Assembly interface stress prediction method

Abstract

The invention discloses an assembly interface stress prediction method, which comprises the steps of remodeling an assembly error parameter into a first assembly surface stress image based on a constructed modal mutual mapping network; and based on the constructed image reconstruction network, performing resolution enhancement operation on the first assembly surface stress image to generate a second assembly surface stress image, the resolution of the second assembly surface stress image being greater than the resolution of the first assembly surface stress image. According to the scheme, through a two-stage architecture, Legendre polynomial compression parameters and a multi-scale residual module capture features, key areas are enhanced in combination with an attention mechanism, collaborative optimization of efficiency and precision is achieved, dimension disasters are solved, model generalization and robustness are improved, and the real-time requirement is met.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence recognition technology, and in particular to a method for predicting assembly interface stress. Background Art

[0002] In the field of high-precision mechanical system assembly, accurate prediction of contact interface stress distribution is directly related to the product's transmission performance, mechanical stability, and service life. In particular, in assembly scenarios such as bolted connections and interference fits of composite or metal components, assembly errors, material nonlinearity, and fluctuations in loading conditions can significantly alter the distribution of the contact stress field. However, the stress fields of complex assembly systems exhibit highly nonlinear and multi-scale characteristics, making it difficult for traditional prediction methods to achieve both computational efficiency and accuracy. Therefore, developing a stress prediction solution that combines high efficiency and high resolution has become a key technical requirement for assembly system design optimization and error compensation.

[0003] Traditional stress prediction schemes primarily rely on finite element analysis (FEA) and single deep learning models. The finite element method, by constructing high-precision meshes and contact models, can accurately calculate stress field distributions. Some studies have attempted to employ end-to-end deep learning models to directly map assembly parameters into stress images. However, when faced with large-scale variations in assembly parameters (such as deviations, interferences, and load fluctuations), the finite element method requires repeated meshing and iterative calculations, resulting in high computational costs. Single deep learning models, however, suffer from unstable training and poor detail recovery capabilities due to the large dimensional span between input parameters and stress images, making them difficult to meet the real-time and precision requirements of engineering scenarios.

[0004] The existing technology has the following core defects:

[0005] Computational efficiency and accuracy are difficult to achieve: Although the finite element method has high accuracy, it takes too long to calculate when dealing with multiple parameter changes, and cannot meet the real-time prediction needs of the assembly site;

[0006] Insufficient dimensional mapping stability: End-to-end deep learning models directly process the mapping of low-dimensional parameters to high-dimensional images, which can easily lead to overfitting due to the curse of dimensionality. This is especially true for areas with insufficient detail restoration in stress concentration areas.

[0007] Weak generalization ability: Traditional solutions are highly dependent on specific working conditions. When the assembly error parameters exceed the training range, the topological structure and amplitude accuracy of the stress prediction results will drop significantly. Summary of the Invention

[0008] An embodiment of the present application provides a method for predicting assembly interface stress.

[0009] Specifically, a method for predicting assembly interface stress includes:

[0010] Based on the constructed modal inter-mapping network, the assembly error parameters are reshaped into the stress image of the first assembly surface;

[0011] Based on the constructed image reconstruction network, a resolution enhancement operation is performed on the first assembly surface stress image to generate a second assembly surface stress image, wherein the resolution of the second assembly surface stress image is greater than the resolution of the first assembly surface stress image.

[0012] A method for predicting assembly interface stress, comprising:

[0013] Performing multi-scale changes on the stress image of the first assembly surface based on the multi-scale feature extraction module to generate multi-scale fusion features;

[0014] Based on the attention enhancement module, channel attention and spatial attention are calculated on the multi-scale fusion features to generate a global correlation feature map;

[0015] The global correlation feature map is resolution enhanced based on the resolution enhancement module to generate a second assembly surface stress image.

[0016] The technical solutions provided in the embodiments of the present application have at least the following beneficial effects:

[0017] (1) Two-stage architecture achieves coordinated optimization of efficiency and accuracy. This solution splits the traditional end-to-end mapping into a hierarchical processing flow of "parameters-low-resolution images-high-resolution images" through a two-stage architecture of "modal inter-mapping network + image reconstruction network". The modal inter-mapping network is based on Legendre polynomial expansion. In actual engineering applications, it can compress 50-dimensional assembly error parameters into 2028-dimensional polynomial coefficients, and quickly reconstruct a 500×500 first assembly surface stress image through a combination of orthogonal functions. The computational complexity is reduced by more than 70% compared with the finite element method (the finite element method requires repeated meshing, while this solution only requires forward calculation). The image reconstruction network is a cascade of multi-scale feature extraction modules and resolution enhancement modules (weights). In actual engineering implementation, it can ensure a resolution of 1000×1000 while controlling the inference time within 50ms, meeting the real-time requirements of the assembly site.

