Structural topology optimization method and device based on unsupervised deep learning

Abstract

The invention discloses a structure topological optimization method and device based on unsupervised deep learning, and the method comprises the steps: inputting a multi-dimensional structure parameter into a trained structure topological optimization model based on unsupervised learning, and obtaining a topological optimization structure; wherein the trained structure topological optimization model is obtained by performing unsupervised training according to sample structure parameters without labels, and the structure topological optimization model comprises a depth automatic encoder, a merging layer, a decoder and a mask dot lamination layer; multi-dimensional structure parameters are input to the structure topology optimization model through the encoder, the merging layer and the input port of the mask dot lamination layer. The structure topology optimization model used in the optimization method provided by the invention is trained based on unsupervised deep learning, so that the requirement on the quality of a sample in the training process is not high, the occupied computing resources are less, but the optimization effect is better; therefore, the optimization method provided by the invention can achieve a better optimization effect with lower cost.

Description

Technical Field

[0001] The present invention belongs to the technical field of structural topology optimization, and specifically relates to a structural topology optimization method and device based on unsupervised deep learning. Background Art

[0002] Structural topology optimization aims to find the optimal distribution of materials (structural topology) and achieve specific structural performance requirements under given constraints. It is considered the most challenging structural optimization method. In the past three decades, theoretical achievements and related algorithms in the field of structural topology optimization have developed rapidly and are widely used in engineering manufacturing fields such as machinery, automobiles, aerospace, and oceans. They are also used to design microscopic materials with specific mechanical or electromagnetic properties.

[0003] With the rapid advancement of computer technology and the explosive growth of data, machine learning algorithms represented by deep learning have achieved unprecedented development, and their related research results have been widely used in many fields such as computer, finance, life sciences and traditional industries. At present, a large number of research works have been devoted to exploring the application of machine learning in topology optimization, aiming to improve the computational efficiency of topology optimization and realize real-time topology optimization design.

[0004] However, the models used in current deep learning-based structural topology optimization methods are often trained through supervised learning. During the training process, a large number of labeled sample data sets with high quality are required, and a large amount of memory set computing resources are consumed during training, resulting in the high cost of current deep learning-based structural topology optimization methods. Summary of the invention

[0005] The embodiments of the present invention provide a method and device for structural topology optimization based on unsupervised deep learning, which can solve the problem of high cost of current structural topology optimization methods based on deep learning.

[0006] In a first aspect, an embodiment of the present invention provides a structural topology optimization method based on unsupervised deep learning, the method comprising:

[0007] Input the multi-dimensional structural parameters into the trained structural topology optimization model based on unsupervised learning to obtain the topology optimized structure;

[0008] Among them, the trained structural topology optimization model is obtained by unsupervised training based on the structural parameters of unlabeled samples. The structural topology optimization model includes a deep autoencoder, a merging layer, a decoder, and a masked dot product layer; the multi-dimensional structural parameters are input into the structural topology optimization model through the input ports of the deep autoencoder, the merging layer, and the masked dot product layer.

[0009] In a second aspect, an embodiment of the present invention provides a structural topology optimization device based on unsupervised deep learning, the device comprising a processing unit; the processing unit is used to:

[0010] Input the multi-dimensional structural parameters into the trained structural topology optimization model based on unsupervised learning to obtain the topology optimized structure;

[0011] Among them, the trained structural topology optimization model is obtained by unsupervised training based on the structural parameters of unlabeled samples. The structural topology optimization model includes a deep autoencoder, a merging layer, a decoder, and a masked dot product layer; the multi-dimensional structural parameters are input into the structural topology optimization model through the input ports of the deep autoencoder, the merging layer, and the masked dot product layer.

[0012] In a third aspect, an embodiment of the present invention provides an electronic device, comprising a processor and a memory, wherein the memory is used to store computer programs; the processor can be used to execute the computer program (instructions) stored in the memory to implement the method of the first aspect above.

