A method for predicting a displacement field of a vehicle body covering

Abstract

The application relates to a displacement field prediction method of a vehicle body covering, and relates to the technical field of vehicle development. The method comprises the following steps: acquiring node coordinates of multiple nodes of a vehicle body covering to be analyzed; performing normalization processing on the node coordinates; inputting the normalized node coordinates into a pre-trained displacement field prediction model; acquiring displacement values of each node at multiple time steps output by the displacement field prediction model; and constructing a displacement field of the vehicle body covering according to the displacement values of each node at each time step. The application can realize rapid acquisition of the displacement field of the vehicle body covering while ensuring prediction accuracy, and effectively improves the efficiency and practicability of structure analysis in the process of multi-vehicle parallel development.

Description

Technical Field

[0001] This application relates to the field of vehicle development technology, and in particular to a method for predicting the displacement field of a vehicle body panel. Background Technology

[0002] Currently, with the continuous improvement of regulations on pedestrian protection for automobiles, body panels such as engine hoods require extensive structural safety analysis during the design phase to assess their displacement response and energy absorption performance under impact conditions. In existing engineering practice, explicit dynamic finite element methods are typically used to simulate the structure of body panels. By establishing a fine mesh model and applying load time histories, the displacement field and related performance indicators of the structure at different time steps are obtained. However, this method is computationally complex, requires high mesh quality and time step size, and has a long simulation time, making it difficult to meet the needs of parallel development of multiple vehicle models and multiple schemes.

[0003] To improve computational efficiency, some solutions introduce machine learning-based surrogate models that predict key performance indicators by inputting structural and operational parameters. However, these methods rely on manually extracted geometric parameters, making it difficult to fully represent the complex features of 3D structures. Furthermore, most only output a small number of scalar results, failing to directly obtain spatially distributed displacement information across the entire field. Simultaneously, significant differences exist between different vehicle models, and existing models generally lack cross-model unified modeling capabilities, making it difficult to generalize and apply them across different structures. Summary of the Invention

[0004] Therefore, it is necessary to provide a method for predicting the displacement field of a vehicle body panel to address the aforementioned technical problems. The method includes: Obtain the node coordinates of multiple nodes of the vehicle body panel to be analyzed; The node coordinates are normalized. Input the normalized node coordinates into the pre-trained displacement field prediction model; Obtain the displacement values ​​of each node at multiple time steps output by the displacement field prediction model; The displacement field of the vehicle body panel is constructed based on the displacement values ​​of each node at each time step.

[0005] As an optional implementation, obtaining the node coordinates of multiple nodes of the vehicle body panel to be analyzed includes: Node sampling is performed on the vehicle body panel to be analyzed to obtain multiple nodes; The three-dimensional coordinates of the node are used as the node coordinates.

[0006] As an optional implementation, the normalization process for the node coordinates includes: Calculate the unified geometric center and scale factor corresponding to the coordinates of each node; The node coordinates are translated according to the unified geometric center; The translated node coordinates are scaled according to the scale factor to obtain the normalized node coordinates.

[0007] As an optional implementation, the training method for the displacement field prediction model includes: Obtain finite element simulation results of body panels of multiple vehicle models under preset power conditions; Based on the finite element simulation results, a training sample is constructed, which includes normalized node coordinates and corresponding node displacement time series. The training samples from multiple vehicle models are input into a deep operator network for joint training to obtain the displacement field prediction model.

[0008] As an optional implementation, constructing training samples based on the finite element simulation results includes: Node sampling is performed on each of the aforementioned vehicle body panels to obtain the node coordinates corresponding to a preset number of nodes; Extract the triaxial displacement data of each node at multiple time steps from the finite element simulation results; Calculate the displacement scalar of each node at each time step based on the triaxial displacement data; For each node, the node coordinates and the corresponding displacement scalars are arranged according to the node dimension and the time dimension to form the node displacement time series.

[0009] As an optional implementation, extracting the triaxial displacement data of each node at multiple time steps from the finite element simulation results includes: For each of the aforementioned vehicle models' body panels, extract multiple geometric shape samples of the body panel. According to a preset number of time steps, the structural response of each node is discretized within a preset time interval to obtain multiple time steps; For each of the geometric shape samples, obtain the triaxial displacement vectors of the multiple nodes at each of the time steps.

[0010] As an optional implementation, the formula for calculating the displacement scalar of each node at each time step based on the triaxial displacement data is as follows: ; in, u main ( i,t j ) is a nodei At time step t j The displacement scalar on, u( i,t j ) is a node i At time step t j The displacement vector on.

[0011] As an optional implementation, the method further includes: Based on the displacement scalars of the nodes at each time step, the principal displacement matrix of the geometric sample is constructed, and the principal displacement matrix is ​​the supervision output of the depth operator network.

