A method and system for predicting seismic ground motion considering building-site effects

Abstract

This invention discloses a method and system for predicting seismic ground motion fields considering building-site effects, belonging to the field of seismic ground motion prediction technology. The method includes acquiring combined data of different site parameters, seismic ground motion parameters, and building parameters to construct a training sample dataset; uniformly encoding building parameters according to their spatial distribution to form building cluster input features; constructing a machine learning model based on the training sample dataset; using the building cluster input features, site parameters, and seismic ground motion parameters as input to the machine learning model to learn the nonlinear mapping relationship of seismic ground motion fields under multiple building conditions; inputting the site parameters, seismic ground motion parameters, and building parameters of the area to be predicted into the machine learning model; and outputting the seismic ground motion response results of the target area based on the machine learning model. The method of this invention improves the accuracy and efficiency of seismic ground motion prediction, enhances the ability to express building-site coupling effects, and has good engineering applicability and promotional value.

Description

Technical Field

[0001] This invention relates to the field of seismic field prediction technology, specifically to a seismic field prediction method and system that takes into account building-site effects. Background Technology

[0002] With the acceleration of urbanization and the widespread distribution of high-density building clusters, the ground motion characteristics under seismic loading exhibit more complex spatial heterogeneity. Traditional seismic ground motion studies are mainly based on the free-field assumption, analyzing and predicting ground motions through site soil parameters and seismic input characteristics. However, numerous studies have shown that there is a significant interaction effect between building structures and foundation soil, namely, soil-structure interaction (SSI), which alters the propagation path and energy distribution of seismic waves. Building on this, with the improvement of computing power, the finite element method (FEM) numerical simulation is widely used in seismic ground motion analysis. Simultaneously, machine learning methods are gradually being introduced into the field of earthquake engineering to establish complex nonlinear mapping relationships between input parameters and seismic ground motion responses, thereby improving prediction efficiency and accuracy.

[0003] Despite some progress in seismic field prediction, existing technologies still have significant shortcomings. Traditional methods are mostly based on free-field assumptions or single-structure analysis, neglecting or simplifying building-site interactions in multi-building environments. This makes it difficult to accurately reflect the spatial disturbance effect of building clusters on the seismic field, leading to deviations between predictions and actual conditions. While numerical simulation-based methods can characterize complex physical processes, they are computationally expensive and highly sensitive to parameters, making efficient application difficult on a large scale or under multiple conditions. Some existing machine learning methods are primarily designed for single-point or free-field seismic prediction, lacking effective representation of the spatial distribution characteristics of building clusters and failing to establish a unified modeling mechanism between building cluster characteristics and seismic field response, resulting in insufficient model generalization ability. When dealing with the influence of multiple buildings, existing methods typically employ linear superposition or empirical correction methods, which struggle to accurately describe the nonlinear coupling relationships between buildings, thus limiting the improvement of prediction accuracy. Therefore, how to introduce machine learning methods to construct a unified nonlinear mapping model, considering building-site interactions, to achieve efficient and accurate prediction of seismic fields under multi-building conditions has become an urgent technical problem to be solved. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is that existing seismic ground motion prediction methods have insufficient prediction accuracy due to neglecting the influence of building clusters on the seismic ground motion, are difficult to effectively characterize the spatial distribution of multiple buildings and their nonlinear coupling relationship, have high computational costs and insufficient model generalization ability, and how to construct a nonlinear seismic ground motion prediction model based on machine learning under the condition of considering building-site effects to achieve efficient and accurate prediction.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a seismic field prediction method considering building-site effects, comprising acquiring combined data of different site parameters, seismic motion parameters, and building parameters and constructing a training sample dataset; uniformly encoding the building parameters according to their spatial distribution to form building cluster input features; constructing a machine learning model based on the training sample dataset; using the building cluster input features, site parameters, and seismic motion parameters as input to the machine learning model to learn the nonlinear mapping relationship of the seismic field under multiple building conditions; inputting the site parameters, seismic motion parameters, and building parameters of the area to be predicted into the machine learning model; and outputting the seismic field response results of the target area based on the machine learning model.

[0007] As a preferred embodiment of the seismic field prediction method considering building-site effects described in this invention, the construction of the training sample dataset includes: combining and sampling site parameters, seismic motion parameters, and building parameters through a preset parameter space; inputting each parameter combination into a numerical analysis model to obtain the corresponding seismic response results; establishing the correspondence between input parameters and output response; standardizing all sample data to ensure that data of different dimensions are on a uniform scale; and triggering an outlier identification and removal mechanism when abnormal shifts occur in the data distribution, thereby completing the construction of the training sample dataset.

[0008] As a preferred embodiment of the seismic field prediction method considering building-site effects described in this invention, the method for forming the building cluster input features includes: acquiring the spatial location information, geometric parameters, and dynamic parameters of multiple buildings within the study area; calculating the degree of mutual influence between buildings based on the spatial distance relationship between them; performing unified feature mapping on the geometric parameters and dynamic parameters of each building, and generating the overall features of the building cluster through weighted aggregation; performing weighted processing on the features of each building based on the degree of influence, and uniformly aggregating all building features to form the overall input features of the building cluster; when the number of buildings exceeds a preset threshold, dividing the building cluster into several sub-regions according to their spatial location, performing feature aggregation on the buildings in each sub-region, and then merging them to obtain the building cluster input feature vector.

