A method and system for predicting structural performance of an integrated die-cast rear floor

Abstract

The application belongs to the technical field of lightweight structure design, and specifically discloses a structure performance physical field prediction method and system for an integrated die-cast rear floor. According to the application, a sub-region graph structure is constructed according to a region division result; a physical field low-dimensional hidden variable is calculated based on an encoder; an enhanced feature vector is generated according to a key force transmission path and a node-level feature matrix through prior guided attention; and an iteration prediction is performed based on a target prediction framework, and a structure performance full physical field of the integrated die-cast rear floor is determined according to the physical field of each sub-region. In the above manner, a time series prediction model is not directly constructed on a high-dimensional node physical field, but the complete physical field of each time step is first compressed into a low-dimensional global hidden variable through an encoder, then an iteration prediction is performed based on a target prediction framework, and the predicted hidden variable is reconstructed into a full physical field through a decoder, so as to reduce the input and output dimensions, thereby effectively improving the efficiency and accuracy of predicting the structure performance full physical field.

Description

Technical Field

[0001] This application belongs to the field of lightweight structural design technology, and more specifically, relates to a method and system for predicting the physical field of structural performance of an integral die-cast floor. Background Technology

[0002] The structure of a one-piece die-cast floor is extremely complex, with numerous design variables and strong coupling, making traditional finite element simulation-based optimization processes extremely time-consuming. A single high-fidelity collision simulation requires the creation of 500,000 to 1.5 million elements, resulting in extremely long computation times. A full-cycle simulation can involve hundreds of iterations, severely limiting design iteration efficiency. Furthermore, the one-piece die-cast floor must simultaneously meet requirements for bending stiffness, torsional stiffness, modal frequencies, rear-impact intrusion, peak force, energy absorption, and connection point loads. Trade-offs exist between these performance characteristics, requiring the model to simultaneously predict multiple physical fields and engineering parameters.

[0003] Existing data-driven surrogate model methods suffer from several technical shortcomings when addressing the aforementioned problems. First, different design schemes for the rear floor often employ different mesh discretization schemes with varying numbers of nodes. Traditional intrinsic orthogonal decomposition or principal component analysis dimensionality reduction methods require all samples to have a completely consistent number of nodes and topological correspondence, which is almost impossible to achieve in practical engineering. Furthermore, purely data-driven neural network models often deviate from the true constitutive relationship in high strain rate and high damage regions, exhibiting problems such as excessively high or low stress and damage evolution that does not conform to physical laws. In addition, existing physics prediction models are mostly trained using GPU cluster acceleration, lacking optimization for prediction CPU architectures. Therefore, the efficiency and accuracy of the above methods in predicting the full physics field of the structural performance of a monolithic die-cast rear floor are relatively low. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this application aims to provide a method and system for predicting the physical field of structural performance of a die-cast integrated floor. This method addresses the problems of low efficiency and accuracy in predicting the full physical field of structural performance of a die-cast integrated floor, which are caused by the requirement that all samples have the same number of nodes and topological correspondence, the fact that prediction results in high strain rate and high damage areas often deviate from the true constitutive relationship, and the lack of optimization for the prediction CPU architecture.

[0005] To achieve the above objectives, in a first aspect, this application provides a method for predicting the physical field of structural performance of a die-cast integrated floor, comprising: Obtain the current CAE file of the integrated die-cast floor, divide the current CAE file into regions based on a preset geometric partitioner, and construct a sub-region map structure based on the region division results; Based on the encoder, the low-dimensional latent variables of the physical field are calculated according to the sub-region graph structure, and the low-dimensional latent variables of the physical field in each region are merged into global latent variables. Extract the key force transmission path from the loading point to the installation point of the integral die-cast floor from the current CAE file, and generate an enhanced feature vector based on the key force transmission path and the node-level feature matrix through prior-guided attention. Based on the target prediction framework, iterative prediction is performed according to the enhanced feature vector, and when the maximum time step is reached, the predicted global latent variable sequence is projected to different sub-regions. The structural performance full physical field of the integral die-cast floor is determined according to the physical field of each sub-region.

[0006] In one embodiment, the steps of obtaining the current CAE file of the integral die-cast floor, dividing the current CAE file into regions based on a preset geometric partitioner, and constructing a sub-region map structure based on the region division results include: Obtain the current CAE file of the integrated die-cast rear floor, and select the finite element model of the rear floor corresponding to the center point of the design space based on the preset geometric partitioner and the current CAE file. The region is divided by assigning labels to each element in the finite element model of the rear floor to its respective sub-region, and the sub-region envelope is calculated based on the node coordinates of each element after the labels are assigned. Calculate the symbolic distance from the centroid coordinates of each unit to the envelope of the sub-region, and assign regions to each unit based on the symbolic distance; Obtain the node set of the sub-region after the assignment unit, and determine the internal edge set and adjacency matrix based on the node set. Construct the sub-region graph structure based on the node set, the internal edge set, and the adjacency matrix.

[0007] In one embodiment, the step of calculating low-dimensional latent variables of the physical field based on the encoder and the sub-region graph structure, and fusing the low-dimensional latent variables of the physical field of each region into global latent variables, includes: Obtain the calculation result file based on the current CAE file of the integral die-cast floor, and calculate the physical feature matrix of the nodes in the sub-region at the current time based on the calculation result file. Based on the encoder, low-dimensional latent variables of the physical field are calculated according to the physical feature matrix and the adjacency matrix in the sub-region graph structure; The query vector, key vector, and value vector of each sub-region are calculated based on the learnable projection matrix, and the attention weights are calculated based on the query vector and the key vector. Calculate the context vector for each sub-region based on the attention weights and the value vectors; The context vector is concatenated with the low-dimensional latent variables of the physical field, and the concatenated latent variables are fused into global latent variables based on the region fusion network.