[0018] (2) Accuracy assurance of polynomial compression and residual learning The output layer of the modal inter-mapping network adopts a multi-Gaussian process structure, and realizes the global approximation of the stress field through Legendre polynomial expansion. Compared with the finite element method, for example, in a typical application scenario, the stress amplitude error is ≤3.2% under typical working conditions such as bolt connection; the multi-scale residual module of the image reconstruction network can capture local details such as the bottom of the thread and the global stress trend of the contact surface through the parallel calculation of 3×3 / 5×5 / 7×7 convolution kernels, and retain the original structural information through residual jump connection, so that the structural similarity index (SSIM) of the stress image of the second assembly surface is improved to above 0.92, which is better than the 0.78 of the traditional deep learning model.

[0019] (3) Dimension Matching Strategy for Parameter Compression and Feature Dimension Upscaling The feature mapping layer of the modal inter-mapping network uses a cascaded fully connected layer. For example, the 50-dimensional assembly error vector is layer-by-layer upgraded to a 2048-dimensional semantic feature through the LeakyReLU and Swish activation functions, and then mapped to a 2028-dimensional polynomial coefficient through a multi-Gaussian process structure. The input-output dimension ratio is optimized from the traditional 1:20,000 to 1:40, effectively alleviating the curse of dimensionality. Experimental data show that this design reduces the fluctuation range of the loss function during model training from ±25% to ±8%, avoiding the overfitting problem caused by the large dimensional span of the end-to-end model.

[0020] (4) Attention Mechanism Enhances Feature Expression in Key Regions The attention enhancement module of the image reconstruction network integrates the ACMix module and the self-attention mechanism. Channel attention enhances the features of stress concentration areas such as bolt holes through global average pooling (response value increased by 40%), spatial attention focuses on high-gradient areas such as the edges of the contact surface, and the self-attention mechanism models the long-range effect of preload on the distal contact surface, ensuring the symmetry of the stress distribution of symmetrical structures. This significantly improves the ability to restore details in stress concentration areas. For example, the edge clarity of the stress concentration zone at the bottom of the thread is three times higher than that of the traditional model, solving the problem of detail loss in key areas of the end-to-end model.

[0021] (5) The polynomial basis function's physical prior embedding modal intermapping network uses Legendre polynomials as an orthogonal basis, which adapts very well to the smoothness and symmetry constraints of the stress field. Compared with a purely data-driven end-to-end model, the stress field topology retention rate is improved from 58% to 89% under unseen assembly conditions (e.g., interference exceeds the training range by ±20%). The multi-Gaussian process structure introduces uncertainty estimation into the coefficient generation process, which improves the model's predictive robustness by 50% under small sample conditions, effectively solving the traditional solution's dependence on specific working conditions.

[0022] (6) Multi-scale Feature Fusion and Global Correlation Modeling The multi-scale feature extraction module of the image reconstruction network can capture stress features at a scale of 50-500 pixels in a specific application scenario through the cascade of initial feature extraction and multi-scale residual modules. The resolution enhancement module, for example, uses the PixelShuffle operation to achieve 2x lossless upsampling, which significantly enhances the model's adaptability to different assembly errors (such as bolt preload deviation and interference fit tolerance). Measured data show that when the assembly parameter combination exceeds the training set range, the stress prediction error growth rate of this scheme is 0.8% / parameter dimension, which is much lower than the 3.5% / parameter dimension of the traditional model, and has excellent generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0024] Figure 1 Flowchart of the assembly interface stress prediction method provided for the implementation of this application.

[0025] Figure 2 Schematic diagram of the assembly interface stress prediction system provided for the implementation of this application. DETAILED DESCRIPTION

[0026] To make the purpose, technical solutions, and advantages of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the specific embodiments of this application and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0027] like Figure 1 and 2 As shown, the assembly interface stress prediction includes: a modal inter-mapping network and an image reconstruction network, and the corresponding assembly interface stress prediction method includes:

[0028] Based on the constructed modal inter-mapping network, the assembly error parameters are reshaped into the stress image of the first assembly surface;

[0029] Based on the constructed image reconstruction network, a resolution enhancement operation is performed on the first assembly surface stress image to generate a second assembly surface stress image, wherein the resolution of the second assembly surface stress image is greater than the resolution of the first assembly surface stress image.

[0030] Optionally, the modality inter-mapping network includes an input layer, a feature mapping layer, and an output layer;

[0031] Based on the constructed modal inter-mapping network, the assembly error parameters are reshaped into the stress image of the first assembly surface, including:

[0032] Based on the input layer, the assembly error parameters are vectorized to obtain the assembly error vector;

[0033] Based on the feature mapping layer, the assembly error vector is mapped to the image semantic features based on the image channels;

[0034] Based on the output layer, image reconstruction is performed according to the image feature coefficients to obtain the first assembly surface stress image.

[0035] Optionally, the feature mapping layer is a cascaded multi-stage fully connected layer;

[0036] Based on the feature mapping layer, the assembly error vector is mapped to image semantic features based on image channels, including:

[0037] Based on the cascaded multi-level fully connected layers, the assembly error vector is subjected to layer-by-layer feature extraction of the mapping relationship between assembly participation and stress distribution, and the extracted features are subjected to layer-by-layer dimensionality upgrading to obtain image semantic features based on image channels.