[0013] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed, the method of the first aspect described above can be implemented.

[0014] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: since the structural topology optimization model used in the optimization method provided by the present invention is trained based on unsupervised deep learning, the quality requirements for samples during the training process are not high, and the occupied computing resources are relatively small; therefore, the cost of the optimization method provided by the present invention is relatively low. On the other hand, for the same computing resources, the training effect provided by the present invention is far better than the training effect of the existing supervised learning, and is particularly suitable for topology optimization problems with more variables in the problem description. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic diagram of a structural topology optimization model provided by an embodiment of the present invention;

[0016] Figure 2 A schematic diagram of a design domain provided by an embodiment of the present invention;

[0017] Figure 3 A schematic diagram of a first tensor provided by an embodiment of the present invention;

[0018] Figure 4 A schematic diagram of a scenario for prediction of a structural topology optimization model provided by an embodiment of the present invention;

[0019] Figure 5A flowchart of a training method for a structural topology optimization model provided by an embodiment of the present invention;

[0020] Figure 6 A schematic diagram of a training scenario of a structural topology optimization model provided in an embodiment of the present invention;

[0021] Figure 7 A comparative schematic diagram of strain energy deviation provided by an embodiment of the present invention;

[0022] Figure 8 A schematic diagram for comparing volume deviations provided by an embodiment of the present invention;

[0023] Fig. 9 A comparative schematic diagram of model reliability provided by an embodiment of the present invention;

[0024] Fig.10 A schematic diagram for comparing prediction effects provided by an embodiment of the present invention;

[0025] Fig.11 A flowchart of an implementation method of a structural topology optimization method based on unsupervised deep learning provided in an embodiment of the present invention;

[0026] Fig.12 A schematic diagram of the structure of a structural topology optimization device based on unsupervised deep learning provided by an embodiment of the present invention;

[0027] Fig.13 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0028] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present invention. However, it should be clear to those skilled in the art that the present invention may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present invention.

[0029] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0030] It should also be understood that the term "and / or" used in the present description and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0031] As used in the present specification and the appended claims, the term "if" may be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" may be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]", depending on the context.

[0032] In addition, in the description of the present specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0033] References to "one embodiment" or "some embodiments" etc. described in the present specification mean that one or more embodiments of the present invention include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0034] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0035] Example 1

[0036] Figure 1 The structure diagram of a structural topology optimization model provided by an embodiment of the present invention is shown. As an example but not limitation, the model 100 may include a deep autoencoder 110, a merging layer 120, a decoder 130, and a masked dot product layer 140.

[0037] In some embodiments, see Figure 1 , the deep autoencoder 110 can encode the first tensor to extract the features of the first tensor to obtain an encoded feature map, and then the merging layer 120 merges the encoded feature map and the second tensor into a new feature map, namely, a spliced ​​feature map. The decoder 130 decodes the spliced ​​feature map to gradually enlarge the size of the spliced ​​feature map and outputs a decoded feature map of the target size. The masked dot product layer 140 masks the decoded feature map based on the third tensor and outputs a prediction result (i.e., a topology optimized structure).

[0038] In a possible implementation, all design parameters may be embedded in the first tensor and the third tensor after masking, while part of the design parameters may be embedded in the second tensor.

[0039] For example, the structural topology optimization model can be used to predict the topological structure with the minimum strain energy according to the design parameters. The design parameters can include the design domain size, load conditions and boundary conditions of the topological structure.

[0040] For example, the structural topology optimization model is used to predict the topology of bridge components, and the design parameters may include the parameters shown in Table 1 below. Figure 2 , where the dark area of ​​the square element discretization is the design domain, and the variable A r The two concentrated forces F1 and F2 are applied to the grid at positions (X1, Y1) and (X2, Y2) at angles θ1 and θ2 relative to the horizontal. The vertical pulleys restrict lateral movement, and the horizontal pulleys restrict longitudinal movement. At the same time, the left edge of the structure can only move in the vertical direction.