[0009] As a preferred embodiment of the seismic field prediction method considering building-site effects described in this invention, the construction of the machine learning model includes: selecting a machine learning model with nonlinear fitting capabilities as the seismic field prediction model, and using the relationship between input features and output seismic response as the learning objective; dividing the sample data into training data and validation data, and monitoring the prediction error in real time during the training process of the seismic field prediction model; automatically triggering hyperparameter optimization or parameter initialization of the seismic field prediction model when the validation error exceeds a preset positive threshold; gradually optimizing the parameters of the seismic field prediction model through continuous iterative training, so that the prediction result gradually approaches the true response; and determining that the seismic field prediction model has reached convergence and stopping training when the error change of the seismic field prediction model tends to stabilize during the training process, thus completing the construction of the machine learning model.

[0010] As a preferred embodiment of the seismic field prediction method considering building-site effects described in this invention, the nonlinear mapping relationship includes: integrating building cluster characteristics, site parameters, and seismic parameters to form complete input data and inputting it into a machine learning model; coupling various input parameters through a nonlinear mapping mechanism within the machine learning model; establishing a correspondence between building cluster characteristics and seismic field response to characterize the nonlinear variation law of the seismic field under multi-building conditions.

[0011] As a preferred embodiment of the seismic field prediction method considering building-site effects described in this invention, the step of inputting the site parameters, seismic motion parameters, and building parameters of the area to be predicted into the machine learning model includes: acquiring site conditions and building distribution information of the target area, performing feature processing and standardization operations; when the input data is missing or incomplete, estimating and completing the missing information based on data from neighboring areas or statistical patterns; preferentially using the weighted average of neighboring area data as the completion result, and when neighboring data is insufficient, using the statistical mean under similar site conditions for completion; inputting the processed input data into the machine learning model for prediction calculation; during the prediction process, imposing physical rationality constraints on the output results of the machine learning model to ensure that the output results meet the preset value range and spatial continuity requirements.

[0012] As a preferred embodiment of the seismic field prediction method considering building-site effects described in this invention, the step of outputting the seismic field response results of the target area based on the machine learning model includes: mapping the discrete seismic response results output by the machine learning model to a spatial grid and establishing a spatial distribution of the seismic field; selecting an interpolation method based on the distance relationship between adjacent grid nodes; using a linear interpolation method to generate transition values ​​when a node is located within an adjacent grid cell; and using a weighted interpolation method based on distance attenuation to calculate values ​​when a node spans multiple grid cells; during the generation of a continuous distribution, when abrupt changes or discontinuities are detected in a local area, a neighborhood averaging method is used for smoothing and correction of abnormal areas; finally, a continuous and stable seismic field distribution result is formed, and the corresponding seismic parameters are output.

[0013] As a preferred embodiment of the seismic field prediction system considering building-site effects described in this invention, it includes: a data processing module, a model training module, and a prediction output module; the data processing module is used to acquire combined data of different site parameters, seismic motion parameters, and building parameters and construct a training sample dataset, and uniformly encode the building parameters according to their spatial distribution to form building cluster input features; the model training module is used to construct a machine learning model based on the training sample dataset, and uses the building cluster input features, site parameters, and seismic motion parameters as input to the machine learning model to learn the nonlinear mapping relationship of the seismic field under multiple building conditions; the prediction output module is used to input the site parameters, seismic motion parameters, and building parameters of the area to be predicted into the machine learning model, and output the seismic field response results of the target area according to the machine learning model.

[0014] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement a method for predicting seismic fields taking into account building-site effects.

[0015] 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 seismic ground fields taking into account building-site effects.

[0016] The beneficial effects of this invention are as follows: The seismic ground motion prediction method considering building-site effects provided by this invention achieves a unified expression of multi-source factors by constructing a training sample dataset; it introduces a spatial coding and influence weighting mechanism for building clusters to effectively characterize the interaction between buildings, compensating for the shortcomings of traditional methods that ignore building cluster effects; further, it models the nonlinear relationships of multiple factors through machine learning models to achieve efficient prediction of seismic ground motion response; in the prediction stage, it improves the model's adaptability and result reliability in complex engineering environments through consistent data processing, missing value completion, and physical constraint mechanisms; simultaneously, it transforms discrete prediction results into continuous seismic ground motion fields through spatial interpolation and smoothing, further outputting a set of seismic ground motion parameter results. This invention not only improves the accuracy and efficiency of seismic ground motion prediction but also enhances the ability to express building-site coupling effects and has good engineering applicability and promotional value. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is an overall flowchart of a seismic field prediction method considering building-site effects provided in Embodiment 1 of the present invention.

[0019] Figure 2 This is a modeling training diagram for a ground motion field prediction method considering building-site effects provided in Embodiment 1 of the present invention.

[0020] Figure 3 This is a schematic diagram of a computer device for a seismic field prediction method considering building-site effects, as provided in Embodiment 3 of the present invention. Detailed Implementation

[0021] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0022] Example 1, referring to Figures 1-2 As an embodiment of the present invention, a method for predicting seismic ground motion fields considering building-site effects is provided, comprising:

[0023] S1: Obtain combined data of different site parameters, ground motion parameters and building parameters and construct a training sample dataset. Encode the building parameters according to their spatial distribution to form the input features of the building group.

