Design seismic oscillation simulation method based on machine learning

Abstract

The invention relates to a design seismic oscillation simulation method based on machine learning. Data, a physical model and intelligent optimization are fused. The method specifically comprises the steps that a seismic oscillation characteristic mother wave construction method considering magnitude, epicentral distance and site conditions is proposed, and it is ensured that regional seismic oscillation characteristics are reasonably represented; adopting a particle swarm optimization algorithm to realize quick convergence of simulated seismic oscillation to multi-target parameters; a time-frequency combined constraint mechanism is introduced, and the non-stationary characteristic of an original record is kept through intensity envelope control. According to the method, complete dependence on a statistical model is broken through, actual seismic oscillation is taken as a characteristic mother wave, and the seismic oscillation simulation method which can conform to a regional seismic environment and has both physical rationality and engineering applicability is constructed.

Description

Technical Field

[0001] The present application relates to a design earthquake motion simulation method based on machine learning, which is applicable to the technical field of engineering earthquake resistance. Background Art

[0002] In engineering seismic design, the reliability of seismic motion input directly determines the rationality of the structural seismic design. Engineering practice has shown that even with sophisticated seismic design methods, if the input seismic motion cannot accurately represent the intensity, frequency, and energy time-varying characteristics of the actual earthquake action, the calculated results are likely to deviate significantly from the actual structural seismic response. Currently, strong earthquake records are significantly uneven in regional distribution and are difficult to cover for all types of structural seismic design requirements. There is an urgent need to construct design seismic motion time histories that conform to the characteristics of the regional seismic environment through artificial simulation methods.

[0003] At present, the design seismic motion simulation methods used in engineering practice are mainly engineering simulation methods based on the empirical statistical relationship of seismic motion parameters and random process theory. This method does not explore the physical background that causes irregular seismic motion time history, but instead focuses on the mathematical characterization of its waveform. The goal is to make the final result of the seismic motion simulation consistent with the result of the empirical statistical relationship. Therefore, the computational workload is small and the efficiency is high. However, the accuracy of the description of seismic motion characteristics by this method depends on the rationality of the empirical statistical relationship. For example, the intensity envelope widely used in engineering practice can only describe the change of seismic motion amplitude over time. The time-frequency envelope function can simultaneously characterize the non-stationary characteristics of seismic motion intensity and frequency, but there are too many parameters that need to be calibrated, and the adaptability of the model is relatively poor. Therefore, the existing technology still faces bottlenecks such as insufficient adaptability of empirical models and difficulty in handling parameter uncertainties, which restrict the engineering applicability of seismic motion simulation technology. Summary of the Invention

[0004] This application provides a machine learning-based method for simulating design seismic motions. By combining characteristic mother wavelet selection with an intelligent optimization algorithm, this method effectively characterizes the correlation between the non-stationary characteristics of seismic motions and the seismic geological environment of a specific engineering site while preserving the temporal characteristics of regional seismic motions. The seismic motions simulated using this method not only achieve high-precision fitting of the target acceleration response spectrum but also enable approximate quantitative control of the time-varying characteristics of seismic motion intensity and the non-stationary characteristics of frequency.

[0005] This application relates to a design earthquake simulation method based on machine learning, comprising the following steps:

[0006] (1) Construct multiple acceleration data sets based on ground motion characteristics and classify and preprocess the actual ground motion data;

[0007] (2) Based on the seismological parameters of the target ground motion, the ground motion records with isomorphic characteristics are screened from the established ground motion dataset by similarity matching method to construct the characteristic mother wave matrix;

[0008] (3) Based on the time-frequency characteristics of the target ground motion, a three-segment intensity envelope model is adopted and its parameterized expression is determined through empirical derivation. The intensity envelope model is used to constrain the stationary duration of the selected characteristic mother wave matrix. The characteristic mother wave matrix after duration adjustment is used as the input signal for the subsequent synthesis and fitting of ground motions.

