Neural network training data set generation method, device, equipment and storage medium

Abstract

The application relates to the technical field of data sampling, in particular to a neural network training data set generation method and device, equipment and a storage medium, which comprises the following steps: acquiring shape data, material data and a target load position space of a virtual model of a part; determining boundary constraint conditions of the virtual model of the part according to the shape data and the material data; performing layered sampling of load points in the target load position space to obtain a plurality of sampling points and sampling parameters; constructing a simulation working condition according to the boundary constraint conditions and the sampling parameters of the plurality of sampling points, and performing topological structure simulation under the simulation working condition to determine a part topological structure corresponding to the simulation working condition according to a simulation result; generating a training data set based on the simulation working condition and the corresponding model topological structure, training a neural network model by using the training data set, and generating a part topological structure under a target real working condition by using the trained neural network model. Thus, the problems of poor quality of a training data set in related technologies are solved.

Description

Technical Field

[0001] This application relates to the field of data sampling technology, and in particular to a method, apparatus, device and storage medium for generating neural network training datasets. Background Technology

[0002] Topology optimization is a technique that uses mathematical methods to optimize the distribution of materials within a given design space to meet specific constraints. Its core objective is to achieve efficient material utilization while ensuring performance. In recent years, with the development of artificial intelligence (AI), the combination of AI and topology optimization has become a research hotspot, including density-based prediction and data-driven end-to-end prediction. The core of these methods is to utilize GANs (Generative Adversarial Networks), CNNs (Convolutional Neural Networks), and fully connected networks to learn the mapping relationship between "operating conditions and topology." These methods heavily rely on topology optimization results as training datasets and require a huge number of samples; even simple models require thousands to tens of thousands of samples for multiple rounds of training.

[0003] However, existing common methods such as Poisson sampling and random sampling produce datasets of low quality, which are difficult to effectively reflect the mapping relationship between working conditions and model stress and strain. They perform reasonably well in training simple models, but pose a significant challenge when dealing with complex models. Summary of the Invention

[0004] This application provides a method, apparatus, device, and storage medium for generating neural network training datasets to solve problems such as poor quality of training datasets in related technologies.

[0005] The first aspect of this application provides a method for generating a neural network training dataset, comprising the following steps: acquiring shape data, material data, and target load location space of a virtual model of a component; determining boundary constraints of the virtual model of the component based on the shape data and material data; performing layered sampling of load points in the target load location space to obtain multiple sampling points and sampling parameters; constructing a simulated working condition based on the boundary constraints and the sampling parameters of the multiple sampling points; performing topology simulation based on the boundary constraints and the sampling parameters of the multiple sampling points; determining the component topology corresponding to the simulated working condition based on the simulation results, wherein the simulated working condition is used to simulate the real working condition; generating a training dataset based on the simulated working condition and the corresponding model topology; training a neural network model using the training dataset; and generating the component topology under the target real working condition using the trained neural network model.

[0006] Optionally, load point hierarchical sampling is performed in the target load location space, including: obtaining the mesh topology of the virtual model of the component, wherein the mesh topology includes elements and nodes; dividing the target load location space into multiple regions, and randomly sampling the elements in each region to obtain sampling points; randomly obtaining the load direction and applying load force to the sampling points in each region, and generating sampling parameters for the corresponding sampling points based on the coordinates of the sampling points and the corresponding load force and load direction.

[0007] Optionally, before constructing the simulated working condition based on the boundary constraints and the sampling parameters of multiple sampling points, the method further includes: filtering the loading position, structural failure, and topological solvability of the multiple sampling points based on the boundary constraints.

[0008] Optionally, the process of selecting loading locations includes: identifying constraint nodes and constraint degrees of freedom in boundary constraints; constructing a node adjacency graph based on the mesh topology of the virtual component model; identifying non-design domains of the virtual component model; calculating the distance between constraint nodes and sampling points, and removing sampling points if the distance is less than a preset distance; detecting whether the element containing the sampling point is located in a non-design domain, and removing the sampling point if it is; and determining whether the sampling point is in a structurally isolated region based on the node adjacency graph, and removing the sampling point if it is in a structurally isolated region.

[0009] Optionally, the screening process for structural failure includes: obtaining the transmission path from the sampling point to the constraint node based on the mesh topology; traversing each node in the transmission path, finding the multi-layer elements adjacent to each node, and grouping the multi-layer elements into an element set; calculating the equivalent cross-sectional area of ​​the target cylindrical model based on the total volume of the element set and the length of the transmission path, calculating the maximum energy of the virtual model of the component under different types of deformation based on the equivalent cross-sectional area, and obtaining the external force work of the sampling point under different types of deformation; if the external force work is greater than the maximum energy, the sampling point is removed.

[0010] Optionally, the topological solvability screening process is as follows: identify the constraint degrees of freedom of multiple constraint nodes in the boundary constraint conditions, and determine the resultant force constraint direction and resultant moment constraint direction of multiple constraint nodes based on the constraint degrees of freedom; select reference points in the virtual model of the components; calculate the resultant load force based on the load force of the sampling points, and calculate the resultant load moment about the reference point based on the resultant load force; if the direction of the resultant load force does not match the resultant force constraint direction, or the direction of the resultant load moment does not match the resultant moment constraint direction, then remove the sampling points with mismatched directions based on the resultant force constraint direction or the resultant moment constraint direction.

[0011] Optionally, after filtering multiple sampling points based on boundary constraints for loading location, structural failure, and topological solvability, the method further includes: performing diversity and sensitivity assessments on the filtered sampling points; generating diversity weights and sensitivity weights for the corresponding sampling points based on the diversity and sensitivity assessment results; determining target sampling points based on the diversity and sensitivity weights of the sampling points; and constructing simulated operating conditions based on the target sampling points and their corresponding sampling parameters and boundary constraints.

