PSCFST bridge seismic acceleration response prediction method based on DL-iTransform

Abstract

The invention discloses a PSCFST bridge seismic acceleration response prediction method based on DL-iTransform. The method comprises the following steps: firstly, obtaining vibration acceleration time sequence data input by the south side and the north side of a bridge and normalizing the data; inputting the data into LSTM, extracting short-term time dependence characteristics through a gating mechanism, and outputting a hidden state sequence; then, on the basis of the sequence, utilizing a differential attention mechanism to calculate feature differences of adjacent time steps, dynamically distribute weights and focus key sudden change moment features to obtain a weighted feature sequence; inputting the sequence into iTransform, capturing long-term dependence and multivariable correlation through a global self-attention mechanism of the sequence, and outputting global feature representation; and finally, through full-connection layer regression prediction and reverse normalization, a bridge superstructure seismic acceleration response prediction value is obtained. According to the method, the feature difference of adjacent time steps is dynamically calculated and the weight is allocated through a differential attention mechanism, and the key sudden change moment of the seismic oscillation response is adaptively focused without signal preprocessing.

Description

Technical Field

[0001] The present invention relates to the technical field of bridge engineering seismic performance analysis, and in particular to a PSCFST bridge seismic acceleration response prediction method based on DL-iTransformer. Background Art

[0002] Segmentally assembled, self-centering concrete-filled steel tube (PSCFST) bridges, due to their excellent seismic toughness and self-resetting capabilities, have become a research hotspot for bridge engineering in high-intensity earthquake zones. These bridges significantly reduce post-earthquake residual displacement and repair costs by integrating the precast segments with the concrete-filled steel tube, combined with the reset mechanism of post-tensioned prestressed steel strands. However, the dynamic response of PSCFST bridges under strong earthquakes exhibits highly nonlinear characteristics, involving slip at the steel tube-concrete interface, opening and closing of segment joints, and nonlinear behavior of energy dissipation devices. This poses significant challenges to predicting the seismic acceleration response.

[0003] Early prediction methods relied on physics-driven models, such as nonlinear time-history analysis and finite element modeling. These methods required detailed simulation of material and geometric nonlinearities and boundary conditions, resulting in high computational complexity and reliance on expert experience. With the development of artificial intelligence (AI), data-driven methods have gradually become mainstream. For example, the VMD-CNN-BiLSTM model, which combines variational mode decomposition (VMD), convolutional neural networks (CNN), and bidirectional long short-term memory networks (BiLSTM), captures the multi-scale characteristics of earthquake motion. However, its complex architecture still leaves room for optimization and suffers from the following technical drawbacks:

[0004] 1. Pre-set dependency and computational redundancy of signal decomposition: The VMD-CNN-BiLSTM model relies on VMD decomposition to extract frequency features. However, the number of decomposition layers must be manually determined using the energy difference criterion. Parameter selection directly affects prediction accuracy, and multiple iterative optimizations result in high computational costs, making it difficult to meet real-time monitoring requirements.

[0005] Although the Conv1D-LSTM model cancels VMD, the convolution kernel parameters of Conv1D still need to be tuned through grid search, which increases the preprocessing complexity.

[0006] 2. Insufficient ability to capture dynamic features: Existing models, such as BiLSTM and LSTM, process time series data through fixed gating mechanisms, resulting in rigid weight allocation for sudden changes such as seismic peaks and high-frequency vibrations. For example, the VMD-CNN-BiLSTM model achieved an RMSE of 0.0163 at an RSN292-120% amplitude. The fitting error for nonlinear sudden changes is still high, exposing its inability to dynamically focus on "critical time steps."

[0007] 3. The "VMD+CNN+BiLSTM" three-module serial architecture has functional overlap. The IMF components after VMD decomposition need to be input into CNN processing separately, resulting in too many model parameters and time-consuming training.

[0008] Although Conv1D-LSTM simplifies the process, the local feature extraction of Conv1D and the time series modeling of LSTM are still serial processing, and the modeling ability of multivariate correlation is limited.

