A deep learning-based structural seismic response prediction method

Abstract

The application discloses a kind of structure earthquake response prediction methods based on deep learning, belong to the cross technical field of structure engineering and artificial intelligence, including: obtaining the structure information of target structure and site category information, construct finite element model, obtain multiple seismic acceleration time history data to carry out nonlinear time history analysis, extract the displacement response time history of key position, form data set;CNN-LSTM-Attention hybrid neural network model is constructed and trained, and seismic response prediction model is obtained;The seismic acceleration time history to be predicted is input into model, and the output structure seismic prediction displacement response time history data is output.The application realizes the fast, accurate prediction of the seismic displacement response of prefabricated assembly pile plate structure by fusing convolutional neural network, long short-term memory network and attention mechanism, the prediction accuracy is high, the calculation efficiency is significantly better than traditional finite element analysis method, and can provide efficient technical support for structure seismic design and performance evaluation.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of structural engineering and artificial intelligence, and in particular to a deep learning-based method for predicting structural seismic response. Background Technology

[0002] Precast assembled pile-slab structures are prefabricated structural systems consisting of precast columns and precast bearing slabs connected by rigid sleeves. Due to their high degree of prefabrication, high rigidity, and low settlement, they have been increasingly widely used in transportation infrastructure construction. The main areas of damage to this structure are the connection between the columns and the bearing slab, as well as the base of the columns; their seismic performance directly affects the safety and service life of the structure.

[0003] Currently, scholars have conducted relevant research on the seismic performance of this type of structure. However, although traditional nonlinear time history analysis methods are theoretically rigorous, they suffer from high computational costs and low analysis efficiency, making it difficult to meet the needs of rapid decision-making in engineering practice.

[0004] In recent years, researchers have applied convolutional neural networks (CNNs) and long short-term memory networks (LSTMs) to predict structural seismic response, achieving some success. However, no reports have yet described a technical solution that integrates CNNs, LSTMs, and attention mechanisms to construct a hybrid neural network model and applies it to predict the seismic displacement response of precast pile-slab structures. Most existing methods are designed for conventional building structures; due to the unique characteristics of precast pile-slab structures, which are characterized by the connection of precast columns and load-bearing slabs via rigid sleeves, these methods are difficult to apply directly. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based method for predicting structural seismic responses, aiming to address the technical problems of low computational efficiency and difficulty in meeting the requirements for rapid prediction of large-scale seismic responses in existing finite element methods. This method constructs a hybrid neural network model integrating convolutional neural networks, long short-term memory networks, and attention mechanisms, and trains the model using historical ground motion data to achieve accurate and efficient prediction of the seismic displacement response of the structure under test, providing rapid decision support for structural seismic design and performance evaluation.

[0006] To achieve the above objectives, this invention provides a deep learning-based method for predicting structural seismic response, comprising the following steps: S1. Obtain the structural information and site category information of the target structure. The structural information includes the connection information of precast columns, precast bearing plates and rigid sleeves. The site category information includes the site category and characteristic period. Construct a finite element model, obtain multiple ground motion acceleration time history data for nonlinear time history analysis, extract the displacement response time history of key parts, the key parts being the top of the precast columns, and form a dataset. S2. Construct a CNN-LSTM-Attention hybrid neural network model, train it using the dataset, and obtain an earthquake response prediction model; S3. Input the time history of the ground motion acceleration to be predicted into the seismic response prediction model, and output the time history data of the structural seismic predicted displacement response.

[0007] Preferably, the target structure in S1 is a prefabricated assembled pile-slab structure; S1 specifically includes: S11. Obtain the structural information and site category information of the target structure, and construct a finite element model; the target structure includes at least one precast column and one precast bearing plate, and the top of the precast column is fixedly connected to the precast bearing plate through a rigid sleeve. S12. Acquire multiple ground motion acceleration time history data, select them according to the characteristic period corresponding to the site category information, and select them from historical strong earthquake records of different magnitudes and different epicentral distances; fill or truncate each ground motion acceleration time history data to make all data have a uniform time step length and sampling frequency. S13. Input each ground motion acceleration time history data into the finite element model for nonlinear time history analysis, extract the displacement response time history at the top of the precast column, and form a dataset with the ground motion acceleration time history as input and the corresponding displacement response time history as output.

[0008] Preferably, in S13, the dataset is divided into a training set, a validation set, and a prediction set in a ratio of 7:1.5:1.5.

