Method for characterizing positioning rigidity of airplane thin-wall component based on generative model

Abstract

The invention provides an aircraft thin-wall component positioning rigidity characterization method based on a generative model, and relates to the field of aircraft equipment intelligent manufacturing, and the method comprises the steps: generating a process characteristic graph corresponding to a rigidity field through constructing a three-dimensional digital model and actual process parameters; acquiring a data set for model training; constructing a generative model based on an image translation mechanism and a conditional generative adversarial network; fine training is carried out on the model by inputting brand new data of'unseen 'of the network; the complete refined training model can adapt to various components, so that the time cost of rigidity field simulation is effectively reduced. According to the method, an effective solution is provided based on the process design of the stiffness field, digital twinning, augmented reality visualization and the like.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing of aircraft equipment, and in particular to a method for characterizing the positioning stiffness of aircraft thin-walled components based on a generative model. Background Art

[0002] The new generation of aircraft exhibits cross-generational development characteristics such as long life, long range, strong stealth, and high maneuverability. The requirements for key indicators such as assembly dimensional accuracy and mechanical properties are increasing exponentially. At this time, aircraft structural parts tend to be integrated, large-scale, and complex, and a large number of complex thin-walled components have emerged. Complex thin-walled components are characterized by large size and weak rigidity. They are very prone to deformation during the positioning process, resulting in the continuous accumulation of deviations during the assembly process, seriously affecting the overall assembly quality of the aircraft. Currently, aviation manufacturing companies mainly rely on manual experience and pre-finite element simulation methods to evaluate and control the possible deformation of thin-walled components during the positioning process. However, the model that relies too much on experience lacks theoretical guidance, and pre-finite element simulation is very time-consuming.

[0003] To address this issue, this paper proposes a generative model based on a conditional generative adversarial network and an image translation mechanism to correlate thin-walled component positioning features and positioning stiffness. By constructing a dataset using pre-processed finite element batch processing, the generative model is trained to rapidly characterize component stiffness. Summary of the Invention

[0004] To address the above-mentioned problems in the prior art, this application proposes a generative model-based method for characterizing the positioning stiffness of aircraft thin-walled components, comprising the following steps: S1, by constructing a three-dimensional digital model and actual process parameters, a process characteristic map corresponding to the stiffness field of the aircraft thin-walled component is generated, thereby obtaining a data set for model training; S2, constructs a generative model for aircraft thin-walled components based on image translation mechanism and conditional generative adversarial network; S3, using the data set generated in step S1 to train the generative model of step S2; after the generative model is trained, it is used as a pre-trained model, and the pre-trained model is refined by inputting new data.

[0005] In one embodiment, stiffness field data of thin-walled aircraft components are obtained by batch processing an input text file and a finite element software solver.

[0006] In one embodiment, obtaining the dataset includes the following steps: S11, establishing a finite element simulation model of an aircraft thin-walled component in a positioned state, applying a force along the assembly load direction at the assembly connection point, and then saving all positioned nodes as a node set; S12, writing the node set into a database file and writing the finite element model into a text file; S13, reading the database file obtained in step S12, randomly combining several groups of positioning solutions with different numbers of positioning nodes and different combination forms, and generating random positioning information; S14, reading the text file obtained in step S12, searching for the location where the positioning information is stored, and replacing the initial positioning information with the random positioning information generated in step S13, thereby generating several text files; S15, batch-calculating the text files generated in step S14 through the secondary development interface to obtain simulation result files corresponding to each text file; S16, opening the simulation result file obtained in step S15 through the secondary development interface, intercepting the deformation field of the aircraft thin-walled component in the minimum stiffness direction, which is the positioning stiffness field of the aircraft thin-walled component; S17, read the different positioning schemes generated by S13, combine the positioning node position in each positioning scheme with the geometric shape of the aircraft thin-walled component, the position of the reinforcement rib and the position of the assembly load application, and draw a process feature diagram.

[0007] In one embodiment, the generative model is a deep learning model, including a global generator, a local generator, a multi-scale discriminator, a residual block, and an improved convolutional block attention mechanism module.

