Machine learning based unrecorded site ground motion simulation method and device

Abstract

The application discloses a kind of based on machine learning's recordless site seismic simulation method and equipment, method includes: by the measured seismic record of several stations near target site under earthquake event, the pseudo-acceleration response spectrum of target site is predicted;With the pseudo-acceleration response spectrum as target spectrum, the measured seismic acceleration time history of station near target site is fitted, and the seismic acceleration time history of target site is obtained;The seismic acceleration time history obtained by fitting is transformed, and the amplitude spectrum gray scale diagram is obtained, and it is input into the generator of trained generative adversarial network with random noise, and the pseudo-amplitude spectrum image with similar characteristics of amplitude spectrum gray scale diagram is generated;Reconstruct the phase spectrum image corresponding to pseudo-amplitude spectrum image;Based on pseudo-amplitude spectrum image and reconstructed phase spectrum image, the one-dimensional seismic acceleration time history of target site is obtained.The application can realize the seismic simulation data generation of any recordless target site.

Description

Technical Field

[0001] This invention relates to seismic motion simulation technology, and more particularly to a method and apparatus for simulating seismic motion in unrecorded sites based on machine learning. Background Technology

[0002] Earthquake monitoring instruments, as core tools for recording seismic wave characteristics and analyzing earthquake causes, are a crucial technical support for modern earthquake prevention and disaster reduction systems. According to the latest statistics, the total number of earthquake monitoring stations worldwide has exceeded ten thousand, mainly distributed in seismically active areas such as the Circum-Pacific Seismic Belt and the Eurasian Seismic Belt. However, the existing monitoring network suffers from significant uneven distribution: developed countries have a high station density in seismically active areas, while large monitoring gaps remain in the ocean areas covering 71% of the Earth's surface, remote mountainous regions, and developing countries.

[0003] In the field of earthquake engineering, seismic acceleration time histories are key input parameters for assessing the seismic response of structures. Despite continuous improvements in global earthquake monitoring capabilities, the existing network of stations still struggles to effectively cover all seismic risk areas. For building structures lacking direct seismic observation records, post-earthquake performance assessment and nonlinear seismic response analysis face the challenge of insufficient foundational data. Therefore, generating reliable seismic acceleration time histories for sites without direct seismic observation records has become a pressing problem to be solved.

[0004] Current mainstream methods for simulating seismic ground motion time histories can be divided into two categories: The first category is numerical simulation methods based on physical mechanisms, which simulate the propagation process of seismic waves by establishing refined models (such as fault models, velocity structure models, etc.) that include source characteristics, propagation paths, and site effects. Although this method can accurately reflect wave characteristics, it has stringent requirements for input data and is computationally expensive. The second category is coherence function methods based on statistical properties, which construct spatial relationship models of seismic ground motions using cross-power spectral density functions and auto-power spectral density functions. Although this method simplifies the calculation process, it still requires complete site parameters and propagation characteristic data, and has significant limitations in practical applications.

[0005] With the development of artificial intelligence technology, machine learning has demonstrated unique advantages in the field of seismic motion simulation. Compared to traditional methods, data-driven machine learning techniques can break free from theoretical constraints on specific coherence functions or power spectra, and establish predictive models by autonomously mining data features. Furthermore, machine learning methods can efficiently capture the spatial variability of seismic motions, maintain the physical rationality of simulation results, and significantly improve computational efficiency, supporting the rapid generation of large-scale seismic motion time histories. However, existing methods still face challenges under complex geological conditions: when the spacing between stations is large, leading to significant differences in soil layers, the propagation characteristics of seismic waves will change significantly, limiting the applicability of traditional methods. Currently, there are related patents for near-fault seismic motion fitting, which can simulate artificial seismic motions that fit the target response spectrum. However, given the known measured seismic motion records for some sites, there is still no reliable method for artificially simulating seismic motion records at arbitrary locations under a given seismic event.

[0006] The published patent CN 114442153 A discloses a method for fitting near-fault ground motion. This method artificially simulates near-fault ground motion that matches the response spectrum. However, it requires prior knowledge of the response spectrum to fit the motion. For sites without any records, it is still impossible to simulate ground motion. Summary of the Invention

[0007] To address the problems existing in the prior art, the purpose of this invention is to provide a machine learning-based method and device for simulating ground motion at a target site without recording data, based on seismic information from nearby stations.

