A method for predicting the seismic response of subway stations based on deep learning

Abstract

The present invention discloses a method for predicting the seismic response of subway stations based on deep learning. The method first obtains the subway station structure information and site category information required for predicting the seismic time-history response of the subway station, establishes a seismic ground motion time-history data-free field soil displacement response-subway station seismic response data set and preprocesses the data therein, uses the preprocessed data set to train the DB-CNN-Attention network model, determines the site category of the target subway station, obtains the target subway station structure information, and obtains the seismic ground motion time-history data according to actual needs; uses the trained DB-CNN-Attention network model to output the prediction result of the seismic time-history response data of the target subway station. The prediction of the present invention has higher accuracy and high computational efficiency, and can be used for post-earthquake damage estimation, seismic disaster scenario reconstruction and pre-earthquake seismic capacity assessment.

Description

Technical Field

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

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

[0003] At present, there are mainly three 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 information such as the geometry, materials, and loads of the subway station, and then selects appropriate seismic ground motion time histories as input conditions and loads them into the finite element dynamic calculation model of the subway station to conduct nonlinear dynamic time history analysis, and the time-varying responses such as displacements, velocities, and accelerations of the subway station structure under seismic action can be calculated. This method can simulate the complex response behavior of the subway station structure under seismic ground motion excitation and provide data support for seismic design. However, the numerical simulation method requires the establishment of a complex finite element model and consumes a large amount of time for nonlinear dynamic time history analysis. In addition, in the post-earthquake emergency assessment, due to the long calculation time, it may not be able to 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 the seismic response of subway station structures to predict the damage state, peak response, and time history response of the station structure, etc., due to their ability to handle complex problems with high precision and high efficiency. However, the existing studies in this field still face the following problems: 1. Some studies only consider the input seismic ground motion information of the subway station. The input seismic ground motion information is usually simple intensity parameters, rather than seismic ground motion time histories or time-frequency diagrams, which can more comprehensively reflect the characteristics and differences of seismic ground motions. 2. The research objects of most deep learning models are limited to stations in specific regions, and it is impossible to simultaneously consider the site category information and seismic ground motion information of subway stations to predict the comprehensive seismic response. When the research area changes, the model needs to be retrained according to new data, lacking generalization ability.

[0005] In summary, due to the limitations of the input information, the existing prediction methods for the seismic response of subway stations have low prediction accuracy, and the model needs to be retrained when adapting to the station structures in different regions, resulting in poor applicability. Therefore, it is of great significance to propose a new prediction method that combines more comprehensive subway station area information and ground motion information to improve the 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 existing prediction methods for the seismic response of subway station structures and the need for the model to be retrained when the prediction object changes, and to propose a prediction method for the seismic response of subway stations based on deep learning.

[0007] The technical solution adopted by the present invention to solve the above technical problems is as follows: A prediction method for the seismic response of subway stations based on deep learning, comprising the following steps:

[0008] Step 1: Obtain the subway station structure information and site category information required for predicting the seismic time-history response of the subway station, establish a finite element calculation model for the subway station structures of different site categories and a one-dimensional equivalent linearized site corresponding to the site category; input the ground motion time-history data into the finite element calculation model, conduct 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 ground motion time-history data; input the ground motion time-history data into the one-dimensional equivalent linearized site to calculate the free-field soil displacement response data of each ground motion time-history data under different site categories, and establish a ground motion time-history data-free-field soil displacement response-subway station seismic response data set;

[0009] Step 2: Preprocess the data in the ground motion time-history data-free-field soil displacement response-subway station seismic response data set, construct 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;

[0010] Step 3: Determine the site category of the target subway station, obtain the target subway station structure information, and obtain the ground motion time-history data according to actual needs; preprocess the target subway station structure information and the ground motion time-history data, and input the preprocessed data into the trained DB-CNN-Attention network model to output the prediction result of the seismic time-history response data of the target subway station.

[0011] Further preferably, the DB-CNN-Attention network model includes a ground motion 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 ground motion time history data encoding module extracts features from the ground motion time history data to obtain a ground motion 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 ground motion feature vector and the free-field soil displacement response feature vector to obtain a fused feature vector. Subsequently, the fused feature is input into the dual-branch attention module for weighting, and the weighted feature is input into the decoding module for decoding to finally obtain the subway station seismic response prediction result.

[0012] Further preferably, the ground motion time history data encoding module and the free-field soil displacement response data encoding module have the same structure, including 3 convolutional pooling modules. Each convolutional pooling module is provided with 1 convolutional layer and 1 pooling layer, and the convolutional layer is followed by the activation function ReLU.