[0008] In one embodiment, the creation of a deep learning model includes the following steps: S21, based on U-Net and ResNet structures, combines the convolutional block attention mechanism to build a global generator; S22, based on the U-Ne structure, build a local generator; S23, construct a multi-scale discriminator to distinguish whether the generated images are “real” or “fake” at different scales; S24, introduces multiple losses to comprehensively evaluate the similarity between the generated stiffness field and the actual stiffness field.

[0009] In one embodiment, the global generator construction process includes the following steps: S211, downsampling the input process feature map, which includes multiple cycles of "convolution → instance normalization → activation"; S212: After the input process feature map is downsampled, it is further extracted through multiple consecutive residual blocks; each residual block includes a cycle of "convolution → instance normalization → activation → convolution → instance normalization → improved convolution block attention mechanism"; S213, upsampling the process feature map after feature extraction, and the upsampling includes multiple cycles of "deconvolution → instance normalization → activation".

[0010] In one embodiment, the local generator construction process includes the following steps: S221, downsampling the output image in the global generator; downsampling includes multiple cycles of "convolution → instance normalization → activation → convolution → instance normalization → activation → average pooling → instance normalization → activation"; S222, performing multiple cycles of “convolution → instance normalization → activation → convolution → instance normalization → activation” on the downsampled feature map; S223, upsampling is performed to restore the image to the same size as the input image; upsampling includes multiple cycles of "deconvolution → instance normalization → activation → convolution → instance normalization → activation → convolution → instance normalization → activation"; S224, perform the final convolution, and the activation operation is completed using the hyperbolic tangent function.

[0011] In one embodiment, the multi-scale discriminator construction process includes the following steps: S231, processing the image at different scales, including the original scale and the scale after one downsampling operation, where the downsampling operation reduces the image scale to 1 / 4 of the original scale; S232, downsampling is performed at each scale to obtain feature maps; downsampling includes multiple cycles of "convolution → instance normalization → activation"; S233, after completing the downsampling of the two scales, the discriminator uses the sigmoid function to perform discrimination.

[0012] In one embodiment, in step S232, the loop operation is: S2221, in the first three convolution operations, the convolution kernel size is 3×3 and the stride is 2; S2222, in the last convolution operation, the convolution kernel size is 3×3 and the stride is 1; At S2223, the activation operation is performed using a linear rectifier unit with a leaky activation function.

[0013] In one embodiment, step S3 includes the following steps: S31, trains the generative model built in S2; S32, setting multiple evaluation indicators, including: mean square error, mean absolute error, determination coefficient, geometric mean square error, peak signal-to-noise ratio, structural similarity index and perceptual error; S33, construct an aircraft thin-walled component migration dataset according to the method in S1; S34, using the trained generative model as a pre-trained model and performing refined training using the migration dataset established in S33; S35, use multiple evaluation indicators in S32 to evaluate the model after fine-tuning training.

[0014] The above technical features can be combined in various suitable ways or replaced by equivalent technical features, as long as the purpose of the present invention can be achieved.

[0015] The present invention provides a generative model-based positioning stiffness characterization method for aircraft thin-walled components. Compared with the prior art, it has at least the following beneficial effects: (1) Rapid visualization of the positioning stiffness of thin-walled aircraft components is achieved with high efficiency; (2) Achieved high-resolution generation of the positioning stiffness field of thin-walled aircraft components with high accuracy; (3) The generation of positioning stiffness fields for various thin-walled components of aircraft is realized with a high degree of generalization; (4) The generation of “end-to-end” positioning stiffness of aircraft thin-walled components was realized, replacing finite element simulation.

