Impact force time history response prediction method and system based on AEGCN-Dynamiformer, storage medium and computer equipment

Abstract

The invention discloses an impact force time history response prediction method and system based on AEGCN-Dynamiformer, a storage medium and computer equipment, and the prediction method comprises the following steps: S1, constructing a database with component parameters, the component parameters comprising geometric parameters, material parameters, boundary load conditions and impact parameters; s2, converting the component parameters in the database into graph structure data and performing normalization processing to form a data set for training; s3, a network neural network is adopted to extract structure topological features, and a feature fusion module is adopted to fuse the extracted structure topological features and the time sequence coding features so as to construct an AEGCN-Dynamiformer deep neural network; s4, carrying out training optimization on the AEGCN-Dynamiformer deep neural network by adopting the data set, and carrying out training optimization on the AEGCN-Dynamiformer deep neural network; and S5, inputting parameters of a to-be-predicted structure into the trained and optimized AEGCN-Dynamiformer deep neural network, so as to generate an impact force time history curve under a corresponding impact working condition. According to the method, the reliability analysis and parameter optimization research efficiency can be remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of impact-resistant design, and in particular to a method, device, storage medium and computer equipment for predicting the time-history response of impact force based on AEGCN-Dynamiformer. Background Art

[0002] Bridge piers and columns will be subject to collisions from vehicles, ships, etc. during their service life, so it is very necessary to design the structure to be impact-resistant.

[0003] Currently, there are two main methods for studying the impact resistance of structures: experimental research and finite element analysis. Experimental research mainly uses devices such as drop hammer impact, pendulum impact, and horizontal trolley impact to obtain the true mechanical response of structural components under impact. However, experimental research consumes a lot of time, manpower, and costs, and this method is not directly adopted in the design of structural impact resistance. The finite element analysis method can achieve high-precision simulation of the response of the structure under impact by calibrating existing test results. However, in impact resistance design, it is necessary to consider the impact combinations that the structure may be subjected to in its environment. In other words, a large amount of parameter analysis of the structure is required to obtain the failure probability of the structure. However, conducting a large number of finite element analyses requires a lot of computing power and time.

[0004] To rapidly analyze the reliability of structures under impact, existing standards and research have proposed the use of proxy models such as empirical formulas, mass-spring models, and OpenSees models. These proxy models focus on a certain type of mechanical response of a structure or component under impact and convert this mechanical response into a performance indicator of the structure. Among these, many studies have focused on summarizing impact force responses. For example, Tin V. Do et al. established a dual-peak force impact model for a small truck impacting a reinforced concrete column, and Sherif El-Tawil established an impulse force model for a large truck impacting a bridge pier. However, these models still have limitations and are not applicable to different types of impact accidents. Summary of the Invention

[0005] The present invention aims to overcome the shortcomings of the above-mentioned prior art and provides a method for predicting the time-history response of impact force based on AEGCN-Dynamiformer. This method for predicting the time-history response of impact force based on AEGCN-Dynamiformer is time-saving, highly efficient, and widely applicable.

[0006] The second object of the present invention is to provide an impact force time history response prediction system based on AEGCN-Dynamiformer.

[0007] A third object of the present invention is to provide a storage medium.

[0008] A fourth object of the present invention is to provide a computer device.

[0009] The first object of the present invention is achieved by the following technical solution: an impact force time history response prediction method based on AEGCN-Dynamiformer, comprising the following steps:

[0010] S1. Constructing a database with component parameters, where the component parameters include geometric parameters, material parameters, boundary load conditions, and impact parameters;

[0011] S2. Convert component parameters in the database into graph structure data and perform normalization processing to form a data set for training;

[0012] S3, using the network neural network to extract the structural topological features, and using the feature fusion module to fuse the extracted structural topological features and temporal coding features to construct the AEGCN-Dynamiformer deep neural network;

[0013] S4. Use the dataset to train and optimize the AEGCN-Dynamiformer deep neural network;

[0014] S5. Input the parameters of the structure to be predicted into the trained and optimized AEGCN-Dynamiformer deep neural network to generate the impact force time history curve under the corresponding impact condition.