[0012] As an optional implementation, the deep operator network includes a geometric coding network, a spatial coding network, an intermediate fusion module, a time coefficient generation network, and a field decoding module. The step of inputting training samples from multiple vehicle models into the deep operator network for joint training to obtain the displacement field prediction model includes: The normalized node coordinates are input into the geometric coding network to obtain global geometric features; The normalized coordinates of each node are input into the spatial coding network to obtain the spatial features corresponding to each node. Based on the global geometric features and the spatial features, the intermediate features are processed by the intermediate fusion module to obtain intermediate features, which are then statistically summarized into a global response vector. Based on the global response vector, the time coefficient generation network processes the data to obtain time coefficient vectors corresponding to multiple time steps. The intermediate features are mapped into a spatial coefficient vector by the field decoding module; Based on the spatial coefficient vector and the time coefficient vector, the predicted displacement value of each node at each time step is determined; Based on the predicted displacement values ​​of each node at each time step and the node displacement time series, the loss function value is determined; The parameters of the deep operator network are updated based on the loss function value.

[0013] As an optional implementation, constructing the displacement field of the vehicle body panel based on the displacement values ​​of each node at each time step includes: The displacement values ​​of each node at each time step are arranged according to the node dimension and the time dimension to obtain the displacement field matrix; The displacement field matrix is ​​mapped back to the node positions of the vehicle body panel to be analyzed, thereby constructing the displacement field of the vehicle body panel at multiple time steps.

[0014] This application provides a displacement field prediction method for a vehicle body panel. The method includes: obtaining the node coordinates of multiple nodes of the vehicle body panel to be analyzed; normalizing the node coordinates; inputting the normalized node coordinates into a pre-trained displacement field prediction model; obtaining the displacement values ​​of each node at multiple time steps output by the displacement field prediction model; and constructing the displacement field of the vehicle body panel based on the displacement values ​​of each node at each time step.

[0015] The technical solutions provided by the embodiments of this application bring at least the following beneficial effects: First, by acquiring the node coordinates of multiple nodes of the body panel to be analyzed and normalizing these coordinates, the geometric information of different structural dimensions and vehicle models is uniformly mapped to the same scale space. This eliminates the geometric scale differences between different vehicle models, avoids the problem of needing to model different vehicle models separately in existing technologies, and improves the applicability of the model. Second, the normalized node coordinates are directly input into a pre-trained displacement field prediction model, enabling the model to establish a mapping relationship between node coordinates and structural response based on the spatial distribution of nodes. Compared to existing finite element methods that require explicit mesh model construction and stepwise time integration calculations, this application can obtain the prediction result through a single forward calculation, thereby significantly reducing computational complexity and time. Furthermore, by acquiring the displacement values ​​of each node at multiple time steps output by the displacement field prediction model and constructing the displacement field of the body panel based on these displacement values, a holistic response characterization of the structure in both time and spatial dimensions is achieved. Compared to existing surrogate models that can only output a single or limited number of scalar indicators, this application can directly obtain the two-dimensional displacement distribution of nodes and time, providing more complete basic data for structural deformation analysis, performance evaluation, and structural optimization. Therefore, the embodiments of this application can achieve rapid acquisition of the displacement field of body panels while ensuring prediction accuracy, effectively improving the efficiency and practicality of structural analysis in the parallel development of multiple vehicle models.

[0016] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1A flowchart illustrating a displacement field prediction method for a vehicle body panel provided in this application embodiment; Figure 2 This is a schematic diagram of the geometric model of the hair cover provided in an embodiment of this application; Figure 3 A flowchart illustrating an example of a displacement field prediction method for a vehicle body panel provided in an embodiment of this application. Detailed Implementation

[0019] 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.

[0020] The displacement field prediction method for a vehicle body panel provided in this application will be described in detail below with reference to specific implementation methods. Figure 1 A flowchart illustrating a displacement field prediction method for a vehicle body panel provided in this application embodiment is shown below. Figure 1 As shown, the specific steps are as follows: S101, Obtain the node coordinates of multiple nodes of the vehicle body panel to be analyzed.

[0021] In practice, the node coordinates of multiple nodes can be extracted from the CAD model, finite element mesh model or pre-sampled representative node set corresponding to the body panel to be analyzed, and the node coordinates of each node can be organized in the order of node number to form a node coordinate set.

[0022] As an optional implementation, the specific method for obtaining the node coordinates of multiple nodes of the vehicle body panel to be analyzed in S101 is as follows: Node sampling is performed on the body panel to be analyzed to obtain multiple nodes, and the three-dimensional coordinates of the nodes are used as node coordinates.

[0023] In implementation, based on the finite element mesh model or geometric model of the body panel to be analyzed, node sampling can be performed on its surface or overall structure to obtain multiple nodes that can represent the structural geometric features. For example, according to preset sampling rules, a preset number of nodes can be selected from the mesh nodes, so that the selected nodes spatially cover key structural locations such as the rear end area of ​​the hood, the reinforcing rib area, and the front end area of ​​the hood, while avoiding nodes being too concentrated or unevenly distributed. Subsequently, the coordinate values ​​of the selected nodes in three-dimensional space are recorded and organized as node coordinates. The embodiments of this application can retain the main geometric features of the body panel while reducing the number of nodes, thereby reducing the data scale of subsequent model calculations.

[0024] S102, normalize the node coordinates.

