Aircraft composite material structure assembly stress prediction method based on deep learning

Abstract

The invention relates to an aircraft composite material structure assembly stress prediction method based on deep learning, is used for composite material structure assembly stress prediction analysis, and belongs to the assembly field of aeronautical manufacturing. And a new thought and method are provided for predicting the assembly stress of the composite material structure. According to the method, the improved conditional generative adversarial network is introduced, so that the assembly stress field of the composite material structure can be autonomously learned and generated under a limited data condition. According to the method, the internal stress state can be deduced according to the surface strain data measured by the DIC, and a new technical means is provided for assembly quality control and assembly stress evaluation of a composite material structure.

Description

Technical Field

[0001] This invention relates to a deep learning-based method for predicting assembly stress in aircraft composite material structures, which is used for stress prediction and analysis in composite material structure assembly and belongs to the assembly field of aerospace manufacturing. Background Technology

[0002] Composite materials, with their excellent mechanical properties, including high specific strength, high specific stiffness, and good fatigue resistance, have been widely used in aircraft structures. Compared with traditional metallic materials, composite materials also have lower density, which helps to reduce aircraft weight, improve fuel efficiency, and enhance structural durability. Bolted connections have strong load-bearing capacity and, compared with riveting, do not cause impact damage to composite material structures; compared with adhesive bonding, bolted connections are more reliable, therefore, bolting is one of the commonly used connection methods for composite material structures.

[0003] In the aerospace manufacturing field, metal structural components are primarily manufactured using CNC machine tools to ensure high geometric accuracy. Composite material structural components commonly undergo autoclaving. Due to the complex physical and chemical reactions that occur during the molding process, the final shape often deviates from the theoretical shape, resulting in gaps between the composite and metal connections. Furthermore, misalignment in assembly positions can also occur during the assembly of metal structures, leading to gaps between the metal and composite materials. If bolts are used for forced assembly in these situations, the composite components will experience significant bending deformation. This not only affects the aerodynamic shape of the aircraft structure but may also cause localized stress concentration around the bolt holes, resulting in damage to the composite material. Reasonable assembly stress is a prerequisite for ensuring the reliability and load transfer efficiency of the aircraft structure. However, improper assembly stress will inevitably have adverse effects on the structural performance and lifespan after assembly, such as plastic deformation or even damage, and premature stress corrosion failure. Current composite panel assembly primarily uses geometric accuracy as a quality evaluation indicator, lacking assessment of assembly stress. Therefore, it is necessary to evaluate and measure the assembly stress during the assembly process of composite panel components. The assembly stress refers to the stress generated in the composite material under the combined action of bolt preload, initial deformation of the composite material, and the gap between the two assembly structures.

[0004] For stress testing of composite materials, both destructive and non-destructive stress testing can be used. Destructive stress testing can damage components, affecting their overall strength, and therefore cannot be applied in actual assembly. Non-destructive stress testing, based on different physical principles, can be summarized as: X-ray method, spectroscopic method, and ultrasonic method. Among these, X-ray and spectroscopic methods require expensive equipment and have limited experimental resources, preventing large-scale application in actual assembly. Ultrasonic methods use large probes, cannot measure stress in small areas around bolts, and the results only provide the average gravitational force in the material thickness direction (0-2 mm), failing to accurately measure the stress in each layer. Therefore, directly using stress testing equipment cannot accurately measure the stress field of composite materials. However, for strain and deformation testing of composite materials, accurate detection can be achieved using a three-dimensional image correlation (3DDIC) system or a three-dimensional laser scanner. Therefore, how to assess the assembly stress level of composite panel panels and how to correlate measurable strain with unmeasurable stress has always been a research challenge.