[0024] Furthermore, constructing the training sample dataset includes: combining and sampling site parameters, ground motion parameters, and building parameters through a preset parameter space; inputting each parameter combination into a numerical analysis model to obtain the corresponding ground motion response results; establishing the correspondence between input parameters and output responses; standardizing all sample data to ensure that data of different dimensions are on a uniform scale; and triggering an outlier identification and removal mechanism when abnormal shifts occur in the data distribution, thus completing the construction of the training sample dataset.

[0025] It should also be noted that a preferred approach to constructing the training sample dataset specifically includes, firstly, combining and sampling site parameters, seismic motion parameters, and building parameters within a predefined parameter space. The predefined parameter space is set based on the geological conditions, building type distribution, and seismic motion input range of the target study area to ensure that the sampled data covers common engineering scenarios while reflecting the differences in seismic field response under different parameter combinations. For the first... 1 sample, original input parameter vector , is represented as:

[0026]

[0027] in, Indicates the first The original input parameter vector of each sample; Indicates the sample number. , Indicates the total number of samples; Indicates the first The site parameter vector corresponding to each sample Indicates the first The ground motion parameter vector corresponding to each sample Indicates the first Each sample corresponds to a set of building parameters. The parameter sampling process is used not only to form input combinations under different working conditions, but also to establish the basic coverage of the sample distribution. By combining site, ground motion, and building parameters, the dataset can simultaneously reflect the comprehensive impact of "changes in site conditions," "changes in input ground motion," and "changes in building distribution" on the target ground motion field, providing sufficient data support for machine learning models to learn complex coupling effects. After obtaining the input parameter combinations, each set of parameter combinations is input into the numerical analysis model to obtain the corresponding ground motion response results. Based on site parameters Determine soil dynamic property parameters based on structural parameters Construct a structural dynamic model to obtain the system mass matrix. Damping matrix Stiffness matrix According to seismic motion parameters Structural seismic input excitation A dynamic analysis model based on building-site interaction is constructed, represented as follows:

[0028]

[0029] in, Represents a time variable. Indicates the first A sample at time... The acceleration response vector, Indicates the first A sample at time... The velocity response vector, Indicates the first Each sample at time... The displacement response vector; using Newmark- The method performs time integration and updates the displacement response iteratively, as follows:

[0030]

[0031] in, The time step is represented by . By discretely integrating over the entire duration of the seismic action, the complete dynamic response time history of each computational node within the target area can be obtained. After obtaining the response time history, key response indicators characterizing the seismic motion are extracted to construct the output response vector, represented as:

[0032]

[0033] in, Indicates the first The output response vector corresponding to each sample Indicates the first Peak ground acceleration for each sample, Indicates peak ground speed. Indicates the first The maximum value of the displacement response time history of each sample. Through the seismic response results, a one-to-one correspondence between each set of input parameters and the response results can be established, forming the original training sample dataset. , is represented as:

[0034]

[0035] Since different features in the original samples typically have different physical dimensions and numerical ranges, failing to standardize the scale can easily lead to certain high-order features unreasonably dominating the learning results during the training of the seismic field prediction model. Therefore, after establishing the original sample set, the input features are standardized. For any input feature, the corresponding standardization result is... , is represented as:

[0036]

[0037] in, Represents the original dataset The Middle One original input feature value, express The corresponding standardized results This represents the feature index, covering all input dimensions from the set of site parameters, seismic motion parameters, and building parameters. Indicates the first Features in the original dataset The mean of the middle, Indicates the first Features in the original dataset The standard deviation is calculated. After processing, a standardized sample set is obtained. The standardized features are mapped to a uniform numerical scale, which is beneficial for the subsequent stable training of the seismic ground motion prediction model within a common feature space.

[0038] Furthermore, to improve data quality and avoid interference from outliers in the training process, outlier identification is performed on the standardized samples. For the first... In the nth sample One feature, through standard scores Outlier identification is represented as follows:

[0039]

[0040] in, Indicates the first The first sample The original input feature values ​​of each feature. express Standard scores of the original input feature values. When satisfying... At that time, the judgment of the first The sample at the th Abnormal offsets occur in several feature dimensions, among which, This is the anomaly detection threshold. If a sample exhibits an abnormal offset in any input dimension, it is marked as an anomaly and removed from the standardized sample set. This yields the final training sample dataset. , is represented as:

[0041]

[0042] in, This represents the training sample dataset after standardization and outlier handling.

[0043] It should be noted that the process of forming the building cluster input features includes: acquiring the spatial location information, geometric parameters, and dynamic parameters of multiple buildings within the study area; calculating the degree of mutual influence between buildings based on their spatial distance relationships; performing unified feature mapping on the geometric and dynamic parameters of each building, and generating the overall features of the building cluster through weighted aggregation; weighting the features of each building based on the degree of influence, and then uniformly aggregating all building features to form the overall input features of the building cluster; when the number of buildings exceeds a preset threshold, dividing the building cluster into several sub-regions based on their spatial location, aggregating the features of the buildings in each sub-region separately, and then merging them to obtain the building cluster input feature vector.