[0008] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0009] A machine learning-based method for simulating ground motion in unrecorded sites includes the following steps:

[0010] (1) The pseudo-acceleration response spectrum of the target site is predicted by using the actual ground motion records of several stations near the target site under the earthquake event;

[0011] (2) Using the pseudo-acceleration response spectrum as the target spectrum, the measured ground motion acceleration time history of the stations near the target site is fitted to obtain the ground motion acceleration time history of the target site.

[0012] (3) Transform the fitted earthquake acceleration time history to obtain an amplitude spectrum grayscale image, and input the random noise into the generator of the trained generative adversarial network to generate a pseudo amplitude spectrum image with similar features to the amplitude spectrum grayscale image; wherein, the generative adversarial network includes a generator and a discriminator. During training, the generator is used to generate a new pseudo amplitude spectrum image based on random noise, and the discriminator is used to determine the similarity between the pseudo amplitude spectrum image generated by the generator and the amplitude spectrum grayscale image, and backpropagate according to the similarity to make the pseudo amplitude spectrum image generated by the generator close to the amplitude spectrum grayscale image;

[0013] (4) Process the generated pseudo amplitude spectrum image to reconstruct the phase spectrum image corresponding to the pseudo amplitude spectrum image;

[0014] (5) Combine the generated pseudo amplitude spectrum image and the reconstructed phase spectrum image to perform inverse short-time Fourier transform to obtain the one-dimensional ground motion acceleration time history of the target site.

[0015] Furthermore, step (1) specifically includes:

[0016] (1.1) Obtain the measured ground motion information from several stations near the target site, including the normalized station location coordinates and the average shear wave velocity of the top 30m soil layer of the site where the station is located. The logarithm and the corresponding pseudo-acceleration response spectrum;

[0017] (1.2) The Gaussian process regression model is established as follows:

[0018]

[0019] In the formula, This represents the pseudo-acceleration response spectrum. Represents the Gaussian process function. The covariance function representing the Gaussian process is obtained through training. This indicates the actual ground motion information measured by the station. This represents the normalized x, y, and z coordinates of the station location. Indicates station The logarithm of ; This indicates the seismic motion information of the target site. This represents the normalized x, y, and z coordinates of the target site location. Indicates the target site The logarithm of ;

[0020] (1.3) Input the measured ground motion information from several stations into the Gaussian process regression model for training to obtain the trained Gaussian process regression model;

[0021] (1.4) Input the ground motion information of the target site into the Gaussian process regression model to obtain the pseudo-acceleration response spectrum of the target site.

[0022] Furthermore, step (2) specifically includes:

[0023] (2.1) Obtain the time history of measured ground motion from several stations near the target site;

[0024] (2.2) A wavelet function is superimposed on the measured ground motion time history at the station to obtain a new ground motion time history, so that the pseudo-acceleration response spectrum corresponding to the new ground motion time history matches the pseudo-acceleration response spectrum of the target site. The new ground motion time history is the ground motion acceleration time history of the target site.

[0025] Furthermore, the calculation process for the amplitude spectrum grayscale image is as follows:

[0026] A short-time Fourier transform is performed on the fitted seismic acceleration time history. By adjusting the window length and overlap length, an amplitude spectrum image of a preset size is obtained, and then normalized to obtain a grayscale image of the amplitude spectrum.

[0027] Furthermore, the generator includes an input layer, a fully connected layer, two upsampling layers, and four convolutional layers connected in sequence. A batch normalization operation and a LeakyReLU activation function are provided between the third and fourth convolutional layers, and a Tanh activation function is provided after the fourth convolutional layer.

[0028] Furthermore, the discrimination includes a sequentially connected input layer, four convolutional layers, one flattening layer, and one fully connected layer. Each convolutional layer is followed by batch normalization, LeakyReLU activation function, and random deactivation processing.

[0029] Furthermore, step (4) specifically includes:

[0030] (4.1) Perform inverse normalization on the generated pseudo amplitude spectrum image to obtain the amplitude spectrum color image;

[0031] (4.2) The phase spectrum image corresponding to the amplitude spectrum color image is reconstructed using the Griffin-Lim algorithm.

[0032] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described above.

[0033] A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the above-described method.

[0034] A computer program product includes a computer program / instructions that, when executed by a processor, implement the above-described method.