[0005] With the development of artificial intelligence (AI) technology, machine vision algorithms have achieved fruitful results in predicting the mechanical properties of composite materials, structural optimization and design, and damage prediction. This provides new ideas and methods for solving the problem of predicting assembly stress in composite material structures. Among them, generative AI, as an AI technology capable of autonomously learning and generating new data, has powerful data simulation and creation capabilities, making it particularly suitable for scenarios with insufficient or difficult-to-obtain data, such as the assembly stress of composite material structures. In generative AI, the most commonly used model is Generative Adversarial Networks (GANs) based on Nash equilibrium theory. GANs continuously optimize the quality of data generated by the generator through a competitive mechanism between the generator and the discriminator, until the discriminator can no longer distinguish between generated and real data, thereby achieving highly realistic data generation and providing a new method for predicting assembly stress in composite material structures. Summary of the Invention

[0006] The technical problem solved by this invention is: how to provide a method for predicting the assembly stress of composite material structures.

[0007] According to one aspect of this application, a deep learning-based method for predicting assembly stress in aircraft composite material structures is provided.

[0008] S1: Data preparation and processing. A model of the aircraft composite material structure connection simulation component was built using Abaqus software. Different boundary conditions were set according to the possible situations in reality. The strain and stress of the aircraft composite material structure connection simulation component under different boundary conditions were calculated to obtain the finite element analysis results.

[0009] S2: Based on the finite element analysis results obtained in S1, the surface strain field and internal stress field around the holes of the simulated aircraft composite material structure connection are extracted. The surface strain field and internal stress field are processed using a unified color code to obtain a dataset. After enhancement, the dataset is divided into a training set and a test set.

[0010] S3: Design a conditional generative adversarial network model, train the network using the training set, and test the accuracy of the conditional generative adversarial network model using the test set;

[0011] The training set and test set are the augmented datasets in S2;

[0012] S4: The surface strain field around the hole of the aircraft composite material structure connection simulation part, which is measured by digital image correlation technology, is used as the input of the conditional generative adversarial network model. The conditional generative adversarial network model outputs the corresponding internal stress field around the hole of the aircraft composite material structure connection simulation part.

[0013] S11: The model is built using hexahedral linear reduced integral elements. In the model of the aircraft composite material structure connection simulation component, each layup corresponds to one layer of elements.

[0014] S12: In the model of the aircraft composite material structure connection simulation component, the periphery of the hole in the aircraft composite material structure connection simulation component is locally meshed to increase the accuracy of the stress field data inside the hole.

[0015] S21: Processing surface strain and internal stress fields using a unified color code specifically includes: finding the maximum and minimum values ​​of the surface strain and internal stress fields respectively, and setting the upper and lower limits of all surface strain and internal stress fields to these maximum and minimum values, so that the color corresponds one-to-one with the strain or stress.

[0016] S22: Manually augment the dataset by cropping out the parts of the images containing holes to create the augmented dataset. Specifically, this includes mapping the images in the dataset to a size of 1024x1024 and cropping the images using a cropping box with a size of 0.5 to 0.8 times the image size.

[0017] S31: The conditional generative adversarial network model includes a generator and a discriminator;

[0018] The generator receives the surface strain field and outputs the internal stress field. Its function is to learn the distribution of real data samples and generate new samples that are as close as possible to the real data.

[0019] The input to the discriminator is the true stress field and the false stress field merged along the channel dimension. After passing through the downsampling module, it gives a score relative to the real image. The true stress field and the false stress field refer to the internal stress field in the dataset and the internal stress field generated by the generator, respectively.

[0020] S32: The generator consists of two parts: a U-Net network and a super-resolution network. The input surface strain field passes through 5 convolutional modules in the U-Net network and downsampling modules between each convolutional module, and then enters the residual module. After passing through the residual module, the image size is restored to its original size by the upsampling module and the convolutional module.

[0021] The convolution module consists of two layers of two-dimensional convolution, instance normalization layer, and LeakyReLU activation function stacked twice, and Dropout is added afterward to prevent overfitting.

[0022] The residual module consists of two superimposed layers of two-dimensional convolution, instance normalization layer and ReLU activation function, and is connected by skip connections.

[0023] The downsampling module consists of a two-dimensional convolution, an instance normalization layer, and a LeakyReLU activation function, to achieve double downsampling.

