Stiffness distribution guided intelligent design method for positioning layout of aeronautical thin-walled component

Abstract

The application provides a rigidity distribution guided intelligent design method for positioning layout of an aerospace thin-walled component, establishes a finite element simulation model of the thin-walled component in a positioning state, saves a positioning point feasible region as a to-be-selected node set, and obtains an input file of the finite element simulation; based on the input file of the initial simulation, the positioning point number in the input file is randomly and batchingly rewritten to form a large number of input files corresponding to different positioning layouts; according to the input file and a geometric model of the thin-walled component, a positioning layout feature map expressing the component shape, the reinforcing rib and the positioning point is drawn; the finite element simulation is performed on all the generated input files in batches, and the deformation of each node in the thickness direction of the component in each simulation result is output, the rapid and accurate evaluation of the rigidity can be realized, the problem of low efficiency of the finite element simulation is avoided, and therefore the shape-property collaborative control of the thin-walled structure of the aerospace in a positioning section is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of aerospace design optimization, in particular to a stiffness distribution guided intelligent design method for positioning and layout of an aerospace thin-walled component. BACKGROUND

[0002] There are a series of large and complex thin-walled components in the structure of a new generation of aircraft. Such components are distributed throughout the fuselage, wings and tail. The thin-walled structure has weak rigidity, and in the assembly process, deformation may occur due to unreasonable positioning and layout, thereby causing excessive deviation of bolt holes used for assembly and connection, and causing serious stress concentration. In order to ensure the dimensional accuracy of the structure assembly, the component should first be ensured not to deform excessively in the positioning state. The existing positioning and layout method often takes minimizing the deformation of the part as the design target, and is only applicable to regular-shaped structures. However, in a complex assembly environment, it is difficult to determine the actual load condition of the component, and therefore, when performing finite element simulation, the simulation result may not be accurate. In addition, complex boundary conditions will greatly increase the size of matrix operation, resulting in reduced simulation efficiency. Therefore, the deformation guided positioning and layout design method is not suitable for aircraft thin-walled components. The root cause of solving this problem is to control a physical quantity directly related to deformation and independent of specific boundary conditions, and at the same time, the physical quantity can simultaneously affect the deformation behavior and stress distribution. In other words, an inherent property of a component in the positioning state is needed, which can reflect the deformation ability and stress distribution of the component under various boundary conditions, to guide the design of positioning and layout. For this purpose, the present application proposes to control the global distribution of stiffness to be uniform, which can indirectly control the thin-walled component not to deform greatly, while satisfying the overall balanced low stress state of the component. SUMMARY

[0003] The technical problem to be solved by the present application is to establish a conditional generative adversarial network for end-to-end generation of physical fields, and to realize fast prediction of physical fields, in view of the deficiencies and needs of the prior art. Subsequently, the pre-trained generative adversarial network is embedded as a solver in NSGA-III (Non-dominated Sort Genetic Algorithm III) for process design. Combined with the NSGA-III process design, the conditional generative adversarial network can quickly and accurately predict the physical field of the process design, and the process design can be optimized by the predicted physical field. Figure 1 The specific method, process and content implemented by the present application are as follows:

[0004] S1, a finite element simulation model of the thin-walled component in a positioning state is established, and the positioning point feasible region is saved as a set of to-be-selected nodes, and an input file for finite element simulation is obtained; based on the input file of the initial simulation, the positioning point numbers in the input file are randomly and batch-wise rewritten to form a large number of input files corresponding to different positioning layouts; according to the input file and the geometric model of the thin-walled component, a positioning layout feature map expressing the component shape, reinforcing ribs and positioning points is drawn; all the generated input files are batch-wise subjected to finite element simulation, and the deformation of each node in the thickness direction of the component in each simulation result is output; a smooth grid is created through existing data points for interpolation; an instance feature encoder, a local generator and a global generator are constructed, and on the basis of a convolution block attention module, minimum pooling is added to help the model identify weak or subtle features; the loss of the deep learning network is defined;

[0005] S2, considering the pixel points of the component plane region, a global uniformity index is constructed to describe the global distribution uniformity of stiffness; a local smoothness index is constructed to measure the degree of excessive sharpness of each pixel point in the neighborhood; an extreme difference index is constructed to quantify the relative difference between the maximum value and the minimum value, and to impose constraints on the positioning layout generated in the iteration process;