[0009] (4) Through the data-driven PSO particle swarm optimization algorithm, guided by the target response spectrum, the weight coefficient is adaptively optimized to fit the target response spectrum;

[0010] (5) Perform error evaluation on the fitted target response spectrum until the earthquake velocity history that meets the target is obtained.

[0011] Wherein, in step (1), the following steps are included:

[0012] (1.1) The earthquake acceleration records in the NGA-west2 database are grouped according to moment magnitude Mw, epicenter distance R, and site average shear wave velocity Vs30 to form the general earthquake acceleration dataset GAdataset;

[0013] (1.2) Based on the PEER database, we screened the ground motion records containing velocity pulses near the fault, constructed the velocity pulse dataset VAdataset, divided the sites according to Vs30, and established the screening criteria for strong earthquake events;

[0014] (1.3) Establish regional characteristic datasets (RAdatasets) for different earthquake regions, and form specific target datasets by integrating magnitude, distance, and simplified site conditions;

[0015] (1.4) The data from different data sets are preprocessed, and the long and short time window method is used to determine the starting point of the acceleration time history.

[0016] Preferably, in step (1.4), the data length is unified by the envelope function model in formula (1):

[0017]

[0018] Where t1 is the duration of the rising phase, t s is the duration of the stable phase; c is a parameter used to control the speed of intensity attenuation in the descending phase, and its value range is 0.1-2.0; k is a parameter indicating the termination of the earthquake motion, and its value range is 0.1-0.5;

[0019] Then, use formula (2) to obtain the duration t of the acceleration time history decline segment c ;

[0020]

[0021] Finally, a seismic data matrix with a unified format is obtained.

[0022] Among them, in step (3), the characteristic mother wavelet selection and envelope function parameter calibration include the following steps:

[0023] (3.1) Based on the magnitude, epicenter distance and site conditions of the target ground motion, select the adapted dataset from GAdataset, VAdataset and RAdataset;

[0024] (3.2) Through multi-dimensional parameter matching and data quality assessment, a characteristic mother wavelet set a reflecting the target characteristics is constructed. i (t)=(a1(t),a2(t),…a n (t));

[0025] (3.3) In view of the non-stationary characteristics of ground motion, the three-stage intensity envelope model in formula (3) is used to perform time history correction with the target ground motion data.

[0026]

[0027] Where t1 and t2 represent the start and end time of the strong earthquake steady phase, respectively. d Represents the total duration of the earthquake motion.

[0028] Wherein, in step (4), the following steps are included:

[0029] (4.1) Define the objective function to quantify the difference between the synthetic acceleration response spectrum and the target spectrum;

[0030] (4.2) Based on the PSO algorithm, a random particle swarm is generated in the initialization phase, and each particle represents a set of weight coefficient vectors;

[0031] (4.3) Based on the objective function, the PSO algorithm obtains a set of optimal weight coefficients after several iterative optimizations, and then obtains the acceleration time history that meets the target response spectrum.

[0032] Preferably, in step (4.2), the iterative update of the particle's velocity v and position x satisfies equations (5) and (6):

[0033] v ij (t+1)=ωv ij (t)+c1r1(t)[p ij (t)-x ij(t)]+c2r2(t)[p gj (t)-x ij (t)] (5)

[0034] x ij (t+1)=x ij (t)+v ij (t+1) (6)

[0035] Where ω represents the inertia weight; p ij (t) represents the individual optimal position of particle i at the tth iteration, j represents the jth dimension of the solution vector; p gj (t) represents the global optimal position; c1 and c2 are learning factors, c1 = c2 = 2.5; r1 and r2 are two random variables in the particle velocity iteration formula, each taking a random value in the interval [0,1].