[0013] Further preferably, the dual-branch feature fusion module includes two branches. One branch reduces the dimension of the ground motion feature vector through a convolutional layer with a kernel size of 3×1, followed by a batch normalization layer to extract deep features. The other branch reduces the dimension of the free-field soil displacement response feature vector through a convolutional layer with a kernel size of 2×1, followed by a batch normalization layer to extract deep features. The deep features of the two branches are multiplied element-wise and then the channel average value is calculated through 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 ground motion feature or the free-field soil displacement response feature is dynamically selected to obtain the fused feature.

[0014] Further preferably, the dual-branch attention module includes a convolutional 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 convolutional group; the temporal attention module first applies an average pooling layer and a max pooling layer to reduce the dimensionality of the fused features and generate two sets of features, then concatenates the two sets of features, and then compresses the concatenated features into a single-channel feature through a convolutional layer, and then generates temporal attention weights by normalizing the temporal weights through a Sigmoid activation function; the channel attention module applies an average pooling layer and a max pooling layer in the channel dimension to reduce the dimensionality of the fused features and generate two sets of features, sums the two sets of features element-wise, and then inputs them into a multi-layer perceptron to model the mutual dependence between channels, adaptively discriminates the importance of each channel, and then generates channel attention weights by a Sigmoid activation function; finally, the temporal attention weights and the channel attention weights are multiplied and applied to the features processed by the convolutional group to generate dual-branch attention enhanced features.

[0015] Further preferably, the decoding module consists of a flattening layer, a fully connected layer, and the activation function ReLU arranged in sequence.

[0016] Further preferably, the ground motion time history data is selected from the Pacific Earthquake Engineering Research Center ground motion database by means of random sampling.

[0017] Further preferably, the preprocessing of the ground motion time history data includes: unifying the duration and sampling frequency of the ground motion time history data; the preprocessing methods of the seismic time history response data and the free-field soil displacement response data are the same, and the training data is normalized and scaled to between [0, 1].

[0018] Further preferably, the DB-CNN-Attention network model is optimized by an Adam optimizer.

[0019] The beneficial effects of the present invention are as follows:

[0020] In the prediction of the seismic response of subway stations, the present invention takes into account the subway station structure information and site category available in the actual scenario, and uses a deep learning algorithm to establish a DB-CNN-Attention network model suitable for predicting the seismic time history response of subway stations. Compared with the existing methods, the present invention obtains higher prediction accuracy when only using the subway station structure information and site category information, and these information are easy to obtain and closer to the actual application requirements.

[0021] Meanwhile, 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, greatly improving the actual application efficiency of the model. In addition, this method has high computational efficiency and can effectively serve earthquake damage estimation of subway stations, earthquake disaster scene reconstruction, and seismic capacity assessment of subway stations before earthquakes. Description of the Drawings

[0022] By referring to 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 drawings, several embodiments of the present invention are shown in an exemplary rather than restrictive manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0023] Figure 1 is a schematic diagram of a partially randomly selected seismic acceleration response spectrum.

[0024] Figure 2 is a schematic diagram of the DB-CNN-Attention network model.

[0025] Figure 3 is a schematic diagram of the double-branch special fusion module.

[0026] Figure 4 is a schematic diagram of the double-branch attention module.

[0027] Figure 5 is the iteration curve of the DB-CNN-Attention network model.

[0028] Figure 6 is the seismic wave randomly selected for verifying the present invention.

[0029] Figure 7 is the prediction effect diagram of the seismic time-history response of a three-story and three-span subway station under Class II site conditions.

[0030] Figure 8 is the prediction effect diagram of the seismic time-history response of a three-story and three-span subway station under Class IV site conditions. Detailed Embodiments

[0031] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0032] It should be understood that when terms such as "first", "second", etc. are used in the claims, specification and drawings of the present invention, they are only used to distinguish different objects, rather than to describe a specific order. The terms "comprising" and "including" used in the specification 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 combinations.

[0033] A subway station seismic response prediction method based on deep learning includes Step 1 to Step 3, which are specifically described below.

