Method and device for predicting assembly deformation of aircraft composite material structures based on surrogate model

Abstract

The present invention discloses a method for predicting the assembly deformation of aircraft composite material structures based on a proxy model. By establishing a numerical simulation reference model of aircraft composite materials, a data basis is provided for a data-driven perception algorithm. A two-stage proxy model framework is proposed, and an attention-enhanced convolutional compression autoencoder is designed to extract nonlinear structures from high-dimensional discrete field variable data. By achieving efficient compression and precise reconstruction of high-dimensional field variable data, the ultimate prediction accuracy of the proxy model is ensured. A variety of indicators are introduced to illustrate the effect of the proposed model. The present invention also provides a device for predicting the assembly deformation of aircraft composite material structures. The method provided by the present invention can reconstruct the displacement and stress field according to the actual status of each digital tooling at the assembly site, so as to solve the problem of real-time and precise prediction of the displacement and stress field at the aircraft composite material structure assembly site.

Description

Technical Field

[0001] The present invention belongs to the field of aircraft automated assembly, and in particular relates to a method and device for predicting aircraft composite material structure assembly deformation based on an agent model. Background Art

[0002] Aircraft are highly complex products, and their assembly process is crucial to ensuring final quality and safety. Thousands of parts are assembled through multiple stages to form the aircraft's main structure. The assembly process primarily involves steps such as positioning, hole making, and connection. Each stage of the process also places stringent requirements on the aircraft's structural appearance. Furthermore, due to the widespread use of carbon fiber reinforced composites (CFRPs) in aircraft, the manufacturing characteristics of their laminated layup place even higher demands on the assembly process. In addition to focusing on the shape accuracy of composite structures, their stress distribution should also be included in assembly quality evaluation criteria. However, there is currently a lack of full-field stress sensing technologies and methods suitable for assembly sites.

[0003] Therefore, intelligent assembly technology based on data-driven approaches has emerged. However, due to the low production volumes and long cycle times of aircraft manufacturing, it is difficult to directly collect large amounts of data from the production process to construct data-driven models. Numerical simulation models, due to their accuracy and low cost, have become increasingly widely used in actual manufacturing processes. However, numerical models of composite structures introduce higher nonlinearities and significantly increase the discrete dimensions of field variables, resulting in longer solution times and making it difficult for traditional numerical simulations to quickly respond to the changing conditions in the actual assembly site. These shortcomings severely limit their application in aircraft assembly processes. Surrogate models, driven by simulation data, provide an effective solution for simulating and analyzing complex systems. Surrogate models bridge the gap between the parameter space and model results without focusing on the details of model construction. However, the construction of surrogate models is also prone to insufficient prediction accuracy. Generally, the greater the complexity of the problem, the more simulation samples are required. In summary, there is an urgent need to develop a method for constructing accurate surrogate models using limited data to achieve comprehensive perception of the displacement and stress fields of composite structures at the assembly site.

[0004] Patent document CN117973142A discloses a method for assembly control of thin-walled composite parts driven by numerical simulation. This method uses the construction of the stiffness distribution law of thin-walled composite parts and the prediction of assembly deformation of thin-walled composite parts under the action of continuous error accumulation as important supports for numerical simulation. Through the shape control and compensation of thin-walled composite parts under multiple assembly conditions, the key assembly processes are predicted and optimized.

[0005] Patent document CN115630558A discloses a method for predicting the assembly deformation of composite components, which includes the following steps: S1. Obtaining point cloud data of the composite component to be predicted, and preprocessing the acquired point cloud data to obtain noise-free point cloud data N; S2. Building a deep neural network prediction model based on a multivariate linear regression model and an RBF neural network; S3. Training the deep neural network prediction model to obtain an optimal deep neural network prediction model; S4. Using the optimal deep neural network prediction model to predict the deformation of the composite component to be predicted. Summary of the Invention

[0006] The purpose of the present invention is to provide a method and device for predicting deformation of aircraft composite structure assembly based on a proxy model. The method can reconstruct the displacement and stress fields according to the actual status of each digital tooling at the assembly site, so as to solve the problem of real-time and accurate prediction of displacement and stress fields at the aircraft composite structure assembly site.

