Subway station earthquake response prediction method based on deep learning

Abstract

The invention discloses a subway station earthquake response prediction method based on deep learning. The method comprises the following steps: firstly, acquiring subway station structure information and site category information required by subway station earthquake time history response prediction, establishing an earthquake oscillation time history data-free field soil body displacement response-subway station earthquake response data set, and preprocessing the data in the data set; training the DB-CNN-Attention network model by using the preprocessed data set, determining the site category of a target subway station, obtaining the structural information of the target subway station, and obtaining seismic oscillation time history data according to the actual demand; and using the trained DB-CNN-Attention network model to output an earthquake time history response data prediction result of the target subway station. The method is higher in prediction precision and high in calculation efficiency, and can be used for post-earthquake damage estimation, earthquake damage scene reconstruction and pre-earthquake earthquake resistance evaluation.

Description

Technical Field

[0001] The present invention relates to the field of subway station earthquake response prediction. More specifically, the present invention relates to a subway station earthquake response prediction method based on deep learning. Background Art

[0002] With the accelerated urbanization process, the rapid expansion of urban subway systems, especially the large-scale construction of subway stations, the seismic safety of their structures has gradually become a research focus in the engineering and academic circles. Due to their unique buried environment and complex stress characteristics, the response behavior of subway stations under earthquakes is significantly different from that of ground buildings. The performance of these subway stations under earthquakes directly affects the scale of economic losses and casualties caused by earthquakes. Therefore, it is urgent to conduct structural seismic response analysis of subway station structures and strengthen the weak parts of seismic resistance found.

[0003] At present, there are three main research methods for predicting the seismic response of subway station structures, namely model tests, theoretical analysis and numerical simulation. Among them, the numerical simulation method is the most important research method. This method establishes a finite element dynamic calculation model based on the geometry, materials and load information of the subway station, and then selects the appropriate seismic time history as the input condition to load it into the finite element dynamic calculation model of the subway station. The nonlinear dynamic time history analysis is performed on it, and the time-varying response of the subway station structure under the action of the earthquake, such as displacement, velocity and acceleration, can be calculated. This method can simulate the complex response behavior of the subway station structure under seismic excitation and provide data support for seismic design. However, the numerical simulation method requires the establishment of a complex finite element model and takes a lot of time to perform nonlinear dynamic time history analysis. In addition, in post-earthquake emergency assessments, due to the long calculation time, it may not meet the high timeliness requirements.

[0004] In recent years, with the development of deep learning methods, more and more studies have applied deep learning methods to the prediction of seismic response of subway station structures, relying on their ability to handle complex problems with high precision and high efficiency, to predict the damage state, peak response and time-history response of station structures. However, existing research in this field still faces the following problems: 1. Some studies only consider the input seismic information of subway stations, which is usually a simple intensity parameter, rather than a seismic time history or time-frequency diagram, which can more comprehensively reflect the characteristics and differences of seismic motion. 2. The research objects of most deep learning models are limited to stations in a specific area, and it is impossible to simultaneously consider the site category information and seismic information of subway stations to predict the comprehensive seismic response. When the study area changes, the model needs to be retrained based on the new data, and lacks generalization ability.

[0005] In summary, the existing subway station earthquake response prediction method has low prediction accuracy due to the limitations of input information, and the model needs to be retrained when adapting to station structures in different regions, and its applicability is poor. Therefore, it is of great significance to propose a new prediction method that combines more comprehensive subway station area information and seismic motion information to improve prediction accuracy and enhance the adaptability of the model. Summary of the invention

[0006] The purpose of the present invention is to solve the problems of low prediction accuracy of the seismic response of subway station structures in existing prediction methods and the need to retrain the model when the prediction object changes, and to propose a subway station seismic response prediction method based on deep learning.