[0006] S3, a large number of paired positioning layout process feature maps-component stiffness cloud map data sets are obtained and labeled; the deep learning model is trained using the labeled data set;

[0007] S4, a flexible tooling with reconfigurable positioning layout is built, and a test platform is built, taking the assembly hole in the test structure as the key measurement point, measuring the position of the point to verify whether the positioning design method can reduce the positioning deformation, measuring the positions of any three points around the hole to obtain the center of the circle, and taking the average value of three groups of each point as the final hole position.

[0008] In one embodiment, there are mainly two strategies for drawing the stiffness cloud map. One is: for components without holes or circular structures, different colors are directly assigned to different nodes, and a continuous cloud map is drawn through color interpolation. The second is: for components with holes or circular structures, the edge points of the component are detected, and the distance and angle of any given point to other points are calculated.

[0009] In one embodiment, the edge points are clustered by a density-based clustering algorithm to identify the inside and outside of the component, and the convex hull is calculated to obtain the outer boundary and the inner boundary, and it is detected whether the grid node is located in the inner convex hull, and if so, no interpolation is performed.

[0010] In one embodiment, the encoder extracts features in the scale of the original stiffness cloud map through down-sampling and up-sampling.

[0011] In one embodiment, a set of residual blocks in series is added to the deepest layer of the encoder-decoder structure, and an improved convolution block attention module is incorporated in the residual blocks.

[0012] In one embodiment, the local generator adopts a progressive multi-scale processing strategy to gradually refine the stiffness field from low resolution to high resolution. Each local generator first performs a down-sampling operation on the input image at the current resolution to extract features, then fuses these features with the features of the previous level, and finally restores the resolution of the fused features through up-sampling processing. Because the scales of features at different levels are not the same, the local generator does not explicitly define the input and output image sizes. In addition, the local generators used at different levels do not share parameters.

[0013] In one embodiment, a hierarchical PatchGAN architecture is used to construct the discriminator, which evaluates the authenticity of the stiffness field at multiple resolution levels; the output single-channel feature map is used for authenticity judgment, forming a multi-scale discrimination mechanism from high resolution to low resolution.

[0014] In one embodiment, constraints are imposed on the positioning layout generated during the iteration process to ensure the physical consistency of the generated scheme; the constraints are divided into feasibility constraints and positioning relationship constraints, the feasibility constraints ensure that the positioning layout appears in the part area, and the positioning relationship constraints ensure that the generated positioning layout meets the requirements.

[0015] The above technical features can be combined in various suitable ways or replaced by equivalent technical features, as long as the purpose of the application can be achieved.

[0016] The stiffness distribution guided intelligent design method for positioning layout of an aviation thin-walled component provided by the application has at least the following beneficial effects compared with the prior art:

[0017] (1) The stiffness can be quickly and accurately evaluated, and the low efficiency of finite element simulation is avoided;

[0018] (2) The inherent properties of the component, i.e., stiffness, can be used to avoid the difficulty of setting complex boundary conditions;

[0019] (3) The stiffness can be used to associate the shape accuracy and stress distribution level, and shape-property synergy can be achieved.

[0020] The stiffness distribution guided intelligent design method for shape-property synergistic positioning layout of an aerospace thin-walled component provided by the application can realize shape-property synergistic control of the aerospace thin-walled structure in positioning layout. BRIEF DESCRIPTION OF DRAWINGS

[0021] The application will be described in more detail below based on the embodiments and with reference to the accompanying drawings.

[0022] Figure 1 Batch generation process for inp file;

[0023] Figure 2 Structure for instance encoder;

[0024] Figure 3 Structure for global generator;

[0025] Figure 4 Structure for local generator;

[0026] Figure 5 Structure for discriminator;

[0027] Figure 6 Non-dominated sorting genetic algorithm III flow chart for embedding pre-training deep learning model;

[0028] Figure 7 Flexible tooling for configuration and positioning layout reconfigurable;

[0029] Figure 8 Rear fuselage structure of certain type of aircraft;

[0030] Figure 9 Positioning layout feature map;

[0031] Figure 10 Stiffness cloud chart. DETAILED DESCRIPTION

[0032] The application will be further described below with reference to the drawings.

