Method and system for generating regional ground motion time histories

Abstract

The application discloses a regional seismic motion time history generation method and system, comprising the following steps: obtaining existing seismic motion physical simulation data information and preprocessing to construct a training data set; constructing a regional seismic motion time history generation primary model based on an XGBoost model and a physical information neural network model; training the constructed regional seismic motion time history generation primary model by using the constructed training data set to obtain a regional seismic motion time history generation model; and generating the seismic motion time history of a target region by using the obtained regional seismic motion time history generation model. The application constructs and trains the regional seismic motion time history generation model based on the XGBoost model and the physical information neural network model, not only realizes the generation of the regional seismic motion time history data, but also has higher reliability, better accuracy and better effect.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of seismology, and particularly relates to a regional seismic motion time history generation method and system. BACKGROUND

[0002] For the study of earthquakes, it can help people understand the process of earthquakes, thereby greatly reducing the disasters caused by earthquakes. Therefore, the study of earthquakes is of great significance.

[0003] In the process of earthquake research, accurately and reliably generating ground motion of a specific scenario has always been the basis for seismic hazard analysis, risk assessment, seismic structure design and disaster prevention. At present, the traditional regional seismic motion time history generation scheme mainly includes a physical-based deterministic simulation scheme, a random scheme, a hybrid scheme and an empirical ground motion prediction equation scheme. The physical-based deterministic simulation scheme generally realizes mechanical fidelity through three-dimensional wave propagation modeling, but due to the deterministic solving paradigm, unresolved source physical characterization and small-scale crustal heterogeneity, this scheme needs a large amount of calculation process, which is extremely time-consuming and laborious, and the performance of this scheme is also relatively poor in integrating uncertainty analysis and high-frequency (>1 Hz) waveform synthesis. The random scheme assumes that the ground motion is Gaussian white noise, and based on random vibration theory, iteratively adjusts the matching degree of the simulated response spectrum and the target response spectrum to generate synthetic ground motion records that meet the target spectral characteristics; this scheme has good calculation efficiency, but it needs to make random phase assumption, and also sacrifices physical coherence, and has limited ability in accurately simulating long-period motion. The hybrid scheme generally combines deterministic low-frequency (<1 Hz) simulation and random high-frequency (≥1 Hz) simulation to realize full waveform synthesis; however, this scheme depends on the source model for calibration, so it is still not practical in actual application; moreover, this process needs extensive parameterization, so the complexity of this scheme is high, and the scalability is poor. The empirical ground motion prediction equation scheme allows researchers to select ground motion from a database (such as PEER NGA-West2) that matches the spectral characteristics near the structural natural frequency; however, this scaling can cause waveform distortion, thereby destroying the unstable ground motion characteristics; moreover, this scheme produces mismatches in energy time distribution or other characteristics; in addition, when the structure exhibits nonlinear behavior, the time history analysis of the scaled motion used by this scheme may deviate from the actual structure response, thereby greatly affecting the accuracy and reliability of the scheme. SUMMARY

[0004] One of the purposes of the present application is to provide a regional seismic motion time history generation method with high reliability, good accuracy and good effect.

[0005] The second objective of this invention is to provide a system for implementing the method for generating time histories of regional ground motions.

[0006] The method for generating regional ground motion time histories provided by this invention includes the following steps:

[0007] S1. Obtain existing physical simulation data of seismic motion;

[0008] S2. Preprocess the data obtained in step S1 to construct a training dataset;

[0009] S3. Based on the XGBoost model and the physical information neural network model, a primary model for generating regional ground motion time histories is constructed; wherein, the XGBoost model is used to predict the Fourier amplitude spectrum, pseudospectral acceleration and Arias intensity, and the physical information neural network model is used to generate seismic wave data;

[0010] S4. Using the training dataset constructed in step S2, train the primary model for generating regional ground motion time histories constructed in step S3 to obtain the regional ground motion time histories generation model.

[0011] S5. Using the regional ground motion time history generation model obtained in step S4, the ground motion time history of the target area is generated.

[0012] Step S1 specifically includes the following steps:

[0013] Seismic motion physical simulation data information was obtained from the CyberShake seismic motion physical simulation dataset;

[0014] The acquired data includes source parameters, path parameters, and site parameters;

[0015] Among them, the source parameters include magnitude. Fault width W, fault length L, dip angle Dip, slip angle Rake, focal depth and the depth of the fracture top ;

[0016] Path parameters include rupture distance Joyner-Boore distance Epicenter distance Fault normal coordinates Parallel coordinates of the fault ;

[0017] Site parameters include the time-averaged shear wave velocity over the upper 30 meters. , depth up to 1.0 km / s shear wave velocity level and depth up to 2.5 km / s shear wave velocity level .

[0018] The step S3 comprises the following steps:

[0019] Based on the XGBoost model and the physical information neural network model, a regional seismic ground motion time history generation primary model is constructed;

[0020] The weak learner is iteratively constructed based on the XGBoost model, and is used for correcting the residual generated in the previous iteration; the target of each new weak learner is to optimize the prediction along the steepest descent direction of the loss function;

[0021] A time series generation module is constructed based on the information neural network model; the time series generation module is used for receiving the output of the weak learner, and after nonlinear mapping of a plurality of hidden layers, outputs the distribution sampling of the seismic ground motion acceleration time series at the output layer to generate the seismic wave data.

[0022] The weak learner is iteratively constructed based on the XGBoost model, and specifically comprises the following steps:

[0023] The update rule of the prediction value in the tth iteration is represented as

[0024]

[0025] In the formula, is the prediction value in the tth iteration; is the learning rate, used for controlling the step length; is the feature vector of the ith sample; is the weak learner trained in the tth iteration, and n is the number of training samples; is the differentiable loss function; is the true value label of the ith sample; is the prediction value of the ith sample by the weak model in this round; is the feature vector of the ith sample;

[0026] The following formula is used as the objective function:

[0027]

[0028] In the formula, is the first-order gradient of the loss function with respect to the current prediction; is the second-order gradient of the loss function with respect to the current prediction; is the regularization term, and , is the complexity cost of adding a new leaf, T is the number of leaves of the current tree, is the L2 regularization coefficient of the leaf weight, is the vector of the leaf weight, is the square of L2 norm;

[0029] the weight of leaf node j Take derivative and set the derivative to zero to get the optimal leaf weight is where is the index set of all samples of the jth leaf node;

[0030] Substitute the optimal leaf weight into the objective function and calculate the difference before and after splitting to get the split gain G as where is the instance set of the left child node, is the instance set of the right child node, and I is the instance set of the parent node;

[0031] By maximizing the split gain G, the optimal split feature and split point are identified;

[0032] Discretize the feature value into K bins by the following binning function to reduce the split points to be traversed:

[0033]

[0034] where is the bin number of sample i on feature j, is the quantile point boundary;

[0035] If there are missing values, first calculate the split gain for non-missing samples, then direct the missing values to the left and right child nodes respectively and calculate the gain of both, and finally select the direction with the maximum gain as the "default direction";

[0036] Use the following formula as the loss function:

[0037]

[0038] where is the loss function value; y is the observed value; is the predicted value; is the target quantile number;

[0039] Use the following formula as the covariance matrix learning formula:

[0040]

[0041] where is the regularization of the correlation between outputs; is the trace operator used to calculate the sum of diagonal elements in the matrix; is the multi-task weight matrix; is the target inter-covariance matrix.

