A method for simulating ground motion based on extracting Green's function from environmental noise field

Abstract

The present invention relates to the field of seismic wave simulation, and specifically discloses a method for simulating earthquake motion based on extracting Green's function from an environmental noise field, comprising: S1. collecting single-station data recorded from an environmental field and performing preprocessing, the preprocessing including detrending and deaveraging, truncation and resampling, removing instrument response, bandpass filtering, and normalization processing; S2. calculating the cross-correlation function of all station pairs; S3. extracting Green's function from an environmental noise field using a neural network based on wavelet basis functions and unsupervised learning technology; S4. synthesizing long-period earthquake motion based on the extracted Green's function. The technical solution of the present invention can achieve high spatial coverage, continuous real-time monitoring and synthesis, and provide higher resolution of underground media.

Description

Technical Field

[0001] The invention relates to the field of seismic wave simulation, and in particular to a seismic motion simulation method based on extracting Green's function from an environmental noise field. Background Art

[0002] Accurate simulation of ground motion is crucial in seismological research and engineering seismic design. Traditional ground motion simulation methods rely on detailed knowledge of underground medium parameters, such as velocity structure models and source mechanisms. These methods usually require a large amount of seismic data collection and complex geological exploration work, which is costly and time-consuming. In addition, in order to obtain sufficient seismic data, a large number of seismic stations or sensors need to be deployed in the target area, which is particularly difficult in areas where there are few people or it is difficult to deploy stations.

[0003] Currently, the method of synthesizing seismic waves is usually limited by the density of seismic stations, making it difficult to obtain accurate information about the structure of small-scale underground media. In addition, traditional methods require a lot of data processing and calculation, resulting in time delays in the synthesis results, making it difficult to achieve continuous real-time monitoring and synthesis.

[0004] In the process of continuous technological development, although some studies have tried to improve traditional methods, they have not been able to completely solve the above problems. Therefore, there is an urgent need for a new ground motion synthesis method that can overcome the defects of traditional methods, achieve high spatial coverage, continuous real-time monitoring and synthesis, and provide higher underground medium resolution. Summary of the invention

[0005] The present invention provides a seismic motion simulation method based on extracting Green's function from an environmental noise field, which can achieve high spatial coverage, continuous real-time monitoring and synthesis, and provide higher underground medium resolution.

[0006] In order to solve the above technical problems, this application provides the following technical solutions:

[0007] A method for simulating ground motion based on extracting Green's function from an environmental noise field, comprising:

[0008] S1. Collect single-station data of environmental field records and perform preprocessing, including detrending and demeaning, truncation and resampling, instrument response removal, bandpass filtering and normalization;

[0009] S2. Calculate the cross-correlation function of all station pairs;

[0010] S3. Extract Green's function in the environmental noise field using neural network based on wavelet basis function and unsupervised learning technology;

[0011] S4. Synthesis of long-period ground motion based on the extracted Green's function.

[0012] Furthermore, in S3, a machine learning method is used to extract Green's function, a neural network based on wavelet basis functions is used for preliminary screening, and then the Green's function is further divided by an unsupervised learning method, and finally the far-field long-period wave is synthesized; the specific steps are as follows:

[0013] S3-1. Selection of basic wavelet functions and input features

[0014] The basic wavelet function is selected as follows

[0015]

[0016] Select the following basic parameters to make the data input terminal xi:

[0017] ω: frequency,

[0018] ts: starting time,

[0019] td: length of the wavelet,

[0020] p: polarization value (±1),

[0021] A: The amplitude ratio of the wavelet to the original correlation function,

[0022] T: predicted and estimated surface wave arrival time ratio,

[0023] F: Deviation between noise correlation function and wavelet between ts and te,

[0024] R: Deviation between the noise correlation function and the wavelet over the entire time period,

[0025] N0: sampling point;

[0026] y i It only contains 0 or 1, indicating whether to remove the related function;

[0027] S3-2. Use artificially selected related functions as training sets, set the neural network structure, and train the network;

