A)~(5C) all represent the maximum inter-story drift angle (MIDR) and the vertical axis all represent the floor, S1 all represent the MIDR calculated from the ground motion data generated by the conditional generative adversarial network model, S2 all represent the MIDR of the building under the measured ground motion, and S3 all represent the MIDR calculated from the ground motion data generated by the unguided deep learning model.

[0059] Among them, for (5A), the magnitude is 4.9~5.1. M w =[4.9, 5.1], epicenter distance from EpiD is 20km~60km, site parameters V S30 The speed range is between 270 m / s and 360 m / s, and the label is 3. Similarly, for (5B), the magnitude is 5.1 to 5.3. M w =[5.1, 5.3], epicenter distance from EpiD is 20km~60km, site parameters V S30 The speed range is between 270 m / s and 360 m / s, and the label is 3. For (5C), the magnitude is 5.3 to 5.5. M w =[5.3, 5.5], epicenter distance from EpiD is 20km~60km, site parameters V S30 The speed range is 270m / s to 360m / s and the label is 3.

[0060] Furthermore, when the earthquake damage label information is label 4 (i.e., earthquake damage level 4), the generated maximum inter-story drift angle (MIDR) is as follows: Figure 6As shown; where (6A)~(6C) all represent the maximum inter-story drift angle (MIDR) and the vertical axis all represent the floor, S1 all represent the MIDR calculated from the ground motion data generated by the conditional generative adversarial network model, S2 all represent the MIDR of the building under the measured ground motion, and S3 all represent the MIDR calculated from the ground motion data generated by the unphysical deep learning model.

[0061] For (6A), the magnitude is 6.3~6.5. M w =[6.3, 6.5], epicenter distance from EpiD is 100km~140km, site parameters V S30 The speed range is between 270 m / s and 360 m / s, and the label is 4. Similarly, for (6B), the magnitude is 6.5 to 6.7. M w =[6.5, 6.7], epicenter distance from EpiD is 100km~140km, site parameters V S30 The speed range is between 270 m / s and 360 m / s, and the label is 4. For (6C), the magnitude is 6.7 to 6.9. M w =[6.7, 6.9], epicenter distance from EpiD is 100km~140km, site parameters V S30 The speed range is 270m / s to 360m / s and the label is 4.

[0062] Therefore, according to such Figures 5-6 The results show that the physically guided conditional generative adversarial network model significantly outperforms the unguided deep learning model, thus verifying the reliability of the generated data in engineering applications. It should be noted that label=4 indicates a maximum inter-layer displacement angle greater than 0.03345°, and label=3 indicates a maximum inter-layer displacement angle greater than 0.005° but not greater than 0.03345°.

[0063] (4) This solution addresses the scarcity of extreme earthquake damage data. This method can generate earthquake motion data that cause severe structural damage in batches by specifying "high earthquake damage labels". This effectively solves the problems of few large earthquake records and unbalanced samples in actual observations, and provides high-quality augmented data for structural seismic performance assessment and deep learning earthquake damage prediction models.

[0064] Based on the above method embodiments, this invention also provides a physical-guided seismic ground motion data generation system that matches building seismic damage, such as... Figure 7 As shown, the system includes: The acquisition module 71 is used to acquire target earthquake parameters and target earthquake damage label information; wherein, the target earthquake parameters include the magnitude, epicentral distance and site parameters of the target earthquake, and the target earthquake damage label information is used to characterize the earthquake damage level information of the target earthquake; The calculation module 72 is used to input the target earthquake parameters and target earthquake damage label information into a pre-trained conditional generative adversarial network model so that the conditional generative adversarial network model outputs the target ground motion acceleration time history of the target earthquake; wherein, the conditional generative adversarial network model includes a generator, a discriminator and a physical performance evaluator, and the physical performance evaluator is a pre-trained building earthquake damage classification proxy model.

[0065] The physical-guided seismic ground motion data generation system based on building damage matching provided in this invention uses a conditional generative adversarial network (GAN) model. In addition to a generator and discriminator, the GAN model incorporates a pre-trained building damage classification proxy model as a physical performance evaluator, thus realizing a physical-guided GAN model. This solves the physical consistency problem in the seismic ground motion data generation process and improves the accuracy of the generated data. Furthermore, by using target earthquake parameters and target damage label information as conditional variables input to the GAN model, the system can generate not only seismic waves consistent with specific geological environments but also target ground motion acceleration time histories with specified destructive forces. This addresses the problem of uncontrollable disaster-causing effects of generated seismic ground motions, further improving the applicability of the data and providing more reliable technical support for urban seismic resilience assessment and rapid post-earthquake assessment.