[0035] Compared with the prior art, the beneficial effects of this invention are:

[0036] (1) Based on the hybrid machine learning framework of Gaussian process regression and deep convolutional generative adversarial network, this application conducts innovative research on the problem of ground motion simulation in unrecorded sites and proposes a two-stage simulation strategy of "response spectrum guidance-time history generation". While ensuring the physical rationality of ground motion, it significantly improves the computational efficiency and is applicable to ground motion prediction under any geological conditions. While improving the simulation accuracy, it reduces the dependence on prior data and enhances the adaptability to complex site conditions, providing a brand-new technical path for data-driven simulation in the field of earthquake engineering.

[0037] (2) This application constructs a seismic motion generation system based on the synergistic optimization of spectral features and temporal features, breaking through the technical bottlenecks of traditional methods in terms of physical model dependence and statistical constraints. Single methods are limited by their theoretical basis and application scope, making it difficult to take into account the physical characteristics and spatial variability of seismic motions, and they also suffer from problems such as power spectrum distortion and coherence function deviation.

[0038] (3) This application can not only efficiently generate seismic motion time histories that conform to the characteristics of specified earthquake events, but also maintain the spatial correlation and physical consistency of seismic motion, providing reliable data support for seismic analysis of building structures, seismic risk assessment and optimization of seismic design of engineering. Attached Figure Description

[0039] Figure 1 A flowchart illustrating the machine learning-based ground motion simulation method for unrecorded sites provided by this invention.

[0040] Figure 2 The geographical distribution of the stations;

[0041] Figure 3 The pseudo-acceleration response spectrum predicted by the Gaussian process regression model in the east-west direction for station S1;

[0042] Figure 4 To compare the fitted ground motion acceleration time history with the fitted ground motion response spectrum and the predicted response spectrum, (a) shows the ground motion acceleration time history fitted with the superimposed wavelet function, and (b) shows the ground motion response spectrum fitted with the superimposed wavelet function and the predicted response spectrum.

[0043] Figure 5 The images show the amplitude spectrum grayscale images before and after normalization. (a) is the amplitude spectrum image before normalization, and (b) is the amplitude spectrum grayscale image after normalization.

[0044] Figure 6 The generator architecture diagram provided by this invention;

[0045] Figure 7 The discriminator provided for this invention will be patterned;

[0046] Figure 8 (a) represents the iteration error of the Griffin-Lim algorithm and the reconstructed ground motion time history. (b) represents the reconstructed ground motion time history of the Griffin-Lim algorithm with the optimal number of iterations.

[0047] Figure 9 For the comparison of results from station S1, (a) compares the pseudo-acceleration response spectra of the measured east-west ground motion at station S1 with those of 150 artificially simulated ground motions, and (b) compares the peak acceleration and duration of the measured east-west ground motion at station S1 with those of 150 artificially simulated ground motions. Detailed Implementation

[0048] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0049] Example 1

[0050] This invention provides a machine learning-based method for simulating ground motion in unrecorded sites, such as... Figure 1 As shown, it includes the following steps:

[0051] (1) The pseudo-acceleration response spectrum of the target site is predicted by using the actual ground motion records of several stations near the target site under the earthquake event.

[0052] This step specifically includes:

[0053] (1.1) Obtain the measured ground motion information from several stations near the target site, including the normalized station location coordinates and the average shear wave velocity V of the top 30m soil layer of the site where the station is located. s30 The logarithm and the corresponding pseudo-acceleration response spectrum.

[0054] In the experiment, the 2019 Mw 7.1 Ridgecrest earthquake in California, USA, was used as the target seismic event. The measured ground motion data were obtained from 149 community seismic stations within the California Institute of Technology's seismic network. Data from 148 stations were used as the training set, and data from the remaining station was used as the test set. The geographical distribution of the 149 stations is shown below. Figure 2As shown. Due to the large geographical distance between different stations and the significant differences in soil conditions at each station, the input information in the Gaussian process regression model of this invention includes the location coordinates of the station and the average shear wave velocity of the top 30m soil layer at the site where the station is located. Considering the significant differences in soil characteristics at the sites of different stations, in order to reduce... The degree of dispersion during the training process is subjected to a logarithmic transformation to obtain the measured seismic motion information. , This represents the normalized x, y, and z coordinates of the station location. = To avoid the adverse effects of data of different magnitudes on the training process of the Gaussian process regression model, the data for these four dimensions were normalized before training to ensure that the data distribution of each dimension has a mean of zero and a standard deviation of one. The pseudo-acceleration response spectrum is a pseudo-acceleration response spectrum with a 5% damping ratio, containing 85 specific periods between 0.1 seconds and 20 seconds. The pseudo-spectral acceleration values ​​differ significantly between long and short periods, and such different magnitudes of data may cause large biases and instabilities during the training of the machine learning model. Therefore, a logarithmic transformation was performed on the pseudo-spectral acceleration values ​​of measured ground motions from different stations to reduce the numerical differences between different periods. This resulted in the training dataset... ,in The test dataset consists of the four-dimensional site information of the input station and the pseudo-acceleration response spectrum observations of the measured ground motion at that station over 85 specified periods. The test dataset comprises the site information and pseudo-spectral acceleration data of stations S1 and S2. ,in .