[0007] The technical solution adopted by the present invention to solve the above technical problems is: a subway station earthquake response prediction method based on deep learning, comprising the following steps: Step 1: Obtain the subway station structure information and site category information required for the prediction of the earthquake time history response of the subway station, establish a finite element calculation model and a one-dimensional equivalent linearized site corresponding to the site category for the subway station structure of different site categories; input the seismic time history data into the finite element calculation model, perform nonlinear time history response analysis on various subway station structures, and obtain the seismic response data of the subway station structure of different site categories under each seismic time history data; input the seismic time history data into the one-dimensional equivalent linearized site to calculate the free field soil displacement response data of each seismic time history data under different site categories, and establish the seismic time history data-free field soil displacement response-subway station seismic response data set; Step 2: Preprocess the data in the earthquake time history data-free field soil displacement response-subway station earthquake response data set, build a DB-CNN-Attention network model, and use the preprocessed data set to train the DB-CNN-Attention network model to obtain a trained DB-CNN-Attention network model; Step 3: Determine the site category of the target subway station, obtain the structural information of the target subway station, and obtain the seismic time history data according to actual needs; preprocess the structural information and seismic time history data of the target subway station, and input the preprocessed data into the trained DB-CNN-Attention network model to output the earthquake time history response data prediction results of the target subway station.

[0008] Further preferably, the DB-CNN-Attention network model includes a seismic time-history data encoding module, a free-field soil displacement response data encoding module, a dual-branch feature fusion module, a dual-branch attention module and a decoding module. The seismic time-history data encoding module performs feature extraction on the seismic time-history data to obtain a seismic feature vector. The free-field soil displacement response data encoding module performs feature extraction on the free-field soil displacement response data to obtain a free-field soil displacement response feature vector. The dual-branch feature fusion module fuses the seismic feature vector and the free-field soil displacement response feature vector to obtain a fused feature vector. Subsequently, the fused features are input into the dual-branch attention module for weighting, and the weighted features are input into the decoding module for decoding, so as to finally obtain the earthquake response prediction result of the subway station.

[0009] Further preferably, the seismic time-history data encoding module and the free-field soil displacement response data encoding module have the same structure, including three convolutional pooling modules, the convolutional pooling module is provided with one convolutional layer and one pooling layer, and the convolutional layer is post-activated by a ReLU activation function.

[0010] Further preferably, the dual-branch feature fusion module includes two branches, one branch performs dimensionality reduction processing on the seismic motion feature vector through a convolution layer with a convolution kernel size of 3×1, followed by a batch normalization layer to extract deep features; the other branch performs dimensionality reduction processing on the free-field soil displacement response feature vector through a convolution layer with a convolution kernel size of 2×1, followed by a batch normalization layer to extract deep features, the deep features of the two branches are fused by dot multiplication, and then the channel average is calculated by the mean function mean to generate a correlation matrix, and the activation function Sigmoid is used to evaluate the correlation between features. Finally, according to the value of the correlation matrix, the seismic motion feature or the free-field soil displacement response feature is dynamically selected to obtain the fused feature.

[0011] Further preferably, the dual-branch attention module includes a convolution group, a temporal attention module and a channel attention module; the fused features output by the dual-branch feature fusion module are re-extracted through the convolution group; the temporal attention module first applies the average pooling layer and the maximum pooling layer to reduce the dimension of the fused features and generate two sets of features, then splices the two sets of features, and then compresses the spliced ​​features into single-channel features through the convolution layer, and then normalizes the temporal weights through the Sigmoid activation function to generate temporal attention weights; the channel attention module applies the average pooling layer and the maximum pooling layer in the channel dimension to reduce the dimension of the fused features and generate two sets of features, sums the two sets of features element by element, and then inputs the multi-layer perceptron to model the interdependence between channels, adaptively judge the importance of each channel, and then generates the channel attention weight by the Sigmoid activation function; finally, the temporal attention weight and the channel attention weight are multiplied and applied to the features processed by the convolution group to generate the dual-branch attention enhanced features.

[0012] Further preferably, the decoding module consists of a flattening layer, a fully connected layer and an activation function ReLU which are arranged in sequence.

[0013] Further preferably, the seismic time history data are selected from the Pacific Earthquake Engineering Research Center seismic database using a random sampling method.

[0014] Further preferably, the preprocessing of the seismic time history data includes: unifying the duration and sampling frequency of the seismic time history data; the preprocessing method of the seismic time history response data and the free field soil displacement response data is consistent with this, the training data is normalized, and the training data is scaled to between [0, 1].

[0015] Further preferably, the DB-CNN-Attention network model is optimized by an Adam optimizer.

