Optimization design method for complex part profile based on multi-field node feature weighted aggregation

Abstract

The application discloses a complex part contour optimization design method based on multi-field node feature weighted aggregation. First, according to the complex part contour geometry design variables, temperature and contact surface friction coefficient to be optimized, a batch of finite element simulations are carried out to obtain a training data set. Then, a deep map sampling aggregation neural network fusing space and temperature attention is constructed, and then the deep map sampling aggregation neural network fusing space and temperature attention is trained by using the training data set. After the training is completed, a stress prediction model is obtained. Finally, combined with the stress prediction model, an iterative optimization is carried out to minimize the maximum equivalent stress of the complex part as an optimization objective, and the optimal geometric parameters of the current complex part contour in the target temperature interval are obtained. The application realizes the rapid reconstruction of the stress distribution state of the complex part, improves the optimization speed while fully guaranteeing the accuracy of the optimization result, and greatly saves the calculation resources.

Description

TECHNICAL FIELD

[0001] The application relates to an optimization design method for a complex part profile, in particular to a complex part profile optimization design method based on multi-field node feature weighted aggregation. BACKGROUND

[0002] In the industrial field, as a core component of mechanical and electronic equipment, aerospace vehicles and other industrial products, the performance of complex parts directly affects the reliability, work efficiency and service life of the entire equipment and system. As the most important structural feature of complex parts, the profile not only determines the contour of the part itself, but also directly affects the transmission performance and mechanical performance between parts.

[0003] In addition, due to the combined action of long-term high temperature and high pressure environment and various loads, the profile as the working area of the cooperation between parts is prone to stress concentration due to frequent contact and extrusion. Such stress concentration can easily lead to fatigue failure and even fracture of the part, and thus cause extremely serious safety accidents. Therefore, in order to effectively alleviate stress concentration and improve the stability and reliability of complex part work, it is very important to optimize the profile design of complex parts.

[0004] The optimization design of the profile of the complex part is to find a set of profile design variables that meet the requirements of the objective function under certain constraints, that is, the optimal solution. At the same time, since the input (geometric parameters) and output (stress and strain values) are implicitly expressed, the optimal solution cannot be simply obtained according to the gradient information, so an optimization algorithm is needed to optimize and solve the design problem.

[0005] At present, there are mainly two methods for the optimization design of the profile of the complex part. One is to combine finite element analysis with optimization algorithm, which regards the finite element analysis software as the relationship model between the design variables and the objective function, and drives the finite element analysis software to perform iterative calculation and optimization through commercial optimization software or optimization control program. With the help of the optimization algorithm, the value of the design variable is constantly updated until the final convergence is completed, so as to obtain the optimal profile parameter design value of the complex part. However, the solution of the optimization problem of the profile of the complex part involves complex processes such as modeling of the research object, thermal-structural coupling finite element simulation analysis and post-processing of the results, and a large amount of computing resources and time are needed to realize a complete analysis. Therefore, when establishing the optimization design model of the profile of the complex part, an approximate model of the related index is often used instead of the actual simulation process to ensure the smooth progress of the optimization process.

[0006] With the wide application of data analysis, the method of combining proxy model and optimization algorithm has gradually become the mainstream of the current hull form optimization design, that is, first calculate a group of data through parameterized modeling, then build a proxy model, and then establish the mapping relationship from input to output, and then optimize through intelligent optimization algorithm. The commonly used proxy models are response surface model, Kriging model, support vector regression and neural network, etc. With the continuous development of deep learning technology, using various neural networks as proxy models to realize the prediction from input to output has gradually become the mainstream. In the offline stage, the input samples are used to train the neural network model, and in the online stage, the test samples are quickly predicted, which greatly improves the speed of optimization design and saves the system overhead.

[0007] Unlike traditional graph neural networks, which require operating on the features of all nodes in the graph during model training, GraphSAGE can randomly sample the information of neighboring nodes by iteratively aggregating the features of neighboring nodes, and through the aggregation function of each layer, the information is non-linearly processed with its own features to form a new feature vector. In addition, GraphSAGE uses a small batch multi-round training method, which effectively solves the problem of solution failure caused by hardware computing power limitations and reduces the dependence of traditional graph convolutional neural networks on full graph information, making it particularly suitable for processing large graph structures with a large number of nodes and complex adjacency relationships, and therefore very suitable for predicting finite element grid node features.