[0016] This paper provides a method for rapidly generating stiffness fields for thin-walled aircraft components based on a generative deep learning model. The comprehensive and refined training model can be adapted to various components, effectively reducing the time cost of stiffness field simulation. This paper also provides effective solutions for process design, digital twins, and augmented reality visualization based on stiffness fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The present invention will be described in more detail below based on embodiments and with reference to the accompanying drawings, wherein: Figure 1 This is the overall implementation process of the new method proposed in the present invention; Figure 2 is the global generator structure of the model proposed in this invention; Figure 3 It is the local generator structure of the model proposed in the present invention; Figure 4 This is the multi-scale discriminator structure of the model proposed in this invention; Figure 5 It is the preferred embodiment of the present invention. Figure 6 This is a process characteristic diagram of an embodiment of the present invention; Figure 7 The process of obtaining the stiffness field of the embodiment of the present invention; Figure 8 is the stiffness field of an embodiment of the present invention; Figure 9 This is the model training result of the embodiment of the present invention; Figure 10 This is the loss evolution curve of the training process of an embodiment of the present invention; Figure 11This is a feature comparison between the stiffness field generated after training and the real stiffness field according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The present invention will be further described below with reference to the accompanying drawings.

[0019] The present invention provides a method for characterizing the positioning stiffness of thin-walled aircraft components based on a generative model. The technical problem to be solved is to address the deficiencies and needs of the above-mentioned existing technologies, establish the concept of stiffness field, introduce conditional generative adversarial networks, and realize the rapid generation of stiffness field. Figure 1 The specific methods, processes and contents of the present invention are as follows: S1, finite element batch processing and dataset construction based on Abaqus / Python; S2, generative model construction based on image translation mechanism and conditional generative adversarial network; S3, model pre-training and refinement training.

[0020] In step S1, batch processing of the input inp file and the Abaqus solver is performed in parallel to complete numerous finite element simulations and generate a large number of stiffness fields. Simultaneously, process characteristic maps corresponding to the stiffness fields are generated using the component's 3D digital model and actual process parameters. This process then acquires the dataset used for model training. This process specifically includes the following steps: S11, establishing a finite element simulation model of the thin-walled component in a positioned state, applying a force of 1 N along the assembly load direction at the assembly connection point, then saving all positioned nodes as a first node set, and saving all nodes that can be used as positioning nodes as a second node set; S12, writing the feasible node set of the positioning node into the xlsx file and the finite element model into the inp file; S13, using Python language, reading the xlsx file obtained in step S12, and randomly combining several groups of positioning solutions with different numbers and combinations of positioning points; S14, using Python language, read the inp file obtained in step S12, retrieve the location where the positioning point information is stored, and replace the initial positioning information with the random positioning information generated in step S13, thereby generating several inp files; S15, batch-calculating all the inp files generated in step S14 through the Abaqus / Python secondary development interface to obtain the simulation result odb file corresponding to each inp file; S16, opening the odb file obtained in step S15 through the Abaqus / Python secondary development interface, intercepting the deformation field of the component in the minimum stiffness direction, which is the positioning stiffness field of the component; S17, read the different positioning schemes generated in S13, combine the positioning point positions in each positioning scheme with the component geometry, the reinforcement position and the assembly load application position, and draw a process characteristic diagram.

[0021] In a further embodiment, in step S2, a deep learning model consisting of a global generator, a local generator, a multi-scale discriminator, a residual block, and an improved convolutional block attention mechanism is established to establish a one-to-one correspondence between the process feature map and the stiffness field. Specifically, the following steps are included: S21, based on the U-Net and ResNet structures, combined with the Convolutional Block Attention Mechanism (CBAM), built a global generator.

[0022] In a further embodiment, in step S21, combined with Figure 2 , the global generator construction process includes the following steps: S211, downsample the input process feature map. The downsampling consists of three cycles of "convolution → instance normalization → activation".

[0023] In a further embodiment, in step S211, the specific operations of the loop are: S2111, in each convolution operation, the convolution kernel size is 3×3, the stride is 2, and the padding is 1; At S2112, the activation operation is performed using the activation function Rectified Linear Unit (ReLU).

[0024] S212, after the input process feature map is downsampled, further feature extraction is performed through 10 consecutive residual blocks; each residual block consists of a cycle of "convolution → instance normalization → activation → convolution → instance normalization → improved CBAM".

[0025] In a further embodiment, in step S212, the specific operations of the loop are: S2121, in each convolution operation, the convolution kernel size is 3×3, the stride is 1, and the padding is 1; At S2122, the activation operation is performed using the activation function Rectified Linear Unit (ReLU). S2123, the improved CBAM is derived from the traditional CBAM framework.