[0016] The beneficial effects of the present invention are: In the prediction of subway station earthquake response, the present invention takes into account the subway station structure information and site category available in actual scenarios, and uses deep learning algorithms to establish a DB-CNN-Attention network model suitable for subway station earthquake time-history response prediction. Compared with existing methods, the present invention achieves higher prediction accuracy when only using subway station structure information and site category information, and this information is easy to obtain and closer to actual application needs.

[0017] At the same time, the trained DB-CNN-Attention network model takes into account the information of different site categories of subway station structures. By introducing the free-field soil displacement characteristics, the model has strong generalization ability and can be applied to subway station structures of different site categories without retraining, which greatly improves the practical application efficiency of the model. In addition, this method has high computational efficiency and can effectively serve the damage estimation of subway stations after earthquakes, reconstruction of earthquake damage scenes, and seismic capacity assessment of subway stations before earthquakes. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein: Figure 1 This is a schematic diagram of some randomly selected earthquake acceleration response spectra.

[0019] Figure 2 This is a schematic diagram of the DB-CNN-Attention network model.

[0020] Figure 3 It is a schematic diagram of a dual-branch special fusion module.

[0021] Figure 4 It is a schematic diagram of the dual-branch attention module.

[0022] Figure 5 It is the iteration curve of the DB-CNN-Attention network model.

[0023] Figure 6 It is a seismic wave randomly selected to verify the present invention.

[0024] Figure 7 This is the earthquake time-history response prediction effect diagram of a three-story, three-span subway station in a Class II site.

[0025] Figure 8 This is the earthquake time-history response prediction effect diagram of a three-story, three-span subway station in a Class IV site. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0027] It should be understood that when the terms "first", "second", etc. are used in the claims, descriptions, and drawings of the present invention, they are only used to distinguish different objects, rather than to describe a specific order. The terms "include" and "comprise" used in the description and claims of the present invention indicate the presence of the described features, wholes, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or their collections.

[0028] A method for predicting earthquake response of subway stations based on deep learning includes steps one to three, which are described in detail below.

[0029] Step 1: Obtain the subway station structure information and site category information required for the prediction of the subway station seismic time-history response, establish a finite element calculation model and a one-dimensional equivalent linearized site corresponding to the site category for the subway station structure of different site categories, select a number of seismic time-history data from the seismic motion database, and divide them into a training set, a validation set, and a test set; input the selected seismic time-history data into the finite element calculation model, perform nonlinear time-history response analysis on various subway station structures, and obtain the seismic response data of the subway station structures of different site categories under each seismic time-history data; input the seismic time-history data into the one-dimensional equivalent linearized site to calculate the free-field soil displacement response data of each seismic time-history data under different site categories, and establish a seismic time-history data set - free-field soil displacement response - subway station seismic response.

[0030] Seismic waves with different peak ground accelerations (PGA) and different magnitudes are selected from the seismic database. In this embodiment, 160 seismic waves (i.e., seismic time history data) with large differences are randomly selected from the NGA-West2 seismic database (PEER database) of the Pacific Earthquake Engineering Research (PEER) Center. Figure 1 The figure shows some randomly selected earthquake acceleration response spectra. This group of data has large differences in PGA, spectral characteristics and duration, which can be used to verify the generalization ability of the neural network model for predicting different earthquake waves. The PGA range of the obtained earthquake wave is 0.05-0.30 g. In order to simplify the calculation process, the sampling interval of the original earthquake wave acceleration time history is adjusted to 0.02 s. The preprocessed earthquake motion time history data is divided into three data sets: training set, validation set and test set, with data proportions of 70%, 10% and 20% respectively.

[0031] In order to accurately establish one-dimensional equivalent linearized sites corresponding to different site types and site categories, this embodiment needs to collect four types of parameter information corresponding to different site types, such as soil thickness, shear wave velocity, soil density and nonlinear characteristics, and then input these four types of site type parameters into the numerical calculation software to establish the corresponding one-dimensional equivalent linearized site model.

