A method and device for predicting the seismic response of a site

Abstract

An embodiment of the present invention provides a method and device for predicting the seismic response of a site, belonging to the technical field of ground motion simulation. The corresponding method includes: determining the pulse parameters and residual parameters of the pulse-type ground motion according to the scenario parameters of the pulse-type ground motion; wherein, the scenario parameters include: the moment magnitude, fault distance, site shear wave velocity and directivity effect parameter of the pulse-type ground motion; generating a time history model of the pulse-type ground motion according to the pulse parameters and residual parameters to simulate the pulse-type ground motion.

Description

Technical Field

[0001] This application belongs to the field of earthquake engineering technology, particularly earthquake disaster prevention technology, and specifically relates to a method and apparatus for predicting site seismic response. Background Technology

[0002] It is easy to understand that in the field of earthquake engineering, accurately predicting the seismic response of a site is of great significance for the seismic design of engineering projects. The response of the site's soil overburden layer under seismic loading is crucial to the seismic safety of buildings constructed on the site.

[0003] Existing site seismic response analysis methods mainly rely on physical mechanism-based numerical simulation methods, such as the finite element method and the finite difference method. Traditional physical mechanism-based methods for site seismic response analysis often involve a large number of model assumptions. Establishing a detailed physical model requires a large number of in-situ dynamic and static parameters of the site soil. Due to the inherent complexity and spatial variability of soil and rock materials, accurately obtaining these parameters is very difficult and costly. Summary of the Invention

[0004] One objective of this invention is to construct a method and apparatus for predicting site seismic response, in order to solve the problem of predicting the site seismic response of shaft stations in areas lacking data.

[0005] Another object of the present invention is to provide a device for predicting site seismic response. Another object of the present invention is to provide an electronic device comprising a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the above-described site seismic response prediction method. A further object of the present invention is to provide a readable medium storing a computer program thereon, the computer program being executed by a processor to implement the steps of the above-described site seismic response prediction method.

[0006] To address the technical problems in the background section of this application, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a method for predicting site seismic response, comprising:

[0008] Acquire event parameters of the target ground motion and information parameters of the target site;

[0009] The pre-generated seismic response prediction model is fine-tuned using transfer learning based on a portion of the event parameters of the target ground motion;

[0010] The earthquake response prediction model is used to predict the response of the target ground motion at the target site based on the event parameters of the target ground motion, the information parameters, and the earthquake response prediction model fine-tuned by transfer learning; wherein the earthquake response prediction model is based on a deep learning algorithm and is generated by training on the event parameters of multiple historical ground motions and the responses of multiple sites corresponding to the multiple historical ground motions.

[0011] In one embodiment of the present invention, the seismic response prediction model includes:

[0012] The ground motion coding module is used to extract the features of the event parameters of the multiple historical ground motions, the features of the responses corresponding to the multiple historical ground motions, and the features of the event parameters of the target ground motion.

[0013] The site information encoding module is used to extract the features of the information parameters of the multiple sites and the features of the information parameters of the target site.

[0014] An information interaction module is used to fuse the characteristics of the event parameters of the target ground motion and the characteristics of the information parameters of the target site; and

[0015] The features of event parameters of the multiple historical ground motions are integrated with the features of information parameters of the multiple sites corresponding to the multiple historical ground motions.

[0016] In one embodiment of the present invention, fine-tuning a pre-generated seismic response prediction model using transfer learning based on a portion of the event parameters of the target ground motion includes:

[0017] Multiple sets of minor earthquake data are selected from the time parameters of the target ground motion; wherein the peak ground acceleration of the minor earthquake data is less than 100 gal.

[0018] The information interaction module is fine-tuned using transfer learning based on the multiple sets of minor earthquake data.

[0019] In one embodiment of the present invention, the information interaction module includes a first fully connected module and a second fully connected module. When the seismic response prediction model predicts the response of the target ground motion corresponding to the target site, the information interaction module is used to:

[0020] The first fully connected module maps the characteristics of the event parameters of the target ground motion.

[0021] The information parameters of the target site are mapped through the second fully connected module, wherein the activation function of the second fully connected module is the sigmoid function;

[0022] The output of the second fully connected module is used as a coefficient and multiplied with the output of the first fully connected module to generate fused features.

[0023] In one embodiment of the present invention, training the seismic response prediction model using event parameters of the plurality of historical seismic ground motions and the responses of the plurality of sites corresponding to the plurality of historical seismic ground motions includes:

[0024] Based on the event parameters of the multiple historical ground motions and the responses of the multiple sites corresponding to the multiple historical ground motions, the response spectra of downhole seismic acceleration and the response spectra of surface seismic acceleration at each site are determined.

[0025] The response spectra of the downhole seismic acceleration at each site are set as the input to the seismic response prediction model;

[0026] The response spectrum of the surface seismic acceleration of each site is set as the output of the seismic response prediction model;

[0027] The hyperparameters of the earthquake response prediction model are optimized using a Bayesian optimization method based on the input and the output to train the earthquake response prediction model.

[0028] In one embodiment of the present invention, the event parameters include the response spectrum, earthquake magnitude, focal depth, and epicentral distance of the triaxial seismic loads on the target site and the plurality of sites.

[0029] The information parameters include one-dimensional shear wave velocity and compression wave velocity structure.

[0030] In a second aspect, the present invention provides a site seismic response prediction device, the device comprising:

[0031] The parameter acquisition module is used to acquire event parameters of the target ground motion and information parameters of the target site.