[0033] The application provides a stiffness distribution guided intelligent design method for positioning layout of an aviation thin-walled component, establishes a conditional generative adversarial network for generating a physical field in an end-to-end mode, and realizes rapid prediction of the physical field. Subsequently, a pre-trained generative adversarial network is taken as a solver, and is embedded into NSGA-III (Non-dominated Sort Genetic Algorithm III) for process design. Combined with the characteristics of the aviation thin-walled component, the application realizes the intelligent design of the positioning layout of the aviation thin-walled component. Figure 1 The specific method, process and content implemented by the application are as follows:

[0034] S1, physical field prediction model construction based on the conditional generative adversarial network;

[0035] S2, NSGA-III optimization model construction embedded with a pre-training solver;

[0036] S3, deep learning model training and optimization model training method;

[0037] S4, test method.

[0038] In step S1, including S11 dataset construction and S12 deep learning modeling, wherein the S11 dataset construction process mainly includes the following steps:

[0039] S111, a finite element simulation model of the thin-walled component in the positioning state is established, and the positioning point feasible region is saved as a to-be-selected node set, and an input file (inp file) of finite element simulation is obtained;

[0040] S112, based on the initial simulation inp file, the positioning point number in the inp file is randomly and batched rewritten to form a large number of inp files corresponding to different positioning layouts, and the specific process is as shown in Figure 1 , including:

[0041] S1121, for the initial inp file, the coordinates of the first (initial i =1) node are read; i

[0042] S1122, it is judged whether the coordinates are in the thickness direction plane, if yes, continue to S1123, if not, repeat S1121 until the coordinates are in the thickness direction plane;

[0043] S1123, the node information is output to an excel document;

[0044] S1124, it is judged whether the current node number i is the maximum node number n , if yes, the process is ended, if not, the node number+1, and turn to S1121, repeat S1121-S1124 until i is the maximum node number n；

[0045] S1125, copy y inp files, read the first k line (initial k =1) of the first j inp file (initial j =1), and search whether there is a “ position ” word in the line, if yes, a plurality of nodes are randomly selected from the excel output from S1123 to replace the node number of the first j +1 line, if not, the line number+1, and continue to search for the “ position ” word;

[0046] S1126, the first k inp file is output;

[0047] S1127, it is judged whether k the number of inp files is y ​Equal, end flow, complete all inp file generation, if not, inp file number +1, go to S1125, repeat S1125~S1127, until k With y Equal.

[0048] S113, according to the inp file and the thin-walled component geometric model, draw the positioning layout feature map expressing the component shape, stiffener and positioning point.

[0049] S114, batch finite element simulation is performed on all generated inp files, and the deformation of each node in the thickness direction of the component in each simulation result is output. The reciprocal of the deformation of each node is taken as the absolute stiffness of the node, and the absolute stiffness of all components is normalized according to the following formula:

[0050] ;

[0051] In the formula, k norm is the normalized stiffness value, k is the current node stiffness value, k min is the global minimum stiffness value, k max is the global maximum stiffness value.

[0052] S115, there are mainly two strategies for drawing the stiffness cloud map. One: for components without openings or circular ring structures, different colors can be directly assigned to different nodes, and a continuous cloud map can be drawn through color interpolation.

[0053] S116, the second: for components with openings or circular rings, the edge points of the component are detected, and the distance p i x pi , y pi and the angle p j x pj , y pj from other points d ij θ ij are calculated:

[0054] ;

[0055] Wherein, the atan2 function is the four-quadrant arctangent function. On this basis, every θ degree is detected once, and if p ​​​i At a given distance d If there are not enough enough neighboring points within the distance, the point is considered p i as an edge point.

[0056] S117, the edge points are clustered by density-based clustering algorithm to identify the inside and outside of the component. The neighborhood N ζ p i and core point criterion are defined as:

[0057] ;

[0058] where, D is the point set, ζ is the neighborhood range, MinPts is the minimum number of points containing the point.