[0024] The upsampling module first performs double upsampling using nearest neighbor interpolation, followed by two-dimensional convolution, instance normalization layer, and LeakyReLU activation function;

[0025] The image passes through the U-Net network layer and then enters the super-resolution module;

[0026] The super-resolution module consists of two smaller super-resolution modules. Each smaller super-resolution module consists of two residual modules and one upsampling module, and its function is to increase the resolution of the image by four times, thereby making the generated image clearer.

[0027] S33: The input to the discriminator is the true stress field and the false stress field merged along the channel dimension. It determines whether the input sample comes from real data or generated data. The true stress field and the false stress field refer to the internal stress field in the dataset and the internal stress field generated by the generator, respectively. Then, it goes through four downsampling modules. Each downsampling module consists of a two-dimensional convolution, a batch normalization layer, and a LeakyReLU activation function. Then, it goes through another convolution to reduce the dimensionality. Finally, it goes through a global average pooling layer to give its score relative to the real image and map it to the range of 0 to 1.

[0028] The generator's loss function is a weighted sum of GAN loss and L1 loss, with L1 loss typically having a higher weight to ensure that the generated image is similar to the real image at the pixel level. GAN loss calculates the mean square error between the discriminator's prediction of the generated image pair and 1, while L1 loss calculates the L1 distance between the generated image and the real image.

[0029] The total loss of the discriminator is a weighted sum of the true loss and the false loss. Typically, the false loss has a higher weight to encourage the discriminator to better identify the generated images. The true loss is calculated as the mean square error between the discriminator's prediction of the real image pair and 1, while the false loss is calculated as the mean square error between the discriminator's prediction of the generated image pair and 0.

[0030] S34: Input the surface strain field in the test set into the pre-trained generator model, and the generator outputs the corresponding internal stress field;

[0031] S35: Calculate the structural similarity index and peak signal-to-noise ratio between the internal stress field generated by the generator and the internal stress field in the test set to evaluate the quality of the generated image.

[0032] The software environment set up is a learning framework based on PyTorch under the Win11 system;

[0033] The hardware environment used was an NVIDIA 4070 Ti SUPER graphics card.

[0034] S41: Use a fixture to position and clamp the aircraft composite structure connection simulator. After the bolts are tightened, use the DIC system to measure the surface strain field around the holes of the aircraft composite structure connection simulator.

[0035] The beneficial effects of this invention are: it provides a new approach and method for predicting assembly stress in composite material structures. By introducing an improved conditional generative adversarial network (GAN), this invention can autonomously learn and generate the assembly stress field of composite material structures under limited data conditions. This method can infer the internal stress state based on surface strain data measured by DIC, providing a new technical means for assembly quality control and assembly stress assessment of composite material structures. Attached Figure Description

[0036] Figure 1 This is the overall flowchart for assembly stress prediction;

[0037] Figure 2 This is an overall architecture diagram of the improved conditional generative adversarial network;

[0038] Figure 3 This is the network structure diagram of the generator;

[0039] Figure 4 This is the network structure diagram of the discriminator. Detailed Implementation

[0040] The present application is described in detail below with reference to the embodiments, but the present application is not limited to these embodiments.

[0041] Example 1

[0042] In this embodiment, a carbon fiber epoxy resin laminate is used as an example to predict the assembly stress of the simulated connection between the laminate and the aluminum alloy plate. The composite material selected is T300 carbon fiber epoxy resin composite material, with a single layer thickness of 0.2 mm, 18 symmetrically laid layers, and a layup sequence of [45° / 90° / -45° / 0° / 90° / 0° / -45° / 90° / 45°]s. The remaining dimensions of the laminate are designed according to the specifications for tensile specimen dimensions in ASTM D5961 and the specifications for countersunk bolt hole dimensions in the national standard GB / T 152.2-2014. The aluminum alloy plate material is 7075-T651; countersunk bolts, model HB1-132-2002, are used; and hexagonal self-locking nuts, model GB1337-1988, are selected. The surface strain field of the laminate after the bolts are tightened is measured using a 3D-DIC system.

[0043] The method for predicting assembly stress in composite material structures involved in this embodiment is as follows: Figure 1 As shown, perform the following steps:

[0044] Step 1: Use Abaqus software to establish a finite element model of the composite material structure connection simulation component. Set different boundary conditions according to the possible situations in reality, and calculate the strain and stress of the structure under different boundary conditions.

