Elastic Bragg breakwater structure response prediction method

Abstract

The invention discloses an elastic Bragg breakwater structure response prediction method, and belongs to the technical field of ocean engineering structure dynamic response prediction, and the method comprises the steps: obtaining the front and rear wave surface elevation and motion response data of a breakwater; a double-flow attention mechanism based on physical prior guidance is constructed, and wave-structure interaction is decoupled into two parallel attention flows of structural dynamic evolution and wave-structure coupling feedback; a self-attention mechanism is combined with a time decay physical prior coding structure motion feature, and a bidirectional cross attention mechanism based on a resonance characteristic is combined with a wave-structure phase relation physical prior coding coupling feature; the double-flow features are adaptively integrated through a global context fusion module, and a multi-step prediction sequence is output through linear layer decoding; and a time-frequency domain joint loss function optimization model is adopted, so that the prediction precision is improved. According to the method, physical prior is explicitly embedded into an attention mechanism and is improved by more than 16% compared with an optimal baseline model, and the generalization ability is enhanced by more than 13% under the working condition that waves do not appear.

Description

Technical Field

[0001] This invention belongs to the field of dynamic response prediction technology for marine engineering structures, and specifically relates to a method for predicting the response of an elastic Bragg breakwater structure. Background Technology

[0002] Elastic Bragg breakwaters enhance energy dissipation through the synergistic effects of elastic deformation, wave breaking, and Bragg resonance mechanisms; however, the bidirectional coupling between their structural dynamics and the wave field is highly complex. Traditional theoretical models (such as linear wave theory) struggle to handle nonlinear coupling, while purely data-driven deep learning models (such as LSTM and Transformer) suffer from the following limitations: (1) Lack of time decay characteristics: The unmodeled structural damping effect (i.e., the current state is mainly affected by recent history) leads to prediction results that violate physical laws; (2) Insufficient modeling of resonance phenomenon: Traditional attention mechanism is based on feature similarity calculation, which cannot capture the resonance characteristics determined by the wave-structure phase relationship; (3) Frequency domain information is ignored: learning is supervised only in the time domain, which leads to distortion of the predicted spectrum and cannot guarantee physical conservation.

[0003] In the field of marine engineering, data is scarce and operating conditions are highly variable. Models that rely solely on data-driven approaches are prone to overfitting to specific conditions and have poor generalization ability. There is an urgent need for a prediction method that integrates physical priors, explicitly embedding physical laws such as damping characteristics and phase resonance into the model architecture to achieve high-precision and high-generalization response prediction.

[0004] The motion response of an elastic structure in waves is a highly complex dynamic process, at its core being the bidirectional interaction between wave excitation and structural motion. Structural motion alters the surrounding wave field, and the altered wave field, in turn, influences structural motion. This tight coupling induces resonance at specific frequencies, leading to a sharp amplification of the structural response. This physical mechanism exhibits distinct time-decrease characteristics (damping effect) and phase-synchronization characteristics (resonance effect). Traditional deep learning models, lacking corresponding inductive biases, struggle to effectively extract these key physical features from data.

[0005] Structural response prediction in complex marine environments places extremely high demands on model capabilities due to the complexity of its physical mechanisms and the high-dimensional nonlinearity of the data. However, the field of marine engineering has limited field-measured data samples. Simply relying on data-driven complex network models for training not only leads to difficulty in convergence but also to overfitting to specific operating conditions, resulting in poor generalization ability and insufficient predictive robustness in unfamiliar sea conditions. Summary of the Invention

[0006] This invention proposes a method for predicting the structural response of an elastic Bragg breakwater. It embeds physical laws such as damping characteristics and phase resonance into a dual-stream attention network in the form of structured parameters, solving the problems of missing time decay characteristics, insufficient modeling of resonance phenomena, and neglect of frequency domain information in traditional models.

[0007] The technical solution of the present invention is as follows: A method for predicting the response of an elastic Bragg breakwater structure includes the following steps: Step 1: Arrange multiple wave path detection devices in the experimental water tank, and conduct large-scale simulations under various irregular wave conditions to generate a large amount of wave elevation time series data and structural displacement time series data. Step 2: Construct a two-stream attention network prediction model based on physical prior guidance; the model includes two parallel attention streams, a global context fusion module, and a linear layer; the two parallel attention streams are the first attention stream and the second attention stream, respectively; Step 3: Design a hybrid time-frequency domain loss function to optimize the two-stream attention network prediction model.