[0055] (1.2) The Gaussian process regression model is established as follows:

[0056]

[0057] In the formula, y= This represents the pseudo-acceleration response spectrum. Represents the Gaussian process function. The covariance function representing the Gaussian process is obtained through training. This indicates the actual ground motion information measured by the station. This represents the normalized x, y, and z coordinates of the station location. Indicates station The logarithm of ; This indicates the seismic motion information of the target site. This represents the normalized x, y, and z coordinates of the target site location. Indicates the target site The logarithm of; where,

[0058]

[0059] This represents the pseudo-acceleration response spectrum of the target site. express , express The prior mean vector, Let represent the prior mean vector of the target site. express The covariance matrix, and express and The covariance matrix between them express The covariance matrix.

[0060] (1.3) Input the measured ground motion information from several stations into the Gaussian process regression model for training to obtain the trained Gaussian process regression model;

[0061] training dataset By inputting the Gaussian process regression model for training, we can obtain:

[0062]

[0063]

[0064] Indicates predicted value The variance of the Gaussian distribution.

[0065] (1.4) Input the ground motion information of the target site into the Gaussian process regression model to obtain the pseudo-acceleration response spectrum of the target site;

[0066] Specifically, the pseudo-acceleration response spectrum of the target site can be expressed as:

[0067]

[0068] in, This represents the input to the training set; Represents seismic motion information at the target site; input from a known training dataset. and output Under these conditions, you can input The pseudo-acceleration response spectrum of the target site was obtained. .

[0069] In this example, the response spectrum predicted by the east-west Gaussian process regression model at station S1 differs from the predicted response spectrum of the measured ground motion at the station as follows: Figure 3 As shown.

[0070] (2) Using the pseudo-acceleration response spectrum as the target spectrum, the measured ground motion acceleration time history of the stations near the target site is fitted to obtain the ground motion acceleration time history of the target site.

[0071] Step (2) specifically includes:

[0072] (2.1) Obtain the time history of measured ground motion from several stations near the target site;

[0073] (2.2) A wavelet function is superimposed on the measured ground motion time history at the station to obtain a new ground motion time history, so that the pseudo-acceleration response spectrum corresponding to the new ground motion time history matches the pseudo-acceleration response spectrum of the target site. The new ground motion time history is the ground motion acceleration time history of the target site.

[0074] The wavelet function used is as follows:

[0075]

[0076] in It is frequency Wavelet function of time; It is the moment when the maximum response of the mono-degree system occurs at the j-th frequency in the initial earthquake time history; The time when the wavelet function reaches its maximum value is the time when its response function reaches its maximum value. The time difference between them; n is an integer representing the number of cycles of the sine function, used to adjust the duration of the wavelet function; It is a parameter related to frequency and cycle number. When determining this parameter, it is necessary to ensure that the velocity and displacement time histories of the wavelet function do not exhibit zero-line drift.

[0077] In this example, the measured ground motion acceleration time histories from 10 surrounding stations were adjusted so that the acceleration response spectrum of the adjusted time histories matched the pseudo-acceleration response spectrum of the target site. The fitted ground motion acceleration time histories are as follows: Figure 4 As shown in (a), the duration of the ground motion time history is 200 s, and the time interval is 0.02 seconds, meaning each ground motion time history consists of 10001 data points. The comparison between the response spectrum of the fitted ground motion acceleration time history and the target response spectrum is shown below. Figure 4 As shown in (b).