[0026] Preferably, the channel attention module uses global average pooling, global maximum pooling, and global minimum pooling to reduce the dimensionality of the input feature map, outputting a feature map of size 1×1. The spatial attention module performs average pooling, maximum pooling, and minimum pooling along the channel dimension, forming a feature map with three channels. Furthermore, in the spatial attention module, a 3x3 convolution kernel is used instead of the traditional 7x7 convolution kernel.

[0027] S213, upsampling the process feature map after feature extraction, and the upsampling consists of three cycles of "deconvolution → instance normalization → activation".

[0028] In a further embodiment, in step S213, the specific operations of the loop are: S2131, in each convolution operation, the deconvolution kernel size is 3×3, the stride is 2, and the padding is 1; At S2132, the activation operation is performed using the activation function Rectified Linear Unit (ReLU).

[0029] S22, based on the U-Ne structure, constructed a local generator.

[0030] In a further embodiment, in step S22, combined with Figure 3 , the local generator construction process includes the following steps: S221, downsample the output image in the global generator of S21. Downsampling includes four cycles of "convolution → instance normalization → activation → convolution → instance normalization → activation → average pooling → instance normalization → activation".

[0031] In a further embodiment, in step S221, the specific operations of the loop are: S2211, in each convolution operation, the convolution kernel size is 3×3, the stride is 1, and the padding is 1; At S2212, the activation operation is performed using the activation function Rectified Linear Unit (ReLU). At S2213, average pooling is performed using a convolution operation with a kernel size of 2×2 and a stride of 2.

[0032] S222: Perform two cycles of “convolution → instance normalization → activation → convolution → instance normalization → activation” on the downsampled feature map.

[0033] In a further embodiment, in step S222, the specific operations of the loop are: S2221, in each convolution operation, the convolution kernel size is 3×3, the stride is 1, and the padding is 1; At S2222, the activation operation is performed using the activation function Rectified Linear Unit (ReLU).

[0034] S223: Upsampling is performed to restore the image to the same size as the input image. Upsampling includes four cycles of "deconvolution → instance normalization → activation → convolution → instance normalization → activation → convolution → instance normalization → activation".

[0035] In a further embodiment, in step S223, the specific operation of restoring the size is: S2231, in the deconvolution operation, the deconvolution kernel size is 2×2 and the stride is 2; At S2232, the activation operation is performed using the activation function Rectified Linear Unit (ReLU). S2233, in each convolution operation, the convolution kernel size is 3×3, the stride is 1, and the padding is 1.

[0036] S224, perform the final convolution with a kernel size of 1×1 and a stride of 1. The activation operation is completed using the hyperbolic tangent function (Tanh).

[0037] S23, build a multi-scale discriminator to judge whether the generated images are "real" or "fake" at different scales.

[0038] In a further embodiment, in step S23, combined with Figure 4 , the multi-scale discriminator construction process includes the following steps: S231 processes the image at different scales, including the original scale and a scale that has been downsampled once. The downsampling operation reduces the image scale to 1 / 4 of the original scale. S232: Downsampling is performed at each scale to obtain feature maps. Downsampling includes 3-4 cycles of "convolution → instance normalization → activation".

[0039] In a further embodiment, in step S232, the specific operations of the loop are: S2221, in the first three convolution operations, the convolution kernel size is 3×3 and the stride is 2; S2222, in the last convolution operation, the convolution kernel size is 3×3 and the stride is 1; At S2223, the activation operation is performed using the activation function Leaky Rectified Linear Unit (Leaky ReLU).

[0040] S233, after completing the downsampling of the two scales, uses the Sigmoid function to perform the discriminator's judgment.

[0041] S24, introduces multiple losses to comprehensively evaluate the similarity between the generated stiffness field and the actual stiffness field.

