Large-scale structure dynamic response prediction method based on hybrid deep learning model

Abstract

The invention relates to a large-scale structure dynamic response prediction method based on a hybrid deep learning model. The method comprises the following steps: establishing a finite element model of a large-scale structure; a vibration exciter is used for measuring modal parameters of the large-scale structure, and the finite element model of the large-scale structure is corrected according to the modal parameters of the large-scale structure; simulating the dynamic response of the large-scale structure under the dynamic load by using the corrected finite element model; meanwhile, a sensor is used for actually collecting the dynamic response of the large-scale structure under the dynamic load; constructing a training set through analog simulation and actually acquired dynamic response of the large-scale structure under the dynamic load; training the hybrid deep learning model according to the constructed training set; and inputting the dynamic load acting on the large-scale structure into the trained deep learning model to predict the dynamic response of the large-scale structure. According to the method, the prediction efficiency and accuracy are improved, the generalization ability of the model is enhanced, and the method can adapt to various different environments and conditions.

Description

Technical Field

[0001] The present invention relates to the field of structural dynamic response prediction, and in particular to a large-scale structural dynamic response prediction method based on a hybrid deep learning model. Background Art

[0002] With the rapid development of modern industry and cities, large structures such as skyscrapers, bridges, transmission towers, dams and industrial facilities have become key pillars of modern social infrastructure and technological progress. These structures not only play a vital role in economic development, supporting the efficient operation of multiple fields such as transportation, energy and communications, but also play an indispensable role in ensuring public safety, providing employment opportunities and promoting scientific and technological innovation. The stability and durability of large structures are directly related to human well-being and environmental sustainability. Therefore, in-depth research on the dynamic response of large structures under dynamic loads is crucial to ensure their safe operation.

[0003] In the field of dynamic response prediction of large structures, traditional prediction methods face many challenges. On the one hand, dynamic response prediction based solely on finite element models is often difficult to accurately reflect the behavior of large structures under real dynamic loads due to the complexity of the actual structure and the approximate processing in the modeling process. The uncertainty of factors such as the material properties and boundary conditions of the actual structure leads to deviations between the prediction results of the finite element model and the actual situation.

[0004] On the other hand, although relying solely on data collected by sensors for prediction can reflect the actual response of the structure, it lacks a deep understanding of the intrinsic physical mechanism of the structure, and the data is often affected by factors such as noise and sensor accuracy.

[0005] Existing deep learning models also have limitations when dealing with such problems. A single deep learning model, such as a convolutional neural network (CNN), may not be able to fully capture the feature information at different scales when processing large-scale structural dynamic response data with complex spatiotemporal characteristics; although the long short-term memory network (LSTM) has good processing capabilities for time series data, it is insufficient in the richness of feature extraction. Therefore, a method that can comprehensively consider the advantages of structural physical models and deep learning is needed to improve the accuracy and reliability of large-scale structural dynamic response prediction. Summary of the invention

[0006] In order to solve the problems existing in the background technology, the present invention provides a large-scale structure dynamic response prediction method based on a hybrid deep learning model, comprising:

[0007] S1: Establish finite element models of large structures;

[0008] S2: Use a shaker to measure the modal parameters of a large structure, and modify the finite element model of the large structure based on the modal parameters of the large structure;

[0009] S3: Use the modified finite element model to simulate the dynamic response of the large structure under dynamic loads; at the same time, use the dynamic response of the large structure under dynamic loads actually collected by sensors;

[0010] S4: Construct a training set by simulating and actually collecting the dynamic responses of large structures under dynamic loads;

[0011] S5: training a hybrid deep learning model according to the constructed training set, wherein the hybrid deep learning model includes: a multi-scale convolutional neural network MS-CNN, a multi-head attention mechanism Multi-Head Attention and a long short-term memory network LSTM; the multi-scale convolutional neural network MS-CNN is used to extract multi-scale features of input training samples, and the multi-head attention mechanism Multi-Head Attention is used to perform attention enhancement processing on the extracted multi-scale features; the long short-term memory network LSTM is used to predict the dynamic response of a large structure according to the features after attention enhancement;

[0012] S6: The dynamic loads acting on the large structure are input into the trained deep learning model to predict the dynamic response of the large structure.