[0009] Preferably, in S11: The precast column is a hollow precast pipe pile. The precast pipe pile is equipped with prestressed steel bars and ordinary steel bars along its length, and its lower end is fixed to the foundation. The steel sleeve is a steel connector pre-embedded in the top of the precast column or in the precast bearing plate, and is consolidated by cast-in-place concrete or mechanical connection. When constructing the finite element model, the precast columns adopt force-based nonlinear beam-column elements with multiple Gaussian integration points along the length direction. The concrete adopts the Concrete01 uniaxial material model, the steel bars and prestressed steel bars adopt the Steel01 bilinear material model, the precast bearing plate adopts shell elements, and the bottom of the structure simulates the structure-foundation interaction through zero-length elements.

[0010] Preferably, the target structure includes multiple prefabricated columns spaced longitudinally. In S13, the displacement response time history of the top of each prefabricated column is extracted so that the output dimension of the dataset corresponds to the number of prefabricated columns. The target structure is a multi-span continuous structure, with the load-bearing plates of each span forming a continuous whole. The prefabricated columns are arranged symmetrically or at equal intervals below each span, and the load-bearing plate sections between adjacent prefabricated columns form a span.

[0011] Preferably, S2 specifically includes: S21. Construct a CNN-LSTM-Attention hybrid neural network model, which includes a convolutional neural network module, a long short-term memory network module, and an attention mechanism module. The convolutional neural network module consists of two one-dimensional convolutional layers. The first convolutional layer has 64 filters and a kernel size of 5, while the second convolutional layer has 128 filters and a kernel size of 3. Each convolutional layer is followed by a ReLU activation function and batch normalization. The output of the convolutional neural network module is flattened in terms of feature dimension by a fully connected layer to form a temporal feature sequence, which is then input to the long short-term memory network module. The long short-term memory network module consists of two LSTM layers, each with 128 hidden units. The attention mechanism module receives the output sequence of the LSTM layer, calculates the attention weights through a fully connected layer, normalizes them using Softmax, and then sums the output sequence by weight to obtain the attention context vector. Finally, it outputs the displacement response prediction value through a fully connected layer. S22. Train the hybrid neural network model using the dataset until it converges to obtain the earthquake response prediction model. During training, a Dropout strategy is used with a Dropout ratio of 0.3. The optimizer is AdamW, with a first-order moment decay rate of 0.9, a second-order moment decay rate of 0.999, and a weight decay coefficient of 10. -6 Add an L2 regularization term to the loss function, with the regularization coefficient set to 10. -6 An early stop strategy is adopted, which stops training when the validation loss does not decrease after 50 consecutive training rounds, with the minimum number of training rounds set to 100 rounds.

[0012] Preferably, in S3, the structural seismic prediction displacement response time history data includes the displacement response time history and peak displacement of the top of all precast columns.

[0013] Preferably, it also includes S4, prediction accuracy evaluation; S4 specifically includes: The root mean square error, mean absolute error, and coefficient of determination are used as evaluation indicators to assess the prediction accuracy of the hybrid neural network model. The formula for calculating the root mean square error is: ; in, The root mean square error, For the first Actual values ​​of structural seismic response for each sample. For the first The model prediction value for each sample. The number of samples; The formula for calculating the mean absolute error is: ; in, Mean absolute error; The formula for calculating the coefficient of determination is: ; in, As the coefficient of determination, This is the mean of the actual values ​​of all samples.

[0014] Therefore, this invention employs the aforementioned deep learning-based structural seismic response prediction method. By integrating the spatial feature extraction capabilities of CNN, the temporal dependency modeling capabilities of LSTM, and the key information focusing capabilities of the attention mechanism, it achieves accurate prediction of the seismic displacement response of highway pile-slab bridge structures. Finite element simulation verification shows that the prediction set's coefficient of determination... The accuracy and speed of predicting structural seismic response are both high, with a root mean square error of approximately 0.001 and a mean absolute error of approximately 0.0001. The error between predicted and actual displacement is approximately 1 mm, and the single prediction time far exceeds that of traditional finite element analysis. This method significantly improves the accuracy and speed of predicting structural seismic response and verifies the feasibility and engineering applicability of the model on a general-purpose hardware platform.