[0044] It should also be noted that a preferred approach to forming the input features of a building cluster specifically involves, after constructing the training sample dataset, further structuring the building parameters. Building parameters are typically composed of discrete attributes of multiple individual buildings; directly concatenating these attributes does not effectively reflect the spatial distribution relationships and group interaction characteristics between buildings. Therefore, it is necessary to re-encode the building set in each sample group to form input features that reflect the overall impact of the building cluster. For the first... The nth sample, the nth The parameter vector of a building , is represented as:

[0045]

[0046] in, For building indexing, , Indicates the first The number of buildings in each sample Indicates the first In the nth sample The parameter vector of a building, Indicates the first In the nth sample The height of the building Indicates the first In the nth sample The equivalent stiffness of a building. Indicates the first In the nth sample Equivalent damping of a building; Indicates the first In the nth sample The spatial coordinate vectors of each building are used. To characterize the potential impact of spatial proximity between buildings, the calculation of inter-building distance and the degree of impact is introduced. For the first... Any two buildings in the sample and spatial distance , is represented as:

[0047]

[0048] in, Indicates the first In the nth sample The spatial coordinate vector of a building. Representation and architecture The corresponding building index has a value of and ; Indicates the first Buildings in each sample With architecture The Euclidean distance between them. Based on the engineering understanding that the building-site effect gradually decreases with increasing spatial distance, the degree of mutual influence between buildings. , is represented as:

[0049]

[0050] in, Indicates the first Buildings in each sample For architecture Influence weight; This is the distance attenuation coefficient, used to adjust the rate at which the spatial distance weakens the influence of building interactions; the greater the spatial distance, the weaker the coupling effect. Through a continuous attenuation method, feature construction becomes smoother and more consistent with actual physical laws. Based on the degree of influence, the [missing value] can be further calculated. The overall influence weight of an individual building within a building complex , is represented as:

[0051]

[0052] in, Indicates the first In the nth sample The cumulative impact of one building on the remaining buildings.

[0053] After obtaining the influence weights between buildings, a unified feature mapping is performed on the parameters of individual buildings. The original building parameters are converted into a unified feature vector, represented as:

[0054]

[0055] in, Indicates the first In the nth sample The feature vector after mapping each building; This represents the building feature mapping function. Further, based on the comprehensive influence weights, the features of individual buildings are weighted and aggregated to obtain the overall feature vector of the building complex. , is represented as:

[0056]

[0057] in, Indicates the first The overall feature vector of a building cluster for each sample. When the number of buildings exceeds a preset threshold. At that time, the building complex is spatially partitioned. Specifically, when the following conditions are met... At that time, the study area was divided into sections based on the building's floor plan location. Each subregion corresponds to a subset of buildings. ,in, Indicates a sub-region index. , Indicates the first The number of sub-regions obtained by dividing the building complex in each sample Indicates the first In the nth sample The set of building indexes contained in each subregion. For the... Local features of each sub-region , is represented as:

[0058]

[0059] in, Indicates the first In the nth sample The local building features of each sub-region are first preserved, and then the features of all sub-regions are merged to obtain the global building cluster features. , is represented as:

[0060]

[0061] After generating the building complex features, they will be concatenated with the corresponding site parameters and seismic motion parameters to form the final input vector for the seismic motion field prediction model. , is represented as:

[0062]

[0063] in, Indicates the first These samples are ultimately used as input vectors for training the seismic motion field prediction model. This results in the final training dataset:

[0064]

[0065] in, This indicates the number of valid samples remaining after outlier removal. This represents the final training dataset.

[0066] It should also be noted that this invention constructs a training sample dataset by acquiring combined data of different parameters, achieving unified modeling of multi-source influencing factors; it performs combined sampling through a preset parameter space, enabling the samples to cover various geological conditions, seismic inputs, and building distribution scenarios, establishing a representative input space; it generates corresponding seismic response results through a numerical analysis model, realizing the mapping relationship between input parameters and output responses; it standardizes the sample data, mapping different dimensional features to a unified scale, avoiding the problem of feature dominance imbalance; and it introduces an outlier identification and removal mechanism to improve data quality. By spatially structuring and encoding building parameters, it establishes an overall feature representation of the building complex, and through weighted aggregation and partitioning mechanisms, it achieves effective compression and representation of the spatial distribution features of the building complex. This provides a sufficient and reliable data foundation for the model to learn complex building-site coupling effects, improving the model training effect and physical consistency.

[0067] S2: Based on the training sample dataset, a machine learning model is built. The input features of the building complex, site parameters, and ground motion parameters are used as inputs to the machine learning model to learn the nonlinear mapping relationship of the ground motion field under multiple building conditions.

[0068] Furthermore, the construction of the machine learning model includes: selecting a machine learning model with nonlinear fitting capabilities as the seismic field prediction model, using the relationship between input features and output seismic response as the learning objective; dividing the sample data into training data and validation data, and monitoring the prediction error in real time during the training process of the seismic field prediction model; automatically triggering hyperparameter optimization or parameter initialization of the seismic field prediction model when the validation error exceeds a preset positive threshold; gradually optimizing the parameters of the seismic field prediction model through continuous iterative training, so that the prediction results gradually approach the true response; and determining that the seismic field prediction model has reached convergence and stopping training when the error change of the seismic field prediction model tends to stabilize during the training process, thus completing the construction of the machine learning model.