[0042] In a further embodiment, in step S24, the multiple losses include the following: S241, define the generator loss based on the least squares generative adversarial network (LSGAN). It is shown in the following formula: ; S242, define L1 loss. As shown in the following formula: ; S243, extract features through the pre-trained VGG19 model, thereby introducing perceptual loss. As shown in the following formula: ; S244, define the discriminator loss for the generated image and the real image. The superposition of the two losses is used as the final discriminator loss. It is shown in the following formula: .

[0043] Preferably, the losses defined by S241 to S243 are weighted summed as the total loss of the generator, as shown in the following formula: .

[0044] In a further embodiment, in step S3, model training is divided into two steps. In the first step, the deep learning model proposed in the second step is trained using the dataset generated in the first step. In the second step, the model trained in the first step is used as a pre-trained model, and the model is refined by inputting new data that the network has never seen before to achieve generalization. Because the pre-trained model is relatively complete, refined training can be completed using a smaller sample size.

[0045] S31, train the model built in S2.

[0046] S32 uses multiple evaluation metrics, including mean squared error (MSE) and mean absolute error (MAE) for local evaluation, coefficient of determination (R2), geometric mean squared error (GMSE), and peak signal-to-noise ratio (PSNR) for global evaluation, structural similarity index (SSIM), and perceptual error (PE). Perceptual error is extracted using a pre-trained VGG19. Different convolutional layers are integrated into three slices to evaluate the feature matching between the generated stiffness field and the ground-truth stiffness field. These metrics are defined as follows: ; ; ; ; ; .

[0047] S33, investigate representative thin-walled components in other aircraft structures, construct a data set according to the method in S1, and form a migration data set.

[0048] S34: Use the trained model as a pre-training model and perform refined training using the migration dataset established in S33 according to the training method described in S31.

[0049] S35, further evaluate the refined trained model according to the method described in S32.

[0050] In one embodiment, the rear fuselage structure of the aircraft, such as Figure 5 The structure consists of a central frame, reinforced wall panels, and inner and outer wall panels, all of which are thin-walled components.

[0051] Specifically, Figure 1 An optional process is given, which includes the following steps: Step 1: Create process feature maps and stiffness fields for each part under several different positioning conditions. Step 2: Write the implementation code for the conditional generative adversarial network model. Step 3: Train the model and evaluate it.

[0052] Step 11: Make several process feature maps of each part under different positioning conditions. Specifically, Figure 6 : (1) The black closed lines depict the collective shape of the parts; (2) The green lines depict the location distribution of the part's reinforcement ribs; (3) The blue node represents the location where the current part is connected to other parts after positioning, and is also the location where the unit assembly load is applied in the finite element; (4) The red node represents the positioning point of the current part, which is also the position where the fixed constraint is applied in the finite element; the constraint position is read through the inp file in step 12 (1).

[0053] Step 12: Create several stiffness fields for each part under different positioning conditions. Specifically, Figure 7 : (1) Create inp files in batches, modify the node number combination under the positioning point set of the original inp file, and form several inp files; (2) Conduct batch finite element simulation; (3) Intercept the deformation simulation result diagram of the part in the thickness direction; Figure 8 As shown, the stiffness field of parts generated in batches.

[0054] Step 21: Use Python and PyTorch to write the implementation code of the model described in S2. Specifically: (1) Use a global generator to extract the features of the process feature graph. The input of the global generator is the process feature graph, and the output is the process feature graph after feature extraction. (2) Generate the stiffness field using a local generator. The input of the local generator is the process feature map after feature extraction, and the output is the stiffness field. (3) Use the discriminator to judge the similarity between the generated stiffness field and the true stiffness field. If the discriminator determines that the generated stiffness field is true, it is considered that at this time, the stiffness field generated by the generator is real enough so that the discriminator cannot distinguish its true from false.

[0055] Step S22: Use Python and PyTorch to write the implementation code for the model loss described in S2. Specifically: (1) The generator loss is used to measure whether the generated stiffness field can be regarded as a real stiffness field by the discriminator; (2) L1 loss is used to evaluate the absolute difference between the generated stiffness field and the true stiffness field in pixel space; (3) The perceptual loss introduced based on the pre-trained VGG19 network is used to compare the feature expressions of the generated stiffness field and the real stiffness field and evaluate the similarity between them. This loss evaluation method takes into account the high-level features of the stiffness field. The loss is calculated by comparing the differences between the generated stiffness field and the real stiffness field at each slice; (4) The discriminator loss is used to measure whether the discriminator can accurately judge that the generated stiffness field is false.