[0036] The present application proposes a method for simulating design seismic motions based on machine learning, which integrates data, physical models and intelligently optimized design seismic motion synthesis methods. Specifically, it includes: proposing a method for constructing a seismic motion characteristic mother wave that takes into account the magnitude, epicenter distance and site conditions to ensure the reasonable characterization of regional seismic motion characteristics; using a particle swarm optimization (PSO) algorithm to achieve rapid convergence of simulated seismic motions for multiple target parameters such as response spectrum and duration; introducing a joint time-frequency constraint mechanism to maintain the non-stationary characteristics of the original record through intensity envelope control. The method of the present application breaks through the complete reliance on statistical models, and uses actual seismic motions as characteristic mother waves to construct a seismic motion simulation method that conforms to the regional seismic environment and has both physical rationality and engineering applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flow chart of the design earthquake motion simulation method of the present application.

[0038] Figure 2 Schematic diagram showing the processing of earthquake acceleration data.

[0039] Figure 3 The basic flow chart of synthesizing ground motion using the PSO algorithm is shown.

[0040] Figure 4 Schematic diagram showing the target response spectrum fitting in the embodiment.

[0041] Figure 5 Schematic diagram showing the target acceleration fitting situation in the embodiment.

[0042] Figure 6 A schematic diagram showing the speed of synthesis in the examples. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solutions and advantages of this application more clear, the following will describe the embodiments of this application in detail with reference to the accompanying drawings. It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of this application can be combined with each other in any way.

[0044] A method for designing earthquake motion simulation based on machine learning according to the present application includes the following steps:

[0045] (1) Construct multiple acceleration data sets based on ground motion characteristics and classify and preprocess the actual ground motion data;

[0046] (2) Based on the seismological parameters of the target ground motion, the ground motion records with isomorphic characteristics are screened from the established ground motion database by similarity matching method to construct the characteristic mother wave matrix;

[0047] (3) Based on the time-frequency characteristics of the target ground motion, a three-segment intensity envelope model is adopted and its parameterized expression is determined through empirical derivation. The intensity envelope function is used to constrain the duration of the stationary segment of the selected characteristic mother wave matrix. The characteristic mother wave matrix after duration adjustment is used as the input signal for the subsequent synthesis and fitting of ground motions.

[0048] (4) Through the data-driven PSO particle swarm optimization algorithm, guided by the target response spectrum, the weight coefficient is adaptively optimized to fit the target response spectrum;

[0049] (5) Perform error evaluation on the fitted target response spectrum until the earthquake acceleration time history that meets the target is obtained.

[0050] like Figure 1 As shown, a design earthquake simulation method based on machine learning according to the present application includes the following steps:

[0051] (1) Construct multiple acceleration data sets based on seismic motion characteristics and classify and preprocess the actual seismic motion data

[0052] Firstly, the earthquake records in the NGA-west2 database of the Pacific Earthquake Engineering Center were grouped according to the moment magnitude Mw, epicenter distance R and site average shear wave velocity Vs30 recorded by earthquake monitoring tools to form the general earthquake acceleration dataset GAdataset.

[0053] Secondly, based on the PEER database, near-fault seismic motion records containing velocity pulses were screened to construct the velocity pulse dataset VAdataset. The sites were divided according to Vs30 and strong earthquake event screening criteria were formulated. A regional feature dataset RAdataset was established for different earthquake regions, and multiple feature subsets were formed by comprehensively considering magnitude, distance and simplified site conditions.

[0054] Finally, the data from different datasets are pre-processed, and the long-short window method is used to determine the starting point of the acceleration time history. For example, as shown in FIG. 1, the data length can be unified by the envelope function model in formula (1), and finally the time interval is standardized to 0.005 seconds by using cubic spline interpolation, so as to ensure the data format. Figure 2 Figure 2 In the figure, the abscissa represents the time period, and the unit is second; and the ordinate is acceleration, and the unit is g.

[0055] The formula of the descending segment of the intensity envelope function is as follows:

[0056]

[0057] In the formula, t1 is the duration of the ascending segment, t s is the duration of the stable segment; c is a parameter for realizing the constraint of the descending speed, which is used to control the attenuation speed of the descending segment, and the value range can be 0.1-2.0, and the preferred value is 0.15; k is a parameter representing the termination of the ground motion, that is, when the value of the descending segment function is k, the ground motion is terminated, and the value range of k can be 0.1-0.5, and the preferred value is 0.3. Then, the duration t c of the descending segment of the acceleration time history is obtained by using formula (2). After this processing, the ground motion data matrix with unified format can be obtained, which can be directly applied to the subsequent process.