[0078] This step can fit the same number of ground motion acceleration time histories as the earthquake ground motion records in step (1). For example, if step (1) uses N earthquake ground motions, then step (2) can fit N ground motion time histories. Since the number of measured ground motions is small, only a small number of ground motions at the target site can be obtained through step (2). These ground motion data are then used as a training set to generate a large number of new artificial ground motion data through the subsequent generative adversarial network model.

[0079] (3) Transform the fitted ground motion acceleration time history to obtain an amplitude spectrum grayscale image, and input it as random noise into the generator of the trained generative adversarial network to generate a pseudo amplitude spectrum image with similar features to the amplitude spectrum grayscale image.

[0080] The calculation process of the amplitude spectrum grayscale image is as follows: A short-time Fourier transform is performed on the fitted seismic acceleration time history. To completely preserve the time and frequency domain information of the seismic motion, the short-time Fourier transform does not truncate the time and frequency of the original seismic motion time history. Considering the convolution operation of the generator and discriminator, the window length is set to 198 data points, and the overlap length is 96 data points. The amplitude spectrum grayscale image obtained by the short-time Fourier transform is a two-dimensional square image with a size of 100×100 pixels. Figure 5 As shown in (a), due to the significant differences in peak ground acceleration among different earthquake time histories, the amplitude spectra after short-time Fourier transform also exhibit large numerical differences. To reduce the instability that may occur during model training caused by this discreteness of input data, this invention normalizes the two-dimensional time-frequency domain amplitude spectra corresponding to the 10 earthquakes in the training set. The normalized amplitude spectra are used as the input of the model in the form of single-channel grayscale images. Figure 5 (b) shows the grayscale image of the seismic amplitude spectrum used for training the generative adversarial network model.

[0081] The generative adversarial network includes a generator and a discriminator. During training, the generator is used to generate new pseudo amplitude spectrum images based on random noise, and the discriminator is used to determine the similarity between the pseudo amplitude spectrum image generated by the generator and the amplitude spectrum grayscale image, and backpropagates according to the similarity to make the pseudo amplitude spectrum image generated by the generator close to the amplitude spectrum grayscale image.

[0082] The deep convolutional network model structure of the generator is as follows: Figure 6As shown, after inputting a one-dimensional random noise variable of length 100 into the generator, it is first expanded into an 80000×1 matrix through a fully connected layer, and then the matrix is ​​resized into a 25×25×128 three-dimensional tensor. Next, its size is increased to 50×50×128 through a first upsampling operation. Subsequently, it undergoes two convolution operations, maintaining its size at 50×50×128. A second upsampling operation is then performed, expanding the matrix to a size of 100×100×128. Next, through a third convolution, the size of the matrix is ​​adjusted to 100×100×64. Finally, after a fourth convolution, the output is a 100×100 matrix. The data of this matrix is ​​plotted as a grayscale image, which is the final output of the generator, i.e., the pseudo-amplitude spectrum image. Batch normalization is performed on the output after the first three convolution operations to accelerate neural network training and improve stability, with the momentum parameter of batch normalization set to 0.8. Next, the batch-normalized result is activated using the LeakyReLU function. After the fourth convolutional operation, the output is activated directly using the Tanh function.

[0083] The deep convolutional neural network model structure of the discriminator is as follows: Figure 7 As shown, the discriminator first inputs a pseudo-amplitude spectrum image of size 100×100. After processing through the first convolutional layer, a 3D tensor of size 50×50×32 is generated. Then, through a second convolutional operation, the tensor size is reduced to 25×25×64. Subsequently, a third convolutional operation transforms it into a 13×13×128 3D tensor, followed by a fourth convolution to obtain a 13×13×256 tensor. Next, this 3D tensor is flattened into a vector of length 42364, and finally, a scalar is output through a fully connected layer as the score of the pseudo-amplitude spectrum image. After each convolutional operation, batch normalization, activation, and random deactivation are performed sequentially. In batch normalization, the momentum parameter is set the same as in the generator. The activation function for the nonlinear transformation is the LeakyReLU function. The parameter for random deactivation is set to 0.25, meaning that some neuron outputs are randomly discarded with a probability of 0.25 to reduce the risk of model overfitting. When the flattened vector is processed into the final output through a fully connected layer, the Sigmoid function is used as the activation function to ensure that the discriminator's output is between 0 and 1. The magnitude of the output value determines whether the amplitude spectrum image input to the discriminator is real or generated by the generator; the closer the score is to 1, the closer the discriminator considers the image to be real. Furthermore, the discriminator first needs to learn the features of the amplitude spectrum grayscale image to ensure that the output score of the amplitude spectrum grayscale image is 1. Then, the pseudo-spectral image generated by the generator is input into the discriminator, which gives the pseudo-spectral image a score. Based on the score, the generator and discriminator continuously optimize the parameters of the convolutional neural network, ultimately achieving a dynamic balance.