[0007] To achieve the first objective of the present invention, the following technical solution is provided: a method for predicting aircraft composite structure assembly deformation based on a proxy model, comprising the following steps:

[0008] Acquire aircraft assembly information, including the specific positioning and connection methods of the assembly structure, and determine the range of motion of each digital tooling based on the acquired aircraft assembly information;

[0009] Sampling is performed according to the motion range of each digital tooling to construct a uniform and discretely distributed input parameter sample set;

[0010] Constructing a numerical simulation reference model based on the material properties of the assembly structure and specific ply information, and meshing the numerical simulation reference model;

[0011] Setting the boundary condition form of the numerical simulation reference model according to the acquired aircraft assembly information, and using each sample point in the sample set as the boundary condition of the numerical simulation reference model to solve the stress and displacement distribution of the structure under different boundary conditions;

[0012] Each input parameter sample point and the corresponding stress and displacement distribution form a data set;

[0013] A two-stage surrogate model was constructed. The first stage of the surrogate model used a compressed convolutional autoencoder, which includes an encoder for compressed data and a decoder for reconstructed data. The autoencoder then extracted nonlinear features of stress and displacement distributions through self-supervised learning to obtain low-dimensional embeddings of high-dimensional field variables in a latent variable space. In the second stage, a multilayer perceptron was used to establish a mapping relationship between aircraft assembly information and the low-dimensional embeddings.

[0014] During prediction, the aircraft assembly information is input into a multi-layer perceptron to obtain a low-dimensional embedding, which is then input into the decoder in the compressed convolutional autoencoder to obtain a high-dimensional stress and displacement field.

[0015] The proxy model is trained using the dataset to obtain an assembly deformation proxy model for predicting stress and displacement distributions.

[0016] The present invention establishes a numerical simulation model of a composite material structure, takes into account its intra-laminar damage and inter-laminar failure, adopts Latin hypercube sampling based on genetic algorithm to design parameter space sample points, establishes a composite material structure assembly deformation dataset through parametric modeling technology, and uses the composite material structure assembly deformation dataset to train a target proxy model to obtain the optimal proxy model.

[0017] Specifically, the original Latin hypercube sampling is used as the initial value of the motion range sampling of the digital tooling, and the average distance from each design point to other design points is used as the optimization target to obtain a uniform and discretely distributed sample set through genetic algorithm.

[0018] Specifically, the numerical simulation reference model predicts the mechanical properties based on the material direction and thickness of each layer, the damage within the single layer material, and the delamination damage between layers.

[0019] Specifically, the delamination damage is analyzed using the 3D Hashin criterion as the failure criterion, which specifically includes:

[0020] Fiber tensile failure:

[0021]

[0022] Fiber compression failure:

[0023]

[0024] Resin tensile failure

[0025]

[0026] Resin compression failure:

[0027]

[0028] Among them, X 11,t and X 22,t Indicates the tensile strength of the material in direction 1 and direction 2, X 11,c and X 22,c represent the compressive strength in direction 1 and direction 2, respectively, and X 12 and X 23 represents shear strength; σ11 , σ 22 , σ 33 represents the stress components in the 1st, 2nd, and 3rd directions in the fiber-reinforced region; σ 12 , σ 13 , σ 23 represents the shear stress component in the fiber reinforced region; f f,t 、f f,c 、f m,t and f m,c It is an indicator of failure of the fiber and matrix resin, where damage occurs when the index is greater than 1;

[0029] The cohesion element based on the bilinear hybrid constitutive model is used to simulate the delamination behavior. The specific process is as follows:

[0030]

[0031] where K i (i=n,st, represents the stiffness of the cohesion unit, t i (i=n,s,t) and δ i (i = n, s, t) are the nominal stress and displacement in the normal or shear direction, respectively;

[0032] The initial delamination of the delamination behavior is predicted based on the quadratic stress criterion, neglecting the influence of the normal stress:

[0033]

[0034] where N, S, and T are the strengths in the normal and in-plane shear directions, respectively, and the damage evolution of delamination is predicted using the energy-based Benzeggah-Kenane criterion:

[0035]

[0036] Among them G c is the total fracture energy, G iC (i=Ι,ΙΙ) are normal and shear fracture energies respectively, where G S and G T are the out-of-plane shear and total dissipated energy in three directions, respectively, and η is the BK ratio.