[0069] It should also be noted that a preferred approach to constructing a machine learning model specifically includes building a seismic field prediction model to approximate the relationship between the input and output, expressed as:

[0070]

[0071] in, This indicates that the seismic ground motion prediction model is effective for the first... The predicted output vector for the nth sample. To express complex nonlinear relationships, the seismic ground motion prediction model is constructed as a multi-layer nonlinear mapping structure. For the nth sample... Layer, output features , is represented as:

[0072]

[0073] in, Indicates the first The sample at the th The output feature vector of the layer, Indicates the first One sample in The output feature vector of the layer, when hour, Input for the seismic ground motion prediction model Indicates the first Layer weight matrix vector This represents the bias vector. It should be noted that the first layer primarily performs a linear combination of the original input; the middle layers progressively extract higher-order features; and the final layer outputs the prediction result. Therefore, the multi-layer nonlinear mapping structure can progressively transform the multi-source information in the original input into high-dimensional features that can interpret the seismic response. After the structure of the seismic field prediction model is determined, the parameters of the seismic field prediction model are trained using sample data. The dataset... Divided into training set and validation set:

[0074]

[0075] in, Indicates the number of training samples. This indicates the number of validation samples. The training set is used to update the parameters of the seismic ground motion prediction model, and the validation set is used to evaluate the generalization ability of the model. To quantify the difference between the model's predictions and the actual results, a loss function is introduced. , is represented as:

[0076]

[0077] This reflects the overall prediction error of the seismic ground motion prediction model under the current parameter conditions. A smaller value indicates a higher degree of model fit. During training, the parameters of the seismic ground motion prediction model are updated iteratively. The parameter update rule for the seismic ground motion prediction model in the next iteration is expressed as follows:

[0078]

[0079] in, express Parameters of the seismic field prediction model in the next iteration. express Parameters of the seismic field prediction model in the next iteration. The learning rate controls the step size for each update. This represents the iteration number index. The parameter update process combines the coupling relationship between building cluster features and site parameters to constrain or adjust the gradient direction. After each round of training, the performance of the seismic field prediction model is evaluated using a validation set.

[0080]

[0081] When satisfied hour, This indicates a preset positive threshold, signifying a significant increase in the error of the seismic ground motion prediction model on the validation set. In this case, adjusting the learning rate or reinitializing some parameters allows the model to re-enter a stable training state. As training progresses, when the following condition is met... This indicates that the loss function has stabilized and the parameters of the seismic ground motion prediction model have basically converged. Indicates the first The training loss function for the next iteration. Indicates the first The training loss function for the next iteration. This represents the threshold value for the change in the loss function, used to determine whether the seismic field prediction model has reached convergence. Training stops when the seismic field prediction model reaches convergence, resulting in a machine learning model.

[0082] It should be noted that the nonlinear mapping relationship includes: unifying and integrating building cluster characteristics, site parameters, and seismic motion parameters to form complete input data and inputting it into the machine learning model; coupling various input parameters through a nonlinear mapping mechanism within the machine learning model; establishing the correspondence between building cluster characteristics and seismic field response to characterize the nonlinear variation law of seismic field under multi-building conditions.

[0083] It should also be noted that a preferred scheme for the nonlinear mapping relationship specifically includes, after completing the training of the machine learning model, further explaining from a mechanistic perspective how the machine learning model achieves nonlinear mapping under multiple building conditions, with the input data uniformly represented as follows:

[0084]

[0085] By unifying the input data representation, the dispersed multi-source information is converted into a standardized input form that can be directly received by the machine learning model, providing a unified entry point for nonlinear coupling operations. After obtaining the input vector, the machine learning model does not directly map the linear data to the output, but first performs multiple layers of nonlinear transformation internally to extract the coupling features between different factors. This invention, through multiple layers of nonlinear mapping, gradually transforms the unified input vector into hidden features that can characterize deep coupling relationships, expressed as:

[0086]

[0087] in, Indicates the first The hidden features of the machine learning model for each sample. After obtaining the hidden features of the machine learning model, the model further establishes an overall mapping relationship between multi-source inputs and ground motion response, used to characterize the variation of the ground motion response of the target area with input factors under the condition of considering building-site effects. The overall mapping relationship is expressed as:

[0088]

[0089] Based on the overall mapping relationship, the machine learning model not only processes discrete samples but also forms a class of continuous prediction rules where the input determines the output. This can be used to describe how the seismic field response changes under different site conditions, different seismic inputs, and different building layouts. After establishing the overall mapping relationship, this invention can be applied to the area to be predicted. Following the same input organization method as the training phase, the relevant parameters of the target area are organized to ensure that the input data for prediction is structurally consistent with the input data for training. When predicting the target area, the site parameters, seismic parameters, and building features are also concatenated into an input feature vector, represented as:

[0090]

[0091] in, This represents the output vector corresponding to the target region. Represents the target area site parameter vector; This represents the input ground motion parameter vector; This represents the feature vector of the building complex in the target area.

[0092] It should also be noted that this invention selects a machine learning model structure with nonlinear fitting capabilities, enabling the machine learning model to characterize the high-dimensional coupling laws of seismic field responses under multiple building conditions. By dividing the samples into training and validation sets, and monitoring the prediction error in real time during training, the generalization ability of the machine learning model is dynamically evaluated. When the validation error exceeds a preset threshold, a hyperparameter optimization or parameter initialization mechanism is automatically triggered, allowing the machine learning model to reach a local optimum. By iteratively updating the machine learning model parameters and gradually reducing the loss function, the predicted results are made closer to the true response. Simultaneously, the convergence condition is determined by error changes, ensuring the stability and termination of the machine learning model training process. Furthermore, by constructing a multi-layer nonlinear mapping structure, the input features are coupled layer by layer within the machine learning model, and feature extraction is achieved, fully expressing the complex relationship between the spatial effects of building clusters and site responses. This realizes the transformation from structured data to a high-precision prediction model, giving the model not only fitting capabilities but also good stability and generalization performance.