[0015] The database in step S1 is constructed by collecting internationally public impact test data sets, numerical simulation data sets, and engineering case data.

[0016] The process of converting the component parameters in the database into graph structure data and normalizing them in step S2 is as follows:

[0017] The structure is discretized into node features and edge features, and corresponding mechanical properties are assigned to the node features and edge features. The node features include node height, constraint conditions, node mass, impact mass, impact velocity, and axial load; the edge features include cross-sectional information, material information, and unit length; and the node feature edges, edge features, and impact force time history data are standardized.

[0018] Step S3 includes the following specific steps:

[0019] S31, using graph neural network to extract structural topological features;

[0020] S32. Tensor concatenation of the structural embedding vector generated by the graph neural network and the temporal coding features; the Performer fast attention mechanism is used on the encoder side to process the impact temporal features, and the gated recurrent unit is used on the decoder side to fuse the structural embedding vector and the temporal features, and long-term temporal dependency modeling is achieved through the encoder-decoder architecture to form the AEGCN-Dynamiformer deep neural network.

[0021] The specific process of step S31 is as follows:

[0022] An adaptive edge feature encoder based on the attention mechanism is designed. The dynamic mapping relationship between edge features and node features is established through the NNConv graph convolution layer. Multi-layer GCN is used for feature propagation. The overall structural embedding vector is generated through global pooling to complete the extraction of structural topological features.

[0023] In step S32, the encoder side includes multiple layers of encoders, each layer of encoders including:

[0024] The first layer normalization performs layer normalization on the input tensor to stabilize the subsequent FAVOR+attention calculation and alleviate the gradient vanishing problem;

[0025] The FAVOR+ mechanism establishes dynamic correlations across time steps and captures long-range dependencies of impulse responses;

[0026] The first residual connection retains the original input information and alleviates the degradation problem of deep networks;

[0027] The second layer normalizes the attention output again to provide the feedforward network with feature input that conforms to the Gaussian distribution;

[0028] Feedforward neural network, which enhances nonlinear expression capability and especially processes shock pulse characteristics;

[0029] The second residual connection ensures that the gradient is effectively returned.

[0030] The decoder module in step S32 includes:

[0031] The first gate controls the recurrent unit layer, initializes the target sequence processing, models the temporal dependencies of the decoder input, and extracts low-order temporal patterns;

[0032] The second gate controls the recurrent unit layer, capturing the dependency between the structural topological features containing temporal information and the target sequence. It jointly models the output of the first control recurrent unit layer with the graph temporal feature vector passed by the encoder to output high-order fusion features.

[0033] A decoder feed-forward network is used to map high-order fused features to the target space for full prediction.

[0034] The second object of the present invention is achieved by the following technical solution: an impact force time history response prediction system based on AEGCN-Dynamiformer, comprising:

[0035] The data acquisition and processing module is used to collect component parameters and process the component parameters to form a data set;

[0036] Neural network module, used to build the AEGCN-Dynamiformer deep neural network and train and optimize the AEGCN-Dynamiformer deep neural network;

[0037] The execution module is used to deploy the AEGCN-Dynamiformer deep neural network to detect and analyze the predicted structure to generate the impact force time history curve under the corresponding impact condition.

[0038] The third object of the present invention is achieved by the following technical solution: a storage medium stores a program, and when the program is executed by the processor, the impact force time history response prediction method based on AEGCN-Dynamiformer described in the first object is achieved.

[0039] The fourth purpose of the present invention is achieved through the following technical solution: a computer device includes a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the impact force time-history response prediction method based on AEGCN-Dynamiformer described in the first purpose is achieved.