[0012] A second aspect of this application provides a neural network training dataset generation apparatus, comprising: an acquisition module for acquiring shape data, material data, and target load location space of a virtual model of a component; a first determination module for determining boundary constraints of the virtual model of the component based on the shape data and material data; a sampling module for performing layered sampling of load points in the target load location space to obtain multiple sampling points and sampling parameters; a second determination module for constructing a simulated working condition based on the boundary constraints and the sampling parameters of the multiple sampling points, performing topology simulation based on the boundary constraints and the sampling parameters of the multiple sampling points, and determining the component topology corresponding to the simulated working condition based on the simulation results, wherein the simulated working condition is used to simulate a real working condition; and a generation module for generating a training dataset based on the simulated working condition and the corresponding model topology, training a neural network model using the training dataset, and generating the component topology under the target real working condition using the trained neural network model.

[0013] Optionally, the sampling module is further used to: obtain the mesh topology of the virtual model of the component, wherein the mesh topology includes elements and nodes; divide the target load location space into multiple regions, and randomly sample the elements in each region to obtain sampling points; for the sampling points in each region, randomly obtain the load direction and apply the load force, and generate the sampling parameters of the corresponding sampling points based on the coordinates of the sampling points and the corresponding load force and load direction.

[0014] Optionally, it also includes a filtering module, used to filter the loading position, structural failure and topological solvability of multiple sampling points based on boundary constraints before constructing the simulation conditions based on the boundary constraints and the sampling parameters of multiple sampling points.

[0015] Optionally, the process of selecting loading locations includes: identifying constraint nodes and constraint degrees of freedom in boundary constraints; constructing a node adjacency graph based on the mesh topology of the virtual component model; identifying non-design domains of the virtual component model; calculating the distance between constraint nodes and sampling points, and removing sampling points if the distance is less than a preset distance; detecting whether the element containing the sampling point is located in a non-design domain, and removing the sampling point if it is; and determining whether the sampling point is in a structurally isolated region based on the node adjacency graph, and removing the sampling point if it is in a structurally isolated region.

[0016] Optionally, the screening process for structural failure includes: obtaining the transmission path from the sampling point to the constraint node based on the mesh topology; traversing each node in the transmission path, finding the multi-layer elements adjacent to each node, and grouping the multi-layer elements into an element set; calculating the equivalent cross-sectional area of ​​the target cylindrical model based on the total volume of the element set and the length of the transmission path, calculating the maximum energy of the virtual model of the component under different types of deformation based on the equivalent cross-sectional area, and obtaining the external force work of the sampling point under different types of deformation; if the external force work is greater than the maximum energy, the sampling point is removed.

[0017] Optionally, the topology solvability screening process includes: identifying the constraint degrees of freedom of multiple constraint nodes in the boundary constraint conditions, and determining the resultant force constraint direction and resultant moment constraint direction of the multiple constraint nodes based on the constraint degrees of freedom; selecting reference points in the virtual model of the components; calculating the resultant load force based on the load force of the sampling points, and calculating the resultant load moment about the reference point based on the resultant load force; if the direction of the resultant load force does not match the resultant force constraint direction, or the direction of the resultant load moment does not match the resultant moment constraint direction, then removing the sampling points with mismatched directions based on the resultant force constraint direction or the resultant moment constraint direction.

[0018] Optionally, it also includes: an evaluation module, used to perform diversity evaluation and sensitivity evaluation on the selected sampling points after filtering the loading position, structural failure and topological solvability based on the boundary constraints; generate diversity weights and sensitivity weights for the corresponding sampling points based on the diversity evaluation results and sensitivity evaluation results; determine the target sampling point based on the diversity weights and sensitivity weights of the sampling point; and construct the simulation condition based on the target sampling point and the corresponding sampling parameters and boundary constraints.

[0019] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to perform the neural network training dataset generation method as described above.

[0020] A fourth aspect of this application provides a computer-readable storage medium having a computer program or instructions stored thereon, which are executed by a processor to perform the neural network training dataset generation method as described above.

[0021] A fifth aspect of this application provides a computer program product, including a computer program or instructions, which, when executed, implement the neural network training dataset generation method as described in the above embodiments.

[0022] Therefore, this application has at least the following beneficial effects: This application's embodiments can acquire shape data, material data, and target load location space of a virtual model of a component. Boundary conditions are set based on the shape and material data, and load points within the target load location space are sampled hierarchically to obtain multiple sampling points and sampling point parameters. A simulated working condition is constructed based on the sampling parameters and boundary constraints, and a topological structure simulation is performed. A training dataset is then generated based on the simulation results and the simulated working condition. Hierarchical sampling improves the quality of the sampled data, thereby improving the quality of the training dataset and enhancing the training efficiency and effectiveness of the neural network model. This solves the technical problem of poor training dataset quality in related technologies.

[0023] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0024] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a method for generating a neural network training dataset according to an embodiment of this application; Figure 2 Here is a complete execution flowchart of the neural network training dataset generation method provided according to the embodiments of this application; Figure 3 This is an example diagram of a neural network training dataset generation apparatus provided according to an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0025] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0026] Before describing the solution of this application, let me first introduce the relevant technologies of this application to help in understanding the solution.

[0027] Topology optimization is a mathematical method used to optimize material distribution within a given design space to satisfy specific constraints. Methods include density methods, level set methods, and hard-kill methods. Among these, the Simp method of the density method is the most commonly used topology optimization method and is widely applied in engineering. Simp assigns a relative density variable of 0 to 1 to each element in the design domain to represent material variations and reflect stiffness changes. The element stiffness matrix of each element is calculated and then concatenated to obtain the global stiffness matrix, thus yielding the stress and strain of the model under specific conditions. This method has two problems: first, it relies on calculating the element stiffness matrix of each element, which becomes slow when the number of elements is large; second, when the working conditions change, existing calculation results cannot be reused and need to be recalculated.