[0013] The present invention has at least the following beneficial effects

[0014] The present invention uses an exciter to measure the modal parameters of a large structure, and accordingly modifies the finite element model so that the finite element model can more accurately reflect the actual characteristics of the large structure, providing a more reliable basis for subsequent simulations and effectively reducing prediction deviations caused by model errors. The modified finite element model is used to simulate the dynamic response of a large structure under dynamic loads, and sensors are used to actually collect response data at the same time; and a training set is constructed based on these two types of data. This method of combining simulation and measured data fully utilizes the advantages of data from different sources, makes up for the shortcomings of a single data source, enriches the information of the training data, and helps to improve the generalization ability and prediction accuracy of the model. The present invention adopts a hybrid deep learning model including multi-scale convolutional neural network (MS-CNN), multi-head attention mechanism (Multi-Head Attention) and long short-term memory network (LSTM). MS-CNN can extract multi-scale features of input training samples, capture the structural dynamic response characteristics at different scales, and enrich the feature information; the multi-head attention mechanism performs attention enhancement processing on the extracted multi-scale features, so that the model can pay more attention to key features and improve the expression ability of features; LSTM predicts the dynamic response of large structures based on the features after attention enhancement, gives full play to its processing ability for time series data, and effectively integrates different types of feature information, thereby improving the accuracy and reliability of the prediction of the dynamic response of large structures. It provides a convenient and efficient prediction method for practical engineering applications, helps engineers to timely understand the response of structures under different load conditions, and provides strong support for structural design, maintenance and safety assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0016] Figure 2 A schematic diagram of a power transmission tower dynamic model in an embodiment of the present invention;

[0017] Figure 3 A wind speed time history diagram and a wind power spectrum comparison diagram in an embodiment of the present invention;

[0018] Figure 4 A network structure diagram of a deep learning hybrid model in an embodiment of the present invention;

[0019] Figure 5 It is a prediction diagram of the dynamic response (displacement response) of a transmission tower in an embodiment of the present invention;

[0020] Figure 6 It is a prediction diagram of the dynamic response (stress response) of a transmission tower in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0022] See also Figure 1 The present invention provides a large-scale structure dynamic response prediction method based on a hybrid deep learning model, comprising:

[0023] S1: Establish finite element models of large structures;

[0024] Preferably, the establishment of the finite element model of the large structure includes: constructing a three-dimensional model of each component of the large structure in SolidWorks according to the design drawings of the large structure, and after completing the SolidWorks modeling, importing the three-dimensional model of the large structure into Abaqus, and establishing the finite element model of the large structure by dividing the grid, defining material properties, and setting contact and constraint conditions.

[0025] See also Figure 2 Taking the transmission tower as an example, in this embodiment, a proportional physical model of the transmission tower is established according to the detailed design drawings of the transmission tower. Then, the numerical models of the various components of the transmission tower (tower body, tower arm, connectors, etc.) are created one by one using Solid Works software according to the design drawings. After the modeling of all components is completed, they are imported into Abaqus, a professional finite element analysis software, and the finite element model of the structure is constructed through steps such as meshing, defining material properties, setting contact and constraint conditions, etc.

[0026] S2: Use a shaker to measure the modal parameters of a large structure, and modify the finite element model of the large structure based on the modal parameters of the large structure;

[0027] Preferably, step S2 comprises:

[0028] S21: applying a dynamic load to a large structure using a vibration exciter; the dynamic load includes: an excitation force of known frequency and amplitude;

[0029] Install the vibration exciter at the transmission tower site. The vibration exciter is a device that can generate an excitation force of a specific frequency and amplitude. According to the pre-set experimental plan, select a suitable vibration exciter and fix it at a specific position of the transmission tower, such as a standard section of the tower body.