[0038] Optionally, the output layer includes a multi-Gaussian process structure and an image reconstruction module,

[0039] Based on the output layer, image reconstruction is performed according to the image feature coefficients to obtain the stress image of the first assembly surface, including:

[0040] Based on the multi-Gaussian process structure, the image semantic features based on the image channel are expanded into Legendre polynomials to obtain the Legendre polynomial coefficients corresponding to the three channels L, A, and B of the image;

[0041] Optionally, based on the image reconstruction module, the Legendre polynomial coefficients corresponding to the three channels of the image L, A, and B are substituted into a two-dimensional orthogonal function to perform image reconstruction, and a first assembly surface stress image represented by LAB three-channel color is obtained.

[0042] Preferably, in a specific scenario, the above solution is preferably or alternatively implemented as follows:

[0043] 1. Input layer: structured vectorization of assembly parameters

[0044] Based on the input layer, the discrete assembly error parameters are converted into standardized vectors suitable for neural network processing, solving the structuring problem of input data.

[0045] For example, including the bolt preload deviation ΔF p , interference fit tolerance Δd, material elastic modulus fluctuation ΔE and other n-dimensional assembly error parameters (n = 10 ~ 50).

[0046] The specific implementation process is as follows:

[0047] 1. Parameter vectorization: Map the parameter set to an assembly error vector through a fully connected layer The expression is:

[0048] x=FC in (ΔF p ,Δd,ΔE,...), for this reason, it can be seen that through the linear transformation matrix (m is the number of parameters) and the bias vector b in Implement parameter reorganization: x = W in ·[ΔF p ,Δd,ΔE] T +b in Among them, W in Each element w in ij Indicates the jth physical parameter (such as ΔF p ) is the contribution weight of the i-th feature dimension, b in This process is used to compensate for the base offset of different parameters. This process maps parameters of different physical dimensions (such as force, length, modulus) into a unified feature space.

[0049] 2. Standardization: Use Batch Normalization for normalization:

[0050] To this end, dynamic normalization of the input parameters is implemented, where μ and σ are the mean vector and standard deviation vector of the parameters in the training batch, respectively, ∈ = 10 -5 is a smoothing factor to prevent the denominator from being zero. Through the affine transformation γ·x norm +β (γ and β are learnable scaling and translation parameters), preserving the variability of the data distribution.

[0051] To this end, a standardized assembly error vector is obtained to provide a unified input format for subsequent feature extraction.

[0052] 2. Feature Mapping Layer: Multi-level Fully Connected Semantic Feature Dimensionality Upgrading

[0053] Based on the feature mapping layer, through layer-by-layer feature extraction and dimensionality enhancement, a nonlinear mapping relationship between assembly parameters and stress distribution is established to solve the problem of large dimensional span.

[0054] For example, the feature mapping layer includes a cascade of four fully connected layers (256-512-1024-2048), each of which contains an activation function and a regularization operation. The detailed process is as follows:

[0055] Feature extraction process:

[0056] 1. Primary feature extraction (256 dimensions):

[0057] h1=LeakyReLU(FC1(x norm )), This layer captures the basic relationship between material parameters and stress amplitude through 256 neurons, and the weight matrix Each element w in ij represents the influence of the jth input feature on the ith hidden neuron. The negative slope parameter α = 0.2 ensures that negative input values ​​are not completely suppressed, preserving the negative impact of parameter fluctuations.

[0058] 2. Intermediate feature fusion (512 dimensions):

[0059] h2=Swish(FC2(h1)), Swish(x)=x·sigmoid(x), the Swish function realizes intelligent response to different assembly forms through adaptive gating mechanism. As an activation gate, when the input x is large, the gate is close to 1 and the function is approximately linear; when the input x is small, the gate is attenuated and the nonlinearity is enhanced.

[0060] 3. High-level semantic generation (1024 dimensions):

[0061] h3 = BN(FC3(h2)) + h2. The residual connection decomposes the stress field into a "basic solution + modified solution" model. h2 can be considered the basic stress field based on a simplified material mechanics model, while FC3(h2) is a correction term that takes into account complex factors such as geometric nonlinearity and contact effects. Batch Normalization standardizes the distribution of activation values ​​to ensure that the stress characteristics learned by each network layer have the same statistical properties.

[0062] 4. Dimension upscaling (2048 dimensions):

[0063] h4 = Dropout(0.2)(FC4(h3)). Dropout randomly inactivates neurons with a probability of 0.2. This process can be regarded as a simulation of material inhomogeneity. During training, each neuron has a probability of 0.2 of being ignored, forcing the network to learn more robust feature representations.

[0064] 3. Output layer: Multi-Gaussian process and image reconstruction module

[0065] (1) Multi-Gaussian Process (MOGP) Structure: Generation of Legendre Polynomial Coefficients

[0066] The output layer is expanded through Legendre polynomials to compress the 2048-dimensional semantic features into 2028-dimensional polynomial coefficients, and physical priors are introduced to solve the dimensionality curse.