[0093] S3: Input the site parameters, seismic motion parameters and building parameters of the area to be predicted into the machine learning model, and output the seismic response results of the target area based on the machine learning model.

[0094] Furthermore, the site parameters, seismic motion parameters, and building parameters of the area to be predicted are input into the machine learning model. This includes obtaining site conditions and building distribution information of the target area, performing feature processing and standardization. When the input data is missing or incomplete, the missing information is estimated and completed based on data from neighboring areas or statistical patterns. The weighted average of neighboring area data is preferentially used as the completion result, and when neighboring data is insufficient, the statistical mean under similar site conditions is used for completion. The processed input data is then input into the machine learning model for prediction calculation. During the prediction process, physical rationality constraints are imposed on the output results of the machine learning model to ensure that the output results meet the preset value range and spatial continuity requirements.

[0095] It should also be noted that a preferred approach for inputting site parameters, seismic motion parameters, and building parameters of the area to be predicted into the machine learning model specifically includes the following: since standardized data is used during the training phase of the machine learning model, a consistent data processing method must be adopted during the prediction phase. The standardized result corresponding to any input feature is expressed as:

[0096]

[0097] in, The target region input vector after standardization is represented as the first... 1 eigenvalue, The unnormalized nth element in the target region input vector Each feature value is used; through standardization, the original input in the prediction stage is mapped to the same feature space as that in the training stage, providing a unified input basis for subsequent machine learning model inference. In practical engineering, some regions may have missing parameters or incomplete data. Weighted estimation is performed using corresponding features from neighboring training samples, expressed as:

[0098]

[0099] in, Indicates the neighboring region's first The training sample of the th training sample 1 eigenvalue, Indicates the neighboring training sample number. This represents the number of nearby reference samples. To further reflect the difference in the impact of nearby samples on the completion result, weights are constructed based on the spatial distance between the target region and each nearby sample. Indicates the first The weight coefficients of each sample are expressed as:

[0100]

[0101] in, Indicates the target area and the first Spatial distance between training samples This is the distance attenuation coefficient. When... hour, The threshold for the number of neighboring reference samples is represented by the statistical mean, and is expressed as follows:

[0102]

[0103] in, Indicates the first The average value of each feature in the training data. After completing the input construction, the standardized target region input vector is input into the trained machine learning model to obtain the initial discrete seismic response result of the target region, expressed as:

[0104]

[0105] in, This represents the discrete ground motion response output of the machine learning model. While the machine learning model output reflects the nonlinear mapping relationship between the input parameters and the ground motion response, because machine learning models are inherently data-driven, they may produce abnormal predictions that do not conform to engineering reality, or even exceed the physically permissible range, when the local sample distribution is sparse or the input features exceed the training distribution range. Directly using the prediction results may lead to unreasonable abrupt changes or distortions in subsequent ground motion field construction. Therefore, a physical rationality constraint step is introduced after the machine learning model prediction to restrict the machine learning model output results. While preserving the nonlinear mapping capability of the machine learning model, engineering experience or physical boundary conditions are introduced to correct the prediction results, improving the stability and engineering usability of the results. The first... Each component is subjected to dimension-wise constraint processing, represented as:

[0106]

[0107] in, Indicates the constraint after the first Discrete ground motion response results for each output component The first Discrete ground motion response results for each output component Indicates the first The maximum allowed value for each output component Indicates the first The minimum allowed value of each output component; through constraint processing, the output results of the machine learning model are made to satisfy the nonlinear mapping relationship and at the same time conform to the physical value range of the seismic response, thereby avoiding the impact of local abnormal prediction values ​​on the subsequent spatial distribution construction and improving the engineering reliability and stability of the prediction results.

[0108] It should be noted that the output of the seismic ground response results of the target area based on the machine learning model includes mapping the discrete seismic ground response results output by the machine learning model to a spatial grid and establishing the spatial distribution of the seismic ground field; between adjacent grid nodes, the interpolation method is selected according to the distance relationship. When the node is located within an adjacent grid cell, a linear interpolation method is used to generate transition values. When the node spans multiple grid cells, a weighted interpolation method based on distance attenuation is used for calculation; during the generation of the continuous distribution, when abrupt changes or discontinuities are detected in a local area, a neighborhood averaging method is used for smoothing and correction of abnormal areas; finally, a continuous and stable seismic ground field distribution result is formed, and the corresponding seismic ground parameters are output.

[0109] It should also be noted that a preferred approach based on the seismic response results of the target area output by the machine learning model specifically includes: dividing the target area into regular grids; establishing a unified spatial bearing capacity framework for the seismic response results, so that discrete predicted values ​​can correspond one-to-one with specific spatial locations, facilitating interpolation calculations and visualization of regional distribution. The target area is divided into multiple regular grid nodes, and all grid nodes are organized into a node set. , is represented as:

[0110]

[0111] in, Indicates the first The coordinates of each grid node Indicates the grid node number, , This represents the total number of grid nodes. After grid generation, the obtained machine learning model predictions are mapped to the corresponding grid nodes. Although the output of the machine learning model provides discrete response results for the target region, only by assigning values ​​to specific grid node locations can the response values ​​have a clear spatial meaning. Specifically, for the _ The seismic response value of each grid node is expressed as:

[0112]