[0015] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a deep learning-based structural seismic response prediction method according to the present invention. Figure 2 This is a schematic diagram of the structure of a Long Short-Term Memory (LSTM) neural network unit according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the attention mechanism framework according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the CNN-LSTM-Attention hybrid neural network model according to an embodiment of the present invention; Figure 5 This is a graph of the loss function during the training process of the hybrid neural network model in an embodiment of the present invention; Figure 6 These are comparison diagrams of the displacement predicted by the hybrid neural network model and the actual displacement by the finite element method in the embodiments of the present invention: (a) is a comparison of the displacement time history of column #1-L under the first ground motion; (b) is a comparison of the displacement time history of column #6-L under the first ground motion; (c) is a comparison of the displacement time history of column #1-L under the second ground motion; and (d) is a comparison of the displacement time history of column #6-L under the second ground motion. Detailed Implementation

[0017] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 like Figure 1 As shown, this invention provides a deep learning-based method for predicting structural seismic response, comprising the following steps: S1. Obtain structural information and site category information of the target structure, construct a finite element model, obtain multiple ground motion acceleration time history data for nonlinear time history analysis, extract displacement response time histories of key parts, and form a dataset.

[0020] The target structure is a prefabricated assembled pile-slab structure. Specifically, S1 includes the following sub-steps: S11. Obtain the structural information and site category information of the target structure, and construct a finite element model; the target structure includes at least one precast column and one precast bearing plate, and the top of the precast column is fixedly connected to the precast bearing plate through a rigid sleeve.

[0021] In this embodiment, the target structure is a highway pile-slab bridge structure. The precast columns are hollow precast pipe piles, with prestressed steel bars and ordinary steel bars arranged along the length of the precast pipe piles, and the lower end is fixed to the foundation; the rigid sleeve is a steel connector pre-embedded in the top of the precast column or in the precast bearing plate, and is consolidated by cast-in-place concrete or mechanical connection.

[0022] When constructing the finite element model, the precast columns adopt force-based nonlinear beam-column elements with multiple Gaussian integration points along the length direction. The concrete adopts the Concrete01 uniaxial material model, the steel bars and prestressed steel bars adopt the Steel01 bilinear material model, the precast bearing plate adopts shell elements, and the bottom of the structure simulates the structure-foundation interaction through zero-length elements.

[0023] The material parameters used in this embodiment are shown in the table below: Table 1 Material Parameters (Unit: MPa)

[0024] in, This indicates the ultimate compressive strength of the reinforcing steel. Indicates the compressive strength of concrete; This indicates the ultimate compressive strength of concrete; This represents the elastic modulus, and the unit is MPa. This represents the peak compressive strain of the concrete. It represents the ultimate compressive strain of concrete.

[0025] S12. Acquire multiple ground motion acceleration time history data, select them according to the characteristic period corresponding to the site category information, and select them from historical strong earthquake records of different magnitudes and epicentral distances; fill or truncate each ground motion acceleration time history data to make all data have a uniform time step length and sampling frequency.

[0026] Based on the "Code for Seismic Design of Highway Engineering" (JTG / T2231-01-2020) and the actual site conditions of the project, the horizontal design acceleration response spectrum for the corresponding site type is generated. The E2 design response spectrum (with a seismic return period of 2000 years) can be calculated using the following formula: ; ; in, To design the peak acceleration of the response spectrum, The characteristic period of the site where the structure is located; the characteristic period of the ground motion at the site in this embodiment. It is 0.35s. Indicates the natural period of the structure. Indicates the start period of the platform segment. This is the seismic importance coefficient. Indicates the site coefficient. This is the damping adjustment coefficient. This represents the basic horizontal peak ground acceleration.

[0027] Based on the generated response spectrum, 172 raw ground motion data were selected from the Pacific Earthquake Engineering Research Center (PEER) strong-motion database, with magnitudes ranging from 5.6 to 6.9 and epicentral distances ranging from 30 to 60 kilometers. The sampling frequency for each ground motion acceleration time history data was 50 Hz. The maximum length of 7650 sample points was selected as the uniform length. For ground motion records with fewer than 7650 sample points, zeros were padded to the end of the time history to reach the uniform length.

[0028] S13. Input each ground motion acceleration time history data into the finite element model for nonlinear time history analysis, extract the displacement response time history at the top of the precast column, and form a dataset with the ground motion acceleration time history as input and the corresponding displacement response time history as output.

[0029] The target structure includes multiple precast columns (e.g., 6 columns) spaced longitudinally. In S13, the displacement response time history of the top of each precast column is extracted so that the output dimension of the dataset corresponds to the number of precast columns. The target structure is a multi-span continuous structure (e.g., 10 spans, total length 60m). The load-bearing plates of each span are continuous as a whole. The precast columns are arranged symmetrically or at equal intervals below each span. The load-bearing plate section between adjacent precast columns forms a span.