[0059] S118, the convex hull is used to calculate the outer boundary and the inner boundary, and its properties are as follows:

[0060] ;

[0061] where, P is the point set p i composed of the convex hull, q is a point in the convex hull conh P λ n is the non-negative weight of the nth point.

[0062] S119, a smooth grid is created by the existing data points for interpolation, and it is detected whether the grid node is located in the inner convex hull, if so, no interpolation is performed. Thus, it is ensured that the interpolation is only performed within the actual component range. In order to avoid the contour lines of the cloud map being too sparse, the interval between the contour lines is set to a small value.

[0063] Further, the S12 deep learning modeling process includes the following steps:

[0064] S121, an instance feature encoder is constructed. The encoder performs feature extraction on the scale of the original stiffness cloud map in the form of down-sampling-up-sampling. The structure of the encoder is shown in Figure 2 ;

[0065] ​​S122, a global generator is constructed, which is similar to the encoder, using an encoder-decoder structure. At the same time, a set of residual blocks in series is added to the deepest layer of the structure, and an improved Convolutional Block Attention Module (iCBAM) is integrated into the residual blocks. The structure of the global generator is shown in Figure 3 .

[0066] S123, a local generator is constructed using a U-Net structure. Similar to the global generator, a set of residual blocks integrating the improved convolution block attention module is added to the deepest layer of the U-Net. The local generator adopts a progressive multi-scale processing strategy, gradually refining the stiffness field from low resolution to high resolution. Each local generator first performs down-sampling operation on the input image of the current resolution to extract features, then fuses these features with the features of the previous level, and finally restores the resolution of the fused features through up-sampling processing. Because the feature scales of different levels are not the same, the local generator does not explicitly define the input and output image sizes. In addition, the local generators used at different levels do not share parameters. The structure of the local generator is shown in Figure 4 ;

[0067] S124, on the basis of the traditional CBAM, a minimum pooling is added to help the model identify weak or subtle features to form iCBAM. By capturing these low-response features, the model's learning ability for the stiffness distribution pattern in small areas can be improved;

[0068] S125, a hierarchical PatchGAN architecture is used to construct the discriminator. This architecture evaluates the authenticity of the stiffness field at multiple resolution levels. The output single-channel feature map is used for authenticity judgment, forming a multi-scale discrimination mechanism from high resolution to low resolution. The structure of the discriminator is shown in Figure 5 ;

[0069] S126, the loss of the deep learning network is defined, including:

[0070] S1261, the generator loss based on LSGAN:

[0071]

[0072] wherein, E x~Pgen represents the expectation of the input sampled from the generated data distribution, D ( G ( x )) is the determination result of the discriminator for the generated sample.

[0073] S1262, the loss of the discriminator based on LSGAN: LPixel error loss of 1:

[0074]

[0075] where, E x~Pdata denotes the expectation of the input sampled from the real data distribution, G ( x ) is the output of the generator, x real denotes the real stiffness field of the input x . ||·||1 is the L 1-norm.

[0076] S1263, perceptual loss based on VGG19:

[0077]

[0078] where, ϕ i is the feature of the i layer of the VGG network. ω i is the layer weight.

[0079] S1264, discriminator loss:

[0080]

[0081] In step S2, the stiffness distribution evaluation index design S21 and the positioning layout design model construction with engineering constraints S22 are included, wherein S21 mainly includes the following steps:

[0082] S211, considering the non-white area, i.e. the pixel points of the component plane area, a global uniformity index (G) is constructed to describe the global distribution uniformity of the stiffness: I GU

[0083]

[0084] where, M and N are the number of pixel rows and the number of pixel columns, respectively; f i,j denotes the value corresponding to the pixel point at the position ( i , j ); f is the global mean of the physical quantity, which can be described as:

[0085]

[0086] S212, define the neighborhood range of the pixel points, and construct a local smoothness index (L) to describe the local smoothness of the stiffness: I ​LS ) for measuring the degree of excessive sharpness of each pixel point in the neighborhood:

[0087]

[0088] where N( i , j ) represents the neighborhood set of the pixel point at position ( i , j ); | N| is the number of pixel points in the neighborhood; f p,q represents the physical quantity value of the pixel points in the neighborhood. Taking the four-neighborhood and eight-neighborhood as examples. For the four-neighborhood, we have:

[0089]

[0090] For the eight-neighborhood, we have:

[0091]

[0092] S213, constructing an extreme difference index ( I ED ), quantifying the relative difference between the maximum and minimum values:

[0093]

[0094] where f max and f min are the maximum and minimum values of the physical quantity, respectively.