[0034] Step 1: Obtain the subway station structure information and site category information required for predicting the subway station seismic time history response. Establish a finite element calculation model for the subway station structures of different site categories and a one-dimensional equivalent linearized site corresponding to the site category. Select several seismic ground motion time history data from the seismic ground motion database, and divide them into a training set, a validation set and a test set; input the selected seismic ground motion time history data into the finite element calculation model, and perform nonlinear time history response analysis on various subway station structures to obtain the subway station seismic response data of the subway station structures of different site categories under each seismic ground motion time history data; input the seismic ground motion time history data into the one-dimensional equivalent linearized site to calculate the free-field soil displacement response data of each seismic ground motion time history data under different site categories, and establish a seismic ground motion time history data-free-field soil displacement response-subway station seismic response data set.

[0035] Seismic waves with different peak ground accelerations (PGA) and different magnitudes are selected from the seismic ground motion database. In this embodiment, 160 seismic waves (i.e., seismic ground motion time history data) with large differences are randomly selected from the NGA-West2 seismic ground motion database (PEER database) of the Pacific Earthquake Engineering Research (PEER) Center. Figure 1 The partial seismic acceleration response spectra randomly selected are shown. This set of data has large differences in PGA, spectral characteristics and duration, and can be used to verify the generalization ability of the neural network model for predicting different seismic waves. The PGA range of the obtained seismic waves is 0.05 - 0.30 g. To simplify the calculation process, the sampling interval of the original seismic wave acceleration time history is adjusted to 0.02 s. The preprocessed seismic ground motion time history data is divided into three data sets: a training set, a validation set and a test set, and the data ratios are 70%, 10% and 20% respectively.

[0036] In this embodiment, in order to accurately establish the one-dimensional equivalent linearized site corresponding to different site types, it is necessary to collect four types of parameter information, namely the soil layer thickness, shear wave velocity, soil unit weight, and nonlinear characteristics corresponding to different site types. Then, based on these four types of site type parameters, they are input into numerical calculation software to establish the corresponding one-dimensional equivalent linearized site model.

[0037] Construct a typical finite element calculation model according to the structural parameters of the subway station. The modeling method is as follows: The subway station structure is simulated using plane strain elements. The soil components and concrete components are simulated using four-node bilinear plane strain quadrilateral elements. The steel bars are simulated using bar elements embedded in the concrete. A series of springs are used to simulate the interaction between the soil and the subway station structure during an earthquake. The contact between the soil and the subway station structure adopts hard contact. The friction coefficient of the contact surface between the soil and the subway station structure is taken as 0.4. The boundary conditions of the finite element calculation model adopt the static-dynamic coupling processing technology.

[0038] Input the above three selected data sets (training set, validation set, and test set) into the established finite element calculation model of the subway station structure for nonlinear time history analysis; at the same time, input the data sets into the one-dimensional equivalent linearized site model to calculate the free-field soil displacement response, obtain the inter-story displacement response corresponding to the ground motion record and the corresponding free-field soil deformation, and establish a ground motion time history data-free-field soil displacement response-subway station seismic response data set.

[0039] Step 2: Preprocess the data in the ground motion time history data-free-field soil displacement response-subway station seismic response data set, construct 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.

[0040] As Figure 2As shown in the figure, the DB-CNN-Attention network model includes a ground motion 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 ground motion time history data encoding module extracts features from the ground motion time history data to obtain a ground motion 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 ground motion feature vector and the free-field soil displacement response feature vector to obtain a fused feature vector. Subsequently, the fused feature is input into the dual-branch attention module for weighting, and the weighted feature is input into the decoding module for decoding to finally obtain the subway station seismic response prediction result. 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 input layer dimension: [1, 2000], output layer output dimension [1, 2000].

[0041] The ground motion time history data encoding module and the free-field soil displacement response data encoding module have the same structure, including 3 convolutional pooling modules. Each convolutional pooling module is provided with 1 convolutional layer and 1 pooling layer. The convolutional layer is followed by an activation function ReLU, and the size of the convolutional kernel of the convolutional layer is set to 3×1; the size of the pooling kernel of the pooling layer is 2×1, and the stride is 1 for both. The pooling operation is max pooling. The number of kernels of the convolutional layers in the three convolutional pooling modules is sequentially set to 32, 64, and 128.

[0042] As Figure 3 shown, the dual-branch feature fusion module includes two branches. One branch reduces the dimension of the ground motion feature vector through a convolutional layer with a convolutional 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 free-field soil displacement response feature vector through a convolutional layer with a convolutional 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 dot-multiplied and fused, and then the channel average value is calculated through 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 ground motion feature or the free-field soil displacement response feature is dynamically selected to obtain the fused feature.