[0041] Specifically, see Table 1, where parameter A r , f belong to the design domain dimensions, θ1, X1, Y1, θ2, X2, Y2 belong to load conditions, and H1, V1, H2, V2 belong to boundary conditions.

[0042] In one example, the first tensor I1 and the third tensor I3 may be embedded in [A r fθ1X1 Y1θ2X2 Y2 H1 V1H2 V2], the size of the first tensor I1 can be 33×65×5, and the size of the third tensor I3 can be 32×64×1.

[0043] For details, see Figure 3 , in the first dimension of I1[A x ,A y ,0] can be embedded in the design domain size, and the mask value in the valid design domain can be 1, corresponding to Figure 1 The dark part in the invalid design domain can be 0. x is the length of the design domain, A y is the width of the design domain.

[0044] For details, see Figure 3In order to embed the external forces F1 and F2 in the second and third dimensions of the first tensor respectively, the mask values ​​in the second and third dimensions can be set to satisfy: I1[X1 Y1 1] = cosθ1, I1[X1 Y1 2] = sinθ1, I1[X2 Y2 1] = cosθ2, I1[X2 Y2 2] = sinθ2. In particular, when [X1 Y1] and [X2 Y2] are the same, let I1[X1Y1 1] = cosθ1+cosθ2, I1[X1 Y1 2] = sinθ1+sinθ2, and the other values ​​are 0.

[0045] For details, see Figure 3 In order to embed boundary conditions in the fourth and fifth dimensions of the first tensor, the mask value of the position where horizontal movement is prohibited in the fourth dimension can be set to 1, and the rest can be set to 0; the mask value of the position where vertical movement is restricted in the fifth dimension can be set to 1, and the rest can be set to 0.

[0046] Table 1

[0047] Dimensions variable meaning scope 1 <![CDATA[A r ]]> Aspect Ratio 1～31 2 f Quality score 0.3～0.4 3 <![CDATA[θ1]]> <![CDATA[The angle of the acting force F1, relative to the horizontal line]]> 0～2π 4 <![CDATA[X1]]> <![CDATA[Horizontal position of the acting force F1]]> <![CDATA[1~64-A r ]]> 5 <![CDATA[Y1]]> <![CDATA[Vertical position of the acting force F1]]> 1～31 6 <![CDATA[θ2]]> <![CDATA[The angle of the force F2, relative to the horizontal line]]> 0～2π 7 <![CDATA[X2]]> <![CDATA[Horizontal position of the acting force F2]]> <![CDATA[1~64-A r ]]> 8 <![CDATA[Y2]]> <![CDATA[Vertical position of the acting force F2]]> 1～31 9 <![CDATA[H1]]> The horizontal position of the applied X-DoF constraint <![CDATA[33-A r ~63-A r ]]> 10 <![CDATA[V1]]> The horizontal position of the applied Y-DoF constraint <![CDATA[33-A r ~63-A r ]]> 11 <![CDATA[H2]]> The vertical position of the applied X-DoF constraint 1～31 12 <![CDATA[V2]]> The vertical position of the applied Y-DoF constraint 1～31

[0048] In one example, the second tensor I2 may have a quality score f embedded therein, and its size may be 4×8×1.

[0049] In one possible implementation, see Figure 4 , the deep autoencoder 110 may include multiple layers of convolutional layers and pooling layers. The input size of the convolutional layer is H×W×C, the convolution kernel size is k1×k1×F1, and the number of output channels is F1. It can use a nonlinear activation function, such as ReLU, to maintain or change the input size and extract the primary features of the first tensor. The input size of the subsequent convolutional layer changes according to the output of the previous layer, the convolution kernel size is k2×k2×F2, the number of output channels is F2, the degree of input size change and the activation function remain unchanged, and higher-level features are extracted. The pooling layer can reduce the size of the feature map to H'×W'×C', and extract more abstract features through pooling. Through multiple rounds of convolution and pooling, the number of convolution kernels is gradually increased and the size of the feature map is reduced, multi-level features are extracted, and finally an encoded feature map can be obtained.