[0066] Optionally, the system further includes: constructing a building earthquake damage database; wherein the building earthquake damage database includes: multiple earthquake parameters, multiple ground motion acceleration time histories, and multiple earthquake damage label information; training a deep learning classification model based on the building earthquake damage database to obtain a building earthquake damage classification proxy model; and training an established conditional generative adversarial network model based on the building earthquake damage database to obtain a trained conditional generative adversarial network model.

[0067] Optionally, a deep learning classification model is trained based on a building earthquake damage database to obtain a building earthquake damage classification proxy model, including: converting multiple ground motion acceleration time histories into wavelet time-frequency matrices based on wavelet transform; using the wavelet time-frequency matrices as input to the deep learning classification model and multiple earthquake damage label information as output to the deep learning classification model, and training the deep learning classification model to obtain the building earthquake damage classification proxy model.

[0068] Optionally, the established conditional generative adversarial network (GAN) model is trained based on a building seismic damage database to obtain a trained GAN model. This includes: inputting multiple seismic parameters and multiple seismic damage label information into the generator to make the generator output ground motion acceleration time history; until the generator's total loss function converges to obtain the trained GAN model; wherein, the total loss function includes a generative adversarial loss function and a physical mechanism loss function, the physical mechanism loss function is generated by evaluating the wavelet time-frequency matrix of the ground motion acceleration time history output by the generator based on a physical performance evaluator, and the generative adversarial loss function is generated by discriminating the ground motion acceleration time history output by the generator based on a discriminator.

[0069] Optionally, a building seismic damage database is constructed, including: acquiring multiple strong ground motion record data; wherein the strong ground motion record data includes: earthquake parameters and ground acceleration data, and the earthquake parameters include: magnitude, epicentral distance, and site parameters of the strong earthquake; preprocessing the ground acceleration data to obtain the ground acceleration data time history; calculating the seismic response index of the target building under the action of each strong ground motion record data through nonlinear time history analysis; classifying the seismic response index into seismic damage levels and generating corresponding seismic damage label information; and constructing a building seismic damage database based on the earthquake parameters, ground acceleration time history, and corresponding seismic damage label information.

[0070] Optionally, the seismic response index includes the maximum inter-story drift angle. The seismic response index is classified into seismic damage levels, and corresponding seismic damage label information is generated. This includes: classifying the seismic damage level into multiple seismic damage levels based on the maximum inter-story drift angle and multiple preset intervals, and generating corresponding seismic damage label information for each seismic damage level.

[0071] Optionally, the ground motion acceleration data is preprocessed to generate a ground motion acceleration data time history, including: unifying the duration of all ground motion acceleration data corresponding to the same seismic parameter to a fixed length, and sampling the data using a fixed frequency to obtain the ground motion acceleration data time history corresponding to the seismic parameter.

[0072] The physical-guided seismic ground motion data generation system for matching building earthquake damage provided in this embodiment of the invention has the same technical features as the physical-guided seismic ground motion data generation method for matching building earthquake damage provided in the above embodiments, so it can also solve the same technical problems and achieve the same technical effects.

[0073] This invention also provides an electronic device, including a processor and a memory. The memory stores a computer program that can be executed by the processor, and the processor executes the computer program to implement the above-described method for generating ground motion data that is based on physical guidance and matches building earthquake damage.

[0074] See Figure 8As shown, the electronic device includes a processor 100 and a memory 101. The memory 101 stores a computer program that can be executed by the processor 100. The processor 100 executes the computer program to implement the above-described physical-guided method for generating seismic ground motion data that matches building earthquake damage.

[0075] Furthermore, Figure 8 The electronic device shown also includes a bus 102 and a communication interface 103, with the processor 100, the communication interface 103 and the memory 101 connected via the bus 102.

[0076] The memory 101 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 103 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 102 may be an ISA (Industrial Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Enhanced Industry Standard Architecture) bus, etc. These buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 8 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0077] Processor 100 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 100 or by instructions in software form. Processor 100 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a readily available storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 101, and the processor 100 reads the information from memory 101 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0078] This embodiment also provides a computer-readable storage medium storing a computer program, which is executed by a processor to perform the above-described method for generating ground motion data based on physical guidance and matching building earthquake damage.

