Machine tool structure optimization method based on self-guided online learning and high-efficiency sampling

Abstract

The present invention discloses a method for optimizing machine tool structures based on self-guided online learning and high-efficiency sampling. The method comprises: constructing a two-dimensional / three-dimensional model of the machine tool component to be optimized and an optimization objective function; constructing a neural network integrating a probability distribution convolution layer and a loss function including an information divergence value; inputting an initial sample group of design variables into the optimization objective function to obtain an objective function value and inputting the value into the network for training until the loss function converges to obtain a machine tool structure optimization neural network; optimizing the network and self-guidedly generating a new sample group, thereby iteratively training the network to obtain the optimal solution for the design variables and optimizing the machine tool structure. The present invention is applicable to both two-dimensional and three-dimensional designs. By integrating a probability distribution convolution layer in topology optimization under given constraints, and using Gaussian processes and self-guided online learning to replace traditional gradient optimization, the predicted optimal solution converges to the global optimal solution, accelerating the convergence speed, and ultimately achieving machine tool structure optimization.

Description

Technical Field

[0001] The present invention relates to a machine tool structure optimization method, and relates to the technical field related to structural design, and in particular to a machine tool structure optimization method based on self-guided online learning and high-efficiency sampling. Background Art

[0002] Topology optimization is a method for deriving an optimal design within a design domain that satisfies given loads and boundary conditions. In topology optimization, an object is represented as a continuous material distribution model rather than as a traditional discrete solid. This method utilizes computer algorithms and mathematical optimization techniques to automatically explore and find the optimal material distribution to maximize design objectives and minimize unnecessary material usage.

[0003] Currently, topology optimization methods can be broadly divided into two categories: density description models, such as the variable density method and the progressive optimization method, and boundary description models, such as the level set method and the moving deformable component method. Variable density methods are the most widely used due to their intuitive mathematical model, simple implementation, and computational efficiency. While topology optimization offers advantages over conceptual design, most methods are based on finite element analysis, and the numerous design variables involved in the calculation process limit their application due to computational costs.

[0004] Artificial intelligence encompasses systematic technologies that use computers to simulate human behavior. Machine learning, a subset of AI, aims to learn meaningful patterns from data using statistical methods. Deep learning, a subset of machine learning, seeks to improve learning capabilities by training multi-layer neural network structures from the data itself. Combining deep learning techniques with traditional topology optimization methods can achieve more efficient and accurate material distribution prediction and structural optimization. Deep learning models can learn from large amounts of historical data and topology optimization results, extracting hidden features and patterns. This accelerates the optimization process, reduces computational costs, and generates more complex and optimized topologies, enabling more innovative and optimized engineering designs. Deep learning algorithms currently used in topology optimization can be categorized into four main types: supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning. Neural networks fall under the category of supervised learning. Research on using deep learning to improve topology optimization techniques includes accelerating iterations, enabling non-iterative optimization, meta-modeling, reducing the dimensionality of the design space, improving optimizers, and enabling generative design and post-processing.

[0005] Although various deep learning methods have demonstrated their applicability to topology optimization problems in numerous experiments, they have also exposed several limitations during computational processes. First, most current topology optimization methods incorporating deep learning are not scalable; that is, a method is not transferable between 2D and 3D. However, for real-world large-scale problems, application to 3D design is essential. Second, when applying deep learning methods to topology optimization, data acquisition costs and model training efficiency should be considered. Third, because physical information is not accurately reflected, optimization results are predicted based on pixel-to-pixel similarity, and thus the connectivity of the optimization results may not be guaranteed. Fourth, the robustness of neural networks used for topology optimization remains to be verified and improved, and evaluation metrics require further refinement. Finally, computational cost and accuracy are often a trade-off. Although deep learning aims for acceleration, its accuracy is still inferior to that of traditional topology optimization methods. Using finite element analysis to improve accuracy during the intermediate steps still results in high computational cost and increased computational time. Summary of the Invention

[0006] To address the problems mentioned in the background art, the present invention provides a method for machine tool structure optimization based on self-guided online learning and high-efficiency sampling. The method utilizes self-guided online machine learning to build a deep neural network by adding probability distribution relationships to the weight parameters of the convolutional layer. The deep neural network is used to learn the case topology optimization process, gradually adjusting the range of sample points generated based on the set objective function and boundary conditions, achieving self-guided learning of topology optimization during a high-efficiency sampling process.

[0007] The technical solution adopted in the present invention is:

[0008] The machine tool structure optimization method based on self-guided online learning and high-efficiency sampling of the present invention includes:

[0009] 1) Construct a 2D / 3D model of the machine tool components to be optimized and an optimization objective function based on material volume constraints; key machine tool components mainly include the column, spindle, and bed.