[0067] The two-dimensional 25th-order Legendre polynomial expansion is:

[0068]

[0069] Among them, P i (x) is the i-th order Legendre polynomial, satisfying orthogonality:

[0070]

[0071] It can be seen that the expansion decomposes the stress field into a superposition of different frequency components, where i and j represent the order of the polynomial, which controls the frequency of spatial variation. i,j It represents the amplitude of the corresponding frequency component, which determines the contribution intensity of the stress field at this frequency.

[0072] The implementation principle of MOGP is as follows:

[0073] 1. Gaussian process regression: Mapping the 2048-dimensional feature h4 into a coefficient matrix Assume that the coefficient generation process follows a Gaussian distribution:

[0074]

[0075] where k(·) is the square exponential kernel function:

[0076]

[0077] It represents the signal variance and controls the overall fluctuation range of the output value. The length scale parameter l controls the smoothness of the function and determines how large the distance in the input space will lead to a significant change in the output.

[0078] 2. Uncertainty estimation: Through the covariance matrix of the Gaussian process, for each coefficient c i,j,k Provides probability distribution estimates.

[0079] Covariance matrix ∑ i,j,k The elements in ∑ ij Represents coefficient c i and c j The covariance between them reflects the statistical dependence between them. ii Represents coefficient c i The variance of , that is, the degree of uncertainty.

[0080] (2) Image reconstruction module: two-dimensional orthogonal function combination

[0081] Based on the image reconstruction module, the polynomial coefficients are substituted into the orthogonal function to reconstruct a 500×500 pixel LAB stress image, realizing the physical mapping from parameters to images.

[0082] Coordinate mapping: Normalize the image coordinates (u,v)∈[0,500]×[0,500] to the Legendre polynomial domain [-1,1]×[-1,1]:

[0083]

[0084] This mapping linearly scales the physical coordinates (u, v) to the standard domain of the Legendre polynomials [-1, 1], ensuring the validity of the polynomial expansion. The Jacobian matrix J = diag(2 / 500, 2 / 500) describes the local scaling factor of the coordinate transformation and is used in subsequent stress gradient calculations.

[0085] Image generation formula:

[0086]

[0087] I L (u, v) represents the brightness channel in the LAB color space, corresponding to the stress amplitude; I A (u,v) and I B (u, v) denotes the green-red and blue-yellow hue channels, respectively, corresponding to the stress gradient directions. The pixel value of each channel is obtained by summing the products of a 26×26 Legendre coefficient matrix and the corresponding basis function.

[0088] 4. Loss Function: Weighted Optimization of MSE and SSIM

[0089] Technical purpose: To optimize pixel accuracy and structural authenticity simultaneously to ensure that the reconstructed image conforms to physical laws.

[0090] Objective function:

[0091] in:

[0092] 1. Mean Squared Error (MSE):

[0093]

[0094] This loss function measures the difference between the predicted image I and the reference image The squared error at each pixel location, where H, W, and C represent the height, width, and number of channels of the image, respectively. MSE is sensitive to large errors and helps ensure the accuracy of stress amplitude.

[0095] 2. Structural Similarity (SSIM):

[0096]

[0097] μ I and Represents images I and The local mean of reflects the brightness information; and They represent local variance and reflect contrast information; represents the local covariance, reflecting structural information. C1 and C2 are stable constants that prevent the denominator from being zero.

[0098] In summary, in the above solution, based on the dimension conversion chain:

[0099]

[0100] This process achieves a triple mapping from "physical parameter space" to "mathematical feature space" and then to "engineering image space." Through layer-by-layer abstraction and dimensionality reduction, the model maps high-dimensional assembly parameters into low-dimensional polynomial coefficients rich in physical information, ultimately reconstructing an intuitive stress image.

[0101] Furthermore, the orthogonality and symmetry of Legendre polynomials accurately adapt to the smoothness and physical constraints of the stress field. The Legendre polynomials satisfy the orthogonality condition, making basis functions of different orders independent and avoiding redundant representation. Their symmetry matches the physical properties of the stress field (such as rotational and mirror symmetry), enabling a natural representation of stress distribution patterns in engineering structures.

[0102] Furthermore, in an aircraft engine turbine disk bolt assembly scenario, the network achieved stress image reconstruction within 50 milliseconds, a 200-fold improvement in computational efficiency compared to the finite element method. By leveraging pre-trained physical prior knowledge and an efficient neural network architecture, the model can rapidly map assembly parameters to stress distribution, avoiding the time-consuming meshing and iterative solution processes of traditional finite element methods and enabling real-time engineering analysis.

[0103] Optionally, the image reconstruction network includes: a multi-scale feature extraction module, an attention enhancement module, and a resolution enhancement module.

[0104] Based on the constructed image reconstruction network, a resolution enhancement operation is performed on the first assembly surface stress image to generate a second assembly surface stress image, including:

[0105] Performing multi-scale changes on the stress image of the first assembly surface based on the multi-scale feature extraction module to generate multi-scale fusion features;

[0106] Based on the attention enhancement module, channel attention and spatial attention are calculated on the multi-scale fusion features to generate a global correlation feature map;

[0107] The global correlation feature map is resolution enhanced based on the resolution enhancement module to generate a second assembly surface stress image.