[0040] The present invention has the following advantages over the prior art:

[0041] This method is applicable to the impact-resistant design of civil engineering structures. It can rapidly generate load-time history curves under different impact conditions, providing key input parameters for performance-based impact-resistant design. Compared to the finite element method, prediction time is reduced from hours to seconds, significantly improving the efficiency of reliability analysis and parameter optimization research.

[0042] The present invention uses the AEGCN-Dynamiformer deep neural network to predict the time-history response of impact force, which can fully consider the characteristic parameters of the structure and significantly improve the calculation efficiency of the time-history response of the structural impact force, thereby being applicable to different types of impact accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The figure is a flow chart of the method for predicting the time-history response of impact force based on AEGCN-Dynamiformer according to the present invention.

[0044] Figure 2 Schematic diagram of the model of the advanced graph neural network (AEGCN) of the present invention.

[0045] Figure 3 Schematic diagram of the Dynamiformer model in step S323 of the present invention;

[0046] Figure 4 Schematic diagram of a process for predicting the time history curve of impact force of reinforced concrete columns based on AEGCN-Dynamiformer in an embodiment of the present invention;

[0047] Figure 5 This is the impact force time history curve obtained by feeding part of the training set into the model in the embodiment of the present invention;

[0048] Figure 6 This is the impact force time history curve obtained by feeding part of the verification set into the model in the embodiment of the present invention;

[0049] Figure 7 The figure is a comparison chart of the fitting effects of the method of the present invention and different existing prediction methods. DETAILED DESCRIPTION

[0050] The present invention will be further described below with reference to the accompanying drawings and examples.

[0051] Example

[0052] like Figures 1 to 3 As shown in FIG, the impact force time history response prediction method based on AEGCN-Dynamiformer includes the following steps:

[0053] S1. Construct a database containing component parameters, including geometric parameters, material parameters, boundary load conditions, and impact parameters. The database in step S1 is constructed by collecting internationally available impact test data sets, numerical simulation data sets, and engineering case data. The database is constructed in the following manner:

[0054] Different research objects require different geometric parameters and material parameters. Geometric parameters and material parameters can be parameters related to the mechanical properties of components. For example, when studying steel tube concrete components, the geometric parameters to be collected should include cross-sectional shape, cross-sectional dimensions, steel tube thickness, component length, etc., and the material information should include steel strength, steel elastic modulus, concrete strength, concrete elastic modulus, etc.; when studying reinforced concrete components reinforced with carbon fiber reinforced plastic (CFRP) materials, the geometric parameters to be collected should include cross-sectional shape, cross-sectional dimensions, reinforcement ratio, stirrup ratio, concrete cover thickness, CFRP sheet bonding length, CFRP sheet bonding width, number of CFRP sheet bonding layers, component length, etc., and the material information should include longitudinal reinforcement strength, longitudinal reinforcement elastic modulus, stirrup strength, stirrup elastic modulus, CFRP sheet performance parameters, impregnation adhesive performance parameters, etc. The same type of components described above includes two types of components: beams and columns.

[0055] When taking reinforced concrete components as the research object, internationally published reinforced concrete component impact test data sets, numerical simulation data sets (finite element models verified by mesh convergence) and engineering case data should be collected. The collection parameters are as follows: geometric parameters include cross-sectional shape, cross-sectional size, component length, reinforcement ratio, stirrup ratio, concrete cover thickness, etc.; material parameters include longitudinal reinforcement yield strength, longitudinal reinforcement ultimate strength, stirrup reinforcement yield strength, stirrup reinforcement ultimate strength, reinforcement elastic modulus, concrete strength, concrete elastic modulus, etc.; boundary load conditions should include the fixing method at both ends of the component, the applied axial compressive load, etc.; impact parameters include impact position, impact velocity and impact mass, etc.; the data sample size N should be as large as possible.