[0030] In order to fully obtain the modal parameters of the transmission tower, the exciter is set to generate dynamic loads of different frequencies and amplitudes. For example, the exciter is first set to generate an excitation force with a frequency of 5 Hz and an amplitude of 100 N to apply a dynamic load to the transmission tower. This excitation force will cause the transmission tower to vibrate.

[0031] S22: using a sensor to measure a dynamic response of the large structure, and measuring modal parameters of the large structure through the sensor, wherein the modal parameters of the large structure include: a vibration form of the large structure under different loads, a natural frequency of the vibration, and an attenuation rate of the vibration;

[0032] Multiple sensors, such as acceleration sensors, displacement sensors, etc., are installed at key locations on the transmission tower.

[0033] When the exciter applies a dynamic load, the transmission tower begins to vibrate, and the sensors measure the dynamic response of the tower in real time. For example, the acceleration sensor measures the acceleration change of the tower during vibration, and the displacement sensor measures the displacement change of various parts of the tower.

[0034] By analyzing and processing the data collected by the sensor, the modal parameters of the transmission tower can be obtained. Specifically, by analyzing the characteristics of the vibration signal, the vibration mode of the tower under different loads can be identified, such as whether it is overall vibration or local vibration, the direction and amplitude of vibration, etc. The natural frequency of vibration, that is, the vibration frequency of the tower itself during free vibration, can be determined by spectrum analysis and other methods; by observing the attenuation of the vibration signal over time, the attenuation rate of the vibration can be calculated.

[0035] S23: By adjusting the mesh of the finite element model, defining material properties, and setting contact and constraint condition parameters, the finite element model is corrected so that the modal parameters of the large structure measured and identified by the sensor are consistent with the theoretical modal parameters calculated by the finite element model within a preset error range.

[0036] In this embodiment, in the finite element model analysis software Abaqus, the theoretical modal parameters calculated by the previously established transmission tower finite element model are checked. The modal parameters identified by the sensor measurement are compared with the theoretical modal parameters calculated by the finite element model. If a large difference is found between the two, the number of grids is appropriately increased or decreased, and the shape and size distribution of the grid are changed. Check whether the material parameters previously input are accurate. If the elastic modulus of the steel material actually used may have a certain deviation, readjust the material property parameters such as the elastic modulus, calculate the modal parameters of the finite element model again, and observe the degree of closeness to the actual measured values. Check whether the contact setting between the tower sections is reasonable and whether the contact stiffness of the bolt connection is consistent with the actual situation. For example, the contact stiffness between the bolt and the connecting plate is appropriately adjusted, or the constraint conditions of the tower bottom are modified to make it more consistent with the actual installation and stress state, and then the modal parameters are recalculated. By continuously adjusting these parameters, the theoretical modal parameters calculated by the finite element model are consistent with the actual modal parameters identified by the sensor measurement within a preset error range.

[0037] S3: Use the modified finite element model to simulate the dynamic response of the large structure under dynamic loads; at the same time, use the dynamic response of the large structure under dynamic loads actually collected by sensors;

[0038] Preferably, the dynamic response data includes: one or more of displacement, stress and strain of measurement points on the large structure.

[0039] S4: Construct a training set by simulating and actually collecting the dynamic responses of large structures under dynamic loads;

[0040] Preferably, the training samples in the training set include: dynamic loads V acting on large structures and corresponding dynamic responses U;

[0041]

[0042] Among them, v ti Indicates that at t i Dynamic loads on large structures at time u ti Indicates that at t i Dynamic response of dangerous areas of large structures at all times.