[0028] S3-3. Use the trained neural network model to screen all relevant functions;

[0029] S3-4. Arrange it into a format suitable for PCA analysis. Each correlation function can be represented as an eigenvector, where each element represents the amplitude or energy of the waveform at a certain time point or frequency. Use the PCA method to reduce the dimension of the correlation function data; PCA can extract the main change patterns in the data and represent them as a set of principal components. By calculating the eigenvectors and eigenvalues ​​of the covariance matrix, the number and importance of the principal components can be determined;

[0030] S3-5. According to the results of PCA, the most representative principal component is selected as the characteristic of the seismic wave;

[0031] S3-6. Use the mixed Gaussian model to cluster the standardized features; divide them into multiple Gaussian distribution components, each component represents a cluster; select the appropriate number of clusters based on the BIC score and perform model training; assign each seismic wave to the most likely cluster through the mixed Gaussian model, each cluster represents a group;

[0032] S3-7. Take the average value of each cluster as the final Green's function representative value;

[0033] S3-8. Modification of Green's function based on soil layer effects.

[0034] Furthermore, the average value of the first cluster in S3-7 is selected as the approximate representation of the Green's function; then the total Green's function has the following expression:

[0035]

[0036] Since the earthquake source has a certain depth, and the virtual earthquake source assumed by the above method is on the surface, correction is needed.

[0037] Furthermore, r1(0), r1(h) and l1(0), l1(h) are introduced to express the Green's function correction factors of the two surface waves located at the soil surface and at a depth of h, respectively. They are calculated by the spectral method. The corrected Green's functions have the following forms:

[0038]

[0039] in

[0040] c R , c L and U R , U L These are two expressions for the phase velocity and group velocity of surface waves;

[0041] Plot the r1, l1 values ​​when the period is T = 2s and 5s, ∑ n Represents the superposition of Green's functions under different periodic conditions.

[0042] Furthermore, S3-8 is based on the ground motion in the frequency domain after the modified Green's function. The expression is as follows:

[0043]

[0044] In the above formula, is the seismic moment tensor, and Represented as a spatial position vector.

[0045] The basic scheme principle and beneficial effects are as follows: The technical principle of the present invention is based on the theory of seismology and signal processing. Ambient noise, although seemingly random, actually contains information about underground structures and earth medium characteristics.

[0046] The waveform in the ambient noise field is subjected to multiple scattering and reflection by the underground medium, carrying detailed information about the underground structure. By analyzing the cross-correlation between the ambient noise recorded by different stations, common waveform features caused by the characteristics of the underground medium can be identified.

[0047] In the present invention, non-target signals and noise in the data are eliminated through operations such as detrending and deaveraging, so as to prepare for subsequent analysis. The cross-correlation function is used to analyze the similarity between records of different stations, and common signals in the environmental noise are extracted, which contain information about the underground structure. Machine learning technology is applied to identify and extract Green's functions in the environmental noise, which describe the propagation characteristics of seismic waves in the underground medium. Based on the extracted Green's function, the seismic motion under specific conditions is synthesized to provide simulation data for earthquake engineering and risk assessment. By calculating the cross-correlation function between multiple station pairs, the present invention can identify similar waveform patterns that appear simultaneously in different locations, which reflect the continuity and heterogeneity of the underground medium. The multi-scale analysis of the cross-correlation function using wavelet transform can reveal the characteristics of the underground medium at different time and frequency scales, thereby improving the resolution of the underground structure. A large amount of environmental noise data is intelligently screened and feature extracted using a neural network and unsupervised learning technology based on wavelet basis functions, and key signals reflecting the characteristics of the underground medium are automatically identified, further enhancing the ability to identify the details of the underground structure.

[0048] Compared with the traditional method that requires active source to excite seismic waves, the present invention can utilize the widespread environmental noise, saving a lot of data collection costs and time. The present invention is suitable for remote areas where it is difficult to deploy seismic stations, and improves the coverage of seismic monitoring. By continuously collecting environmental noise data, the present invention can achieve real-time monitoring and rapid response of seismic motion. By using the interrelated information extracted from multiple stations, the present invention can more accurately analyze the structural characteristics of the underground medium.