[0056] S2. Convert component parameters in the database into graph structure data and perform normalization processing to form a data set for training;

[0057] The process of converting the component parameters in the database into graph structure data and normalizing them in step S2 is as follows:

[0058] The structure is discretized into node features and edge features, and corresponding mechanical properties are assigned to the node features and edge features. The node features include node height, constraint conditions, node mass, impact mass, impact velocity, and axial load; the edge features include cross-sectional information, material information, and unit length; and the node feature edges, edge features, and impact force time history data are standardized.

[0059] In this embodiment, the cross-section of the reinforced concrete components in the database is consistent along the length of the component. Therefore, when the reinforced concrete component parameters are converted into graph structure data, the component is simplified into three node features and two edge features. The three node features are the two endpoints and the impact point of the component, such as Figure 4If a component contains cross-sectional abrupt changes, longitudinal reinforcement cutoff points, or areas with denser stirrup reinforcement, node features can be added at both ends of these areas. Node features include node height, constraints, node mass, impact mass, impact velocity, and axial load; edge features include section position, cross-sectional area, moment of inertia, section modulus, reinforcement ratio, stirrup strength, longitudinal reinforcement strength, concrete strength, cover thickness, and edge length. Node features, edge features, and impact force time history data are standardized to construct a training dataset.

[0060] S3, using the network neural network to extract the structural topological features, and using the feature fusion module to fuse the extracted structural topological features and temporal coding features to construct the AEGCN-Dynamiformer deep neural network;

[0061] Step S3 includes the following specific steps:

[0062] S31, using a graph neural network to extract structural topological features; the specific process of step S31 is as follows:

[0063] An adaptive edge feature encoder based on the attention mechanism is designed. The dynamic mapping relationship between edge features and node features is established through the NNConv graph convolution layer. Multi-layer GCN is used for feature propagation. The overall structural embedding vector is generated through global pooling to complete the extraction of structural topological features.

[0064] Specifically, in the edge feature processing stage, the original edge features are first mapped to a high-dimensional feature space through a dimensional transformation layer. This step uses linear transformation to broaden the expressive power of features, making it easier to capture the potential relationships between different features in high-dimensional space. Subsequently, a multi-head self-attention mechanism is used to divide the high-dimensional edge features into multiple subspaces (i.e., "attention heads"), and each subspace independently calculates the synergistic effects between different feature combinations, such as the synergistic effect between steel bars and concrete. Finally, a two-layer MLP network in a feedforward neural network (FNN) is used to perform nonlinear transformation on the fused features to further enhance the model's ability to model complex synergistic effects, and finally obtain enhanced edge features.

[0065] Then, based on PyTorch's NNConv layer, the dynamically weighted edge features are coupled with the node features. This layer takes the enhanced edge features as input and dynamically generates a corresponding weight matrix for each edge. This weight matrix, along with the node's adjacency relationships, guides node feature updates. For example, the updated features of a node incorporate the material synergy information of all its adjacent edges and dynamically adjust the aggregation ratio based on the edge weights. This means that the update of a node's features depends not only on its own features but also on the features of the edges connected to it. This dynamic coupling allows the model to more accurately capture the complex relationships between nodes and edges in the graph structure.

[0066] Furthermore, a two-layer GCNConv is used for feature propagation. GCNConv is a graph convolutional network layer that aggregates the features of a node and its neighbors to generate new node features. Through two layers of GCNConv feature propagation, information about the entire graph structure is gradually aggregated, so that each node's features include information about its neighboring nodes. Finally, a global pooling layer compresses all node features into a single vector, forming a graph embedding. This vector represents the global characteristics of the structure, providing compact and information-rich input for downstream tasks.

[0067] S32. Tensor concatenation of the structural embedding vector generated by the graph neural network and the temporal coding features; the Performer fast attention mechanism is used on the encoder side to process the impact temporal features, and the gated recurrent unit is used on the decoder side to fuse the structural embedding vector and the temporal features, and long-term temporal dependency modeling is achieved through the encoder-decoder architecture to form the AEGCN-Dynamiformer deep neural network.