[0009] 4. Data preprocessing complexity and generalization limitations: Existing technologies rely on preprocessing steps such as linear interpolation to fill missing values and maximum-minimum normalization, and they are highly dependent on high-quality data. In real-world projects, seismic data may be affected by noise or have insufficient samples, resulting in reduced model robustness. Summary of the Invention

[0010] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a PSCFST bridge seismic acceleration response prediction method based on DL-iTransformer.

[0011] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0012] The present invention provides a PSCFST bridge earthquake acceleration response prediction method based on DL-iTransformer, comprising the following steps:

[0013] S1. Data preprocessing: Obtain input seismic acceleration time series data on the south and north sides of the bridge, normalize the time series data, and obtain standardized input data;

[0014] S2. Short-term time series feature extraction: The standardized input data is input into the LSTM. Through the LSTM's forget gate, input gate, and output gate mechanism, the short-term time dependency features of the earthquake motion data are extracted and the LSTM hidden state sequence is output.

[0015] S3, differential attention feature focusing: Based on the LSTM hidden state sequence, the differential attention mechanism is used to calculate the feature differences between adjacent time steps, dynamically assign attention weights, focus on the key mutation moment features of the earthquake response, and obtain a weighted feature sequence;

[0016] S4. Global Time Series Modeling: The weighted feature sequence is input into the improved Transformer model iTransformer. The global self-attention mechanism of iTransformer captures the long-term temporal dependencies and multivariate correlations of the ground motion data and outputs a global feature representation.

[0017] S5. Response prediction: The global feature representation is input into the fully connected layer for regression prediction, and the prediction results are denormalized to obtain the predicted value of the earthquake acceleration response of the bridge superstructure.

[0018] As a preferred technical solution of the present invention, in step S2, the forget gate, input gate, and output gate of the LSTM are calculated as follows:

[0019] The forget gate calculates the degree of forgetting of the unit state at the previous moment at the current moment:

[0020] f t =σ(W f ·[h t-1 ,x t ]+b f );

[0021] Input gate i t Calculate the input information at the current moment:

[0022] i t =σ(W i ·[h t-1 ,x t ]+b i );

[0023] Output gate o t Calculate the hidden state output at the current moment:

[0024] o t =σ(W o ·[h t-1 ,x t ]+b o );

[0025] In the above formula, σ is the sigmoid activation function, W f 、W i 、W O is the weight matrix, b f 、b i 、b o is the bias vector, h t-1 is the hidden state at the previous moment, x t Enter data for the current moment.

[0026] As a preferred technical solution of the present invention, the calculation process of the differential attention mechanism in step S3 includes:

[0027] The LSTM hidden state sequence H=[h1,h2,…,h T ] is mapped into query vector Q, key vector K and value vector V;

[0028] Calculate the basic attention weight matrix where d k is the key vector dimension;

[0029] A learnable scalar λ is introduced to adjust the attention difference and calculate the differential attention weight:

[0030]

[0031] The value vector V is weighted and summed by the differential attention weight β to obtain a weighted feature sequence that focuses on key mutation features.

[0032] As a preferred technical solution of the present invention, the global self-attention mechanism of iTransformer in step S4 is specifically as follows:

[0033] The dimension inversion embedding method is used to embed the south and north side earthquake acceleration time series data as independent variable sequences into tokens, preserving the independence between variables.

[0034] Calculate the correlation between different time steps and variables through multi-head self-attention:

[0035] MultiHead(Q,K,V)=Concat(head1,…,head h )W o ;

[0036] Each attention head

[0037] The FFN performs nonlinear transformation on the multi-head attention output to generate a global feature representation.

[0038] As a preferred technical solution of the present invention, the normalization process in step S1 adopts the maximum-minimum normalization method to normalize the data to the range of [0, 1]. The specific formula is:

[0039]

[0040] Among them, x min and x max are the minimum and maximum values of the input data respectively;

[0041] The denormalization process in step S5 is the inverse process of normalization, which restores the actual physical dimension of the predicted value. The specific formula is:

[0042] x pred =x norm ·(x max -x min )+x min .

[0043] As a preferred technical solution of the present invention, the method further includes a model training step:

[0044] The training dataset is constructed using shaking table test data and finite element simulation data. The input data is the south / north side earthquake acceleration time series, and the output data is the bridge superstructure acceleration response time series.