[0010] 2) Construct a neural network that integrates a probability distribution convolution layer. The input dimension of the neural network that integrates the probability distribution convolution layer is determined according to the dimension of the two-dimensional / three-dimensional model of the machine tool component and the number of its design variables. At the same time, a loss function including the information divergence value of the neural network that integrates the probability distribution convolution layer is constructed.

[0011] 3) Randomly generate several groups of initial sample points of each design variable in the two-dimensional / three-dimensional model of the machine tool component and obtain several groups of initial sample groups after normalization processing, and input each group of initial sample groups into the optimization objective function to obtain the objective function value of each group of initial sample groups.

[0012] 4) Each group of initial sample groups and their objective function values ​​are input into the neural network fused with the probability distribution convolution layer for training until the loss function converges to obtain the trained neural network fused with the probability distribution convolution layer as the machine tool structure optimization neural network.

[0013] 5) The Gaussian process model and expected expectation acquisition function are used to optimize the machine tool structure optimization neural network, obtain the optimal candidate points of each design variable to form an optimal candidate group, and self-guide to generate a batch of new sample groups near the optimal candidate group.

[0014] 6) Repeat the same operations of the initial sample group in steps 3)-6) on the optimal candidate group and the new sample group, thereby iteratively training the machine tool structure optimization neural network until the preset maximum number of iterations is reached, and then stop the iterative training. The optimal candidate points of each design variable at this time are obtained as the optimal solutions of the design variables of the machine tool components, so as to design the machine tool structure and achieve machine tool structure optimization.

[0015] In the step 1), in the two-dimensional / three-dimensional model, the area where the structure of the machine tool component to be optimized is located is the design domain, which is discretized into a number of finite element nodes. The design variable at each finite element node is a density value, and the density value is 1 or 0.

[0016] In step 1), the optimization objective function based on the material volume constraint is specifically as follows:

[0017]

[0018] Among them, ρ is the density value set of each finite element node, ρ1,ρ2,…,ρ i ,…,ρ N are the density values ​​of the 1st, 2nd, …, i, …, Nth finite element nodes respectively; F(ρ) is the objective function value of the density value set ρ of each finite element node; is the structural flexibility of the machine tool component under the density value set ρ of each finite element node. Under the action of a single solid mechanics field, the objective function is generally defined as the structural flexibility; G0(ρ) is the material volume constraint of the machine tool component under the density value set ρ of each finite element node, w i is the weight of the density value of the i-th finite element node in the overall volume integral, V0 is the preset volume threshold of the machine tool component; other possible constraints in the optimization objective function except the material volume constraint are all less than or equal to zero; S is the value range of the density value of each finite element node, S = [0,1].

[0019] In the step 2), the neural network integrating the probability distribution convolution layer includes m first block structures and a second block structure connected in sequence, the first block structure includes a probability distribution convolution layer, a batch normalization layer and a Softplus activation function connected in sequence, and the second block structure includes a fully connected layer and a probability distribution linear layer connected in sequence; the input of the neural network integrating the probability distribution convolution layer is the density value set ρ at each finite element node, and the output of the neural network integrating the probability distribution convolution layer includes the objective function mean f(ρ)_a, the objective function standard deviation f(ρ)_δ and the information divergence value KL of the density value set ρ at each finite element node.

[0020] The probability distribution convolution layer is specifically as follows:

[0021]

[0022] Among them, b j A is the output of the probability distribution convolution layer in the jth first block structure of the neural network that integrates the probability distribution convolution layer; k is the receptive field of the kth group of initial samples in the neural network of the input fusion probability distribution convolution layer, * is the convolution operation, μ k is the mean of the kth group of initial samples; j is the variance weight parameter of the probability distribution convolution layer in the jth first block structure, ε j Obey the normal distribution of (0, 1), α k is the component parameter of the variance of the kth group of initial samples in the neural network input fusion probability distribution convolution layer, and ⊙ is the component multiplication operation.

[0023] In the step 2), the input of the neural network fused with the probability distribution convolution layer includes a total of four dimensions. The value of the first dimension is the number of initial sample groups of the neural network input fused with the probability distribution convolution layer. When the model constructed by the machine tool component to be optimized is a two-dimensional model, the value of the second dimension is 1, and the values ​​of the third and fourth dimensions are respectively the number of design variables in the length and width dimensions of the two-dimensional model, and the dimensional product is consistent with the total number of design variables; when the model constructed by the machine tool component to be optimized is a three-dimensional model, the values ​​of the second, third and fourth dimensions are respectively the number of design variables in the high length and width dimensions of the two-dimensional model, and the dimensional product is consistent with the total number of design variables.