[0008] Furthermore, in step 1, the collected wave elevation time series data is used as wave excitation sequence data. The structural displacement time series data is used as the structural motion response sequence data. ;in, For time steps, The feature dimension is used; the continuous time series data is divided into multiple input-output sample pairs by using the sliding time window method.

[0009] Furthermore, in step 2, the first attention stream is a damped bidirectional self-attention module; this module introduces a learnable time decay bias to correct the attention weights in the standard self-attention calculation; the specific working process of the damped bidirectional self-attention module is as follows: First, the structural motion response sequence data... Projection as query matrix Key matrix Sum matrix : (1); in, , , The projective weight matrix is ​​a learnable matrix. For any two time steps and Its standard scaled dot product attention score for: (2); in, It is the softmax function; and These are query matrices Bond matrix The row and number The row number corresponds to the time step number; It is the transpose symbol; Key matrix The lth line; Total number of rows; The attenuation coefficient is calculated based on the time interval between the two time steps. : (3); in, and These are learnable parameters. Represents the sampling interval; The decay coefficient is combined with the attention score to generate a time-decay-weighted attention weight. : (4); Finally, the output of the damped bidirectional self-attention module is obtained. : (5); in, Value matrix The j-th line.

[0010] Furthermore, in step 2, the second attention stream is a resonant-guided bidirectional cross-attention module; this module modulates the standard attention weights by introducing a phase bias term controlled by learnable parameters; at the same time, a two-stage computational paradigm is used to model the bidirectionality of wave-structure interaction. The phase offset term is generated based on the phase relationship between waves and the structure, using the following formula: (6); in, For query matrix row i Bond matrix row j Phase offset term; These are learnable parameters; The modulated attention weights are obtained by modulating the standard attention weights using the phase bias term. : (7); in, For query matrix row i Bond matrix Line 1 Phase offset term; Finally, the value matrix Perform a weighted summation: (8); in, For the attention function with added phase bias term; For the two-stage computational paradigm, the first stage involves processing the structural motion response sequence data. Projection as query matrix Wave excitation sequence data Projection as a key matrix Sum matrix: (9); in, , , The first stage is the learnable weight matrix; the first stage output. for: (10); The second phase will use wave excitation sequence data. Projection as query matrix The output of the first stage Projection as a key matrix Sum matrix : (11); in, , , The weight matrix is ​​the learnable matrix for the second stage; the output of the second stage. for: (12); Second stage output This is the final output of the resonance-guided bidirectional cross-attention module.

[0011] Furthermore, in step 2, the specific working process of the global context fusion module and the linear layer is as follows: First, extract the global context information. : (13); in, It is the softmax function; The weight matrix is ​​a learnable matrix; Then global context information Extended to length L This information is dynamically integrated into the output of the damped bidirectional self-attention module, enabling deep interaction between the two types of physical information. The final linear layer outputs the predicted motion of the elastic Bragg breakwater structure. : (14); in, Linear layer; This is a concatenation function; This is the global context information after extension of length; The final prediction result is a multi-moment-time prediction sequence of the motion response of the elastic Bragg breakwater structure, which includes the displacement of the structure in the horizontal and vertical directions.

[0012] Furthermore, in step 3, the hybrid time-frequency domain loss function consists of two weighted parts: one part is the time-domain loss. The other part is the frequency domain loss. ; The formula for calculating the temporal loss is: (15); in, To predict the time step; This represents the predicted value at time step i. Represents the actual value of time step i; The formula for calculating frequency domain loss is: (16); (17); in, The predicted value is obtained after performing a Fast Fourier Transform. The true value obtained after performing a Fast Fourier Transform; For Fast Fourier Transform; The number of main frequency components retained after frequency domain decomposition; Let be the a-th frequency component of the predicted value at time step i; The a-th frequency component of the true value at time step i; Hybrid time-frequency domain loss The calculation formula is: (18); in, This is a hyperparameter.

[0013] The beneficial technical effects brought about by this invention are as follows.