[0042] The time series generation module based on the information neural network model includes the following steps:

[0043] The constructed time series generation module is used to generate Fourier amplitude spectra, pseudospectral accelerations, and Arias intensities;

[0044] The Fourier amplitude spectrum is calculated as follows:

[0045]

[0046] In the formula The frequency domain complex value spectrum; The Fourier amplitude spectrum, that is The modulus; N is the total number of sampling points of the signal; For a discrete-time signal; k is the frequency index; n is the time index; and k corresponds to the physical frequency. for , The sampling frequency;

[0047] The formula for calculating pseudospectral acceleration is:

[0048]

[0049]

[0050]

[0051]

[0052] In the formula Pseudospectral acceleration; The natural period of the structure; The damping ratio; This represents the time-domain acceleration response at the m-th sampling point; Frequency domain acceleration response; The acceleration transfer function of a single-degree-of-freedom system at discrete frequency points The value at; Let be the natural angular frequency of the target, and ;

[0053] The formula for calculating the intensity of Arias is:

[0054]

[0055] In the formula Let be the cumulative Arias intensity at step p, and when hour The total Arias intensity is given by g; g is the acceleration due to gravity. It is a time-domain discrete signal; For time step.

[0056] The training described in step S4 includes the following steps:

[0057] Visualize the iterative waveforms to more easily observe the model's behavior patterns and performance characteristics;

[0058] For the generated Fourier amplitude spectrum, pseudospectral acceleration, and Arias intensity, a loss function is constructed, and the loss weights corresponding to the Fourier amplitude spectrum, pseudospectral acceleration, and Arias intensity are dynamically adjusted to optimize the training effect;

[0059] The AdamW optimizer was used to train the model.

[0060] The training specifically includes the following steps:

[0061] The following formula is used as the loss function:

[0062]

[0063] In the formula, LOSS is the value of the loss function; For Fourier amplitude spectrum loss; Weights for Fourier amplitude spectrum loss; For pseudospectral acceleration loss; Weights for pseudospectral acceleration loss; For Arias intensity loss; The Arias intensity loss weights; The calculation formula is , This represents the total number of samples. For the true value of the i-th sample, This is the predicted value for the i-th sample;

[0064] At the start of training, the weights were set to satisfy... During training, when When the value is less than the set threshold, reduce... The value increases simultaneously and The value;

[0065] When training a model using the AdamW optimizer, the update rules include:

[0066] The gradient momentum is estimated using the following formula. and gradient square momentum :

[0067]

[0068]

[0069] wherein is a first exponential decay rate; is the gradient of the t-th step; is a second exponential decay rate;

[0070] The bias correction is performed using the following equation:

[0071]

[0072]

[0073] wherein is the bias-corrected first moment estimate; is the bias-corrected second moment estimate; is the decay rate of the first moment; is the decay rate of the second moment;

[0074] The parameter update with decoupled weight decay is performed using the following equation:

[0075]

[0076] wherein is the model parameter after the t-th optimization step; is a constant set for numerical stability; is the weight decay coefficient.

[0077] The application further provides a system for implementing the regional seismic time history generation method, comprising a data acquisition module, a data processing module, a model construction module, a model training module and a data generation module; the data acquisition module, the data processing module, the model construction module, the model training module and the data generation module are connected in sequence; the data acquisition module is used for acquiring existing seismic physical simulation data information and uploading the data information to the data processing module; the data processing module is used for pre-processing the acquired data information according to the received data information, constructing a training data set and uploading the data information to the model construction module; the model construction module is used for constructing a regional seismic time history generation primary model based on an XGBoost model and a physical information neural network model according to the received data information, and uploading the data information to the model training module; wherein the XGBoost model is used for predicting Fourier amplitude spectrum, pseudo spectrum acceleration and Arias intensity, and the physical information neural network model is used for generating seismic wave data; the model training module is used for training the constructed regional seismic time history generation primary model by using the constructed training data set according to the received data information, obtaining a regional seismic time history generation model and uploading the data information to the data generation module; and the data generation module is used for generating the seismic time history of a target region by using the obtained regional seismic time history generation model according to the received data information.

[0078] The regional seismic time history generation method and system provided by the application construct a regional seismic time history generation model based on an XGBoost model and a physical information neural network model and train the model, which not only realizes the generation of regional seismic time history data, but also has higher reliability, better accuracy and better effect. BRIEF DESCRIPTION OF DRAWINGS

[0079] Figure 1 It is a method flowchart of the method of the application.

[0080] Figure 2 It is a data amount distribution diagram of the database of the method embodiment of the application under different parameters; wherein, Figure 2 (a) is a data amount distribution diagram of source parameters and site parameters, Figure 2 (b) is a data amount distribution diagram of path parameters and site parameters.

[0081] Figure 3 It is a representative waveform evolution sequence diagram of the random sampling acceleration record of the method embodiment of the application.

[0082] Figure 4 It is a synchronous evolution diagram of the multi-index characteristics of the acceleration record of the method embodiment of the application; wherein, Figure 4 (a) is a target pseudo spectrum acceleration diagram; Figure 4(b) is a plot of target Arias Intensity; Figure 4 (c) is a plot of target Fourier Amplitude Spectrum; Figure 4 (d) is a plot of weighted total loss; Figure 4 (e) is a plot of individual target index loss.

[0083] Figure 5 FAS of ground motion including basin effect for embodiments of the method of the present application; wherein, Figure 5 (a) is a plot for R rup = 20 km, M w = 7.5, Figure 5 (b) is a plot for R rup = 20 km, M w = 7.95, Figure 5 (c) is a plot for R rup = 20 km, M w = 8.4, Figure 5 (d) is a plot for R rup = 40 km, M w = 7.5, Figure 5 (e) is a plot for R rup = 40 km, M w = 7.95, Figure 5 (f) is a plot for R rup = 40 km, M w = 8.4, Figure 5 (g) is a plot for R rup = 60 km, M w = 7.5, Figure 5 (h) is a plot for R rup = 60 km, M w = 7.95, Figure 5 (i) is a plot for R rup = 60 km, M w = 8.4, Figure 5 (j) is a plot for R rup = 80 km, M w = 7.5, Figure 5 (k) is a plot for R rup = 80 km, M w = 7.95, Figure 5 (l) is a plot for R rup = 80 km, M w = 8.4.

[0084] Figure 6Figures showing PSA for ground motion including basin effects for embodiments of the method of the present application at different magnitude and fault distances; where, Figure 6 (a) is a plot for R rup = 20 km, M w = 7.5, Figure 6 (b) is a plot for R rup = 20 km, M w = 7.95, Figure 6 (c) is a plot for R rup = 20 km, M w = 8.4, Figure 6 (d) is a plot for R rup = 40 km, M w = 7.5, Figure 6 (e) is a plot for R rup = 40 km, M w = 7.95, Figure 6 (f) is a plot for R rup = 40 km, M w = 8.4, Figure 6 (g) is a plot for R rup = 60 km, M w = 7.5, Figure 6 (h) is a plot for R rup = 60 km, M w = 7.95, Figure 6 (i) is a plot for R rup = 60 km, M w = 8.4, Figure 6 (j) is a plot for R rup = 80 km, M w = 7.5, Figure 6 (k) is a plot for R rup = 80 km, M w = 7.95, Figure 6 (l) is a plot for R rup = 80 km, M w = 8.4.