[0024] In step 2), the loss function including the information divergence value of the neural network fused with the probability distribution convolution layer is specifically as follows:

[0025] loss=loss1+ω·loss2

[0026]

[0027]

[0028] Among them, loss1 and loss2 are the root mean square error loss and information divergence loss respectively; ω is the preset first weight parameter of the neural network fused with the probability distribution convolution layer, which is determined by the order of magnitude difference between the two loss functions; n is the number of initial sample groups in the neural network input into the fused probability distribution convolution layer; f k (ρ) is the target function network prediction value of the kth group of initial samples, F k (ρ) is the objective function value of the kth group of initial samples obtained in the input optimization objective function; KL is the information divergence value; q θ is the approximate distribution, w is the preset second weight parameter of the neural network fused with the probability distribution convolution layer, D is the distribution parameter, p is the prior distribution, and || is the symbol of the conditional probability distribution.

[0029] The first part of the loss function, loss1, is the mean square error between the neural network prediction value and the finite element calculation result value of the training sample point; the second part of the loss function, loss2, is the information divergence value KL of the network output of the convolutional layer with the probability distribution added. The variational inference method Bayes by backprop is used here, which uses an approximate distribution and the information divergence value to make the network parameters closer and closer to the true posterior distribution.

[0030] In step 5), the Gaussian process model and the expected expectation acquisition function are used to optimize the machine tool structure optimization neural network, as follows:

[0031] 5.1) Obtain the objective function mean f(ρ)_a and objective function standard deviation f(ρ)_δ of the density value set ρ at each finite element node output by the machine tool structure optimization neural network in step 4), thereby constructing a Gaussian process model.

[0032] 5.2) Use the expected improvement (EI) acquisition function to perform a global minimum search on the Gaussian process model to obtain the current optimal candidate group.

[0033] 5.3) Determine whether the current optimized machine tool structure optimization neural network converges or reaches a preset iteration number threshold. If so, output the current optimal candidate group as the final optimal candidate group of the design variables, that is, the final optimal density value of each finite element node. If not, add the current optimal candidate group to each group of initial sample groups as the input of the current optimized machine tool structure optimization neural network, and repeat the same operations of the machine tool structure optimization neural network in steps 5.1)-5.3) until the final optimal candidate group of the design variables is obtained.

[0034] When the difference between the objective function network prediction value of the optimal candidate group predicted by the optimized machine tool structure optimization neural network and the objective function value of the optimal candidate group obtained according to the optimized objective function is less than the preset difference threshold, it is judged that the optimized machine tool structure optimization neural network has converged.

[0035] In the step 5), disturbances are added near the optimal candidate group, that is, according to the genetic algorithm, disturbances such as mutation, crossover and convolution are added to the optimal candidate group, thereby obtaining several new sample groups.

[0036] Adding perturbations near the optimal candidate group allows the algorithm to increase the number of sample points based on the previous training sample points. Referring to the idea of ​​genetic algorithm, the added perturbations can be classified into mutation, crossover and convolution; mutation means replacing one or more design variables with random numbers for this set of optimal solutions, that is, changing the density value of one or several finite element nodes, crossover means exchanging several elements in this set of design variables, that is, exchanging the density values ​​of several finite element nodes, and convolution means applying a given convolution kernel to this set of design variables, that is, performing a convolution operation on the density values ​​of each finite element node in one of the areas.

[0037] An electronic device of the present invention includes: a memory and a processor coupled to each other, wherein the memory stores program data, and the processor calls the program data to execute the method described above.

[0038] The beneficial effects of the present invention are:

[0039] By adjusting the network input dimension, the present invention can be applied to both two-dimensional design and three-dimensional design, and can significantly shorten the optimization time while ensuring the design accuracy, without being limited by the design dimension. The present invention converts the design variables in the topology optimization problem into pixel points in the image detection problem by integrating the probability distribution convolution layer in the topology optimization under given constraints, and uses the convolution kernel of the convolution layer to make the connection between the design variables closer, thereby improving the calculation accuracy and efficiency of the topology optimization; by replacing the traditional gradient optimization method with the Gaussian process model and the expected expectation acquisition function, the network optimization can overcome the problem of falling into the local minimum, thereby making the method applicable to large-scale complex problems. By integrating self-guided online learning, the present invention can accurately generate a new batch of sample points in the area of ​​interest of the current problem, thereby reducing the amount of model calculation, making the predicted optimal solution converge to the global optimal solution, accelerating the convergence speed, and ultimately achieving machine tool structure optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Flowchart of the machine tool structure optimization method of the present invention;

[0041] Figure 2This is a diagram of the neural network structure that integrates the probability distribution convolution layer;

[0042] Figure 3 This is a flowchart of the neural network optimization method used;

[0043] Figure 4 An example diagram of the perturbation form used to generate new sample points for self-guided generation;

[0044] Figure 5 This is a before-and-after comparison of the structural topology optimization results for a specific machine tool component implementation case. Figure 5 (a) is a schematic diagram of a three-dimensional machine tool column implementation case. Figure 5 (b) is a schematic diagram of the structural topology optimization design variable division for the implementation case of a three-dimensional machine tool column. Figure 5 (c) is a schematic diagram of the structural topology optimization results of the three-dimensional machine tool column implementation case. DETAILED DESCRIPTION

[0045] Exemplary embodiments will be described in detail herein, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, like numbers in different figures represent like or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all possible embodiments consistent with the present invention. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present invention, as detailed in the appended claims.