[0025] In implementation, the node coordinates can be standardized by numerical scaling to map the node coordinates corresponding to body panels of different models and sizes to a preset range, resulting in normalized node coordinates. For example, node coordinates can be translated and scaled based on a preset reference coordinate system; each dimension of the coordinates can be normalized based on the statistical distribution of the node coordinates; or the node coordinates can be scaled according to a preset proportional relationship. By normalizing the node coordinates, the impact of differences in absolute size, installation position, and spatial distribution of body panels of different models on the model input can be reduced, allowing the geometric information of different models to be expressed in a relatively consistent coordinate space. This facilitates the displacement field prediction model in learning a more stable correspondence between node coordinates and displacement response, and improves the model's applicability and generalization ability across different models.

[0026] As an optional implementation, the specific steps for normalizing the node coordinates in S102 are as follows: S1021, calculate the unified geometric center and scale factor corresponding to the coordinates of each node.

[0027] In implementation, the coordinate data of all nodes in multiple vehicle model samples during the training phase can be pre-collected, and a unified geometric center and scale factor can be determined based on the coordinate data. The unified geometric center represents the central position of the overall distribution of coordinates across multiple vehicle model nodes, while the scale factor represents the scale range of this distribution. For example, the formula for calculating the global geometric center c is as follows: ; in, c It is a three-dimensional vector representing the coordinates of the geometric center. c x Are all nodes in x Average position in the direction, c y Are all nodes in y Average position in the direction, c z Are all nodes in z Average position in the direction. Assuming there are N nodes, the node coordinates are as follows: ( x 1 ,y 1 ,z 1), ( x 2 ,y 2 ,z 2), ( x i ,y i ,z i )…(x N ,y N ,z N ); but: ; ; .

[0028] There are various ways to determine the scale factor, and this application does not limit the specific method used in the embodiments. For example, the maximum absolute value of all node coordinates can be taken, ensuring that the normalized value falls between -1 and 1. Alternatively, the scale factor can be determined by calculating the bounding box size. s=max ( x max -x min , y may -y min , z maz -z min The maximum distance from all node coordinates to the geometric center can also be used as a scale factor.

[0029] S1022, translate the node coordinates according to the unified geometric center.

[0030] In practice, the original coordinates of each node in the body panel to be analyzed can be calculated by performing coordinate difference operations with the geometric center. This allows the original node coordinates to be translated as a whole to a position relative to the unified geometric center. For example, for each node, the coordinate components corresponding to the unified geometric center can be subtracted in each coordinate dimension, so that the translated node coordinates reflect the spatial position of the node relative to the unified geometric center. For example, nodes... i The coordinates are ( x i ,y i ,z i The coordinates after translation are ( ). x i -c , y i -c , z i -c ).

[0031] S1023, the translated node coordinates are scaled according to the scale factor to obtain the normalized node coordinates.

[0032] In implementation, the coordinates of each node after translation can be divided by a scaling factor to uniformly compress or stretch the range of node coordinate values, resulting in normalized node coordinates. For example, when body panels of different car models differ in length, width, or overall dimensions, a scaling factor can be used to uniformly scale the translated node coordinates, mapping the node coordinates of different car models to a similar numerical range. Scaling by the scaling factor further reduces the impact of size differences between different car models on the model input, making it easier for the displacement field prediction model to learn a stable mapping relationship between node coordinates and displacement response during training and inference, thereby improving the model's prediction stability and cross-car model generalization ability. (Example: Scaled node coordinates) i The coordinates are .

[0033] Finally, the normalized N nodes can be arranged into a matrix by rows, as shown below: .

[0034] S103 inputs the normalized node coordinates into the pre-trained displacement field prediction model.

[0035] In implementation, the normalized nodal coordinates can be input as a matrix into a pre-trained displacement field prediction model, which then outputs the time-series response corresponding to that set of nodal coordinates. This embodiment differs from the traditional explicit dynamic finite element method's time-step iterative solution; instead, it establishes a mapping relationship between nodal coordinates and the displacement field through a pre-trained model. Therefore, during the inference phase, only the normalized nodal coordinates need to be input, allowing the model to directly infer the spatiotemporal response of the structure under preset conditions, significantly simplifying the prediction process.

[0036] As an optional implementation method, the specific steps of the training method for the displacement field prediction model in S103 are as follows: S1031, obtain the finite element simulation results of the body panels of multiple vehicle models under preset power conditions.

[0037] In practice, finite element simulation calculations can be performed on the body panels of multiple vehicle models under uniform or standardized dynamic conditions to obtain the displacement response results of each node at multiple time steps.

[0038] S1032, based on the finite element simulation results, a training sample is constructed, which includes normalized node coordinates and the corresponding node displacement time series.

[0039] In implementation, node sampling can be performed on the body panels of each vehicle model to obtain a predetermined number of representative nodes, and the three-dimensional coordinates of these nodes can be extracted. Subsequently, the node coordinates are normalized to obtain normalized node coordinates. Simultaneously, displacement data of the corresponding nodes at multiple time steps are extracted from the finite element simulation results, and a node displacement time series is formed based on this data. Finally, the normalized node coordinates are mapped one-to-one with the corresponding node displacement time series to construct training samples. Through this method, the geometric information of the structure and its temporal response information can be uniformly expressed, allowing the training samples to simultaneously contain spatial distribution characteristics and temporal evolution characteristics, thus enabling the model to learn the mapping relationship between node coordinates and the displacement field.