[0085] Figure 7 Figures showing probabilistic acceleration for embodiments of the method of the present application at M w = 7.45; where, Figure 7 (a1) is a plot of observed acceleration for R rup = 20 km, Figure 7 (a2) is a plot of observed acceleration for R rup = 40 km, Figure 7 (a3) is a plot of observed acceleration for R rup= 20 km, Figure 7 (b1) is R rup = 20 km, Figure 7 (b2) is R rup = 40 km, Figure 7 (b3) is R rup = 60 km, Figure 7 (b4) is R rup = 20 km, Figure 7 (b5) is R rup = 40 km, Figure 7 (b6) is R rup = 60 km, Figure 7 (b7) is R rup = 20 km, Figure 7 (b8) is R rup = 40 km, Figure 7 (b9) is R rup = 60 km, Figure 7 (b10) is R rup = 20 km, Figure 7 (b11) is R rup = 40 km, Figure 7 (b12) is R rup = 60 km, Figure 7 (b13) is R rup = 20 km, Figure 7 (b14) is R rup = 40 km, Figure 7 (b15) is R rup = 60 km, Figure 7 (b16) is R rup = 20 km, Figure 7(b17) is R rup Sixth plot of acceleration generated from 5%-95% confidence intervals for = 40 km, Figure 7 (b18) is R rup Sixth plot of acceleration generated from 5%-95% confidence intervals for = 40 km, Figure 7 (b19) is R rup Seventh plot of acceleration generated from 5%-95% confidence intervals for = 20 km, Figure 7 (b20) is R rup Seventh plot of acceleration generated from 5%-95% confidence intervals for = 40 km, Figure 7 (b21) is R rup Seventh plot of acceleration generated from 5%-95% confidence intervals for = 60 km, Figure 7 (c1) is R rup Pseudo-acceleration spectrum and Arias intensity plot generated from acceleration plot for = 20 km, Figure 7 (c2) is R rup Pseudo-acceleration spectrum and Arias intensity plot generated from acceleration plot for = 40 km, Figure 7 (c3) is R rup Pseudo-acceleration spectrum and Arias intensity plot generated from acceleration plot for = 60 km, Figure 7 (c4) is R rup Fourier amplitude spectrum plot generated from acceleration plot for = 20 km, Figure 7 (c5) is R rup Fourier amplitude spectrum plot generated from acceleration plot for = 40 km, Figure 7 (c6) is R rup Fourier amplitude spectrum plot generated from acceleration plot for = 60 km.

[0086] Figure 8 M = 7.45; wherein, w Probability acceleration plot for = 7.45; wherein, Figure 8 (a1) is R rup Acceleration plot observed for = 20 km, Figure 8 (a2) is R rup Acceleration plot observed for = 40 km, Figure 8 (a3) is R rup Acceleration plot observed for = 60 km, Figure 8 (b1) is R rup First plot of acceleration generated from 5%-95% confidence intervals for = 20 km, Figure 8 (b2) is R rup= 40 km, first schematic of acceleration generated from 5%-95% confidence intervals, Figure 8 (b3) is R rup = 60 km, first schematic of acceleration generated from 5%-95% confidence intervals, Figure 8 (b4) is R rup = 20 km, second schematic of acceleration generated from 5%-95% confidence intervals, Figure 8 (b5) is R rup = 40 km, second schematic of acceleration generated from 5%-95% confidence intervals, Figure 8 (b6) is R rup = 60 km, second schematic of acceleration generated from 5%-95% confidence intervals, Figure 8 (b7) is R rup = 20 km, third schematic of acceleration generated from 5%-95% confidence intervals, Figure 8 (b8) is R rup = 40 km, third schematic of acceleration generated from 5%-95% confidence intervals, Figure 8 (b9) is R rup = 60 km, third schematic of acceleration generated from 5%-95% confidence intervals, Figure 8 (b10) is R rup = 20 km, fourth schematic of acceleration generated from 5%-95% confidence intervals, Figure 8 (b11) is R rup = 40 km, fourth schematic of acceleration generated from 5%-95% confidence intervals, Figure 8 (b12) is R rup = 60 km, fourth schematic of acceleration generated from 5%-95% confidence intervals, Figure 8 (b13) is R rup = 20 km, fifth schematic of acceleration generated from 5%-95% confidence intervals, Figure 8 (b14) is R rup = 40 km, fifth schematic of acceleration generated from 5%-95% confidence intervals, Figure 8 (b15) is R rup = 60 km, fifth schematic of acceleration generated from 5%-95% confidence intervals, Figure 8 (b16) is R rup = 20 km, sixth schematic of acceleration generated from 5%-95% confidence intervals, Figure 8 (b17) is R rup = 40 km, sixth schematic of acceleration generated from 5%-95% confidence intervals, Figure 8 (b18) is R rup = 60 km, sixth schematic of acceleration generated from 5%-95% confidence intervals, Figure 8 (b19) is R rup = 20 km, 5%-95% confidence interval generated acceleration seventh plot, Figure 8 (b20) is R rup = 40 km, 5%-95% confidence interval generated acceleration seventh plot, Figure 8 (b21) is R rup = 60 km, 5%-95% confidence interval generated acceleration seventh plot, Figure 8 (c1) is R rup = 20 km, pseudo-acceleration spectrum and Arias intensity plot generated for acceleration plot, Figure 8 (c2) is R rup = 40 km, pseudo-acceleration spectrum and Arias intensity plot generated for acceleration plot, Figure 8 (c3) is R rup = 60 km, pseudo-acceleration spectrum and Arias intensity plot generated for acceleration plot, Figure 8 (c4) is R rup = 20 km, Fourier amplitude spectrum plot generated for acceleration plot, Figure 8 (c5) is R rup = 40 km, Fourier amplitude spectrum plot generated for acceleration plot, Figure 8 (c6) is R rup = 60 km, Fourier amplitude spectrum plot generated for acceleration plot.

[0087] Figure 9 M = 7.45 for the method embodiment of the invention; wherein, w = 20 km, observed acceleration plot, Figure 9 (a1) is R rup = 20 km, observed acceleration plot, Figure 9 (a2) is R rup = 40 km, observed acceleration plot, Figure 9 (a3) is R rup = 60 km, observed acceleration plot, Figure 9 (b1) is R rup = 20 km, 5%-95% confidence interval generated acceleration first plot, Figure 9 (b2) is R rup = 40 km, 5%-95% confidence interval generated acceleration first plot, Figure 9 (b3) is R rup = 60 km, 5%-95% confidence interval generated acceleration first plot, Figure 9 (b4) is R rup= 20 km, 5%-95% confidence interval generated second plot of acceleration, Figure 9 (b5) is R rup = 40 km, 5%-95% confidence interval generated second plot of acceleration, Figure 9 (b6) is R rup = 60 km, 5%-95% confidence interval generated second plot of acceleration, Figure 9 (b7) is R rup = 20 km, 5%-95% confidence interval generated third plot of acceleration, Figure 9 (b8) is R rup = 40 km, 5%-95% confidence interval generated third plot of acceleration, Figure 9 (b9) is R rup = 60 km, 5%-95% confidence interval generated third plot of acceleration, Figure 9 (b10) is R rup = 20 km, 5%-95% confidence interval generated fourth plot of acceleration, Figure 9 (b11) is R rup = 40 km, 5%-95% confidence interval generated fourth plot of acceleration, Figure 9 (b12) is R rup = 60 km, 5%-95% confidence interval generated fourth plot of acceleration, Figure 9 (b13) is R rup = 20 km, 5%-95% confidence interval generated fifth plot of acceleration, Figure 9 (b14) is R rup = 40 km, 5%-95% confidence interval generated fifth plot of acceleration, Figure 9 (b15) is R rup = 60 km, 5%-95% confidence interval generated fifth plot of acceleration, Figure 9 (b16) is R rup = 20 km, 5%-95% confidence interval generated sixth plot of acceleration, Figure 9 (b17) is R rup = 40 km, 5%-95% confidence interval generated sixth plot of acceleration, Figure 9 (b18) is R rup = 60 km, 5%-95% confidence interval generated sixth plot of acceleration, Figure 9 (b19) is R rup = 20 km, 5%-95% confidence interval generated seventh plot of acceleration, Figure 9 (b20) is R rup = 40 km, 5%-95% confidence interval generated seventh plot of acceleration, Figure 9 (b21) is R rup = 60 km, the 5%-95% confidence interval generated acceleration seventh diagram, Figure 9 (c1) is R rup = 20 km, the pseudo-acceleration spectrum and Arias intensity diagram generated acceleration diagram, Figure 9 (c2) is R rup = 40 km, the pseudo-acceleration spectrum and Arias intensity diagram generated acceleration diagram, Figure 9 (c3) is R rup = 60 km, the pseudo-acceleration spectrum and Arias intensity diagram generated acceleration diagram, Figure 9 (c4) is R rup = 20 km, the Fourier amplitude spectrum diagram generated acceleration diagram, Figure 9 (c5) is R rup = 40 km, the Fourier amplitude spectrum diagram generated acceleration diagram, Figure 9 (c6) is R rup = 60 km, the Fourier amplitude spectrum diagram generated acceleration diagram.