[0046] The terms used in this invention are for the purpose of describing specific embodiments only and are not intended to limit the invention. The singular forms "a," "the," and "the" used in this invention and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0047] It should be understood that although the terms first, second, third, etc. may be used in the present invention to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other.

[0048] like Figure 1 As shown, the machine tool structure optimization method based on self-guided online learning and high-efficiency sampling of the present invention is specifically as follows:

[0049] 1) Construct a 2D / 3D model of the machine tool component to be optimized and an optimization objective function based on material volume constraints. Key machine tool components primarily include the column, spindle, and bed. Within the 2D / 3D model, the region containing the component's structure is the design domain, which is discretized into a number of finite element nodes. The design variables at each finite element node are density values, which are either 1 or 0.

[0050] The optimization objective function based on material volume constraint is as follows:

[0051]

[0052] Among them, ρ is the density value set of each finite element node, ρ1,ρ2,…,ρ i ,…,ρ N are the density values ​​of the 1st, 2nd, …, i, …, Nth finite element nodes respectively; F(ρ) is the objective function value of the density value set ρ of each finite element node; is the structural flexibility of the machine tool component under the density value set ρ of each finite element node. Under the action of a single solid mechanics field, the objective function is generally defined as the structural flexibility; G0(ρ) is the material volume constraint of the machine tool component under the density value set ρ of each finite element node, w i is the weight of the density value of the i-th finite element node in the overall volume integral, V0 is the preset volume threshold of the machine tool component; other possible constraints in the optimization objective function except the material volume constraint are all less than or equal to zero; S is the value range of the density value of each finite element node, S = [0,1].

[0053] 2) Construct a neural network that integrates a probability distribution convolution layer. The input dimension of the neural network that integrates the probability distribution convolution layer is determined according to the dimension of the two-dimensional / three-dimensional model of the machine tool component and the number of its design variables. At the same time, a loss function including the information divergence value of the neural network that integrates the probability distribution convolution layer is constructed.

[0054] like Figure 2 As shown, the neural network fused with the probability distribution convolution layer includes m first block structures and a second block structure connected in sequence. The first block structure includes a probability distribution convolution layer, a batch normalization layer and a Softplus activation function connected in sequence, and the second block structure includes a fully connected layer and a probability distribution linear layer connected in sequence. The input of the neural network fused with the probability distribution convolution layer is the density value set ρ at each finite element node, and the output of the neural network fused with the probability distribution convolution layer includes the objective function mean f(ρ)_a, the objective function standard deviation f(ρ)_δ and the information divergence value KL of the density value set ρ at each finite element node.

[0055] The probability distribution convolution layer is as follows:

[0056]

[0057] Among them, b j A is the output of the probability distribution convolution layer in the jth first block structure of the neural network that integrates the probability distribution convolution layer; k is the receptive field of the kth group of initial samples in the neural network of the input fusion probability distribution convolution layer, * is the convolution operation, μ k is the mean of the kth group of initial samples; j is the variance weight parameter of the probability distribution convolution layer in the jth first block structure, ε j Obey the normal distribution of (0, 1), α k is the component parameter of the variance of the kth group of initial samples in the neural network input fusion probability distribution convolution layer, and ⊙ is the component multiplication operation.

[0058] The input of the neural network fused with the probability distribution convolution layer includes a total of four dimensions. The value of the first dimension is the number of initial sample groups of the neural network input fused with the probability distribution convolution layer. When the model constructed by the machine tool component to be optimized is a two-dimensional model, the value of the second dimension is 1, and the values ​​of the third and fourth dimensions are the number of design variables in the length and width dimensions of the two-dimensional model, respectively. The product of the dimensions is consistent with the total number of design variables; when the model constructed by the machine tool component to be optimized is a three-dimensional model, the values ​​of the second, third and fourth dimensions are the number of design variables in the high length and width dimensions of the two-dimensional model, respectively. The product of the dimensions is consistent with the total number of design variables.

[0059] The loss function of the neural network fused with the probability distribution convolution layer, including the information divergence value, is as follows:

[0060] loss=loss1+ω·loss2

[0061]

[0062]

[0063] Among them, loss1 and loss2 are the root mean square error loss and information divergence loss respectively; ω is the preset first weight parameter of the neural network fused with the probability distribution convolution layer, which is determined by the order of magnitude difference between the two loss functions; n is the number of initial sample groups in the neural network input into the fused probability distribution convolution layer; f k (ρ) is the target function network prediction value of the kth group of initial samples, F k (ρ) is the objective function value of the kth group of initial samples obtained in the input optimization objective function; KL is the information divergence value; q θis the approximate distribution, w is the preset second weight parameter of the neural network fused with the probability distribution convolution layer, D is the distribution parameter, p is the prior distribution, and || is the symbol of the conditional probability distribution.

