Aluminum alloy integral stiffened wallboard forming springback prediction method based on deep learning

Abstract

The invention belongs to the technical field of integral stiffened wallboard structures, and relates to an aluminum alloy integral stiffened wallboard forming springback prediction method based on deep learning. Comprising the following steps: S1, forming a discrete characterization method by combining structural characteristics of the complex integral stiffened wallboard and key parameters of a forming process; s2, obtaining sample data through a finite element simulation method, and establishing a springback result database of different stiffened wallboard structures; s3, establishing and training a deep learning neural network model; and S4, performing springback prediction on a new sample by using the trained model. According to the method, the age forming springback of the aluminum alloy integral wall plate with the multi-curvature, multi-thickness and high-rib characteristics can be quickly and accurately predicted, and technical auxiliary support can be well provided for forming of the large complex integral stiffened wall plate.

Description

Technical Field

[0001] This invention belongs to the technical field of integral stiffened wall panel structure, and relates to a method for predicting the springback of integral stiffened aluminum alloy wall panels based on deep learning. Background Technology

[0002] Large, integral stiffened panels for aerospace applications are crucial load-bearing structures for aircraft wings and fuselages, and their design and manufacturing technologies represent a significant direction for technological development in the aerospace equipment field. Currently, the most mature manufacturing technologies for integral stiffened panels for aerospace applications include shot peening and aging forming. Finite element method (FEM) simulation is the primary means of predicting springback during panel forming. However, with the increasing size and complexity of panel structures, the time cost of FEM calculations has increased significantly. For panel components with complex structures such as multiple curvatures, varying thicknesses, and high stiffness, the springback profile exhibits poor determinism and may even fail to converge, making it impossible to accurately predict their springback.

[0003] In recent years, with the development of artificial intelligence, machine learning, due to its strong learning ability and low time cost, has been increasingly applied to the prediction of metal component forming, and has already been used in sheet metal, pipe and other forming processes. Using deep learning technology to build neural networks and applying them to the prediction of springback in the forming of large integral stiffened panels can greatly improve early-stage R&D efficiency and is an effective way to solve the aforementioned problems. Summary of the Invention

[0004] Purpose of the invention Based on the above analysis, the purpose of this invention is to provide a deep learning-based method for predicting the springback of integral stiffened aluminum alloy panels. This method can discretize and characterize the complex integral stiffened panel structure, and combine deep learning technology to achieve fast and accurate prediction of the springback of the panel, effectively reducing computation time and resource consumption.

[0005] Technical solution A deep learning-based method for predicting the springback during forming of integral stiffened aluminum alloy panels, comprising the following steps: S1: Combining the structural characteristics of complex integral stiffened wall panels with key parameters of forming process, a discrete characterization method is formed; S2: Obtain sample data through finite element simulation and establish a database of springback results for different stiffened panel structures; S3: Build and train a deep learning neural network model; S4: Use the trained model to predict the rebound of new samples.

[0006] Furthermore, step S1 specifically involves: the discrete characterization method for complex stiffened panels refers to using a set of parameters to represent the structural and technological characteristics of any point on the panel. Each set of parameters includes structural characteristic values, technological characteristic values, and springback characteristic values; wherein the structural characteristic values ​​include the skin thickness (δ) at that point. s ), rib height (δ) r ) and its structural change coefficient with adjacent points (θ) x+ θ x- θ z+ θ z- The process characteristic values ​​include forming time (t) and die radius (R), and the springback characteristic value is the springback difference (ΔY) at that point. xz ).

[0007] Furthermore, step S2 specifically involves: defining the skin thickness (δ) according to a conventional reinforced wall panel structure. s ) and rib height (δ) r The forming time (t) and the radius of the mold in the x-direction (R) are limited by process experiments. x ) and the radius of the mold in the z-direction (R) z The range is determined, and then the forming simulation analysis of the stiffened wall panels with different structural features or process parameters is carried out by forming finite element simulation method, and the three-dimensional coordinates of the mesh nodes of the wall panel finite element model are derived. Furthermore, the three-dimensional coordinates include the x-coordinate (x0) and z-coordinate (z0) before node formation, and the springback coordinate (y0) after formation. t ); Furthermore, the rebound difference (ΔY) at the i-th point i xz The value of is calculated using the following formula:

[0008] In addition, the skin thickness, rib height, forming time, and mold radius of point i are respectively expressed as δ i s δ i r t i R i x and R i z When there is no rib at point i, then δ i r =δ i s The structural variation coefficients θ of this point in the x+, x-, z+, and z- directions are... i x+ θ i x- θi z+ and θ i z- They are calculated using the following formulas respectively:

[0009]

[0010]

[0011]

[0012] Where δ x+ δ x- δ z+ and δ z- The ratio of the skin thickness or rib height at each structural abrupt change point i in each direction to the skin thickness or rib height at the current point, and s x+ s x- s z+ and s z- These represent the distances from the structural abrupt change points in each direction.