[0043] Specifically, in this embodiment, the wind-driven dynamic response data of the transmission tower obtained through simulation analysis and experiments is preprocessed to construct the data set required for the deep learning hybrid model. The data set is divided into a training set, a validation set, and a test set in a ratio of 6:3:3, where the ratio of real data to simulation data in the training set is 4:2, and the ratio of real data to simulation data in the validation set and the test set is 2:1.

[0044] See also Figure 4, S5: training a hybrid deep learning model according to the constructed training set, wherein the hybrid deep learning model includes: a multi-scale convolutional neural network MS-CNN, a multi-head attention mechanism Multi-HeadAttention and a long short-term memory network LSTM; the multi-scale convolutional neural network MS-CNN is used to extract multi-scale features of input training samples, and the multi-head attention mechanism Multi-Head Attention is used to perform attention enhancement processing on the extracted multi-scale features; the long short-term memory network LSTM is used to predict the dynamic response of a large structure according to the features after attention enhancement;

[0045] Preferably, the multi-scale convolutional neural network MS-CNN comprises: 4 parallel convolution paths, each convolution path adopts a convolution kernel and a dilation coefficient of a different size to capture the multi-scale information of the input features; the 4 parallel convolution paths are used to perform dynamic loads V∈R acting on the large structure. T×P×D , T is the time dimension, P is the position dimension, and D is the feature dimension. Four scale feature extractions are performed to obtain four scale features. and The output features of each convolutional path are concatenated in the channel dimension to obtain the multi-scale feature F MS ∈R C×H×W , where C = C1 + C2 + C3 + C4, C represents the multi-scale feature F MS The channel dimension, C j Representation feature F j The channel dimension of , j∈{1, 2, 3, 4}.

[0046] In this embodiment, the four parallel convolution paths of the multi-scale convolutional neural network MS-CNN perform multi-scale feature extraction including:

[0047] The first convolution path: Use a 1×1 convolution kernel to extract features from the input and obtain a feature F1 with a dimension of [C1, H, W]. The 1×1 convolution is used for dimensionality reduction and feature channel compression while retaining the detailed information of local features.

[0048] Where, C1 is the number of channels;

[0049]

[0050] K is the size of the convolution kernel, d is the dilation coefficient, P is the size of the edge padding, and S is the step size;

[0051] Second convolution path: Use 3×3 and 1×1 convolution kernels to extract features from the input. The first convolution layer has a dilation factor of 1 to capture medium-scale local features; the second 1×1 convolution is used to further adjust the number of feature channels. Each convolution layer uses edge padding to ensure that the spatial size of the feature map is the same as F1, and finally obtains feature F2 with a dimension of [C2,H,W].

[0052] The third convolutional path: 5×5 and 1×1 convolution kernels are used to extract features from the input. The expansion coefficient of the first convolution layer is 2, which is used to expand the receptive field and capture feature information of a larger scale; the second 1×1 convolution is used to adjust the feature channel. The edge zero padding method is also used to obtain the feature F3 with a dimension of [C3, H, W].

[0053] Fourth convolution path: Use 1×1 and 3×3 convolution kernels to extract features from the input. The first 1×1 convolution is used for dimensionality reduction, and the second 3×3 convolution layer has a dilation coefficient of 1, which is used to extract local features. By padding the edges with zeros, the feature F4 with a dimension of [C4, H, W] is finally obtained. The data input uses a multi-scale convolution method, that is, through convolution kernels of different sizes and dilated convolutions, the result matrices processed by different convolution layers are spliced ​​together in the depth dimension to form a deeper matrix. Smaller convolution kernels focus on the local details of the input data and extract fine feature information; while larger convolution kernels use dilated convolutions to capture wider regional features. Information aggregation at different sizes is achieved, and features are extracted from different scales, making the model more flexible and robust when dealing with complex scenarios.

[0054] In order to further optimize the computational efficiency and improve the model performance, the multi-head attention mechanism is used to accurately screen out the key features closely related to the output. This process effectively reduces the data dimension and reduces redundant information. Subsequently, the generated multi-dimensional tensor is flattened into a one-dimensional tensor and input into the long short-term memory network (LSTM). The loop structure of LSTM can effectively capture the temporal dependencies in the data, thereby effectively integrating the input features and predicting or outputting the corresponding responses, and finally obtaining the dynamic response results of the transmission tower.