[0037] Specifically, the compressed convolutional autoencoder is constructed using a convolutional neural network combined with residual connections and an attention mechanism, and its expression is as follows:

[0038] φ:x→f

[0039] ψ:f→x

[0040]

[0041] Here, φ represents the encoder, f represents the latent space representation, ψ represents the decoder, and x represents the field variables predicted by the proxy model.

[0042] Specifically, the attention mechanism enhances the feature extraction capability of the compressed convolutional autoencoder through three independent branches, including channel attention, spatial attention, and depth attention.

[0043] Specifically, during the training process, the mean square error is used as the training loss function of each neural network, and the coefficient of determination, signal-to-noise ratio, and multi-scale structural similarity index are combined as indicators for optimizing and evaluating the proxy model.

[0044] To achieve the second objective of the present invention, the following technical solution is provided: a device for predicting deformation of an aircraft composite material structure assembly, which is implemented by the above-mentioned agent-based model-based method for predicting deformation of an aircraft composite material structure assembly. The specific implementation process is as follows:

[0045] The aircraft assembly information of the aircraft to be assembled is input into the assembly deformation proxy model to obtain the stress and displacement distribution of each structure of the aircraft to be assembled.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] By establishing a numerical simulation model of aircraft structures, a data foundation is provided for data-driven perception algorithms. A two-stage surrogate model framework is proposed, using an attention-enhanced convolutional compression autoencoder for nonlinear feature acquisition. This achieves efficient compression and precise reconstruction of high-dimensional field variable data, ensuring the ultimate prediction accuracy of the surrogate model.

[0048] In addition, the trained proxy model can simultaneously predict the stress field and displacement field, providing a more comprehensive evaluation index for deformation monitoring during the aircraft assembly process, thus overcoming the difficulty of directly measuring the stress of complex aircraft structures at the assembly site. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A flowchart of the aircraft composite material structure assembly deformation prediction method based on the proxy model provided in this embodiment;

[0050] Figure 2 The projection of the Latin hypercube sampling result based on the genetic algorithm provided in this embodiment in two-dimensional space;

[0051] Figure 3 A schematic diagram of the specific positioning of the aircraft fuselage provided in this embodiment;

[0052] Figure 4Architecture diagram of the attention-enhanced compressed autoencoder provided for this embodiment. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0054] like Figure 1 As shown in FIG, a method for predicting aircraft composite material structure assembly deformation based on a proxy model is provided in this embodiment, and the specific steps are as follows:

[0055] Acquire aircraft assembly information, including the specific positioning and connection method of the assembly structure, and determine the motion range of each digital tooling based on the acquired aircraft assembly information.

[0056] like Figure 2 As shown, this embodiment provides a flowchart for positioning a large aircraft fuselage. The specific process is as follows: considering that the positioning process in aircraft assembly has a significant impact on the final manufacturing accuracy of the entire product, a positioning deformation proxy model of the fuselage structure is constructed. Since the position and shape of the aircraft fuselage are adjusted by CNC positioners, the spatial coordinates of each CNC positioner are defined as the input variables of the proxy model. Each positioner has three degrees of freedom, so the input variables of the proxy model have 12 dimensions.

[0057] The Latin hypercube sampling method based on genetic algorithm optimization is used to sample the motion range of each digital tooling to construct a sample set with uniform and discrete distribution.

[0058] Furthermore, a set of original Latin hypercube sampling results is used as initial values, which are a set of n points in a d-dimensional space.

[0059] The Latin Hypercube sampling scheme is then evaluated according to a fitness function that takes the pairwise Euclidean distance between sampling points as the fitness function and aims to maximize the spatial coverage:

[0060]

[0061] Among them, xi and x j is a point in the Latin hypercube sampling design X, ||x i -x j || is the Euclidean distance between them, with the inverse square penalizing shorter distances.

[0062] The best performing individuals are selected from the current group to generate the next generation. The present invention randomly selects a certain number of individuals from the group and takes the individuals with the highest fitness values ​​as the parents. The generation of offspring can be generated by combining two or more parent samples through a simple crossover operation:

[0063] Child k =α·Parent1+(1-α)·Parent2

[0064] Here, α is a random factor between 0 and 1, indicating that the offspring generated will have a mixture of characteristics from both parents. In addition, random mutations can be introduced into the offspring to maintain genetic diversity and explore new regions of the solution space.