[0032] A typical finite element calculation model is constructed according to the structural parameters of the subway station. The modeling method is as follows: the subway station structure is simulated by plane strain units, the soil components and concrete components are simulated by four-node bilinear plane strain quadrilateral units, the steel bars are simulated by rod units embedded in concrete, and a series of springs are used to simulate the interaction between the soil and the subway station structure during earthquakes. The contact between the soil and the subway station structure adopts hard contact, the contact surface friction coefficient between the soil and the subway station structure is taken as 0.4, and the boundary conditions of the finite element calculation model adopt static and dynamic coupling processing technology.

[0033] The three data sets (training set, validation set and test set) selected above are input into the established finite element calculation model of the subway station structure for nonlinear time history analysis. At the same time, the data sets are input into the one-dimensional equivalent linearized site model for free field soil displacement response calculation, and the inter-layer displacement response and the corresponding free field soil deformation corresponding to the seismic motion record are obtained, and the seismic time history data-free field soil displacement response-subway station seismic response data set is established.

[0034] Step 2: Preprocess the data in the earthquake time history data-free field soil displacement response-subway station earthquake response data set, build a DB-CNN-Attention network model, and use the preprocessed data set to train the DB-CNN-Attention network model to obtain a trained DB-CNN-Attention network model.

[0035] like Figure 2As shown, the DB-CNN-Attention network model includes a seismic time-history data encoding module, a free-field soil displacement response data encoding module, a dual-branch feature fusion module, a dual-branch attention module, and a decoding module. The seismic time-history data encoding module extracts features from the seismic time-history data to obtain a seismic feature vector. The free-field soil displacement response data encoding module extracts features from the free-field soil displacement response data to obtain a free-field soil displacement response feature vector. The dual-branch feature fusion module fuses the seismic feature vector and the free-field soil displacement response feature vector to obtain a fused feature vector. Subsequently, the fused features are input into the dual-branch attention module for weighting, and the weighted features are input into the decoding module for decoding, and finally the subway station earthquake response prediction results are obtained. The decoding module consists of a flattening layer, a fully connected layer, and an output layer. In this embodiment, the number of hidden units in the fully connected layer is set to 128, and the network structure is an input layer dimension: [1,2000], and an output layer output dimension [1,2000].

[0036] The structure of the earthquake time history data encoding module and the free field soil displacement response data encoding module is the same, including three convolutional pooling modules, each of which is equipped with one convolutional layer and one pooling layer. The convolutional layer is post-activated by ReLU, and the convolution kernel size of the convolutional layer is set to 3×1; the size of the pooling kernel of the pooling layer is 2×1, the step size is 1, and the pooling operation is the maximum pooling. The number of kernels of the convolutional layers in the three convolutional pooling modules is set to 32, 64, and 128 respectively.

[0037] like Figure 3 As shown in the figure, the dual-branch feature fusion module includes two branches. One branch reduces the dimension of the earthquake feature vector through a convolution layer with a convolution kernel size of 3×1, followed by a batch normalization layer to reduce feature redundancy and extract deep features; the other branch reduces the dimension of the branch free-field soil displacement response feature vector through a convolution layer with a convolution kernel size of 2×1, followed by a batch normalization layer to reduce feature redundancy and extract deep features. The deep features of the two branches are fused by point multiplication, and the channel average is calculated by the mean function to generate a correlation matrix, and the activation function Sigmoid is used to evaluate the correlation between features. Finally, according to the value of the correlation matrix, the earthquake feature or the free-field soil displacement response feature is dynamically selected to obtain the fused feature.

[0038] Further preferably, Figure 4As shown, the dual-branch attention module includes a convolution group, a temporal attention module and a channel attention module; the convolution group is composed of a convolution layer, a ReLU activation function and another convolution layer, and the fused features output by the dual-branch feature fusion module are re-extracted through the convolution group; the temporal attention module first applies the average pooling layer and the maximum pooling layer to reduce the dimension of the fused features and generate two sets of features, then splices the two sets of features, and then compresses the spliced ​​features into single-channel features through the convolution layer, and then normalizes the temporal weights through the Sigmoid activation function to generate temporal attention weights; the channel attention module applies the average pooling layer and the maximum pooling layer in the channel dimension to reduce the dimension of the fused features and generate two sets of features, sums the two sets of features element by element, and then inputs the multi-layer perceptron (including a fully connected layer for dimensionality reduction, a ReLU activation function and a fully connected layer for dimensionality increase) to model the interdependence between channels, adaptively judge the importance of each channel, and then the Sigmoid The activation function generates the channel attention weight; finally, the temporal attention weight and the channel attention weight are multiplied and applied to the features after convolution group processing to generate the dual-branch attention enhanced features.