[0055] Preferably, the multi-head attention mechanism Multi-Head Attention includes:

[0056] Feature Mapping: Multi-scale features F of the input MS ∈R C×H×W Apply three parallel 1x1 convolution operations to generate features F5∈R C×H×W , F6∈R C×H×W and F7∈R C×H×W, features F5, F6 and F7 are used as query vector Q, key vector K and value vector V in the multi-head self-attention mechanism respectively;

[0057] Multi-head self-attention calculation: The query vector Q, key vector K and value vector V are evenly divided into M groups in the channel dimension, with the number of channels in each group being C / M, to obtain the grouped query vector Q m , key vector K m Sum value vector V m , where m = {1, 2, …, m, …, M};

[0058] For each set of query vectors Q m , key vector K m Sum value vector V m Calculate the attention-enhanced feature Z through the self-attention mechanism m :

[0059]

[0060] Among them, d k To represent the temperature coefficient, softmax represents the activation function;

[0061] Multi-head feature concatenation: The attention-enhanced features of the M groups of self-attention outputs are concatenated in the channel dimension to obtain the feature F after the multi-head attention mechanism Multi-Head Attention is enhanced. att ∈R C×H×W =concat(Z1,Z2,...,Z m ,...,Z M ), concat represents the concatenation function.

[0062] Preferably, the long short-term memory network LSTM comprises:

[0063] Input feature processing: Input feature F att Flatten from shape [C, H, W] to [C, H×W], convert the spatial dimension into sequence data through the flattening operation, forming a feature sequence with a time step of H×W and a feature dimension of C;

[0064] Time series modeling and feature integration: Use two layers of LSTM units to perform time series modeling and feature integration on the flattened feature sequence. The input dimension of the LSTM unit is C, and the hidden layer size is hidden-size. The LSTM unit captures the long-term dependency between features and outputs time series features with a shape of [hidden-size, H×W].

[0065] Feature mapping and prediction: The output of the last time step of LSTM is mapped to the target dimension [S] through a fully connected layer, where S is the dimension of the dynamic response data to be predicted, thereby realizing the dynamic response prediction of large structures.

[0066] S6: The dynamic loads acting on the large structure are input into the trained deep learning model to predict the dynamic response of the large structure.

[0067] Preferably, the loss function used when training the hybrid deep learning model is:

[0068]

[0069] Among them, L MSE represents the loss function, m is the number of samples, u predict is the predicted data, u target For real data.

[0070] Figure 2 The elevation of a wine glass-shaped transmission tower with a total height of 42.8m is given, and the tower is used as an example for analysis. Based on the Davenport wind speed power spectrum, the harmonic superposition method is used to simulate the wind speed. For the proportional model of the transmission tower, the wind tunnel test is used to load the wind speed. At the same time, a dynamic response sensor is installed on the transmission tower in proportion, and the sampling frequency is set to 1Hz to collect the displacement and stress data of the transmission tower under wind load; for the finite element model, the wind speed is converted into wind pressure according to the transmission tower design manual and input into the finite element model for dynamic response analysis. Figure 3 The wind speed time history curve is given and compared with the Davenport pulsating wind speed power spectrum to verify that the simulated wind speed time history curve meets the requirements. The trained weights are loaded into the deep learning hybrid model. According to the dynamic load data measured by the sensor, they are pre-processed and put into the trained deep learning model to predict the dynamic response of the transmission tower. The comparison charts of the time history curves of the maximum displacement and maximum stress positions of the transmission tower and the predicted values ​​are given respectively, as shown in Figure 5 and Figure 6 As shown in the figure, the accuracy of displacement and stress response calculated is 98.87% and 97.65% respectively. At the same time, the prediction accuracy, noise robustness and computational efficiency of the model dynamic response are evaluated.