[0008] In one embodiment, the step of extracting the key force transmission path from the loading point to the installation point of the integral die-cast floor from the current CAE file, and generating an enhanced feature vector based on the key force transmission path and the node-level feature matrix through prior-guided attention, includes: The total loss function is constructed based on the mean square error of the nodal physics field reconstruction, the divergence loss, the constraints of the cast aluminum material damage model, the tension-compression asymmetric yield constraint of the cast aluminum material, and the residual constraints of the implicit space momentum equation. The physical field mapping unit of the integral die-cast floor is trained according to the total loss function, and the mapped global latent variables are obtained based on the physical field mapping unit. The node-level feature matrix is ​​determined based on the mapped global latent variables; Extract the key force transmission path from the loading point to the installation point of the integral die-cast floor from the current CAE file, and generate an enhanced feature vector based on the key force transmission path and the node-level feature matrix through prior-guided attention.

[0009] In one embodiment, the step of determining the node-level feature matrix based on the mapped global latent variables includes: Obtain the global latent variable sequence within the historical time window, and perform position encoding on the mapped global latent variables at each time step based on the global latent variable sequence within the historical time window and the position encoding information using sine and cosine functions; The encoded global latent variables at each time step are broadcast to each node of the integral die-cast floor, and the node coordinates and wall thickness are obtained respectively. Based on multi-layer graph convolution, the global latent variables at the nodes, the node coordinates, and the wall thickness are concatenated into a node-level feature matrix.

[0010] In one embodiment, the step of iteratively predicting based on the enhanced feature vector using the target prediction framework, and projecting the predicted global latent variable sequence to different sub-regions when the maximum time step is reached, and determining the full physical field of the structural performance of the integral die-cast floor based on the physical field of each sub-region, includes: The path enhancement feature sequence and the original global latent variable sequence are concatenated for each time step within the historical time window; Based on the target prediction framework, iterative prediction is performed according to the enhanced feature vector and the concatenation result; When the maximum time step is reached, the predicted global latent variable sequence is projected to different sub-regions; The global latent variables of the different sub-regions are input into the decoder so that the decoder outputs the physical field of the sub-region; The physical fields of each of the sub-regions are spliced ​​together to determine the full physical field of the structural performance of the integral die-cast floor.

[0011] Secondly, this application provides a physical field prediction system for the structural performance of an integral die-cast floor, comprising: The partitioning module is used to obtain the current CAE file of the integrated die-cast floor, partition the current CAE file into regions based on a preset geometric partitioner, and construct a sub-region map structure based on the region partitioning results. The calculation module is used to calculate the low-dimensional latent variables of the physical field based on the encoder and the sub-region graph structure, and to merge the low-dimensional latent variables of the physical field of each region into global latent variables. The generation module is used to extract the key force transmission path from the loading point to the installation point of the integral die-cast floor from the current CAE file, and generate an enhanced feature vector based on the key force transmission path and the node-level feature matrix through prior-guided attention. The prediction module is used to perform iterative prediction based on the target prediction framework and the enhanced feature vector. When the maximum time step is reached, the predicted global latent variable sequence is projected to different sub-regions, and the structural performance full physical field of the integral die-cast floor is determined according to the physical field of each sub-region.

[0012] Thirdly, this application provides an electronic device, comprising: at least one memory for storing a program; and at least one processor for executing the program stored in the memory, wherein when the program stored in the memory is executed, the processor is configured to execute the method described in the first aspect or any possible implementation thereof.

[0013] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to perform the method described in the first aspect or any possible implementation thereof.

[0014] Fifthly, this application provides a computer program product that, when run on a processor, causes the processor to perform the method described in the first aspect or any possible implementation thereof.

[0015] It is understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0016] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art: (1) This application does not directly construct a time series prediction model on the high-dimensional node physical field. Instead, after constructing a sub-region graph structure based on the region division results, it calculates the low-dimensional latent variables of the physical field based on the encoder and the sub-region graph structure. The low-dimensional latent variables of the physical field in each region are then merged into global latent variables. Iterative prediction is then performed based on the target prediction framework. Finally, the predicted latent variables are reconstructed into the full physical field through the decoder. This achieves an indirect prediction path of "dimensionality reduction, prediction, and reconstruction" without directly constructing a time series prediction model on the high-dimensional node physical field. This reduces the input and output dimensions, making the prediction task, which originally required a huge number of parameters and computational resources, lighter. This effectively improves the efficiency and accuracy of predicting the full physical field of the structural performance. At the same time, the latent variables themselves encode the global deformation mode and stress distribution characteristics of the physical field. Predicting the latent variables is equivalent to predicting the key essential properties of the physical field, avoiding the dimensionality curse and overfitting problems faced when directly predicting high-dimensional fields.

[0017] (2) This application elevates physical constraints to the latent space for execution. Specifically, it utilizes the Jacobian matrix or eigenorthogonal decomposition modes of the latent variables output by the decoder to construct a reduced-order basis, and projects the residuals of the high-dimensional momentum equations onto the low-dimensional latent space to obtain reduced-order physical residual terms. The calculation of these reduced-order physical residuals only requires approximation on a small number of randomly sampled nodes, and its computational cost is related to the dimension of the latent variables rather than the number of nodes, so that the addition of physical constraints hardly destroys the lightweight characteristics of the model. Through latent space physical guidance, the model can still learn the evolution laws of latent variables that conform to the conservation of momentum, conservation of energy, and the constitutive relation of cast aluminum, even with limited training samples, thereby indirectly ensuring the physical consistency of the reconstructed physical field.