[0049] By selecting appropriate basic wavelet functions and defining input feature parameters, the present invention can capture seismic wave features in environmental noise and provide rich data dimensions for subsequent machine learning screening. Using a neural network based on wavelet basis function features for preliminary screening, potential seismic wave signals can be efficiently identified from a large number of related functions. Unsupervised learning techniques are further applied to cluster the preliminarily screened signals to ensure that the extracted Green's function is highly representative and accurate. The PCA method is used to reduce the dimensionality of the related function data, extract the main change patterns in the data, simplify the data processing process, and retain key information. Using a mixed Gaussian model to cluster the features can divide the seismic wave signals into different groups, each of which represents a specific type of seismic wave feature. In view of the influence of focal depth and soil layer, a correction factor is introduced to correct the extracted Green's function to ensure that the synthesized seismic motion is closer to the actual situation.

[0050] Based on the preliminary screening of wavelet transform and neural network, the present invention can accurately extract the key features of seismic waves and improve the accuracy of signal processing. Through unsupervised learning and PCA dimensionality reduction, the present invention simplifies the data processing process, while retaining the main information of seismic waves and improving the efficiency of data processing. By introducing a correction factor to correct the Green's function, the present invention can simulate seismic motion more realistically and provide more reliable data support for earthquake engineering and risk assessment. The method of the present invention is applicable to different underground media and source conditions, and has good adaptability and generalization ability.

[0051] In summary, the present invention achieves high spatial coverage, continuous real-time monitoring and synthesis, and provides a higher underground medium resolution effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 Schematic diagram of the calculated station noise field correlation function information.

[0053] Figure 2 This is a sample diagram of stations used in the calculation using LINE1 as an example.

[0054] Figure 3 Schematic diagram of velocity structure information obtained by inverting noise correlation function and seismic correlation function.

[0055] Figure 4 This is a comparison chart of the synthesized long-period seismic waves and the actual seismic waves.

[0056] Figure 5 It is a schematic diagram of the process of the present invention.

[0057] Figure 6 Basic flow chart for extracting Green's function based on machine learning method.

[0058] Figure 7 This is an example diagram of the selection of basic wavelet functions and input features.

[0059] Figure 8 This is a cluster analysis diagram of the standardized features using a Gaussian mixture model (GMM).

[0060] Fig. 9 Example plot for choosing the number of clusters based on BIC scores.

[0061] Fig.10 Schematic diagram of the relationship between r1 and l1 values ​​when the period is T=2s and 5s. DETAILED DESCRIPTION

[0062] The following is further described in detail through specific implementation methods:

[0063] A method for simulating ground motion based on extracting Green's function from environmental noise field (such as Figure 5 shown), including:

[0064] S1. Collect single-station data of environmental field records and perform preprocessing, including detrending and demeaning, truncation and resampling, instrument response removal, bandpass filtering and normalization;

[0065] S2. Calculate the cross-correlation function of all station pairs;

[0066] S3. Extract Green's function in the environmental noise field using neural network based on wavelet basis function and unsupervised learning technology;

[0067] S4. Synthesis of long-period ground motion based on the extracted Green's function.

[0068] Specific use: The present invention uses machine learning methods to extract the Green's function of the environmental noise field and to synthesize long-period seismic waves in the far-field basin. Figure 2 and Figure 3 An example of Kanto Basin in Japan is given to illustrate that the extracted Green's function can reflect the velocity structure characteristics of the medium during earthquake propagation. Figure 1 To calculate the station noise field correlation function information, the Figure 2 18,000 correlation functions based on station records from 2 specific stations in 2015-2016, Figure 2 To calculate the stations used, take LINE1 as an example, Figure 3 The velocity structure information inverted by the noise correlation function and the seismic correlation function is compared with the JIVSM data. The results are very close, indicating the effectiveness of the noise correlation function. Figure 4 The synthetic long-period seismic waves and real seismic waves are compared.