[0030] The dataset was divided into training, validation, and prediction sets in a ratio of 7:1.5:1.5. The specific partition dimensions are shown in the table below. Table 2. Seismic motion dataset partitioning

[0031] The output dimension is 6, corresponding to the column top displacement response time history of the 6 prefabricated columns (#1-L to #6-L) selected along the longitudinal direction in this embodiment; S2. Construct a CNN-LSTM-Attention hybrid neural network model, train it using the dataset, and obtain an earthquake response prediction model.

[0032] In this embodiment, S2 specifically includes the following sub-steps: S21. Construct a CNN-LSTM-Attention hybrid neural network model, which includes a convolutional neural network module, a long short-term memory network module, and an attention mechanism module.

[0033] like Figure 2 As shown, Long Short-Term Memory (LSTM) networks control information flow through gating mechanisms (forget gate, input gate, and output gate), effectively capturing long-term dependency features in sequences. At time steps... (in (where the sequence length is) the first The relevant calculation formulas for the LSTM cells are as follows: ; ; ; ; ; ; in, For the first Layer in time step Input status, Output for the forget gate. For input gate output, Candidate memory units, For output gate output, For the unit state, In hidden state, Indicates the first The weight matrix of each gate unit in the layer; This represents the bias term for the corresponding gating unit. (Symbol) The symbol represents the sigmoid function, and tanh represents the hyperbolic tangent function; This represents the Hadamard product (i.e., the element-wise product between neurons). LSTM improves its modeling ability by learning the dependencies between the input and output sequences.

[0034] like Figure 3 As shown, the attention mechanism, through dynamic weight allocation, selectively strengthens features in the input that have a significant impact on the prediction result, while suppressing redundant or invalid information, thereby significantly improving the model's efficiency in utilizing effective features. In time series modeling, the attention mechanism is often used in conjunction with LSTM networks. Its core idea is to enable the model to selectively focus on the information in the input sequence that is most important to the current task. The calculation formula for the attention mechanism is as follows: ; ; ; in, Key vectors for all time steps The matrix formed This indicates that after processing by the LSTM layer, at time step The output vector, It is the query vector determined by the final hidden state of the encoder; and These represent the weight matrices used to transform the query and key vectors in the attention mechanism, respectively. Indicates the first The similarity between the key vector and the query vector at each time step is derived from their linear transformation and is used to measure the importance of this location information to the current task. This is the scaling factor.

[0035] The obtained similarity scores are normalized using the Softmax function, limiting the resulting attention weights to the [0,1] interval and summing them to 1. ; in, Dummy variables in the summation of the denominator (traversing all time steps) ), Indicates the first position in the input sequence The attention score corresponding to each time step reflects the importance of the features at that time step in the current task. This is the similarity score for the current time step.

[0036] The normalized attention weights are then summed with their corresponding input vectors to generate the final attention context vector. ; in, The weighted feature vectors (weighted at each time step) result in the final attention context vector consisting of all... sum.

[0037] The core of the attention mechanism lies in assigning different importance weights to different features in the input sequence through a weight allocation strategy, thereby highlighting key information that has a significant impact on the prediction results and effectively improving feature utilization efficiency and overall prediction performance.

[0038] like Figure 4 As shown, the convolutional neural network module consists of two one-dimensional convolutional layers. The first convolutional layer has 64 filters and a kernel size of 5, while the second convolutional layer has 128 filters and a kernel size of 3. Each convolutional layer is followed by a ReLU activation function and batch normalization. The input-output mapping relationship of the convolutional layer is as follows: ; ; in, and They represent the first Input and output features of convolutional operations. Indicates the first Layer Each output feature For output channel index, This represents the depth dimension of the first convolutional layer. Indicates the first The weight parameters corresponding to the receptive field in a convolutional layer. Indicates the first The first layer of the input feature map One channel, Indicates the first Layer The bias term for each output channel. This is the output feature index.

[0039] The Long Short-Term Memory (LSTM) network module consists of two LSTM layers, each with 128 hidden units. The attention mechanism module receives the output sequence of the LSTM layer, calculates the attention weights through a fully connected layer and normalizes them using Softmax, then sums the weighted output sequence to obtain the attention context vector, and finally outputs the shift response prediction value through a fully connected layer.