[0045] Since the self-locking nut is tightened by tightening the fixed bolt, the laminate only bears the pressure of the bolt during tightening, and is not subjected to the frictional force generated by the bolt rotation. Therefore, in finite element modeling, we can treat the bolt and nut as a whole and use the bolt-load method in Abaqus to simulate the bolt preload, that is, to apply the preload by shortening the bolt shank. This simulation not only conforms to the actual assembly situation, but also simplifies the model and improves computational efficiency.

[0046] To save computation time while also considering the calculation of contact problems, this study ultimately adopted hexahedral linear reduced integral elements. Each layup corresponds to one element layer in the modeling of the composite component. Local mesh refinement was implemented in the peripore region of the composite material to increase the accuracy of the peripore stress results. Light springs were used to connect the components to bolts to prevent rigid body displacement during the analysis.

[0047] In the actual assembly of composite structural components, the composite structural components with initial deformation are often adjusted to the theoretical shape before subsequent clamping and connection. Therefore, in the laminate structure, the following boundary conditions are also used as variables: initial deviation, assembly gap, and bolt preload, and their specific parameter settings are shown in Table 1. The bolt preload is adjusted using the bolt-load method; the gap size is adjusted by setting the distance between the laminate and the aluminum alloy plate; to simulate the initial deformation of the wall panel, this embodiment adopts the method of applying various forms and sizes of deformation to the sample, and using the deformed sample as the connected part in the assembly. The deformation methods used in this embodiment are: applying bending deformation and torsional deformation to one end of the laminate structure. In the simulation, the composite sample with initial deformation is used as the connected part. During connection, the deformed sample is first adjusted to the theoretical shape before the bolts are tightened.

[0048] Finally, the Hashin failure criterion was added to the Abaqus simulation by calling a subroutine to determine whether the composite laminate was damaged after the bolts were tightened.

[0049] Table 1. Types of boundary condition variables and their parameter settings

[0050]

[0051] Step 2: From the finite element analysis results obtained above, extract the surface strain field and internal stress field around the holes in the composite material structure, and use a unified color coding to unify them, thereby constructing a dataset.

[0052] The purpose of this embodiment in predicting assembly stress is to assess the load-bearing capacity of the laminate, i.e., its proximity to damage, when the laminate is undamaged. For holes that have already been damaged, ultrasonic equipment can be used to detect the damage, and assembly stress prediction is no longer required. Therefore, when creating the dataset, the surface strain and internal stress of undamaged holes are extracted as the dataset based on the calculation results of the Hashin failure criterion. To illustrate the feasibility of the method, this embodiment selects the surface strain field in direction 1 around the hole and the stress field in direction 1 of the 6th layer of the material as the dataset when creating the dataset.

[0053] It is worth noting that when constructing the dataset, a unified color encoding method needs to be used to process the strain and stress fields. That is, the maximum and minimum values ​​of the surface strain field and the internal stress field are found respectively, and then the upper and lower limits of all strain and stress fields are set to these maximum and minimum values. In this way, each color in the image corresponds to a stress or strain value, that is, the color of the image is given meaning, and this meaning is the stress or strain value.

[0054] To ensure the model fully learns the relationship between surface strain and internal stress, manual data augmentation was performed on the images in the dataset: First, the cropped images were mapped to a 1024x1024 size, and a square cropping box of random size and position was created to crop the original image. To enable the model to learn the stress and strain distribution around the hole, the cropped images must include the hole, requiring the cropping box size to be substantial. In this embodiment, the cropping box size is 0.5 times larger and less than 0.8 times the original image size. It is worth noting that for a pair of images, cropping boxes of the same position and size should be used to ensure the model learns the stress-strain relationship at the corresponding locations.

[0055] In this embodiment, a total of 405 pairs of images of the strain and stress fields around the holes were captured. 396 pairs of these images were used as the original data for the training set, and 9 pairs were used as the test set. The 396 pairs of images were then randomly cropped 10 times in sequence, resulting in 3960 pairs of images used as the training set.