[0040] Optionally, finite element simulation samples from various vehicle models can be used to form a training set consisting of 90% training data and 10% validation data. Each sample consists of a normalized nodal coordinate matrix and its corresponding principal displacement matrix.

[0041] As an optional implementation, the specific steps for constructing training samples based on finite element simulation results in S1032 are as follows: S10321, perform node sampling on each body panel to obtain the node coordinates corresponding to a preset number of nodes.

[0042] In implementation, based on the finite element mesh model corresponding to the body panels of each vehicle model, nodes can be sampled to obtain a preset number of representative nodes, and the three-dimensional coordinates of these representative nodes can be extracted as node coordinates. For example, nodes with a relatively uniform distribution can be selected according to preset sampling rules so that the sampling results can characterize the overall geometric contour and local structural features of the body panels. By sampling nodes according to uniform rules for each body panel, the main geometric information can be preserved while reducing the data size. Furthermore, the input data for different vehicle models can maintain consistency in the number and organization of nodes, thus facilitating the subsequent construction of unified training samples.

[0043] S10322 extracts the triaxial displacement data of each node at multiple time steps from the finite element simulation results.

[0044] In implementation, based on the representative nodes obtained in step S10321, triaxial displacement data of each node at multiple time steps can be extracted from the corresponding finite element simulation results. The triaxial displacement data includes displacement components of each node in the x, y, and z directions. For example, the triaxial displacement response of each representative node can be read sequentially throughout the entire simulation time interval to obtain the changes in displacement components at different times. By extracting the triaxial displacement data of nodes at multiple time steps, the evolution process of the structural response in the time dimension and its distribution information in the spatial dimension can be fully preserved.

[0045] As an optional implementation, the specific steps in S10322 for extracting the triaxial displacement data of each node at multiple time steps from the finite element simulation results are as follows: For each vehicle model's body panels, extract multiple geometric shape samples of the body panels; According to the preset time discretization step number, the structural response of each node is discretized within the preset time interval to obtain multiple time steps; For each geometric sample, obtain the three-dimensional displacement vectors of multiple nodes at each time step.

[0046] In implementation, multiple geometric shape samples can be pre-constructed for the body panels of various vehicle models, and finite element simulations can be performed on each sample under the same dynamic conditions. Subsequently, the simulation time interval is divided into multiple time steps according to a preset time discretization number, and the structural response results of each node are read at each time step, thereby obtaining the triaxial displacement vectors of multiple nodes in each geometric shape sample at each time step. For example, for the engine hood of the same vehicle model, multiple geometric shape samples can be formed by changing the local stiffener structure, plate thickness distribution, or shape parameters, and each geometric shape sample can be simulated under a uniform impact condition. After the simulation results are output, the x-direction displacement components, y-direction displacement components, and z-direction displacement components of each representative node at each time step are extracted according to a preset time sampling frequency to form the triaxial displacement vector sequence of the corresponding node. For example, simulation analysis can be performed on the engine hoods of different vehicle models under the same impact condition to obtain data on the corresponding node displacement changes over time, and the simulation results of each vehicle model under different structural schemes or different conditions can be used as training data. For example: Figure 2 This is a schematic diagram of the geometric model of the hair cover provided in the embodiments of this application, such as... Figure 2 As shown, explicit dynamic finite element simulations were performed on the engine hoods or other coverings of three representative vehicle models (model A, model B, and model C) under uniform or standardized power conditions. Each geometric sample is denoted as a case. The geometry can be obtained through parameterization. As shown in the figure, four key geometric features were extracted: the rear z-axis height of the hood (210), the rib width (220), the rib height (230), and the z-axis height of the hood locking groove (240). This resulted in 100 cases for each vehicle model. For each case, the total number of nodes after sampling is denoted as N, and the time interval step number is denoted as T. i The three-dimensional coordinates of the nodes are ( x i , y i , z i All node coordinates are arranged into a matrix by rows. During the simulation, the structural response is discretized over the time interval. t 1, t 2,… t T At each time step, output the three-dimensional displacement vector u of the node. i , t j ).

[0047] As an optional implementation, the formula for calculating the displacement scalar of each node at each time step based on the triaxial displacement data is as follows: ; in, u main ( i,t j ) is a node i At time step t j The displacement scalar on, u( i,t j ) is a node i At time step t j The displacement vector on.

[0048] As an alternative implementation, the principal displacement matrix of the geometric sample can also be constructed based on the displacement scalar of the node at each time step. The principal displacement matrix is ​​the supervision output of the depth operator network.

[0049] In implementation, the displacement scalars of each node at multiple time steps can be arranged according to the node dimension and the time dimension to construct the principal displacement matrix corresponding to the geometric sample. The rows of the matrix correspond to different nodes, and the columns correspond to different time steps. For example, for each geometric sample, the displacement scalars of the 1st to Nth nodes at time steps 1 to T can be arranged sequentially to form a principal displacement matrix of dimension N×T. By uniformly organizing the displacement scalars of nodes at each time step into the principal displacement matrix, the distribution relationship of node responses in both spatial and temporal dimensions can be expressed in a structured form. Furthermore, it can serve as the supervisory output of the deep operator network, enabling the model to learn the mapping relationship between node coordinates and corresponding temporal displacement responses during training, thereby achieving the prediction of the entire field displacement.