[0056] Step S31: Train the model on a device with physical memory ≥ 16G and video memory ≥ 8G. Specifically: (1) Divide the training set, validation set, and test set into 7:2:1 parts; (2) Use Adam optimizer to dynamically adjust neural network parameters; (3) The initial learning rate is set to 0.0001; (4) The number of training rounds is set to 1000, where the initial learning rate is used for the first 200 rounds and the learning rate is linearly decayed for the next 800 rounds. (5) Load 1 picture per batch; (6) Set the weight of feature matching loss to 10.

[0057] S32, trained until the discriminator cannot accurately judge the authenticity of the stiffness field, and it is believed that the generator and discriminator are in Nash equilibrium.

[0058] Step S32, evaluate the model, the training results are as follows Figure 9 , combined with Figure 10 and Figure 11 , specifically: (1) The loss evolution during training is as follows Figure 10 shown.

[0059] For the generator, the loss decreases rapidly during the first 50 training rounds, then stabilizes but still exhibits significant fluctuations. Early in training, the generated stiffness field is of poor quality, resulting in high initial loss. As training progresses, the generator rapidly improves, bringing the generated images closer to real images, leading to a rapid decrease in loss. Furthermore, the discriminator's capabilities are relatively weak during early training, making it easier for the generator to relearn and reduce the loss, which contributes to the rapid decrease in loss. In subsequent training, the loss continues to fluctuate. This is because the generator and discriminator engage in a game of chess. As the images produced by the generator become increasingly realistic, the discriminator gradually improves its ability to distinguish between the generated images. This process leads to fluctuations in loss.

[0060] For the discriminator, early on, the losses on both the training and validation sets exhibited significant fluctuations. However, the overall trend in loss was downward, indicating that the discriminator was gradually learning to distinguish between images generated by the generator and real images. Between 200 and 800 training rounds, the loss remained relatively low, with minimal difference between the losses on the training and validation sets, with occasional minor fluctuations. This suggests that in the mid-term, the discriminator and generator reached a relatively balanced state. The discriminator was able to distinguish between real and generated images remarkably well, while the generator was also able to produce relatively realistic images. Fluctuations in loss were normal due to the competition between the generator and discriminator.

[0061] (2) The generated stiffness field and the real stiffness field show high consistency in all aspects.

[0062] (3) Feature matching perception error such as Figure 11 By pre-training VGG19 and integrating different convolutional layers into 3 slices, we can prove that the feature matching level between the generated stiffness field and the real stiffness field is very high.

[0063] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely illustrative of the principles and applications of the invention. It should be understood that many modifications may be made to the illustrative embodiments, and that other arrangements may be devised, without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in ways other than those described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be employed in conjunction with other described embodiments.

Claims

1. A method for characterizing the positioning stiffness of aircraft thin-walled components based on a generative model, characterized in that: The following steps are involved: S1, by constructing a three-dimensional digital model and actual process parameters, a process characteristic map corresponding to the stiffness field of the aircraft thin-walled component is generated, thereby obtaining a data set for model training; S2, constructs a generative model for aircraft thin-walled components based on image translation mechanism and conditional generative adversarial network; S3, using the data set generated in step S1 to train the generative model of step S2; after the generative model is trained, it is used as a pre-trained model, and the pre-trained model is refined by inputting new data.

2. The method for characterizing the positioning stiffness of aircraft thin-walled components based on a generative model according to claim 1, characterized in that: The stiffness field data of aircraft thin-walled components are obtained by batch processing the input text files and the finite element software solver.