[0028] In recent years, research combining AI with topology optimization has become increasingly prevalent, including density-based prediction methods, data-driven end-to-end prediction, and the use of networks such as GANs, CNNs, and fully connected networks to learn models and map operating conditions to topology. Both methods rely on the model's topology optimization results as training datasets, requiring a large number of samples. Ivan Sosnovik used 10,000 samples for 80 rounds of training on a 2D BB beam with 4800 elements, while Saurabh Banga used 6,000 samples for 100 rounds of training on a 3D beam with 3456 elements. Even the simplest models require a large number of samples for multiple training rounds. Currently, the number of elements in complex automotive component models is in the millions. Common methods in current papers, such as Poisson sampling and random sampling, produce datasets of low quality, failing to effectively reflect the mapping relationship between operating conditions and model stress-strain. While these methods perform reasonably well in training simple models, they pose a significant challenge when dealing with complex models. Therefore, while many AI topology optimization papers discuss 3D irregular models, the experimental subjects are often simple MBB beams.

[0029] Specifically, Figure 1 This is a flowchart illustrating a method for generating a neural network training dataset according to an embodiment of this application.

[0030] like Figure 1 As shown, the method for generating the neural network training dataset includes the following steps: In step S101, the shape data, material data, and target load location space of the virtual model of the component are obtained.

[0031] Among them, the parts are vehicle components, which can be irregularly shaped components, such as irregular beams, steering knuckles, etc. The virtual model of the parts can be a 3D unspecified model; the target load location space is the geometric range in the space of the virtual model of the parts where load points are allowed to be set, excluding non-design domains, constraint areas, etc.

[0032] It should be noted that after obtaining the model, it is necessary to first divide it into meshes and nodes to facilitate subsequent processes such as sampling and filtering of load points in the model.

[0033] In step S102, the boundary constraints of the virtual model of the component are determined based on the shape data and material data.

[0034] Boundary constraints include constraint type (point constraint, line constraint, surface constraint), constraint location, and degrees of freedom, which are used to restrict the rigid body motion of the model.

[0035] In step S103, load point layer sampling is performed in the target load location space to obtain multiple sampling points and sampling parameters.

[0036] It is understood that the embodiments of this application can perform layered sampling of load points in the target load location space to obtain multiple sampling points and sampling parameters.

[0037] Among them, the sampling point is the load point obtained by sampling, and the sampling parameters are the parameters of the load point, including the load magnitude, load unit direction and load position.

[0038] In this embodiment of the application, load point hierarchical sampling is performed in the target load location space, including: obtaining the mesh topology of the virtual model of the component, wherein the mesh topology includes elements and nodes; dividing the target load location space into multiple regions, and randomly sampling the elements in each region to obtain sampling points; randomly obtaining the load direction and applying load force to the sampling points in each region, and generating sampling parameters for the corresponding sampling points based on the coordinates of the sampling points and the corresponding load force and load direction.

[0039] It is understood that in this embodiment of the application, the virtual model of the component is divided into units and nodes after being meshed, and the target load location space is divided. The sample is independently sampled in each region to ensure that the sampling points are evenly distributed. For the sampling points in each region, the load direction is randomly obtained and the load force is applied. Based on the coordinates of the sampling points and the corresponding load force and load direction, the sampling parameters of the corresponding sampling points are generated. The problem of local density or sparseness of random sampling is avoided by layered sampling, and the load points are ensured to cover different geometric features of the model.

[0040] Specifically, the load point hierarchical sampling process is as follows: First, an initial centroid is randomly selected to divide the load location space into Voronoi regions. Then, the geometric center of all load points within each region is calculated and used as the new centroid. This process is repeated, continuously updating the centroid and re-dividing the Voronoi regions, until the migration distance of all centroids is below a preset threshold, at which point the algorithm converges. Ultimately, the load location space is divided into multiple stable Voronoi regions. Within each Voronoi region, node ID and element ID are recorded. During sampling, the element ID is first randomly obtained, and then the centroid method is used to obtain the points within the element. This process is repeated within each region to obtain the load point coordinates for hierarchical sampling. A load magnitude is applied to each load point coordinate, and the direction of the unit force is randomly obtained using spherical coordinates, which is then saved together with the corresponding load coordinates.

[0041] In step S104, a simulated working condition is constructed based on the boundary constraints and the sampling parameters of multiple sampling points. A topology simulation is performed based on the boundary constraints and the sampling parameters of multiple sampling points. The component topology corresponding to the simulated working condition is determined based on the simulation structure. The simulated working condition is used to simulate the real working condition.

[0042] Among them, the simulated working condition is a simulation scenario constructed based on the sampling parameters and boundary constraints of the sampling points, used to simulate the stress state of the components in real use.

[0043] It is understood that, in the embodiments of this application, a simulated working condition is constructed based on boundary constraints and sampling parameters of multiple sampling points, a topology simulation is performed based on the boundary constraints and sampling parameters of multiple sampling points, and the component topology corresponding to the simulated working condition is determined based on the simulation structure.

[0044] In this embodiment of the application, before constructing the simulated working condition based on the boundary constraints and the sampling parameters of multiple sampling points, the method further includes: filtering the loading position, structural failure, and topological solvability of the multiple sampling points based on the boundary constraints.

[0045] It is understood that after obtaining multiple sampling points, the embodiments of this application filter the sampling points, that is, filter out invalid combinations and increase the proportion of valid sampling points. Specifically, the filtering can be performed by a constraint discriminator, which includes a loading position criterion, a structural failure criterion, and a topological solvability criterion.

[0046] The loading location criterion checks the loading location of the load points to avoid invalid loading points. The structural failure criterion uses the energy method to approximately determine whether the current load is within the elastic strain range, avoiding loads that could damage the structure and render the topology invalid. The topology solvability criterion calculates the resultant force of the loads, compares it with the constraint conditions, and determines whether the constraint conditions can resist the resultant force, avoiding topological translations that could lead to no solution.

[0047] In this embodiment, the process of selecting the loading location includes: identifying the constraint nodes and constraint degrees of freedom in the boundary constraints; constructing a node adjacency graph based on the mesh topology of the virtual component model; identifying the non-design domain of the virtual component model; calculating the distance between the constraint node and the sampling point, and removing the sampling point if the distance is less than a preset distance; detecting whether the unit where the sampling point is located is located in the non-design domain, and removing the sampling point if it is located in the non-design domain; and determining whether the sampling point is in a structurally isolated area based on the node adjacency graph, and removing the sampling point if it is in a structurally isolated area.