[0032] The model fine-tuning module is used to perform transfer learning to fine-tune the pre-generated seismic response prediction model based on a portion of the event parameters of the target ground motion;

[0033] The response prediction module is used to predict the response of the target ground motion to the target site based on the event parameters of the target ground motion, the information parameters, and the earthquake response prediction model fine-tuned by transfer learning; wherein the earthquake response prediction model is based on a deep learning algorithm and is generated by training on the event parameters of multiple historical ground motions and the responses of multiple sites corresponding to the multiple historical ground motions.

[0034] Thirdly, the present invention provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of a method for predicting site seismic response.

[0035] Fourthly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of a method for predicting site seismic response.

[0036] Fifthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a method for predicting site seismic response.

[0037] As described above, the site seismic response prediction method and apparatus proposed in this application first obtains the event parameters of the target ground motion and the information parameters of the target site; then, it performs transfer learning fine-tuning on a pre-generated seismic response prediction model based on a portion of the event parameters of the target ground motion; finally, it predicts the response of the target ground motion corresponding to the target site based on the event parameters, information parameters, and the transfer learning fine-tuned seismic response prediction model; wherein, the seismic response prediction model is generated by a deep learning algorithm and trained from the event parameters of multiple historical ground motions and the responses of multiple sites corresponding to multiple historical ground motions.

[0038] The present invention provides a site seismic response prediction method that combines deep learning and transfer learning. It can achieve reliable site seismic response prediction even when measured data of the target site is scarce, and provides reliable technical support for seismic design of engineering projects. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart illustrating the site seismic response prediction method in an embodiment of the present invention. Figure 1 ;

[0041] Figure 2 The block of the earthquake response prediction model in the embodiment of the present invention Figure 1 ;

[0042] Figure 3 The block of the earthquake response prediction model in the embodiment of the present invention Figure 2;

[0043] Figure 4 This is a flowchart illustrating step 200 of the site seismic response prediction method in an embodiment of the present invention.

[0044] Figure 5 This is a flowchart illustrating step 400 of the site seismic response prediction method in an embodiment of the present invention;

[0045] Figure 6 This is a flowchart illustrating step 300 of the site seismic response prediction method in an embodiment of the present invention;

[0046] Figure 7 This is a flowchart illustrating the site seismic response prediction method in a specific embodiment of the present invention.

[0047] Figure 8 A block of a site seismic response prediction device in a specific embodiment of the present invention. Figure 1 ;

[0048] Figure 9 This is a block diagram of the model fine-tuning module 20 in an embodiment of the present invention;

[0049] Figure 10 A block of a site seismic response prediction device in a specific embodiment of the present invention. Figure 2 ;

[0050] Figure 11 This is a block diagram of the model training module 40 in an embodiment of the present invention;

[0051] Figure 12 This is a schematic diagram of the structure of an electronic device in an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0054] It should be noted that the terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover a non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or apparatuses. Without conflict, the embodiments and features in the embodiments of this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0055] The acquisition, storage, use, and processing of data in this application comply with relevant laws and regulations.

[0056] In recent years, researchers have begun to explore data-driven methods to solve the technical problems described in the background section of this application. Deep learning technology can automatically learn the complex nonlinear relationship between seismic motion and site response from a large amount of measured data. Based on a large amount of measured seismic motion data from vertical shaft stations, a series of site seismic response surrogate models have been developed using deep learning methods to achieve high-efficiency and high-precision prediction of site seismic response.

[0057] However, existing deep learning methods still have some problems: First, most methods are developed for specific sites or structures, lacking universality and transferability; second, deep learning models usually require a large amount of training data, while in practical engineering applications, measured seismic motion data for specific sites are often scarce, especially strong earthquake data, making it difficult to directly train deep learning models. Therefore, there is an urgent need to develop a technical solution that can comprehensively consider seismic motion characteristics and site conditions, and still achieve reliable site seismic response prediction under data scarcity conditions, to meet the needs of practical engineering applications. Based on this, embodiments of the present invention provide a specific implementation method for predicting site seismic response, see [link to specific implementation]. Figure 1 The method specifically includes the following:

[0058] Step 100: Obtain the event parameters of the target ground motion and the information parameters of the target site;

[0059] Step 200: Fine-tune the pre-generated seismic response prediction model using transfer learning based on a portion of the event parameters of the target ground motion;

[0060] Step 300: Predict the response of the target ground motion at the target site based on the event parameters of the target ground motion, the information parameters, and the earthquake response prediction model fine-tuned by transfer learning; wherein, the earthquake response prediction model is based on a deep learning algorithm and is generated by training on the event parameters of multiple historical ground motions and the responses of multiple sites corresponding to the multiple historical ground motions.

[0061] As described above, the site seismic response prediction method and apparatus proposed in this application first obtains the event parameters of the target ground motion and the information parameters of the target site; then, it performs transfer learning fine-tuning on a pre-generated seismic response prediction model based on a portion of the event parameters of the target ground motion; finally, it predicts the response of the target ground motion corresponding to the target site based on the event parameters, information parameters, and the transfer learning fine-tuned seismic response prediction model; wherein, the seismic response prediction model is generated by a deep learning algorithm and trained from the event parameters of multiple historical ground motions and the responses of multiple sites corresponding to multiple historical ground motions.

[0062] This invention can fully utilize massive amounts of measured data to train a general model and fine-tune the model using a small amount of target site data, achieving reliable prediction under data scarcity conditions. It significantly improves prediction accuracy and model versatility, providing an effective technical means for seismic design of engineering projects.