[0064] The first part of the loss function, loss1, is the mean square error between the neural network prediction value and the finite element calculation result value of the training sample point; the second part of the loss function, Loss2, is the information divergence value KL of the network output of the convolutional layer with the probability distribution added. The variational inference method of Bayes by backprop is used here, which uses an approximate distribution and the information divergence value to make the network parameters closer and closer to the true posterior distribution.

[0065] 3) Randomly generate several groups of initial sample points of each design variable in the two-dimensional / three-dimensional model of the machine tool component and obtain several groups of initial sample groups after normalization processing, and input each group of initial sample groups into the optimization objective function to obtain the objective function value of each group of initial sample groups.

[0066] 4) Each group of initial sample groups and their objective function values ​​are input into the neural network fused with the probability distribution convolution layer for training until the loss function converges to obtain the trained neural network fused with the probability distribution convolution layer as the machine tool structure optimization neural network.

[0067] 5) Using the Gaussian process model and the expected expectation acquisition function to optimize the machine tool structure optimization neural network, the optimal candidate points of each design variable are obtained to form an optimal candidate group, and a batch of new sample groups are generated by self-guiding near the optimal candidate group. Using the Gaussian process model and the expected expectation acquisition function to optimize the machine tool structure optimization neural network is as follows:

[0068] 5.1) Obtain the objective function mean f(ρ)_a and objective function standard deviation f(ρ)_δ of the density value set ρ at each finite element node output by the machine tool structure optimization neural network in step 4), thereby constructing a Gaussian process model.

[0069] 5.2) Use the expected EI acquisition function to perform global minimum search and optimization on the Gaussian process model to obtain the current optimal candidate group.

[0070] 5.3) Determine whether the current optimized machine tool structure optimization neural network converges or reaches a preset iteration number threshold. If so, output the current optimal candidate group as the final optimal candidate group of the design variables, that is, the final optimal density value of each finite element node. If not, add the current optimal candidate group to each group of initial sample groups as the input of the current optimized machine tool structure optimization neural network, and repeat the same operations of the machine tool structure optimization neural network in steps 5.1)-5.3) until the final optimal candidate group of the design variables is obtained.

[0071] When the difference between the objective function network prediction value of the optimal candidate group predicted by the optimized machine tool structure optimization neural network and the objective function value of the optimal candidate group obtained according to the optimized objective function is less than the preset difference threshold, it is judged that the optimized machine tool structure optimization neural network has converged.

[0072] Add perturbations near the optimal candidate group, that is, according to the genetic algorithm, add perturbations of mutation, crossover and convolution to the optimal candidate group, so as to obtain several groups of new sample groups.

[0073] Adding perturbations near the optimal candidate group allows the algorithm to increase the number of sample points based on the previous training sample points. Referring to the idea of ​​genetic algorithm, the added perturbations can be classified into mutation, crossover and convolution; mutation means replacing one or more design variables with random numbers for this set of optimal solutions, that is, changing the density value of one or several finite element nodes, crossover means exchanging several elements in this set of design variables, that is, exchanging the density values ​​of several finite element nodes, and convolution means applying a given convolution kernel to this set of design variables, that is, performing a convolution operation on the density values ​​of each finite element node in one of the areas.

[0074] 6) Repeat the same operations of the initial sample group in steps 3)-6) on the optimal candidate group and the new sample group, thereby iteratively training the machine tool structure optimization neural network until the preset maximum number of iterations is reached, and then stop the iterative training. The optimal candidate points of each design variable at this time are obtained as the optimal solutions of the design variables of the machine tool components, so as to design the machine tool structure and achieve machine tool structure optimization.

[0075] The specific implementation of the present invention will be described more completely and clearly below with reference to the accompanying drawings and specific embodiments.

[0076] like Figure 1 As shown in FIG, the machine tool structure optimization method based on self-guided online learning and high-efficiency sampling of the present invention includes the following steps:

[0077] 1) Determine the machine tool components to be topologically optimized, the optimization objective function, the material volume constraints, and the corresponding design variables, specifically:

[0078] like Figure 5 (a) and Figure 5As shown in (b), taking the three-dimensional machine tool column as a specific example, the rectangular design domain contained therein can be evenly divided into 8×8×20 grids. Assume that the bottom surface A of the column is the bottom fixed surface, that is, a fixed constraint. In addition to gravity, the model is subjected to the vertical downward gravity B from the spindle box and the upward force brought by the workpiece. Taking the above working condition as an example, the load-bearing surface of the column is simplified to the load-bearing surface of the above working condition, consisting of the C and D load lines and the E load line. The bending moment of the E load line is counterclockwise, the C load line is at the lower edge of the E load line, and the D load line is at the upper edge of the E load line. The proportion of external forces on the C and D load lines is topologically optimized using enumeration. Finally, the working condition of external force on the D load line: external force on the C load line = 2:8 is selected. The resulting topological structure has the smallest deformation. At this time, there are 9×9×21=1701 design variables in the design domain to control the material distribution, that is, the node density ρ i (i=1,2,…,1701). The objective function of this case is to minimize the structural flexibility. The optimization objective function is set as follows:

[0079]

[0080] Where ρ0=[0.5,0.5,…,0.5] T , dimensionless elastic energy It is defined as the ratio of the structural elastic energy E(ρ) of any given material distribution to the elastic energy E(ρ0) of the reference uniform distribution. The elastic energy is calculated by finite element method using Young's modulus in the design domain. The Young's modulus of each node is density-dependent, so the simplified isotropic penalized material method Y(ρ(x))=Y0ρ(x) in the variable density method should be used. 3 +ε[1-ρ(x) 3 ] is calculated, where Y and Y0 are variables and constants, respectively. ρ(x) is the material density at a given location x, obtained by linear interpolation of the nodal values ​​of the element. ε is a small number used to avoid numerical singularities. The material volume constraint is w·ρ≤0.5, where w is the vector of the linear Gaussian quadrature.

[0081] 2) Determine the model input dimension based on the case dimension and the number of design variables, and build a neural network that integrates the probability distribution convolution layer accordingly. Specifically:

[0082] The input dimension of the model is determined. First, the first dimension is the number of training sample points. The initial number is set to 100. The number of sample points increases by 100 with each additional round, that is, dimension 1 increases by 100. The last three dimensions are consistent with the number of design variables of a single layer in the height, length and width directions corresponding to the model, that is, the second, third and fourth dimensions are 21, 9 and 9 respectively. Therefore, the initial input dimension of the model is 100×21×9×9. During the training process, the first dimension will increase with the increase of the data set and become n×21×9×9.

[0083] Then build a neural network framework that integrates the probability distribution convolution layer as follows Figure 2 As shown in the figure, the network mainly consists of Block Structure 1 and Block Structure 2. Block Structure 1 consists of three parts: a probability distribution convolution layer, a batch normalization layer, and a Softplus activation function layer. Block Structure 2 consists of two parts: a fully connected layer and a probability distribution linear layer, and is used before the network output. For the 3D machine tool column case, the network includes 10 Block Structure 1s and Block Structure 2, which is used before the output. The convolution kernel of the probability distribution convolution layer in Block Structure 1 is always 3. The output of the neural network consists of three parts: the objective function mean f(ρ)_a, the objective function standard deviation f(ρ)_δ, and the information divergence value KL.

[0084] The probability distribution convolution layer includes two sequence convolutions. The output of the two sequence convolutions contains the mean and variance. In the first convolution operation A k *μ k In the example, the output b is regarded as the output of the convolutional neural network updated by frequent inference, and the mean μ is learned using the Adam optimizer. k ; In the second convolution operation In, variance Will be learned, since the variance here is also a function of the mean, so only α needs to be learned in the second convolution operation k , which ensures that only one parameter needs to be updated for each convolution operation.

[0085] At the same time, the network loss function including the mean square error between the predicted value and the true value and the information divergence value is constructed.

[0086] 3) Randomly generate 100 sets of initial sample points that satisfy volume constraints for each design variable in the 3D model of the machine tool component and normalize them so that the dimensions are consistent with the number of design variables. Perform finite element calculations on the design variables corresponding to the sample points to obtain the corresponding structural flexibility values, i.e., the objective function values.

[0087] 4) Use the Gaussian process model and the expected acquisition function to optimize the trained network, specifically:

[0088] The global minimum of the trained neural network is searched. The mapping relationship between the input and output of the neural network is the target function to be searched. Consider constructing a Gaussian process and using the expected acquisition function to search for the global minimum, as follows:

[0089]

[0090]

[0091] Where EI(ρ) is the acquisition function value between the current Gaussian process model and the current optimal observation group; μ(ρ) is the posterior expected value of the Gaussian process model constructed based on the current sample data set; The current optimal observation group The corresponding optimal observation value, Φ(Z) and φ(Z) are the cumulative distribution function and probability density function of the standard normal distribution, respectively. Z is the variable obtained by normalizing the Gaussian process distribution based on the current optimal observation value; σ(ρ) is the posterior standard deviation of the Gaussian process model constructed based on the current sample data set.

[0092] The structure of the neural network optimization method used is as follows Figure 3 As shown, the details are as follows:

[0093] a) Randomly generate 10 initial training groups. The design variables in the initial training groups all have values ​​in the range [0, 1]. Perform the same data normalization process as in the network training process to obtain the training data set, so that the value range is adjusted to [-1, 1].