[0058]

[0059] (2) Selecting characteristic mother waves and determining the parameters of the intensity envelope function

[0060] Based on the seismological parameters of the target ground motion, including the magnitude Mw, the epicentral distance R and the average shear wave velocity Vs30 of the site, the ground motion records with similar characteristics are selected from the established ground motion database by the similarity matching method. Specifically, taking the target ground motion as the benchmark, a data subset with highly similar ground motion characteristics can be selected in the VAdataset dataset, and a characteristic mother wave matrix is constructed; based on the time-frequency characteristics of the target ground motion, a three-segment intensity envelope function model is used, and the parameterized expression is determined by empirical derivation and other methods; the holding time of the stable segment of the selected characteristic mother wave matrix is adjusted by using the intensity envelope model; and the characteristic mother wave matrix after the holding time adjustment is used as an input signal for the subsequent ground motion synthesis and fitting research based on matching pursuit or time-frequency decomposition. This step provides a function set that meets the engineering characteristics of ground motion by constructing and holding time adjusting the characteristic mother wave matrix driven by physics.

[0061] Further, the selection of the characteristic mother wave and the calibration of the parameters of the envelope model based on the systematic process include the following steps: ​

[0062] (2.1) Based on the magnitude, epicenter distance, and site conditions of the target ground motion, select the appropriate dataset from GAdataset, VAdataset, and RAdataset. RAdataset is preferred for regional feature simulation, while VAdataset is selected for near-field pulse effect analysis.

[0063] (2.2) Through multi-dimensional parameter matching and data quality assessment, a characteristic mother wavelet set a reflecting the target characteristics is constructed. i (t)=(a1(t),a2(t),…a n Multidimensional parameter matching involves matching earthquake magnitude, epicenter distance, and site conditions with target seismic data. Data quality assessment involves evaluating the integrity of the waveform of the selected characteristic mother wave data, thereby determining whether the mother wave data recorded by the instrument is abnormal.

[0064] (2.3) In view of the non-stationary characteristics of seismic motion, the three-stage intensity envelope model in formula (3) is used to perform time history correction with the target seismic motion data.

[0065]

[0066] Where t1 and t2 control the start and end time of the strong earthquake steady phase respectively, t2-t1 represents the duration of the steady phase of the earthquake; t d represents the total duration of the earthquake motion; c is the attenuation factor of the descending segment of the aforementioned earthquake motion time history, which is used to control the attenuation rate of the descending segment.

[0067] (3) Fitting target response spectrum based on PSO algorithm

[0068] This step uses the data-driven PSO particle swarm optimization algorithm to achieve adaptive optimization of the weight coefficients with the target response spectrum as the guide. Specifically, the basic process of this step is as follows: Figure 3 As shown, the following steps are included:

[0069] (3.1) The objective function is defined to quantify the difference between the synthetic acceleration response spectrum and the target spectrum, and its expression is:

[0070]

[0071] Where, is the target acceleration response spectrum, S g (T i ) is the calculated response spectrum of the fitted earthquake motion, T i is the i-th period point corresponding to the response spectrum; N is the total value of the response spectrum points.