[0095] Further, the process of constructing the positioning layout design model with engineering constraints S22 includes the following steps:

[0096] S221, using NSGA-III as a multi-objective optimizer, constructing a real-time solver by combining the deep learning model constructed in the pre-training step S1 with the stiffness evaluation index in step S21. The construction process is shown in Figure 6 , which specifically includes the following steps:

[0097] S2211, initialization, then determine whether the sub-generation is generated when the generation number is 1, if yes, then the generation number +1, if not, perform non-dominated sorting, selection, crossover, mutation, and then generate the sub-generation;

[0098] S2212, combine the parent and child generations, then determine whether a new parent generation is generated, if yes, perform selection, crossover, mutation, and then generate the sub-generation, if not, call the solver composed of the pre-trained deep learning model and the evaluation index, perform non-dominated sorting, create reference points, perform adaptive normalization, based on niche preservation, generate a new parent population;

[0099] In S2213, determine whether the maximum generating algebra has been reached. If not, determine whether to terminate. If yes, terminate. If not, determine whether the termination condition has been met. If yes, terminate. If not, increment the algebra by 1 and go to S2211. Repeat S2211 to S2213 until the termination condition is met.

[0100] S222, to find the optimal solution more quickly and accurately, constraints are imposed on the positioning layout generated during the iteration process to ensure the physical consistency of the generated scheme. The constraints are divided into feasibility constraints and positioning relationship constraints. Feasibility constraints ensure that the positioning layout appears within the part area, while positioning relationship constraints ensure that the generated positioning layout meets basic engineering requirements.

[0101] Step S3 includes S31 deep learning model training and S32 model optimization training, where S31 mainly includes the following steps:

[0102] S311, A large dataset of paired positioning layout process feature maps - component stiffness cloud maps is obtained through S113 and S115 and then labeled.

[0103] S312 uses the labeled dataset obtained in S311 to train the deep learning model established in S12.

[0104] Furthermore, the S32 optimization model training mainly includes the following steps:

[0105] S321. Based on the stiffness evaluation index, set the optimization objective. Specifically, it is desirable for the global distribution of the stiffness field to be as uniform as possible, the local transitions to be as smooth as possible, and the global consistency to be as high as possible. Therefore, the global uniformity index, the local smoothness index, and the extreme value difference index should be maximized.

[0106] S322 sets the population size, number of iterations, initial population size, and number of reference points.

[0107] The test method in step S4 mainly includes the following:

[0108] S41, flexible tooling with reconfigurable configuration and positioning layout, such as Figure 7 As shown, it includes: a tooling base 411; a heading movement module and a spanwise movement module 412, connected to the positioner base, and capable of free movement in the heading and spanwise directions via a linear slide rail-slider system; a flexible positioning end 413 based on a vacuum suction cup; and a connection structure 414 between the vacuum suction cup and the reconfigurable tooling structure, the connection structure being capable of free movement in the longitudinal direction on the tooling structure.

[0109] S42, Test methods, including:

[0110] S421, build a test platform, including: flexible tooling and test pieces, three-coordinate articulated arm measurement equipment, measurement data reading notebook computer;

[0111] S422, take the assembly hole in the test structure as the key measurement point, measure the position to verify whether the positioning design method can reduce the positioning deformation, and calculate the center of the circle from the positions of any three points around the hole, and take the average value of three groups at each point as the final hole position;

[0112] S423, compare the difference between the nominal position of the key measurement point under the original positioning layout and the designed positioning layout;

[0113] S424, establish a finite element simulation model, and compare the difference between the average and maximum values of the stress distribution of the component under the positioning state under the original positioning layout and the designed positioning layout.

[0114] In one embodiment, a stiffness distribution guided intelligent design method for positioning layout of an aeronautical thin-walled component includes the following steps:

[0115] Step 1, make positioning layout feature maps and stiffness cloud maps of each part under different positioning conditions; Step 2, build a deep learning model for generating stiffness cloud maps from positioning layout feature maps and train it; Step 3, embed the component into the pre-trained deep learning model and train the NSGA-III optimization model of the evaluation index; Step 4, test.