3. The method for characterizing the positioning stiffness of aircraft thin-walled components based on a generative model according to claim 2, characterized in that: The acquisition of the dataset includes the following steps: S11, establishing a finite element simulation model of an aircraft thin-walled component in a positioned state, applying a force along the assembly load direction at the assembly connection point, and then saving all positioned nodes as a node set; S12, writing the node set into a database file and writing the finite element model into a text file; S13, reading the database file obtained in step S12, randomly combining several groups of positioning solutions with different numbers of positioning nodes and different combination forms, and generating random positioning information; S14, reading the text file obtained in step S12, searching for the location where the positioning information is stored, and replacing the initial positioning information with the random positioning information generated in step S13, thereby generating several text files; S15, batch-calculating the text files generated in step S14 through the secondary development interface to obtain simulation result files corresponding to each text file; S16, opening the simulation result file obtained in step S15 through the secondary development interface, intercepting the deformation field of the aircraft thin-walled component in the minimum stiffness direction, which is the positioning stiffness field of the aircraft thin-walled component; S17, read the different positioning schemes generated by S13, combine the positioning node position in each positioning scheme with the geometric shape of the aircraft thin-walled component, the position of the reinforcement rib and the position of the assembly load application, and draw a process feature diagram.

4. The method for characterizing the positioning stiffness of thin-walled aircraft components based on a generative model according to claim 3, wherein: The generative model is a deep learning model, including a global generator, a local generator, a multi-scale discriminator, a residual block and an improved convolutional block attention mechanism module.

5. The method for characterizing the positioning stiffness of thin-walled components of aircraft based on a generative model according to claim 4, characterized in that: The creation of a deep learning model involves the following steps: S21, based on U-Net and ResNet structures, combines the convolutional block attention mechanism to build a global generator; S22, based on the U-Ne structure, build a local generator; S23, construct a multi-scale discriminator to distinguish whether the generated images are "real" or "fake" at different scales; S24, introduces multiple losses to comprehensively evaluate the similarity between the generated stiffness field and the actual stiffness field.

6. The method for characterizing the positioning stiffness of thin-walled aircraft components based on a generative model according to claim 5, characterized in that: The global generator construction process includes the following steps: S211, downsampling the input process feature map, which includes multiple cycles of "convolution → instance normalization → activation"; S212: After the input process feature map is downsampled, it is further extracted through multiple consecutive residual blocks; each residual block includes a cycle of "convolution → instance normalization → activation → convolution → instance normalization → improved convolution block attention mechanism"; S213, upsampling the process feature map after feature extraction, and the upsampling includes multiple cycles of "deconvolution → instance normalization → activation".

7. The method for characterizing the positioning stiffness of thin-walled aircraft components based on a generative model according to claim 5, wherein: The local generator construction process includes the following steps: S221, downsampling the output image in the global generator; downsampling includes multiple cycles of "convolution → instance normalization → activation → convolution → instance normalization → activation → average pooling → instance normalization → activation"; S222, performing multiple cycles of "convolution → instance normalization → activation → convolution → instance normalization → activation" on the downsampled feature map; S223, upsampling is performed to restore the image to the same size as the input image; upsampling includes multiple cycles of "deconvolution → instance normalization → activation → convolution → instance normalization → activation → convolution → instance normalization → activation"; S224, perform the final convolution, and the activation operation is completed using the hyperbolic tangent function.

8. The method for characterizing the positioning stiffness of thin-walled aircraft components based on a generative model according to claim 5, wherein: The multi-scale discriminator construction process includes the following steps: S231, processing the image at different scales, including the original scale and the scale after one downsampling operation, where the downsampling operation reduces the image scale to 1 / 4 of the original scale; S232, downsampling is performed at each scale to obtain feature maps; downsampling includes multiple cycles of "convolution → instance normalization → activation"; S233, after completing the downsampling of the two scales, the discriminator uses the sigmoid function to perform discrimination.

9. The method for characterizing the positioning stiffness of thin-walled aircraft components based on a generative model according to claim 8, wherein: In step S232, the loop operation is: S2221, in the first three convolution operations, the convolution kernel size is 3×3 and the stride is 2; S2222, in the last convolution operation, the convolution kernel size is 3×3 and the stride is 1; At S2223, the activation operation is performed using a linear rectifier unit with a leaky activation function.

Images