[0094] b) The predicted mean and standard deviation of the training data set are obtained through the trained network model, and a Gaussian process model FixedNoiseGP with a fixed noise level is constructed.

[0095] c) Calculate the constructed Gaussian process model using the EI acquisition function and find the optimal candidate group based on the calculated values.

[0096] d) Determine whether convergence or reaching a preset iteration threshold. If so, output the optimal candidate group and its corresponding model output value. If not, add the newly sampled candidate group in step c) to the training data set and jump to step b).

[0097] 5) A new batch of sample points is generated near the global minimum value obtained by the optimization algorithm, and the next round of network training is continued until the maximum number of iterations is reached. Specifically:

[0098] Add perturbations near the set of design variables corresponding to the global optimal solution of the trained network, so that the algorithm increases the number of sample points by 100 on the basis of the previous training sample points. Referring to the idea of ​​genetic algorithm, the added perturbations can be classified as mutation, crossover and convolution; mutation means replacing one or more design variables with random numbers for this set of optimal solutions, crossover means exchanging several elements in this set of design variables, and convolution means applying a given convolution kernel to this set of design variables. The volume constraints of the design variables of the newly generated sample points can be enforced in the next step. Examples of perturbations used in self-guided generation of new sample points are as follows: Figure 4 shown.

[0099] When the network converges or reaches the maximum number of iterations, the design variables corresponding to the global optimal solution obtained by the optimization algorithm are output, and the topology optimization structure can be constructed according to the set density filtering threshold (for example, if the threshold is 0.5, the area representing the density greater than 0.5 is taken as 1, otherwise it is taken as 0), such as Figure 5 As shown in (c), the density value at the blank position is 0, and the density value at the non-blank position is 1.

[0100] The above embodiments are intended only to illustrate the design concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. The scope of protection of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design concepts disclosed in the present invention are within the scope of protection of the present invention.

Claims

1. A machine tool structure optimization method based on self-guided online learning and high-efficiency sampling, characterized in that: include: 1) Construct a 2D / 3D model of the machine tool component to be optimized and an optimization objective function based on material volume constraints; 2) Constructing a neural network that integrates a probability distribution convolution layer, wherein the input dimension of the neural network that integrates the probability distribution convolution layer is determined according to the dimension of the two-dimensional / three-dimensional model of the machine tool component and the number of design variables, and constructing a loss function of the neural network that integrates the probability distribution convolution layer, including an information divergence value; 3) randomly generating several groups of initial sample points of each design variable in the two-dimensional / three-dimensional model of the machine tool component and performing normalization processing to obtain several groups of initial sample groups, and inputting each group of initial sample groups into the optimization objective function to obtain the objective function value of each group of initial sample groups; 4) Inputting each group of initial sample groups and their objective function values ​​into the neural network fused with the probability distribution convolution layer for training until the loss function converges, the trained neural network fused with the probability distribution convolution layer is obtained as the machine tool structure optimization neural network; 5) Using the Gaussian process model and the expected expectation acquisition function to optimize the machine tool structure optimization neural network, the optimal candidate points of each design variable are obtained to form an optimal candidate group, and a batch of new sample groups are generated by self-guiding near the optimal candidate group; 6) Repeat the same operations of the initial sample group in steps 3)-6) on the optimal candidate group and the new sample group, thereby iteratively training the machine tool structure optimization neural network until the preset maximum number of iterations is reached, and then stop the iterative training. The optimal candidate points of each design variable at this time are obtained as the optimal solutions of the design variables of the machine tool components, so as to design the machine tool structure and achieve machine tool structure optimization.

2. The machine tool structure optimization method based on self-guided online learning and high-efficiency sampling according to claim 1 is characterized in that: In the step 1), in the two-dimensional / three-dimensional model, the area where the structure of the machine tool component to be optimized is located is the design domain, which is discretized into a number of finite element nodes. The design variable at each finite element node is a density value, and the density value is 1 or 0.

3. The machine tool structure optimization method based on self-guided online learning and high-efficiency sampling according to claim 2 is characterized in that: In step 1), the optimization objective function based on the material volume constraint is specifically as follows: Among them, ρ is the density value set of each finite element node, ρ1,ρ2,…,ρ i ,…,ρ N are the density values ​​of the 1st, 2nd, …, i, …, Nth finite element nodes respectively; F(ρ) is the objective function value of the density value set ρ of each finite element node; is the structural flexibility of the machine tool component under the density value set ρ of each finite element node; G0(ρ) is the material volume constraint of the machine tool component under the density value set ρ of each finite element node, w i is the weight of the density value of the i-th finite element node in the overall volume integral, V0 is the preset volume threshold of the machine tool component; S is the value range of the density value of each finite element node.