[0050] For example, the first tensor may be input by an embedding layer before the deep autoencoder 110 .

[0051] In one possible implementation, the merging layer may concatenate the encoded feature map and the second tensor in a dimension of -1 to form a new concatenated feature map.

[0052] In one possible implementation, see Figure 4The decoder 130 may include a plurality of deconvolution layers and upsampling layers. The deconvolution layer deconvolves the concatenated feature map, and the upsampling layer gradually enlarges the size of the feature map and finally outputs the decoded feature map.

[0053] For example, see Figure 1 and Figure 4 The output prediction result can be a grayscale image, and the black part is the predicted topology optimization structure.

[0054] Example 2

[0055] The training method of the structural topology optimization model provided in the embodiment of the present invention can be applied to electronic devices such as mobile terminals, personal laptops, supercomputers, etc. The embodiment of the present invention does not impose any restrictions on the specific type of electronic devices.

[0056] Figure 5 The flowchart shown is a method for training a structural topology optimization model provided by an embodiment of the present invention. As an example but not a limitation, the method can be applied in the above electronic device to train the above model 100. The method can include steps S501-S505, and each step is described below.

[0057] S501, inputting the k+1th group of sample structure parameters into the structure topology optimization model after the kth round of training to obtain the k+1th group of sample topology optimization structure.

[0058] Exemplarily, k is a non-negative integer, the structural topology optimization model after the 0th round of updating is the initial structural topology optimization model, and its network parameter weights are the initial weights.

[0059] Alternatively, see Figure 6 , a set of sample structure parameters {p k}.

[0060] In one example, the purpose of the structural topology optimization model is to predict the topology optimized structure with the minimum strain energy and satisfying the volume constraint. If the deformation of the structure is small and the deformation of the structure is linearly related to the force, the minimum strain energy can be equivalent to the minimum linear elastic deformation. At this time, the prediction target of the structural topology optimization model can satisfy the following formula:

[0061]

[0062] Among them, ο(x) is the objective function (i.e., strain energy), K is the global stiffness matrix (used to describe the stiffness of the structure), U is the overall displacement (used to describe the deformation of the structure), x is the design variable vector (i.e., the topology optimization structure), and x i is the density of the i-th unit in the design domain, u i For xi The displacement generated at i For x i The stiffness matrix at , V(x) is the material volume, V0 is the design domain volume, f′ is the preset volume fraction, F is the force vector (external force applied to the structure), x min 、x max x i The minimum and maximum values ​​of the boundary conditions, p is the penalty force, and N is the number of elements in the design domain (i.e. Figure 2 The number of squares in the dark area).

[0063] S502, based on the convergence judgment model, determining whether the structural topology optimization model has completed training according to boundary conditions and the k+1th group of sample topology optimization structures.

[0064] In a possible implementation, the topological optimization structure x of the k+1th group of samples can be k+1 Substitute into the convergence judgment model (i.e. the following formula (1.2)), if x k+1 If the constraints of the convergence determination model are not met, the structural topology optimization model has not been trained, and step S503 can be performed.

[0065] In one example, the convergence determination model may satisfy the following formula:

[0066]

[0067] Among them, h is the unfair output in the structural topology optimization model, sigmoid(·) is the output activation function of the structural topology optimization model, ο(x k+1 ) is x k+1 The strain energy, Represents x k+1 gradient.

[0068] In another possible implementation, if x k+1 If the constraints of the convergence determination model are met, the structural topology optimization model has completed training and step S505 can be performed.

[0069] S503: Determine the k+1th group of weights of the network parameters of the structural topology optimization model according to the k+1th group of sample topology optimization structures.