[0068] In step S32, the encoder side includes multiple layers of encoders, each layer of encoders including:

[0069] The first layer normalization performs layer normalization on the input tensor to stabilize the subsequent FAVOR+attention calculation and alleviate the gradient vanishing problem;

[0070] The FAVOR+ mechanism establishes dynamic correlations across time steps and captures long-range dependencies of impulse responses;

[0071] The first residual connection retains the original input information and alleviates the degradation problem of deep networks;

[0072] The second layer normalizes the attention output again to provide the feedforward network with feature input that conforms to the Gaussian distribution;

[0073] Feedforward neural network, which enhances nonlinear expression capability and especially processes shock pulse characteristics;

[0074] The second residual connection ensures that the gradient is effectively returned.

[0075] The decoder module in step S32 includes:

[0076] The first gate controls the recurrent unit layer, initializes the target sequence processing, models the temporal dependencies of the decoder input (such as historical response sequences or initial tokens), and extracts low-order temporal patterns;

[0077] The second gate controls the recurrent unit layer, capturing the dependency between the structural topological features containing temporal information and the target sequence. It jointly models the output of the first control recurrent unit layer with the graph temporal feature vector passed by the encoder to output high-order fusion features.

[0078] The decoding feedforward network uses two layers of MLP (multi-layer perceptron) (with GELU (activation function) activation in the middle) to enhance the expression ability, which is used to map high-order fusion features to the target space (such as impact force time course value) for complete prediction.

[0079] Specifically, the encoder uses the Performer's fast attention (FAVOR+) mechanism to process the temporal characteristics of the structure. The temporal information is used to achieve a low-rank approximation of the attention matrix based on the FAVOR+ algorithm. The query and key matrices of the attention mechanism are projected into a low-dimensional space using a random feature function (such as the SMERF kernel function). This constructs a low-rank approximate attention matrix, reducing the computational complexity from O(N) of traditional attention. 2 ) is reduced to O(N logN), supporting the processing of long-time-step impact time series data. At the same time, the Johnson-Lindenstrauss lemma and kernel function approximation theory ensure that the approximated network prediction performance (such as RMSE and accuracy) does not drop by more than 5%. Each layer of the encoder sequentially sets the first layer of normalization (Layernrom), the FAVOR+ mechanism, the first residual connection, the second layer of normalization (Layernrom), a two-layer MLP feedforward neural network (the first feedforward neural network and the second feedforward neural network), and the second residual connection. Among them, Layernorm aims to solve the gradient vanishing problem by normalizing the output of neurons and accelerate the training process; residual connections are used to alleviate the gradient explosion problem while preserving the integrity of the original time series information.

[0080] The decoder module uses the gated recurrent unit (GRU) as its core component, combined with residual connections and layer normalization techniques to fuse graph structure information and temporal features. The module consists of two core GRU layers (the first gate-controlled recurrent unit layer and the second gate-controlled recurrent unit layer) and a feedforward network. The initialized target sequence is first processed by the first GRU layer, and then the normalized output is concatenated with the graph structure features extracted by the encoder, and then jointly modeled by the second GRU layer. This design maintains the stability of the gradient flow and realizes the multi-level fusion of graph structure information and temporal dynamic features through the synergy of layer normalization and residual connections. Compared with the traditional Transformer decoder, this architecture uses GRU to replace the self-attention mechanism. While maintaining the ability to model long-term dependencies, it provides a more efficient computational path for specific tasks such as structural response prediction through simplified residual connections and a streamlined network structure.