[0008] The stress field prediction of complex parts is to predict the stress values of different grid nodes given the design variables (geometric parameters), and to reconstruct the stress field. However, only establishing a relationship model from geometric parameters to stress field cannot reflect the adaptability and stability of the prediction results under variable working conditions. For complex parts, the working conditions under actual working conditions will inevitably change. Complex parts often work in high temperature and high pressure environments, so it is necessary to consider temperature as a major factor affecting the stress field distribution when considering the stress results. Therefore, considering the temperature field distribution that significantly affects the stress field distribution, restoring the temperature value of each node and forming a relationship model considering the influence of multiple fields under variable temperature conditions has important research significance for the optimization design of the hull form.

[0009] Meanwhile, since the nodes of the finite element grid are interrelated and influence each other, the traditional GraphSAGE algorithm is difficult to fully mine the role relationship between the nodes and fully characterize the role relationship between the nodes, thereby leading to a not very ideal prediction and optimization effect. Moreover, since the contact area of the complex part frequently rubs and extrudes, and the area with large temperature difference is prone to concentrate thermal stress, more attention needs to be paid to the above area, and the features of the nodes in the area should also be fully considered when the graph neural network is built, otherwise the accuracy of the relationship model will be affected due to the unreasonable allocation of the attention degree of the complex part in different areas, thereby the optimization process has certain limitations. SUMMARY

[0010] In order to solve the problems in the background art, the present application proposes a complex part profile optimization design method based on multi-field node feature weighted aggregation. The method proposed by the present application can overcome the shortcomings of the existing methods, consider the influence of the combined action of multiple fields including temperature field and stress field, strain field and other structure fields on stress prediction and optimization results, realize the weighted aggregation of multi-field node features by introducing spatial attention coefficients and temperature attention coefficients, and distinguish the importance degree of the grid nodes. Higher attention is paid to the grid node information of the frequently contacted part and the large temperature difference part. At the same time, the regional maximum temperature and the regional minimum temperature are introduced as the input items of the model, the uncertainty caused by the temperature is considered, and the fast prediction of the stress and strain field distribution of the complex part in the variable temperature interval is realized. At the same time, the improved scarab beetle optimization algorithm based on the optimal Latin hypercube is used for the optimization design of the profile parameters of the complex part, which enriches the diversity of the population and improves the accuracy of the optimization results.

[0011] To achieve the above purpose, the technical scheme of the present application is as follows:

[0012] I. A complex part profile optimization design method based on multi-field node feature weighted aggregation

[0013] S1: According to the complex part profile geometric design variables to be optimized, the temperature and the friction coefficient of the contact surface, finite element simulation is carried out to obtain the feature matrix, the adjacency matrix and the equivalent stress value and the equivalent strain value of each grid node, and then the training data set is obtained; wherein the geometric design variables and the friction coefficient of the contact surface form the geometric parameters;

[0014] S2: A deep graph sampling aggregation neural network fusing spatial and temperature attention is constructed, and then the deep graph sampling aggregation neural network fusing spatial and temperature attention is trained by using the training data set. After the training is completed, a stress prediction model is obtained;

[0015] S3: combine the stress prediction model, and perform iterative optimization with the optimization target of minimizing the maximum equivalent stress of the complex part to obtain the optimal geometric parameters of the current complex part profile in the target temperature interval.

[0016] In S1, the temperature is specifically the maximum temperature and the minimum temperature in the target region.

[0017] In S2, the aggregation function in the depth map sampling aggregation neural network that fuses spatial and temperature attention satisfies the following formula:

[0018]

[0019] wherein, is the jth feature formed by the center node e after aggregation through its i neighbor nodes, (ne) i is the ith neighbor node of the center node e, L_att[(ne) i ] is a spatial attention coefficient, T_att[(ne) i ] is a temperature attention coefficient, and W is an adjustable weight parameter, is the jth feature of the ith neighbor node, and softmax[] is a normalized exponential function, is the spatial distance between the ith neighbor node and the center node, is the temperature difference between the ith neighbor node and the center node e, i = 1, …, n, and n is the number of neighbor nodes of the center node e.

