CN-121997771-A - Node intelligent generation type method integrating deep learning and topology optimization technology

CN121997771ACN 121997771 ACN121997771 ACN 121997771ACN-121997771-A

Abstract

The invention relates to the technical field of computer aided engineering design, and discloses a node intelligent generation method integrating deep learning and topology optimization technology, which comprises the steps of obtaining an initial node geometric model, performing topology optimization based on a variable density method, and constructing a three-dimensional node voxelized training data set through data enhancement and voxelization; the three-dimensional generation countermeasure network is constructed, training is carried out by utilizing the data set, network parameter constraint is carried out by adopting a composite loss function comprising countermeasure loss, topology regularization loss and diversity constraint loss in the training, a generated voxel model output by a generator is obtained and is spatially aligned with a finite element grid, a preset reward coefficient is given to a unit corresponding to a solid area in the aligned finite element model, and secondary topology optimization calculation is carried out by utilizing an objective function comprising the reward coefficient. The invention converts the discrete three-dimensional matrix into a continuous geometric entity, and eliminates the discrete defects of the model boundary and the fracture phenomenon of the force transmission path.

Inventors

YE JUN
LIAO RONGLIN
LU HONGJIA
QUAN GUAN

Assignees

浙江大学

Dates

Publication Date: 20260508
Application Date: 20260227

Claims (10)

1. The intelligent node generating method integrating the deep learning and topology optimization technology is characterized by comprising the following steps of: Obtaining an initial node geometric model, performing topological optimization based on a variable density method, performing data enhancement and voxelization on the optimized initial node geometric model, and constructing a three-dimensional node voxelization training data set; Constructing a three-dimensional generation countermeasure network comprising a generator and a discriminator, training the three-dimensional generation countermeasure network by utilizing the three-dimensional node voxelized training data set, and constraining the three-dimensional generation countermeasure network by adopting a composite loss function in the training process, wherein the composite loss function comprises countermeasure loss, topology regularization loss and diversity constraint loss; And acquiring a generated voxel model output by the trained generator, performing spatial alignment on the generated voxel model and a finite element grid, endowing a unit corresponding to a solid region in the generated voxel model with a preset rewarding coefficient in the aligned finite element model, and performing secondary topological optimization by using an objective function containing the rewarding coefficient to obtain a final complex node model.
2. The method according to claim 1, wherein when the topology optimization is performed based on a variable density method, a three-field floating projection method is introduced to decouple a density field into a design field for controlling global material distribution, a filtering field for suppressing chequer effect, and a projection field for mapping continuous density into discrete distribution of approximately zero and one.
3. The method for intelligently generating a node fusing deep learning and topology optimization techniques as recited in claim 2, wherein the step of constructing a three-dimensional node voxelized training dataset specifically comprises: In the iterative process of topology optimization based on the variable density method, random disturbance is applied to the unit sensitivity so as to jump out a local extremum; Symmetrically processing the initial node geometric model subjected to topology optimization, and performing rotary transformation on the initial node geometric model subjected to the symmetrical processing around a fixed axis at intervals of a preset angle to expand the sample size; And carrying out voxelization on the initial node geometric model with the expanded sample size, and converting the initial node geometric model into a three-dimensional matrix format suitable for network training to obtain the three-dimensional node voxelization training data set.
4. The method for intelligently generating the node fusing the deep learning and topology optimization technology according to claim 1, wherein the generator comprises a cascaded volume transposition convolution module, a residual block and a self-attention mechanism module, and the generator is used for up-sampling and mapping an input random noise vector layer by layer into a three-dimensional entity structure; The discriminator comprises a volume convolution downsampling layer and a full-connection classifying layer, and is used for extracting multi-scale characteristics and outputting true and false discrimination probability; The first network layer of the discriminator adopts example normalization processing, and the subsequent network layer adopts batch normalization processing in combination.
5. The method for intelligently generating a node by combining deep learning and topology optimization technology according to claim 1, wherein the topology regularization loss at least comprises connectivity loss and continuity loss; The connectivity loss is used for maximizing the consistency of adjacent voxels by calculating the first-order difference of voxel density fields in three main directions and extracting the negative correlation of the maximum communication direction as the loss; The continuity penalty is calculated by introducing a mean convolution kernel, the continuity penalty comprising a first term that penalizes activated voxels in a low density neighborhood using an exponential function, and a second term that constrains the local density to a squared difference penalty term that is at a preset optimum.
6. The method for intelligently generating a node by combining deep learning and topology optimization technology according to claim 5, wherein the topology regularization loss further comprises clustering loss and smoothness loss; the clustering loss is realized by extracting an activated voxel set in the generated voxels, calculating the normalized average distance between each activated voxel and the geometric centroid of the activated voxel set, and minimizing the second moment of material distribution to force the generated voxels to gather towards the center; the smoothness loss is calculated by utilizing a second-order discrete Laplacian operator to calculate a second-order difference of a density field, and abrupt change of the density field is punished to eliminate a boundary ladder effect; And when the composite loss function is configured, the weight value of the clustering loss is configured to be larger than the weight values of the connectivity loss, the continuity loss and the smoothness loss.
7. The method for intelligently generating a node by combining deep learning and topology optimization technology according to claim 1, wherein the diversity constraint loss comprises a diversity loss and a pattern search loss; the diversity loss is calculated to generate cosine similarity among characteristic vectors of samples, and punishment is applied when the cosine similarity exceeds a preset similarity threshold; The pattern search loss is calculated to generate the space distance between samples and the potential space distance between corresponding random noise vectors, and the rejection term and the proportion term are used for forcing the space distance to keep the proportion relation with the potential space distance.
8. The method for intelligently generating a node fusing deep learning and topology optimization techniques as recited in claim 1, wherein the step of spatially aligning the generated voxel model with a finite element mesh specifically comprises: taking the generated voxel model as a source point cloud set, and taking a finite element model containing the finite element grid as a target geometric field; constructing a rigid body transformation model comprising rotation operation, translation operation and fine adjustment scaling operation; And constructing an objective function based on a distance field, carrying out optimization solving by taking the sum of the average value of Euclidean distances from the source point cloud set after the minimization transformation to the surface of the target geometric field and a scaling regularization term as a target, and obtaining optimal transformation parameters to finish the space alignment.
9. The method for intelligently generating a node by combining deep learning and topology optimization technology according to claim 8, wherein the step of performing optimization solution to obtain optimal transformation parameters to complete the spatial alignment specifically comprises: solving a non-convex optimization problem containing the objective function based on the distance field by adopting a multi-starting point differential evolution strategy, and obtaining the optimal transformation parameters; and extracting an entity unit index set in the target geometric field based on the optimal transformation parameters by utilizing a nearest neighbor search algorithm, and generating a density mask array for identifying entity areas.
10. The method for intelligently generating a node by combining deep learning and topology optimization according to claim 9, wherein the step of assigning a preset reward coefficient to a unit corresponding to a solid region in the generated voxel model in the aligned finite element model and performing secondary topology optimization by using an objective function including the reward coefficient specifically comprises: Identifying a solid region in the finite element model based on the density mask array; For the aligned finite element model, if the finite element unit belongs to the entity area, giving an entity rewarding coefficient with the value larger than one; if the finite element unit does not belong to the entity area, giving a default rewarding coefficient with the value equal to one; Constructing a variable density method objective function comprising the entity rewarding coefficient and the default rewarding coefficient, wherein the variable density method objective function is configured to calculate the sum of the rewarding coefficient of each finite element unit, punishment power of unit relative density, transposition of unit displacement vector, and continuous product of unit stiffness matrix and unit displacement vector; And finishing the secondary topological optimization by minimizing the objective function of the variable density method, eliminating boundary discrete defects and obtaining the final complex node model.