[0113] Indicates the first The seismic response values ​​of individual grid nodes are used; however, in practical engineering, it is often necessary to obtain seismic response values ​​at arbitrary spatial locations. Spatial interpolation is introduced based on the node response values ​​to estimate the seismic response at any location based on the response values ​​of neighboring nodes. This extends the discrete node values ​​to a continuous spatial field, enabling the output of this invention to cover any location within the target area, not just the grid nodes. For any spatial location... The seismic response is obtained through interpolation calculation. When located between adjacent grid nodes, linear interpolation is used, expressed as:

[0114]

[0115] in, express The set of adjacent grid nodes, Indicates the first Linear weights for each grid node. When the position... Without being affected by local adjacent nodes, linear interpolation struggles to fully reflect distance attenuation patterns when spanning multiple grid cells. Seismic ground motion responses typically exhibit significant distance correlation during spatial propagation; closer nodes have a stronger influence on the target location, while the influence of farther nodes should gradually decrease. Based on this consideration, this invention employs distance attenuation-weighted interpolation, expressed as:

[0116]

[0117] in, Represents grid nodes With grid nodes distance, The distance decay exponent is used to adjust the weight of the influence of neighboring nodes on the interpolation result. After obtaining continuous interpolation results, it is necessary to further consider local anomalies in the spatial field construction. Machine learning prediction errors, node mapping errors, and the interpolation algorithm itself can all cause abrupt changes or discontinuities in local areas, especially in areas with complex building distributions and drastic changes in site conditions. Without local checks on the continuous field, unreasonable spikes or jumps may appear in the final seismic field. Therefore, a spatial gradient detection step is added after continuous field construction to identify anomalous areas with excessive local variations; for the... The degree of local variation of each grid node is calculated and expressed as:

[0118]

[0119] in, Indicates the first The seismic response values ​​of each neighboring node. Indicates the neighboring node number, Represents grid nodes The neighborhood set; when the spatial gradient at a node exceeds a preset threshold, it indicates an unreasonable local abrupt change, requiring smoothing correction. Specifically, when the following conditions are met:

[0120]

[0121] When smoothing is needed, a neighborhood averaging method is used:

[0122]

[0123] in, This represents the smoothing threshold, used to determine the degree of spatial variation in the seismic response. Smoothing is triggered when local variations exceed this threshold. After obtaining the smoothed response values ​​for each grid node, a continuous seismic field function is further established within the target region. The interpolated and smoothed spatial response results are uniformly denoted as the continuous seismic field function, expressed as:

[0124]

[0125] By establishing a continuous ground motion field function, the ground motion response distribution corresponding to each spatial location within the target area can be obtained. The continuous field function not only describes the variation of the overall ground motion field with spatial location but also serves as a unified basis for subsequent parameter extraction. After establishing the continuous ground motion field function, ground motion parameters at each grid node or target location within the target area are further extracted to form the final output set. The values ​​of the continuous field function at the coordinates of each grid node are organized into the final output set, as follows:

[0126]

[0127] in, This represents the set of seismic motion parameter results corresponding to each grid node within the target area. Indicates the first The set contains the seismic motion parameter values ​​at the coordinates of each grid node. Furthermore, each element in the set corresponds to a seismic motion parameter value at a specific spatial location within the target area, thus enabling direct use for regional seismic field analysis, building seismic response assessment, hazard zoning, or safety verification of key nodes.

[0128] It should also be noted that this invention achieves rapid prediction of seismic ground motion response by inputting the parameters to be tested into a trained machine learning model; by standardizing the input features, it ensures that the predicted data and training data are in the same feature space, avoiding the machine learning model mismatch problem; when there are missing input data, it improves the adaptability of the machine learning model to incomplete data in actual engineering by combining weighted completion and statistical mean completion; by imposing physical rationality constraints on the machine learning output results, it improves the engineering credibility of the results. By mapping discrete prediction results to a regular grid and combining linear interpolation and distance-weighted interpolation methods, it achieves continuous spatial reconstruction of seismic ground motion response; at the same time, by spatial gradient detection and neighborhood smoothing, it eliminates abrupt changes and ensures the continuity and stability of the seismic ground motion distribution; and it extracts the seismic ground motion parameters of each grid node to form a result set that can be directly applied to engineering analysis. This completes the transformation process from machine learning model output to continuous seismic ground motion field to engineering parameter output, enhancing the engineering application value of this invention.

[0129] Example 2, an embodiment of the present invention, provides a seismic field prediction system considering building-site effects, including a data processing module, a model training module, and a prediction output module.

[0130] The data processing module is used to acquire combined data of different site parameters, ground motion parameters, and building parameters and construct a training sample dataset. The building parameters are uniformly encoded according to their spatial distribution to form the input features of the building group. The model training module is used to construct a machine learning model based on the training sample dataset. The input features of the building group, along with the site parameters and ground motion parameters, are used as inputs to the machine learning model to learn the nonlinear mapping relationship of the ground motion field under multiple building conditions. The prediction output module is used to input the site parameters, ground motion parameters, and building parameters of the area to be predicted into the machine learning model and output the ground motion field response results of the target area based on the machine learning model.

[0131] Example 3, referring to Figure 3 This embodiment also provides a computer device applicable to the seismic field prediction method considering building-site effects, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the seismic field prediction method considering building-site effects as proposed in the above embodiment.