[0020] In S3, the network parameters of the stress prediction model are frozen, and the stress prediction model is used as a fitness function in the optimization algorithm to obtain an improved optimization algorithm. The improved optimization algorithm performs iterative updating of the geometric parameters with the optimization target of minimizing the maximum equivalent stress of the complex part until the optimal geometric parameters of the current complex part profile in the target temperature interval are obtained.

[0021] S3 specifically includes:

[0022] First, the initial parameters of the scarab algorithm are determined; then the optimal Latin hypercube function is used to initialize the scarab population distribution, so that the initial scarab population is uniformly distributed in the entire optimization space, thereby obtaining an initialized scarab individual position matrix; then, the stress prediction model with frozen network parameters is used as the fitness function of the scarab algorithm to obtain an improved scarab algorithm; then, based on the initialized scarab individual position matrix, the improved scarab algorithm is used to iteratively update the scarab population position until the convergence condition is finally met, thereby obtaining the optimal geometric parameters of the current complex part profile in the target temperature interval.

[0023] II. A complex part profile optimization design system based on multi-field node feature weighted aggregation

[0024] A data set construction unit is configured to perform finite element simulation according to the complex part profile geometric design variables to be optimized, temperature and contact surface friction coefficient, to obtain a feature matrix, an adjacency matrix and equivalent stress values and equivalent strain values of each grid node, and to obtain a training data set after repeated multiple times.

[0025] A stress prediction unit is configured to store a depth map sampling aggregation neural network fusing spatial and temperature attention, and to train the depth map sampling aggregation neural network fusing spatial and temperature attention by using the training data set, so as to obtain a stress prediction model.

[0026] A geometric parameter optimization unit is configured to combine the stress prediction model, to perform iterative optimization with the minimum complex part equivalent stress as an optimization target, and to obtain optimal geometric parameters of the current complex part profile in the target temperature interval.

[0027] Three, a computer device

[0028] The device comprises a memory and a processor, the memory stores a computer program, and the processor realizes the steps of the complex part profile optimization design method based on multi-field node feature weighted aggregation when executing the computer program.

[0029] Four, a computer readable storage medium

[0030] The medium stores a computer program, and the computer program realizes the steps of the complex part profile optimization design method based on multi-field node feature weighted aggregation when executed by a processor.

[0031] Five, a computer program product

[0032] The product comprises a computer program / instruction, which realizes the steps of the complex part profile optimization design method based on multi-field node feature weighted aggregation when executed by a processor.

[0033] Compared with the prior art, the present application has the following beneficial effects:

[0034] 1) The influence of temperature and contact surface friction condition on the stress field prediction result of the complex part is considered, the stress field distribution prediction for the complex part profile parameters under variable working conditions (temperature interval, friction coefficient) is realized, and then the complex part profile optimization design considering multi-field influence in the variable temperature interval is performed.

[0035] 2) The spatial attention coefficient and the temperature attention coefficient are defined, and the GraphSAGE neural network is defined based on the same, the weighted aggregation of the multi-field node features is realized, the mechanism of the finite element thermal-structure coupling analysis is combined, the grid nodes of the contact part of the complex part and the grid nodes with large temperature difference are particularly concerned when the central node aggregates the neighbor node information, the reasonable distribution of the attention degree of the grid nodes is realized, and the parts with significant influence on the stress results are effectively aggregated. More feature information can be aggregated for important nodes in the network.

[0036] Considering the thermal-structure coupling mechanism of the complex part and using the graph neural network as a surrogate model to complete the optimization design, the stress distribution state of the complex part is quickly reconstructed, the optimization speed is improved, the accuracy of the optimization result is fully guaranteed, and the computing resources are greatly saved.

[0037] 3) In order to solve the problem that the initial dung beetle population generated by DBO is unevenly distributed and slows down the convergence speed of the algorithm, the OLH-DBO is improved based on the optimal Latin hypercube, the distribution quality of the initial population in the search space is improved, the global search ability is strengthened, the optimization result is more accurate, and the solving demand of the complex part profile optimization design problem is met. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The profile parameter optimization model flowchart of the blade root wheel groove.