[0048] Among them, the preset distance can be set according to the specific situation, and there is no specific limitation; the constraint node is the node that restricts the degree of freedom in the boundary constraint condition; the non-design domain is the region in the model that is not allowed to be optimized; the structurally isolated region is the region where the sampling point is located and the main body of the model has poor connectivity (i.e., few reachable nodes) and the force cannot be effectively transmitted; the node adjacency graph is a node connection relationship graph constructed based on the mesh topology structure, which is used to determine the node connectivity.

[0049] Specifically, the process for filtering the loading location in this application embodiment is as follows: First, the distance between the constraint node and the load point (i.e., the sampling point) is compared. If the distance is less than a threshold, the load point is removed. Second, nodes outside the design domain are excluded. The element ID corresponding to the load point is checked to see if it is outside the design domain. If it is, the load point is removed. Finally, the connectivity of the nodes is checked. An adjacency graph is constructed based on the element structure of the tetrahedral mesh, and each node sharing the same element is considered an adjacency. Starting from the candidate loading point, a breadth-first search is used to traverse the adjacent nodes and count the number of reachable nodes. When the number of reachable nodes is lower than a set threshold, the loading point is determined to be in an isolated area of ​​the structure and does not have effective loading transfer capability, so it is removed from the loading candidate set.

[0050] In this embodiment, the screening process for structural failure includes: obtaining the transmission path from the sampling point to the constraint node based on the mesh topology; traversing each node in the transmission path, finding the multi-layer elements adjacent to each node, and grouping the multi-layer elements into an element set; calculating the equivalent cross-sectional area of ​​the target cylindrical model based on the total volume of the element set and the length of the transmission path, calculating the maximum energy of the virtual model of the component under different types of deformation based on the equivalent cross-sectional area, and obtaining the external force work of the sampling point under different types of deformation; if the external force work is greater than the maximum energy, the sampling point is removed.

[0051] Specifically, the screening process for structural damage in this application embodiment is as follows: Starting from each load point, the search proceeds sequentially to each constraint point based on the mesh topology. Specifically, a stepwise expansion search strategy based on Euclidean algorithm is adopted, prioritizing the expansion of adjacent nodes with the smallest distance to the constraint point. The search path is recorded, and while global optimality is not guaranteed, consistency in approximate direction is maintained. In each round of expansion, the Euclidean distance from the candidate nodes adjacent to the current node to the target constraint point is calculated, and nodes with smaller distances are prioritized as the path expansion direction, gradually approaching the target position until the path is fully connected. Based on the path of this node, each node in the path is traversed sequentially, and adjacent multi-layer structural elements are found, and all associated elements are grouped into an element set for the approximate model. The equivalent cross-sectional area is calculated using the total volume of associated elements and the path length. Then, the transmission path from the load point to the constraint point is treated as a circular rod, and the work done by the external force caused by the load is calculated according to the failure criterion of the energy method to determine whether it is less than the maximum energy.

[0052] The calculation methods for maximum tension / compression energy and external force work are as follows: ; ; Where E is the elastic modulus of the material, and L is the volume. It represents the maximum average strain, which can be taken as a small value. It is a load. It is the displacement of the load point, which can be approximated by interpolation of the approximate deformation type from the response history of topology optimization.

[0053] Similarly, there is the maximum energy for bending and twisting.

[0054] In this embodiment, the topology solvability screening process includes: identifying the constraint degrees of freedom of multiple constraint nodes in the boundary constraint conditions, and determining the resultant force constraint direction and resultant moment constraint direction of the multiple constraint nodes based on the constraint degrees of freedom; selecting a reference point in the virtual model of the component; calculating the resultant load force based on the load force of the sampling point, and calculating the resultant load moment about the reference point based on the resultant load force; if the direction of the resultant load force does not match the resultant force constraint direction, or the direction of the resultant load moment does not match the resultant moment constraint direction, then removing the sampling point with the mismatched direction based on the resultant force constraint direction or the resultant moment constraint direction.

[0055] Specifically, the topological solvability criterion in this application is used to evaluate the stability and supportability of a structure under a given load. The screening process for topological solvability is as follows: First, all loads acting on the structure are calculated to obtain the resultant force. A reference point is selected, and the vector sum of the moments generated by all forces on that reference point and any directly applied moments is calculated to determine the total resultant moment. After calculating the resultant force and resultant moment, it is checked whether the direction of the resultant force matches the direction of the resultant force constraint provided by all support points, and whether the direction of the resultant moment matches the direction of the resultant moment constraint provided by all support points. This ensures that in all potential rigid body motion directions, at least one or more support points can provide sufficient constraints to resist the resultant force or resultant moment, removing loads that may cause rigid body movement.

[0056] In this embodiment of the application, after filtering multiple sampling points based on boundary constraints by loading location, structural failure, and topological solvability, the method further includes: performing diversity assessment and sensitivity assessment on the filtered sampling points; generating diversity weights and sensitivity weights for the corresponding sampling points based on the diversity assessment results and sensitivity assessment results; determining target sampling points based on the diversity weights and sensitivity weights of the sampling points; and constructing simulated working conditions based on the target sampling points and their corresponding sampling parameters and boundary constraints.

[0057] Among them, diversity assessment compares the deformation type and stress distribution of the sampling points with the differences of historical samples to evaluate its ability to cover new scenarios; sensitivity assessment measures the degree of influence of the load at the sampling points on structural topology changes (e.g., a small load change at a high-sensitivity point can lead to significant topological differences).