[0014] This invention transforms the model from a pure "data fitting" to "physics-guided reasoning" by explicitly designing physical prior knowledge as part of the network architecture, thus establishing a high-precision prediction from external wave excitation to the dynamic response inside the structure.

[0015] This invention provides a dual-flow decoupled modeling paradigm. By decomposing the complex wave-structure interaction problem into two sub-problems—internal dynamics and external coupling—and modeling them using a dedicated attention flow, the learning difficulty is reduced. Simultaneously, the global context fusion module enables effective information exchange, improving the model's expressive power and prediction accuracy.

[0016] This invention provides a hybrid time-frequency domain optimization strategy. Based on the physical fact that structural responses possess both time-domain and frequency-domain characteristics, a hybrid loss function is used to impose dual constraints on the model. This not only improves the numerical accuracy of the predictions but also ensures the physical authenticity of the prediction results in the frequency domain, thereby significantly enhancing the model's generalization ability and robustness in unseen dynamic environments. Attached Figure Description

[0017] Figure 1 This is a flowchart of the elastic Bragg breakwater structure response prediction method of the present invention.

[0018] Figure 2 This is a schematic diagram of the elastic Bragg breakwater of the present invention.

[0019] Figure 3 This is a comparison chart of the predicted time series curves and the actual value curves of the response of the elastic Bragg breakwater structure using the LSTM model.

[0020] Figure 4 This is a comparison chart of the predicted time series curves and the actual value curves of the elastic Bragg breakwater structure response from the TCN model.

[0021] Figure 5 This is a comparison chart of the predicted time series curves and the actual value curves of the response of the elastic Bragg breakwater structure using the LSTM-CNN model.

[0022] Figure 6 This is a comparison chart of the predicted time series curves and the actual value curves of the elastic Bragg breakwater structure response from the Transformer model.

[0023] Figure 7 This is a comparison chart of the predicted time series curves and the actual value curves of the elastic Bragg breakwater structure response from the Informer model.

[0024] Figure 8 This is a comparison chart of the predicted time series curves and the actual value curves of the elastic Bragg breakwater structure response from the Autoformer model.

[0025] Figure 9 This is a comparison chart of the predicted time series curves and the actual value curves of the elastic Bragg breakwater structure response from the FEDformer model.

[0026] Figure 10 This is a comparison chart of the predicted time series curves and the actual value curves of the elastic Bragg breakwater structure response from the iTransformer model.

[0027] Figure 11 This is a comparison chart of the predicted time series curves and the actual value curves of the elastic Bragg breakwater structure response of the PhysAttnNet model of this invention.

[0028] Figure 12 This is a comparison of the peak-to-trough error distribution and scatter plot of the LSTM model.

[0029] Figure 13 This is a comparison of the peak-to-trough error distribution and scatter plot of the TCN model.

[0030] Figure 14 This is a comparison of the peak-valley error distribution and scatter plot of the LSTM-CNN model.

[0031] Figure 15 This is a comparison of the peak-to-trough error distribution and scatter plot of the Transformer model.

[0032] Figure 16 This is a comparison of the peak-to-trough error distribution and scatter plot of the Informer model.

[0033] Figure 17 This is a comparison of the peak-to-trough error distribution and scatter plot of the Autoformer model.

[0034] Figure 18 This is a comparison of the peak-to-trough error distribution and scatter plot of the FEDformer model.

[0035] Figure 19 This is a comparison of the peak-to-trough error distribution and scatter plot of the iTransformer model.

[0036] Figure 20 This is a comparison of the peak-valley error distribution and scatter plot of the PhysAttnNet model of this invention.

[0037] Figure 21 This is a comparison chart of RMSE of different models under different sea state generalization conditions.

[0038] Figure 22 This is a comparison chart of SMAEs under different sea state generalizations of different models. Detailed Implementation

[0039] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: This invention provides a method for predicting the structural response of an elastic Bragg breakwater. The input features used in this invention include the free surface elevation before and after the elastic Bragg breakwater and the breakwater's motion response data, where the horizontal (X-direction) and vertical (Z-direction) displacements of the structure at future time points are the prediction targets. Due to the complex physical correlation between the structural motion response and its historical state and external excitations, this invention explicitly models the time decay characteristics of the internal dynamics and the bidirectional coupling and resonance characteristics of the external wave-structure interaction. Therefore, a physics-guided dual-stream attention network architecture is adopted to decouple the complex physical processes and encode them separately. The encoder encodes the internal dynamic evolution and external coupling feedback in parallel, extracting dynamic representations containing physical priors; the global context fusion module deeply interacts the two types of physical information; finally, a linear layer (decoding module) outputs a high-precision prediction of the elastic Bragg breakwater's structural motion response at multiple future time points.