[0013] Furthermore, the skin thickness, rib height, structural variation coefficient, forming time, and mold radius at point i are used as input parameters for the sample, expressed as (δ i s δ i r θ i x+ θ i x- θ i z+ θ i z- t i R i x R i z The output parameter is the rebound difference (ΔY). i xz ), forming a sample database {D i |i=1,...,N},D i It includes the input and output parameters for that point.

[0014] Furthermore, step 3 specifically involves: building a rebound prediction model based on a deep learning neural network, including an input layer, a hidden layer, and an output layer.

[0015] Furthermore, the input layer consists of the input parameters in S2, the output layer consists of the output parameters in S2, and the hidden layer is initially set to three layers. The root mean square error of the output parameters is used as the loss function, and the Adam optimization algorithm is used to iteratively optimize the neural network with the goal of minimizing the loss function. 90% of the samples in the sample dataset are used as the training set and 10% of the samples are used as the test set to train the neural network and optimize the hyperparameters, finally obtaining the optimal rebound prediction neural network model.

[0016] Furthermore, step 4 specifically involves: for any new stiffened wall panel component with multiple curvatures, variable thickness, and high-strength structure, first using the discrete representation method in S1 to express the component as a new sample set {P}. i |i=1,...,N},P i The input parameters for the i-th point are included; then, the rebound prediction neural network model trained in S3 is used to predict the new sample set; finally, the prediction results are restored into point cloud data according to the length and width dimensions of the new component, and the rebound surface is drawn in CATIA software.

[0017] The beneficial effects of this application are as follows: This invention applies deep learning neural networks to predict springback during the aging process of integral stiffened aluminum alloy panels. It eliminates the need for finite element modeling, preprocessing, and solving of the structure. The trained model can complete predictions within tens of seconds, effectively solving the problems of low computational efficiency and convergence difficulties inherent in traditional finite element simulation methods when dealing with complex large panels with multiple curvatures, varying thicknesses, and high stiffness. The core innovation of this invention lies in proposing a discrete representation method. This method uses a set of parameters to describe the local structural process characteristics and springback results at any point on the panel, decoupling the model input from the overall shape, size, and specific design of the panel. The structural change coefficient effectively quantifies the local impact of geometrical abrupt changes in adjacent areas on springback. The unique springback difference calculation effectively describes the local springback variation characteristics of the component, thus avoiding interference from the overall dimensions. The trained model can be directly applied to new components with multiple curvatures, varying thicknesses, and high stiffness within the range of commonly used design parameters, overcoming the bottleneck of current simple machine learning methods limited to specific geometric configurations. The prediction results of this invention can also be easily restored based on the component dimensions and the springback profile can be directly reconstructed in software such as CATIA, providing input for mold compensation and process optimization. The discrete characterization method of this invention has important reference value for the prediction of springback in the forming of other complex thin-walled components with neural network assistance, and provides a fast, accurate and universal prediction method for the manufacturing of large aerospace structural components. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2This is a schematic diagram of a single-curvature variable-thickness high-ribbed wall panel according to an embodiment of the method of the present invention; Figure 3 This is a schematic diagram illustrating the method for calculating the structural change coefficient. Figure 4 This is a diagram showing the dimensions (unit: mm) of the reinforced wall panel used in an embodiment of the method of the present invention. Figure 5 This is a comparison chart of the neural network prediction results and simulation prediction results for the springback of the stiffened wall panel formed according to an embodiment of the method of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The features and illustrative embodiments of various aspects of the present invention will now be described in detail. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention may be practiced without requiring some of these specific details. The following description of embodiments is merely intended to provide a better understanding of the invention by illustrating examples of the invention. The invention is by no means limited to any specific setup and method set forth below, but covers any improvements, substitutions, and modifications to the structures, methods, and devices without departing from the spirit of the invention. In the following description, well-known structures and techniques are not shown to avoid unnecessarily obscuring the invention.

[0021] In the description of this invention, it should be noted that the directions or positional relationships indicated by terms such as "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer" are based on the stated directions or positional relationships and are only for the convenience of describing and simplifying the invention, and should not be construed as limiting the invention. Furthermore, the use of ordinal numbers (e.g., "first and second," etc.) is for distinguishing objects and is not limited to this order, and should not be construed as indicating or implying relative importance.