Description

Node intelligent generation type method integrating deep learning and topology optimization technology Technical Field The invention relates to the technical field of computer aided engineering design, in particular to a node intelligent generation method integrating deep learning and topology optimization technology. Background In the field of space structures and complex constructional engineering, the nodes are used as key parts for connecting all stressed members, and the topological configuration directly determines the force transmission efficiency and the safety performance of the whole structure. Traditional node designs rely mainly on empirical trial and error or parameterized modeling based on specific criteria, and it is difficult to achieve optimal configuration of material distribution while meeting complex load boundary conditions. In recent years, topological optimization techniques such as a variable density method based on continuous medium mechanics are widely applied to node design, and the method can obtain a structural configuration with a definite force transmission path through sensitivity analysis and iterative solution. However, such numerical optimizing methods consume very much computing resources, and small variations of initial boundary conditions often require re-execution of full-scale iterations, which makes it difficult to meet the real-time requirements of multi-condition fast scheme comparison in engineering design. With the development of artificial intelligence technology, the generation type design of complex nodes by using deep learning models such as generation countermeasure networks becomes a new technical direction. The method tries to realize second-level prediction of the node configuration by establishing a mapping relation between boundary condition features and topology forms. However, in the prior art, the generation result based on the deep learning usually exists in the form of a three-dimensional discrete voxel matrix, and the underlying logic lacks physical conservation constraint, so that the output generation voxel model is extremely easy to generate local fracture of a force transmission path in a three-dimensional space or generate isolated non-bearing material suspended blocks. Meanwhile, due to inherent discrete characteristics of voxel conversion, the generated entity boundary often presents an obvious step-like sawtooth effect, which not only reduces the geometric accuracy of the structure, but also causes that the model cannot be directly imported into a finite element analysis environment for mechanical verification due to lack of boundary continuity, thereby limiting the reliable application of the generated design in actual engineering manufacture. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a node intelligent generation method integrating deep learning and topology optimization technology, which solves the problems that in the existing node topology optimization design, the output three-dimensional voxel discrete model lacks continuity of a physical force transmission path and surface geometric smoothness constraint, so that the generated structure has boundary stepped discrete defects and internal bearing materials are broken, and the requirement of finite element continuous medium mechanical analysis cannot be directly met. In order to achieve the above purpose, the invention is realized by the following technical scheme: the invention provides a node intelligent generation method integrating deep learning and topology optimization technology, which comprises the following steps: Obtaining an initial node geometric model, performing topological optimization based on a variable density method, performing data enhancement and voxelization on the optimized initial node geometric model, and constructing a three-dimensional node voxelization training data set; Constructing a three-dimensional generation countermeasure network comprising a generator and a discriminator, training the three-dimensional generation countermeasure network by utilizing the three-dimensional node voxelized training data set, and constraining the three-dimensional generation countermeasure network by adopting a composite loss function in the training process, wherein the composite loss function comprises countermeasure loss, topology regularization loss and diversity constraint loss; And acquiring a generated voxel model output by the trained generator, performing spatial alignment on the generated voxel model and a finite element grid, endowing a unit corresponding to a solid region in the generated voxel model with a preset rewarding coefficient in the aligned finite element model, and performing secondary topological optimization by using an objective function containing the rewarding coefficient to obtain a final complex node model. Through the technical scheme, the invention couples single three-dimensional matrix discrete