[0116] Step 11, make positioning layout feature maps and stiffness cloud maps of each part under different positioning conditions, specifically, the positioning layout feature map is as shown in Figure 8 、 Figure 9 :

[0117] (1) The black closed line depicts the collective shape of the part;

[0118] (2) The green line depicts the position distribution of the part's reinforcing ribs;

[0119] (3) The blue node represents the position of the current part after positioning, which is connected to other parts, and is also the position of the element assembly load in the finite element;

[0120] (4) The red node represents the positioning point of the current part, which is also the position of the fixed constraint in the finite element.

[0121] Step 12, make stiffness cloud maps of each part under different positioning conditions, as shown in Figure 10 :

[0122] (1) Batch create inp files, modify the node number combination under the positioning point set of the original inp file, and form several inp files;

[0123] (2) batch finite element simulation is performed;

[0124] (3) the deformation simulation result graph of the part in the thickness direction is intercepted.

[0125] Step 21, using Python language and PyTorch framework, the implementation code of the deep learning model described in S12 is written, specifically:

[0126] (1) using a global generator to extract features of the process feature map, the input of the global generator being the process feature map and the output being the process feature map after feature extraction;

[0127] (2) using a local generator to generate a stiffness field, the input of the local generator being the process feature map after feature extraction and the output being the stiffness field;

[0128] (3) using a discriminator to determine the similarity between the generated stiffness field and the real stiffness field, if the discriminator identifies the generated stiffness field as real, it is considered that the stiffness field generated by the generator is real enough so that the discriminator cannot distinguish it;

[0129] (4) generator loss is used to measure whether the generated stiffness field can be regarded as a real stiffness field by the discriminator;

[0130] (5) L 1loss is used to evaluate the absolute difference between the generated stiffness field and the real stiffness field in the pixel space;

[0131] (6) the perceptual loss introduced based on the pre-trained VGG19 network is used to compare the feature expressions of the generated stiffness field and the real stiffness field and evaluate their similarity. This loss evaluation method considers the high-level features of the stiffness field. By comparing the differences between the generated stiffness field and the real stiffness field on each slice, the loss is calculated;

[0132] (7) discriminator loss is used to measure whether the discriminator can accurately determine that the generated stiffness field is false.

[0133] Step S22, training the model on a device with physical memory ≥ 16G and video memory ≥ 8G. Specifically:

[0134] (1) dividing the training set, validation set and test set according to the ratio of 7:2:1;

[0135] (2) using Adam optimizer to dynamically adjust neural network parameters;

[0136] (3) setting the initial learning rate to 0.0001;

[0137] (4) setting the training round to 1000, of which the first 200 times use the initial learning rate, and in the subsequent 800 times, the learning rate will be linearly decayed;

[0138] (5) Load 1 picture per batch;

[0139] (6) Set the weight of feature matching loss to 10.

[0140] Step 23, training to the discriminator cannot accurately judge the true and false of the stiffness field, and it is considered that the generator and the discriminator reach Nash equilibrium.

[0141] Step 31, using Python language and PyTorch framework, embedding the pre-trained deep learning model in step 23, combining the stiffness evaluation index described in S21 as the NSGA-III model of the solver and applying engineering constraints, specifically:

[0142] (1) Randomly generate a positioning layout and generate a positioning layout feature map, and then call the pre-trained deep learning model to generate a stiffness cloud map, and then calculate the stiffness distribution evaluation index based on the cloud map;

[0143] (2) Apply engineering constraints, specifically:

[0144] The design of the positioning layout is essentially a non-fixed sequence modeling problem. Allowing the optimizer to optimize in a dynamic dimension is more in line with engineering applications. At this time, the variable space χ can be defined as:

[0145]

[0146] Where, Ω is the feasible region of the positioning point. i is the number of positioning points, in the embodiment of the present application, the number of positioning points will not be artificially limited. Only the maximum and minimum positioning points are constrained, so there are i min and i max . p i is the th positioning point, i i is the set of positioning points. P outer

[0147] Further, the feasible region of the positioning point is constrained. For the center frame, the positioning point is strictly required to fall between the inner and outer boundaries. Record the area surrounded by the outer boundary as P outer , the area surrounded by the inner boundary as P inner , then the feasible region is:

[0148]

[0149] For the center frame, the positioning layout is generally symmetrical. Therefore, the positioning layout is constrained aboutx =0 symmetry. Moreover, to ensure the effectiveness of the generated positioning scheme, the distance between the positioning points is required to be at least larger than the diameter of the vacuum chuck.