[0071] In summary, the present invention uses an exciter to measure the modal parameters of a large structure, and accordingly modifies the finite element model so that the finite element model can more accurately reflect the actual characteristics of the large structure, providing a more reliable basis for subsequent simulations, and effectively reducing prediction deviations caused by model errors. The modified finite element model is used to simulate the dynamic response of a large structure under dynamic loads, and sensors are used to actually collect response data at the same time; and a training set is constructed based on these two types of data. This method of combining simulation and measured data fully utilizes the advantages of data from different sources, makes up for the shortcomings of a single data source, enriches the information of the training data, and helps to improve the generalization ability and prediction accuracy of the model. The present invention adopts a hybrid deep learning model including multi-scale convolutional neural network (MS-CNN), multi-head attention mechanism (Multi-HeadAttention) and long short-term memory network (LSTM). MS-CNN can extract multi-scale features from input training samples, capture structural dynamic response features at different scales, and enrich feature information; the multi-head attention mechanism performs attention enhancement processing on the extracted multi-scale features, so that the model can pay more attention to key features and improve the expression ability of features; LSTM predicts the dynamic response of large structures based on the features after attention enhancement, giving full play to its processing ability for time series data, and effectively integrating different types of feature information, thereby improving the accuracy and reliability of the prediction of dynamic response of large structures. It provides a convenient and efficient prediction method for practical engineering applications, helps engineers to timely understand the response of structures under different load conditions, and provides strong support for structural design, maintenance and safety assessment.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A large-scale structural dynamic response prediction method based on a hybrid deep learning model, characterized in that: include: S1: Establish finite element models of large structures; S2: Use a shaker to measure the modal parameters of a large structure, and modify the finite element model of the large structure based on the modal parameters of the large structure; S3: Use the modified finite element model to simulate the dynamic response of the large structure under dynamic loads; at the same time, use the dynamic response of the large structure under dynamic loads actually collected by sensors; S4: Construct a training set by simulating and actually collecting the dynamic responses of large structures under dynamic loads; S5: training a hybrid deep learning model according to the constructed training set, wherein the hybrid deep learning model includes: a multi-scale convolutional neural network MS-CNN, a multi-head attention mechanism Multi-Head Attention and a long short-term memory network LSTM; the multi-scale convolutional neural network MS-CNN is used to extract multi-scale features of input training samples, and the multi-head attention mechanism Multi-Head Attention is used to perform attention enhancement processing on the extracted multi-scale features; the long short-term memory network LSTM is used to predict the dynamic response of a large structure according to the features after attention enhancement; S6: The dynamic loads acting on the large structure are input into the trained deep learning model to predict the dynamic response of the large structure.

2. According to claim 1, a large-scale structural dynamic response prediction method based on a hybrid deep learning model is characterized in that: The method of establishing a finite element model of a large structure includes: constructing a three-dimensional model of each component of the large structure in SolidWorks according to the design drawings of the large structure; after completing the SolidWorks modeling, importing the three-dimensional model of the large structure into Abaqus; and establishing a finite element model of the large structure by dividing the grid, defining material properties, and setting contact and constraint conditions.

3. The large-scale structure dynamic response prediction method based on hybrid deep learning model according to claim 1 is characterized in that: The step S2 comprises: S21: applying a dynamic load to a large structure using a vibration exciter; the dynamic load includes: an excitation force of known frequency and amplitude; S22: using a sensor to measure a dynamic response of the large structure, and measuring modal parameters of the large structure through the sensor, wherein the modal parameters of the large structure include: a vibration form of the large structure under different loads, a natural frequency of the vibration, and an attenuation rate of the vibration; S23: By adjusting the mesh of the finite element model, defining material properties, and setting contact and constraint condition parameters, the finite element model is corrected so that the modal parameters of the large structure measured and identified by the sensor are consistent with the theoretical modal parameters calculated by the finite element model within a preset error range.