[0022] The flowchart of the method of this invention is shown below. Figure 1 As shown, the following is a detailed description using a specific embodiment as an example: S1: Combining the structural characteristics of complex integral stiffened wall panels with key parameters of forming process, a discrete characterization method is formed; S11: For a type of age-formed single-curvature variable-thickness high-ribbed wall panel, such as... Figure 2 As shown, the structural features of a certain unit can be represented by the skin thickness (δ). s ), rib height (δ) r ), x+ direction structural change coefficient (θ) x+ ) and x-direction structural variation coefficient (θ) x- The process characteristics can be represented by forming time (t) and die radius (R), and the springback characteristic is the springback difference (ΔY) at that point. x The structural change coefficient at point i is calculated using the following formula:

[0023]

[0024] Where δ x+ and δ x- The ratio of the skin thickness or rib height at structural abrupt changes in each direction to the skin thickness or rib height at the current point, and s x+ and s x- These represent the distances from the structural abrupt change points in each direction, such as... Figure 3 As shown.

[0025] S2: Obtain sample data through finite element simulation and establish a database of springback results for different stiffened panel structures; S21: 7150 aluminum alloy is selected as the wall panel material, and a two-stage aging forming process is used for the wall panel forming process. Based on experience, the wall panel skin thickness range is limited to 2-8mm, the rib height range is 25-40mm, the rib thickness range is 2-4mm, the forming time range is 2-3 hours, and the mold radius range is 650-4000mm.

[0026] S22: Use ABAQUS software to perform two-stage aging forming simulation of stiffened wall panels with different structural features or process parameters as described above. The finite element mesh of the wall panel should have a sufficiently high mesh density in the x-direction.

[0027] S23: After the simulation calculation is completed, export the three-dimensional coordinates of the mesh nodes of the finite element model of the wall panel, including the coordinates of the nodes before forming (x0) and the springback coordinates after forming (y0). t ), the rebound difference at the i-th point (ΔY) i x The calculation formula for ) is as follows:

[0028] S24: The skin thickness, rib height, structural variation coefficient, forming time, and mold radius derived from each node are used as input parameters for the sample, expressed as (δs δ r θ x+ θ x- The output parameter is the springback difference (ΔY), which is the difference between the springback values ​​(ΔY, Δt, and ΔR). x ), forming a sample database {D i |i=1,...,N},D i It includes the input and output parameters for that point.

[0029] S3: Build and train a deep learning neural network model; S31: Build a rebound prediction model based on deep learning neural network, including input layer, hidden layer and output layer. The input layer is composed of the input parameters in S2, the output layer is composed of the output parameters in S2, and the hidden layer is initially set to three layers. The root mean square error of the output parameters is used as the loss function. The Adam optimization algorithm is used to iteratively optimize the neural network. The optimization objective is to minimize the loss function.

[0030] S32: Use 90% of the samples in the sample dataset as the training set and 10% of the samples as the test set to train the neural network and optimize the hyperparameters, and finally obtain the optimal rebound prediction neural network model.

[0031] S4: Use the trained model to predict the rebound of new samples.

[0032] S41: For a new stiffened wall panel component with multiple curvatures, variable thickness, and high ribs, such as... Figure 4 As shown, a forming time of 2 hours and a mold radius of 480 mm were selected as the forming process parameters. The discrete characterization method in S1 was used to express the component as a new sample set {P}. i |i=1,...,N},P i The input parameters include the i-th point, and the density of discrete points in the x-direction of the wall should be sufficiently large.

[0033] S42: Combine the rebound prediction neural network model trained in S3 to predict the validation sample set.

[0034] S43: Based on the length and width dimensions of the verification component, reconstruct the predicted results into point cloud data, draw the springback profile in CATIA software, and compare it with the simulated springback profile results with the same process parameters. Figure 5As shown, the results indicate that the error between the springback profile of the stiffened panel predicted by the neural network and the springback profile simulated by the finite element method is between 0.93 and -0.7 mm, with 80% of the profile errors being between 0.53 and -0.47 mm. This demonstrates the effectiveness of the neural network model in predicting the springback profile of the finite element method. Moreover, the calculation time for the profile prediction by the neural network model is less than 10 seconds, while the calculation time for the finite element method simulation is over 7200 seconds, resulting in a significant improvement in efficiency.

[0035] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly, encompassing both direct connection and indirect connection via an intermediate medium. Those skilled in the art can understand the specific meaning of these terms in this invention based on the specific circumstances.