[0108] Optionally, the multi-scale feature extraction module includes: an initial feature extraction module, a multi-scale residual module

[0109] Extract stress gradient edge features from the stress image of the first assembly surface based on the initial feature extraction module to obtain a basic feature map;

[0110] Based on the multi-scale residual module, local details and global trend features are extracted from the basic feature map and residual jump connections are performed on it to perform multi-scale changes and generate multi-scale fusion features accordingly.

[0111] Optionally, the attention enhancement module includes an ACMix module (Attention Convolution Mix) and a self-attention mechanism module;

[0112] Based on the ACMix module, the multi-scale fusion features are processed into stress concentration areas and spatial attention focused gradient edges to generate attention-enhanced feature maps;

[0113] Based on the self-attention mechanism module, channel attention and spatial attention calculations are performed on the attention-enhanced feature map to model the impact of preload on the contact surface and generate a global correlation feature map based on it.

[0114] Optionally, the resolution of the global correlation feature map is enhanced based on the resolution enhancement module to generate a second assembly surface stress image, including: based on the upsampling module, spatial detail recovery of the global correlation feature map to enhance the resolution and generate the second assembly surface stress image accordingly.

[0115] Preferably, in a specific scenario, the above solution is preferably or alternatively implemented as follows:

[0116] 1. Technical Implementation of Multi-scale Feature Extraction Module

[0117] 1. Initial feature extraction module

[0118] The gradient edge features of the stress field are extracted based on the convolution operation. The mathematical expression is:

[0119] F0=Conv1(I L )+Conv2(I A )+Conv3(I B )

[0120] Among them, I L ,I A ,I B They are the stress images of LAB channel ( The corresponding stress amplitude is corresponding to the gradient direction); Conv kRepresents the convolution operation of the kth channel, and its kernel function Encode stress gradient sensitive operators (such as Sobel operator) to extract gradient features of stress mutation areas such as bolt hole edges and contact surface boundaries; is the basic feature map, C0=64 is the number of initial feature channels, and each channel corresponds to the stress gradient response in a different direction.

[0121] 2. Multi-scale residual module

[0122] The local details and global trends of the stress field are captured by convolution kernels of different scales. The expression is:

[0123] F m =ResBlock m (F m-1 )+F m-1 ,m=1,2,3

[0124] The single-scale residual block is defined as:

[0125] ResBlock(F)=BN(ReLU((Conv 3×3 (F)))+BN(ReLU(Conv 5×5 (F)))

[0126] Among them, the 3×3 convolution kernel (K 3×3 ) captures the high gradient features of local stress concentration areas such as bolt holes, corresponding to the local singularity in elastic mechanics; the 5×5 convolution kernel (K 5×5 ) extracts the overall stress diffusion trend of the contact surface, which conforms to the stress attenuation law of Saint-Venant’s principle; Residual connection (ResBlock m (F)+F) simulates the "global solution + local correction" superposition principle of stress fields, such as the superposition of global tensile stress caused by preload and local stress concentration in bolt holes.

[0127] 3. Multi-scale fusion feature generation

[0128] F ms =Concat(F1,F2,F3)*W ms

[0129] Among them, F1, F2, and F3 are the features output by different residual blocks (corresponding to 16, 32, and 64 channels respectively), including stress features from micro to macro. The Concat operation splices multi-scale features. To fuse the weight matrix, the contributions of features of different scales are automatically distributed through learning (for example, high stress areas rely more on the detail features of 3×3 convolution).

[0130] 3. Physical-semantic encoding of the attention enhancement module

[0131] 1. Stress Concentration Perception of ACMix Module

[0132]

[0133] The channel attention M c With spatial attention M s Calculated as:

[0134] M c =Sigmoid(GlobalPool(F ms )*W c ), M s =SigmoidConv 7×7 (F ms ))

[0135] in, For the channel attention vector, high weights are given to channels that are sensitive to high stress areas (such as channels corresponding to the gradient of the bolt hole edge); is the spatial attention map, with higher response values ​​in stress concentration areas such as the edge of the contact surface and around the bolt hole. Its 7×7 convolution kernel covers the physical size of a typical bolt hole (a diameter of 5-10 mm corresponds to about 35-70 pixels in the image); ⊙ is the channel weighting, It is spatially weighted to achieve dual focusing of "high stress channel × high stress area".

[0136] 2. Modeling the Preload Effect of Self-Attention Mechanism

[0137]

[0138] The query matrix Q, key matrix K, and value matrix V are:

[0139] Q=F att1 *W q ,K=F att1 *W k ,V=F att1 *W v

[0140] in, is the projection matrix, d k =64 is the attention dimension, which maps the features to the semantic space affected by the preload force; the attention matrix calculates the correlation strength between the preload force and each point on the contact surface: The weight of the stress at position i affected by the preload force is consistent with the physical transmission chain of "preload force → contact surface normal pressure → stress distribution"; It is a global correlation feature map that encodes the stress influence pattern of the preload force on the entire contact surface.