[0043] Further preferably, as Figure 4As shown in the figure, the dual-branch attention module includes a convolution group, a temporal attention module, and a channel attention module. The convolution group consists of a convolution layer, a ReLU activation function, and another convolution layer, which re-extracts features from the fused features output by the dual-branch feature fusion module. The temporal attention module first applies an average pooling layer and a max pooling layer to reduce the dimension of the fused features and generate two sets of features, then concatenates the two sets of features, and then compresses the concatenated features into a single-channel feature through a convolution layer, and then normalizes the temporal weights through a Sigmoid activation function to generate temporal attention weights. The channel attention module applies an average pooling layer and a max 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-wise, and then inputs them into a 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 discriminates the importance of each channel, and then generates channel attention weights through a Sigmoid activation function. Finally, the temporal attention weights and the channel attention weights are multiplied and applied to the features processed by the convolution group to generate dual-branch attention enhanced features.

[0044] In this embodiment, during the training process of the DB-CNN-Attention network model, the batch size is set to 32, and the number of training epochs is set to 1000. A selection mechanism is set. When the mean squared error (MSE) of the DB-CNN-Attention network model is the best at a specified epoch, it is saved as the optimal DB-CNN-Attention network model. The definition of the mean squared error (MSE) is as follows:

[0045] ;

[0046] where N is the dimension of a single seismic ground motion time series sample in the training set, Y i is the actual response value of the subway station structure, and P i is the predicted response value of the subway station structure.

[0047] Adam is used as the optimizer to update the learnable parameters of the DB-CNN-Attention network model. The initial learning rate parameter is set to 1.2×10 -3 , and an exponential decay strategy is adopted to dynamically adjust the learning rate. The training objective of the DB-CNN-Attention network model is to minimize the mean squared error MSE of each batch size.

[0048] Furthermore, the training set data is normalized by maximum-minimum normalization to scale the data between [0, 1]. The scale differences between different features are eliminated, making it easier for the model to train and converge. Figure 5It is the iteration curve of the DB-CNN-Attention network model.

[0049] Step 3: Determine the site category of the target subway station, obtain the structural information of the target subway station, and obtain the ground motion time history data according to actual needs; preprocess the structural information of the target subway station and the ground motion time history data, and input the preprocessed data into the trained DB-CNN-Attention network model to output the prediction result of the seismic time history response data of the target subway station.

[0050] 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 ), which are respectively expressed as follows:

[0051] ;

[0052] ;

[0053] Among them, is the true value of the ground motion acceleration response, is the predicted value of the ground motion acceleration response, represents the mean value of the true value of the ground motion acceleration response, and m is the number of data sample points.

[0054] 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 ground motion database (PEER database) of the Pacific Earthquake Engineering Research (PEER) Center, and input into the corresponding one-dimensional equivalent linearized site model and the finite element calculation model of the subway station structure to obtain the corresponding data to establish a test subset completely independent of the training set for verification. The predicted inter-story displacement response of the subway station in this embodiment is shown in Table 1.

[0055] Table 1

[0056]

[0057] The above data shows that the proposed prediction method has a quite high prediction accuracy. Then, the seismic waves shown in Figure 6 are selected and input into the trained DB-CNN-Attention network to predict the seismic response results of the subway station structure as shown in Figures 7-8 . It can be seen that the proposed prediction method has a quite high prediction accuracy. In the prediction of the seismic time history response of the three-story and three-span subway station in Site Class II, the coefficient of determination R 2 reaches 0.972, and in the prediction of the seismic time history response of the three-story and three-span subway station in Site Class IV, the coefficient of determination can reach 0.967.

[0058] The present invention is used to solve the disadvantages of the finite element seismic response prediction method based on the physical driving method in the prior art, such as slow prediction speed, complex modeling, many repeated operations, and high learning costs; and the disadvantages of the seismic response prediction method based on the data-driven method, such as the expensive time cost of obtaining large-sample training data, poor generalization ability, and poor robustness, and can accurately predict the inter-story displacement response of subway stations under seismic action.