[0150] In the mutation stage, the positioning points on the right side of the symmetry axis are mutated first. The mutated positioning points are required to be a certain distance away from the symmetry axis, otherwise, the two positioning points are too close to each other, which may cause stress concentration. At this time, the mutated positioning points p ’(x’, y’) satisfy:

[0151]

[0152] In the formula, Δ represents the change amount, δ and the range. For the positioning points on the left side of the symmetry axis, the coordinates are directly obtained by symmetry calculation.

[0153] For the inner and outer side wall plates, the positioning layout is also constrained first. Taking the inner side wall plate as an example.

[0154] At this time, the feasible space of the positioning layout becomes three rectangular spaces separated by the reinforcing ribs:

[0155]

[0156] In the formula, the lower subscript rec i represents that the point is located in the th i rectangular space.

[0157] For the two wall plates, theoretically, the positioning points can be arranged arbitrarily in the three rectangular spaces. However, considering the feasibility of actual operation and the relative regularity of the components, it is stipulated that the coordinates of the positioning points in each rectangular space should be consistent. z

[0158] In the mutation stage, the point with the maximum y coordinate in each rectangular region is mutated first, and then the z coordinate of the positioning point in this region is uniquely determined. At this time, the y coordinates of other positioning points can be varied within the specified range. In the positioning layout of the two wall plates, the positioning points are still required to be a certain distance away from the symmetry axis.

[0159] For the reinforcing wall plate, there are similar constraints.

[0160] (3) The optimization target is specified, and the objective function f 1 is established for maximizing I GU , f 2 is established for maximizing I LS , f 3 is established for minimizing I ​ED :

[0161]

[0162] In the formula, PF2SF-GAN is the deep learning model constructed above. Layout i For the first i A positioning layout.

[0163] (4) Specify the number of reference points n re :

[0164]

[0165] In the formula, C This indicates combinations that do not consider the order. m and p represent m One goal p There are several possible division scenarios. In this embodiment of the invention, m and p The reference points are 3 and 12 respectively, therefore the number of reference points and the initial population size should be 91. In addition, the crossover and mutation probabilities are set to 0.6 and 0.1 respectively, and the maximum number of generations is set to 100.

[0166] Step 4, the experiment, specifically includes:

[0167] Step 41: Select the optimal solution from all Pareto fronts obtained in step 31 as the final positioning layout;

[0168] Step 42: Using the reconfigurable tooling with the configuration and positioning layout built by S41, conduct positioning tests on each component;

[0169] Step 43: Measure the position of key measurement points using a coordinate measuring machine, compare the difference between the original positioning layout and the nominal position of the key measurement points under the designed positioning layout, and verify whether the positioning design method can reduce positioning deformation.

[0170] Step 44: Establish a finite element simulation model and compare the difference between the average and maximum stress distribution of the component in the positioning state under the original positioning layout and the designed positioning layout.

[0171] The above-disclosed embodiments are merely preferred embodiments of the present invention, but the present invention is not limited thereto. Any non-creative variations that can be conceived by those skilled in the art, as well as any improvements and modifications made without departing from the principles of the present invention, should be within the protection scope of the present invention.