[0088] Figure 10 Key dimensions, key component cross sections and three-dimensional diagrams of RC frame of the method embodiment of the application; wherein, Figure 10 (a) is the elevation view of overall dimensions and layer heights, Figure 10 (b) is the plan view of beams and columns, Figure 10 (c) is the cross-sectional detail view of beams and columns, Figure 10 (d) is the three-dimensional finite element model diagram of dynamic analysis.

[0089] Figure 11 The diagram of the probability acceleration diagram verified by the structural displacement response of the method embodiment of the application; wherein, Figure 11 (a) is the comparison diagram of peak displacements of each floor, Figure 11 (b) is the comparison diagram of peak inter-story drift angles.

[0090] Figure 12 The acceleration diagram and the base shear distribution diagram of the structure under actual ground motion generated by the method embodiment of the application; wherein, Figure 12 (a) is the acceleration-base shear distribution diagram of column 1, Figure 12 (b) is the acceleration-base shear distribution diagram of column 2, Figure 12 (c) is the acceleration-base shear distribution diagram of column 3, Figure 12 (d) is the acceleration-base shear distribution diagram of column 4, Figure 12 (e) is the acceleration-base shear distribution diagram of column 5, Figure 12 (f) is the acceleration-base shear force distribution diagram of column 6, Figure 12 (g) is the acceleration-base shear force distribution diagram of column 7, Figure 12 (h) is the acceleration-base shear force distribution diagram of column 8, Figure 12 (i) is the acceleration-base shear force distribution diagram of column 9.

[0091] Figure 13 Figures are the schematic diagrams of multi-component internal force responses in a representative column under the ground motion generated by the embodiment of the method of the present application and the actual ground motion; wherein, Figure 13 (a) is a schematic diagram of axial force P, Figure 13 (b) is a schematic diagram of bending moment M z , Figure 13 (c) is a schematic diagram of bending moment M y , Figure 13 (d) is a schematic diagram of torsional moment T.

[0092] Figure 14 Fig. 4 is a schematic diagram of functional modules of the system of the present application. DETAILED DESCRIPTION

[0093] Fig. 5 is a schematic diagram of the method flow of the method of the present application: the regionally time-history generation method disclosed in the present application comprises the following steps: Figure 1

[0094] S1. Obtain existing ground motion physical simulation data information; specifically comprising the following steps:

[0095] Obtain ground motion physical simulation data information from the CyberShake ground motion physical simulation set; specifically a subset of the CyberShake ground motion physical simulation 15.12 version data set; the velocity model mainly used in this model is CVM-S4.26.M01 and Northridge 1D; the former is obtained by full three-dimensional waveform tomography, and the borehole information is used as the supplement of the near-surface velocity, and the minimum value of shear wave velocity (Vs) is 500 m / s; the latter represents a layered model constructed by spatial averaging of CVM-S4.26.M01; M01 has a 100 m grid resolution and smooth interfaces; a hybrid method is used to combine the low-frequency (1 Hz) CyberShake 15.4 seismogram with the high-frequency component of the Graves-Pitarka broadband platform to synthesize the broadband ground motion. The obtained data information includes source parameters, path parameters and site parameters;

[0096] The source parameters include magnitude Mw,

[0097] ​​Fault width W, fault length L, dip angle Dip, rake angle Rake, focal depth and the depth of the top of the fault ;

[0098] Path parameters include rupture distance , Joyner-Boore distance , epicenter distance , fault normal coordinate and fault parallel coordinate ; when the dip angle and the slip angle indicate a pure strike-slip mechanism, the fault normal coordinate ( ) and the fault parallel coordinate ( ) are geometrically symmetric, and absolute values are used due to the lack of directional meaning;

[0099] Site parameters include upper 30m time-averaged shear wave velocity , depth to 1.0km / s S-wave velocity level and depth to 2.5km / s S-wave velocity level .

[0100] S2. Preprocess the data information obtained in step S1 to construct a training data set.

[0101] S3. Based on the XGBoost model and the physical information neural network model, a preliminary model for generating regional seismic time history is constructed; wherein the XGBoost model is used to predict Fourier amplitude spectrum, pseudo-spectral acceleration and Arias intensity, and the physical information neural network model is used to generate seismic wave data; including the following steps:

[0102] Based on the XGBoost model and the physical information neural network model, a preliminary model for generating regional seismic time history is constructed;

[0103] Wherein, the weak learner is iteratively constructed based on the XGBoost model, which is used to correct the residual generated by the previous iteration; the goal of each new weak learner is to optimize the prediction along the steepest descent direction of the loss function;

[0104] Based on the information neural network model, a time series generation module is constructed; the time series generation module is used to receive the output of the weak learner, and after nonlinear mapping through several layers of hidden layers, the distribution sampling of the seismic motion acceleration time series is output at the output layer to generate seismic wave data.

[0105] In specific implementation, the weak learner is iteratively constructed based on the XGBoost model, which specifically includes the following steps:

[0106] The update rule of the predicted value in the t-th iteration is represented as

[0107]

[0108] wherein is the prediction value of the tth iteration; is the learning rate used to control the step size; is the feature vector of the ith sample; is the weak learner trained in the tth iteration, and n is the number of training samples; is the differentiable loss function; is the true label of the ith sample; is the prediction value of the ith sample by the current weak model; is the feature vector of the ith sample;

[0109] The following formula is used as the objective function:

[0110]

[0111] wherein is the first-order gradient of the loss function with respect to the current prediction; is the second-order gradient of the loss function with respect to the current prediction; is the regularization term, and , is the complexity cost of adding a new leaf, T is the number of leaves of the current tree, is the L2 regularization coefficient of the leaf weight, is the vector of the leaf weight, is the square of the L2 norm;

[0112] The weight of the leaf node j is is set to 0, the optimal leaf weight is obtained is wherein is the index set of all samples of the jth leaf node;

[0113] The optimal leaf weight is substituted into the objective function, and the difference between before and after the split is calculated to obtain the split gain G as wherein is the instance set of the left child node, is the instance set of the right child node, and I is the instance set of the parent node;

[0114] By maximizing the split gain G, the optimal split feature and split point are identified, which constitutes the basic mechanism of decision tree construction;

[0115] The feature values are discretized into K bins by the following binning function to greatly reduce the split points that need to be traversed:

[0116]

[0117] where b ij is the bin number of sample i on feature j, and τ is the quantile boundary;

[0118] If there are missing values, first calculate the split gain for non-missing samples, then try to direct the missing values to the left and right child nodes respectively and calculate the gain of both, and finally select the direction with the maximum gain as the "default direction";

[0119] The following formula is used as the quantile loss function:

[0120]

[0121] where is the loss function value; y is the observation value; is the predicted value; is the target quantile;

[0122] The following formula is used as the covariance matrix learning formula:

[0123]

[0124] where is the regularization of the correlation between outputs; is a tracking operator used to calculate the sum of diagonal elements in the matrix; is a multi-task weight matrix; is the inter-target covariance matrix.