[0039] Figure 2 The profile diagram of the blade root wheel groove and the multi-field node feature diagram.

[0040] Figure 3 The feature information diagram of the grid node.

[0041] Figure 4 The graph sampling aggregation network structure diagram of the fusion of spatial and temperature attention.

[0042] Figure 5 The flowchart of the OLH-DBO algorithm.

[0043] Figure 6 The result comparison diagram of the blade root wheel groove before and after optimization. DETAILED DESCRIPTION

[0044] The following takes the blade root wheel groove of the compressor as the specific implementation object of the application and further describes the application:

[0045] The profile optimization design method flowchart of the blade root wheel groove is shown in Figure 1 .

[0046] S1: performing finite element simulation according to the blade root wheel groove profile geometric design variable to be optimized, temperature and contact surface friction coefficient, that is, after parameterized batch generation, obtaining a feature matrix, an adjacency matrix and equivalent stress values and equivalent strain values of each grid node, wherein the feature matrix and the adjacency matrix are taken as inputs of a graph neural network, and the equivalent stress values and the equivalent strain values of each grid node are taken as labels, and then a training data set is obtained; wherein the geometric design variable and the contact surface friction coefficient form a geometric parameter;

[0047] S1 specifically is:

[0048] The research object is simplified, the influence of the axial stress is ignored, and the stress analysis problem is regarded as a two-dimensional plane. The data set input of the blade root wheel groove geometric parameter-stress-strain is composed of seven geometric parameters including the outer arc R Ex / mm of the blade root wheel groove, the inner arc R In / mm of the blade root wheel groove, the tooth width D / mm of the contact section of the blade root wheel groove, the tooth inclination angle a / ° of the contact section of the blade root wheel groove, the width L / mm of the blade root bottom, the maximum temperature t max / ℃ of the region of the blade root wheel groove, the minimum temperature T min / ℃ of the region of the blade root wheel groove and the contact surface friction coefficient μ, a total of eight design variables, wherein the maximum temperature and the minimum temperature are set because the temperature distribution state of the blade root wheel groove is that the temperature gradually increases from the center of the wheel groove bottom to the boundary of the blade root top, and the temperature of the two parts is determined, so that the temperature field distribution of the entire research object (blade root wheel groove) is determined; the equivalent stress σ (thermal stress and contact stress) and the equivalent strain ε (thermal strain and elastic strain) in the entire region of the blade root wheel groove are output, and the value range of the design variable is shown in Table 1:

[0049] Table 1 is a design variable value range table

[0050]

[0051] Table 2 is a constraint variable value range table

[0052]

[0053] The learning task of the model is to establish a mapping relationship from x to y, y=f(x), wherein x is the input geometric parameter set, y is the output stress field and strain field distribution, is the stress field information, is the strain field information:

[0054] x={x1,x2,x3,x4,x5,x6,x7,x8}

[0055]

[0056] y=f(x)

[0057] Based on the above geometric parameters and constraint conditions, the schematic diagram of the blade root wheel groove profile and the multi-field node feature schematic diagram are shown as Figure 2

[0058] Then, based on VBA and python script, the parametric design module of Ansys Workbench is automatically called to perform thermal-structural coupling finite element simulation calculation of the blade root wheel groove, and 1000 groups of calculation results are automatically generated in batches, each group of results containing about 2000 grid nodes. Among them, all grid nodes are triangular mesh, and the grid size is refined in a local encryption manner for the contact part of the blade root and the wheel groove. In addition, the coordinate information [x, y] and temperature field information T of all grid nodes in the output area after the thermal-structural coupling analysis calculation are also needed, and the horizontal coordinate x, the vertical coordinate y and the node temperature T are all taken as the input items of each graph convolution layer. Since the neural network used for stress and strain field prediction is a graph sampling aggregation neural network GraphSAGE, an adjacency matrix of the grid nodes also needs to be constructed to represent the adjacency relationship between the nodes.