[0039] In this embodiment, during the training of the DB-CNN-Attention network model, the batch size is set to 32 and the number of training rounds is set to 1000. A selection mechanism is set, and when the mean square error (MSE) of the DB-CNN-Attention network model is currently the best in the specified round, it is saved as the best DB-CNN-Attention network model. The definition of mean square error (MSE) is as follows: ; Where N is the dimension of a single earthquake time series sample in the training set, Y i is the actual response value of the subway station structure, P i is the predicted structural response value of the subway station.

[0040] Adam is used as the optimizer to update the learnable parameters of the DB-CNN-Attention network model, and the initial learning rate parameter is set to 1.2×10 -3 , and adopts exponential decay strategy to dynamically adjust the learning rate. The training goal of the DB-CNN-Attention network model is to minimize the mean square error MSE of each batch size.

[0041] Furthermore, the training set data is normalized to the range [0, 1] to eliminate the scale differences between different features, making the model easier to train and converge. Figure 5 Iteration curve of the DB-CNN-Attention network model.

[0042] Step 3: Determine the site category of the target subway station, obtain the structural information of the target subway station, and obtain the seismic time history data according to actual needs; preprocess the structural information and seismic time history data of the target subway station, and input the preprocessed data into the trained DB-CNN-Attention network model to output the earthquake time history response data prediction results of the target subway station.

[0043] This embodiment uses three indicators to measure the error of the predicted data: mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R 2 ), respectively expressed as follows: ; ; in, is the true value of the earthquake acceleration response, is the predicted value of the earthquake acceleration response, It represents the mean of the true value of the earthquake acceleration response, and m is the number of data sample points.

[0044] In this embodiment, in order to verify the prediction effect of the model, 25 seismic waves different from the test set are randomly selected from the NGA-West2 seismic database (PEER database) of the Pacific Earthquake Engineering Research (PEER) Center, and input into the corresponding one-dimensional equivalent linearized site model and the subway station structure finite element calculation model, and the corresponding data is obtained to establish a test subset that is completely independent of the training set for verification. The inter-story displacement response of the subway station predicted in this embodiment, the prediction results of the test subset are shown in Table 1.

[0045] Table 1

[0046] The above data show that the proposed prediction method has a fairly high prediction accuracy. Figure 6 The seismic wave input shown in the figure shows the seismic response of the subway station structure predicted by the trained DB-CNN-Attention network. Figure 7-Figure 8 It can be seen that the proposed prediction method has a fairly high prediction accuracy. The determination coefficient R in the earthquake time history response prediction of a three-story, three-span subway station in a Class II site is 2 The determination coefficient in the earthquake time-history response prediction of a three-story, three-span subway station in a Class IV site can reach 0.967.

[0047] The present invention is used to solve the shortcomings of the finite element earthquake response prediction method based on the physical drive method in the prior art, such as slow prediction speed, complex modeling, many repeated operations, and high learning cost; and the shortcomings of the earthquake response prediction method based on the data drive method, such as high time cost for obtaining large sample training data, poor generalization ability and poor robustness. The present invention can accurately predict the inter-story displacement response of subway stations under earthquake action.

[0048] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A subway station earthquake response prediction method based on deep learning, characterized in that: The following steps are involved: Step 1: Obtain the subway station structure information and site category information required for the subway station earthquake time history response prediction, and establish a finite element calculation model and a one-dimensional equivalent linearized site corresponding to the site category for the subway station structure of different site categories; The seismic time history data is input into the finite element calculation model, and nonlinear time history response analysis is performed on various subway station structures to obtain the seismic response data of subway station structures of different site categories under various seismic time history data; the seismic time history data is input into the one-dimensional equivalent linearized site to calculate the free field soil displacement response data of each seismic time history data under different site categories, and the seismic time history data-free field soil displacement response-subway station seismic response data set is established; Step 2: Preprocess the data in the earthquake time history data-free field soil displacement response-subway station earthquake response data set, build a DB-CNN-Attention network model, and use the preprocessed data set to train the DB-CNN-Attention network model to obtain a trained DB-CNN-Attention network model; Step 3: Determine the site category of the target subway station, obtain the structural information of the target subway station, and obtain the seismic time history data according to actual needs; The structural information and earthquake time history data of the target subway station are preprocessed, and the preprocessed data are input into the trained DB-CNN-Attention network model to output the earthquake time history response data prediction results of the target subway station.