[0079] The computer program product of the physical-guided method, system and electronic device for generating seismic ground motion data matching building earthquake damage provided in the embodiments of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0080] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0081] Furthermore, in the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0082] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0083] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0084] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for generating seismic ground motion data based on physical guidance and matching with building seismic damage, characterized in that, The method includes: Obtain target earthquake parameters and target earthquake damage label information; wherein, the target earthquake parameters include the magnitude, epicentral distance and site parameters of the target earthquake, and the target earthquake damage label information is used to characterize the earthquake damage level information of the target earthquake; The target earthquake parameters and the target earthquake damage label information are input into a pre-trained conditional generative adversarial network (GAN) model, so that the GAN model outputs the target ground motion acceleration time history of the target earthquake; wherein, the GAN model includes a generator, a discriminator, and a physical performance evaluator, and the physical performance evaluator is a pre-trained building earthquake damage classification proxy model.

2. The method according to claim 1, characterized in that, The method further includes: Construct a building earthquake damage database; wherein, the building earthquake damage database includes: multiple earthquake parameters, multiple ground motion acceleration time histories, and multiple earthquake damage label information; The deep learning classification model is trained based on the building earthquake damage database to obtain the building earthquake damage classification proxy model; The established conditional generative adversarial network model is trained based on the building earthquake damage database to obtain the trained conditional generative adversarial network model.

3. The method according to claim 2, characterized in that, The process of training a deep learning classification model based on the building earthquake damage database to obtain the building earthquake damage classification proxy model includes: The time histories of multiple ground motion accelerations are converted into wavelet time-frequency matrices based on wavelet transform. The wavelet time-frequency matrix is ​​used as the input to the deep learning classification model, and multiple earthquake damage label information is used as the output of the deep learning classification model. The deep learning classification model is then trained to obtain the building earthquake damage classification proxy model.

4. The method according to claim 2, characterized in that, The process of training the established conditional generative adversarial network model based on the building earthquake damage database to obtain the trained conditional generative adversarial network model includes: Multiple earthquake parameters and multiple earthquake damage label information are input into the generator to make the generator output ground motion acceleration time history; until the total loss function of the generator converges, the trained conditional generative adversarial network model is obtained; wherein, the total loss function includes a generative adversarial loss function and a physical mechanism loss function, the physical mechanism loss function is generated by evaluating the wavelet time-frequency matrix of the ground motion acceleration time history output by the generator according to the physical performance evaluator, and the generative adversarial loss function is generated by discriminating the ground motion acceleration time history output by the generator according to the discriminator.

5. The method according to claim 2, characterized in that, The construction of the building earthquake damage database includes: Acquire multiple strong ground motion record data; wherein, the strong ground motion record data includes: earthquake parameters and ground motion acceleration data, and the earthquake parameters include: magnitude, epicentral distance and site parameters of the strong earthquake; The ground motion acceleration data is preprocessed to obtain the time history of the ground motion acceleration data; The seismic response index of the target building under the action of each of the strong ground motion records was calculated by nonlinear time history analysis. The earthquake response indicators are classified into earthquake damage levels, and corresponding earthquake damage label information is generated; The building earthquake damage database is constructed based on the earthquake parameters, the ground motion acceleration time history, and the corresponding earthquake damage label information.

6. The method according to claim 5, characterized in that, The seismic response index includes the maximum inter-story drift angle. The process of classifying the seismic response index into seismic damage levels and generating corresponding damage label information includes: The degree of earthquake damage is divided into multiple damage levels based on the maximum inter-story drift angle and multiple preset intervals, and corresponding damage label information is generated for each damage level.

7. The method according to claim 5, characterized in that, The preprocessing of the seismic acceleration data to generate the seismic acceleration data time history includes: The duration of all ground motion acceleration data corresponding to the same earthquake parameter is unified to a fixed length, and sampling is performed using a fixed frequency to obtain the time history of the ground motion acceleration data corresponding to the earthquake parameter.

8. A physical-guided seismic ground motion data generation system for matching building earthquake damage, characterized in that, The system includes: The acquisition module is used to acquire target earthquake parameters and target earthquake damage label information; wherein, the target earthquake parameters include the magnitude, epicentral distance and site parameters of the target earthquake, and the target earthquake damage label information is used to characterize the earthquake damage level information of the target earthquake; The calculation module is used to input the target earthquake parameters and the target earthquake damage label information into a pre-trained conditional generative adversarial network model, so that the conditional generative adversarial network model outputs the target ground motion acceleration time history of the target earthquake; wherein, the conditional generative adversarial network model includes a generator, a discriminator and a physical performance evaluator, and the physical performance evaluator is a pre-trained building earthquake damage classification proxy model.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the physical-guided method for generating seismic ground motion data that matches building earthquake damage as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the physical-guided ground motion data generation method for matching building earthquake damage as described in any one of claims 1-7.

Images