[0045] Using RMSE as the loss function, the model parameters of LSTM, differential attention mechanism, and iTransformer are optimized through the backpropagation algorithm:

[0046]

[0047] in, is the predicted value, y i is the true value, and m is the number of samples.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1. Improve prediction accuracy and dynamic feature capture capability: This invention uses the differential attention mechanism to dynamically calculate the feature differences of adjacent time steps and assign weights, adaptively focusing on the key mutation moments of the earthquake response without the need for signal preprocessing; this mechanism can replace VMD decomposition, avoid feature loss caused by improper setting of the decomposition layer number, and improve the ability to capture high-frequency vibration and time-varying features, making R 2 , RMSE and other indicators are further optimized.

[0050] 2. Simplify model structure and reduce computational costs: This paper introduces iTransformer, replacing BiLSTM's bidirectional time series modeling with a global self-attention mechanism, solving the vanishing gradient problem of long sequences and simplifying the model structure. Combined with the basic time series modeling capabilities of LSTM, while retaining time series features, the training time is shortened to less than 1 / 2 of the existing technology, significantly improving engineering practicality.

[0051] 3. Enhanced robustness and generalization performance: This invention achieves multi-component collaborative optimization: LSTM retains the core time series modeling capability to ensure the capture of long-term dependencies on earthquake motions; differential attention dynamically assigns weights to reduce noise interference; iTransformer globally models spatiotemporal characteristics to enhance the prediction stability of extreme earthquake motions; in finite element simulation data, the RMSE can be further reduced to below 0.015, and the MAE is reduced by about 25%, verifying its applicability in high-intensity earthquake zones.

[0052] 4. Reduce manual intervention and data dependence: This invention directly processes raw time series data through an end-to-end deep learning architecture, avoiding signal decomposition and feature engineering, and reducing manual intervention. At the same time, the differential attention mechanism has lower requirements on data quality and can still maintain high accuracy under limited samples, providing a reliable solution for scenarios with insufficient data in actual engineering.

[0053] 5. Technical compatibility and scalability: Through a lightweight architecture, LSTM, differential attention, and iTransformer, this invention supports flexible integration of other modules to adapt to the seismic requirements of different bridge types. In addition, the model can be directly deployed on edge computing devices to meet the low-latency requirements of real-time monitoring scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0055] Figure 1 It is a schematic diagram of the overall architecture of the present invention;

[0056] Figure 2 This is the calculation flow chart of the differential attention mechanism of the present invention; DETAILED DESCRIPTION

[0057] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0058] The present invention provides a PSCFST bridge earthquake acceleration response prediction method based on DL-iTransformer, comprising the following steps:

[0059] S1. Data preprocessing: Obtain input seismic acceleration time series data on the south and north sides of the bridge, normalize the time series data, and obtain standardized input data;

[0060] S2. Short-term time series feature extraction: The standardized input data is input into the LSTM. Through the LSTM's forget gate, input gate, and output gate mechanism, the short-term time dependency features of the earthquake motion data are extracted and the LSTM hidden state sequence is output.

[0061] S3, differential attention feature focusing: Based on the LSTM hidden state sequence, the differential attention mechanism is used to calculate the feature differences between adjacent time steps, dynamically assign attention weights, focus on the key mutation moment features of the earthquake response, and obtain a weighted feature sequence;

[0062] S4. Global Time Series Modeling: The weighted feature sequence is input into the improved Transformer model iTransformer. The global self-attention mechanism of iTransformer captures the long-term temporal dependencies and multivariate correlations of the ground motion data and outputs a global feature representation.

[0063] S5. Response prediction: The global feature representation is input into the fully connected layer for regression prediction, and the prediction results are denormalized to obtain the predicted value of the earthquake acceleration response of the bridge superstructure.

[0064] Furthermore, in step S2, the forget gate, input gate, and output gate of LSTM are calculated as follows:

[0065] The forget gate calculates the degree of forgetting of the unit state at the previous moment at the current moment:

[0066] f t =σ(W f ·[h t-1 ,x t ]+b f );

[0067] Input gate i t Calculate the input information at the current moment:

[0068] i t =σ(W i ·[h t-1 ,x t ]+b i );

[0069] Output gate o t Calculate the hidden state output at the current moment:

[0070] o t =σ(W o ·[h t-1 ,x t ]+b o );

[0071] In the above formula, σ is the sigmoid activation function, W f 、W i 、W O is the weight matrix, b f 、b i 、b o is the bias vector, h t-1 is the hidden state at the previous moment, x t Enter data for the current moment.