[0040] S22. Train the hybrid neural network model using the dataset until it converges to obtain the earthquake response prediction model. During training, a Dropout strategy is used with a Dropout ratio of 0.3; the optimizer is AdamW with a first-order moment decay rate of 0.9, a second-order moment decay rate of 0.999, and a weight decay coefficient of 10. -6 Add an L2 regularization term to the loss function, with the regularization coefficient set to 10. -6 An early stop strategy is adopted, which stops training when the validation loss does not decrease after 50 consecutive training rounds, with the minimum number of training rounds set to 100 rounds.

[0041] like Figure 5 As shown, both training loss and validation loss show a decreasing trend and the distance between them is reasonable. The validation loss decreased from about 0.32 to about 0.1, indicating that the model has good convergence and is well trained without overfitting.

[0042] S3. Input the time history of the ground motion acceleration to be predicted into the seismic response prediction model, and output the time history data of the structural seismic predicted displacement response.

[0043] In this embodiment, the structural seismic prediction displacement response time history data includes the displacement response time history and peak displacement of the tops of all precast columns. For example... Figure 6 (a) and Figure 6 As shown in (b), the maximum horizontal displacement of column #1-L caused by the first ground motion was 10.63 mm, while the model predicted 9.68 mm (error 0.95 mm); the actual displacement of column #6-L was 11.4 mm, while the prediction was 10.31 mm (error 1.09 mm). Figure 6 (c) and Figure 6 As shown in (d), the actual earthquake motion for column #1-L was 12.85 mm, while the predicted motion was 12.05 mm (error 0.8 mm); for column #6-L, the actual earthquake motion was 13.88 mm, while the predicted motion was 12.84 mm (error 1.04 mm). This network model can accurately predict the seismic displacement response of the structure.

[0044] S4. Prediction accuracy assessment.

[0045] In this embodiment, S4, a prediction accuracy evaluation step, is also included. S4 specifically includes: The root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination were used. () is used as an evaluation metric to assess the prediction accuracy of hybrid neural network models.

[0046] The formula for calculating the root mean square error is: ; in, The root mean square error, For the first The actual values ​​of the structural seismic response of each sample (i.e., the displacement response values ​​obtained from finite element analysis). For the first The model prediction value for each sample. The number of samples (in this embodiment) =25). RMSE assigns a higher penalty weight to larger errors, which can effectively reflect the overall degree of deviation of the predicted values.

[0047] The formula for calculating the mean absolute error is: ; in, This represents the mean absolute error. It directly reflects the average level of the absolute difference between the predicted and actual values, and has good interpretability.

[0048] The formula for calculating the coefficient of determination is: ; in, As the coefficient of determination, This is the mean of the actual values ​​of all samples. The value ranges from [0,1], and the closer it is to 1, the better the model fits the actual data.

[0049] The determination coefficients of some prediction set samples in this embodiment are shown in the table below after calculation: Table 3. Prediction set accuracy evaluation results (partial sample)

[0050] The results show that the vast majority of predicted samples All values ​​reached above 0.94, with some reaching above 0.98. The root mean square error and mean absolute error were both around 0.001 and 0.0001, respectively, proving that the model proposed in this invention has extremely high prediction accuracy and generalization ability.

[0051] Therefore, this invention employs the aforementioned deep learning-based structural seismic response prediction method. By integrating convolutional neural networks, long short-term memory networks, and attention mechanisms, a CNN-LSTM-Attention hybrid neural network model is constructed, enabling accurate prediction of the displacement response of prefabricated pile-slab structures under seismic loading. This method effectively reduces the time required for traditional finite element analysis and accurately predicts the seismic response characteristics of the structure under different ground motions in a short time.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A deep learning-based method for predicting structural seismic response, characterized in that, Includes the following steps: S1. Obtain the structural information and site category information of the target structure. The structural information includes the connection information of precast columns, precast bearing plates and rigid sleeves. The site category information includes the site category and characteristic period. Construct a finite element model, obtain multiple ground motion acceleration time history data for nonlinear time history analysis, extract the displacement response time history of key parts, the key parts being the top of the precast columns, and form a dataset. S2. Construct a CNN-LSTM-Attention hybrid neural network model, train it using the dataset, and obtain an earthquake response prediction model; S3. Input the time history of the ground motion acceleration to be predicted into the seismic response prediction model, and output the time history data of the structural seismic predicted displacement response.