[0125] In the above scheme, the second-order Taylor expansion is introduced into the objective function to more accurately capture the curvature information of the loss function and improve the optimization accuracy; the regularization term is explicitly integrated: a dynamic balance is achieved between model complexity and generalization ability; the sparse-aware partitioning algorithm dynamically learns the optimal distribution direction of missing values, enhancing the robustness to incomplete data;

[0126] At the same time, the histogram-based binning strategy accelerates feature split calculation; through quantile loss function optimization, native quantile regression support is realized, so that strict prediction interval estimation is performed; the vector leaf tree model explicitly models the dependency between output variables through covariance matrix learning, thereby improving the synergistic effect of multi-task prediction.

[0127] These technical characteristics are highly consistent with the constraint prediction needs of the present application: the multi-output dependency modeling capability realizes the simultaneous optimization of logical consistency between constraints; the histogram acceleration algorithm significantly improves the model optimization efficiency; the quantile regression function provides a key probabilistic intensity measure estimate for probabilistic seismic hazard analysis.

[0128] In specific implementation, the time series generation module based on the information neural network model comprises the following steps:

[0129] The mapping process of the hidden layer in the neural network can be described as first linear transformation and then nonlinear transformation, which is realized by an activation function acting on the neuron; all the nonlinear activation functions in the present application adopt Gaussian Error Linear Unit (GELU) functions, which have a significant performance advantage compared with other activation functions such as ReLU, ELU and LeakyReLU;

[0130] The time series generation module is used to generate Fourier amplitude spectrum, pseudo-spectrum acceleration and Arias intensity.

[0131] The calculation formula of the Fourier amplitude spectrum is

[0132]

[0133] In the formula, is a frequency-domain complex-valued spectrum; is a Fourier amplitude spectrum, that is, the modulus of N is the total number of sampling points of the signal; is a time-domain discrete signal; k is a frequency-domain index; n is a time-domain index; the physical frequency corresponding to k is is , , is a sampling frequency;

[0134] The calculation formula of the pseudo-spectrum acceleration is

[0135]

[0136]

[0137]

[0138]

[0139] In the formula, is a pseudo-spectrum acceleration; is a structural natural vibration period; is a damping ratio; is a time-domain acceleration response of the mth sampling point; is a frequency-domain acceleration response; is the value of the acceleration transfer function of a single-degree-of-freedom system at the discrete frequency point is a target natural angular frequency, and ;

[0140] The calculation formula of the Arias intensity is​​

[0141]

[0142] wherein is the cumulative Arias intensity of the pth step, and when then is the total Arias intensity; g is the acceleration due to gravity; is the time-domain discrete signal; is the time step.

[0143] The time history generated by the above scheme balances between the physical nature and engineering applicability of ground motion, ensuring physical rationality and structural analysis reliability.

[0144] S4. Using the training dataset constructed in step S2, the regional ground motion time history generation model is trained by using the primary model constructed in step S3, to obtain a regional ground motion time history generation model;

[0145] The training process includes the following steps:

[0146] The iterative waveform is visualized to more conveniently observe the behavior pattern and performance characteristics of the model: in the early training stage, the generated signal is shown as unstructured white noise lacking typical seismic characteristics (such as time-varying energy distribution, envelope shape, or high-frequency oscillation details); the waveform presents a chaotic state, without obvious energy accumulation / decay trend, with uniform amplitude distribution, and unable to distinguish the foreshock, main shock, and aftershock stages; in the middle training stage, with the increase of the number of iterations, the signal begins to form a preliminary envelope profile and appears non-uniform energy distribution; the energy accumulation trend becomes identifiable, but the envelope shape is still incomplete and the high-frequency component is sparse; the weak peak acceleration prototype begins to appear, but the peak position is unstable; local periodic oscillation appears, but the overall frequency distribution is still relatively wide, lacking real seismic spectral characteristics; in the later training stage, the generated signal significantly converges to physical seismic characteristics: the waveform forms a complete envelope structure, the energy follows a three-stage evolution pattern: initial accumulation, PGA peak stage, and subsequent decay, the envelope curve gradually smoothens and the symmetry increases; the high-frequency component is more refined, short-period vibration appears and the matching degree with the seismic spectral characteristics improves; the stability of the final signal is enhanced, and the amplitude fluctuation gradually converges to a physically reasonable range.

[0147] For the generated Fourier amplitude spectrum (FAS), pseudo-spectral acceleration, and Arias intensity (PSA), a loss function is constructed, and the loss weights corresponding to the Fourier amplitude spectrum, pseudo-spectral acceleration, and Arias intensity (FAS, PSA, and AI) are dynamically adjusted to optimize the training effect; in the initial training stage, the adaptive system assigns a larger weight to the Fourier amplitude spectrum (FAS) and a smaller weight to the pseudo-spectral acceleration (PSA) and Arias intensity (AI), so that the Fourier amplitude spectrum (FAS) is given priority in the initial training stage; in the middle training stage, the Fourier amplitude spectrum (FAS) and the pseudo-spectral acceleration (PSA) are given equal priority, and the Arias intensity (AI) is given less priority; in the later training stage, the pseudo-spectral acceleration (PSA) is given priority, and the Fourier amplitude spectrum (FAS) and the Arias intensity (AI) are given less priority. The highest weight priority optimizes the time-domain energy distribution, and establishes the overall waveform morphology by controlling the energy accumulation / decay mode. At the same time, the PSA assigns slightly lower weight to continuously focus on the frequency domain features, and maintains the cross-cycle consistency with the target spectrum amplitude. The FAS serves as a supplementary constraint to refine the high-frequency spectral details under the control of the base frequency provided by the PSA. When the time-domain mode is stable, the adaptive mechanism starts weight redistribution according to the preset decay rate: systematically release part of the weight, and allocate it to the PSA and FAS according to the proportion of instantaneous loss. This dynamic rebalancing strategy achieves two key goals: (1) in the transition stage, by maintaining sufficient weight for the PSA, the macro waveform accuracy is preserved; (2) by enhancing the contribution of PSA / FAS, the spectral features are gradually optimized. Through the dynamic weight adjustment throughout the training, the system avoids the dominance of a single indicator in optimization, ensuring the coordinated convergence of time-domain energy dynamics, frequency-domain spectral features, and high-resolution waveform details. This hierarchical priority mechanism first emphasizes the global structure and then turns to local feature optimization. It effectively avoids the risk of premature convergence to a suboptimal solution, while maintaining the physical reasonableness of the generated acceleration record.

[0148] The AdamW optimizer is used for model training: the AdamW optimizer extends the traditional Adam algorithm by decoupling weight decay from gradient updates; in the customized computation graph, the adaptive balance of multiple loss components can cause significant differences in the gradient amplitude of each loss term; the traditional Adam algorithm entangles weight decay with gradient updates, so that the regularization strength is scaled by the adaptive learning rate; this coupling can disrupt the regularization effect when the loss coefficient changes abruptly, because the effective weight decay is affected by gradient noise; AdamW solves this problem by directly applying weight decay after gradient update, ensuring that the regularization effect is independent of the gradient statistical properties; this stability is crucial for multi-objective optimization, when competitive loss terms lead to a pathological optimization terrain; by separating gradient adaptation and parameter shrinkage, AdamW can maintain robust regularization constraints while adaptively traversing complex loss topologies, ultimately achieving better convergence performance.