[0056] Step 3: Design a conditional generative adversarial network that includes U-Net and super-resolution structure, train the network using the training set, and test the accuracy of the model using the test set.

[0057] The improved generative adversarial network consists of two parts: a generator and a discriminator. The overall structure of the network is as follows: Figure 2 As shown. The generator receives the surface strain field and outputs the internal stress field. Its function is to learn the distribution of real data samples and generate new samples that are as close to the real data as possible. The discriminator is input to the true stress field and the false stress field merged along the channel dimension. It determines whether the input sample comes from real data or generated data. The true stress field and the false stress field refer to the stress field obtained from the Abaqus simulation and the stress field generated by the network's generator, respectively.

[0058] The network structure of the generator is as follows Figure 3As shown, the network consists of two parts: a U-Net network and a super-resolution network. Since the main goal of the network model is to establish the relationship between corresponding positions in two images to achieve accurate transfer from strain field to stress field, a semantic segmentation model, U-Net, is incorporated into the generator's network model to enhance its performance. To further optimize the clarity of the output stress field and ensure the sharpness of the boundaries, a super-resolution network is added after the U-Net network. The input image passes through five convolutional layers in the U-Net network and downsampling modules between each convolutional layer before entering the residual module. After passing through the residual module, the image size is restored to its original size through an upsampling module and a convolutional module. The convolutional module consists of two layers of 2D convolution, instance normalization layers, and LeakyReLU activation functions stacked twice, with Dropout added afterward to prevent overfitting. The residual module consists of two layers of 2D convolution, instance normalization layers, and ReLU activation functions stacked twice, connected by skip connections. The downsampling module consists of 2D convolution, instance normalization layers, and LeakyReLU activation functions, achieving a 2x downsampling. The upsampling module first performs double upsampling using nearest neighbor interpolation, followed by two-dimensional convolution, instance normalization layer, and LeakyReLU activation function.

[0059] During each upsampling, the feature map of the corresponding downsampling module is used as part of the input in order to enable the network to learn the image information lost during the downsampling layer.

[0060] After passing through the U-Net network layers, the image enters the super-resolution module. The super-resolution module consists of two smaller super-resolution modules, each composed of two residual modules and one upsampling module. Its function is to increase the image resolution by four times, thereby making the generated image clearer.

[0061] The structure of the discriminator is as follows Figure 4 As shown, the discriminator's input consists of the true and false stress fields merged along the channel dimension. This is then processed through four downsampling modules, each composed of a 2D convolution, a batch normalization layer, and a LeakyReLU activation function. After another convolution for dimensionality reduction, a global average pooling layer is applied to provide a score relative to the real image, mapping it to a range of 0 to 1.

[0062] The generator's loss function is a weighted sum of the GAN loss and the L1 loss, with the L1 loss typically having a higher weight to ensure that the generated image is pixel-level similar to the real image. The GAN loss calculates the mean squared error (MSE) between the discriminator's prediction of the generated image pair and 1; the L1 loss calculates the L1 distance between the generated image and the real image. The discriminator's total loss is a weighted sum of the true loss and the fake loss, with the fake loss typically having a higher weight to encourage the discriminator to better recognize the generated image. The true loss calculates the mean squared error (MSE) between the discriminator's prediction of the real image pair and 1; the fake loss calculates the mean squared error (MSE) between the discriminator's prediction of the generated image pair and 0.

[0063] The software environment set up in this embodiment is a learning framework based on PyTorch under the Win11 system, and the hardware environment is NVIDIA. 4070Ti SUPER graphics card.

[0064] This embodiment uses the Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR) to evaluate the quality of the generated image. SSIM is a metric that measures the visual similarity between two images. It compares contrast, brightness, and structure, effectively reflecting whether the structural information of an image is preserved. It is suitable for evaluating the visual similarity between the generated image and the real image. PSNR measures image quality by calculating the logarithm of the mean square error between the generated and real images relative to the maximum possible pixel value of the image. It focuses on pixel-level accuracy and directly reflects the sharpness and detail retention of the generated image. Since SSIM and PSNR evaluate image quality from two different perspectives—visual perception and pixel level, respectively—using both metrics comprehensively assesses the performance of generative adversarial networks in generating images, ensuring that the generated images are both visually close to the real images and have high fidelity at the pixel level.