[0063] In some embodiments of the present invention, the event parameters of step 100 include the response spectrum, earthquake magnitude, focal depth, and epicentral distance of the triaxial seismic loads on the target site and the plurality of sites, specifically:

[0064] Triaxial ground motion load refers to the seismic acceleration components in three directions experienced by a structure under seismic loading. These include: X-direction (horizontal), Y-direction (horizontal), and Z-direction (vertical).

[0065] A response spectrum is a graph that describes a structure's response to seismic motion. It shows the maximum response (such as displacement, velocity, or acceleration) of a single-degree-of-freedom (SDOF) system to a given seismic input at different natural periods. There are three types of response spectra: displacement response spectrum, velocity response spectrum, and acceleration response spectrum.

[0066] Epicentral distance refers to the horizontal distance from the epicenter of an earthquake to a specific location. Epicentral distance affects the time it takes for seismic waves to reach that location and the intensity of the ground motion. Generally, the greater the epicentral distance, the weaker the ground motion.

[0067] The information parameters include one-dimensional shear wave velocity and compression wave velocity structure.

[0068] In some embodiments of the present invention, see Figure 2 The earthquake response prediction model includes:

[0069] The seismic motion encoding module 01 is used to extract features of event parameters of the multiple historical seismic motions, features of the responses corresponding to the multiple historical seismic motions, and features of event parameters of the target seismic motion. This module adopts a convolutional neural network structure, which is divided into three convolutional layers and three fully connected layers in practical applications. The convolutional layers use one-dimensional convolution operations. The first convolutional layer has a kernel size of 5 and a number of kernels of 16; the second convolutional layer has a kernel size of 5 and a number of kernels of 32; and the third convolutional layer has a kernel size of 5 and a number of kernels of 64. The number of neurons in the fully connected layers is set to 1024, 512, and 256, respectively. The ReLU activation function is used after each convolutional layer and fully connected layer, and Dropout technology is used to prevent overfitting, with a dropout rate of 0.2.

[0070] The input to the seismic response prediction model includes the response spectrum of the triaxial seismic loads on the site and the corresponding seismic event parameters. To more clearly illustrate the amplification effect of the site and reduce the impact of data noise on the model, the seismic load response spectrum is used as the input. Spectral values ​​are selected over 22 periods: T = [0, 0.01, 0.02, 0.03, 0.05, 0.075, 0.1, 0.15, 0.2, 0.25, 0.3, 0.4, 0.5, 0.75, 1.0, 1.5, 2.0, 3.0, 4.0, 5.0, 7.5, 10]s, where the spectral value at period 0 s represents the PGA (Programme-Generalized Gain). Furthermore, the earthquake magnitude, focal depth, and epicentral distance are also considered. These three parameters are concatenated with the spectral values ​​of each period, resulting in a final seismic motion parameter input data dimension of 6×22.

[0071] The site information encoding module 02 is used to extract the features of the information parameters of the multiple sites and the features of the information parameters of the target site.

[0072] Specifically, the site information encoding module is used to extract features of site information parameters. This module also adopts a convolutional neural network structure, containing three convolutional layers and three fully connected layers. The first convolutional layer has a kernel size of 11 and a number of kernels of 8; the second convolutional layer has a kernel size of 11 and a number of kernels of 16; the third convolutional layer has a kernel size of 11 and a number of kernels of 16. The number of neurons in the fully connected layers is consistent with that in the seismic motion encoding module, and both use the ReLU activation function and Dropout regularization.

[0073] The site information parameters include the one-dimensional shear wave velocity and compressional wave velocity structure of the site, taking equivalent wave velocities at different depths. Since the surface wave velocity structure is more detailed than that of deeper layers, it has a greater impact on the site's seismic response. The wave velocity sampling interval gradually increases from the surface to the depth, starting at a depth of 1m, with sampling intervals of 1m, increasing by 1m every 5 points, for a total of 100 points, reaching a maximum depth of 1050m. Therefore, the input data dimension for the site information parameters is 2×100.

[0074] Information interaction module 03 is used to fuse the characteristics of the event parameters of the target ground motion and the characteristics of the information parameters of the target site; and

[0075] The features of event parameters of the multiple historical ground motions are integrated with the features of information parameters of the multiple sites corresponding to the multiple historical ground motions.

[0076] In some embodiments of the present invention, see Figure 3 The earthquake response prediction model further includes:

[0077] Decoding module 04 maps the fused feature vectors to the final output. The decoding module employs three fully connected layers with 256, 128, and 66 neurons respectively. The last layer outputs a dimension of 66, corresponding to the three-dimensional seismic response spectrum of the site surface (22 periodic points in each direction). The first two layers of the decoding module use the ReLU activation function, while the last layer does not use an activation function to ensure output continuity.

[0078] In some embodiments of the present invention, see Figure 4 Step 200 includes:

[0079] Step 201: Select multiple sets of minor earthquake data from the time parameters of the target ground motion; wherein the peak ground acceleration of the minor earthquake data is less than 100 gal;

[0080] Step 202: Fine-tune the information interaction module using transfer learning based on the multiple sets of small earthquake data.

[0081] In steps 201 and 202, it is understood that in practical engineering applications, measured ground motion data for specific sites are often scarce, especially strong earthquake data, making it difficult to train a deep learning model from scratch. Transfer learning techniques can effectively solve this problem. By utilizing the general feature representations learned by the pre-trained model and fine-tuning it on a small amount of target site data, a reliable prediction model for a specific site can be obtained.

[0082] The core idea of ​​transfer learning is the reuse of model features. The pre-trained model's ground motion encoding module and site information encoding module have already learned general feature representations of ground motion and site information, which have a certain degree of universality across different sites. Therefore, during transfer learning, the parameters of these two encoding modules (ground motion encoding module and site information encoding module) are frozen, and only the information interaction module and decoding module are fine-tuned. This strategy maintains the stability of general features while allowing the model to adapt to the response characteristics of specific sites.