[0072] Furthermore, the calculation process of the differential attention mechanism in step S3 includes:

[0073] The LSTM hidden state sequence H=[h1,h2,…,h T ] is mapped into query vector Q, key vector K and value vector V;

[0074] Calculate the basic attention weight matrix where d k is the key vector dimension;

[0075] A learnable scalar λ is introduced to adjust the attention difference and calculate the differential attention weight:

[0076]

[0077] The value vector V is weighted and summed by the differential attention weight β to obtain a weighted feature sequence that focuses on key mutation features.

[0078] Furthermore, the global self-attention mechanism of iTransformer in step S4 is specifically as follows:

[0079] The dimension inversion embedding method is used to embed the south and north side earthquake acceleration time series data as independent variable sequences into tokens, preserving the independence between variables.

[0080] Calculate the correlation between different time steps and variables through multi-head self-attention:

[0081] MultiHead(Q,K,V)=Concat(head1,…,hed h )W o ;

[0082] Each attention head

[0083] The FFN performs nonlinear transformation on the multi-head attention output to generate a global feature representation.

[0084] Furthermore, the normalization process in step S1 uses the maximum-minimum normalization method to normalize the data to the range of [0, 1]. The specific formula is:

[0085]

[0086] Among them, x min and x max are the minimum and maximum values of the input data respectively;

[0087] The denormalization process in step S5 is the inverse process of normalization, which restores the actual physical dimension of the predicted value. The specific formula is:

[0088] x pred =x norm·(x max -x min )+x min .

[0089] Furthermore, the method also includes a model training step:

[0090] The training dataset is constructed using shaking table test data and finite element simulation data. The input data is the south / north side earthquake acceleration time series, and the output data is the bridge superstructure acceleration response time series.

[0091] The root mean square error (RMSE) is used as the loss function, and the model parameters of LSTM, differential attention mechanism, and iTransformer are optimized through the back propagation algorithm:

[0092]

[0093] in, is the predicted value, y i is the true value, and m is the number of samples.

[0094] Example 1

[0095] Multi-model comparison experiment based on shaking table test data

[0096] The data used is from the PSCFST bridge shaking table test, consisting of 116 time series data sets ranging from 10% to 60% for five types of ground motion: RSN006, RSN292, RSN953, AGM1, and AGM2. Each data set includes a time series of south X-direction and north Y-direction input ground motion accelerations, with a sampling frequency of 256 Hz and a duration of 40 seconds, for a total of 10,240 data points, as well as the corresponding time series of the superstructure acceleration response.

[0097] Comparison group:

[0098] 1. VMD-CNN-BiLSTM: The process is "Variational Mode Decomposition (VMD) → One-Dimensional Convolutional Neural Network (CNN) → Bidirectional Long Short-Term Memory Network (BiLSTM)". The number of VMD decomposition layers must be preset to 4.

[0099] 2. Conv1D-LSTM: Uses one-dimensional convolution to extract spatial features and combines it with LSTM to process temporal dependencies.

[0100] 3. XGBoost: A gradient-boosted tree model as a traditional machine learning baseline.

[0101] The evaluation indicators are:

[0102] Root mean square error RMSE: measures the overall deviation between the predicted value and the true value. The formula is:

[0103]

[0104] Mean absolute error MAE: reflects the average absolute deviation between the predicted value and the true value. The formula is:

[0105]

[0106] Coefficient of determination R 2 : Evaluate how well the model fits the data. The formula is:

[0107]

[0108] First, the data is normalized: the maximum-minimum normalization is applied to the input south / north side seismic acceleration and output response data. The formula is:

[0109]

[0110] Among them, x min and x max are the minimum and maximum values of the input data, respectively. The test set uses the normalization parameters of the training set.

[0111] The data is then divided: 80% of the data is randomly selected from the training set, with 93 groups, including 34 groups of RSN006, 20 groups of RSN953, 14 groups of AGM1, and 14 groups of AGM2.