[0058] It is understood that the embodiments of this application can perform diversity assessment and sensitivity assessment on the screened sampling points, generate diversity weights and sensitivity weights for the corresponding sampling points based on the diversity assessment results and sensitivity assessment results, determine the target sampling points based on the diversity weights and sensitivity weights of the sampling points, retain diverse samples to ensure that the dataset covers multiple stress scenarios (such as bending and torsion), improve the generalization ability of the neural network, retain high-sensitivity samples, focus on load points that have a significant impact on structural optimization, enable the neural network to learn key mapping relationships faster, thereby improving the quality of the training dataset, shortening the neural network training cycle, and optimizing the topology prediction accuracy.

[0059] Specifically, this application's embodiments calculate the structural deformation type and stress / strain magnitude caused by the load, and combine historical data for evaluation to ensure comprehensive coverage of the physical response distribution. Specifically, the approximate linear model and calculation results from the structural failure criterion are reused to obtain the deformation type, displacement, stress, and strain of the structural response. The calculated stress magnitudes are categorized and normalized based on historical maximum and minimum values, and saved along with the deformation type as historical results. By comparing the current calculation results with historical results, the diversity of deformation types and the similarity of calculation results for the same type are determined, and a weight is assigned to the load.

[0060] Sensitive locations are initially screened based on the structural stiffness distribution to ensure that each sampling point triggers a significant structural change. First, the minimum distance from the current candidate load point to all support points is calculated, representing its distance from the support; a smaller distance indicates lower sensitivity. Then, the maximum support distance for all candidate load points is calculated, and the minimum distance for each load point is normalized. Next, the load point search results from the structural failure criterion are reused to count the unit connectivity, representing its local structural thickness. Sparsity is set as the reciprocal of connectivity; higher sparsity indicates a weaker structure and higher sensitivity. The maximum sparsity for all load points is calculated, and the sparsity for each load point is normalized. Finally, a fusion factor is used to integrate distance and sparsity into a sensitivity response weight.

[0061] The global weights are calculated using the target diversity weights and topology sensitivity weights of the optimization filter, the load points with the highest weights are retained, and then all parameters are saved to a local text file in a specific structure.

[0062] In step S105, a training dataset is generated based on the simulated working conditions and the corresponding model topology. The training dataset is used to train a neural network model, and the trained neural network model is used to generate the component topology under the target real working conditions.

[0063] It is understood that the embodiments of this application generate a training dataset based on simulated working conditions and corresponding model topology, use the training dataset to train a neural network model, use the trained neural network model to generate the component topology under the target real working conditions, and avoid low quality of training dataset caused by random sampling through hierarchical sampling and simulation, thereby improving the quality of training dataset and thus improving the efficiency and effect of neural network training model.

[0064] In summary, the complete process of the neural network training dataset generation method in this application embodiment is as follows: Figure 2 As shown, it includes: 1. Boundary setting.

[0065] Boundary setting refers to the process of setting boundary conditions for a model. For a new model, it is necessary to first divide the mesh and nodes, and then select appropriate boundary constraints based on its shape and material characteristics. Boundary constraints consist of constraint type, constraint location, and degrees of freedom. Constraint types include point constraints, line constraints, and surface constraints.

[0066] 2. Layered sampling of load points.

[0067] Layered sampling of load points refers to the sampling process of load points. The number of load points is usually much smaller than the number of model nodes, so additional partitioning of the model structure is required. A load point consists of the load location, load unit direction, and load magnitude. The number of load points needs to be greater than the rated number; these load points will be filtered and selected in later stages.

[0068] In this method, the K-means clustering method can be used to segment the nodes in the model space, dividing the model into different regions. The unit IDs in the region are randomly obtained, and the sampling points are obtained by the centroid method, and the correspondence between sampling points and units is constructed. It is not necessary to determine the positional relationship between the sampling points and the region boundaries.

[0069] 3. Constraint discriminator.

[0070] The constraint discriminator filters out invalid combinations, increasing the proportion of valid samples. The discriminator includes a loading location criterion, a structural failure criterion, and a topological solvability criterion. The loading location criterion checks the loading location of the load points to avoid invalid loading points. The structural failure criterion uses the energy method to approximately determine whether the current load is within the elastic strain range, preventing loads from causing structural damage and rendering the topology invalid. The topological solvability criterion calculates the resultant force of the loads, compares it with the constraint conditions, and determines whether the constraint conditions can resist the resultant force, avoiding translational phenomena in the topological process that could lead to no solution.

[0071] Among them, loading position criteria, structural failure criteria, and topology solvability criteria are used to make preliminary judgments on the sampling results, filter out meaningless and low-quality sampling results, and avoid problems such as no solution, model instability, and rigid displacement in the topology optimization process.

[0072] Specifically, the structural failure criterion uses an energy discrimination method to determine whether the load causes plastic failure to the model. The shortest path from the load to the constraint point is obtained through a greedy algorithm. Then, a search method is used to expand the local neighborhood to obtain a set of elements. The set of elements is approximated as a cylindrical model. The maximum energy of the three deformation types of tension, compression, torsion and bending and the work done by the load are calculated. The two are compared to determine whether the energy condition is met.

[0073] The topological solvability criterion considers the matching relationship between the direction of the resultant force and resultant moment of all loads and the constrained degrees of freedom to determine whether rigid displacement will occur.

[0074] 4. Optimize the filter.

[0075] The optimization filter evaluates the sampling results and selects the best subset as the final result. The optimization filter consists of two parts: a target diversity measure and a topological sensitivity measure. The target diversity measure considers the deformation type and stress caused by the load, compares it with historical samples from the current round, and assigns a diversity weight; the higher the convergence of historical sampling points, the lower the weight. The topological sensitivity measure introduces weights for the force loading point and the structural response sensitivity index, expressing the magnitude of the load's impact on structural changes; a larger weight indicates a greater impact on the structure. The optimization filter considers the results of both diversity and sensitivity response weights and selects samples with high weights.

[0076] Specifically, the target diversity weight uses an approximate model to calculate the type of change and the magnitude of stress and displacement caused by the load. It compares the results with historical data to determine the scarcity of the load change type and the degree of similarity of stress and displacement to other historical values ​​in the corresponding change type. A weight is then given to represent the level of load diversity. The larger the weight, the higher the scarcity and the lower the similarity.