[0065] A certain degree of random mutation is introduced by randomly adjusting the position of samples within the sampling interval:

[0066] x i =x i +δ·(upper bound i -lower bound i )

[0067] Here, δ is a small random perturbation, and the above optimization process is repeated a specified number of times or until the fitness no longer improves significantly.

[0068] According to the material properties of the structure and the specific ply information, a numerical simulation reference model is constructed to mesh the mechanical properties in the assembly space. This embodiment is constructed using the general Abaqus software machine and developed using the finite element method. Consistent with its actual additive manufacturing process, the model is established layer by layer, using a single-layer solid element for each layer and zero-thickness cohesion elements between layers. When establishing the simulation model, damage within the laminate and possible delamination damage between layers are also considered.

[0069] Furthermore, the delamination damage is analyzed using the 3D Hashin criterion as a failure, which specifically includes:

[0070] Fiber tensile failure:

[0071]

[0072] Fiber compression failure:

[0073]

[0074] Resin tensile failure

[0075]

[0076] Resin compression failure:

[0077]

[0078] Among them, X 11,t and X 22,t Indicates the tensile strength of the material in direction 1 and direction 2, X 11,c and X 22,c represent the compressive strength in direction 1 and direction 2, respectively, and X 12 and X 23 represents shear strength; σ 11 , σ 22 , σ 33 represents the stress components in the 1st, 2nd, and 3rd directions in the fiber-reinforced region; σ 12 , σ 13 , σ 23 Represents the shear stress component in the fiber reinforced area. In this embodiment, direction 1 is the longitudinal direction, direction 2 is the transverse direction, and direction 3 is the vertical direction.

[0079] f f,t 、f f,c 、f m,t and f m,c It is an indicator of failure of the fiber and matrix resin, where damage occurs when the index is greater than 1;

[0080] The cohesion element based on the bilinear hybrid constitutive model is used to simulate the delamination behavior. The specific process is as follows:

[0081]

[0082] where K i (i=n,st, represents the stiffness of the cohesion unit, t i (i=n,s,t) and δ i (i = n, s, t) are the nominal stress and displacement in the normal or shear direction, respectively;

[0083] The initial delamination of the delamination behavior is predicted based on the quadratic stress criterion, neglecting the influence of the normal stress:

[0084]

[0085] where N, S, and T are the strengths in the normal and in-plane shear directions, respectively, and the damage evolution of delamination is predicted using the energy-based Benzeggah-Kenane criterion:

[0086]

[0087] Among them G c is the total fracture energy, G iC (i=Ι,ΙΙ) are normal and shear fracture energies respectively, where G S and G T are the out-of-plane shear and total dissipated energy in three directions, respectively, and η is the BK ratio.

[0088] Table 1 shows the material properties provided in this embodiment.

[0089] Table 1

[0090]

[0091] The boundary condition form of the numerical simulation reference model is set according to the acquired aircraft assembly information, and each sample point in the sample set is used as the boundary condition of the numerical simulation reference model to solve the stress and displacement distribution under different boundary adjustments.

[0092] The displacements of all finite element nodes and the stresses at all element integration points are extracted, and all input sample points and displacement and stress field values ​​are integrated to construct a data set.

[0093] A two-stage proxy model is constructed. The first stage of the proxy model uses a compressed convolutional autoencoder, and the second stage uses a multilayer perceptron. During prediction, aircraft assembly information is input into the multilayer perceptron to obtain a low-dimensional embedding. This low-dimensional embedding is then input into the decoder of the autoencoder to obtain the high-dimensional stress and displacement fields. Furthermore, the proxy model in this embodiment uses an attention-enhanced compressed convolutional autoencoder for self-supervised training of the input stress and displacement distributions to obtain the corresponding low-dimensional representations. Due to the characteristics of the autoencoder, after training, both an encoder for dimensionality reduction and a decoder for restoring the data to a high-dimensional space are obtained. An autoencoder is a self-supervised neural network used to learn efficient encoding of data. It consists of two parts: an encoder φ compresses the input data into a latent space representation f, while a decoder ψ attempts to reconstruct the original input data from this latent representation. The goal of the autoencoder is to minimize the reconstruction error, allowing the network to capture the most important features in the data. The principle behind this is expressed as follows.