[0059] The adjacency matrix is an M row by M column square matrix, denoted as A M×M Each node in the graph network is a mesh node divided during finite element calculation, and the information of the edge in the graph network represents the connection relationship between the two end nodes. Based on this, the adjacency matrix of the grid nodes is defined as follows:

[0060]

[0061] Among them, A i,j is the adjacency coefficient between the i-th node and the j-th node in the adjacency matrix.

[0062] Then, 1000 groups of sample sets are divided into training set, validation set and test set according to the ratio of 7:1.5:1.5, among which the training set has 700 groups, the validation set has 150 groups, and the test set has 150 groups.

[0063] S2: Construct a deep graph sampling aggregation neural network (Latt&Tatt-SAGEConv) that fuses spatial and temperature attention. Then, the deep graph sampling aggregation neural network that fuses spatial and temperature attention is trained using the training data set, and a stress prediction model is obtained after the training is completed;

[0064] In GraphSAGE, taking the structure of a triangular mesh element node as an example, each node is connected with four adjacent nodes. In the aggregation process, the center node will aggregate the feature information of the four adjacent nodes and splice the original node information to form its new node feature expression. The feature information of the node is as follows: Figure 3 ​As shown, after updating the node information for downstream node feature prediction, GraphSAGE's aggregation function can be expressed as:

[0065]

[0066] Where f(feature) is the feature of the node; F is the aggregated node feature; e(element) is the node; ne(neighbor element) is the neighbor node; i is the number of neighbor nodes included in each aggregation; j is the number of features of the node / neighbor node; Agg[] is the aggregation function.

[0067] Unlike traditional aggregation methods such as mean aggregation and max-pooling aggregation, this paper introduces two adjustment coefficients, spatial attention weight and temperature attention weight, and optimizes the aggregation function in the GraphSAGE graph network based on the spatial attention and temperature attention of the grid nodes.

[0068] For spatial attention, special consideration is given to the mesh refinement nodes in the contact area between the leaf root and the wheel groove, as they are in the area where the two parts frequently come into contact and are also the key area where the maximum equivalent stress may occur. The weight distribution of different neighboring nodes to the central node during the aggregation process is shown in Table 3.

[0069] Table 3 shows the parameters for the aggregation process based on spatial attention weights.

[0070]

[0071] For spatial attention, the aggregation process is defined as follows:

[0072]

[0073] Among them, L_att[(ne) i [] represents the spatial attention weight, which ranges from (0,1) and is defined as follows:

[0074]

[0075] In the formula The definition is as follows:

[0076]

[0077] Where 'e' represents the central node, (ne) i Let x represent the i-th neighbor node of e; e y e Let the x and y coordinates of the central node e be represented respectively. respectively represent the horizontal and vertical coordinates of the i-th neighbor node of the center node e. for the i-th neighbor node (ne) i of the j-th feature, and softmax[] is a normalized exponential function, aiming to smooth the weights and amplify the importance of important neighbor nodes.

[0078] By defining the spatial attention coefficient, nodes with smaller grid node spacing can be given more weight and more attention at each aggregation, and after multiple aggregation operations, the attention trend of the weight value is constantly developing towards the refined grid node, thereby aggregating more feature information of the leaf root and groove contact part.

[0079] For temperature attention, grid nodes with large temperature difference between the blade root and the wheel groove are particularly considered. According to the calculation formula of thermal stress, the temperature difference and thermal stress have a positive correlation trend, that is, the larger the node temperature difference, the greater the probability of the maximum value of the equivalent stress, and this area also needs special attention. The weight distribution of different neighbor nodes to the center node in the aggregation process is shown in Table 4.

[0080] Table 4 is the aggregation process parameter table based on temperature attention weight

[0081]

[0082] Therefore, for temperature attention, the aggregation process is defined as follows:

[0083]

[0084] where T_att[(ne) i ] is the temperature attention weight, whose value range is (0, 1), and is defined as follows:

[0085]

[0086] In the formula, is defined as follows:

[0087]

[0088] where e represents the center node, (ne) i represents the i-th neighbor node of e; T e represents the temperature of the center node e, represents the temperature of the i-th neighbor node of the center node e, and || represents the absolute value.

[0089] By defining the temperature attention coefficient, the nodes with larger temperature difference can obtain more attention in the aggregation process, and after multiple aggregation operations, the attention trend of the weight value is constantly developing towards the grid nodes with larger temperature gradient, thereby aggregating more feature information of parts prone to thermal stress concentration.