4. The machine tool structure optimization method based on self-guided online learning and high-efficiency sampling according to claim 2 is characterized in that: In the step 2), the neural network integrating the probability distribution convolution layer includes m first block structures and a second block structure connected in sequence, the first block structure includes a probability distribution convolution layer, a batch normalization layer and a Softplus activation function connected in sequence, and the second block structure includes a fully connected layer and a probability distribution linear layer connected in sequence; The input of the neural network fused with the probability distribution convolution layer is the density value set ρ at each finite element node, and the output of the neural network fused with the probability distribution convolution layer includes the objective function mean f(ρ)_a, the objective function standard deviation f(ρ)_δ and the information divergence value KL of the density value set ρ at each finite element node.

5. The machine tool structure optimization method based on self-guided online learning and high-efficiency sampling according to claim 4 is characterized in that: The probability distribution convolution layer is specifically as follows: Among them, b j A is the output of the probability distribution convolution layer in the jth first block structure of the neural network that integrates the probability distribution convolution layer; k is the receptive field of the kth group of initial samples in the neural network of the input fusion probability distribution convolution layer, * is the convolution operation, μ k is the mean of the kth group of initial samples; j is the variance weight parameter of the probability distribution convolution layer in the jth first block structure, α k is the component parameter of the variance of the kth group of initial samples in the neural network input fusion probability distribution convolution layer, and ⊙ is the component multiplication operation.

6. The machine tool structure optimization method based on self-guided online learning and high-efficiency sampling according to claim 2 is characterized in that: In the step 2), the input of the neural network fused with the probability distribution convolution layer includes a total of four dimensions. The value of the first dimension is the number of initial sample groups of the neural network input fused with the probability distribution convolution layer. When the model constructed by the machine tool component to be optimized is a two-dimensional model, the value of the second dimension is 1, and the values ​​of the third and fourth dimensions are respectively the number of design variables in the length and width dimensions of the two-dimensional model; when the model constructed by the machine tool component to be optimized is a three-dimensional model, the values ​​of the second dimension, the third dimension and the fourth dimension are respectively the number of design variables in the high length and width dimensions of the two-dimensional model.

7. The machine tool structure optimization method based on self-guided online learning and high-efficiency sampling according to claim 2 is characterized in that: In step 2), the loss function including the information divergence value of the neural network fused with the probability distribution convolution layer is specifically as follows: loss=loss1+ω·loss2 loss2=KL[q θ (w|D)||p(w)] Among them, loss1 and loss2 are the root mean square error loss and information divergence loss respectively; ω is the preset first weight parameter of the neural network fused with the probability distribution convolution layer; n is the number of initial sample groups input into the neural network fused with the probability distribution convolution layer; f k (ρ) is the target function network prediction value of the kth group of initial samples, F k (ρ) is the objective function value of the kth group of initial samples obtained in the input optimization objective function; KL is the information divergence value; q θ is the approximate distribution, w is the preset second weight parameter of the neural network fused with the probability distribution convolution layer, D is the distribution parameter, and p is the prior distribution.

8. The machine tool structure optimization method based on self-guided online learning and high-efficiency sampling according to claim 4 is characterized in that: In step 5), the Gaussian process model and the expected expectation acquisition function are used to optimize the machine tool structure optimization neural network, as follows: 5.1) Obtain the objective function mean f(ρ)_a and objective function standard deviation f(ρ)_δ of the density value set ρ at each finite element node output by the machine tool structure optimization neural network in step 4), thereby constructing a Gaussian process model; 5.2) Using the expected EI acquisition function to perform global minimum search on the Gaussian process model, the optimal candidate group is obtained; 5.3) Determine whether the current optimized machine tool structure optimization neural network has converged or reached a preset iteration number threshold. If so, output the current optimal candidate group as the final optimal candidate group of design variables, that is, the final optimal density value of each finite element node. If not, add the current optimal candidate group to each group of initial sample groups and use it as the input of the current optimized machine tool structure optimization neural network. Repeat the same operations of the machine tool structure optimization neural network in steps 5.1)-5.3) until the final optimal candidate group of design variables is obtained. When the difference between the objective function network prediction value of the optimal candidate group predicted by the optimized machine tool structure optimization neural network and the objective function value of the optimal candidate group obtained according to the optimized objective function is less than the preset difference threshold, it is judged that the optimized machine tool structure optimization neural network has converged.

9. The machine tool structure optimization method based on self-guided online learning and high-efficiency sampling according to claim 1, characterized in that: In the step 5), disturbances are added near the optimal candidate group, that is, according to the genetic algorithm, disturbances such as mutation, crossover and convolution are added to the optimal candidate group, thereby obtaining several new sample groups.

10. An electronic device, characterized in that: include: A memory and a processor coupled to each other, wherein the memory stores program data, and the processor calls the program data to execute the method according to any one of claims 1 to 9.

Images