[0149] In specific implementation, the following steps can be used for training:

[0150] The following formula is used as the loss function:

[0151]

[0152] In the formula, LOSS is the loss function value; is the Fourier amplitude spectrum loss; is the Fourier amplitude spectrum loss weight; is the pseudo-spectrum acceleration loss; is the pseudo-spectrum acceleration loss weight; is the Arias intensity loss;​​ is the weight of Arias intensity loss; The calculation formula is , is the total number of samples, is the true value of the i-th sample, is the predicted value of the i-th sample;

[0153] At the beginning of training, the set weight satisfies ; during the training process, when the value of is less than the set threshold, the value of is reduced and the values of and are increased;

[0154] When the AdamW optimizer is used to train the model, the update rule includes:

[0155] The gradient momentum and the gradient square momentum are estimated by the following formula:

[0156]

[0157]

[0158] In the formula, is the first exponential decay rate; is the gradient at the t-th step; is the second exponential decay rate;

[0159] The bias correction is performed by the following formula:

[0160]

[0161]

[0162] In the formula, is the first-order moment estimate after bias correction; is the second-order moment estimate after bias correction; is the decay rate of the first-order moment; is the decay rate of the second-order moment;

[0163] The parameter update with decoupled weight decay is performed by the following formula:

[0164]

[0165] In the formula, is the model parameter after optimization at the t-th step; is a set constant for numerical stability; is the weight decay coefficient.

[0166] S5. Adopting the regional ground motion time history generation model obtained in step S4, the generation of the ground motion time history of the target region is completed.

[0167] The effects of the method of the present application are described below in combination with one embodiment:

[0168] Acceleration time histories simulated by CyberShake. Two key data sources were integrated by a MySQL database query system: acceleration graph file header metadata obtained via the Globus platform and project-specific information, resulting in a database containing 157,905 validated acceleration time histories. The database contains basic scenario parameters including M w , R rup , V s30 , Z hyp , Z1, and Z 2.5 . Rupture geometry parameters Z tor , Rake, Dip, W, and L were extracted by parsing event-specific rupture configuration files. Source-site geometry parameters (R epi , R jb , R x , R y ) were calculated through coordinate conversion between the WGS-84 geographic system and the Universal Transverse Mercator (UTM) grid, which required detailed geometric analysis of the source-site relationship. After obtaining the acceleration records, an energy-based window function method was used to truncate the records. All waveforms were subjected to initial energy analysis to determine their 1-99% energy concentration intervals. Records containing more than 3,500 sampling points within these intervals were removed (accounting for 1.0% of the original data set, with specific length distribution and removal threshold detailed in Figure 2 ). The remaining records were uniformly processed using a fixed window length of 3,500 sampling points centered on the energy interval. When the window position exceeded the original record boundary, position adjustment was performed to ensure that the window was completely within the available data range. The truncated waveforms were subjected to baseline correction (third-order polynomial removal) and fourth-order Butterworth bandpass filtering (0.1-10 Hz) to obtain the final acceleration time history. Subsequently, key ground motion parameters (IMs) were extracted from these processed records for analysis. The final database contains 157,905 validated acceleration time histories after screening.

[0169] Figure 2 is the data volume distribution of the database on (a) source parameters and (b) path parameters (lower left) and site parameters (upper right). The red line represents the linear relationship between the characteristics. Source parameters include moment magnitude (M), fault width (W) and length (L), dip angle (Dip), slip angle (Rake), source depth (Z hyp ), and rupture top depth (Z tor ). Path parameters include rupture distance (R rup), Joyner-Boore distance (R jb ), epicentral distance (R epi ), fault-normal coordinate (R x ), and fault-parallel coordinate (R y ). Site parameters consist of 30-meter average shear wave velocity (V s30 ) and depth of shear wave velocity reaching 1.0 / 2.5 km / s (Z1 / Z 2.5 ). The dataset is shown in Fig. 1, with the magnitude M w spanning 7.4-8.5, specifically designed to emphasize the dominant role of large earthquakes in structural damage mechanism and seismic risk assessment framework; and the superior performance of the model on high-complexity samples (i.e., large earthquake events) naturally indicates its capability of handling low-complexity scenarios (e.g., smaller magnitude events). The focal depth Z tor is limited within 0.41 km, and the source depth Z hyp extends to 29 km, jointly confirming the classification of shallow crustal earthquakes. The rupture distance R rup is constrained within 200 km, with intensive near-field sampling (<30 km) being crucial for aseismic structural analysis, and data density gradually decreasing beyond 80 km. The V s30 value ranges from 240-720 m / s, effectively characterizing site conditions from compacted soil to weathered bedrock. The Z1 depth extends to 1120 m, embodying significant site condition differences from bare bedrock outcrop to deep sedimentary basin. Strong correlations exist among the source parameters W and L, the path parameters R rup and R jb , and the site parameters Z1 and Z 2.5 , while the correlation between V s30 and Z1 / Z 2.5 , though weaker, is still statistically significant.

[0170] Figure 3A representative waveform evolution sequence of a randomly sampled acceleration record is demonstrated. As shown, at the early stage of training, the generated signal presents unstructured white noise, lacking typical seismic features (e.g., non-stationary characteristics such as temporal energy variation, envelope shape, or high-frequency oscillation details). The waveform appears chaotic, with no clear energy accumulation / decay trend, and the amplitude distribution is uniform, making it impossible to distinguish the foreshock, main shock, and aftershock stages. At the mid-training stage, as the iteration number increases, the signal starts to form a preliminary envelope profile, with uneven energy distribution. Energy accumulation trend can be observed, but the envelope shape is still incomplete, with sparse high-frequency components. Weak peak acceleration prototypes appear, but the peak position is unstable. Local periodic oscillations appear, although the overall frequency distribution is still broad, lacking true seismic spectral characteristics. At the late training stage, the generated signal significantly converges to physical seismic features. The waveform develops a complete envelope structure, with energy following a three-stage evolution pattern: initial accumulation, followed by a PGA peak stage, and subsequent decay, with the envelope curve gradually becoming smoother and more symmetric. In addition, high-frequency components become more refined, with the appearance of short-period oscillations, improving alignment with seismic spectral characteristics. Finally, the signal exhibits higher stability, with amplitude fluctuations gradually converging to a physically reasonable range.

[0171] Figure 4 The synchronous evolution of multi-index features of the acceleration record in Figure 3 is demonstrated. Figure 4 (a) shows that at early iterations, there is a significant difference between the generated response spectrum and the target, which gradually converges to the target PSA as training progresses. Figure 4 (b) shows that the I a energy accumulation curve approaches the target value, confirming successful modeling of the seismic energy release pattern. Figure 4 (c) reveals that spectral characteristics are learned step by step through FAS evolution. Figure 4 (d) indicates that all loss components monotonically decrease with iterations, showing continuous optimization. a Loss converges fastest, followed by PSA, while FAS converges slowest. The total weighted loss approaches zero, verifying global convergence. Figure 4 (e) shows that I a has the highest weight initially, followed by PSA, while FAS weight remains minimal. However, as training progresses, I a weight dynamically decreases, while PSA and FAS weights increase, indicating a preference for refining time-domain and frequency-domain features in later stages. This behavior is due to the carefully designed adaptive loss coefficient regulator of the multi-objective optimization. Specifically, at the initial training stage, the adaptive system assigns I aThe highest weight prioritizes the temporal energy distribution, establishing the overall waveform morphology by controlling the energy build-up / decay pattern. Meanwhile, the slightly lower weight assigned to the PSA ensures continuous attention to the frequency-domain features, maintaining alignment with the target spectral amplitudes across periods. The FAS component serves as a complementary constraint, refining high-frequency spectral details based on the fundamental frequency control provided by the PSA. Once the temporal pattern is stabilized in the later stage, an adaptive mechanism starts redistributing the weights according to a predefined decay rate: it systematically releases portions of the I a weight and allocates these resources proportionally to the PSA and FAS based on their immediate loss ratio. This dynamic rebalancing strategy achieves two key objectives: (1) preserves the accuracy of the macro waveform by maintaining sufficient I a influence during the transition phase; (2) enables gradual refinement of spectral features by enhancing the contributions of the PSA / FAS. By dynamically adjusting the weights throughout the training process, the system prevents any one metric from dominating the optimization, ensuring the synergistic convergence of temporal energy dynamics, frequency-domain spectral features, and high-resolution waveform details. This hierarchical prioritization mechanism—initially emphasizing global structure and then transitioning to local feature optimization—effectively reduces the risk of prematurely converging to a suboptimal solution while maintaining the physical plausibility of the generated acceleration records.