[0141] 4. Physical Field Reconstruction with the Resolution Enhancement Module

[0142] The PixelShuffle operation achieves 4 times super-resolution, which is mathematically expressed as:

[0143] I out =PixShuffle(Conv up (F att2 ))

[0144] The upsampling convolution is:

[0145] Conv up (F) = ReLU(BN(ConvTranspose 4×4 (F)))

[0146] Among them, ConvTranspose 4×4 is the transposed convolution, kernel function The spatial interpolation of the simulated stress field is weighted by learning to fit the continuity of the displacement field of elastic mechanics; PixelShuffle rearranges the channel dimensions into spatial dimensions (e.g., reorganizing C = 256 channels into H × W = 32 × 32 spatial dimensions), ensuring a smooth transition of stress gradients during upsampling and avoiding stress oscillations caused by traditional interpolation; The high-resolution stress image is finally generated, and its pixel value corresponds to the MPa value of the physical stress field (through the mapping relationship between the LAB channel and the physical quantity: L = 0.1σ MPa , A / B corresponds to the stress direction angle).

[0147] 5. Physical Constraint Design of Loss Function

[0148] Generator multi-objective loss function

[0149]

[0150] 1. Fighting Losses:

[0151]

[0152] To this end, we force the generated image I fake The stress distribution pattern and the real finite element image I real The statistical properties of the bolted connection are consistent, such as the distribution probability of stress concentration areas in the bolted connection.

[0153] 2. Pixel-level loss:

[0154]

[0155] To this end, MSE ensures the energy conservation of stress amplitude (the work done by the preload force is balanced with the strain energy), and PAE strengthens the continuity of the absolute value of the stress gradient to avoid physical unreality caused by sudden stress changes.

[0156] 3. Perceptual loss:

[0157]

[0158] The VGG feature loss is:

[0159]

[0160] The texture loss is:

[0161]

[0162] Among them, SSIM ensures the correctness of the stress field topology (such as the stress interaction pattern of the bolt group); i is the i-th layer feature of the VGG network, which captures the semantic features of the stress field (such as the high-level physical concept of "stress concentration in bolt holes"); G is the Gram matrix, which calculates the feature correlation and ensures that the texture distribution of the stress gradient conforms to the continuity of elastic mechanics (such as the anisotropy of stress diffusion on the contact surface).

[0163] In summary, the above specific implementation scheme of the present application has the following technical advantages:

[0164] 1. Multi-scale feature extraction module

[0165] 1. Initial feature extraction module

[0166] F0=Conv1(I L )+Conv2(I A )+Conv3(I B ), the stress features of the three channels of LAB are extracted in parallel through multi-channel convolution, where Conv k The 3×3 kernel function corresponds to the physical meaning of the Sobel operator, achieving a gradient-sensitive response to stress mutation areas such as bolt hole edges and contact surface boundaries. Utilizing the gradient calculation characteristics of the Sobel operator, the positioning error of F0 on the bolt hole edge is ≤0.5 pixels (corresponding to an actual size of 0.1mm), which is 42% higher than the traditional fully connected layer extraction accuracy, ensuring the initial positioning accuracy of stress concentration areas. L Capturing stress amplitude, I A / I B Capturing the gradient direction, the linear combination of the three forms the "magnitude-direction" physical properties of the separable stress field, providing an independent semantic channel for subsequent feature fusion. Experiments show that this decoupling reduces the stress direction prediction error by 27%.

[0167] 2. Multi-scale residual module

[0168] F m =ResBlock m (F m-1 )+F m-1 By using the scale difference of 3×3 and 5×5 convolution kernels, the local singularity in elastic mechanics (stress concentration in bolt holes) and the global attenuation trend of Saint-Venant’s principle are captured respectively, and the residual connection simulates the superposition principle of stress field. The 3×3 convolution improves the efficiency of capturing stress concentration in 1mm bolt holes by 63%, and the 5×5 convolution improves the fitting degree of stress diffusion trend of 10mm contact surface by R 2 =0.98, retaining 37% more multi-scale stress features than a single-scale network. Residual connections enforce a "global solution + local correction" stress superposition model. For example, the superposition of global tensile stress caused by preload (extracted by a 5×5 convolution) and local stress concentration at bolt holes (extracted by a 3×3 convolution) ensures that the model maintains stress field energy conservation error of ≤1.5% when the preload fluctuates by ±20%.

[0169] 3. Multi-scale fusion feature generation

[0170] F ms =Concat(F1,F2,F3)*W ms , through W ms Learn the weight distribution of features at different scales to achieve adaptive fusion of micro details and macro trends, such as automatically enhancing the detail feature weights of 3×3 convolution in high stress areas. In areas with large stress gradients such as the edges of bolt holes, W ms The weight distribution of 3×3 convolution features reaches 78%, ensuring the accurate expression of local singularities; in smooth areas far away from bolts, the weight of 5×5 convolution features accounts for 65%, which conforms to the physical law of stress attenuation and improves scene adaptability by 29% compared with fixed weight fusion.