[0084] This step can generate a large number of pseudo amplitude spectrum images through a generator, providing a data foundation for subsequent steps.

[0085] (4) Process the generated pseudo amplitude spectrum image to reconstruct the phase spectrum image corresponding to the pseudo amplitude spectrum image.

[0086] This invention employs the Griffin-Lim algorithm for reconstruction. Before reconstruction, the generated amplitude spectrum needs to be inversely normalized to obtain a color amplitude spectrum image. Since the random generation of the initial phase spectrum may lead to a slow convergence speed for the Griffin-Lim algorithm, and considering that the phase spectrum corresponding to the training pseudo-amplitude spectrum image is known, it can be used as the initial phase spectrum, thereby improving the efficiency of the Griffin-Lim algorithm in reconstructing seismic ground motion time histories. The number of iterations is a crucial parameter of the Griffin-Lim algorithm. Taking the first seismic ground motion acceleration time history in the training set as an example, the optimal number of iterations is selected by calculating the average absolute error between the reconstructed signal and the original signal under different iteration numbers of the Griffin-Lim algorithm. The iteration error of the Griffin-Lim algorithm is as follows: Figure 8 As shown in (a), the error of the reconstructed signal is minimized when the iteration reaches 990 times, and the reconstruction effect is best at this point. The comparison between the reconstructed acceleration time history and the original time history is as follows: Figure 8 As shown in (b).

[0087] (5) Combine the generated pseudo amplitude spectrum image and the reconstructed phase spectrum image to perform inverse short-time Fourier transform to obtain the one-dimensional ground motion acceleration time history of the target site.

[0088] In this example, 150 amplitude spectrum images generated by the generator at station S1, and 150 phase spectrum images reconstructed by the Griffin-Lim algorithm, are combined to generate 150 artificial ground motion acceleration time histories through inverse short-time Fourier transform. The pseudo-acceleration response spectrum of the artificial ground motion is compared with the true value as follows: Figure 9 As shown in (a), the duration and peak ground acceleration distribution of the simulated ground motion are as follows: Figure 9 As shown in (b). By Figure 9 (a) and Figure 9 As can be seen from (b), the method proposed in this invention realizes the artificial simulation of ground motion in a site without seismic instrument recording. The characteristics of the simulated ground motion in the frequency domain are close to the measured values, and the peak acceleration and duration of the simulated ground motion fluctuate around the corresponding values ​​of the measured ground motion.

[0089] Example 2

[0090] This invention provides a computer device that provides services for implementing the method described in Embodiment 1. The device may include: a memory storing a computer-executable program; a processor coupled to the memory; and the processor calling the computer-executable program stored in the memory to execute the steps of the method described in Embodiment 1.

[0091] The memory may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The device may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, the memory may be used to read and write non-removable, non-volatile magnetic media (commonly referred to as a "hard disk drive"). A program / utility having a set (at least one) of program modules may be stored in, for example, memory. Such program modules include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. The computer-executable program of the program modules typically performs the functions and / or methods described in the embodiments of the present invention.

[0092] The processor executes various functional applications and data processing by running programs stored in memory, such as the method provided in Embodiment 1 of the present invention.

[0093] The code of a computer executable program can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages.

[0094] Example 3

[0095] This invention provides a storage medium containing a computer-executable program, which, when executed by a computer processor, is used to perform the method of Embodiment 1.

[0096] The storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0097] Of course, the computer-executable program provided in the embodiments of the present invention is not limited to the above-described method operations, but can also perform related operations in the methods provided in any embodiment of the present invention.

[0098] Example 4

[0099] This invention also provides a computer program product, such as an app on a mobile phone or tablet, or an installer on a computer. This product includes a computer program / instructions that, when executed by a processor, implement the method described in Embodiment 1. The code for the computer-executable program used to perform the operations of this invention can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as "C" or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0100] It should be understood that the embodiments and descriptions above are only the principles, main features and advantages of the present invention. Various changes and modifications can be made to the present invention without departing from the spirit and scope of the invention, and all such changes and modifications fall within the protection scope of the present invention.