[0070] In a possible implementation, the model may be updated based on the network parameters, and the k+1th group of weights of the network parameters of the structural topology optimization model may be determined according to the k+1th group of sample topology optimization structures.

[0071] If using This traditional update formula based on the back-propagation algorithm using the chain rule to calculate the partial derivative of e with respect to w and using the gradient decay algorithm to update the weights cannot constrain the volume of the topological optimization structure predicted by the model. Therefore, the gradient projection algorithm can be used to determine the network parameters to update the model.

[0072] Specifically, it can be Make a conversion on P C , when x is along When the change occurs, the optimization constraint (volume constraint) is always satisfied and ο(x) still decreases.

[0073] Exemplarily, the network parameter update model may satisfy the following formula:

[0074]

[0075] Among them, w k+1 is the k+1th group weight of the network parameters of the structural topology optimization model, w k is the kth weight of the network parameters of the structural topology optimization model, α k Indicates that the learning rate during the k-th round of training (i.e., the model performs the k-th round of prediction) is a negative real value, D is a partial differential matrix, P C is the transformation matrix.

[0076] In one example, It can be obtained based on formula (1.1) combined with Hooke's law, which can satisfy the following formula:

[0077]

[0078] Among them, x j is the jth component of x, u j , k j are the j-th components of U and K respectively.

[0079] In one example, strain energy is used as the required energy function, and its derivative with respect to the weight w of the network parameter of the structural topology optimization model is determined to obtain the partial differential matrix D. The specific value of the partial differential matrix D can be calculated by automatic differentiation of the structural topology optimization model.

[0080] For example, D may satisfy the following formula:

[0081]

[0082] in, is the gradient with respect to w, w i is the i-th component of w, x i is the i-th component of x.

[0083] S504, based on the back propagation algorithm, updating the network parameters of the structural topology optimization model according to the k+1th group of weights, completing the k+1th round of training of the structural topology optimization model to obtain the structural topology optimization model after the k+1th round of training.

[0084] Exemplarily, after step S504 is performed, k=k+1 may be set, and step S501 may be continued.

[0085] S505, using the structural topology optimization model after the k-th round of training as the trained structural topology optimization model.

[0086] According to the training method provided by the present invention, unlabeled samples are used for training, the network parameters of the model are updated through the network parameter update model, and the convergence judgment model is used to determine whether the model has completed the training, thereby realizing unsupervised deep learning of the structural topology optimization model; instead of using labeled samples for training, the model is updated based on the loss function determined by the real label and it is determined whether the model has completed the training, which can reduce the requirements on the number and quality of samples and reduce the computing resources consumed during training.

[0087] In order to better illustrate the beneficial effects of the training method provided by the present invention, the following simulation experiment was conducted:

[0088] Simulation experiment 1

[0089] Exemplarily, in a simulation experiment, the model 100 can be trained using the current supervised learning method and the unsupervised learning method provided by the present invention. In the experiment, 10,000 samples can be uniformly sampled from the 12-dimensional sample input space for training. Since 10,000 to 80,000 uniformly distributed training samples are used to supervise the model 100, each sample needs to spend 100 finite element analyses. For such a high-dimensional input space, the neural network is easy to overfit, so 5,000 uniformly sampled data are used as the verification set for supervised learning. For unsupervised learning, 2,000 samples are randomly selected from the sample input space C in each iteration, and the model 100 is trained through 50 to 600 iterations.

[0090] Figure 7 Shown is a comparative schematic diagram of strain energy deviation provided in an embodiment of the present invention.

[0091] See also Figure 7 , Figure 7 (a) shows the variation of strain energy deviation with the number of finite element calculations during supervised learning; Figure 7 (b) shows how the strain energy deviation changes with the number of finite element calculations when unsupervised learning is performed according to the training method provided by the present invention.

[0092] Exemplarily, the number of finite element calculations is proportional to the number of training samples.