[0083] Specifically, for each new site, a small amount of ground motion data is first collected as a fine-tuning dataset. In practical applications, 20-50 sets of minor earthquake data (surface PGA less than 100 gal) are selected for fine-tuning. The reasons for selecting minor earthquake data are: firstly, minor earthquake data is relatively easy to obtain; secondly, minor earthquake responses mainly reflect the linear characteristics of the site, which helps the model learn the basic response patterns of the site.

[0084] In the learning rate settings during fine-tuning, since the pre-trained model has already converged to a good state, a smaller learning rate is needed during fine-tuning to avoid destroying the learned features. The fine-tuning learning rate is set to 1 / 10 to 1 / 5 of the pre-training learning rate, i.e., 0.0001-0.0002. The batch size is adjusted according to the amount of fine-tuning data, set to 16-32. An early stopping strategy is used during fine-tuning to monitor changes in the validation loss. Due to the small amount of fine-tuning data, to avoid overfitting, fine-tuning is completed within a few epochs. After fine-tuning, a small amount of validation data is used to evaluate the model performance to ensure good fine-tuning results.

[0085] In some embodiments of the present invention, the information interaction module includes a first fully connected module and a second fully connected module. When the seismic response prediction model predicts the response of the target ground motion corresponding to the target site, the information interaction module is used for:

[0086] The first fully connected module maps the characteristics of the event parameters of the target ground motion.

[0087] The information parameters of the target site are mapped through the second fully connected module, wherein the activation function of the second fully connected module is the sigmoid function;

[0088] The output of the second fully connected module is used as a coefficient and multiplied with the output of the first fully connected module to generate fused features.

[0089] Specifically, the information interaction module is used to reflect the influence of site conditions on the amplitude and spectral characteristics of ground motion by fusing ground motion features and site features. Both pieces of information pass through corresponding fully connected mapping modules, each consisting of three fully connected layers, mapping their respective inputs to a vector space of the same size. The activation function of the ground motion feature mapping module is ReLU, while the activation function of the site information feature mapping module is sigmoid, ensuring that the output vector is between 0 and 1. These output vectors are used as weighting coefficients and multiplied element-wise with the ground motion feature mapping result to achieve adaptive feature fusion.

[0090] In some embodiments of the present invention, see Figure 5 The seismic response prediction model is trained using event parameters of the multiple historical earthquakes and the responses of the multiple sites corresponding to the multiple historical earthquakes, including:

[0091] Step 401: Based on the event parameters of the multiple historical ground motions and the responses of the multiple sites corresponding to the multiple historical ground motions, determine the response spectrum of downhole seismic acceleration and the response spectrum of surface seismic acceleration for each site.

[0092] The data for this step is primarily sourced from a seismic observation network database. This database contains a large number of strong-motion observation stations equipped with high-precision surface and downhole strong-motion instruments, capable of simultaneously recording triaxial ground motions from both the surface and bedrock. The stations are distributed across diverse site conditions, providing a rich data foundation for constructing general predictive models.

[0093] Preferably, station data with complete site parameter records were selected for the study. The training dataset contains hundreds of thousands of ground motion data records from hundreds of stations across tens of thousands of earthquake events. The site types cover a wide range from hard rock to soft soil, and the earthquake events cover a range of magnitudes from minor to major earthquakes, providing a solid data foundation for large-scale general-purpose deep learning proxy models.

[0094] In addition, prior to step 401, the raw data needs to be preprocessed. The data preprocessing process includes: firstly, calculating the response spectrum of all ground motion acceleration time history data, selecting 22 periods of engineering interest, which cover the main frequency bands for structural dynamics analysis. The response spectrum of underground ground motion is used as part of the input data, while the response spectrum of surface ground motion is used as the output data. For seismic parameter data, due to its log-normal distribution, the natural logarithm is used to transform the focal depth and epicentral distance. For the same station, the site information data in the input data remains consistent. To ensure numerical stability during the training process, all data are normalized to the range of 0-1.

[0095] Step 402: Set the response spectrum of the downhole seismic acceleration at each site as the input to the seismic response prediction model;

[0096] Step 403: Set the response spectrum of the surface seismic acceleration of each site as the output of the seismic response prediction model;

[0097] Step 404: Optimize the hyperparameters of the earthquake response prediction model using the Bayesian optimization method based on the input and the output to train the earthquake response prediction model.

[0098] Specifically, the input and output are randomly divided into training, validation, and test sets. Specifically, the dataset division uses a stratified random sampling method. After mixing the ground motion data from all stations, the data is divided into training, validation, and test sets according to a ratio of 70%-10%-20% respectively, ensuring the balance of site type and magnitude distribution in each subset.

[0099] Hyperparameter optimization employs Bayesian optimization, a highly efficient global optimization algorithm particularly suitable for black-box function optimization problems. The main hyperparameters optimized include: the kernel size of the seismic motion encoding module, the kernel size of the site information encoding module, the number of neurons in the fully connected layer, the learning rate, and the batch size. Bayesian optimization approximates the objective function by constructing a surrogate model and utilizes the acquisition function to balance exploration and exploitation, thereby finding a better combination of hyperparameters within a limited number of evaluation iterations.

[0100] Model training was performed on a GPU-equipped computing platform using the PyTorch deep learning framework. During training, changes in training and validation losses were monitored; training was stopped when the validation loss ceased to decrease to avoid overfitting. After training, the model's performance was evaluated on a test set to ensure good generalization ability.