Claims

1. A method for simulating ground motion in a recordless site based on machine learning, characterized in that, Includes the following steps: (1) The pseudo-acceleration response spectrum of the target site is predicted by using the actual ground motion records of several stations near the target site under the earthquake event; (2) Using the pseudo-acceleration response spectrum as the target spectrum, the measured ground motion acceleration time history of the stations near the target site is fitted to obtain the ground motion acceleration time history of the target site. (3) Transform the fitted earthquake acceleration time history to obtain an amplitude spectrum grayscale image, and input one-dimensional random noise into the generator of the trained generative adversarial network to generate a pseudo amplitude spectrum image with similar features to the amplitude spectrum grayscale image; wherein, the generative adversarial network includes a generator and a discriminator. During training, the generator is used to generate a new pseudo amplitude spectrum image based on random noise, and the discriminator is used to determine the similarity between the pseudo amplitude spectrum image generated by the generator and the amplitude spectrum grayscale image, and backpropagate according to the similarity to make the pseudo amplitude spectrum image generated by the generator close to the amplitude spectrum grayscale image; (4) Process the generated pseudo amplitude spectrum image to reconstruct the phase spectrum image corresponding to the pseudo amplitude spectrum image; (5) Combine the generated pseudo amplitude spectrum image and the reconstructed phase spectrum image to perform inverse short-time Fourier transform to obtain the one-dimensional ground motion acceleration time history of the target site.

2. The method according to claim 1, characterized in that: Step (1) specifically includes: (1.1) Obtain the measured ground motion information from several stations near the target site, including the normalized station location coordinates and the average shear wave velocity of the top 30m soil layer of the site where the station is located. The logarithm and the corresponding pseudo-acceleration response spectrum; (1.2) The Gaussian process regression model is established as follows: ， In the formula, This represents the pseudo-acceleration response spectrum. Represents the Gaussian process function. The covariance function representing the Gaussian process is obtained through training. This indicates the actual ground motion information measured by the station. This represents the normalized x, y, and z coordinates of the station location. Indicates station The logarithm of ; This indicates the seismic motion information of the target site. This represents the normalized target site location coordinates along the x, y, and z axes. Indicates the target site The logarithm of ; (1.3) Input the measured ground motion information from several stations into the Gaussian process regression model for training to obtain the trained Gaussian process regression model; (1.4) Input the ground motion information of the target site into the Gaussian process regression model to obtain the pseudo-acceleration response spectrum of the target site.

3. The method according to claim 1, characterized in that: Step (2) specifically includes: (2.1) Obtain the time history of measured ground motion from several stations near the target site; (2.2) A wavelet function is superimposed on the measured ground motion time history at the station to obtain a new ground motion time history, so that the pseudo-acceleration response spectrum corresponding to the new ground motion time history matches the pseudo-acceleration response spectrum of the target site. The new ground motion time history is the ground motion acceleration time history of the target site.

4. The method according to claim 1, characterized in that: In step (3), the calculation process of the amplitude spectrum grayscale image is as follows: A short-time Fourier transform is performed on the fitted seismic acceleration time history. By adjusting the window length and overlap length, an amplitude spectrum image of a preset size is obtained, and then normalized to obtain a grayscale image of the amplitude spectrum.

5. The method according to claim 1, characterized in that: The generator includes an input layer, a fully connected layer, two upsampling layers, and four convolutional layers connected in sequence. A batch normalization operation and a LeakyReLU activation function are provided between the third and fourth convolutional layers, and a Tanh activation function is provided after the fourth convolutional layer.

6. The method according to claim 1, characterized in that, The discrimination process includes one input layer, four convolutional layers, one flattening layer, and one fully connected layer connected in sequence. Each convolutional layer is followed by batch normalization, LeakyReLU activation function, and random deactivation.

7. The method according to claim 1, characterized in that, Step (4) specifically includes: (4.1) Perform inverse normalization on the generated pseudo amplitude spectrum image to obtain the amplitude spectrum color image; (4.2) The phase spectrum image corresponding to the amplitude spectrum color image is reconstructed using the Griffin-Lim algorithm.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes the computer program to implement the method as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, The computer program / instructions, when executed by a processor, implement the method of any one of claims 1-7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the method of any one of claims 1-7.

Images