[0069] We sample and extract Green's functions based on machine learning methods. First, we use a supervised learning method, that is, a neural network based on wavelet basis functions for preliminary screening. Then, we further divide the Green's functions by unsupervised learning methods, and finally synthesize far-field long-period waves. The basic process is as follows: Figure 6 The specific steps are as follows

[0070] Step 1: Selection of basic wavelet functions and input features

[0071] The basic wavelet function is selected as follows

[0072]

[0073] like Figure 7 As shown, select the following basic parameters to make the data input terminal xi:

[0074] ω: frequency,

[0075] ts: starting time,

[0076] td: length of wavelet,

[0077] p: polarization value (±1),

[0078] A: The amplitude ratio of the wavelet to the original correlation function,

[0079] T: Ratio of predicted and estimated surface wave arrival times

[0080] F: Deviation between noise correlation function and wavelet between ts and te

[0081] R: Deviation between the noise correlation function and the wavelet over the entire time period

[0082] N0: sampling point

[0083] In the figure, the input value N=9 and the output M=2, that is, yi only contains two cases, 0 or 1, indicating whether the related function is eliminated.

[0084] Step 2: Use manually selected related functions as the training set, set the neural network structure, and train the network.

[0085] Step 3: Use the trained neural network model to screen all relevant functions.

[0086] Step 4: Arrange it into a format suitable for PCA analysis. Each correlation function can be represented as an eigenvector, where each element represents the amplitude or energy of the waveform at a certain time point or frequency. Use the PCA method to reduce the dimension of the correlation function data. PCA can extract the main change patterns in the data and represent them as a set of principal components. By calculating the eigenvectors and eigenvalues ​​of the covariance matrix, the number and importance of the principal components can be determined.

[0087] Step 5: Based on the results of PCA, select the most representative principal components as the features of seismic waves. These features should have a high variance contribution rate and be able to retain the information of the original data as much as possible. Standardize the selected features to eliminate the dimensional differences between different features. Standardization can make the contribution of different features to the grouping results more balanced.

[0088] Step 6: Figure 8 As shown, a Gaussian mixture model (GMM) is used to perform cluster analysis on the standardized features. GMM can divide the data into multiple Gaussian distribution components, each of which represents a cluster. The appropriate number of clusters is selected based on the BIC score ( Fig. 9 ), and conduct model training. Through the GMM model, each seismic wave is assigned to the most likely cluster, and each cluster represents a group.

[0089] Step 7: Take the average value of each cluster as the final representative value of the Green's function.

[0090] Step 8: Modify the Green's function based on the influence of soil layers.

[0091] G L is the Love wave Green function G R is the Rayleigh wave Green function. As a simplification, for the far-field vibration, it is assumed that the Rayleigh wave is mainly composed of radial seismic waves, and the Love wave is mainly composed of normal phase seismic waves. The radial and normal phase seismic waves are determined by the above steps 1-7 respectively. We select the average value of the first cluster in step 7 as the approximate representation of the Green function. Then the total Green function has the following expression:

[0092]

[0093] Since the earthquake source has a certain depth, and the virtual earthquake source assumed in the above method is on the surface, correction is needed. According to Aki's seismological theory (Quantitative Seismology), we introduce r1(0), r1(h) and l1(0), l1(h) to express the Green's function correction factors of two surface waves located at the soil surface and at a depth of h, respectively, which are calculated by the spectral method, as follows: Fig.10 As shown, the modified Green's functions have the following forms:

[0094]

[0095] in

[0096] c R , c L and U R , U L The expressions of the two surface wave phase velocity and group velocity are

[0097] by Fig.10 As an example, we plot the values ​​of r1 and l1 when the period is T = 2s and 5s, ∑ n Represents the superposition of Green's functions under different periodic conditions.

[0098] Step 8: Ground motion in the frequency domain based on the modified Green's function expression.