[0065] The surface strain field from the test set is input into the trained generator model, and the generator outputs the corresponding stress field. The SSIM and PSNR of the generated stress field and the stress field from the test set are calculated, yielding an average SSIM of 0.93 and a PSNR of 31.47. The SSIM is very close to 1, indicating structural similarity between the image and the original image; the PSNR value is above 30, indicating relatively high image quality, with some distortion, but acceptable.

[0066] Step 4: Use the strain field on the surface around the hole obtained by DIC measurement as the input to the model, and the model outputs the internal stress diagram.

[0067] In the experiment, the laminate was positioned and clamped using a fixture. After the bolts were tightened, the perforated strain field of the laminate was measured using a DIC system.

[0068] After obtaining the surface strain field around the holes in the laminate using DIC measurement, its maximum and minimum values ​​are first set to the maximum and minimum values ​​set in step two. Then, the strain field around the holes is truncated. It is worth noting that the surface strain field of the composite material in the training set only includes the area around the holes, i.e., the bolt hole area is empty. However, truncating the DIC strain field will also crop the surface of the countersunk bolts, which will interfere with the prediction model. Therefore, before inputting the surface strain field into the model, the bolt surface area in the strain field must be covered with another color, so that the strain field input into the model only includes the strain field around the holes. In this embodiment, indigo blue is used to cover the bolts because the bolt hole area in the training set is indigo blue.

[0069] The processed strain field image of the hole periphery surface is input into the trained generator model, and the model will generate a corresponding stress map, which is the stress map of the 6th layer of the composite material in the 1 direction.

[0070] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions made by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting assembly stress in aircraft composite material structures based on deep learning, characterized in that, S1: Data preparation and processing: Establish a model of the aircraft composite material structure connection simulation component, set different boundary conditions, calculate the strain and stress of the aircraft composite material structure connection simulation component under different boundary conditions, and obtain the finite element analysis results. S2: Based on the finite element analysis results obtained in S1, the surface strain field and internal stress field around the holes of the simulated aircraft composite material structure connection are extracted. The surface strain field and internal stress field are processed using a unified color code to obtain a dataset. After enhancement, the dataset is divided into a training set and a test set. S3: Design a conditional generative adversarial network model, train the network using the training set, and test the accuracy of the conditional generative adversarial network model using the test set; The training set and test set are the augmented datasets in S2; S4: The surface strain field around the hole of the aircraft composite material structure connection simulation part, which is measured by digital image correlation technology, is used as the input of the conditional generative adversarial network model. The conditional generative adversarial network model outputs the corresponding internal stress field around the hole of the aircraft composite material structure connection simulation part.

2. The deep learning-based method for predicting assembly stress in aircraft composite material structures according to claim 1, characterized in that, S11: The model is built using hexahedral linear reduced integral elements. In the model of the aircraft composite material structure connection simulation component, each layup corresponds to one layer of elements. S12: In the model of the aircraft composite material structure connection simulation component, the periphery of the hole in the aircraft composite material structure connection simulation component is locally meshed to increase the accuracy of the stress field data inside the hole.

3. The deep learning-based method for predicting assembly stress in aircraft composite material structures according to claim 1, characterized in that, S21: Processing surface strain and internal stress fields using a unified color code specifically includes: finding the maximum and minimum values ​​of the surface strain and internal stress fields respectively, and setting the upper and lower limits of all surface strain and internal stress fields to these maximum and minimum values, so that the color corresponds one-to-one with the strain or stress. S22: Manually augment the dataset by cropping out the parts of the images containing holes to create the augmented dataset. Specifically, this includes mapping the images in the dataset to a size of 1024x1024 and cropping the images using a cropping box with a size of 0.5 to 0.8 times the image size.