[0077] Topological sensitivity weights are calculated by taking the distance from the load to the constraint and the connectivity of the elements around the load, normalizing the distance and connectivity, and using a weighting factor to calculate the global weight. The weight represents the magnitude of the structural change that the load may cause to the model; the larger the weight, the greater the structural change. 5. Topology optimization.

[0078] The working conditions are constructed using sampling parameters and a model, and then topology optimization is performed to solve the problem.

[0079] 6. Construct the training set.

[0080] The training set is constructed by saving the topology optimization data results and sampling parameters together, and waiting for the neural network to be trained.

[0081] In addition, this application also saves the load points and corresponding stress and strain in response history for subsequent rapid approximate calculations.

[0082] The method for generating a neural network training dataset according to this application is described below through a specific embodiment, using an irregularly shaped beam as an example for illustration, including: 1. The model is an irregular beam. The leading edge method is used to segment the model to obtain the vertices and boundaries of the tetrahedron.

[0083] 2. Boundary conditions are generated using function encoding, defining four constraint types across three categories: cantilever beam, simply supported beam, and modified simply supported beam. Each constraint is represented by nodal coordinates and a corresponding array of degrees of freedom. Before sampling, boundary conditions for each case are randomly obtained, and loading nodes that are duplicated with constraint nodes are checked and removed to ensure that loading and constraints are mutually exclusive.

[0084] 3. First, an initial centroid is randomly selected to divide the load location space into Voronoi regions. Then, the geometric center of all load points within each region is calculated and used as the new centroid. This process is repeated, continuously updating the centroids and re-dividing the Voronoi regions, until the migration distance of all centroids is below a preset threshold, at which point the algorithm converges. Ultimately, the load location space is divided into multiple stable Voronoi regions. Within each Voronoi region, the node ID and element ID are recorded. During sampling, the element ID is first randomly obtained, and then the centroid method is used to obtain the points within the element. This process is repeated within each region to obtain the load point coordinates of the layered sampling. A load magnitude is applied to each load point coordinate, and the direction of the unit force is randomly obtained using spherical coordinates and saved together with the corresponding load coordinates.

[0085] 3. Apply loading location criteria, structural failure criteria, and topological solvability criteria to the load points. These have been described in the above embodiments and will not be repeated here.

[0086] 4. Calculate the structural deformation type and stress / strain magnitude caused by the load, and evaluate using historical data to ensure comprehensive coverage of the physical response distribution. Specifically, reuse the approximate linear model and calculation results from the structural failure criterion to obtain the deformation type, displacement, stress, and strain of the structural response. Classify the calculated stress magnitudes and normalize them according to historical maximum and minimum values, saving them along with the deformation type as historical results. Compare the current calculation results with historical results to determine the diversity of deformation types and the similarity of calculation results for the same type, assigning a weight to the load.

[0087] Sensitive locations are initially screened based on the structural stiffness distribution to ensure that each sampling point triggers a significant structural change. First, the minimum distance from the current candidate load point to all support points is calculated, representing its distance from the support; a smaller distance indicates lower sensitivity. Then, the maximum support distance for all candidate load points is calculated, and the minimum distance for each load point is normalized. Next, the load point search results from the structural failure criterion are reused to count the unit connectivity, representing its local structural thickness. Sparsity is set as the reciprocal of connectivity; higher sparsity indicates a weaker structure and higher sensitivity. The maximum sparsity for all load points is calculated, and the sparsity for each load point is normalized. Finally, a fusion factor is used to integrate distance and sparsity into a sensitivity response weight.

[0088] 5. Calculate the global weights using the target diversity weights and topology sensitivity weights of the optimization filter, retain the load points with the highest weights, and then save all parameters to a local text file according to a specific structure.

[0089] 6. Perform topology optimization calculations using Hyperworks industrial software. Write Hyperworks scripts using Tcl to read parameter sampling results and call a bash program to start Hyperworks in the background to perform topology optimization simulations on the samples. Extract information such as density, element ID, and node coordinates from the results and compile them into a CSV file to complete the dataset construction.

[0090] In this embodiment, the stress, strain, displacement, and corresponding sampling parameters of the simulation results are saved as a response history for use in structural failure criterion interpolation.

[0091] Based on the above description, the method of this application embodiment can be used to improve the quality of training set samples. The samples can fully and comprehensively represent the geometric features of the model object, and at the same time, make a preliminary judgment on sensitivity, try to retain the load in the high-sensitivity region, so that the sampled parameter samples can better reflect the structural characteristics and flexibility response of the model object; it can significantly improve training efficiency and effect, especially for models with a large number of units, the sampling method of this application is more likely to capture high-sensitivity regions, enabling the neural network to learn high-stiffness, reasonable and realistic topology optimization results.

[0092] In contrast, the parameter sample topology results obtained by simple random sampling contain a large number of highly flexible results. When multiple loads are coupled together, it becomes difficult to learn the influence of forces on the structure. For neural networks, these samples are of low quality, affecting learning efficiency. The sampling method in this application filters the sampling results, removing low-quality samples, which improves the learning effect and efficiency of the neural network. According to the neural network training dataset generation method proposed in this application, shape data, material data, and target load location space of a virtual model of a component can be obtained. Boundary conditions are set based on the shape data and material data, and load points are sampled hierarchically within the target load location space to obtain multiple sampling points and sampling point parameters. Simulated working conditions are constructed based on the sampling parameters and boundary constraints, and topological structure simulation is performed. Then, a training dataset is generated based on the simulation results and simulated working conditions. The quality of the sampled data is improved through hierarchical sampling, thereby improving the quality of the training dataset and enhancing the training efficiency and effectiveness of the neural network model.

[0093] Next, the neural network training dataset generation apparatus according to the embodiments of this application is described with reference to the accompanying drawings.

[0094] Figure 3 This is a block diagram of a neural network training dataset generation device according to an embodiment of this application.