2. The subway station earthquake response prediction method based on deep learning according to claim 1 is characterized in that: The DB-CNN-Attention network model includes a seismic time-history data encoding module, a free-field soil displacement response data encoding module, a dual-branch feature fusion module, a dual-branch attention module and a decoding module. The seismic time-history data encoding module extracts features from the seismic time-history data to obtain a seismic feature vector. The free-field soil displacement response data encoding module extracts features from the free-field soil displacement response data to obtain a free-field soil displacement response feature vector. The dual-branch feature fusion module fuses the seismic feature vector and the free-field soil displacement response feature vector to obtain a fused feature vector. The fused features are then input into the dual-branch attention module for weighting, and the weighted features are input into the decoding module for decoding, and finally the earthquake response prediction results of the subway station are obtained.

3. The subway station earthquake response prediction method based on deep learning according to claim 2 is characterized in that: The structure of the earthquake time-history data encoding module is the same as that of the free-field soil displacement response data encoding module, which includes three convolutional pooling modules. The convolutional pooling module is equipped with one convolutional layer and one pooling layer, and the convolutional layer is followed by an activation function ReLU.

4. The subway station earthquake response prediction method based on deep learning according to claim 2 is characterized in that: The dual-branch feature fusion module includes two branches, one branch performs dimensionality reduction processing on the earthquake motion feature vector through a convolution layer, followed by a batch normalization layer to extract deep features; the other branch performs dimensionality reduction processing on the free-field soil displacement response feature vector through a convolution layer, followed by a batch normalization layer to extract deep features, the deep features of the two branches are fused by dot multiplication, and then the channel average is calculated by the mean function mean to generate a correlation matrix, and the activation function Sigmoid is used to evaluate the correlation between features. Finally, according to the value of the correlation matrix, the earthquake motion feature or the free-field soil displacement response feature is dynamically selected to obtain the fused feature.

5. The subway station earthquake response prediction method based on deep learning according to claim 2, characterized in that: The dual-branch attention module includes a convolution group, a temporal attention module and a channel attention module; the fused features output by the dual-branch feature fusion module are re-extracted through the convolution group; The temporal attention module first applies the average pooling layer and the maximum pooling layer to reduce the dimension of the fused features and generate two sets of features, then concatenates the two sets of features, compresses the concatenated features into single-channel features through the convolution layer, and then normalizes the temporal weights through the Sigmoid activation function to generate the temporal attention weights; the channel attention module applies the average pooling layer and the maximum pooling layer in the channel dimension to reduce the dimension of the fused features and generate two sets of features, sums the two sets of features element by element, and then inputs the multi-layer perceptron to model the interdependence between channels and adaptively judge the importance of each channel, and then generates the channel attention weight by the Sigmoid activation function; finally, the temporal attention weight and the channel attention weight are multiplied and applied to the features after convolution group processing to generate the dual-branch attention enhancement feature.

6. The subway station earthquake response prediction method based on deep learning according to claim 1, characterized in that: The seismic time history data were selected from the Pacific Earthquake Engineering Research Center seismic database using a random sampling method.

7. The subway station earthquake response prediction method based on deep learning according to claim 2, characterized in that: The preprocessing of seismic time history data includes: unifying the duration and sampling frequency of seismic time history data; the preprocessing methods of earthquake time history response data and free field soil displacement response data are consistent with this, and the training data are normalized and scaled to [0, 1].

8. The subway station earthquake response prediction method based on deep learning according to claim 2, characterized in that: The DB-CNN-Attention network model is optimized by the Adam optimizer.

9. The subway station earthquake response prediction method based on deep learning according to claim 2, characterized in that: The decoding module is composed of a flattening layer, a fully connected layer and an activation function ReLU which are arranged in sequence.

Images