[0081] S4. Use the dataset to train and optimize the AEGCN-Dynamiformer deep neural network;

[0082] Specifically, this embodiment divides the data set into a training set and a validation set in a ratio of 7:3, and uses the mean square error loss function combined with the Adam optimizer for end-to-end training. The total number of training times is set to 8000, and a cosine annealing learning rate scheduling strategy is introduced to prevent local optimality. During the training process, the validation set loss function is monitored synchronously, and the parameters are input into the optimized AEGCN-Dynamiformer deep neural network to generate the impact force time history curve under the corresponding impact condition. The real-dimensional mechanical response is obtained through denormalization processing, providing load input for the evaluation of the structural impact resistance performance.

[0083] This example verifies the superior performance of the AEGCN-Dynamiformer model by comparing the simulation effects of different prediction methods on the time-history curve of the impact force of reinforced concrete columns. A reinforced concrete column with a cross-sectional size of 260mm×260mm and a height of 1500mm was taken as the research object. Under the working conditions of an impact height of 500mm and an axial pressure of 181.2kN, three methods, OpenSees, LS-DYNA and AEGCN-Dynamiformer, were used for prediction. The specimen has a reinforcement ratio of 1.51%, a stirrup ratio of 0.76%, a longitudinal reinforcement yield strength and an ultimate strength of 451.7MPa and 608.3MPa respectively, a stirrup yield strength and an ultimate strength of 398.3MPa and 594.9MPa respectively, a concrete strength of 19.8MPa, and a concrete cover thickness of 25mm. As Figure 7As shown in the figure, the traditional OpenSees method performs poorly, while both LS-DYNA and AEGCN-Dynamiformer demonstrate good prediction accuracy. More notably, as shown in Table 1, AEGCN-Dynamiformer achieves a significant breakthrough in computational efficiency: compared to OpenSees' 2.574s iterative solution time and LS-DYNA's 600s due to explicit integration and mesh refinement, this model completes the prediction in just 0.655s, a 916-fold improvement in computational efficiency over LS-DYNA. This significantly reduces analysis time while maintaining accuracy, providing an efficient load input solution for structural impact performance assessment.

[0084] Table 1 Comparison of time consumption of different prediction methods

[0085]

[0086] The second object of the present invention is achieved by the following technical solution: an impact force time history response prediction system based on AEGCN-Dynamiformer, comprising:

[0087] The data acquisition and processing module is used to collect component parameters and process the component parameters to form a data set;

[0088] Neural network module, used to build the AEGCN-Dynamiformer deep neural network and train and optimize the AEGCN-Dynamiformer deep neural network;

[0089] The execution module is used to deploy the AEGCN-Dynamiformer deep neural network to detect and analyze the predicted structure to generate the impact force time history curve under the corresponding impact condition.

[0090] The third object of the present invention is achieved by the following technical solution: a storage medium stores a program, and when the program is executed by the processor, the impact force time history response prediction method based on AEGCN-Dynamiformer described in the first object is achieved.

[0091] The fourth purpose of the present invention is achieved through the following technical solution: a computer device includes a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the impact force time-history response prediction method based on AEGCN-Dynamiformer described in the first purpose is achieved.

[0092] The above specific implementation manner is a preferred embodiment of the present invention and does not limit the present invention. Any other changes or other equivalent replacement methods that do not deviate from the technical solution of the present invention are included in the protection scope of the present invention.

Claims

1. The impact force time history response prediction method based on AEGCN-Dynamiformer is characterized by: The following steps are involved: S1. Constructing a database with component parameters, where the component parameters include geometric parameters, material parameters, boundary load conditions, and impact parameters; S2. Convert component parameters in the database into graph structure data and perform normalization processing to form a data set for training; S3, using the network neural network to extract the structural topological features, and using the feature fusion module to fuse the extracted structural topological features and temporal coding features to construct the AEGCN-Dynamiformer deep neural network; S4. Use the dataset to train and optimize the AEGCN-Dynamiformer deep neural network; S5. Input the parameters of the structure to be predicted into the trained and optimized AEGCN-Dynamiformer deep neural network to generate the impact force time history curve under the corresponding impact condition.