[0050] S10323, calculate the displacement scalar of each node at each time step based on the triaxial displacement data.

[0051] In implementation, the three-directional displacement components of each node at each time step can be synthesized to obtain the displacement scalar of that node at the corresponding time step. For example, the magnitude of the combined displacement of each node can be determined based on its three-directional displacement components, and this combined displacement magnitude can be used as the displacement scalar of that node at the corresponding time step. By converting the three-directional displacement data into displacement scalars, the complexity of the output data organization can be reduced while preserving the overall deformation amplitude information of the nodes, making the model easier to learn the correspondence between the spatial position of the nodes and the strength of the structural response.

[0052] S10324, For each node, arrange the node coordinates and corresponding displacement scalars of the node according to the node dimension and time dimension to form the node displacement time series of the node.

[0053] In implementation, the node coordinates of the sampled nodes can be arranged according to node order, and the displacement scalars of the corresponding nodes at multiple time steps can be arranged in chronological order, thereby establishing a one-to-one correspondence between node coordinates and displacement responses, forming a node displacement time series. For example, the displacement scalars of each node at each time step can be used to form the time series response data of that node, and then the time series response data of all nodes can be organized uniformly to form the output samples for model training. By uniformly arranging the node coordinates and corresponding displacement scalars according to the node dimension and the time dimension, the training samples can simultaneously contain geometric input information and time series response information, thereby supporting deep operator networks to learn the mapping relationship from node coordinates to node displacement time series.

[0054] S1033 inputs training samples from multiple vehicle models into a deep operator network for joint training to obtain a displacement field prediction model.

[0055] In implementation, training samples from multiple vehicle models can be uniformly input into a deep operator network. The network learns the correspondence between node coordinates and node displacement time series, and iteratively updates the network parameters based on the input-output relationship in the training samples until the model converges, resulting in a displacement field prediction model. For example, during training, the network parameters can be continuously adjusted to make the node displacement time series output by the model gradually approximate the corresponding finite element simulation results. By using training samples from multiple vehicle models for joint training, the model can simultaneously learn the common and difference features between different vehicle models in a unified parameter space, thereby achieving unified modeling of body panels of different vehicle models and improving the model's predictive ability on new or similar structures.

[0056] As an optional implementation, the deep operator network includes a geometric coding network, a spatial coding network, an intermediate fusion module, a time coefficient generation network, and a field decoding module. The specific steps in S1033 to input training samples from multiple vehicle models into the deep operator network for joint training to obtain the displacement field prediction model are as follows: S10331, input the normalized node coordinates into the geometric coding network to obtain global geometric features.

[0057] In implementation, the normalized coordinates of all nodes can be input as a whole into the geometry encoding network. Feature extraction and aggregation are then performed on the node set to obtain global geometric features representing the overall geometric shape of the vehicle body panel. For example, a fixed-length global feature vector can be obtained by mapping all node coordinates layer by layer and summarizing them along the node dimension. This vector describes the overall contour and geometric distribution characteristics of the structure. By extracting global geometric features, subsequent calculations can not only rely on the local information of individual nodes but also reflect the influence of the overall structural shape on local responses. For example, the geometry encoding network takes the entire set of normalized node coordinates as input and uses a neural network structure suitable for point sets to extract and aggregate features from the node set, outputting a fixed-length global geometric vector g. ; Where g is the global geometric vector. E geom For geometric coding networks, d g The dimension of the global features.

[0058] Optionally, the geometric coding network can be a multilayer feedforward network, a point cloud network, a graph neural network, or other forms.

[0059] S10332, input the normalized coordinates of each node into the spatial coding network to obtain the spatial features corresponding to each node.

[0060] In implementation, the normalized coordinates of each node can be input into the spatial coding network to perform feature mapping on the node's position in three-dimensional space, obtaining the corresponding spatial features. For example, a nonlinear transformation can be applied to the coordinates of each node to map them into a high-dimensional feature representation, thereby enhancing the ability to express complex geometric details. By extracting spatial features from each node separately, feature information reflecting the local spatial position of the node and its relative relationships can be obtained, providing a foundation for subsequent displacement prediction. For example, the spatial coding network takes the coordinates of a single normalized node as input and uses a nonlinear feedforward structure to perform high-dimensional embedding of the node's position in three-dimensional space, obtaining a point-by-point spatial feature vector.

[0061] ; in, For the first i Spatial characteristics of each node E space For spatial coding networks, x i For the first i The coordinates of each node, d s Let N be the dimension of the spatial feature and N be the number of nodes.

[0062] Arranging all nodes row-wise yields the spatial characteristic matrix: ; in, The spatial feature matrix of all nodes, For the first i Spatial characteristics of each node (transposed into a single row).