[0018] (3) After generating the enhanced feature vector, this application introduces a target prediction framework (e.g., the Transformer architecture), using the global latent variable sequence within the historical window as input, and predicts the latent variables for multiple future time steps in parallel through a self-attention mechanism, completely avoiding the sequential dependency of time steps in recurrent networks. This design ensures that the computation time for latent variable prediction does not increase linearly with the sequence length, naturally adapting to CPU parallel computing. At the same time, a few force transmission paths extracted from the mesh diagram are used to enhance the representation of latent variables: the node features on the path are aggregated by one-dimensional convolution, and then fused with the global latent variables through a prior attention mechanism, so that the evolution of latent variables focuses on key mechanical channels. Since the dimension of latent variables is much lower than that of the original physical field, the extraction and fusion of path features only requires processing vectors of hundreds of dimensions, resulting in extremely low computational overhead. Finally, the entire prediction process can be completed in milliseconds, achieving truly lightweight real-time prediction, and the predicted physical field maintains complete information comparable to that of finite element simulation due to the indirect paths.

[0019] In summary, this application obtains the current CAE file of the integral die-cast floor, divides the current CAE file into regions based on a preset geometric partitioner, and constructs a sub-region graph structure based on the region division results; based on an encoder, it calculates low-dimensional latent variables of the physical field according to the sub-region graph structure, and merges the low-dimensional latent variables of the physical field of each region into global latent variables; it extracts the key force transmission path of the integral die-cast floor from the loading point to the installation point from the current CAE file, and generates enhanced feature vectors based on the key force transmission path and node-level feature matrix through prior-guided attention; based on the target prediction framework, iterative prediction is performed according to the enhanced feature vectors, and when the maximum time step is reached, the predicted global latent variable sequence is projected to different sub-regions, and the structural performance full physical field of the integral die-cast floor is determined according to the physical field of each sub-region. By using the above method, instead of directly building a time series prediction model on the high-dimensional node physical field, the complete physical field of each time step is first compressed into low-dimensional global latent variables by the encoder, then iterative prediction is performed based on the target prediction framework, and finally the predicted latent variables are reconstructed into the full physical field by the decoder, so as to reduce the input and output dimensions and thus effectively improve the efficiency and accuracy of the prediction structure performance of the full physical field. Attached Figure Description

[0020] Figure 1 This is one of the flowcharts illustrating the physical field prediction method for the structural performance of an integral die-cast floor provided in this application embodiment; Figure 2 This is a multi-stage schematic diagram of the physical field prediction of structural performance provided in an embodiment of this application; Figure 3 This is the second flowchart illustrating the method for predicting the physical field of structural performance of an integral die-cast floor provided in this application embodiment; Figure 4 This is a schematic diagram of the module structure of the physical field prediction system for the structural performance of the integrated die-cast floor provided in this application embodiment; Figure 5 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0022] In this article, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. The symbol " / " in this article indicates that the related objects are in an "or" relationship; for example, A / B means A or B.

[0023] The terms "first" and "second," etc., used in the specification and claims herein are used to distinguish different objects, not to describe a specific order of objects. For example, "first response message" and "second response message," etc., are used to distinguish different response messages, not to describe a specific order of response messages.

[0024] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0025] Based on this, embodiments of this application provide a method for predicting the physical field of structural performance of an integral die-cast floor, referring to... Figure 1 , Figure 1 This is one of the flowcharts illustrating the method for predicting the structural performance physical field of a die-cast integrated floor provided in this application. In this embodiment, the method for predicting the structural performance physical field of a die-cast integrated floor includes steps S10 to S40: Step S10: Obtain the current CAE file of the integrated die-cast floor, divide the current CAE file into regions based on a preset geometric partitioner, and construct a sub-region map structure based on the region division results.

[0026] It should be noted that the physical field prediction method for the structural performance of the integral die-cast floor in this embodiment is divided into three stages: the data preprocessing stage of the integral die-cast floor, the physical field mapping stage of the integral die-cast floor, and the graph time-series prediction stage based on force transmission path awareness. For details, please refer to... Figure 2 In the data preprocessing stage of the integral die-cast floor, a structure-performance dataset of the integral die-cast floor is established, including but not limited to geometric design parameter vectors, finite element mesh files, and calculation result files. In the physical field mapping stage of the integral die-cast floor, an encoder with sub-region physical fields is constructed. In the graph temporal prediction stage based on force path awareness, a floor force path extractor, a physical field temporal encoder, a feature fusion network, and a target prediction framework are involved. Iterative prediction is performed based on the target prediction framework, and the predicted global latent variable sequence is projected to different sub-regions. The structural performance full physical field of the integral die-cast floor is determined according to the physical field of each sub-region.

[0027] It should be understood that by collecting finite element models and calculation results of at least N different design schemes for the integral die-cast floor, a structure-performance dataset for the integral die-cast floor can be established based on parameters such as geometric design parameter vectors, finite element mesh files, and calculation result files. The geometric design parameter vectors include, but are not limited to, wall thickness distribution parameters, stiffener geometric parameters, boss positions and dimensions, and fillet radii; the finite element mesh files include, but are not limited to, element node coordinate matrices, mesh element topology connections, and element wall thickness attributes; the calculation result files include, but are not limited to, the front... The nodal displacement at each time step Element integral point stress tensor Element equivalent plastic strain In addition, the historical variable output for the properties of cast aluminum materials is the porosity fraction at the unit integration point. and stress components .

[0028] Understandably, after obtaining the current CAE file of the integral die-cast rear floor, the rear floor mesh in the finite element mesh file can be divided into K sub-regions based on the wall thickness and geometric features in the geometric design parameter vector, and solidified into a preset geometric partitioner. This allows for the same sub-region division of different rear floor meshes, and the construction of a sub-region graph structure based on the region division results.