[0093] Specifically, in Figure 7 Each box plot in Figure 2 shows how the distribution of strain energy deviation Δο varies with the number of finite element calculations. On each box, the center mark indicates the median, and the bottom and top edges of the box indicate the 25th and 75th percentiles, respectively. These outliers are plotted separately using a red “+” symbol.

[0094] Depend on Figure 7 As can be seen from (a) in Figure 1, the strain energy deviation of supervised learning decreases slowly with the increase of training samples; even with 80,000 training samples (i.e., 8 million finite element analyses), about 20% of the prediction results are unusable. Figure 7 As can be seen in (b), as the number of training iterations increases, the strain energy deviation of the prediction results obtained by unsupervised learning decreases rapidly. Even after only 500 iterations (i.e., 1 million finite element analyses), only 9% of the prediction results are unusable.

[0095] Figure 8 A comparative schematic diagram of volume deviation provided in an embodiment of the present invention is shown.

[0096] See also Figure 8 , Figure 8 (a) shows the change of volume deviation with the number of finite element calculations during supervised learning; Figure 8 (b) shows how the volume deviation changes with the number of finite element calculations when unsupervised learning is performed according to the training method provided by the present invention.

[0097] Depend on Figure 8 As can be seen from (a) in Figure 1, as the number of training samples increases (i.e., the number of finite element calculations increases), the predicted results ΔV of supervised learning decreases quite slowly. Even with 80,000 training samples (i.e., 8 million finite element analyses), about 13% of the predicted results are still unqualified. In contrast, Figure 8 As can be seen in (b), as the number of training iterations increases, the volume deviation predicted by unsupervised learning decreases rapidly. Even after only 500 iterations (i.e., 1 million FEAs), only 1% of the predictions fail.

[0098] Fig. 9 A comparative schematic diagram of the reliability of a model provided in an embodiment of the present invention is shown.

[0099] See also Fig. 9 , Fig. 9 (a) shows how the reliability of the model changes with the number of finite element calculations during supervised learning; Fig. 9(b) shows how the model reliability changes with the number of finite element calculations when unsupervised learning is performed according to the training method provided by the present invention.

[0100] Depend on Fig. 9 It can be seen that unsupervised learning can improve the reliability of the model by 12% with only 100,000 finite element calculations. Compared with supervised learning, unsupervised learning reduces the unreliable predictions of the model by nearly 30% using only 1 / 80 of the finite element calculations. And when the amount of finite element calculations increases to 1 million, the reliability increases to 92%. That is, compared with supervised learning, unsupervised learning reduces the unreliable predictions of the model by nearly 75% using only 1 / 8 of the finite element calculations.

[0101] Fig.10 A comparative schematic diagram of a prediction effect provided by an embodiment of the present invention is shown.

[0102] See also Fig.10 , Fig.10 (a) is the topological optimization structure output by the model 100 based on supervised learning training. Fig.10 (b) is the topological optimization structure output by the model 100 based on unsupervised learning training. Fig.10 (c) in the figure is the actual optimal topology.

[0103] Depend on Fig.10 It can be seen that even after training with up to 80,000 samples, the prediction results of supervised learning are still very fuzzy. This is because the number of training samples is far from enough for a 12-dimensional problem. For unsupervised learning, clear prediction results were obtained using 500 iterations.

[0104] Therefore, according to the training method provided by the present invention, unlabeled samples are used for training, the network parameters of the model are updated through the network parameters, and the convergence judgment model is used to determine whether the model has completed the training, thereby realizing unsupervised deep learning of the structural topology optimization model; instead of using labeled samples for training, the model is updated based on the loss function determined by the real label and it is determined whether the model has completed the training, which can reduce the requirements on the number and quality of samples and reduce the computing resources consumed during training.

[0105] In addition, the results of the above simulation experiments can also prove that the training method provided by the present invention can improve the accuracy of model prediction and can solve high-dimensional optimization problems with topological structures of more than ten dimensions.