[0090] The aggregation process considering spatial attention and temperature attention is weighted to obtain the final aggregation function:

[0091]

[0092] Wherein, W is an adjustable weight parameter, and its value range is (0, 1).

[0093] The depth GraphSAGE neural network (Latt&Tatt-GraphSAGE) fusing spatial and temperature attention is composed of an input layer, multiple Latt&Tatt-SAGEConv graph convolution layers and an output layer, wherein the graph convolution module can be represented as:

[0094]

[0095] The graph convolution module of each layer includes an aggregation function Latt&Tatt-Aggregate() and a concatenation function CONCAT(), and uses LeakyReLU as an activation function to increase non-linear excitation and improve fitting efficiency, and K is a trainable weight matrix. is the feature of the neighbor node after aggregation, f e (j) is the feature of the center node. Compared with the ReLU function, it solves the problem of neuron death in the negative half of the function, and has the ability to overcome the problems of gradient disappearance and gradient explosion of the ReLU function. At the same time, a dropout layer is added before the last output layer to avoid the occurrence of neural network overfitting phenomenon, and the dropout rate is set to 0.25, and the structure of the network is as shown in Figure 4

[0096] After normalizing the data set, it is divided into training set, validation set and test set, and the network model is trained and tested.

[0097] When training GraphSAGE, not only the feature matrix needs to be input, but also the adjacency matrix needs to be input at the same time to fully express the correlation between nodes.

[0098] The initial parameter settings of the Latt&Tatt-GraphSAGE neural network are as follows: epoch is 1000, batch size is 64, and the initial learning rate is 0.001.

[0099] ​The stress and strain prediction in this invention is a supervised learning task. In each training iteration, the model takes the node feature map and adjacency matrix as input and performs forward propagation calculations based on the current network parameters. Propagation through multiple graph convolutional layers captures the information flow between nodes. Then, a loss function is calculated to quantify the error between the output and the label value. By continuously performing backpropagation, the gradient of the deviation value relative to the model parameters is calculated, and the Adam optimizer uses this gradient information to update the network parameters. Through iterative reduction of the loss function, the prediction model gradually learns how to predict unknown labels based on node features and the structure of the graph network.

[0100] To evaluate the predictive performance of the Latt & Tatt-GraphSAGE neural network model, the true values ​​y of L test samples are evaluated. i Compared with the predicted value The three indicators of the error between them include:

[0101] Root Mean Square Error (RMSE):

[0102]

[0103] The root mean square error (RMSE) is the square root of the ratio of the sum of squares of the deviations between the predicted and the true values ​​to the total number of test samples. It is used to measure the difference between the predicted and the true values. Theoretically, its value is positive, and the closer it is to 0, the better the model's prediction performance.

[0104] Mean Absolute Error (MAE):

[0105]

[0106] The mean absolute error is calculated by dividing the absolute value of the deviation between the predicted value and the true value by the total number of test samples. Theoretically, its value is positive and the closer it is to 0, the better.

[0107] Coefficient of determination (R) 2 ):

[0108]

[0109] Coefficient of determination R 2 R is a statistical indicator used to evaluate the goodness of fit of a regression model. 2 The value of is between (0,1), and R 2 The closer the value is to 1, the better the model's regression prediction fits the sample.

[0110] S3: Combining the stress prediction model, the optimal geometric parameters of the current blade root groove profile are obtained by iterative optimization with the goal of minimizing the equivalent stress of the blade root groove.