[0036] It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other, and the various embodiments can be referenced and cited in each other. The present invention will now be described in detail with reference to and in conjunction with the embodiments.

[0037] The above detailed embodiments are a description of the present invention. It should not be considered that the specific embodiments of the present invention are limited to these descriptions. For those skilled in the art, several simple deductions and substitutions can be made without departing from the concept of the present invention, and all of these should be considered to fall within the protection scope of the present invention.

Claims

1. A method for predicting springback during forming of integral stiffened aluminum alloy panels based on deep learning, characterized in that the steps include... include: Step S1: Combining the structural characteristics of complex integral stiffened panel and key parameters of forming process, a discrete characterization method is formed; Step S2: Obtain sample data through finite element simulation and establish a database of springback results for different stiffened panel structures; Step S3: Build and train a deep learning neural network model; Step S4: Use the trained model to predict the rebound of new samples.

2. The method as described in claim 1, characterized in that, The specific step S1 is as follows: The discrete characterization method of complex stiffened wall panels refers to using a set of parameters to represent the structural and technological characteristics of any point on the wall panel. Each set of parameters includes structural characteristic values, technological characteristic values, and springback characteristic values. The structural characteristic value includes the skin thickness δ at that point. s Fiber height δ r and its structural variation coefficient θ with adjacent points x+ θ x- θ z+ θ z- The process characteristic values ​​include forming time t and mold radius R, and the springback characteristic value is the springback difference ΔY at that point. xz .

3. The method as described in claim 2, characterized in that, Step S2 specifically involves: defining the skin thickness δ according to a conventional reinforced wall panel structure. s and rib height δ r The range is defined by process experiments, specifying the forming time t and the mold radius R in the x-direction. x and the radius R of the mold in the z-direction z The range is then determined, and the forming simulation analysis of the stiffened wall panels with different structural features or process parameters is carried out using the forming finite element simulation method, and the three-dimensional coordinates of the mesh nodes of the wall panel finite element model are derived.

4. The method as described in claim 3, characterized in that, The three-dimensional coordinate system includes the x-coordinate (x0) and z-coordinate (z0) before node formation, and the springback coordinate (y) after formation. t .

5. The method as described in claim 4, characterized in that, The rebound difference ΔY at the i-th point i xz The value is calculated using the following formula: In addition, the skin thickness, rib height, forming time, and mold radius of point i are respectively represented by δ. i s δ i r t i R i x and R i z When there is no rib at point i, then δ i r =δ i s The structural variation coefficients θ of this point in the x+, x-, z+, and z- directions are... i x+ θ i x- θ i z+ and θ i z- They are calculated using the following formulas respectively: Where δ x+ δ x- δ z+ and δ z- The ratio of the skin thickness or rib height at each structural abrupt change point i in each direction to the skin thickness or rib height at the current point, and s x+ s x- s z+ and s z- These represent the distances from the structural abrupt change points in each direction.

6. The method as described in claim 5, characterized in that, The skin thickness, rib height, structural variation coefficient, forming time, and mold radius at point i are used as input parameters for the sample, expressed as δ. i s δ i r θ i x+ θ i x- θ i z+ θ i z- t i R i x R i z The output parameter is the rebound difference ΔY. i xz To form a sample database {D i |i=1,...,N},D i It includes the input and output parameters for that point.

7. The method as described in claim 6, characterized in that, Step S3 specifically involves: building a rebound prediction model based on a deep learning neural network, including an input layer, a hidden layer, and an output layer.

8. The method as described in claim 7, characterized in that, The input layer consists of the input parameters from step S2, and the output layer consists of the output parameters from step S2. The hidden layer is initially set to three layers. The root mean square error of the output parameters is used as the loss function. The Adam optimization algorithm is used to iteratively optimize the neural network, with the optimization objective being to minimize the loss function. 90% of the samples in the sample dataset are used as the training set, and 10% of the samples are used as the test set to train the neural network and optimize the hyperparameters, finally obtaining the optimal rebound prediction neural network model.

9. The method as described in claim 8, characterized in that, Step S4 specifically involves: for any new stiffened wall panel component with multiple curvatures, variable thickness, and high rib structure, first using the discrete representation method in S1 to express the component as a new sample set {P}. i |i=1,...,N},P i The input parameters for the i-th point are included; then, the rebound prediction neural network model trained in S3 is used to predict the new sample set; finally, the prediction results are restored into point cloud data according to the length and width dimensions of the new component, and the rebound surface is drawn in CATIA software.

Images