[0063] Optionally, spatial coding networks can employ structures such as multilayer feedforward networks with periodic activation functions or position coding networks to enhance the ability to represent high-frequency geometric details.

[0064] S10333, based on global geometric features and spatial features, is processed by an intermediate fusion module to obtain intermediate features, which are then statistically summarized into a global response vector.

[0065] In implementation, global geometric features can be fused with the spatial features corresponding to each node to generate intermediate features that simultaneously contain both overall structural information and local spatial information. For example, global geometric features can be used to modulate the spatial features of each node, allowing global morphological information corresponding to different structures to influence the feature representation of each node. Through this fusion process, the correlation between the overall structural morphology and the responses of local nodes can be established, thereby improving the model's ability to express responses to complex structures. For example, the intermediate fusion module is used to fuse the global geometric vector g with pointwise spatial features to form intermediate features that combine local spatial information and overall geometric information. ; in, h i For the first i The intermediate features of each node, F fusion The fusion function for the intermediate fusion module can include operations such as channel scaling, offsetting, gating, or residual joining. d h The dimension of the intermediate features.

[0066] Subsequently, the intermediate features of all nodes are statistically summarized along the node dimension (e.g., averaged) to obtain the global response vector: ; Where r is the global response vector. h i For the first i The intermediate features of each node. The global response vector can characterize the overall response pattern of the structure under standard operating conditions, providing input for subsequent time coefficient generation.

[0067] S10334, based on the global response vector, is processed by a time coefficient generation network to obtain time coefficient vectors corresponding to multiple time steps.

[0068] In implementation, coefficient values ​​corresponding to each time step can be obtained based on the overall structural response characteristics represented by intermediate features, which are used to characterize the structural response changes over time. By introducing time coefficients, the temporal evolution characteristics of the structural response can be modeled without explicitly inputting the load time history, thereby simplifying the model input format and improving prediction efficiency. For example, a time coefficient generation network takes the global response vector as input and outputs a time coefficient matrix through several layers of linear / nonlinear transformations. ; Where C is the time coefficient matrix and T is the number of time steps. F time For time coefficient generation networks, r is the global response vector. d c Let C be the time coefficient channel dimension. The t-th row of C corresponds to the time coefficient vector α at the t-th time step. t : ; The time coefficient generation network can implicitly learn the time response pattern of a structure under standard operating conditions without explicitly inputting the load time history or additional operating condition parameters during the inference phase.

[0069] S10335 maps intermediate features into spatial coefficient vectors through the field decoding module.

[0070] In implementation, time-step calculations can be performed based on the time coefficients and the features corresponding to the nodes to obtain the displacement response of the nodes throughout the entire time interval. Through this process, spatial and temporal features can be jointly modeled to predict the displacement of nodes in both time and space dimensions. For example, intermediate features can first be mapped to a spatial coefficient vector for field decoding, in the following form: ; in, Let k be the predicted displacement of the i-th node at time step t, and k be the feature channel index. For time coefficient, This is the spatial coefficient vector.

[0071] S10336, based on the spatial coefficient vector and the time coefficient vector, determines the predicted displacement value of each node at each time step.

[0072] In practice, at each time step t, the predicted principal displacement of a node can be obtained by integrating the time coefficient vector and the node spatial coefficient vector. By arranging the predicted values ​​of all nodes and all time steps in matrix form, the predicted displacement field matrix can be obtained: ; This enables operator mapping from normalized nodal coordinates to the full-time principal displacement field.

[0073] S10336, Based on the predicted displacement values ​​of each node at each time step and the node displacement time series, determine the value of the loss function.

[0074] In implementation, the displacement values ​​of each node at each time step predicted by the model can be compared with the corresponding node displacement time series, and the loss function value can be calculated based on the difference between the two. For example, the differences between the predicted results and the actual results of each node at each time step can be summarized to obtain the loss function value reflecting the prediction error. By constructing the loss function and measuring the difference between the prediction results and the finite element simulation results, a basis can be provided for subsequent model parameter updates, allowing the model to gradually approximate the actual structural response, thereby improving the accuracy of displacement field prediction.

[0075] S10337, Update the parameters of the deep operator network based on the loss function value.

[0076] In implementation, the gradients of each parameter of the deep operator network with respect to the loss function value can be calculated using backpropagation based on the loss function value. The network parameters are then iteratively updated based on these gradients to reduce the discrepancy between the model output and the node displacement time series. For example, in each training round, the predicted displacement values ​​of each node at each time step can be obtained first based on the current network parameters. Then, the loss function value is determined based on the predicted displacement values ​​and the corresponding node displacement time series. Subsequently, the parameters in the geometric coding network, spatial coding network, intermediate fusion module, time coefficient generation network, and field decoding module are updated based on the loss function value until the loss function value meets the preset convergence condition or reaches the preset training round. Through this method, the deep operator network can continuously adjust its parameter values ​​during training, allowing the model to gradually learn the mapping relationship between the normalized node coordinates and the node displacement time series, thereby improving the prediction accuracy of the obtained displacement field prediction model for the displacement response of the vehicle body panels.

[0077] S104: Obtain the displacement values ​​of each node at multiple time steps output by the displacement field prediction model.