[0059] Those skilled in the art can easily understand that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting the seismic response of subway stations based on deep learning, characterized in that, It includes the following steps: Step 1: Obtain the subway station structure information and site category information required for predicting the seismic time history response of the subway station. Establish a finite element calculation model for the subway station structures of different site categories and a one-dimensional equivalent linearized site corresponding to the site category; Input the seismic ground motion time history data into the finite element calculation model, conduct nonlinear time history response analysis on various subway station structures, and obtain the seismic response data of subway station structures of different site categories under each seismic ground motion time history data; Input the seismic ground motion time history data into the one-dimensional equivalent linearized site to calculate the free-field soil displacement response data of each seismic ground motion time history data under different site categories, and establish a seismic ground motion time history data-free field soil displacement response-subway station seismic response dataset; Step 2: Preprocess the data in the seismic ground motion time history data-free field soil displacement response-subway station seismic response dataset, construct a DB-CNN-Attention network model, and use the preprocessed dataset to train the DB-CNN-Attention network model to obtain a trained DB-CNN-Attention network model; The DB-CNN-Attention network model includes a seismic ground motion time history data encoding module, a free-field soil displacement response data encoding module, a double-branch feature fusion module, a double-branch attention module, and a decoding module. The seismic ground motion time history data encoding module extracts features from the seismic ground motion time history data to obtain a seismic ground motion 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 double-branch feature fusion module fuses the seismic ground motion feature vector and the free-field soil displacement response feature vector to obtain a fused feature vector; Subsequently, the fused feature is input into the double-branch attention module for weighting, and the weighted feature is input into the decoding module for decoding to finally obtain the prediction result of the subway station seismic response; Step 3: Determine the site category of the target subway station, obtain the target subway station structure information, and obtain the seismic ground motion time history data according to actual needs; Preprocess the target subway station structure information and the seismic ground motion time history data, and input the preprocessed data into the trained DB-CNN-Attention network model to output the prediction result of the seismic time history response data of the target subway station.

2. The method for predicting the seismic response of a subway station based on deep learning according to claim 1, wherein The structures of the seismic ground motion time history data encoding module and the free-field soil displacement response data encoding module are the same, including 3 convolutional pooling modules. Each convolutional pooling module is provided with 1 convolutional layer and 1 pooling layer, and the convolutional layer is followed by the activation function ReLU.

3. The method for predicting the seismic response of a subway station based on deep learning according to claim 1, characterized in that The double-branch feature fusion module includes two branches. One branch reduces the dimension of the ground motion feature vector through a convolutional layer, followed by a batch normalization layer to extract deep features. The other branch reduces the dimension of the free-field soil displacement response feature vector through a convolutional layer, followed by a batch normalization layer to extract deep features. The deep features of the two branches are multiplied element-wise and fused, and then the channel average value is calculated through the mean function to generate a correlation matrix. The sigmoid activation function is used to evaluate the correlation between features. Finally, according to the value of the correlation matrix, the ground motion feature or the free-field soil displacement response feature is dynamically selected to obtain the fused feature.

4. The method for predicting the seismic response of a subway station based on deep learning according to claim 1, characterized in that, The double-branch attention module includes a convolutional group, a temporal attention module, and a channel attention module. The fused feature output by the double-branch feature fusion module is re-extracted through the convolutional group. The temporal attention module first applies the average pooling layer and the max pooling layer to reduce the dimension of the fused feature and generate two sets of features. Then the two sets of features are concatenated, and the concatenated features are compressed into a single-channel feature through a convolutional layer. Then the sigmoid activation function is used to normalize the temporal weights to generate the temporal attention weights. The channel attention module applies the average pooling layer and the max pooling layer in the channel dimension to reduce the dimension of the fused feature and generate two sets of features. The two sets of features are added element-wise and then input into a multi-layer perceptron to model the mutual dependence between channels and adaptively judge the importance of each channel. Subsequently, the sigmoid activation function generates the channel attention weights. Finally, the temporal attention weights and the channel attention weights are multiplied and applied to the feature processed by the convolutional group to generate the double-branch attention enhanced feature.

5. The method for predicting the seismic response of a subway station based on deep learning according to claim 1, wherein The ground motion time history data is selected from the Pacific Earthquake Engineering Research Center (PEER) ground motion database by using the random sampling method.

6. The method for predicting the seismic response of a subway station based on deep learning according to claim 1, wherein The preprocessing of the ground motion time history data includes: unifying the duration and sampling frequency of the ground motion time history data. The preprocessing methods of the earthquake time history response data and the free-field soil displacement response data are the same. The training data is normalized and scaled to the range of [0, 1].

7. The method for predicting the seismic response of a subway station based on deep learning according to claim 1, wherein The DB-CNN-Attention network model is optimized by the Adam optimizer.

8. The method for predicting the seismic response of a subway station based on deep learning according to claim 1, wherein The decoding module consists of a flattening layer, a fully connected layer, and the ReLU activation function, which are arranged in sequence.

Images