[0072] (3.2) Based on the PSO algorithm, a random particle swarm is generated in the initialization phase. Each particle represents a set of weight coefficient vectors λ. The iterative updates of the particle velocity v and position x satisfy equations (5) and (6):

[0073] v ij (t+1)=ωv ij (t)+c1r1(t)[p ij (t)-x ij (t)]+c2r2(t)[p gj (t)-x ij (t)] (5)

[0074] x ij (t+1)=x ij (t)+v ij (t+1) (6)

[0075] Where ω represents the inertia weight; p ij (t) represents the individual optimal position of particle i at the tth iteration, j represents the jth dimension of the solution vector. For example, if the optimization problem is D-dimensional, the weight coefficient vector λ=[λ1,λ2,…,λ D ], then j∈{1,2,…,D}; p gj (t) represents the global optimal position; c1 and c2 are learning factors, which can be set to c1 = c2 = 2.5; r1 and r2 are two independently generated random variables in the particle velocity iteration formula, each randomly taking values ​​in the interval [0, 1] to increase the diversity of the search process;

[0076] (3.3) Based on the objective function, the PSO algorithm obtains a set of optimal weight coefficient vectors λ after several iterations of optimization. The determination of the weight coefficient λ can be controlled by error to ensure the fitting accuracy of multiple objectives, and then obtain the acceleration time history that meets the target response spectrum. Right now

[0077]

[0078] Where a i (t) represents the characteristic acceleration.

[0079] (4) Evaluate the error of the fitting results until the earthquake acceleration time history that meets the target is obtained.

[0080] Equation (4) is used to calculate whether the relative fitting error between the synthetic spectrum and the target spectrum in the full period is below the allowable threshold. If not, step (3) is readjusted and iteratively executed. If so, equations (8) and (9) are used to calculate whether the difference between the synthetic ground motion at time t1′ and t2′ and the target ground motion at time t1 and t2 meets the preset threshold p. If so, a ground motion acceleration time history that meets the target is obtained. If not, step (2) is repeated to re-verify the intensity envelope function parameters until a ground motion acceleration time history that meets the target is obtained.

[0081]

[0082] Example

[0083] Taking the record of an actual earthquake station as an example, its target response spectrum is as follows: Figure 4 As shown by the black line in . The earthquake information recorded by the station is moment magnitude Mw = 6.3, epicenter distance R = 68.89km, and site average shear wave velocity Vs30 = 465.55m / s. Based on the target seismic motion parameters and pulse characteristics, 47 near-fault pulse earthquake records with similar magnitude, epicenter distance and site conditions were screened from VAdataset to construct a characteristic mother wave data set. Furthermore, the three-segment intensity envelope parameters of the target acceleration time history are calculated using formula (3), and the rising segment cutoff time t1 = 17.1s, the stable segment end time t2 = 27.1s, the attenuation factor c = 0.15 and the total duration t d =90s. After the characteristic mother wave is modified based on the above envelope function, the particle swarm optimization (PSO) algorithm is introduced to iteratively optimize the mother wave weight coefficient with the target response spectrum as the convergence standard, and finally the synthetic acceleration time history and velocity time history matching the target spectrum are obtained. Figure 4 As shown in Figure 2, the maximum relative fitting error of its response spectrum is 14.68%. Figure 5 The figure shows the fitting effect of acceleration duration, where the horizontal axis represents the time period in seconds and the vertical axis represents the acceleration in g. Figure 6 The resulting velocity is shown, with the horizontal axis representing the time period in seconds and the vertical axis representing the velocity in meters per second. The figure shows a velocity pulse phenomenon. The seismic response spectrum, acceleration duration, intensity, and frequency nonstationary characteristics all meet the target requirements.

[0084] Although the embodiments disclosed in this application are as described above, the contents described are merely embodiments adopted to facilitate understanding of this application and are not intended to limit this application. Any person skilled in the art of the art to which this application belongs may make any modifications and changes in the form and details of the implementation without departing from the spirit and scope disclosed in this application. However, the scope of patent protection of this application shall still be based on the scope defined by the attached claims.