4. The large-scale structural dynamic response prediction method based on a hybrid deep learning model according to claim 3 is characterized in that: The dynamic response data includes one or more of displacement, stress and strain of a measurement point on the large structure.

5. The large-scale structural dynamic response prediction method based on a hybrid deep learning model according to claim 4 is characterized in that: The training samples in the training set include: dynamic loads V acting on large structures and corresponding dynamic responses U; Among them, v ti Indicates that at t i Dynamic loads on large structures at time u ti Indicates that at t i Dynamic response of dangerous areas of large structures at all times.

6. The large-scale structural dynamic response prediction method based on a hybrid deep learning model according to claim 1 is characterized in that: The multi-scale convolutional neural network MS-CNN includes: 4 parallel convolution paths, each of which uses convolution kernels and dilation coefficients of different sizes to capture multi-scale information of input features; the 4 parallel convolution paths are used to simulate the dynamic load V∈R acting on the large structure. T×P×D , T is the time dimension, P is the position dimension, and D is the feature dimension. Four scale feature extractions are performed to obtain four scale features. and The output features of each convolutional path are concatenated in the channel dimension to obtain the multi-scale feature F MS ∈R C×H×W , where C = C1 + C2 + C3 + C4, C represents the multi-scale feature F MS The channel dimension, C j Representation feature F j The channel dimension of , j∈{1, 2, 3, 4}.

7. The large-scale structure dynamic response prediction method based on hybrid deep learning model according to claim 6 is characterized in that: The multi-head attention mechanism Multi-Head Attention includes: Feature Mapping: Multi-scale features F of the input MS ∈R C×H×W Apply three parallel 1x1 convolution operations to generate features F5∈R C×H×W , F6∈R C×H×W and F7∈R C×H×W , features F5, F6 and F7 are used as query vector Q, key vector K and value vector V in the multi-head self-attention mechanism respectively; Multi-head self-attention calculation: The query vector Q, key vector K and value vector V are evenly divided into M groups in the channel dimension, with the number of channels in each group being C / M, to obtain the grouped query vector Q m , key vector K m Sum value vector V m , where m = {1, 2, …, m, …, M}; For each set of query vectors Q m , key vector K m Sum value vector V m Calculate the attention-enhanced feature Z through the self-attention mechanism m : Among them, d k To represent the temperature coefficient, softmax represents the activation function; Multi-head feature concatenation: The attention-enhanced features of the M groups of self-attention outputs are concatenated in the channel dimension to obtain the feature F after the multi-head attention mechanism Multi-Head Attention is enhanced. att ∈R C×H×W =concat(Z1,Z2,...,Z m ,...,Z M ), concat represents the concatenation function.

8. The large-scale structure dynamic response prediction method based on hybrid deep learning model according to claim 7 is characterized in that: The long short-term memory network LSTM include: Input feature processing: Input feature F att Flatten from shape [C, H, W] to [C, H×W], convert the spatial dimension into sequence data through the flattening operation, forming a feature sequence with a time step of H×W and a feature dimension of C; Time series modeling and feature integration: Use two layers of LSTM units to perform time series modeling and feature integration on the flattened feature sequence. The input dimension of the LSTM unit is C, and the hidden layer size is hidden-size. The LSTM unit captures the long-term dependency between features and outputs time series features with a shape of [hidden-size, H×W]. Feature mapping and prediction: The output of the last time step of LSTM is mapped to the target dimension [S] through a fully connected layer, where S is the dimension of the dynamic response data to be predicted, thereby realizing the dynamic response prediction of large structures.

9. The large-scale structural dynamic response prediction method based on a hybrid deep learning model according to claim 1 is characterized in that: The loss function used when training the hybrid deep learning model is: Among them, L MSE represents the loss function, m is the number of samples, u predict is the predicted data, u target For real data.

Images