[0101] The entire model has tens of millions of parameters. The loss function used during model training is mean squared error (MSE), the optimizer is the Adam algorithm, the initial learning rate is set to 0.001, a learning rate decay strategy is adopted, the batch size is set to 256, and an early stopping mechanism is used during training, stopping training when the validation set loss does not decrease for 10 consecutive epochs.

[0102] In some embodiments of the present invention, the event parameters include the response spectrum, earthquake magnitude, focal depth, and epicentral distance of the triaxial seismic loads on the target site and the plurality of sites.

[0103] The information parameters include one-dimensional shear wave velocity and compression wave velocity structure.

[0104] In some embodiments of the present invention, see Figure 6 Step 300 includes:

[0105] Step 301: Input the event parameters of the target ground motion and the information parameters of the target site into the earthquake response prediction model after transfer learning fine-tuning;

[0106] Step 302: Obtain the predicted results of the three-dimensional seismic response of the target site.

[0107] After fine-tuning in steps 301 and 302, the seismic response prediction model can be used to predict the seismic response of the site. The prediction process includes three steps: input data preparation, model inference, and result evaluation.

[0108] During the input data preparation phase, the event parameters of the target ground motion and the information parameters of the target site need to be organized according to the format required by the model. Event parameters include response spectrum data, magnitude, focal depth, and epicentral distance, which need to be preprocessed using the standardization methods employed during training. The site information parameters are the wave velocity profiles of the site, which need to be sampled and organized according to the model input dimension requirements.

[0109] During the model inference phase, the preprocessed input data is fed into the seismic response prediction model, and the prediction results are obtained through forward propagation. Since the model output is normalized data, it needs to be denormalized to obtain the actual seismic response values.

[0110] In the results evaluation phase, the seismic response characteristics of the site are assessed based on the prediction results. To quantify the prediction accuracy of the model, the root mean square error (RMSE) and mean absolute error (MAE) are used for evaluation.

[0111]

[0112]

[0113] in, For the model to predict the surface response in the period The acceleration spectrum value on For the corresponding reference value, N This represents the total number of samples taken in a period.

[0114] Understandably, the prediction results from step 302 are used to assess the site's seismic response characteristics and amplification effects. Specifically, the prediction results can be used to analyze various aspects of the site's response characteristics: by comparing the response spectra of the surface and bedrock, the site's response spectrum amplification factor can be calculated to identify the site's dominant period and maximum amplification factor; by analyzing the response differences under earthquakes of different intensities, the nonlinear response characteristics of the site can be assessed; and the influence of different seismic parameters on the site response can be analyzed to provide guidance for the site's seismic safety assessment.

[0115] In one specific embodiment, the present invention also provides a specific implementation of a method for predicting site seismic response, see [link to implementation details]. Figure 7 Specifically, it includes the following:

[0116] S1: Construct a multi-input deep learning model that includes event parameters of ground motion, response, and site information parameters.

[0117] This multi-input deep learning model includes:

[0118] Model building module: Used to build multi-input deep learning models that include seismic motion parameters and site information parameters. This module includes:

[0119] The seismic motion coding unit is used to instantiate the seismic motion feature extraction subnetwork. It calls the deep learning framework API to build three one-dimensional convolutional layers (kernel size 5, number of channels 16→32→64) and three fully connected layers (number of neurons 1024→512→256). It is configured with ReLU activation function and Dropout parameter of 0.2 to extract high-level semantic features from 6×22-dimensional seismic motion input.

[0120] The site information coding unit is used to instantiate the site information feature extraction subnetwork, construct three convolutional layers (kernel size 11, number of channels 8→16→16) and corresponding fully connected layers, perform feature learning on the depth variation characteristics of the 2×100-dimensional site wave velocity profile, and output a feature vector of the same dimension as the ground motion coding unit.

[0121] The information interaction unit is used to instantiate the feature fusion subnetwork. It contains two parallel three-layer fully connected mapping branches. Through a learnable weight matrix, it maps ground motion and site features to a unified 256-dimensional feature space. The site branch uses sigmoid activation to output normalized weights to realize the adaptive modulation mechanism of the site's response to earthquakes.

[0122] The decoding unit is used to instantiate the response prediction subnetwork and construct a three-layer fully connected structure with an output dimension of 256→128→66. The final output layer corresponds to the response spectrum prediction values ​​for 22 cycles in each of the three directions. Linear activation is used to ensure the continuity of the output.

[0123] Pre-training module: Used for pre-training deep learning models based on massive amounts of measured seismic ground motion data. This module includes:

[0124] The data management unit is used to establish and maintain a large-scale seismic motion database, realize unified storage and indexing of multi-source heterogeneous data, support rapid retrieval by multiple dimensions such as station type, magnitude range, and site conditions, and ensure the diversity and representativeness of training data.

[0125] The data preprocessing unit is used to perform batch data preprocessing tasks, including calling the signal processing library to calculate the response spectrum, implementing adaptive data transformation strategies (logarithmic transformation, normalization), performing data quality checks and outlier removal, and generating a standardized model input format.

[0126] The hyperparameter optimization unit is used to perform automated hyperparameter search. It integrates a Bayesian optimization algorithm library, defines the hyperparameter search space and optimization objective, and automatically finds the optimal network configuration and training parameters by evaluating candidate parameter combinations in parallel.

[0127] The model training unit is used to execute the training process of deep learning models, including loss calculation, gradient update, performance monitoring, etc., to ensure the stability and efficiency of training on large-scale datasets.