[0029] Further, step S10 includes: obtaining the current CAE file of the integral die-cast floor, and selecting the finite element model of the floor corresponding to the center point of the design space based on the current CAE file using a preset geometric partitioner; dividing the floor into regions by assigning sub-region labels to each element in the finite element model of the floor, and calculating the sub-region envelope based on the node coordinates of each element after assigning labels; calculating the symbolic distance from the centroid coordinates of each element to the sub-region envelope, and assigning regions to each element based on the symbolic distance; obtaining the node set of the sub-region after assigning the unit, and determining the internal edge set and adjacency matrix based on the node set, and constructing the sub-region graph structure based on the node set, the internal edge set, and the adjacency matrix.

[0030] It should be understood that, for the preset geometric partitioner, it is necessary to select the finite element model of the rear floor corresponding to the center point of the design space based on the current CAE file, and use this finite element model of the rear floor as a reference template. For use as a reference template Each unit in In other words, assigning labels to its sub-regions. For each sub-region, collect reference templates. Chinese label Calculate the 3D convex hull of all elements and all node coordinates. If a region is concave, then an axis-aligned bounding box hierarchical tree is used to record the spatial occupancy distribution of the centroids of the units within that region.

[0031] It should also be emphasized that for the input rear floor finite element model, it is necessary to traverse each element within it. Calculate its centroid coordinates And calculate the centroid coordinates. To the envelope of each sub-region The sign distance is used to assign regions to each unit, specifically: if there exists a unique... Make the centroid coordinates Located in the sub-region envelope Interior (i.e., symbolic distance) ), then the unit Assigned to the area If the centroid coordinates If the location is in the region where multiple convex hulls overlap, then calculate the centroid coordinates. Based on the Euclidean distance to the mean center of the centroid points of each overlapping region, the region label with the closest distance is selected for assignment. If the centroid coordinates... Located outside all convex hulls (all symbolic distances) Then calculate the centroid coordinates. The minimum distance to each convex hull boundary surface is used to select the region label with the smallest distance for assignment. In addition, for sub-regions, including the rear longitudinal beam force transmission area, stiffener mesh area, perimeter flange and installation point area, the region label of each node is recorded, and the connecting edges between regions are extracted.

[0032] It is understandable that after the unit assignment is completed, each sub-region is defined based on the node set, the internal edge set, and the adjacency matrix. k The sub-region graph structure is constructed as follows: ,in, This represents the set of nodes in the region, with a quantity of . ; Represents the set of internal edges, if both endpoints belong to the set of nodes. Furthermore, if a cell is shared within the original mesh, then an edge is considered to exist at that location; Represents the adjacency matrix. If node and If there is an edge, then the corresponding Set the value to 1 if it is a boundary condition, otherwise set it to 0; also record the set of cross-region edges. , i.e., node Belonging to subregion k, node Belongs to sub-region Furthermore, the two nodes share the same cell edge in the original mesh.

[0033] It is also important to emphasize that, based on the computational results file, it is necessary to extract time-series physical field indices at the node level for different regions, including but not limited to displacement, stress, and strain. Specifically, the mean and standard deviation of the nodal displacement field physical quantities are calculated over the entire dataset, and Z-score standardization is performed. Since the number of nodes varies in different regions, the standardization parameters are stored separately for each region. For physical quantities defined at the element center (such as element integration point stress and plastic strain), they are mapped to the nodes through shape function interpolation or area-weighted averaging; ultimately forming... k Ft(k) is the physical feature matrix of the nodes in the region at time t, where D is the physical feature dimension of each node.

[0034] Step S20: Based on the encoder, calculate the low-dimensional latent variables of the physical field according to the sub-region graph structure, and merge the low-dimensional latent variables of the physical field of each region into global latent variables.

[0035] It is understandable that, in each sub-region k After constructing the sub-region graph structure, encoders are built for each region to calculate the low-dimensional latent variables of the physical field in different regions. Global latent variables refer to those generated by... k The global latent variable, formed by fusing low-dimensional latent variables of the physical fields in each sub-region, can be represented as: .

[0036] Further, step S20 includes: obtaining a calculation result file based on the current CAE file of the integral die-cast floor, and calculating the physical feature matrix of the nodes in the sub-region at the current time based on the calculation result file; calculating the low-dimensional latent variables of the physical field based on the encoder, the physical feature matrix, and the adjacency matrix in the sub-region graph structure; calculating the query vector, key vector, and value vector of each sub-region based on the learnable projection matrix, and calculating the attention weight based on the query vector and the key vector; calculating the context vector of each sub-region based on the attention weight and the value vector; concatenating the context vector with the low-dimensional latent variables of the physical field, and fusing the concatenated latent variables into a global latent variable based on the region fusion network.

[0037] It should be understood that, for each sub-region k In terms of encoder Based on the feature matrix of the node and adjacency matrix As input, output the physical field of the region at time step Low-dimensional hidden variables of physical fields The distribution function, including the mean and logarithmic variance The specific encoder It consists of two graph convolutional layers, then graph global pooling and fully connected layers are used to map the encoded features to the latent space, and finally reparameterized sampling is used to obtain the results, as shown in the following equation: .

[0038] in, Represents low-dimensional hidden variables of physical fields. Represents element-wise product. It follows a standard normal distribution. This represents the mean. It represents the logarithmic variance.

[0039] Understandably, when fusing low-dimensional latent variables of a physical field, it is necessary to calculate the query vector, key vector, and value vector for each sub-region based on the learnable projection matrix. Specifically: query vector key vector value vector ,in , , This represents the learnable projection matrix. For attention weights, it represents the region. For the region The importance of this lies in the fact that, after calculating the context vectors of each sub-region, they can be compared with the low-dimensional latent variables of the physical field. The data is then stitched together and then processed by MLP to obtain the fused regional features. ,in, As a whole, it represents a projected multilayer perceptron. This represents the context vector of sub-region k, and the global latent variables are obtained by average pooling the features of all regions. Latent variables of subregions Merge into global hidden variables .