[0111] S3 is specifically:

[0112] The minimum maximum equivalent stress is selected as the objective function of the optimization design, the thermal stress and contact stress of the blade root raceway are reduced as much as possible under the premise of meeting all constraint conditions under the influence of multiple fields, and according to the design variables, constraint conditions and objective function, the mathematical model of the blade root raceway optimization design is as follows, the specific meaning is to find a group of x in the interval of 400 (x7) to 600 ℃ (x8), under the condition of meeting the constraint variables W(x) and H(x), the maximum stress value in the stress field is minimized:

[0113] Find x={x1,x2,x3,x4,x5,x6}

[0114]

[0115]

[0116] Wherein, x is the design variable of the blade root raceway, x1, x2, x3, x4, x5, x6, x7, x8 are respectively the contact section tooth width D / mm of the blade root raceway, the contact section tooth inclination angle α / ° of the blade root raceway, the outer arc R Ex / mm of the blade root raceway, the inner arc R In / mm of the blade root raceway, the bottom width L / mm of the blade root raceway, the friction coefficient μ of the contact surface, the minimum temperature T min / ℃ of the region, the maximum temperature T max / ℃ of the region of the blade root raceway, The maximum stress value in the stress field, W(x) is the geometric constraint variable of the blade root raceway in the horizontal direction, The horizontal constraint limit is generally 0.9 times the initial design value, y(x) is the geometric constraint variable of the blade root raceway in the vertical direction, The vertical constraint limit is generally 0.9 times the initial design value.

[0117] The expression of the constraint variable can be derived by means of plane geometry in the horizontal direction and the vertical direction, which is shown in Table 5:

[0118] Table 5 is the expression of the constraint variable

[0119]

[0120] Meanwhile, since the dung beetle algorithm population initialization adopts a random generation mode, the dung beetle population may gather in a certain area, so that the algorithm is prone to fall into local optimum and poor convergence accuracy during iterative search. In order to solve these problems, the optimal Latin hypercube (Optimal Latin Hypercube) is introduced to initialize the distribution of the dung beetle population, instead of the original random initialization strategy, so that the dung beetle population is relatively uniformly distributed in the entire optimization space, while increasing the diversity of the initial population and avoiding the problem of falling into local optimum during the search process. The improved dung beetle initialization function is as follows:

[0121] [x n ,y n ]=OLH(Maxnum,lb,ub)

[0122] Wherein, ub is the upper limit of search, lb is the lower limit of search, Maxnum is the population number, [x n ,y n ] is the updated position matrix of the dung beetle individual.

[0123] The parameter setting of the dung beetle optimization algorithm is as follows: the initial number of dung beetles is 100, the maximum iteration number is 300, and the improved optimization algorithm is as shown in Figure 5 When the optimization process reaches the convergence condition, a set of optimal parameter combinations can be obtained as shown in Table 6:

[0124] Table 6 is the geometric design variable value before and after optimization

[0125]

[0126]

[0127] The comparison of the results before and after the optimization of the blade root wheel groove is as shown in Figure 6 The maximum equivalent stress of the blade root wheel groove area is reduced from 202.22Mpa to 167.6Mpa, which is reduced by 17.1% compared with the initial configuration, so it can be seen that the maximum equivalent stress of the blade root wheel groove is significantly improved, which fully embodies the effectiveness and feasibility of the optimization design method proposed in the application.

[0128] The application also proposes a blade root wheel groove profile optimization design system based on multi-field node feature weighted aggregation, comprising:

[0129] A data set construction unit is used for finite element simulation according to the blade root wheel groove profile geometric design variables to be optimized, temperature and contact surface friction coefficient, to obtain a feature matrix, an adjacency matrix and equivalent stress values and equivalent strain values of each grid node, and after repeated multiple times, a training data set is obtained;

[0130] The stress prediction unit is configured to store a depth map sampling aggregation neural network fusing spatial and temperature attention, and train the depth map sampling aggregation neural network fusing spatial and temperature attention by using a training data set, so as to obtain a stress prediction model.

[0131] The geometry parameter optimization unit is combined with the stress prediction model to perform iterative optimization with the minimum blade root race groove equivalent stress as an optimization target, so as to obtain optimal geometry parameters of the current blade root race groove profile in the target temperature interval.

[0132] The present application further provides a computer device, which comprises a memory and a processor, the memory stores a computer program, and the processor implements the steps of the blade root race groove profile optimization design method based on multi-field node feature weighted aggregation when executing the computer program.

[0133] The present application further provides a computer readable storage medium, which stores a computer program, and the computer program implements the steps of the blade root race groove profile optimization design method based on multi-field node feature weighted aggregation when executed by a processor.