[0095] like Figure 3As shown, the neural network training dataset generation device 10 includes: an acquisition module 100, a first determination module 200, a sampling module 300, a second determination module 400, and a generation module 500.

[0096] The module comprises the following components: an acquisition module 100 for acquiring shape data, material data, and target load location space of a virtual model of a component; a first determination module 200 for determining boundary constraints of the virtual model of the component based on the shape data and material data; a sampling module 300 for performing layered sampling of load points in the target load location space to obtain multiple sampling points and sampling parameters; a second determination module 400 for constructing a simulated working condition based on the boundary constraints and sampling parameters of multiple sampling points, performing topology simulation based on the boundary constraints and sampling parameters of multiple sampling points, and determining the component topology corresponding to the simulated working condition based on the simulation results, wherein the simulated working condition is used to simulate the real working condition; and a generation module 500 for generating a training dataset based on the simulated working condition and the corresponding model topology, training a neural network model using the training dataset, and generating the component topology under the target real working condition using the trained neural network model.

[0097] In this embodiment, the sampling module 300 is further configured to: acquire the mesh topology of the virtual model of the component, wherein the mesh topology includes elements and nodes; divide the target load location space into multiple regions, and randomly sample the elements in each region to obtain sampling points; for the sampling points in each region, randomly acquire the load direction and apply a load force, and generate sampling parameters for the corresponding sampling points based on the coordinates of the sampling points and the corresponding load force and load direction.

[0098] In this embodiment of the application, the apparatus 10 further includes a screening module.

[0099] The filtering module is used to filter multiple sampling points based on boundary constraints, including loading location, structural failure, and topological solvability, before constructing the simulation conditions based on boundary constraints and sampling parameters of multiple sampling points.

[0100] In this embodiment, the process of selecting the loading location includes: identifying the constraint nodes and constraint degrees of freedom in the boundary constraints; constructing a node adjacency graph based on the mesh topology of the virtual component model; identifying the non-design domain of the virtual component model; calculating the distance between the constraint node and the sampling point, and removing the sampling point if the distance is less than a preset distance; detecting whether the unit where the sampling point is located is located in the non-design domain, and removing the sampling point if it is located in the non-design domain; and determining whether the sampling point is in a structurally isolated area based on the node adjacency graph, and removing the sampling point if it is in a structurally isolated area.

[0101] In this embodiment, the screening process for structural failure includes: obtaining the transmission path from the sampling point to the constraint node based on the mesh topology; traversing each node in the transmission path, finding the multi-layer elements adjacent to each node, and grouping the multi-layer elements into an element set; calculating the equivalent cross-sectional area of ​​the target cylindrical model based on the total volume of the element set and the length of the transmission path, calculating the maximum energy of the virtual model of the component under different types of deformation based on the equivalent cross-sectional area, and obtaining the external force work of the sampling point under different types of deformation; if the external force work is greater than the maximum energy, the sampling point is removed.

[0102] In this embodiment, the topology solvability screening process includes: identifying the constraint degrees of freedom of multiple constraint nodes in the boundary constraint conditions, and determining the resultant force constraint direction and resultant moment constraint direction of the multiple constraint nodes based on the constraint degrees of freedom; selecting a reference point in the virtual model of the component; calculating the resultant load force based on the load force of the sampling point, and calculating the resultant load moment about the reference point based on the resultant load force; if the direction of the resultant load force does not match the resultant force constraint direction, or the direction of the resultant load moment does not match the resultant moment constraint direction, then removing the sampling point with the mismatched direction based on the resultant force constraint direction or the resultant moment constraint direction.

[0103] In this embodiment of the application, the apparatus 10 further includes an evaluation module.

[0104] The evaluation module is used to perform diversity and sensitivity evaluations on the selected sampling points after filtering them for loading location, structural failure, and topological solvability based on boundary constraints; generate diversity weights and sensitivity weights for the corresponding sampling points based on the diversity and sensitivity evaluation results; determine the target sampling points based on the diversity and sensitivity weights of the sampling points; and construct the simulated working conditions based on the target sampling points and their corresponding sampling parameters and boundary constraints.

[0105] It should be noted that the foregoing explanation of the embodiment of the neural network training dataset generation method also applies to the neural network training dataset generation device of this embodiment, and will not be repeated here.

[0106] According to the neural network training dataset generation device proposed in the embodiments of this application, the device can acquire shape data, material data, and target load location space of virtual models of parts, set boundary conditions based on shape data and material data, perform hierarchical sampling of load points in the target load location space to obtain multiple sampling points and sampling point parameters, construct simulated working conditions based on sampling parameters and boundary constraints, perform topological structure simulation, and then generate training datasets based on simulation results and simulated working conditions. By hierarchical sampling, the quality of sampled data is improved, thereby improving the quality of training datasets and improving the training efficiency and effect of neural network models.

[0107] Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include: The memory 401, the processor 402, and the computer program stored on the memory 401 and capable of running on the processor 402.

[0108] When processor 402 executes the program, it implements the neural network training dataset generation method provided in the above embodiments.

[0109] Furthermore, electronic devices also include: Communication interface 403 is used for communication between memory 401 and processor 402.

[0110] The memory 401 is used to store computer programs that can run on the processor 402.

[0111] Memory 401 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0112] If the memory 401, processor 402, and communication interface 403 are implemented independently, then the communication interface 403, memory 401, and processor 402 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized into address buses, data buses, control buses, etc. For ease of representation, Figure 4 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0113] Optionally, in a specific implementation, if the memory 401, processor 402, and communication interface 403 are integrated on a single chip, then the memory 401, processor 402, and communication interface 403 can communicate with each other through an internal interface.

[0114] Processor 402 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0115] This application also provides a computer-readable storage medium storing a computer program or instructions thereon, which, when executed by a processor, implements the above-described method for generating a neural network training dataset.

[0116] This application also provides a computer program product, including a computer program or instructions, which, when executed, implement the above-described method for generating a neural network training dataset.