[0106] Example 3

[0107] Fig.11The flowchart shown is an implementation flow chart of a structural topology optimization method based on unsupervised deep learning provided by an embodiment of the present invention. As an example but not a limitation, the optimization method can be applied to the above-mentioned electronic device, and the optimization method can include steps S1101-S1102. Each step is described below.

[0108] S1101, obtaining multi-dimensional structural parameters.

[0109] Exemplarily, the parameter types of the multi-dimensional structural parameters can be found in Table 1 above.

[0110] S1102, inputting the multi-dimensional structural parameters into a trained structural topology optimization model based on unsupervised learning to obtain a topology optimized structure.

[0111] Exemplarily, the structural topology optimization model may be a model 100 trained based on the above training method.

[0112] Since the structural topology optimization model used in the optimization method provided by the present invention is trained based on unsupervised deep learning, the number and quality of samples are not required to be high during the training process, and the computing resources occupied are relatively small; therefore, the cost of the optimization method provided by the present invention is relatively low.

[0113] Example 4

[0114] Fig.12 The structure diagram of a structural topology optimization device based on unsupervised deep learning provided by an embodiment of the present invention is shown. As an example but not a limitation, the device 1200 may include an acquisition unit 1210 and a processing unit 1220.

[0115] Exemplarily, the acquisition unit 1210 can be used to acquire multi-dimensional structural parameters, and the processing unit 1220 can be used to input the multi-dimensional structural parameters into a trained structural topology optimization model based on unsupervised learning based on the above-mentioned optimization method to obtain a topologically optimized structure.

[0116] Exemplarily, the structural topology optimization model may be a model 100 trained based on the above training method.

[0117] Since the structural topology optimization model used in the present device 1200 is trained based on unsupervised deep learning, the quantity and quality of samples are not required to be high during the training process, and the computing resources occupied are relatively small; therefore, the operating cost of the optimization device provided by the present invention is relatively low.

[0118] Example 5

[0119] Fig.13 FIG. 1 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. Fig.13The electronic device 1300 shown may include: at least one processor 1310 ( Fig.13 Only one processor is shown in the figure), a memory 1320, and a computer program 1330 stored in the memory 1320 and executable on the at least one processor 1310, wherein the processor 1310 implements the steps of any of the above-mentioned method embodiments when executing the computer program 1330.

[0120] The electronic device 1300 may be a processing device such as a robot that can implement the above method. The embodiment of the present invention does not impose any limitation on the specific type of the electronic device.

[0121] Those skilled in the art will understand that Fig.13 The electronic device 1300 is merely an example and does not constitute a limitation on the electronic device. The electronic device 1300 may include more or fewer components than shown in the figure, or may combine certain components, or may include different components. For example, the electronic device 1300 may also include an input and output interface.

[0122] The processor 1310 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASTC), field programmable gate arrays (FPGA), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0123] The memory 1320 may be an internal storage unit in some embodiments, such as a hard disk or a memory. The memory 1320 may also be an external storage device in other embodiments, such as a plug-in hard disk, a smart memory card (SmartMemory Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. Further, the memory 1320 may also include both an internal storage unit and an external storage device. The memory 1320 is used to store an operating system, an application, a boot loader (Boot Loader), data, and other programs, such as the program code of the computer program, etc. The memory 1320 may also be used to temporarily store data that has been output or is to be output.

[0124] It should be understood that the order of execution of the steps in the above embodiment does not necessarily mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0125] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into a processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0126] An embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned method embodiments can be implemented.

[0127] An embodiment of the present invention provides a computer program product. When the computer program product runs on an electronic device, the electronic device can implement the steps in the above-mentioned method embodiments when executing the computer program product.

[0128] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium can at least include: any entity or device that can carry the computer program code to the camera / terminal device, recording medium, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, RandomAccess Memory), electric carrier signal, telecommunication signal and software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk or an optical disk. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electric carrier signals and telecommunication signals.