[0040] Step S30: Extract the key force transmission path from the loading point to the installation point of the integral die-cast floor from the current CAE file, and generate an enhanced feature vector based on the key force transmission path and the node-level feature matrix through prior-guided attention.

[0041] It should be understood that after extracting the key force transmission path from the loading point to the installation point of the integral die-cast floor from the current CAE file, it is solidified into a force transmission path extractor. The node-level feature matrix refers to the node-level feature matrix obtained by two layers of graph convolution, and the enhanced feature vector refers to the single feature vector formed by prior-guided attention fusion.

[0042] Step S40: Based on the target prediction framework, iterative prediction is performed according to the enhanced feature vector, and when the maximum time step is reached, the predicted global latent variable sequence is projected to different sub-regions, and the structural performance full physical field of the integral die-cast floor is determined according to the physical field of each sub-region.

[0043] It should be noted that the target prediction framework refers to a prediction framework that iteratively predicts the global latent variable sequence. This target prediction framework can be a Transformer prediction framework. After iterative prediction based on the target prediction framework, the last step or several steps of the prediction are combined with the historical window and the above steps are repeated until the maximum time step is reached. At this point, the predicted global latent variable sequence needs to be projected to different sub-regions. Based on the physical field of each sub-region, the full physical field of the structural performance of the integral die-cast floor is determined, realizing an indirect prediction path of "first dimensionality reduction, then prediction, and then reconstruction". This avoids directly building a time series prediction model on the high-dimensional node physical field, thereby reducing the input and output dimensions and making the prediction task, which originally required a huge number of parameters and computational resources, more lightweight. At the same time, the latent variables themselves encode the global deformation mode and stress distribution characteristics of the physical field. Predicting the latent variables is equivalent to predicting the key essential properties of the physical field, avoiding the curse of dimensionality and overfitting problems faced when directly predicting high-dimensional fields.

[0044] Further, step S40 includes: concatenating the path enhancement feature sequence and the original global latent variable sequence for each time step within the historical time window; performing iterative prediction based on the target prediction framework, according to the enhanced feature vector and the concatenation result; when the maximum time step is reached, projecting the predicted global latent variable sequence to different sub-regions; inputting the global latent variables of the different sub-regions into the decoder so that the decoder outputs the physical field of the sub-region; and concatenating the physical fields of each sub-region to determine the full physical field of the structural performance of the integral die-cast floor.

[0045] It should be understood that this embodiment can be applied to a target prediction framework, using the path enhancement feature sequence and the original global latent variable sequence at each time step to predict future targets. The latent variable sequence of each step is used for prediction, and the trained Transformer prediction framework is fixed. Specifically, the path enhancement feature sequence is used for each time step within the historical time window. and the original global hidden variable sequence The data is concatenated and used as input to the Transformer encoder. It can be calculated using the following formula: .

[0046] in, This represents the path enhancement feature sequence at each time step within a historical time window. This represents the sequence of original global hidden variables within a historical time window.

[0047] It's also important to emphasize that the Transformer encoder's structure involves a learnable linear projection that maps the input to the model dimension, followed by positional encoding, and finally, stacking multiple layers of Transformer encoders to obtain the final output. In contrast, the Transformer decoder in the object prediction framework employs a single-step direct prediction strategy. Specifically, it performs global pooling of the encoder output in the time dimension to obtain a fixed-length context vector, and then directly outputs the future value through a multilayer perceptron. The hidden variable sequence of the step is as follows: .

[0048] in, Indicates the future The sequence of latent variables for each step.

[0049] It should also be emphasized that, for the target prediction framework, the loss function used for training can be expressed as: .

[0050] in, This represents the loss function used to train the target prediction framework. Indicates the predicted value. Represents the actual value.

[0051] This embodiment acquires the current CAE file of the integrated die-cast floor, divides the current CAE file into regions based on a preset geometric partitioner, and constructs a sub-region graph structure based on the region division results; based on the encoder, it calculates the low-dimensional latent variables of the physical field according to the sub-region graph structure, and merges the low-dimensional latent variables of the physical field of each region into global latent variables; it extracts the key force transmission path of the integrated die-cast floor from the loading point to the installation point from the current CAE file, and generates enhanced feature vectors based on the key force transmission path and node-level feature matrix through prior-guided attention; based on the target prediction framework, iterative prediction is performed according to the enhanced feature vectors, and when the maximum time step is reached, the predicted global latent variable sequence is projected to different sub-regions, and the structural performance full physical field of the integrated die-cast floor is determined according to the physical field of each sub-region. By using the above method, instead of directly building a time series prediction model on the high-dimensional node physical field, the complete physical field of each time step is first compressed into low-dimensional global latent variables by the encoder, then iterative prediction is performed based on the target prediction framework, and finally the predicted latent variables are reconstructed into the full physical field by the decoder, so as to reduce the input and output dimensions and thus effectively improve the efficiency and accuracy of the prediction structure performance of the full physical field.

[0052] In one specific implementation, this application provides steps for generating enhanced feature vectors. Please refer to... Figure 3 , Figure 3 This is the second flowchart illustrating the method for predicting the physical field of structural performance of a die-cast integrated floor provided in this application. Step S30 includes steps S301 to S304: Step S301: Construct the total loss function based on the reconstructed mean square error of the nodal physical field, divergence loss, cast aluminum material damage model constraints, cast aluminum material tension-compression asymmetric yield constraints, and latent space momentum equation residual constraints.

[0053] It should be noted that decoders for each region also need to be built. This maps the latent variables back to the physical field of the region. Specifically, for the region... Each node decoder Construct condition vector ,in For node coordinates, The thickness at the node. The hidden physical variables at the nodes are then generated using a multi-layer MLP. physical feature vector Finally, the reconstructed physical feature matrix is ​​obtained by performing parallel computation on all nodes. .