[0128] Transfer Learning Module: Used for fine-tuning pre-trained models using transfer learning. This module includes:

[0129] The parameter freezing unit is used to load pre-trained model weights and selectively freeze network layer parameters according to the transfer strategy to maintain the stability of general features.

[0130] The data adaptation unit is used to collect and preprocess a small amount of ground motion data from the target site to build a fine-tuning dataset.

[0131] The training control unit manages the fine-tuning training process, sets a small learning rate and batch size, implements early stopping mechanisms and learning rate decay strategies, and monitors loss changes and gradient updates during the fine-tuning process.

[0132] Prediction Module: Used to predict the seismic response of the target site using a site-specific model. This module includes:

[0133] The data input unit is used to receive and preprocess target ground motion parameters and site information parameters, and provides a standardized data input interface that supports multiple formats of ground motion parameters and site information input.

[0134] The inference computation unit is used to perform forward propagation computation of the model and obtain the original prediction results;

[0135] The result post-processing unit is used to perform post-processing operations such as inverse normalization and format conversion on the prediction results.

[0136] The performance evaluation unit is used to calculate prediction accuracy indicators, evaluate model performance, perform in-depth analysis based on prediction results, calculate site amplification factor, identify dominant cycles, evaluate nonlinear effects, and generate visualization results such as response spectrum and amplification factor diagram, which facilitates the interpretation of results and decision support by engineering technicians.

[0137] S2: The deep learning model is pre-trained based on massive measured ground motion data to obtain the pre-trained model.

[0138] S3: Obtain a small number of ground motion event parameters for the target site, perform transfer learning to fine-tune the pre-trained model, and obtain a target site-specific model.

[0139] S4: Predict the seismic response of the target site using a site-specific model.

[0140] Based on the same inventive concept, this application also provides a site seismic response prediction device, which can be used to implement the method described in the above embodiments, as shown in the following embodiments. Since the principle of the site seismic response prediction device in solving the problem is similar to that of the site seismic response prediction method, the implementation of the site seismic response prediction device can refer to the implementation of the site seismic response prediction method, and repeated details will not be elaborated further. As used below, the terms "unit" or "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0141] Embodiments of the present invention provide a specific implementation of a site seismic response prediction device capable of realizing a site seismic response prediction method, see below. Figure 8 A site seismic response prediction device specifically includes the following components:

[0142] The parameter acquisition module 10 is used to acquire the event parameters of the target ground motion and the information parameters of the target site.

[0143] The model fine-tuning module 20 is used to perform transfer learning fine-tuning on the pre-generated seismic response prediction model based on a portion of the event parameters of the target ground motion;

[0144] The response prediction module 30 is used to predict the response of the target ground motion corresponding to the target site based on the event parameters of the target ground motion, the information parameters, and the earthquake response prediction model fine-tuned by transfer learning; wherein, the earthquake response prediction model is based on a deep learning algorithm and is generated by training the event parameters of multiple historical ground motions and the responses of multiple sites corresponding to the multiple historical ground motions.

[0145] In one embodiment of the present invention, the seismic response prediction model includes:

[0146] The ground motion coding module is used to extract the features of the event parameters of the multiple historical ground motions, the features of the responses corresponding to the multiple historical ground motions, and the features of the event parameters of the target ground motion.

[0147] The site information encoding module is used to extract the features of the information parameters of the multiple sites and the features of the information parameters of the target site.

[0148] An information interaction module is used to fuse the characteristics of the event parameters of the target ground motion and the characteristics of the information parameters of the target site; and

[0149] The features of event parameters of the multiple historical ground motions are integrated with the features of information parameters of the multiple sites corresponding to the multiple historical ground motions.

[0150] In one embodiment of the present invention, see Figure 9 The model fine-tuning module 20 includes:

[0151] Minor earthquake data selection unit 20a is used to select multiple sets of minor earthquake data from the time parameters of the target ground motion; wherein the peak ground acceleration of the minor earthquake data is less than 100 gal;

[0152] The model fine-tuning unit 20b is used to perform the transfer learning fine-tuning of the information interaction module based on the multiple sets of small earthquake data.

[0153] In one embodiment of the present invention, the information interaction module includes a first fully connected module and a second fully connected module. When the seismic response prediction model predicts the response of the target ground motion corresponding to the target site, the information interaction module is used to:

[0154] The first fully connected module maps the characteristics of the event parameters of the target ground motion.

[0155] The information parameters of the target site are mapped through the second fully connected module, wherein the activation function of the second fully connected module is the sigmoid function;

[0156] The output of the second fully connected module is used as a coefficient and multiplied with the output of the first fully connected module to generate fused features.

[0157] In one embodiment of the present invention, see Figure 10 A site seismic response prediction device, further comprising:

[0158] Model training module 40 is used to train the seismic response prediction model using event parameters of the multiple historical ground motions and the responses of multiple sites corresponding to the multiple historical ground motions. See [link / reference] Figure 11 The model training module 40 includes:

[0159] The response spectrum determination unit 40a is used to determine the response spectrum of the downhole seismic acceleration and the response spectrum of the surface seismic acceleration of each site based on the event parameters of the multiple historical ground motions and the responses of the multiple sites corresponding to the multiple historical ground motions.

[0160] Input setting unit 40b is used to set the response spectrum of the downhole seismic acceleration of each site as the input of the seismic response prediction model;

[0161] The output setting unit 40c sets the response spectrum of the surface seismic acceleration of each site as the output of the seismic response prediction model.

[0162] The model training unit 40d is used to optimize the hyperparameters of the earthquake response prediction model based on the input and the output using a Bayesian optimization method, so as to train the earthquake response prediction model.