[0132] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0133] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the seismic field prediction method considering building-site effects as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

Claims

1. A method for predicting seismic ground motion fields considering building-site effects, characterized in that, include: Acquire combined data of different site parameters, seismic motion parameters and building parameters and construct a training sample dataset. Unify the encoding of building parameters according to spatial distribution to form the input features of building groups. A machine learning model is built based on the training sample dataset. The input features of the building complex, site parameters, and ground motion parameters are used as inputs to the machine learning model to learn the nonlinear mapping relationship of the ground motion field under multiple building conditions. The site parameters, seismic motion parameters, and building parameters of the area to be predicted are input into the machine learning model, and the seismic response results of the target area are output based on the machine learning model.

2. The seismic field prediction method considering building-site effects as described in claim 1, characterized in that: The construction of the training sample dataset includes, Site parameters, seismic motion parameters, and building parameters are sampled in combination using a preset parameter space. Each set of parameters is input into the numerical analysis model to obtain the corresponding seismic response results. Establish the correspondence between input parameters and output responses; All sample data are standardized to bring data of different dimensions to a uniform scale; When an abnormal shift occurs in the data distribution, an outlier identification and removal mechanism is triggered to complete the construction of the training sample dataset.

3. The seismic field prediction method considering building-site effects as described in claim 2, characterized in that: The input features for forming a building complex include, Acquire spatial location information, geometric parameters, and dynamic parameters of multiple buildings within the study area; Calculate the degree of mutual influence between buildings based on the spatial distance relationship between them; The geometric and dynamic parameters of each building are uniformly mapped, and the overall features of the building complex are generated through weighted aggregation. Based on the degree of influence, each building feature is weighted and processed, and all building features are aggregated to form the overall input features of the building group. When the number of buildings exceeds a preset threshold, the building complex is divided into several sub-regions based on spatial location. The features of the buildings in each sub-region are aggregated separately and then merged to obtain the building complex input feature vector.

4. The seismic field prediction method considering building-site effects as described in claim 3, characterized in that: The construction of the machine learning model includes, A machine learning model with nonlinear fitting capability was selected as the seismic ground motion prediction model, and the relationship between input features and output seismic ground motion response was taken as the learning objective. The sample data is divided into training data and validation data, and the prediction error is monitored in real time during the training process of the seismic field prediction model. When the verification error exceeds a preset positive threshold, the hyperparameter optimization or parameter initialization of the seismic field prediction model is automatically triggered. By continuously iterating and training, the parameters of the seismic ground motion prediction model are gradually optimized, so that the prediction results gradually approach the real response. During the training process of the seismic ground motion prediction model, when the error of the seismic ground motion prediction model tends to stabilize, it is determined that the seismic ground motion prediction model has reached the convergence state and training is stopped, thus completing the construction of the machine learning model.

5. The seismic field prediction method considering building-site effects as described in claim 4, characterized in that: The nonlinear mapping relationship includes, The characteristics of the building complex, site parameters, and seismic motion parameters are integrated to form complete input data, which is then fed into the machine learning model. The machine learning model uses a non-linear mapping mechanism to couple various input parameters. Establish the correspondence between building cluster characteristics and seismic field response, and characterize the nonlinear variation law of seismic field under multi-building conditions.

6. The seismic field prediction method considering building-site effects as described in claim 5, characterized in that: The process of inputting site parameters, seismic motion parameters, and building parameters of the area to be predicted into the machine learning model includes... Acquire site conditions and building distribution information for the target area, and perform feature processing and standardization operations; When the input data is missing or incomplete, the missing information is estimated and filled in based on data from neighboring regions or statistical patterns. The weighted average of neighboring area data is preferred as the completion result. When neighboring data is insufficient, the statistical mean under similar site conditions is used for completion. The processed input data is then fed into a machine learning model for prediction calculations. During the prediction process, physical rationality constraints are imposed on the output of the machine learning model to ensure that the output meets the preset value range and spatial continuity requirements.

7. The seismic field prediction method considering building-site effects as described in claim 6, characterized in that: The output of the ground motion field response results of the target area based on the machine learning model includes... The discrete ground motion response results output by the machine learning model are mapped to a spatial grid, and the spatial distribution of the ground motion field is established. Between adjacent grid nodes, the interpolation method is selected according to the distance relationship. When the node is located within an adjacent grid cell, a linear interpolation method is used to generate the transition value. When the node spans multiple grid cells, a weighted interpolation method based on distance attenuation is used for calculation. During the generation of a continuous distribution, when abrupt changes or discontinuities are detected in a local area, a neighborhood averaging method is used for smoothing to correct the abnormal areas. The final result is a continuous and stable seismic field distribution, and the corresponding seismic parameters are output.

8. A seismic field prediction system considering building-site effects, employing the seismic field prediction method considering building-site effects as described in any one of claims 1 to 7, characterized in that: It includes a data processing module, a model training module, and a prediction output module; The data processing module is used to acquire combined data of different site parameters, seismic motion parameters and building parameters and construct a training sample dataset. The building parameters are uniformly encoded according to the spatial distribution relationship to form the input features of the building group. The model training module is used to build a machine learning model based on the training sample dataset. It takes the input features of the building complex, site parameters, and ground motion parameters as input to the machine learning model to learn the nonlinear mapping relationship of the ground motion field under multiple building conditions. The prediction output module is used to input the site parameters, seismic motion parameters and building parameters of the area to be predicted into the machine learning model, and output the seismic field response results of the target area based on the machine learning model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the seismic field prediction method considering building-site effects as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the seismic field prediction method considering building-site effects as described in any one of claims 1 to 7.

Images