[0054] It should be understood that the total loss function refers to the loss function used to train the physical field mapping unit of the integral die-cast floor. This total loss function can be constructed based on the mean square error of the nodal physical field reconstruction, divergence loss, cast aluminum material damage model constraints, cast aluminum material tension-compression asymmetric yield constraints, and latent space momentum equation residual constraints, specifically: .

[0055] in, Represents the total loss function. This represents the mean square error of the nodal physics reconstruction. ,express KL Divergence loss, This indicates the constraints of the GTN damage model for cast aluminum materials. This indicates the asymmetric yield constraint under tension and compression in cast aluminum materials. This represents the residual constraint of the implicit space momentum equation. This represents the preset hyperparameters used to balance the weights of the various loss terms. Training uses the Adam optimizer with an initial learning rate of [value missing]. The learning rate is halved if the validation set loss does not decrease for 5 consecutive epochs, and the total loss decreases by less than [a certain percentage]. Training should be stopped at this time.

[0056] On the one hand, constraints on the GTN damage model for cast aluminum materials Specifically, it is defined as the residual between the predicted stress triaxiality and the equivalent plastic strain relationship and the GTN yield surface, and its calculation method is as follows: .

[0057] in, Represents the GTN yield function. Indicates hydrostatic stress. Indicates equivalent stress. Represents equivalent plastic strain. This represents the equivalent void volume fraction; all the above physical quantities are calculated using the nodal stress and strain fields output by the decoder. This represents a subset of nodes randomly sampled in high-stress regions of the structure to reduce computational overhead. This represents the total number of nodes randomly sampled in the high-stress region of the structure, i.e. The total number of sample points.

[0058] On the other hand, for the tension-compression asymmetric yield constraint of cast aluminum materials, it is specifically defined as the residual of the predicted stress state deviating from the CPB06 yield criterion, and its calculation formula is as follows: .

[0059] in, The yield stress function, which takes into account tension-compression asymmetry, can be calibrated based on uniaxial tension-compression experimental data of cast aluminum materials.

[0060] On the other hand, regarding the residual constraints of the latent space momentum equation... In other words, the calculation method is as follows: for each training batch, M latent variables z are randomly sampled, and the corresponding physical field is decoded. Calculate the nodal acceleration field in the critical region of the structure. Constructing the residual of the discrete momentum equation Where M represents the mass matrix, C represents the damping matrix, and P represents the internal force vector. Using the Jacobian matrix of the decoder with respect to the latent variables, the residual is projected back into the latent space, specifically: .

[0061] in, The Jacobian matrix of the decoder output with respect to the latent variables is efficiently calculated through automatic differentiation.

[0062] Step S302: Train the physical field mapping unit of the integral die-cast floor according to the total loss function, and obtain the mapped global latent variables based on the physical field mapping unit.

[0063] Step S303: Determine the node-level feature matrix based on the mapped global latent variables.

[0064] Further, step S303 includes: obtaining the global latent variable sequence within the historical time window, and performing position encoding on the mapped global latent variables at each time step based on the global latent variable sequence within the historical time window and the position encoding information using sine and cosine functions; broadcasting the encoded global latent variables at each time step to each node of the integral die-cast floor, and obtaining the node coordinates and wall thickness respectively; and concatenating the global latent variables at the node, the node coordinates, and the wall thickness into a node-level feature matrix based on multi-layer graph convolution.

[0065] Understandably, for obtaining mapped global latent variables based on physical field mapping units, temporal feature encoding is required. Specifically, this involves encoding the historical time window. Global hidden variable sequence within As input, where This represents the window length. For the mapped global latent variables at each time step, position encoding is required, specifically: .

[0066] in, This represents encoded global implicit variables. Indicates time information, , This indicates the positional encoding information using sine and cosine functions.

[0067] It should also be emphasized that, after obtaining the encoded global latent variables Then, it is broadcast to each node of the integrated die-cast floor, spliced ​​with the node coordinates and wall thickness, and obtained as a node-level feature matrix through two layers of graph convolution.

[0068] Step S304: Extract the key force transmission path from the loading point to the installation point of the integral die-cast floor from the current CAE file, and generate an enhanced feature vector based on the key force transmission path and the node-level feature matrix through prior-guided attention.

[0069] Understandably, after extracting the key force transmission path from the loading point to the mounting point of the integrated die-cast rear floor from the current CAE file, it is solidified into a force transmission path extractor. This embodiment uses a rear collision scenario as an example for explanation. The loading point is the connection area between the rear bumper and the rear floor, and the mounting points are the front ends of the left / right rear longitudinal beams, the left / right suspension mounting points, etc. Specifically, the entire rear floor mesh diagram is regarded as an undirected weighted graph. ,in This is the set of mesh nodes for the entire rear floor. Let the set of edges have weights. Let the energy transfer efficiency of the unit be denoted by the following formula, which is calculated as follows: .

[0070] in, Indicates the average wall thickness of the element. Indicates the principal curvature at the location.

[0071] It should be understood that for each node pair Using A The algorithm finds the path with the highest energy transfer efficiency. And record the node sequence for each path. , Representing a path The number of nodes in the sequence; for each node sequence, its path importance is calculated using the following formula. Specifically: .

[0072] in, Representing an edge Average energy transfer efficiency across all training samples. Indicates the first q A key force transmission path.

[0073] It should also be emphasized that the importance of the computational path Then, based on path importance node feature sequence Considered to be of length The sequence is processed by applying multiple 1D convolutional kernels to extract multi-scale features, and then global max pooling is performed to obtain the path feature vector. This operation processes all paths and time steps in parallel, and combines the features of all paths at the current time step. Prior-guided attention is fused into a single enhanced feature vector. Specifically:

[0074] .

[0075] in, Represents a single enhanced feature vector. , This represents the parameters that can be learned during network training. Indicates the importance of the path. Represents the path feature vector. Indicates the first q The path importance coefficient of a key force transmission path.