[0040] This invention provides a modeling method based on physical priors and an efficient optimization strategy based on a hybrid time-frequency domain loss function. By embedding physical mechanisms such as time decay, bidirectional coupling, and resonance as inductive biases into the attention module, the model is provided with strong physical constraints, effectively improving its generalization ability and prediction robustness under unseen conditions. This solves the problems of unclear physical meaning and poor generalization ability in traditional data-driven models.

[0041] like Figure 1 As shown, this embodiment of the invention provides a method for predicting the motion response of elastic structures based on a physically guided dual-stream attention network. The method mainly includes the following steps: Step 1: By arranging multiple wave path detection devices (WG1 to WG12) in the experimental water tank, the free liquid surface elevation change data of the area before and after the elastic Bragg breakwater, as well as the motion response data of the breakwater body, are collected synchronously and used as the feature sample dataset for subsequent model input. The embodiments of the present invention first calibrate and verify the SPH numerical model through physical water tank experiments to ensure that it can reproduce the interaction process between waves and elastic structures with high fidelity.

[0042] To construct sufficient training samples, this embodiment utilizes the validated SPH numerical model to conduct large-scale simulations under various irregular wave conditions, generating a large amount of wave elevation time-series data and structural displacement time-series data to construct training, validation, and test datasets. Multiple wave sensors are incorporated into the numerical model to collect wave elevation time-series data in front of and behind the structure.

[0043] The construction of feature samples includes: using the collected wave elevation time series data as external wave excitation sequence data. (External input) Using structural displacement time series data as structural motion response sequence data. (Endogenous input); among which, For time steps, The feature dimension is defined by the sliding time window method, which divides continuous time series data into multiple input-output sample pairs for subsequent model training and evaluation.

[0044] Step 2: Construct a two-stream attention network prediction model based on physical prior guidance. The core of this model is to decouple the complex wave-structure interaction into two parallel attention streams for modeling: a first attention stream and a second attention stream. The first attention stream is used to simulate and capture the dynamic evolution within the structure, while the second attention stream is used to simulate and capture the external wave-structure coupling feedback. This invention utilizes the aforementioned two-stream attention mechanism to encode feature samples and extract dynamic response representations containing physical priors. Finally, through a global context fusion module and a linear layer, a predicted sequence of structural motion responses at multiple future time points is output based on the dynamic response representations. The specific process is as follows: For internal dynamic evolution, the first attention stream in this embodiment of the invention employs a damped bidirectional self-attention module. This module introduces a learnable time decay bias to correct the attention weights in standard self-attention calculations. This bias is parameterized by a negative logarithmic function proportional to the time step difference. This bias aims to amplify the influence of recent historical states on the current prediction, thereby simulating the damping effect in structural dynamics, where the current state is more influenced by recent rather than distant historical states. The input to the damped bidirectional self-attention module is structural motion response sequence data, and the specific working process is as follows: First, the structural motion response sequence data... Projection as query matrix Key matrix Sum matrix : (1); in, , , is a learnable projection weight matrix.

[0045] For any two time steps and Its standard scaled dot product attention score for: (2); in, It is the softmax function; and These are query matrices Bond matrix The row and number The row number corresponds to the time step number; It is the transpose symbol; Key matrix The lth line; This represents the total number of rows.

[0046] The attenuation coefficient is calculated based on the time interval between the two time steps. : (3); in, and These are learnable parameters. = 0.05 represents the sampling interval.

[0047] The decay coefficient is combined with the attention score to generate a time-decay-weighted attention weight. : (4); Finally, the output of the damped bidirectional self-attention module is obtained. : (5); in, Value matrix The j-th line.