2. The method according to claim 1, characterized in that, The target structure described in S1 is a prefabricated assembled pile-slab structure; S1 specifically includes: S11. Obtain the structural information and site category information of the target structure, and construct a finite element model; the target structure includes at least one precast column and one precast bearing plate, and the top of the precast column is fixedly connected to the precast bearing plate through a rigid sleeve. S12. Acquire multiple ground motion acceleration time history data, select them according to the characteristic period corresponding to the site category information, and select them from historical strong earthquake records of different magnitudes and different epicentral distances; fill or truncate each ground motion acceleration time history data to make all data have a uniform time step length and sampling frequency. S13. Input each ground motion acceleration time history data into the finite element model for nonlinear time history analysis, extract the displacement response time history at the top of the precast column, and form a dataset with the ground motion acceleration time history as input and the corresponding displacement response time history as output.

3. The method according to claim 2, characterized in that, In S13, the dataset is divided into a training set, a validation set, and a prediction set in a ratio of 7:1.5:1.

5.

4. The method according to claim 2, characterized in that, In S11: The precast column is a hollow precast pipe pile. The precast pipe pile is equipped with prestressed steel bars and ordinary steel bars along its length, and its lower end is fixed to the foundation. The steel sleeve is a steel connector pre-embedded in the top of the precast column or in the precast bearing plate, and is consolidated by cast-in-place concrete or mechanical connection. When constructing the finite element model, the precast columns adopt force-based nonlinear beam-column elements with multiple Gaussian integration points along the length direction. The concrete adopts the Concrete01 uniaxial material model, the steel bars and prestressed steel bars adopt the Steel01 bilinear material model, the precast bearing plate adopts shell elements, and the bottom of the structure simulates the structure-foundation interaction through zero-length elements.

5. The method according to claim 2, characterized in that, The target structure includes multiple prefabricated columns spaced longitudinally. In S13, the displacement response time history of the top of each prefabricated column is extracted so that the output dimension of the dataset corresponds to the number of prefabricated columns. The target structure is a multi-span continuous structure, with the load-bearing plates of each span forming a continuous whole. The prefabricated columns are arranged symmetrically or at equal intervals below each span, and the load-bearing plate sections between adjacent prefabricated columns form a span.

6. The method according to claim 1, characterized in that, S2 specifically includes: S21. Construct a CNN-LSTM-Attention hybrid neural network model, which includes a convolutional neural network module, a long short-term memory network module, and an attention mechanism module. The convolutional neural network module consists of two one-dimensional convolutional layers. The first convolutional layer has 64 filters and a kernel size of 5, while the second convolutional layer has 128 filters and a kernel size of 3. Each convolutional layer is followed by a ReLU activation function and batch normalization. The output of the convolutional neural network module is flattened in terms of feature dimension by a fully connected layer to form a temporal feature sequence, which is then input to the long short-term memory network module. The long short-term memory network module consists of two LSTM layers, each with 128 hidden units. The attention mechanism module receives the output sequence of the LSTM layer, calculates the attention weights through a fully connected layer, normalizes them using Softmax, and then sums the output sequence by weight to obtain the attention context vector. Finally, it outputs the displacement response prediction value through a fully connected layer. S22. Train the hybrid neural network model using the dataset until it converges to obtain the earthquake response prediction model. During training, a Dropout strategy is used with a Dropout ratio of 0.

3. The optimizer is AdamW, with a first-order moment decay rate of 0.9, a second-order moment decay rate of 0.999, and a weight decay coefficient of 10. -6 Add an L2 regularization term to the loss function, with the regularization coefficient set to 10. -6 An early stop strategy is adopted, which stops training when the validation loss does not decrease after 50 consecutive training rounds, with the minimum number of training rounds set to 100 rounds.

7. The method according to claim 1, characterized in that, In S3, the structural seismic prediction displacement response time history data includes the displacement response time history and peak displacement of the top of all precast columns.

8. The method according to claim 1, characterized in that, It also includes S4, prediction accuracy evaluation; S4 specifically includes: The root mean square error, mean absolute error, and coefficient of determination are used as evaluation indicators to assess the prediction accuracy of the hybrid neural network model. The formula for calculating the root mean square error is: ； in, The root mean square error, For the first Actual values ​​of structural seismic response for each sample. For the first The model prediction value for each sample. The number of samples; The formula for calculating the mean absolute error is: ； in, Mean absolute error; The formula for calculating the coefficient of determination is: ； in, As the coefficient of determination, This is the mean of the actual values ​​of all samples.

Images