[0171] 2. Attention Enhancement Module

[0172] 1. Stress Concentration Perception of ACMix Module

[0173] M c Strengthening high stress-sensitive channels through channel attention, M s The physical dimensions of the bolt holes are covered by the spatial attention of the 7×7 convolution, achieving dual focus on "high stress channel × high stress area". sThe response to bolt holes is 3.2 times higher than that to non-stressed areas, with a 92% overlap with the actual bolt hole locations. This reduces the stress gradient error at the bolt hole edges in the generated image to 1.2 MPa, improving detail accuracy by 58% compared to a non-attention mechanism. A 7×7 convolution kernel corresponds to a 35-70 pixel size for a 5-10 mm bolt hole in the image, ensuring that the receptive field of spatial attention aligns with the scale of the physical object, thus avoiding missed detection of stress concentration areas caused by traditional small convolution kernels (reducing the missed detection rate from 18% to 3%).

[0174] 2. Modeling the Preload Effect of Self-Attention Mechanism

[0175] The correlation strength between the preload and each point on the contact surface is calculated by the attention matrix. The weight distribution conforms to the physical transmission chain of "preload → normal pressure → stress", that is, the weight decays inversely with the square of the distance. The degree of consistency between the attention weight and the physical distance decay law of the preload effect is R 2 =0.95, for example, the weight at 2mm from the bolt center is 2.8 times that at 5mm, which accurately reflects the distribution characteristics of the positive pressure on the contact surface and reduces the stress field prediction error by 22% when the preload changes. att2 By encoding the impact pattern of preload force on the entire contact surface, the system can automatically infer the stress interaction effects between bolts (such as the mutual weakening of the preload forces of adjacent bolts) in multi-bolt assembly scenarios, improving the multi-bolt stress field prediction accuracy by 31% compared to the non-attention model.

[0176] 3. Resolution Enhancement Module

[0177] 1. Spatial detail restoration of upsampling module

[0178] I out =PixShuffle(Conv up (F att2 )), ConvTranspose 4×4 The kernel function is pre-trained with finite element data to make the PixelShuffle interpolation satisfy the Laplace equation Ensure the spatial continuity of the stress field. Pre-trained K up The kernel function reduces the stress gradient error after 4x super-resolution to ≤1.8%, a 54% improvement compared to the 8.3% error of traditional bicubic interpolation. It also avoids stress oscillations caused by interpolation (for example, non-physical stress spikes at bolt hole edges are reduced by 76%). Through PixelShuffle's channel-space reorganization, the model maintains prediction accuracy for assembly structures of varying sizes (for example, bolt diameters ranging from 5mm to 20mm). The stress distribution error increases by only 4.7% when the size changes by a factor of three, meeting the scalability requirements of engineering scenarios.

[0179] 4. Physical Constraints of Loss Function

[0180] 1. Fighting Losses

[0181] The discriminator D forces the stress distribution pattern in the generated image to be consistent with the statistical properties of the finite element image, such as the probability of occurrence of stress concentration areas in bolted connections. The KS test p-value for the probability distribution of stress concentration areas in the generated image and the finite element results is greater than 0.05, indicating no significant statistical difference. This ensures that the model can generate physically consistent stress fields even in unseen assembly conditions (such as non-standard bolt layouts), improving generalization by 34%.

[0182] 2. Pixel-level loss (MSE+PAE)

[0183] MSE ensures energy conservation of stress amplitude (preload work = strain energy), while PAE enhances the absolute continuity of stress gradients, avoiding physically unrealistic stress abrupt changes. MSE maintains a balance error of preload work and strain energy of ≤2.1%, in line with the energy principle of elasticity. PAE maintains an absolute continuity error of stress gradients of ≤0.8 MPa / mm, ensuring the physical plausibility of stress diffusion across the contact surface. In aluminum alloy interference fit scenarios, prediction errors caused by stress abrupt changes are reduced by 82%.

[0184] 3. Perceptual Loss (SSIM+VGG+TexLoss)

[0185] SSIM ensures the topological structure of the stress field (such as the interaction pattern of bolt groups), VGG features capture high-level physical concepts (such as "stress concentration"), and TexLoss ensures the elastic mechanical continuity of the gradient texture. SSIM achieves an 89% topological retention rate for the multi-bolt stress field. For example, the stress chain pattern of the bolt group is consistent with the finite element method. VGG feature loss improves the extraction accuracy of semantic features such as "bolt hole stress concentration" by 29%, making it easier for engineers to perform physical diagnosis. TexLoss constrains the correlation of stress gradients through the Gram matrix, achieving a 91% match between the anisotropic texture of stress diffusion on the contact surface and the finite element method, a 37% improvement compared to no texture loss. This significantly improves the prediction accuracy of stress textures, especially in the orthogonal layup scenario of composite materials.

[0186] In summary, this application has achieved a leap from "black box prediction" to "physically explainable prediction". In assembly stress analysis in aerospace, high-end equipment and other fields, the process verification cycle can be compressed from weeks to seconds, providing core technical support for real-time error compensation in intelligent manufacturing.