[0099]

[0100]

[0101] In the above formula, is the seismic moment tensor, and Represented as a spatial position vector.

[0102] The above are only embodiments of the present invention. The invention is not limited to the fields involved in this implementation case. The common sense such as the known specific structures and characteristics in the scheme is not described in detail here. The ordinary technicians in the relevant field know all the common technical knowledge in the technical field to which the invention belongs before the application date or the priority date, can obtain all the existing technologies in the field, and have the ability to apply the conventional experimental means before that date. The ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can be made, which should also be regarded as the protection scope of the present invention, which will not affect the effect of the implementation of the present invention and the practicality of the patent. The protection scope required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to explain the content of the claims.

Claims

1. A method for simulating earthquake motion based on extracting Green's function from an environmental noise field, characterized in that: include: S1. Collect single-station data of environmental field records and perform preprocessing, including detrending and demeaning, truncation and resampling, instrument response removal, bandpass filtering and normalization; S2. Calculate the cross-correlation function of all station pairs; S3. Extract Green's function in the environmental noise field using neural network based on wavelet basis function and unsupervised learning technology; S4. Synthesis of long-period ground motion based on the extracted Green's function.

2. The method for simulating earthquake motion based on extracting Green's function from environmental noise field according to claim 1, characterized in that: In S3, a machine learning method is used to extract Green's function, a neural network based on wavelet basis functions is used for preliminary screening, and then an unsupervised learning method is used to further divide the Green's function, and finally a far-field long-period wave is synthesized; the specific steps are as follows: S3-1. Selection of basic wavelet functions and input features The basic wavelet function is selected as follows Select the following basic parameters to make the data input terminal xi: ω: frequency, y i It only contains 0 or 1, indicating whether to remove the related function; S3-2. Use artificially selected related functions as training sets, set the neural network structure, and train the network; S3-3. Use the trained neural network model to screen all relevant functions; S3-4. Arrange into a format suitable for PCA analysis. Each correlation function can be represented as an eigenvector, where each element represents the amplitude or energy of the waveform at a certain time point or frequency. Use the PCA method to reduce the dimension of the correlation function data; PCA can extract the main change patterns in the data and represent them as a set of principal components; by calculating the eigenvectors and eigenvalues ​​of the covariance matrix, the number and importance of the principal components can be determined; S3-5. According to the results of PCA, the most representative principal component is selected as the characteristic of the seismic wave; S3-6. Use the mixed Gaussian model to cluster the standardized features; divide them into multiple Gaussian distribution components, each component represents a cluster; select the appropriate number of clusters based on the BIC score and perform model training; assign each seismic wave to the most likely cluster through the mixed Gaussian model, each cluster represents a group; S3-7. Take the average value of each cluster as the final Green's function representative value; S3-8. Modification of Green's function based on soil layer effects.

3. The method for simulating earthquake motion based on extracting Green's function from environmental noise field according to claim 2, characterized in that: The average value of the first cluster in S3-7 is selected as the approximate representation of the Green's function; then the total Green's function has the following expression: Since the earthquake source has a certain depth, and the virtual earthquake source assumed by the above method is on the surface, correction is needed.

4. The method for simulating earthquake motion based on extracting Green's function from environmental noise field according to claim 3, characterized in that: r1(0), r1(h) and l1(0), l1(h) are introduced to express the Green's function correction factors of the two surface waves located at the soil surface and at a depth of h, respectively. They are calculated by the spectral method. The corrected Green's functions have the following forms: in c R , c L and U R , U L These are two expressions for the phase velocity and group velocity of surface waves; Plot the r1, l1 values ​​when the period is T = 2s and 5s, ∑ n Represents the superposition of Green's functions under different periodic conditions.

5. The method for simulating earthquake motion based on extracting Green's function from environmental noise field according to claim 4, characterized in that: S3-8 is based on the ground motion in the frequency domain after the modified Green's function. The expression is as follows: In the above formula, is the seismic moment tensor, and Represented as a spatial position vector.

Images