[0094] φ:x→f

[0095] ψ:f→x

[0096]

[0097] like Figure 4 As shown, the autoencoder in this embodiment adopts a convolutional neural network as the main structure. This autoencoder is also called a convolutional compression autoencoder. By using convolution layers and pooling layers, feature extraction and downsampling are performed on the input data, thereby achieving effective encoding and reconstruction of high-dimensional data.

[0098] Residual blocks are integrated into the network to enhance its performance. The key to residual blocks is the use of skip connections, which add the input of the hidden layer directly to the output. During training, these skip connections can promote the flow of gradients, thereby alleviating the vanishing gradient problem and making it possible to build deeper networks.

[0099] The triplet attention mechanism is used to enhance the feature representation capabilities of convolutional neural networks by focusing attention across spatial and channel dimensions. The triplet attention mechanism consists of three branches: two branches are dedicated to capturing the interaction between the channel dimension C and the spatial dimension W / H, and the remaining branch is responsible for calculating the traditional spatial attention weights.

[0100] The first branch builds attention between the height dimension and the channel dimension, given an input tensor X∈R C×H×W , first rotate it 90° counterclockwise along the H axis. The shape of the rotated tensor is (W×H×C), which is recorded as Substituting it into the following formula, we can obtain the feature map corrected by attention. Finally, the feature is rotated 90° clockwise along the H axis to restore the data to the original input shape of X.

[0101]

[0102] Among them A hc is the feature modified by attention. w This consists of a Z-pool module and a standard convolutional layer, as shown in the following formula. σ is a nonlinear Sigmoid activation function. The activation function generates the channel attention weights, and ⊙ represents the Hadamard product between the attention weights and the input features.

[0103]

[0104] Where 0d refers to the maximum and average pooling operations in the 0th dimension.

[0105] Similarly, in the second branch, the input feature is first rotated 90° counterclockwise along the W axis, and the resulting tensor is recorded as Substituting it into the following formula can obtain the attention-corrected feature, and finally the feature is also rotated clockwise along the W axis to restore the original shape.

[0106]

[0107] Among them A wc is the attention-modified feature, f h There is f w The same structure also consists of a Z-pool module and a standard convolutional layer.

[0108] For the last branch, the original input is directly substituted into the following without rotating the input features.

[0109] A wh (X)=σ(f c (X))⊙X

[0110] Finally, the features generated by each branch are aggregated through average pooling to complete the calculation of the entire attention module.

[0111] The proxy model uses a multi-layer perceptron to establish the relationship between input parameters and low-dimensional embeddings to achieve accurate mapping of input parameters to high-dimensional stress and displacement fields. The specific principles are as follows:

[0112] After the data is compressed in the first stage, when constructing the proxy model in the low-dimensional space, a relatively simple multi-layer perceptron can be used to fit the mapping relationship between the input parameters and the latent variables. The multi-layer perceptron structure in the second stage of the proxy model is designed to have four hidden layers, with 64 neural nodes in each layer.

[0113] Once the mapping between input parameters and latent variables is established, a proxy model of the displacement and stress field of the aircraft structure under the action of boundary conditions can be established. During the assembly process, the control parameters are input into the second-stage multi-layer perceptron to obtain the corresponding low-dimensional representation of the displacement and stress field. This low-dimensional representation is then input into the decoder of the attention-enhanced compressed autoencoder to achieve real-time perception of the entire displacement and stress field of the aircraft structure.

[0114] The surrogate model is trained using the dataset. The training process uses the mean square error as the training loss function for each neural network, and combines the coefficient of determination, signal-to-noise ratio, and multi-scale structural similarity index as indicators for optimizing and evaluating the surrogate model. Since the displacement of each node and the principal stress at the integration point of the discrete field quantity contain three components, after preprocessing, the high-dimensional field variable data can be converted into a three-dimensional array similar to an image. The mean square error is a commonly used indicator used to quantify the difference between the original data and its compressed form. Its calculation formula is:

[0115]

[0116] Among them, I orig (i, j) represents the displacement or stress value at the node or integration point (i, j) in the field quantity data, I comp (i, j) represents the value of the compressed and reconstructed field quantity data at the same location, while m and n represent the number of rows and columns in the transformed field quantity data array, respectively. The loss function measures the mean squared error of each node value, reflecting the degree of distortion or loss caused by data compression. Lower values ​​of the loss function indicate higher accuracy in the reconstructed field quantity data, effectively preserving the quality of the original data while compressing it.