Claims

1. A method for intelligent design of positioning layout of aerospace thin-walled components guided by stiffness distribution, characterized in that, Includes the following steps: S1. Establish a finite element simulation model of the thin-walled component in the positioning state, save the feasible region of the positioning point as a set of candidate nodes, and obtain the input file for the finite element simulation. ； Ω represents the feasible region of the location point; the area enclosed by the outer boundary is... P outer The area enclosed by the inner boundary is P inner , P i For the set of location points; Based on the initial simulation input file, the location point numbers in the input file are randomly and batch-rewritten to form a large number of input files corresponding to different location layouts; according to the input files and the geometric model of the thin-walled component, location layout feature maps expressing the component shape, stiffeners and location points are drawn; finite element simulations are performed on all generated input files in batches, and the deformation of each node in the thickness direction of the component is output in each simulation result; a smooth mesh is created using existing data points for interpolation; an instance feature encoder, a local generator and a global generator are constructed, and minimum pooling is added to the convolutional block attention module to help the model identify weak or subtle features; the loss of the deep learning network is defined. S2, considering the pixels in the planar region of the component, construct a global uniformity index to describe the uniformity of stiffness distribution globally; define the neighborhood range of each pixel and construct a local smoothness index. Used to measure how sharp the numerical value of each pixel is within its neighborhood; ； Where N( i , j )express( i , j The set of neighborhoods of a given pixel; |N| is the number of pixels in the neighborhood; f p,q Represents the physical quantity value of a pixel in the neighborhood; M and N These represent the number of pixel rows and the number of pixel columns, respectively. f i,j express( i , j The value corresponding to the pixel at that location; It is the global mean of the physical quantity; Construct an extreme value difference index to quantify the relative difference between the maximum and minimum values, and impose constraints on the positioning layout generated during the iteration process; S3. Obtain a large dataset of paired positioning layout process feature maps and component stiffness cloud maps, and annotate them; use the annotated dataset to train a deep learning model. S4, a flexible tooling with reconfigurable positioning layout, is used to build a test platform. The assembly hole in the test structure is used as the key measurement point. The position of the point is measured to verify whether the positioning design method can reduce positioning deformation. The center of the circle is obtained by measuring the position of any three points around the hole. Three sets of measurements are taken at each point and the average value is used as the final hole position.

2. The intelligent design method for the positioning layout of aerospace thin-walled components guided by stiffness distribution according to claim 1, characterized in that, The following strategies are used to draw stiffness contour maps: For components without openings or ring structures, different colors are directly assigned to different nodes, and continuous contour maps are drawn through color interpolation; For components with openings or ring structures, the edge points of the component are detected, and the distance and angle from any given point to other points are calculated.

3. The intelligent design method for the positioning layout of aerospace thin-walled components guided by stiffness distribution according to claim 1, characterized in that, Edge points are clustered using density-based clustering algorithm to identify the interior and exterior of components. The exterior and interior boundaries are calculated using convex hull. If a grid node is located within the interior convex hull, no interpolation is performed.

4. The intelligent design method for positioning and layout of aerospace thin-walled components guided by stiffness distribution according to claim 1, characterized in that, The encoder extracts features at the scale of the original stiffness contour map through downsampling-upsampling.

5. The intelligent design method for positioning and layout of aerospace thin-walled components guided by stiffness distribution according to claim 1, characterized in that, A set of cascaded residual blocks is added to the deepest layer of the encoder-decoder structure, and an improved convolutional block attention module is incorporated into the residual blocks.

6. The intelligent design method for the positioning layout of aerospace thin-walled components guided by stiffness distribution according to claim 1, characterized in that, The local generator adopts a progressive multi-scale processing strategy to refine the stiffness field step by step from low resolution to high resolution. Each local generator first performs a downsampling operation on the input map at the current resolution to extract features, then fuses these features with the features of the previous level, and the fused features are then upsampled to restore the resolution.

7. The intelligent design method for the positioning layout of aerospace thin-walled components guided by stiffness distribution according to claim 1, characterized in that, A hierarchical PatchGAN architecture is used to build the discriminator, which evaluates the authenticity of the stiffness field at multiple resolution levels. The output single-channel feature map is used for authenticity judgment, forming a multi-scale discrimination mechanism from high resolution to low resolution.

8. The intelligent design method for the positioning layout of aerospace thin-walled components guided by stiffness distribution according to claim 1, characterized in that, Constraints are applied to the positioning layout generated during the iteration process to ensure the physical consistency of the generated scheme. The constraints are divided into feasibility constraints and positioning relationship constraints. Feasibility constraints ensure that the positioning layout appears in the part area, and positioning relationship constraints ensure that the generated positioning layout meets the requirements.

Images