[0078] In implementation, the displacement field prediction model outputs displacement values ​​of each node at multiple discrete time steps (such as a displacement scalar sequence), or node-time two-dimensional displacement data formed after model decoding. For each node, its displacement value changing over time can be obtained. For example, the displacement results from time step 1 to time step T can be output sequentially according to node order, thereby forming a time-series displacement prediction result corresponding one-to-one with the input nodes. The displacement field prediction model provided in this application does not only output a single scalar index such as the maximum depression value or peak displacement, but directly outputs the response results of all nodes at multiple time steps, thus simultaneously preserving spatial distribution information and temporal evolution information.

[0079] S105, construct the displacement field of the vehicle body panel based on the displacement values ​​of each node at each time step.

[0080] In implementation, the displacement values ​​of each node at each time step can be arranged according to the node dimension and the time dimension to form a displacement field matrix. This displacement field matrix is ​​then mapped back to the node positions corresponding to the body panel to be analyzed, obtaining the displacement field of the body panel at multiple time steps. Optionally, it can also be restored to a finer grid through interpolation to generate displacement cloud maps or time-series dynamic display results. For example, a time-series displacement matrix can be first formed for the sampling nodes on the engine hood surface, and then the displacement value at each moment in the matrix can be backfilled to the corresponding node coordinate position to obtain the overall displacement distribution at the corresponding moment. The principle is that the displacement value of a single node only represents the discrete point response, while reconstructing the displacement results of all nodes at all time steps according to their spatial positions can form a displacement field characterizing the deformation state of the entire body panel, thereby supporting subsequent pedestrian protection performance evaluation and deformation analysis of key areas.

[0081] As an optional implementation, the specific steps in S105 for constructing the displacement field of the vehicle body panel based on the displacement values ​​of each node at each time step are as follows: S1051, arrange the displacement values ​​of each node at each time step according to the node dimension and the time dimension to obtain the displacement field matrix.

[0082] In implementation, the displacement values ​​of each node at multiple time steps can be organized according to a preset node order and time order to form a corresponding displacement field matrix. Rows in the matrix can correspond to different nodes, and columns can correspond to different time steps. For example, for N nodes sampled from a vehicle body panel to be analyzed, the displacement values ​​of nodes 1 to N at time steps 1 to T can be arranged sequentially to form a displacement field matrix of dimension N×T. By uniformly arranging the displacement values ​​according to the node and time dimensions, discrete node displacement results can be transformed into spatiotemporal response data with a regular organizational form, thus providing a foundation for subsequent displacement field reconstruction and visualization.

[0083] S1052, map the displacement field matrix back to the node positions of the body panel to be analyzed, and construct the displacement field of the body panel at multiple time steps.

[0084] In implementation, the displacement values ​​of each node in the displacement field matrix at each time step can be used to backfill the corresponding node positions of the body panel to be analyzed, thereby constructing the displacement field of the body panel at multiple time steps. For example, for each time step, a correspondence can be established between the displacement values ​​of each node at that time step and the spatial position of each node, thus forming the overall displacement distribution of the body panel at that time step. Repeating the above process for multiple time steps yields the dynamic displacement field of the body panel over the entire preset time interval. By mapping the displacement field matrix back to the node positions, the actual distribution of the node response in the spatial structure can be restored. This allows the prediction results to not only reflect the displacement changes of individual nodes but also characterize the overall deformation of the entire body panel during the time evolution process, facilitating subsequent structural response analysis and performance evaluation.

[0085] Optionally, after generating the predicted displacement field, key evaluation indicators (such as maximum displacement in a specified area, peak displacement at a specific measuring point, etc.) can be extracted from the predicted displacement field according to pre-defined pedestrian protection evaluation standards. This allows for direct determination of whether the structure meets design requirements and guides structural optimization work such as stiffener arrangement and plate thickness distribution. In the above process, the inference stage does not require re-solving the finite element method or inputting explicit load-time curves or material parameters. Only nodal coordinate information is needed to complete the overall response prediction, greatly reducing the time cost of evaluating multiple vehicle models and multiple schemes.

[0086] As an optional implementation method, Figure 3 A flowchart illustrating an example of a displacement field prediction method for vehicle body panels provided in this application embodiment is shown below. Figure 3 As shown, the specific steps are as follows: S301, Multi-vehicle model finite element data acquisition and node sampling: Finite element simulation is performed on the body panels of various vehicle models under a unified dynamic condition, outputting the results of node displacement changes over time, and uniformly sampling the nodes. S302, Principal Displacement Calculation: Calculate the principal displacement of the nodes based on the finite element simulation results; S303, Coordinate Normalization: Perform unified normalization on the 3D node coordinates across multiple vehicle models to obtain a geometric representation at a unified scale. S304, Constructing a coordinate-driven depth operator network: Using normalized node coordinates as input, construct a depth operator model consisting of a geometric coding network, a spatial coding network, an intermediate fusion module, a time coefficient generation module, and a field decoding module; S305, multi-vehicle joint training and validation to obtain displacement field prediction model: deep operator network is jointly trained on multiple vehicle models and multi-working condition samples, and the overall prediction effect and cross-vehicle generalization ability are evaluated on multi-vehicle validation sets. S306, Input the coordinates of the new vehicle model's hood nodes: For any new vehicle model or structural scheme with similar geometric changes, only the node coordinates need to be provided to quickly generate the entire field displacement time series for pedestrian protection and structural optimization.