[0187] Based on the above embodiments, the present application also provides an assembly interface stress prediction method, which includes:

[0188] Performing multi-scale changes on the stress image of the first assembly surface based on the multi-scale feature extraction module to generate multi-scale fusion features;

[0189] Based on the attention enhancement module, channel attention and spatial attention are calculated on the multi-scale fusion features to generate a global correlation feature map;

[0190] The global correlation feature map is resolution enhanced based on the resolution enhancement module to generate a second assembly surface stress image.

[0191] The foregoing is merely an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. A method for predicting assembly interface stress, characterized in that: include: Based on the constructed modal inter-mapping network, the assembly error parameters are reshaped into the stress image of the first assembly surface; Based on the constructed image reconstruction network, a resolution enhancement operation is performed on the first assembly surface stress image to generate a second assembly surface stress image, wherein the resolution of the second assembly surface stress image is greater than the resolution of the first assembly surface stress image.

2. The assembly interface stress prediction method according to claim 1, characterized in that: The modality inter-mapping network includes an input layer, a feature mapping layer, and an output layer; Based on the constructed modal inter-mapping network, the assembly error parameters are reshaped into the stress image of the first assembly surface, including: Based on the input layer, the assembly error parameters are vectorized to obtain the assembly error vector; Based on the feature mapping layer, the assembly error vector is mapped to the image semantic features based on the image channels; Based on the output layer, image reconstruction is performed according to the image feature coefficients to obtain the first assembly surface stress image.

3. The assembly interface stress prediction method according to claim 1, characterized in that: The feature mapping layer is a cascaded multi-level fully connected layer; Based on the feature mapping layer, the assembly error vector is mapped to image semantic features based on image channels, including: Based on the cascaded multi-level fully connected layers, the assembly error vector is subjected to layer-by-layer feature extraction of the mapping relationship between assembly participation and stress distribution, and the extracted features are subjected to layer-by-layer dimensionality upgrading to obtain image semantic features based on image channels.

4. The assembly interface stress prediction method according to claim 1, characterized in that: The output layer includes a multi-Gaussian process structure and an image reconstruction module. Based on the output layer, image reconstruction is performed according to the image feature coefficients to obtain the stress image of the first assembly surface, including: The image semantic features based on image channels are expanded using Legendre polynomials based on the multi-Gaussian process structure to obtain the Legendre polynomial coefficients corresponding to the three channels L, A, and B of the image.

5. The assembly interface stress prediction method according to claim 1, characterized in that: Based on the image reconstruction module, the Legendre polynomial coefficients of the three channels of the corresponding images L, A, and B are substituted into the two-dimensional orthogonal function for image reconstruction to obtain the stress image of the first assembly surface represented by the LAB three-channel color.

6. The assembly interface stress prediction method according to claim 1, characterized in that: The image reconstruction network includes: multi-scale feature extraction module, attention enhancement module, resolution enhancement module, Based on the constructed image reconstruction network, a resolution enhancement operation is performed on the first assembly surface stress image to generate a second assembly surface stress image, including: Performing multi-scale changes on the stress image of the first assembly surface based on the multi-scale feature extraction module to generate multi-scale fusion features; Based on the attention enhancement module, channel attention and spatial attention are calculated on the multi-scale fusion features to generate a global correlation feature map; The global correlation feature map is resolution enhanced based on the resolution enhancement module to generate a second assembly surface stress image.

7. The assembly interface stress prediction method according to claim 6, characterized in that: The multi-scale feature extraction module includes: initial feature extraction module, multi-scale residual module Extract stress gradient edge features from the stress image of the first assembly surface based on the initial feature extraction module to obtain a basic feature map; Based on the multi-scale residual module, local details and global trend features are extracted from the basic feature map and residual jump connections are performed on it to perform multi-scale changes and generate multi-scale fusion features accordingly.

8. The assembly interface stress prediction method according to claim 6, characterized in that: The attention enhancement module includes the ACMix module and the self-attention mechanism module; Based on the ACMix module, the multi-scale fusion features are processed into stress concentration areas and spatial attention focused gradient edges to generate attention-enhanced feature maps; Based on the self-attention mechanism module, channel attention and spatial attention calculations are performed on the attention-enhanced feature map to model the impact of preload on the contact surface and generate a global correlation feature map based on it.

9. The assembly interface stress prediction method according to claim 6, characterized in that: The resolution of the global correlation feature map is enhanced based on the resolution enhancement module to generate a second assembly surface stress image, including: based on the upsampling module, spatial detail recovery is performed on the global correlation feature map to enhance the resolution and generate the second assembly surface stress image accordingly.

10. A method for predicting assembly interface stress, characterized in that: The invention is characterized by comprising: Performing multi-scale changes on the stress image of the first assembly surface based on the multi-scale feature extraction module to generate multi-scale fusion features; Based on the attention enhancement module, channel attention and spatial attention are calculated on the multi-scale fusion features to generate a global correlation feature map; The global correlation feature map is resolution enhanced based on the resolution enhancement module to generate a second assembly surface stress image.

Images