[0076] This embodiment constructs a total loss function based on the mean square error of the node physics reconstruction, divergence loss, constraints of the cast aluminum material damage model, constraints of the tensile-compressive asymmetric yield of the cast aluminum material, and constraints of the residuals of the latent space momentum equation. A physics mapping unit for the integral die-cast floor is trained based on this total loss function, and mapped global latent variables are obtained based on the physics mapping unit. A node-level feature matrix is ​​determined based on the mapped global latent variables. The key force transmission path from the loading point to the installation point of the integral die-cast floor is extracted from the current CAE file, and enhanced feature vectors are generated based on the key force transmission path and the node-level feature matrix using prior-guided attention. Through the above method, after constructing the total loss function, the physics mapping unit for the integral die-cast floor is trained, and the mapped global latent variables are temporally encoded. After obtaining the encoded global latent variables, they are broadcast to each node of the integral die-cast floor, concatenated with node coordinates and wall thickness, and subjected to two layers of graph convolution to obtain the node-level feature matrix. Finally, enhanced feature vectors are generated through prior-guided attention, thereby effectively improving the accuracy of generating enhanced feature vectors.

[0077] The physical field prediction system for the structural performance of the integral die-cast floor provided in this application is described below. The physical field prediction system for the structural performance of the integral die-cast floor described below can be referred to in conjunction with the physical field prediction method for the integral die-cast floor described above. Please refer to... Figure 4 , Figure 4 This is a schematic diagram of the module structure of the integrated die-cast floor structure performance physical field prediction system provided in this application embodiment, including: The partitioning module T10 is used to obtain the current CAE file of the integrated die-cast floor, partition the current CAE file into regions based on a preset geometric partitioner, and construct a sub-region map structure based on the region partitioning results.

[0078] The calculation module T20 is used to calculate the low-dimensional latent variables of the physical field based on the encoder and the sub-region graph structure, and to merge the low-dimensional latent variables of the physical field of each region into global latent variables.

[0079] The generation module T30 is used to extract the key force transmission path from the loading point to the installation point of the integral die-cast floor from the current CAE file, and generate an enhanced feature vector based on the key force transmission path and the node-level feature matrix through prior-guided attention.

[0080] The prediction module T40 is used to perform iterative prediction based on the target prediction framework and the enhanced feature vector. When the maximum time step is reached, the predicted global latent variable sequence is projected to different sub-regions, and the structural performance full physical field of the integral die-cast floor is determined according to the physical field of each sub-region.

[0081] This embodiment acquires the current CAE file of the integrated die-cast floor, divides the current CAE file into regions based on a preset geometric partitioner, and constructs a sub-region graph structure based on the region division results; based on the encoder, it calculates the low-dimensional latent variables of the physical field according to the sub-region graph structure, and merges the low-dimensional latent variables of the physical field of each region into global latent variables; it extracts the key force transmission path of the integrated die-cast floor from the loading point to the installation point from the current CAE file, and generates enhanced feature vectors based on the key force transmission path and node-level feature matrix through prior-guided attention; based on the target prediction framework, iterative prediction is performed according to the enhanced feature vectors, and when the maximum time step is reached, the predicted global latent variable sequence is projected to different sub-regions, and the structural performance full physical field of the integrated die-cast floor is determined according to the physical field of each sub-region. By using the above method, instead of directly building a time series prediction model on the high-dimensional node physical field, the complete physical field of each time step is first compressed into low-dimensional global latent variables by the encoder, then iterative prediction is performed based on the target prediction framework, and finally the predicted latent variables are reconstructed into the full physical field by the decoder, so as to reduce the input and output dimensions and thus effectively improve the efficiency and accuracy of the prediction structure performance of the full physical field.

[0082] It is understood that the detailed functional implementation of each of the above modules can be found in the description of the aforementioned method embodiments, and will not be repeated here.

[0083] It should be understood that the above-described device is used to execute the methods in the above embodiments. The implementation principle and technical effect of the corresponding program modules in the device are similar to those described in the above methods. The working process of the device can be referred to the corresponding process in the above methods, and will not be repeated here.

[0084] Based on the methods in the above embodiments, this application provides an electronic device, please refer to... Figure 5 , Figure 5 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application.

[0085] It should be noted that the system may include: a processor 10, a communications interface 20, a memory 30, and a communication bus 40. The processor 10, communications interface 20, and memory 30 communicate with each other via the communication bus 40. The processor 10 can invoke logical instructions stored in the memory 30 to execute the methods described in the above embodiments.

[0086] Furthermore, the logical instructions in the aforementioned memory 30 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0087] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0088] Based on the methods in the above embodiments, this application provides a computer program product that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0089] It is understood that the processor in the embodiments of this application can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, field-programmable gate arrays, or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.

[0090] The method steps in this application embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory, flash memory, read-only memory, programmable read-only memory, erasable programmable read-only memory, electrically erasable programmable read-only memory, registers, hard disks, portable hard disks, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor.

[0091] It is understood that the various numerical designations used in the embodiments of this application are merely for descriptive convenience and are not intended to limit the scope of the embodiments of this application. Those skilled in the art will readily understand that the above descriptions are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for predicting the structural performance physical field of a die-cast integrated floor, characterized in that, include: Obtain the current CAE file of the integrated die-cast floor, divide the current CAE file into regions based on a preset geometric partitioner, and construct a sub-region map structure based on the region division results; Based on the encoder, the low-dimensional latent variables of the physical field are calculated according to the sub-region graph structure, and the low-dimensional latent variables of the physical field in each region are merged into global latent variables. Extract the key force transmission path from the loading point to the installation point of the integral die-cast floor from the current CAE file, and generate an enhanced feature vector based on the key force transmission path and the node-level feature matrix through prior-guided attention. Based on the target prediction framework, iterative prediction is performed according to the enhanced feature vector, and when the maximum time step is reached, the predicted global latent variable sequence is projected to different sub-regions. The structural performance full physical field of the integral die-cast floor is determined according to the physical field of each sub-region.