[0117] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0118] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0119] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0120] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or more of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0121] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

Claims

1. A method for generating a neural network training dataset, characterized in that, Includes the following steps: Acquire shape data, material data, and target load location space of the virtual model of the component; The boundary constraints of the virtual model of the component are determined based on the shape data and the material data. Load point layer sampling is performed in the target load location space to obtain multiple sampling points and sampling parameters; A simulated working condition is constructed based on the boundary constraints and the sampling parameters of the multiple sampling points. A topology simulation is performed based on the boundary constraints and the sampling parameters of the multiple sampling points. The component topology corresponding to the simulated working condition is determined based on the simulation results. The simulated working condition is used to simulate the real working condition. A training dataset is generated based on the simulated working conditions and the corresponding model topology. A neural network model is trained using the training dataset. The trained neural network model is then used to generate the component topology under the target real working conditions. The step of performing layered sampling of load points in the target load location space includes: Obtain the mesh topology of the virtual model of the component, wherein the mesh topology includes elements and nodes; The target load location space is divided into multiple regions, and sampling points are obtained by randomly sampling the units in each region; For each region's sampling points, the load direction is randomly obtained and a load force is applied. Based on the coordinates of the sampling points and the corresponding load force and load direction, sampling parameters for the corresponding sampling points are generated. Before constructing the simulated working condition based on the boundary constraints and the sampling parameters of the plurality of sampling points, the method further includes: Based on the boundary constraints, the multiple sampling points are respectively screened for loading location, structural failure, and topological solvability. After performing loading location screening, structural failure screening, and topological solvability screening on the multiple sampling points based on the boundary constraints, the method further includes: Diversity and sensitivity assessments were performed on the selected sampling points; Based on the diversity assessment results and sensitivity assessment results, generate the diversity weights and sensitivity weights for the corresponding sampling points; The target sampling points are determined based on the diversity weight and sensitivity weight of the sampling points. The simulated working condition is constructed based on the target sampling points and the corresponding sampling parameters and boundary constraints.

2. The method for generating a neural network training dataset according to claim 1, characterized in that, The process of filtering the loading location includes: Identify the constraint nodes and constraint degrees of freedom in the boundary constraints; A node adjacency graph is constructed based on the mesh topology of the virtual model of the components; Identify the non-design domains of the virtual model of the component; Calculate the distance between the constraint node and the sampling point; if the distance is less than a preset distance, remove the sampling point. Detect whether the unit where the sampling point is located is located within the non-design domain. If it is located within the non-design domain, remove the sampling point. Based on the node adjacency graph, it is determined whether the sampling point is in a structurally isolated region. If it is in the structurally isolated region, the sampling point is removed.

3. The method for generating a neural network training dataset according to claim 2, characterized in that, The screening process for structural failure includes: The transmission path from the sampling point to the constraint node is obtained based on the mesh topology; Traverse each node in the transmission path, find the multi-layer unit adjacent to each node, and group the multi-layer units into a unit set; The equivalent cross-sectional area of ​​the target cylindrical model is calculated based on the total volume of the unit set and the length of the transmission path. The maximum energy of the virtual model of the component under different types of deformation is calculated based on the equivalent cross-sectional area, and the external force work of the sampling point under different types of deformation is obtained. If the external force work is greater than the maximum energy, then the sampling point is removed.

4. The method for generating a neural network training dataset according to claim 1, characterized in that, The process for filtering topological solvability includes: Identify the constraint degrees of freedom of multiple constraint nodes in the boundary constraint conditions, and determine the resultant force constraint direction and resultant moment constraint direction of the multiple constraint nodes based on the constraint degrees of freedom; Select a reference point in the virtual model of the component; The resultant load force is calculated based on the load force at the sampling point, and the resultant load moment about the reference point is calculated based on the resultant load force. If the direction of the resultant load force does not match the direction of the resultant force constraint, or the direction of the resultant load moment does not match the direction of the resultant moment constraint, then the sampling points with mismatched directions are removed based on the direction of the resultant force constraint or the direction of the resultant moment constraint.

5. A neural network training dataset generation device, characterized in that, include: The acquisition module is used to acquire the shape data, material data, and target load location space of the virtual model of the parts; The first determining module is used to determine the boundary constraint conditions of the virtual model of the component based on the shape data and the material data; The sampling module is used to perform layered sampling of load points in the target load location space to obtain multiple sampling points and sampling parameters; The second determining module is used to construct a simulated working condition based on the boundary constraints and the sampling parameters of the multiple sampling points, perform topology simulation based on the boundary constraints and the sampling parameters of the multiple sampling points, and determine the component topology corresponding to the simulated working condition based on the simulation results, wherein the simulated working condition is used to simulate the real working condition. The generation module is used to generate a training dataset based on the simulated working conditions and the corresponding model topology, train a neural network model using the training dataset, and generate the component topology under the target real working conditions using the trained neural network model. The sampling module is further used to: obtain the mesh topology of the virtual model of the component, wherein the mesh topology includes elements and nodes; divide the target load location space into multiple regions, and randomly sample the elements in each region to obtain sampling points; randomly obtain the load direction and apply load force to the sampling points in each region, and generate sampling parameters for the corresponding sampling points based on the coordinates of the sampling points and the corresponding load force and load direction; The filtering module is used to filter multiple sampling points based on boundary constraints, including loading location, structural failure, and topological solvability, before constructing the simulation conditions based on boundary constraints and sampling parameters of multiple sampling points. The evaluation module is used to perform diversity and sensitivity evaluations on the selected sampling points after filtering them for loading location, structural failure, and topological solvability based on boundary constraints; generate diversity weights and sensitivity weights for the corresponding sampling points based on the diversity and sensitivity evaluation results; determine the target sampling points based on the diversity and sensitivity weights of the sampling points; and construct the simulated working conditions based on the target sampling points and their corresponding sampling parameters and boundary constraints.

6. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the neural network training dataset generation method as described in any one of claims 1-4.

7. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, The computer program or instructions are executed by a processor to implement the neural network training dataset generation method as described in any one of claims 1-4.

Images