[0129] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0130] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

Claims

1. A structural topology optimization method based on unsupervised deep learning, characterized in that: include: Input the multi-dimensional structural parameters into the trained structural topology optimization model based on unsupervised learning to obtain the topology optimized structure; The trained structural topology optimization model is obtained by unsupervised training based on unlabeled sample structural parameters, and the structural topology optimization model includes a deep autoencoder, a merging layer, a decoder, and a masked dot product layer; The multi-dimensional structural parameters are input into the structural topology optimization model through the input ports of the deep autoencoder, the merging layer and the masked dot product layer.

2. The method according to claim 1, characterized in that The deep autoencoder is used to encode a first tensor to extract features of the first tensor to obtain an encoded feature map, wherein the first tensor is obtained by performing mask processing on all the multidimensional structural parameters; The merging layer is used to merge and connect the encoded feature map and the second tensor to obtain a concatenated feature map, wherein part of the multidimensional structure parameters are embedded in the second tensor; The decoder is used to decode the spliced ​​feature map to obtain a decoded feature map; The masked dot product layer is used to mask the decoded feature map based on a third tensor to identify the design domain of the topology optimization structure to obtain the topology optimization structure, wherein the third tensor is obtained by masking all of the multidimensional structure parameters.

3. The method according to claim 1, characterized in that The training method of the topology optimization model includes: Input the k+1th group of sample structure parameters into the structural topology optimization model after the kth round of training to obtain the k+1th group of sample topology optimization structure; Based on the convergence judgment model, determining whether the structural topology optimization model has completed training according to boundary conditions and the k+1th group of sample topology optimization structures; If the structural topology optimization model has not completed training, then updating the model based on the network parameters, and determining the k+1th group of weights of the network parameters of the structural topology optimization model according to the k+1th group of sample topology optimization structures; Based on the back propagation algorithm, the network parameters of the structural topology optimization model are updated according to the k+1th group of weights, and the k+1th round of training of the structural topology optimization model is completed to obtain the structural topology optimization model after the k+1th round of training.

4. The method according to claim 3, characterized in that The convergence determination model satisfies the following formula: Where x represents the prediction result output by the structural topology optimization model, x min 、x max are the minimum and maximum values ​​in the boundary conditions, sigmoid(·) is the activation function, h is the unfair output in the structural topology optimization model, Represents the topological optimization structure x of the k+1th group of samples k+1 The gradient of ο(x k+1 ) is x k+1 The strain energy.

5. The method according to claim 4, characterized in that The network parameter update model satisfies the following formula: Among them, w k+1 is the k+1th group weight, w k is the kth group weight of the network parameters of the structural topology optimization model, α k represents the learning rate during the kth round of training, D is a partial differential matrix, P C is the transformation matrix.

6. The method according to claim 4 or claim 5, characterized in that: The following formula can be satisfied: Among them, p is the penalty force, x j is the jth component of x, u j is the jth component of the overall displacement U, k j is the jth component of the global stiffness matrix K.

7. The method according to claim 5, characterized in that The partial differential matrix satisfies the following formula: in, is the gradient of the weight w of the network parameter of the structural topology optimization model, w i is the i-th component of w, x i is the i-th component of x.

8. A structural topology optimization device based on unsupervised deep learning, characterized in that: The invention comprises a processing unit, wherein the processing unit is configured to: Input the multi-dimensional structural parameters into the trained structural topology optimization model based on unsupervised learning to obtain the topology optimized structure; Among them, the trained structural topology optimization model is obtained by unsupervised training based on unlabeled sample structural parameters, and the structural topology optimization model includes a deep autoencoder, a merging layer, a decoder, and a masked dot product layer; the multi-dimensional structural parameters are input into the structural topology optimization model through the input ports of the deep autoencoder, the merging layer and the masked dot product layer.

9. An electronic device comprising a memory, a processor and a computer program stored in the memory, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the electronic device, the method according to any one of claims 1 to 7 is implemented.

Images