[0048] To address external coupling feedback, the second attention stream in this embodiment employs a resonance-guided bidirectional cross-attention module. This module modulates the standard attention weights by introducing a phase bias term controlled by a learnable parameter. This bias aims to capture the phase correlation between wave excitation and structural response, thereby identifying resonance phenomena. Simultaneously, it employs a two-stage computational paradigm: the first stage calculates the wave's effect on the structure, and the second stage calculates the structure's feedback to the wave, thus explicitly modeling the bidirectional nature of wave-structure interaction. The input to the resonance-guided bidirectional cross-attention module is wave excitation sequence data. and structural motion response sequence data This module establishes a two-way coupling relationship between external wave excitation and structural motion response data; the specific working process is as follows: A phase bias term is generated based on the phase relationship between waves and the structure. This phase bias term is derived from the phase function. Calculation. For the query matrix. row i Bond matrix row j Phase bias term for: (6); in, These are learnable parameters.

[0049] The modulated attention weights are obtained by modulating the standard attention weights using the phase bias term. : (7); in, For query matrix row i Bond matrix Line 1 Phase offset term; Finally, the value matrix Perform a weighted summation: (8); in, For the attention function with added phase bias term; Value matrix The j-th line.

[0050] For the two-stage computational paradigm, the first stage involves processing the structural motion response sequence data. Projection as query matrix Wave excitation sequence data Projection as a key matrix Sum matrix : (9); in, , , This is the learnable weight matrix for the first stage. The output of the first stage. for: (10); The second phase will use wave excitation sequence data. Projection as query matrix The output of the first stage Projection as a key matrix Sum matrix : (11); in, , , This is the learnable weight matrix for the second stage. The output of the second stage. for: (12); Second stage output This is the final output of the resonance-guided bidirectional cross-attention module; Finally, the outputs of the two attention streams are integrated through a global context fusion module. This module extracts contextual information representing the global state from the output of the resonance-guided bidirectional cross-attention module to summarize key information about wave-structure interactions. The specific process is as follows: First, extract the global context information. : (13); in, It is the softmax function; This is a learnable weight matrix.

[0051] Then global context information Extended to length L This information is dynamically integrated into the output of the damped bidirectional self-attention module, enabling deep interaction between the two types of physical information. The final linear layer outputs the predicted motion of the elastic Bragg breakwater structure. : (14); in, Linear layer; This is a concatenation function; This is the global context information after extension of length; The final prediction result is a multi-moment-time prediction sequence of the motion response of the elastic Bragg breakwater structure, which includes the displacement of the structure in the horizontal (X direction) and vertical (Z direction).

[0052] Step 3: Design a hybrid time-frequency domain loss function to efficiently optimize the prediction model of the two-stream attention network; Considering that the structural motion response needs to be accurate not only in the time domain but also in the frequency domain, this embodiment employs a hybrid time-frequency domain loss function for model optimization. This loss function consists of two weighted parts: one part is the time-domain loss. Mean squared absolute error is used to ensure the numerical accuracy of the predicted and actual values ​​at each time point: (15); in, To predict the time step; This represents the predicted value at time step i. This represents the actual value of time step i.

[0053] The other part is frequency domain loss. By analyzing the predicted sequence and the real sequence Perform a Fast Fourier Transform: (16); Then, the mean square absolute error between the two spectra is calculated to ensure that the spectral characteristics of the predicted response are consistent with the actual situation: (17); in, The predicted value is obtained after performing a Fast Fourier Transform. The true value obtained after performing a Fast Fourier Transform; For Fast Fourier Transform; The number of main frequency components retained after frequency domain decomposition; Let be the a-th frequency component of the predicted value at time step i; The a-th frequency component of the true value at time step i; The time-domain loss ensures the point-by-point accuracy of the predicted sequence in terms of numerical values; the frequency-domain loss penalizes the difference in the spectrum between the predicted and true sequences by performing a Fourier transform on them, thus ensuring the physical authenticity of the prediction result; finally, the time-domain loss and the frequency-domain loss are weighted and summed to obtain the hybrid time-frequency domain loss. To collaboratively optimize the time-domain accuracy and frequency-domain fidelity of the model: (18); in, This is a hyperparameter.

[0054] To verify the effectiveness and robustness of the method proposed in this invention, the following model performance verification was performed.

[0055] The dual-stream attention network prediction model was trained using a training set and then validated using a test set. Finally, the trained model was validated in terms of both zero-shot generalization ability and material property adaptability.