Claims

1. A design earthquake simulation method based on machine learning, characterized in that: The following steps are involved: (1) Construct multiple acceleration data sets based on ground motion characteristics and classify and preprocess the actual ground motion data; (2) Based on the seismological parameters of the target ground motion, the ground motion records with isomorphic characteristics are screened from the established ground motion dataset by similarity matching method to construct the characteristic mother wave matrix; (3) Based on the time-frequency characteristics of the target ground motion, a three-segment intensity envelope model is adopted and its parameterized expression is determined through empirical derivation. The intensity envelope model is used to constrain the stationary duration of the selected characteristic mother wave matrix. The characteristic mother wave matrix after duration adjustment is used as the input signal for the subsequent synthesis and fitting of ground motions. (4) Through the data-driven PSO particle swarm optimization algorithm, guided by the target response spectrum, the weight coefficient is adaptively optimized to fit the target response spectrum; (5) Perform error evaluation on the fitted target response spectrum until the earthquake velocity history that meets the target is obtained.

2. The design earthquake simulation method based on machine learning according to claim 1, characterized in that: In step (1), the following steps are included: (1.1) The earthquake acceleration records in the NGA-west2 database are grouped according to moment magnitude Mw, epicenter distance R, and site average shear wave velocity Vs30 to form the general earthquake acceleration dataset GAdataset; (1.2) Based on the PEER database, we screened the ground motion records containing velocity pulses near the fault, constructed the velocity pulse dataset VAdataset, divided the sites according to Vs30, and established the screening criteria for strong earthquake events; (1.3) Establish regional characteristic datasets (RAdatasets) for different earthquake regions, and form specific target datasets by integrating magnitude, distance, and simplified site conditions; (1.4) The data from different data sets are preprocessed, and the long and short time window method is used to determine the starting point of the acceleration time history.

3. The design earthquake simulation method based on machine learning according to claim 2, characterized in that: In step (1.4), the data length is unified by the envelope function model in formula (1): Where t1 is the duration of the rising phase, t s is the duration of the plateau; c is a parameter used to control the speed of intensity decay in the descending section, and its value range is 0.1-2.0; k is a parameter indicating the termination of the earthquake motion, and its value range is 0.1-0.5; Then, use formula (2) to obtain the duration t of the acceleration time history decline segment c ; Finally, a seismic data matrix with a unified format is obtained.

4. The design earthquake simulation method based on machine learning according to any one of claims 2-3, characterized in that: In step (3), the characteristic mother wavelet selection and envelope function parameter calibration include the following steps: (3.1) Based on the magnitude, epicenter distance and site conditions of the target ground motion, select the adapted dataset from GAdataset, VAdataset and RAdataset; (3.2) Through multi-dimensional parameter matching and data quality assessment, a characteristic mother wavelet set a reflecting the target characteristics is constructed. i (t)=(a1(t),a2(t),…a n (t)); (3.3) In view of the non-stationary characteristics of ground motion, the three-stage intensity envelope model in formula (3) is used to perform time history correction with the target ground motion data. Where t1 and t2 represent the start and end time of the strong earthquake steady phase, respectively. d Represents the total duration of the earthquake motion.

5. The design earthquake simulation method based on machine learning according to claim 4 is characterized in that: In step (4), the following steps are included: (4.1) Define the objective function to quantify the difference between the synthetic acceleration response spectrum and the target spectrum; (4.2) Based on the PSO algorithm, a random particle swarm is generated in the initialization phase, and each particle represents a set of weight coefficients; (4.3) Based on the objective function, the PSO algorithm obtains a set of optimal weight coefficients after several iterative optimizations, and then obtains the acceleration time history that meets the target response spectrum.

6. The design earthquake simulation method based on machine learning according to claim 5, characterized in that: In step (4.2), the iterative update of the particle's velocity v and position x satisfies equations (5) and (6): v ij (t+1)=ωv ij (t)+c1r1(t)[p ij (t)-x ij (t)]+c2r2(t)[p gj (t)-x ij (t)] (5) x ij (t+1)=x ij (t)+v ij (t+1) (6) Where ω represents the inertia weight; p ij (t) represents the individual optimal position of particle i at the tth iteration, j represents the jth dimension of the solution vector; p gj (t) represents the global optimal position; c1 and c2 are learning factors, c1 = c2 = 2.5; r1 and r2 are two random variables in the particle velocity iteration formula, each taking a random value in the interval [0,1].

Images