2. The method for predicting the physical field of structural performance as described in claim 1, characterized in that, The steps of obtaining the current CAE file of the integral die-cast floor, dividing the current CAE file into regions based on a preset geometric partitioner, and constructing a sub-region map structure based on the region division results include: Obtain the current CAE file of the integrated die-cast rear floor, and select the finite element model of the rear floor corresponding to the center point of the design space based on the preset geometric partitioner and the current CAE file. The region is divided by assigning labels to each element in the finite element model of the rear floor to its respective sub-region, and the sub-region envelope is calculated based on the node coordinates of each element after the labels are assigned. Calculate the symbolic distance from the centroid coordinates of each unit to the envelope of the sub-region, and assign regions to each unit based on the symbolic distance; Obtain the node set of the sub-region after the assignment unit, and determine the internal edge set and adjacency matrix based on the node set. Construct the sub-region graph structure based on the node set, the internal edge set, and the adjacency matrix.

3. The method for predicting the physical field of structural performance as described in claim 1, characterized in that, The step of calculating low-dimensional latent variables of the physical field based on the encoder and the sub-region graph structure, and fusing the low-dimensional latent variables of the physical field of each region into global latent variables, includes: Obtain the calculation result file based on the current CAE file of the integral die-cast floor, and calculate the physical feature matrix of the nodes in the sub-region at the current time based on the calculation result file. Based on the encoder, low-dimensional latent variables of the physical field are calculated according to the physical feature matrix and the adjacency matrix in the sub-region graph structure; The query vector, key vector, and value vector of each sub-region are calculated based on the learnable projection matrix, and the attention weights are calculated based on the query vector and the key vector. Calculate the context vector for each sub-region based on the attention weights and the value vectors; The context vector is concatenated with the low-dimensional latent variables of the physical field, and the concatenated latent variables are fused into global latent variables based on the region fusion network.

4. The method for predicting the physical field of structural performance as described in claim 1, characterized in that, The step of extracting the key force transmission path from the loading point to the installation point of the integral die-cast floor from the current CAE file, and generating enhanced feature vectors based on the key force transmission path and the node-level feature matrix through prior-guided attention, includes: The total loss function is constructed based on the mean square error of the nodal physics field reconstruction, the divergence loss, the constraints of the cast aluminum material damage model, the tension-compression asymmetric yield constraint of the cast aluminum material, and the residual constraints of the implicit space momentum equation. The physical field mapping unit of the integral die-cast floor is trained according to the total loss function, and the mapped global latent variables are obtained based on the physical field mapping unit. The node-level feature matrix is ​​determined based on the mapped global latent variables; Extract the key force transmission path from the loading point to the installation point of the integral die-cast floor from the current CAE file, and generate an enhanced feature vector based on the key force transmission path and the node-level feature matrix through prior-guided attention.

5. The method for predicting the physical field of structural performance as described in claim 4, characterized in that, The step of determining the node-level feature matrix based on the mapped global latent variables includes: Obtain the global latent variable sequence within the historical time window, and perform position encoding on the mapped global latent variables at each time step based on the global latent variable sequence within the historical time window and the position encoding information using sine and cosine functions; The encoded global latent variables at each time step are broadcast to each node of the integral die-cast floor, and the node coordinates and wall thickness are obtained respectively. Based on multi-layer graph convolution, the global latent variables at the nodes, the node coordinates, and the wall thickness are concatenated into a node-level feature matrix.

6. The method for predicting the physical field of structural performance as described in any one of claims 1 to 5, characterized in that, The steps of iteratively predicting based on the enhanced feature vector using the target prediction framework, and projecting the predicted global latent variable sequence to different sub-regions when the maximum time step is reached, and determining the full physical field of the structural performance of the integral die-cast floor based on the physical field of each sub-region, include: The path enhancement feature sequence and the original global latent variable sequence are concatenated for each time step within the historical time window; Based on the target prediction framework, iterative prediction is performed according to the enhanced feature vector and the concatenation result; When the maximum time step is reached, the predicted global latent variable sequence is projected to different sub-regions; The global latent variables of the different sub-regions are input into the decoder so that the decoder outputs the physical field of the sub-region; The physical fields of each of the sub-regions are spliced ​​together to determine the full physical field of the structural performance of the integral die-cast floor.

7. A physical field prediction system for the structural performance of an integral die-cast floor, characterized in that, include: The partitioning module is used to obtain the current CAE file of the integrated die-cast floor, partition the current CAE file into regions based on a preset geometric partitioner, and construct a sub-region map structure based on the region partitioning results. The calculation module is used to calculate the low-dimensional latent variables of the physical field based on the encoder and the sub-region graph structure, and to merge the low-dimensional latent variables of the physical field of each region into global latent variables. The generation module is used to extract the key force transmission path from the loading point to the installation point of the integral die-cast floor from the current CAE file, and generate an enhanced feature vector based on the key force transmission path and the node-level feature matrix through prior-guided attention. The prediction module is used to perform iterative prediction based on the target prediction framework and the enhanced feature vector. When the maximum time step is reached, the predicted global latent variable sequence is projected to different sub-regions, and the structural performance full physical field of the integral die-cast floor is determined according to the physical field of each sub-region.

8. An electronic device, characterized in that, include: At least one memory for storing computer programs; At least one processor is configured to execute a program stored in the memory, wherein when the program stored in the memory is executed, the processor is configured to perform the method as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is run on the processor, it causes the processor to perform the method as described in any one of claims 1-6.

10. A computer program product, characterized in that, When the computer program product is run on a processor, the processor causes the processor to perform the method as described in any one of claims 1-6.

Images