[0056] Test set verification: This validation aims to examine the model's predictive ability on the test set. In practice, the trained model is applied to the test set under the same sea conditions as the training set. The results are shown in Table 1. The comparison models listed in this table are all mainstream time series prediction neural network architectures, including: LSTM, which models long-term temporal dependencies through a gating mechanism; TCN, which uses dilated causal convolution to capture local and global temporal patterns in parallel; LSTM-CNN, a hybrid structure combining convolutional layers to extract local features and LSTM to model temporal dynamics; Transformer, which achieves long-distance dependency modeling based on a self-attention mechanism; Informer, which introduces sparse attention and distillation mechanisms to improve the prediction efficiency of long sequences; Autoformer, which integrates sequence decomposition and self-attention to enhance the ability to model temporal structures; FEDformer, which guides attention through frequency domain transformation to strengthen the capture of periodic components; iTransformer, which treats multivariate channel dimensions as sequence inputs to model cross-variable dependencies; and the proposed dual-stream attention network prediction model PhysAttnNet, whose core lies in modeling structural dynamics evolution and wave-structure coupling feedback through damped bidirectional self-attention and resonant guided cross-attention, respectively, and achieving deep interaction of physical information through global fusion. All models were compared fairly under the same dataset and evaluation metrics to objectively verify the technical effects of this invention. As shown in Table 1, the model of this invention, in the task of predicting the response of an elastic Bragg breakwater structure, has a mean absolute error (MAE) of 0.1862, a root mean square error (RMSE) of 0.3218, a coefficient of determination (R²) of 0.8446, and a frequency domain mean absolute error (SMAE) of 0.9250, significantly outperforming the eight neural network models compared. A direct comparison with the best-performing comparative model on the test set—Informer (MAE = 0.2240)—shows a relative reduction in prediction error of 16.9% for the model of this invention.

[0057] Table 1. Performance comparison of the method of the present invention on the test set. .

[0058] Figures 3-11 A comparison of the predicted time-series curves and actual value curves of the response of elastic Bragg breakwater structures under different models. Figures 12-20 This section provides a comparison of the peak-to-trough error distributions and scatter plots for different models. From... Figures 3-20 It is evident that the model of this invention, in predicting the vertical displacement of the elastic Bragg breakwater structure, exhibits a high degree of agreement between its prediction time series curve and the measured value, with the most concentrated peak-valley error distribution and the highest correlation in the scatter plot. This is significantly superior to baseline models such as Long Short-Term Memory Network and Transformer, fully demonstrating the superiority of the method of this invention in prediction accuracy.

[0059] Zero-shot generalization capability validation: This validation aims to examine the model's predictive ability under entirely new sea states without prior training. In practice, the trained model is directly applied to multiple representations of different sea states (defined by different significant wave heights H). s Spectral peak period T p On the test dataset (defined by combination), this process does not involve any retraining or fine-tuning of model parameters. The results are as follows: Figure 21 and Figure 22 As shown, the model of this invention achieves the lowest root mean square error and mean absolute error values ​​under all sea state conditions, with a generalization ability enhanced by more than 13%. The results indicate that the method of this invention, by incorporating physical priors, successfully learns universal physical laws rather than statistical patterns specific to the training data, thus possessing strong generalization ability in unknown dynamic environments.

[0060] Material property suitability verification: This validation aims to examine the model's adaptability to changes in the structure's physical properties. Specifically, by altering the key material parameter of the structure—the elastic modulus (E)—new datasets with varying stiffness characteristics were generated. The trained model was then fine-tuned on these new datasets. The performance comparison results are shown in Table 2. The model of this invention significantly outperforms the compared neural network models on different structural materials. The results demonstrate that even when the physical properties of the structure change, the model proposed in this invention still significantly outperforms other baseline models.

[0061] Table 2 Comparison of the adaptability of the method of the present invention to different structural materials .

[0062] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for predicting the response of an elastic Bragg breakwater structure, characterized by, The method comprises the following steps: Step 1, arranging a plurality of wave path detection devices in an experimental water tank, carrying out large-scale simulation under a plurality of irregular wave conditions, and generating a large amount of wave height time series data and structure displacement time series data; Step 2, constructing a double-flow attention network prediction model based on physical prior guidance; the model comprises two parallel attention flows, a global context fusion module and a linear layer; the two parallel attention flows are a first attention flow and a second attention flow respectively; Step 3, designing a hybrid time-frequency domain loss function to optimize the double-flow attention network prediction model.