[0172] Figure 5 and Figure 6 demonstrates the physical trends exhibited by the proposed model after training. Specifically, Figure 5 demonstrates the FAS, which contains the ground motion of the basin effect (quantified by the parameter Z1), under different magnitude (M w ) and fault distance (R rup ). Observations include: (1) the FAS amplitude significantly increases with M w , with larger earthquakes releasing more low-frequency energy in proportion; (2) the spectral amplitude generally decays with distance, with the high-frequency components decaying faster than the low-frequency components; (3) the basin amplification effect increases with the Z1 value, with thick sedimentary basins showing significant low-frequency amplification (long-period effect). These observations are consistent with established empirical knowledge about the changes in various parameters. Notably, the model also successfully captures the complex interaction effects between the source, path, and basin parameters. Figure 6 demonstrates the FAS under different M w and R rupUnder these conditions, a PSA including basin effects (Z1 parameter) was used. Key observations include: (1) Total seismic energy increases significantly with increasing magnitude, particularly enhancing the long-period (>1s) component. This is consistent with Brune-type source models, where large earthquakes exhibit lower corner frequencies and longer energy release periods, shifting the spectral peak to longer periods. Short-period amplification remains moderate or even stable due to the more dispersed distribution of high-frequency energy. (2) Increased distance leads to a decrease in the overall spectrum due to geometric diffusion and inelastic decay. High-frequency components decay faster, significantly reducing short-period values, while long-period components decay more slowly, producing a flatter far-field spectrum with a slight rightward shift in peak value. (3) Larger Z1 values ​​(deeper basins) enhance amplification from low-velocity sediments. Multiple basin reflections prolong wave duration, significantly increasing long-period spectral values ​​(>1s), exhibiting a “basin pulsation” effect. These single-parameter trends are consistent with empirical expectations, while the model also accurately captures complex interaction effects.

[0173] Figures 7-9 The above figure compares the actual acceleration records in the test set with those in R. rup The figure below shows probabilistic acceleration maps generated under the scenario of approximately [20, 40, 60] km, while the figure below shows a comparison of their corresponding IMs (only a portion of the generated acceleration maps are shown due to space limitations). The generated probabilistic acceleration maps were constrained to within ±1 standard deviation of the target IM distribution using 300 random samples. Observations show that the generated waveforms exhibit strong consistency with actual records in terms of peak ground acceleration (PGA) amplitude and main waveform patterns at different magnitudes and distances. The envelope of the generated waveforms effectively covers the rise and decay phases of the actual waveforms, indicating that the model can capture the temporal characteristics of actual ground motion. Notably, although the PGA is not explicitly constrained, the generated results still show strong consistency with the observed data. This indicates that the model can learn the implicit correlations between seismic parameters from the training data, demonstrating excellent generalization ability. This implicit correlation further suggests that the selected constraints adequately characterize key waveform features, especially engineering-related features such as seismic design parameters, thus eliminating the need for explicit PGA constraints.

[0174] To further verify the effectiveness of the generated probabilistic ground motion, Figure 8 Chinese R rup Synthetic seismic waves at a range of approximately 20 km were input into a 6-story reinforced concrete (RC) frame structure. The peak structural responses, including force and displacement responses, induced by generated and actual ground motions were compared. The frame model was implemented in OpenSeesPy. Figure 10Key dimensions, key member cross-sections, and a three-dimensional sketch of the RC frame are presented. Column and beam elements use three material models: unconfined concrete for the cover concrete (Concrete01), Kent-Park confined concrete for the core concrete (Concrete01), and longitudinal reinforcement for the steel bars (Steel01). The concrete confinement uses the Kent-Scott-Park model. The parameters for the steel bars include: yield strength F y = 400 MPa, elastic modulus E = 206,000 MPa, and strain hardening ratio b = 0.005. Figure 11 (a) Peak displacements are compared across different stories. The results show that the peak displacement increases significantly with the story height, indicating that high-rise buildings are more susceptible to large seismic displacements. This trend is consistently observed under both recorded and generated ground motions. The envelope region represents the response boundary (i.e., the uncertainty range) induced by the generated waves under the R rup ≈ 20 km scenario, which is a result of mapping the confidence interval at the wave level to the structural response level. All responses induced by the actual waves fall within this range, confirming that the uncertainty of the generated waves can directly inform the structural seismic risk assessment. The confidence interval widens as the story height increases, revealing the increased sensitivity of high-rise structures to ground motion variations. Figure 11 (b) The inter-story drift ratios are compared between the generated and recorded ground motions. Larger inter-story drift ratios occur at lower stories, and the values are smaller for the upper stories—a pattern that is consistently captured by both the observed and generated waves. The red observed curves are densely covered by the gray generated curves, indicating that the model can reproduce realistic structural responses.

[0175] Figure 12 The base shear distributions of the structure are compared under the generated acceleration histories and actual ground motions. The results show that the base shear distributions induced by the generated seismic waves (colored shaded regions) fully encompass all the observed values (red dots within the shaded regions), and most of the observed data points align closely with the central tendency of the generated results (blue curves). This indicates that the model can reproduce realistic base shear distributions. Figure 13 The multi-component internal force responses in a representative column—including axial force, bending moment, and torsional moment—are compared under the generated and actual ground motions. The force responses induced by the generated waves consistently cover the actual values in all components, confirming the effectiveness of the model in reconstructing key internal force distributions.

[0176] As Figure 14The system of the application is shown in the functional module diagram: the system for implementing the regional seismic time history generation method disclosed in the application comprises a data acquisition module, a data processing module, a model construction module, a model training module and a data generation module; the data acquisition module, the data processing module, the model construction module, the model training module and the data generation module are connected in sequence; the data acquisition module is used for acquiring existing seismic physical simulation data information and uploading the data information to the data processing module; the data processing module is used for pre-processing the acquired data information according to the received data information, constructing a training data set and uploading the data information to the model construction module; the model construction module is used for constructing a regional seismic time history generation primary model based on an XGBoost model and a physical information neural network model according to the received data information, and uploading the data information to the model training module; wherein the XGBoost model is used for predicting Fourier amplitude spectrum, pseudo spectrum acceleration and Arias intensity, and the physical information neural network model is used for generating seismic wave data; the model training module is used for training the constructed regional seismic time history generation primary model by using the constructed training data set according to the received data information, obtaining a regional seismic time history generation model and uploading the data information to the data generation module; and the data generation module is used for generating the seismic time history of a target region by using the obtained regional seismic time history generation model according to the received data information.