4. The deep learning-based method for predicting assembly stress in aircraft composite material structures according to claim 1, characterized in that, S31: The conditional generative adversarial network model includes a generator and a discriminator; The generator receives the surface strain field and outputs the internal stress field. Its function is to learn the distribution of real data samples and generate new samples that are as close as possible to the real data. The input to the discriminator is the true stress field and the false stress field merged along the channel dimension. After passing through the downsampling module, it gives a score relative to the real image. The true stress field and the false stress field refer to the internal stress field in the dataset and the internal stress field generated by the generator, respectively.

5. The deep learning-based method for predicting assembly stress in aircraft composite material structures according to claim 4, characterized in that, S32: The generator consists of two parts: a U-Net network and a super-resolution network. The input surface strain field passes through 5 convolutional modules in the U-Net network and downsampling modules between each convolutional module, and then enters the residual module. After passing through the residual module, the image size is restored to its original size by the upsampling module and the convolutional module. The convolution module consists of two layers of two-dimensional convolution, instance normalization layer, and LeakyReLU activation function stacked twice, and Dropout is added afterward to prevent overfitting. The residual module consists of two superimposed layers of two-dimensional convolution, instance normalization layer and ReLU activation function, and is connected by skip connections. The downsampling module consists of a two-dimensional convolution, an instance normalization layer, and a LeakyReLU activation function, to achieve double downsampling. The upsampling module first performs double upsampling using nearest neighbor interpolation, followed by two-dimensional convolution, instance normalization layer, and LeakyReLU activation function; The image passes through the U-Net network layer and then enters the super-resolution module; The super-resolution module consists of two smaller super-resolution modules. Each smaller super-resolution module consists of two residual modules and one upsampling module, and its function is to increase the resolution of the image by four times, thereby making the generated image clearer.

6. The deep learning-based method for predicting assembly stress in aircraft composite material structures according to claim 4, characterized in that, S33: The input to the discriminator is the true stress field and the false stress field merged along the channel dimension. It determines whether the input sample comes from real data or generated data. The true stress field and the false stress field refer to the internal stress field in the dataset and the internal stress field generated by the generator, respectively. Then, it goes through four downsampling modules. Each downsampling module consists of a two-dimensional convolution, a batch normalization layer, and a LeakyReLU activation function. Then, it goes through another convolution to reduce the dimensionality. Finally, it goes through a global average pooling layer to give its score relative to the real image and map it to the range of 0 to 1.

7. The deep learning-based method for predicting assembly stress in aircraft composite material structures according to claim 4, characterized in that, The generator's loss function is a weighted sum of GAN loss and L1 loss, with L1 loss typically having a higher weight to ensure that the generated image is similar to the real image at the pixel level. GAN loss calculates the mean square error between the discriminator's prediction of the generated image pair and 1, while L1 loss calculates the L1 distance between the generated image and the real image. The total loss of the discriminator is a weighted sum of the true loss and the false loss. Typically, the false loss has a higher weight to encourage the discriminator to better identify the generated images. The true loss is calculated as the mean square error between the discriminator's prediction of the real image pair and 1, while the false loss is calculated as the mean square error between the discriminator's prediction of the generated image pair and 0.

8. The deep learning-based method for predicting assembly stress in aircraft composite material structures according to claim 4, characterized in that, S34: Input the surface strain field in the test set into the pre-trained generator model, and the generator outputs the corresponding internal stress field; S35: Calculate the structural similarity index and peak signal-to-noise ratio between the internal stress field generated by the generator and the internal stress field in the test set to evaluate the quality of the generated image.

9. The deep learning-based method for predicting assembly stress in aircraft composite material structures according to claim 4, characterized in that, The software environment set up is a learning framework based on PyTorch under the Win11 system; The hardware environment used was an NVIDIA 4070 Ti SUPER graphics card.

10. The deep learning-based method for predicting assembly stress in aircraft composite material structures according to claim 1, characterized in that, S41: Use a fixture to position and clamp the aircraft composite structure connection simulator. After the bolts are tightened, use the DIC system to measure the surface strain field around the holes of the aircraft composite structure connection simulator.

Images