2. The impact force time history response prediction method based on AEGCN-Dynamiformer according to claim 1 is characterized in that: The database in step S1 is constructed by collecting internationally public impact test data sets, numerical simulation data sets, and engineering case data.

3. The impact force time-history response prediction method based on AEGCN-Dynamiformer according to claim 1, characterized in that: The process of converting the component parameters in the database into graph structure data and normalizing them in step S2 is as follows: The structure is discretized into node features and edge features, and corresponding mechanical properties are assigned to the node features and edge features. The node features include node height, constraint conditions, node mass, impact mass, impact velocity, and axial load; the edge features include cross-sectional information, material information, and unit length; and the node feature edges, edge features, and impact force time history data are standardized.

4. The impact force time-history response prediction method based on AEGCN-Dynamiformer according to claim 1, characterized in that: Step S3 includes the following specific steps: S31, using graph neural network to extract structural topological features; S32. Tensor concatenation of the structural embedding vector generated by the graph neural network and the temporal coding features; the Performer fast attention mechanism is used on the encoder side to process the impact temporal features, and the gated recurrent unit is used on the decoder side to fuse the structural embedding vector and the temporal features, and long-term temporal dependency modeling is achieved through the encoder-decoder architecture to form the AEGCN-Dynamiformer deep neural network.

5. The impact force time-history response prediction method based on AEGCN-Dynamiformer according to claim 4, characterized in that: The specific process of step S31 is as follows: An adaptive edge feature encoder based on the attention mechanism is designed. The dynamic mapping relationship between edge features and node features is established through the NNConv graph convolution layer. Multi-layer GCN is used for feature propagation. The overall structural embedding vector is generated through global pooling to complete the extraction of structural topological features.

6. The impact force time history response prediction method based on AEGCN-Dynamiformer according to claim 4 is characterized in that: In step S32, the encoder side includes multiple layers of encoders, each layer of encoders including: The first layer normalization performs layer normalization on the input tensor to stabilize the subsequent FAVOR+attention calculation and alleviate the gradient vanishing problem; The FAVOR+ mechanism establishes dynamic correlations across time steps and captures long-range dependencies of impulse responses; The first residual connection retains the original input information and alleviates the degradation problem of deep networks; The second layer normalizes the attention output again to provide the feedforward network with feature input that conforms to the Gaussian distribution; Feedforward neural network, which enhances nonlinear expression capability and especially processes shock pulse characteristics; The second residual connection ensures that the gradient is effectively returned.

7. The impact force time-history response prediction method based on AEGCN-Dynamiformer according to claim 4, characterized in that: The decoder module in step S32 includes: The first gate controls the recurrent unit layer, initializes the target sequence processing, models the temporal dependencies of the decoder input, and extracts low-order temporal patterns; The second gate controls the recurrent unit layer, capturing the dependency between the structural topological features containing temporal information and the target sequence. It jointly models the output of the first control recurrent unit layer with the graph temporal feature vector passed by the encoder to output high-order fusion features. A decoder feed-forward network is used to map high-order fused features to the target space for full prediction.

8. The impact force time history response prediction system based on AEGCN-Dynamiformer is characterized by: include: The data acquisition and processing module is used to collect component parameters and process the component parameters to form a data set; Neural network module, used to build the AEGCN-Dynamiformer deep neural network and train and optimize the AEGCN-Dynamiformer deep neural network; The execution module is used to deploy the AEGCN-Dynamiformer deep neural network to detect and analyze the predicted structure to generate the impact force time history curve under the corresponding impact condition.

9. A storage medium, characterized in that A program is stored, and when the program is executed by a processor, the impact force time history response prediction method based on AEGCN-Dynamiformer according to any one of claims 1 to 7 is realized.

10. Computer device, characterized in that The invention comprises a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the impact force time history response prediction method based on AEGCN-Dynamiformer according to any one of claims 1 to 7 is implemented.

Images