[0112] Test set: The remaining 20% of the data consists of 23 groups, containing only RSN292-60% amplitude data. Finally, linear interpolation is used to fill in the missing data points. The model parameter settings are shown in Table 1:

[0113]

[0114] The final experimental results of Table 1 are shown in Table 2:

[0115]

[0116] Table 2

[0117] It can be seen that the R 2 Compared with VMD-CNN-BiLSTM, it is improved by 3.3% and MAE is reduced by 54.1%, indicating that it can capture the high-frequency components of earthquake motion, such as the peak acceleration at 15s, more accurately. The peak error rate is only 5.2%, while the existing technologies are all over 10%.

[0118] The training time is shortened by 54.1% compared with VMD-CNN-BiLSTM, mainly due to the elimination of VMD decomposition and the parallel computing characteristics of iTransformer.

[0119] To verify the role of the differential attention mechanism and iTransformer, the following ablation experiment is designed to remove the differential attention (only LSTM+iTransformer):

[0120] R 2 It dropped to 0.942, the RMSE increased to 0.0145 (+30.4%), and the MAE increased to 0.0084 (+35.5%).

[0121] Conclusion: Differential attention is distributed through dynamic weights, formula:

[0122]

[0123] Effectively focus on key mutation moments, such as increasing the weight of peak points by 20%. After removal, the model's response ability to sudden vibrations is significantly reduced.

[0124] Replacing iTransformer with BiLSTM (DL-BiLSTM): Training time increased to 250s, an increase of +35.1%, R 2 It dropped to 0.935, and the RMSE increased to 0.0138, with an increase of +35.1% and +23.2%.

[0125] Conclusion: iTransformer's global self-attention mechanism can parallelize the temporal correlation of multiple variables, improving its modeling capability by approximately 5% compared to BiLSTM's unidirectional temporal processing.

[0126] Example 2

[0127] Large-amplitude ground motion prediction experiment (finite element simulation data)

[0128] For the strong earthquake scenario not covered by the shaking table test, RSN292-120% amplitude, the OpenSees fiber model was used to generate data to verify the generalization ability of the model under extreme conditions.

[0129] The dataset uses the PSCFST bridge finite element model, inputs the RSN292 ground motion and amplifies it to 70%-120% amplitude, generating 80 sets of large-amplitude data, each set lasting 40 seconds and with a sampling frequency of 256 Hz.

[0130] The data feature input is the south / north side seismic acceleration, with a peak value of 1.8g, and the output is the acceleration response of the superstructure, including nonlinear vibration caused by the opening and closing of segment joints.

[0131] The model was pre-trained using 116 sets of shaker test data and then fine-tuned using 80 sets of simulated data. The accuracy of the peak response prediction was focused on the extreme 120% amplitude.

[0132] The experimental results are shown in Table 3:

[0133]

[0134] Table 3

[0135] It can be seen that the peak error of the proposed method at 120% amplitude is 7.8%, which is only 1 / 2 of that of VMD-CNN-BiLSTM. Thanks to the global self-attention mechanism of iTransformer, which models the dependencies of long sequences, it avoids the gradient vanishing problem of BiLSTM under strong earthquakes. The gradient norm is maintained at 0.5-1.0, while that of BiLSTM drops to below 0.1.

[0136] In the high-frequency vibration segment of 20-25s with a frequency >10Hz, the RMSE of the present invention is reduced by 52% compared with the prior art, proving that the differential attention mechanism effectively filters low-frequency noise.

[0137] Example 3

[0138] Robustness test under different earthquake motion amplitudes

[0139] In order to verify the adaptability of the model under different amplitudes, 10%, 30%, 60%, and 120% amplitude data of RSN292 earthquake motion were selected, with 20 groups each, and the performance fluctuations of the present invention and VMD-CNN-BiLSTM were compared.

[0140] The comparison of model robustness under different amplitudes is shown in Table 4:

[0141]

[0142] Table 4

[0143] It can be seen from this that the method of the present invention has the following R 2 All of them are above 0.95, with fluctuation range ≤1%, while R 2 The fluctuation is 1.6%, from 0.916 to 0.932, indicating that DL-iTransformer is more robust to changes in ground motion amplitude.