[0117] And R 2 Also known as the coefficient of determination, it is a common indicator for evaluating the predictive accuracy of a single feature in regression analysis. In the case of multi-output regression tasks, it is calculated by R 2 The value of can be used to analyze the relative importance and value of different features. In addition, the R 2 Statistical analysis of the values ​​can also be used to optimize the constructed regression model, which helps to refine and enhance the model.

[0118] For a given feature, is the predicted value of the i-th sample, y i is the corresponding label, R 2 The value is obtained by using the following formula in N s Calculated on samples:

[0119]

[0120] in R 2 The maximum value of R is 1.0, indicating that the model fits the data perfectly. 2 Negative values ​​indicate poor model performance.

[0121] The structural similarity index (SSI) is a widely used image quality assessment metric that mimics the human visual system's ability to extract structural information from images. While designed to measure the similarity between two images, the high-dimensional field variable data studied in this paper can be converted into image-like three-dimensional arrays, making the SSI also applicable to predictive evaluation of field variable data reconstruction. The SSI is calculated by multiplying the brightness, contrast, and structural similarity between samples x and y.

[0122] SSIM(x,y)=[l(x,y) α c(x,y) β s(x,y) γ ]

[0123] Where l(x,y) is a brightness comparison, c(x,y) is a comparison based on contrast, and s(x,y) is a structure comparison.

[0124] Based on the structural similarity index, the multi-scale structural similarity index improves it. This method takes the integration of the reference image and the distorted image signal as input, and then iteratively applies a low-pass filter to downsample the image by a factor of 0.5. The resolution of the original image is denoted as Scale1, and the resolution of the image after M-1 iterations is denoted as Scale M The contrast index c(x,y) and structure index s(x,y) are calculated at each scale, and the brightness index l(x,y) is calculated at the final scale. By integrating the results of measurements at different scales, the comprehensive index MS-SSIM is obtained.

[0125]

[0126] where α M , β j , γ j is a coefficient used to adjust the relative importance of different components.

[0127] The signal-to-noise ratio (SNR) is a key metric used to quantify the quality of compressed and reconstructed data. It is calculated by dividing the signal power by the noise power that affects the fidelity of the reconstructed data. The formula for calculating the SNR is as follows:

[0128]

[0129] Among them I orig (i, j) is the displacement or stress value at the node or integration point (i, j) in the field data, I comp (i, j) represents the value of the compressed and reconstructed field data at the same position, and m and n represent the number of rows and columns in the transformed field data array, respectively. This paper uses logarithmic decibels to express the signal-to-noise ratio, which more clearly characterizes the reconstruction quality. Higher signal-to-noise ratios indicate lower noise levels and better reconstruction quality.

[0130] This embodiment further provides an aircraft composite material structure assembly deformation prediction device, which is implemented by the aircraft composite material structure assembly deformation prediction method based on the proxy model provided in the above embodiment. The specific implementation process is as follows:

[0131] The aircraft assembly information of the aircraft to be assembled is input into the assembly deformation proxy model to obtain the stress and displacement distribution of each structure of the aircraft to be assembled.

[0132] In addition, the terms "upper", "lower", "inner", "outer", "front", and "back" are used for descriptive purposes only and should not be understood as indicating or implying relative importance. Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the present invention.

[0133] Of course, the above description is only a specific embodiment of the present invention and is not intended to limit the scope of implementation of the present invention. Any equivalent changes or modifications made based on the structure, features and principles described in the scope of the patent application of the present invention should be included in the scope of the patent application of the present invention.