[0163] In one embodiment of the present invention, the event parameters include the response spectrum, earthquake magnitude, focal depth, and epicentral distance of the triaxial seismic loads on the target site and the plurality of sites.

[0164] The information parameters include one-dimensional shear wave velocity and compression wave velocity structure.

[0165] Embodiments of this application also provide a specific implementation of an electronic device capable of implementing all steps in the site seismic response prediction method described in the above embodiments. See [link to implementation details]. Figure 12 The electronic devices specifically include the following:

[0166] Processor 1201, memory 1202, communications interface 1203, and bus 1204;

[0167] The processor 1201, memory 1202, and communication interface 1203 communicate with each other via bus 1204; the communication interface 1203 is used to realize information transmission between server-side devices and client-side devices and other related devices.

[0168] The processor 1201 is used to call the computer program in the memory 1202. When the processor executes the computer program, it implements all the steps in the site seismic response prediction method in the above embodiments. For example, when the processor executes the computer program, it implements the following steps:

[0169] Step 100: Obtain the event parameters of the target ground motion and the information parameters of the target site;

[0170] Step 200: Fine-tune the pre-generated seismic response prediction model using transfer learning based on a portion of the event parameters of the target ground motion;

[0171] Step 300: Predict the response of the target ground motion at the target site based on the event parameters of the target ground motion, the information parameters, and the earthquake response prediction model fine-tuned by transfer learning; wherein, the earthquake response prediction model is based on a deep learning algorithm and is generated by training on the event parameters of multiple historical ground motions and the responses of multiple sites corresponding to the multiple historical ground motions.

[0172] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the site seismic response prediction method in the above embodiments. The computer-readable storage medium stores a computer program that, when executed by a processor, implements all steps of the site seismic response prediction method in the above embodiments. For example, when the processor executes the computer program, it implements the following steps:

[0173] Step 100: Obtain the event parameters of the target ground motion and the information parameters of the target site;

[0174] Step 200: Fine-tune the pre-generated seismic response prediction model using transfer learning based on a portion of the event parameters of the target ground motion;

[0175] Step 300: Predict the response of the target ground motion at the target site based on the event parameters of the target ground motion, the information parameters, and the earthquake response prediction model fine-tuned by transfer learning; wherein, the earthquake response prediction model is based on a deep learning algorithm and is generated by training on the event parameters of multiple historical ground motions and the responses of multiple sites corresponding to the multiple historical ground motions.

[0176] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, hardware + program embodiments are relatively simple in description because they are fundamentally similar to method embodiments; relevant parts can be referred to the descriptions in the method embodiments.

[0177] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0178] While this application provides method operation steps as shown in the embodiments or flowcharts, more or fewer operation steps may be included based on conventional or non-inventive labor. The order of steps listed in the embodiments is merely one possible execution order among many and does not represent the only execution order. In actual device or client product execution, the method can be executed in the order shown in the embodiments or drawings or in parallel (e.g., in a parallel processor or multi-threaded processing environment).

[0179] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing the embodiments of this specification, the functions of each module can be implemented in one or more software and / or hardware components, or a module that performs the same function can be implemented by a combination of multiple sub-modules or sub-units. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between devices or units, and may be electrical, mechanical, or other forms.

[0180] Those skilled in the art will also know that, besides implementing the controller using purely computer-readable program code, the same functions can be achieved by logically programming the method steps, making the controller function as logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers (PLCs), and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the devices within it used to implement various functions can also be considered structures within that hardware component. Alternatively, the devices used to implement various functions can be considered as both software modules implementing the method and structures within a hardware component.

[0181] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0182] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0183] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, system embodiments are basically similar to method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. In the description of this specification, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the embodiments in this specification. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0184] The above description is merely an embodiment of the present specification and is not intended to limit the embodiments of the present specification. For those skilled in the art, various modifications and variations can be made to the embodiments of the present specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the embodiments of the present specification should be included within the scope of the claims of the embodiments of the present specification.

Claims

1. A method for predicting site seismic response, characterized in that, include: Acquire event parameters of the target ground motion and information parameters of the target site; The pre-generated seismic response prediction model is fine-tuned using transfer learning based on a portion of the event parameters of the target ground motion; The earthquake response prediction model is used to predict the response of the target site to the target ground motion based on the event parameters of the target ground motion, the information parameters, and the earthquake response prediction model after fine-tuning by transfer learning; wherein the earthquake response prediction model is based on a deep learning algorithm and is generated by training on the event parameters of multiple historical ground motions and the responses of multiple sites corresponding to the multiple historical ground motions. The earthquake response prediction model includes: The seismic motion encoding module is used to extract features of event parameters of the multiple historical seismic ground motions, features of the responses corresponding to the multiple historical seismic ground motions, and features of event parameters of the target seismic ground motion. The seismic motion encoding module adopts a convolutional neural network structure, consisting of three convolutional layers and three fully connected layers. The convolutional layers use one-dimensional convolution operations. The first layer has a kernel size of 5 and a kernel count of 16; the second layer has a kernel size of 5 and a kernel count of 32; and the third layer has a kernel size of 5 and a kernel count of 64. The number of neurons in the fully connected layers is set to 1024, 512, and 256, respectively. A ReLU activation function is used after each convolutional layer and fully connected layer, with a dropout rate of 0.