[0144] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A PSCFST bridge seismic acceleration response prediction method based on DL-iTransformer, characterized by: The following steps are involved: S1. Data preprocessing: Obtain input seismic acceleration time series data on the south and north sides of the bridge, normalize the time series data, and obtain standardized input data; S2. Short-term time series feature extraction: The standardized input data is input into the LSTM. Through the LSTM's forget gate, input gate, and output gate mechanism, the short-term time dependency features of the earthquake motion data are extracted and the LSTM hidden state sequence is output. S3, differential attention feature focusing: Based on the LSTM hidden state sequence, the differential attention mechanism is used to calculate the feature differences between adjacent time steps, dynamically assign attention weights, focus on the key mutation moment features of the earthquake response, and obtain a weighted feature sequence; S4. Global Time Series Modeling: The weighted feature sequence is input into the improved Transformer model iTransformer. The global self-attention mechanism of iTransformer captures the long-term temporal dependencies and multivariate correlations of the ground motion data and outputs a global feature representation. S5. Response prediction: The global feature representation is input into the fully connected layer for regression prediction, and the prediction results are denormalized to obtain the predicted value of the earthquake acceleration response of the bridge superstructure.

2. The PSCFST bridge seismic acceleration response prediction method based on DL-iTransformer according to claim 1 is characterized in that: In step S2, the calculation method of the forget gate, input gate and output gate of LSTM is as follows: The forget gate calculates the degree of forgetting of the unit state at the previous moment at the current moment: f t =σ(W f ·[h t-1 ,x t ]+b f ); Input gate i t Calculate the input information at the current moment: i t =σ(W i ·[h t-1 ,x t ]+b i ); Output gate o t Calculate the hidden state output at the current moment: the t =σ(W o ·[h t-1 ,x t ]+b o ); In the above formula, σ is the sigmoid activation function, W f 、W i 、W O is the weight matrix, b f 、b i 、b o is the bias vector, h t-1 is the hidden state at the previous moment, x t Enter data for the current moment.

3. The PSCFST bridge seismic acceleration response prediction method based on DL-iTransformer according to claim 1 is characterized in that: The calculation process of the differential attention mechanism in step S3 includes: The LSTM hidden state sequence H=[h1,h2,…,h T ] is mapped into query vector Q, key vector K and value vector V; Calculate the basic attention weight matrix where d k is the key vector dimension; A learnable scalar λ is introduced to adjust the attention difference and calculate the differential attention weight: The value vector V is weighted and summed by the differential attention weight β to obtain a weighted feature sequence that focuses on key mutation features.

4. The PSCFST bridge seismic acceleration response prediction method based on DL-iTransformer according to claim 1 is characterized in that: The global self-attention mechanism of iTransformer in step S4 is specifically as follows: The dimension inversion embedding method is used to embed the south and north side earthquake acceleration time series data as independent variable sequences into tokens, preserving the independence between variables. Calculate the correlation between different time steps and variables through multi-head self-attention: MultiHead(Q,K,V)=Concat(head1,…,head h )W O ; Each attention head The FFN performs nonlinear transformation on the multi-head attention output to generate a global feature representation.

5. The PSCFST bridge seismic acceleration response prediction method based on DL-iTransformer according to claim 1 is characterized in that: The normalization process in step S1 adopts the maximum-minimum normalization method to normalize the data to the range of [0, 1]. The specific formula is: Among them, x min and x max are the minimum and maximum values of the input data respectively; The denormalization process in step S5 is the inverse process of normalization, which restores the actual physical dimension of the predicted value. The specific formula is: x pred =x norm ·(x max -x min )+x min 。 6. The PSCFST bridge seismic acceleration response prediction method based on DL-iTransformer according to claim 1 is characterized in that: The method also includes a model training step: The training dataset is constructed using shaking table test data and finite element simulation data. The input data is the south / north side earthquake acceleration time series, and the output data is the bridge superstructure acceleration response time series. Using RMSE as the loss function, the model parameters of LSTM, differential attention mechanism, and iTransformer are optimized through the backpropagation algorithm: in, is the predicted value, y i is the true value, and m is the number of samples.

Images