[0134] The present application further provides a computer program product, which comprises computer programs / instructions, and the computer programs / instructions implement the steps of the blade root race groove profile optimization design method based on multi-field node feature weighted aggregation when executed by a processor.

[0135] The above embodiments are used to explain and illustrate the present application, but not to limit the present application, and any modifications and changes made to the present application within the spirit and protection scope of the claims of the present application all fall within the protection scope of the present application.

Claims

1. A method for optimizing the profile design of complex parts based on weighted aggregation of multi-field node features, characterized in that, Includes the following steps: S1: Based on the geometric design variables, temperature, and contact surface friction coefficient of the complex part profile to be optimized, finite element simulation is performed to obtain the feature matrix, adjacency matrix, and equivalent stress and equivalent strain values ​​of each mesh node, thereby obtaining the training dataset; among them, the geometric design variables and contact surface friction coefficient constitute the geometric parameters; S2: Construct a depth map sampling aggregation neural network that integrates spatial and temperature attention. Then, train the depth map sampling aggregation neural network that integrates spatial and temperature attention using the training dataset. After training, obtain the stress prediction model. S3: Combining the stress prediction model, the optimization objective is to minimize the maximum equivalent stress of the complex part and iteratively optimize to obtain the optimal geometric parameters of the current complex part profile within the target temperature range. In S2, the aggregation function in the depth map sampling aggregation neural network that integrates spatial and temperature attention satisfies the following formula: in, As the central node Through its The first neighbor node is formed after aggregation. One characteristic, As the central node The One neighboring node, Spatial attention coefficient, Temperature attention coefficient, For adjustable weight parameters, For the first The first neighbor node feature, For normalized exponential functions, For the first The spatial distance between each neighboring node and the central node For the first Each neighboring node and the central node temperature difference, , As the central node The number of neighboring nodes; In S3, the network parameters of the stress prediction model are frozen and the stress prediction model is used as the fitness function in the optimization algorithm to obtain an improved optimization algorithm. The improved optimization algorithm takes minimizing the maximum equivalent stress of the complex part as the optimization objective and iteratively updates the geometric parameters until the optimal geometric parameters of the current complex part profile within the target temperature range are obtained. Specifically, S3 is: First, the initial parameters of the dung beetle algorithm are determined. Then, the optimal Latin hypercube function is used to initialize the dung beetle population distribution, ensuring that the initial dung beetle population is evenly distributed throughout the optimization space, thereby obtaining the initialized dung beetle individual position matrix. Next, the stress prediction model of the frozen network parameters is used as the fitness function of the dung beetle algorithm to obtain an improved dung beetle algorithm. Then, based on the initialized dung beetle individual position matrix, the improved dung beetle algorithm is used to iteratively update the dung beetle population position until the convergence condition is finally met, thereby obtaining the optimal geometric parameters of the current complex part profile within the target temperature range.

2. The method for optimizing the profile design of complex parts based on weighted aggregation of multi-field node features according to claim 1, characterized in that, In S1, the temperature specifically refers to the maximum and minimum temperatures in the target area.

3. A complex part profile optimization design system based on multi-field node feature weighted aggregation for implementing the optimization design method of claim 1, characterized in that, include: The dataset construction unit is used to perform finite element simulation based on the geometric design variables of the complex part profile to be optimized, temperature and contact surface friction coefficient, to obtain the feature matrix, adjacency matrix and equivalent stress value and equivalent strain value of each mesh node. After repeating this process multiple times, the training dataset is obtained. The stress prediction unit is used to store the depth map sampling aggregation neural network that integrates spatial and temperature attention, and to train the depth map sampling aggregation neural network that integrates spatial and temperature attention using the training dataset, thereby obtaining the stress prediction model. The geometric parameter optimization unit is used to combine the stress prediction model to iteratively optimize the model with the goal of minimizing the equivalent stress of complex parts, and obtain the optimal geometric parameters of the current complex part profile within the target temperature range.

4. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the complex part profile optimization design method based on multi-field node feature weighted aggregation as described in claim 1 or 2.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the complex part profile optimization design method based on multi-field node feature weighted aggregation as described in claim 1 or 2.

6. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the complex part profile optimization design method based on multi-field node feature weighted aggregation as described in claim 1 or 2.

Images