[0087] It should be understood that, although Figure 1 and Figure 3 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 and Figure 3 At least some of the steps in the process may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but may be executed at different times. The execution order of these steps or stages is not necessarily sequential, but may be executed in turn or alternately with other steps or at least some of the steps or stages in other steps.

[0088] It is understood that the same / similar parts between the various embodiments of the methods described above in this specification can be referred to each other. Each embodiment focuses on the differences from other embodiments, and relevant parts can be referred to the description of other method embodiments.

[0089] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0090] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0091] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0092] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for predicting the displacement field of a vehicle body panel, characterized in that, The method includes: Obtain the node coordinates of multiple nodes of the vehicle body panel to be analyzed; The node coordinates are normalized. Input the normalized node coordinates into the pre-trained displacement field prediction model; Obtain the displacement values ​​of each node at multiple time steps output by the displacement field prediction model; The displacement field of the vehicle body panel is constructed based on the displacement values ​​of each node at each time step.

2. The method according to claim 1, characterized in that, The process of obtaining the node coordinates of multiple nodes of the vehicle body panel to be analyzed includes: Node sampling is performed on the vehicle body panel to be analyzed to obtain multiple nodes; The three-dimensional coordinates of the node are used as the node coordinates.

3. The method according to claim 1, characterized in that, The normalization process for the node coordinates includes: Calculate the unified geometric center and scale factor corresponding to the coordinates of each node; The node coordinates are translated according to the unified geometric center; The translated node coordinates are scaled according to the scale factor to obtain the normalized node coordinates.

4. The method according to claim 1, characterized in that, The training method for the displacement field prediction model includes: Obtain finite element simulation results of body panels of multiple vehicle models under preset power conditions; Based on the finite element simulation results, a training sample is constructed, which includes normalized node coordinates and corresponding node displacement time series. The training samples from multiple vehicle models are input into a deep operator network for joint training to obtain the displacement field prediction model.

5. The method according to claim 4, characterized in that, The construction of training samples based on the finite element simulation results includes: Node sampling is performed on each of the aforementioned vehicle body panels to obtain the node coordinates corresponding to a preset number of nodes; Extract the triaxial displacement data of each node at multiple time steps from the finite element simulation results; Calculate the displacement scalar of each node at each time step based on the triaxial displacement data; For each node, the node coordinates and the corresponding displacement scalars are arranged according to the node dimension and the time dimension to form the node displacement time series.

6. The method according to claim 4, characterized in that, The extraction of triaxial displacement data of each node at multiple time steps from the finite element simulation results includes: For each of the aforementioned vehicle models' body panels, extract multiple geometric shape samples of the body panel. According to a preset number of time steps, the structural response of each node is discretized within a preset time interval to obtain multiple time steps; For each of the geometric shape samples, obtain the triaxial displacement vectors of the multiple nodes at each of the time steps.

7. The method according to claim 5, characterized in that, The formula for calculating the displacement scalar of each node at each time step based on the triaxial displacement data is as follows: ； in, u main ( i,t j ) is a node i At time step t j The displacement scalar on, u( i,t j ) is a node i At time step t j The displacement vector on.

8. The method according to claim 6, characterized in that, The method further includes: Based on the displacement scalars of the nodes at each time step, the principal displacement matrix of the geometric sample is constructed, and the principal displacement matrix is ​​the supervision output of the depth operator network.

9. The method according to claim 4, characterized in that, The deep operator network includes a geometric coding network, a spatial coding network, an intermediate fusion module, a time coefficient generation network, and a field decoding module. The step of inputting training samples from multiple vehicle models into the deep operator network for joint training to obtain the displacement field prediction model includes: The normalized node coordinates are input into the geometric coding network to obtain global geometric features; The normalized coordinates of each node are input into the spatial coding network to obtain the spatial features corresponding to each node. Based on the global geometric features and the spatial features, the intermediate features are processed by the intermediate fusion module to obtain intermediate features, which are then statistically summarized into a global response vector. Based on the global response vector, the time coefficient generation network processes the data to obtain time coefficient vectors corresponding to multiple time steps. The intermediate features are mapped into a spatial coefficient vector by the field decoding module; Based on the spatial coefficient vector and the time coefficient vector, the predicted displacement value of each node at each time step is determined; Based on the predicted displacement values ​​of each node at each time step and the node displacement time series, the loss function value is determined; The parameters of the deep operator network are updated based on the loss function value.

10. The method according to claim 1, characterized in that, The step of constructing the displacement field of the vehicle body panel based on the displacement values ​​of each node at each time step includes: The displacement values ​​of each node at each time step are arranged according to the node dimension and the time dimension to obtain the displacement field matrix; The displacement field matrix is ​​mapped back to the node positions of the vehicle body panel to be analyzed, thereby constructing the displacement field of the vehicle body panel at multiple time steps.

Images