[0134] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the above-described embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above-described embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein with equivalents. Such modifications, changes, or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A method for predicting aircraft composite structure assembly deformation based on a proxy model, characterized in that: The following steps are involved: Acquire aircraft assembly information, including the specific positioning and connection methods of the assembly structure, and determine the range of motion of each digital tooling based on the acquired aircraft assembly information; Sampling is performed according to the motion range of each digital tooling to construct a uniform and discretely distributed input parameter sample set; Constructing a numerical simulation reference model based on the material properties of the assembly structure and specific ply information, and meshing the numerical simulation reference model; Setting the boundary condition form of the numerical simulation reference model according to the acquired aircraft assembly information, and using each sample point in the sample set as the boundary condition of the numerical simulation reference model to solve the stress and displacement distribution of the structure under different boundary conditions; Each input parameter sample point and the corresponding stress and displacement distribution form a data set; A two-stage surrogate model was constructed. The first stage of the surrogate model used a compressed convolutional autoencoder, which includes an encoder for compressed data and a decoder for reconstructed data. The autoencoder then extracted nonlinear features of stress and displacement distributions through self-supervised learning to obtain low-dimensional embeddings of high-dimensional field variables in a latent variable space. In the second stage, a multilayer perceptron was used to establish a mapping relationship between aircraft assembly information and the low-dimensional embeddings. During prediction, the aircraft assembly information is input into a multi-layer perceptron to obtain a low-dimensional embedding, which is then input into the decoder in the compressed convolutional autoencoder to obtain a high-dimensional stress and displacement field. The proxy model is trained using the dataset to obtain an assembly deformation proxy model for predicting stress and displacement distributions.

2. The method for predicting aircraft composite material structure assembly deformation based on a proxy model according to claim 1, characterized in that: The original Latin hypercube sampling is used as the initial value for the motion range sampling of the digital tooling, and the average distance from each design point to other design points is used as the optimization target to obtain a uniform and discretely distributed sample set through genetic algorithm.

3. The method for predicting aircraft composite material structure assembly deformation based on a proxy model according to claim 1, characterized in that: The numerical simulation reference model predicts mechanical properties based on the material direction and thickness of each layer, as well as damage within a single layer of material and delamination damage between layers.

4. The method for predicting aircraft composite material structure assembly deformation based on a proxy model according to claim 3, characterized in that: The delamination damage is analyzed using the 3D Hashin criterion as the failure criterion, which specifically includes: Fiber tensile failure: Fiber compression failure: Resin tensile failure Resin compression failure: Among them, X 11,t and X 22,t Indicates the tensile strength of the material in direction 1 and direction 2, X 11,c and X 22,c represent the compressive strength in direction 1 and direction 2, respectively, and X 12 and X 23 represents shear strength; σ 11 , σ 22 , σ 33 represents the stress components in the 1st, 2nd, and 3rd directions in the fiber-reinforced region; σ 12 , σ 13 , σ 23 represents the shear stress component in the fiber reinforced region; f f,t 、f f,c 、f m,t and f m,c It is an indicator of failure of the fiber and matrix resin. When the index is greater than 1, it is judged that damage will occur. The cohesion element based on the bilinear hybrid constitutive model is used to simulate the delamination behavior. The specific process is as follows: where K i (i=n,s,t) represents the stiffness of the cohesion unit, t i (i=n,s,t) and δ i (i = n, s, t) are the nominal stress and displacement in the normal or shear direction, respectively; The initial delamination of the delamination behavior is predicted based on the quadratic stress criterion, neglecting the influence of the normal stress: where N, S, and T are the strengths in the normal and in-plane shear directions, respectively, and the damage evolution of delamination is predicted using the energy-based Benzeggah-Kenane criterion: Among them G c is the total fracture energy, G iC (i=Ι,ΙΙ) are normal and shear fracture energies respectively, where G S and G T are the out-of-plane shear and total dissipated energy in three directions, respectively, and η is the BK ratio.

5. The method for predicting aircraft composite material structure assembly deformation based on a proxy model according to claim 1, characterized in that: The compressed convolutional autoencoder is constructed using a convolutional neural network combined with residual connections and an attention mechanism, and its expression is as follows: φ:x→f ψ:f→x Here, φ represents the encoder, f represents the latent space representation, ψ represents the decoder, and x represents the field variables predicted by the proxy model.

6. The method for predicting aircraft composite material structure assembly deformation based on a proxy model according to claim 5, characterized in that: The proposed attention mechanism enhances the feature extraction capability of the compressed convolutional autoencoder through three independent branches, including channel attention, spatial attention, and depth attention.

7. The method for predicting aircraft composite material structure assembly deformation based on a proxy model according to claim 1, characterized in that: During the training process, the mean square error is used as the training loss function of each neural network, and the determination coefficient, signal-to-noise ratio and multi-scale structural similarity index are combined as indicators for optimizing and evaluating the proxy model.

8. An aircraft composite material structure assembly deformation prediction device, characterized in that: This is achieved by the aircraft composite material structure assembly deformation prediction method based on the proxy model as described in any one of claims 1 to 7.

Images