2. The site information encoding module is used to extract the features of the information parameters of the multiple sites and the features of the information parameters of the target site. An information interaction module is used to fuse the characteristics of the event parameters of the target ground motion and the characteristics of the information parameters of the target site; and The features of event parameters of the multiple historical ground motions and the features of information parameters of the multiple sites corresponding to the multiple historical ground motions are integrated; The information interaction module includes a first fully connected module and a second fully connected module. When the seismic response prediction model predicts the response of the target ground motion corresponding to the target site, the information interaction module is used for: The first fully connected module maps the characteristics of the event parameters of the target ground motion. The information parameters of the target site are mapped through the second fully connected module, wherein the activation function of the second fully connected module is the sigmoid function; The output of the second fully connected module is used as a coefficient and multiplied with the output of the first fully connected module to generate fused features; The information interaction module reflects the influence of site conditions on the amplitude and spectral characteristics of ground motion by fusing ground motion features and site features. Both ground motion features and site features are processed by corresponding fully connected mapping modules. The first and second fully connected modules are each composed of three fully connected layers, mapping their respective inputs to a vector space of the same size. The activation function of the ground motion feature mapping module is ReLU, and the activation function of the site information feature mapping module is sigmoid, so that the output vector is between 0 and 1. These are used as weight coefficients and multiplied element-wise with the ground motion feature mapping result to achieve adaptive feature fusion.

2. The prediction method according to claim 1, characterized in that, The pre-generated seismic response prediction model is fine-tuned using transfer learning based on a portion of the event parameters of the target ground motion, including: Multiple sets of minor earthquake data are selected from the time parameters of the target ground motion; wherein the peak ground acceleration of the minor earthquake data is less than 100 gal. The information interaction module is fine-tuned using transfer learning based on the multiple sets of minor earthquake data.

3. The prediction method according to claim 1, characterized in that, The seismic response prediction model is trained using event parameters of the multiple historical seismic ground motions and the responses of the multiple sites corresponding to the multiple historical seismic ground motions, including: Based on the event parameters of the multiple historical ground motions and the responses of the multiple sites corresponding to the multiple historical ground motions, the response spectra of downhole seismic acceleration and the response spectra of surface seismic acceleration at each site are determined. The response spectra of the downhole seismic acceleration at each site are set as the input to the seismic response prediction model; The response spectrum of the surface seismic acceleration of each site is set as the output of the seismic response prediction model; The hyperparameters of the earthquake response prediction model are optimized using a Bayesian optimization method based on the input and the output to train the earthquake response prediction model.

4. The prediction method according to any one of claims 1 to 3, characterized in that, The event parameters include the response spectrum, earthquake magnitude, focal depth, and epicentral distance of the triaxial ground motion loads applied to the target site and the multiple sites. The information parameters include one-dimensional shear wave velocity and compression wave velocity structure.

5. A device for predicting site seismic response, characterized in that, include: The parameter acquisition module is used to acquire event parameters of the target ground motion and information parameters of the target site. The model fine-tuning module is used to perform transfer learning to fine-tune the pre-generated seismic response prediction model based on a portion of the event parameters of the target ground motion; The response prediction module is used to predict the response of the target ground motion to the target site based on the event parameters of the target ground motion, the information parameters, and the earthquake response prediction model fine-tuned by transfer learning; wherein, the earthquake response prediction model is based on a deep learning algorithm and is generated by training the event parameters of multiple historical ground motions and the responses of multiple sites corresponding to the multiple historical ground motions. The earthquake response prediction model includes: The seismic motion encoding module is used to extract features of event parameters of the multiple historical seismic ground motions, features of the responses corresponding to the multiple historical seismic ground motions, and features of event parameters of the target seismic ground motion. The seismic motion encoding module adopts a convolutional neural network structure, consisting of three convolutional layers and three fully connected layers. The convolutional layers use one-dimensional convolution operations. The first layer has a kernel size of 5 and a kernel count of 16; the second layer has a kernel size of 5 and a kernel count of 32; and the third layer has a kernel size of 5 and a kernel count of 64. The number of neurons in the fully connected layers is set to 1024, 512, and 256, respectively. A ReLU activation function is used after each convolutional layer and fully connected layer, with a dropout rate of 0.

2. The site information encoding module is used to extract the features of the information parameters of the multiple sites and the features of the information parameters of the target site. An information interaction module is used to fuse the characteristics of the event parameters of the target ground motion and the characteristics of the information parameters of the target site; and The features of event parameters of the multiple historical ground motions and the features of information parameters of the multiple sites corresponding to the multiple historical ground motions are integrated; The information interaction module includes a first fully connected module and a second fully connected module. When the seismic response prediction model predicts the response of the target ground motion corresponding to the target site, the information interaction module is used for: The first fully connected module maps the characteristics of the event parameters of the target ground motion. The information parameters of the target site are mapped through the second fully connected module, wherein the activation function of the second fully connected module is the sigmoid function; The output of the second fully connected module is used as a coefficient and multiplied with the output of the first fully connected module to generate fused features; The information interaction module reflects the influence of site conditions on the amplitude and spectral characteristics of ground motion by fusing ground motion features and site features. Both ground motion features and site features are processed by corresponding fully connected mapping modules. The first and second fully connected modules are each composed of three fully connected layers, mapping their respective inputs to a vector space of the same size. The activation function of the ground motion feature mapping module is ReLU, and the activation function of the site information feature mapping module is sigmoid, so that the output vector is between 0 and 1. These are used as weight coefficients and multiplied element-wise with the ground motion feature mapping result to achieve adaptive feature fusion.

6. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the site seismic response prediction method according to any one of claims 1 to 4.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the site seismic response prediction method according to any one of claims 1 to 4.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method for predicting site seismic response as described in any one of claims 1 to 4.

Images