Claims

1. A method for generating time histories of regional ground motions, characterized by It comprises the following steps: S1. Obtain existing seismic motion physical simulation data information; S2. Preprocess the data information obtained in step S1 to construct a training data set; S3. Based on the XGBoost model and the physical information neural network model, a regional seismic motion time history generation primary model is constructed; wherein the XGBoost model is used to predict Fourier amplitude spectrum, pseudo-spectrum acceleration and Arias intensity, and the physical information neural network model is used to generate seismic wave data; comprising the following steps: Based on the XGBoost model and the physical information neural network model, a regional seismic motion time history generation primary model is constructed; Wherein, the weak learner is iteratively constructed based on the XGBoost model, which is used to correct the residual generated by the previous iteration; the goal of each new weak learner is to optimize the prediction along the steepest descent direction of the loss function; The time series generation module is constructed based on the physical information neural network model; the time series generation module is used to receive the output of the weak learner, and after nonlinear mapping of several hidden layers, the distribution sampling of the seismic motion acceleration time series is output at the output layer to generate seismic wave data; Wherein, the weak learner is iteratively constructed based on the XGBoost model, which comprises the following steps: The update rule of the predicted value in the tth iteration is represented as In the formula is the prediction value of the tth iteration; is the learning rate for controlling the step size; is the feature vector of the ith sample; is the weak learner trained in the tth iteration, and , is the number of training samples; is the differentiable loss function; is the true value label of the ith sample; is the prediction value of the ith sample by the weak model in this round; is the feature vector of the ith sample; The following formula is used as the objective function: wherein is the first order gradient of the loss function with respect to the current prediction; is the second order gradient of the loss function with respect to the current prediction; is a regularization term, and , is a complexity cost of adding a new leaf, T is the number of leaves of the current tree, is an L2 regularization coefficient for leaf weights, is a vector of leaf weights, is the square of the L2 norm; the weight of leaf node j derivation and setting the obtained derivative to 0 to obtain the optimal leaf weight is wherein is the index set of all samples of the jth leaf node The optimal leaf weight is substituted into the objective function, and the difference before and after the splitting of the objective function is calculated to obtain the splitting gain G as wherein is the instance set of the left child node, is the instance set of the right child node, and I is the instance set of the parent node. By maximizing the split gain G, the optimal split feature and split point are identified; The feature values are discretized into K bins through the following binning function to reduce the split points that need to be traversed: wherein is the bin number of sample i on feature j, is the quantile boundary; If there are missing values, first calculate the split gain for non-missing samples, then direct the missing values to the left and right child nodes and calculate the gain of both, and finally select the direction with the maximum gain as the "default direction"; The following formula is used as the loss function: In the formula is a loss function value; y is an observed value; is a predicted value; is a target quantile; The following formula is used as the covariance matrix learning formula: wherein is a regularization of the correlation between outputs; is a trace operator for computing the sum of the diagonal elements of a matrix; is a multitask weight matrix; is an inter-target covariance matrix; S4. The training data set constructed in step S2 is used to train the regional seismic motion time history generation primary model constructed in step S3 to obtain a regional seismic motion time history generation model; S5. The regional seismic motion time history generation model obtained in step S4 is used to complete the generation of the seismic motion time history of the target region.

2. The regional ground motion time history generation method according to claim 1, characterized by The step S1 specifically comprises the following steps: Obtain seismic motion physical simulation data information from the CyberShake seismic motion physical simulation set; The obtained data information includes source parameters, path parameters and site parameters; Wherein, the source parameters include magnitude , fault width W, fault length L, dip angle Dip, rake angle Rake, source depth and fracture top depth ; Path parameters include rupture distance , Joyner-Boore distance , epicentral distance , fault normal coordinate , and fault parallel coordinate ; Site parameters include upper 30m time-averaged shear wave velocity , depth to 1.0 km / s S-wave velocity level and depth to 2.5 km / s S-wave velocity level .

3. The regional ground motion time history generation method according to claim 2, characterized by The time series generation module constructed based on the physical information neural network model specifically comprises the following steps: The constructed time series generation module is used to generate Fourier amplitude spectrum, pseudo-spectrum acceleration and Arias intensity; Wherein, the calculation formula of Fourier amplitude spectrum is wherein is the frequency domain complex valued spectrum; is the Fourier amplitude spectrum, i.e. the modulus of ; N is the total number of samples of the signal; is the time domain discrete signal; k is the frequency domain index; n is the time domain index; k corresponds to the physical frequency is the , is the sampling frequency; The calculation formula of pseudo-spectrum acceleration is wherein is the pseudo-spectral acceleration; is the natural period of the structure; is the damping ratio; is the time-domain acceleration response at the mth sampling point; is the frequency-domain acceleration response; is the value of the acceleration transfer function of a single degree of freedom system at the discrete frequency point ; and is the target natural angular frequency, and ; The calculation formula of Arias intensity is wherein is the cumulative Arias intensity for the pth step, and when is the total Arias intensity; g is the acceleration due to gravity; is the total Arias intensity; g is the acceleration due to gravity; is the time-domain discrete signal; is the time step.

4. The regional ground motion time history generation method according to claim 3, characterized by The training of step S4 comprises the following steps: Visualize the iteration waveform to more conveniently observe the behavior pattern and performance characteristics of the model; A loss function is constructed for the generated Fourier amplitude spectrum, pseudo-spectral acceleration and Arias intensity, and the loss weights corresponding to the Fourier amplitude spectrum, pseudo-spectral acceleration and Arias intensity are dynamically adjusted to optimize the training effect; The model is trained by using an AdamW optimizer.

5. The regional ground motion time history generation method according to claim 4, characterized by The training specifically includes the following steps: The following formula is used as the loss function: where LOSS is a loss function value; is a Fourier amplitude spectrum loss; is a Fourier amplitude spectrum loss weight; is a pseudo-spectrum acceleration loss; is a pseudo-spectrum acceleration loss weight; is an Arias intensity loss; is an Arias intensity loss weight; is a calculation formula of , is a total number of samples, is a real value of the i-th sample, is a predicted value of the i-th sample; At the beginning of the training, the set weight satisfies ; during the training, when the value of is less than a set threshold, the value of is decreased and the values of and are increased simultaneously; When the model is trained by using an AdamW optimizer, the update rule includes: The gradient momentum is estimated using the following equation and the gradient square momentum : wherein is a first exponential decay rate; is the gradient of the tth step; is a second exponential decay rate; The following formula is used for bias correction: wherein is the bias-corrected first moment estimate; is the bias-corrected second moment estimate; is the decay rate of the first moment; is the decay rate of the second moment; The following formula is used for parameter update with decoupled weight decay: In the formula is the model parameter optimized in the tth step; is a constant set for numerical stability; is a weight decay coefficient.

6. A system for implementing the method for generating time histories of regional ground motions according to any one of claims 1 to 5, characterized in that The system comprises a data acquisition module, a data processing module, a model construction module, a model training module and a data generation module; the data acquisition module, the data processing module, the model construction module, the model training module and the data generation module are connected in series; the data acquisition module is configured to acquire existing seismic motion physical simulation data information and upload the data information to the data processing module; The data processing module is configured to pre-process the acquired data information according to the received data information, construct a training data set, and upload the data information to the model construction module; The model construction module is configured to construct a regional seismic motion time history generation primary model based on an XGBoost model and a physical information neural network model according to the received data information, and upload the data information to the model training module; the XGBoost model is configured to predict Fourier amplitude spectrum, pseudo-spectral acceleration and Arias intensity, and the physical information neural network model is configured to generate seismic wave data; the model training module is configured to train the constructed regional seismic motion time history generation primary model by using the constructed training data set according to the received data information, obtain a regional seismic motion time history generation model, and upload the data information to the data generation module; and the data generation module is configured to generate a target regional seismic motion time history by using the obtained regional seismic motion time history generation model according to the received data information.

Images