2. The method of claim 1, wherein the method further comprises: In the step 1, the collected wave elevation time series data is taken as the wave excitation sequence data In the step 2, the structural displacement time series data is taken as the structural motion response sequence data ; wherein, is the time step number, is the feature dimension; the continuous time series data is cut into multiple input-output sample pairs by the sliding time window method.

3. The method of claim 1, wherein the method further comprises: In step 2, the first attention flow is a damping bidirectional self-attention module; in the standard self-attention calculation, a learnable time decay bias is introduced to correct the attention weight; the specific working process of the damping bidirectional self-attention module is as follows: First, structural motion response sequence data Projection as query matrix , key matrix and value matrix : (1)； wherein, , , is a learnable projection weight matrix; For any two time steps and whose standard scaled dot-product attention score is: (2)； wherein, is a softmax function; and are the first and the first row of the query matrix and the key matrix respectively, the row index corresponding to the time step index; is a transpose symbol; is the l-th row of the key matrix ; and is the total number of rows. The decay coefficient is calculated according to the time interval of two time steps : (3)； wherein, and are learnable parameters, denotes a sampling interval; combining the decay coefficients with attention scores to generate time-decay weighted attention weights : (4)； Finally, the output of the damping bidirectional self-attention module is obtained : (5)； wherein is the jth row of the value matrix is the jth row of the value matrix 4. The elastic Bragg breakwater structure response prediction method according to claim 3, wherein, In step 2, the second attention flow is a resonance-guided bidirectional cross-attention module; the module modulates the standard attention weight by introducing a phase bias term controlled by a learnable parameter; at the same time, a two-stage calculation paradigm is adopted to model the bidirectionality of wave-structure interaction; The phase bias term is generated according to the phase relationship between the wave and the structure, and the formula is as follows: (6)； wherein, is a query matrix ith row and a key matrix jth row a phase bias term; are learnable parameters; modulate the standard attention weight through the phase bias term, to obtain a modulated attention weight : (7)； wherein is a query matrix ith row and a key matrix ith row a phase bias term; Finally, the value matrix is summed up with weighting: (8)； wherein, is an attention function with a phase bias term added; For the two-stage computational paradigm, the first stage projects the structural motion response sequence data onto the query matrix , the wave excitation sequence data onto the key matrix and the value matrix: (9)； wherein, , , is a first stage learnable weight matrix; the output of the first stage is: (10)； The second stage projects the wave excitation sequence data onto the query matrix The output of the first stage is projected onto the key matrix and the value matrix ​ (11)； wherein, , , is a second stage learnable weight matrix; the output of the second stage is is: (12)； Output of the second stage Final output for the resonance guided bidirectional cross attention module.

5. The method of claim 4, wherein the method further comprises: In step 2, the specific working process of the global context fusion module and the linear layer is as follows: First, global context information is extracted : (13)； wherein, is a softmax function; is a learnable weight matrix; Then global context information Extended to length L This information is dynamically integrated into the output of the damped bidirectional self-attention module, enabling deep interaction between the two types of physical information. The final linear layer outputs the predicted motion of the elastic Bragg breakwater structure. : (14)； wherein, is a linear layer; is a stitching function; is the global context information after the length extension; The final prediction result is a future multi-time elastic breakwater structure motion response prediction sequence, which includes the horizontal and vertical displacements of the structure.

6. The method of claim 1, wherein the method further comprises: In step 3, the mixed time-frequency domain loss function is composed of two parts: one part is the time domain loss , and the other part is the frequency domain loss ; The calculation formula of the time domain loss is as follows: (15)； wherein, is the prediction time step; represents the predicted value of time step i, represents the true value of time step i; The calculation formula of the frequency domain loss is as follows: (16)； (17)； wherein is the predicted value after performing a fast Fourier transform; is the true value after performing a fast Fourier transform; is the fast Fourier transform; is the number of main frequency components retained after the frequency domain decomposition; is the a-th frequency component of the predicted value at time step i; is the a-th frequency component of the true value at time step i; Mixed time-frequency domain loss The calculation formula is: (18)； wherein is a hyperparameter.

Images