A method based on cGAN to solve the scissor-fork problem in flap optimization

By generating flap positions that meet the scissor-fork requirement using a conditional generative adversarial neural network (cGAN), the scissor-fork problem in flap optimization is solved, improving the aerodynamic performance and stability of the optimization algorithm.

CN115203827BActive Publication Date: 2026-11-17BEIHANG UNIV
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202210836307.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-07-15
Publication Date
2026-11-17
Estimated Expiration
2042-07-15

AI Technical Summary

Technical Problem

In flap optimization design, existing methods are unable to effectively avoid the scissor-fork problem, which leads to frequent negative feedback or optimization space compression of the optimization algorithm, affecting the optimal solution for flap position.

Method used

Conditional Generative Adversarial Neural Network (cGAN) is used to generate flap positions that meet the scissor-fork requirements. Through the game between the generator and the discriminator, rich data that meets the constraints is generated, reducing invalid aerodynamic calculations and optimization algorithm divergence.

Benefits of technology

It improves aerodynamic performance during flap optimization, avoids algorithm divergence due to negative feedback, and maintains the richness of optimization variables and the accessibility of the optimal solution.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115203827B_ABST
    Figure CN115203827B_ABST
Patent Text Reader

Abstract

This invention discloses a cGAN-based method for solving the scissor-cross problem in flap optimization. It employs a conditional generative adversarial neural network (cGAN), where a generator and discriminator engage in a game-like interaction. Using a large amount of data conforming to the scissor-cross condition as samples, the generator can output data that meets the scissor-cross requirement, facilitating aerodynamic optimization. This method allows for greater consideration of aerodynamic performance improvement during flap optimization, preventing the optimization algorithm from diverging due to a large amount of negative feedback from non-scissor-cross data. Simultaneously, it avoids excessive compression of the sample space, which could lead to a decrease in the performance of the optimal solution.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of aircraft design and relates to the application of artificial intelligence technology in aircraft design. Specifically, it relates to a method based on cGAN to solve the scissor-fork problem in flap optimization. Background Technology

[0002] Lift enhancement devices for large aircraft play a crucial role in improving their takeoff and landing performance and are a key technology in large aircraft design. The design of lift enhancement devices for large aircraft must not only meet aerodynamic requirements but also satisfy comprehensive performance standards, including mechanical and structural aspects. Existing large aircraft trailing-edge lift enhancement devices all employ a two-flap configuration. During takeoff and landing, when the flaps are deployed, the flaps must be aligned as much as possible. Figure 1 The two flaps shown are not allowed to be arranged in a scissor-like pattern, with their front and back edges intersecting and their vertical alignment misaligned. However, in the optimized design of the flap position, the rotation axes of the inner and outer flaps (the inner section uses two reference points) are not allowed. and This indicates that the outer segment is... and ), deflection angle (the inner segment is θ) in outer segment θ out The parameters of the outer flap are independent of each other, which means that the flap position given by the optimization algorithm is likely not to meet the scissor-fork requirement. The essence of the problem is that the optimization variable range of the outer flap changes with the parameters of the inner flap. This patent proposes a method based on conditional adversarial generative neural networks, enabling the optimization algorithm to generate an outer flap position that meets the scissor-fork requirement, thereby satisfying the aerodynamic optimization needs.

[0003] In flap optimization, the time and computational cost of aerodynamic calculations are the most significant factors. However, given the flap parameters (…),… θ in θ out Calculating the spatial coordinates of the takeoff and landing positions of the flaps is very easy, and the computational cost is negligible compared to aerodynamic calculations. Therefore, the optimization process should consider minimizing invalid or meaningless aerodynamic calculations caused by non-compliance with the scissor-fork condition. Essentially, the reasonable range of parameters for the outer flap changes with the parameters for the inner flap, and the optimization algorithm should sample appropriate outer flap parameters after providing the inner flap parameters.

[0004] Currently, there are two main approaches to solving this problem:

[0005] 1. During optimization, if the configuration given by the optimization algorithm does not meet the scissor-fork requirement, a large negative feedback is given to the optimization algorithm.

[0006] 2. Fitting the outer flap parameters. Using a fitting method, the outer flap is treated as a function of the inner flap, i.e.:

[0007]

[0008] This method treats the outer flap as a "follower" and no longer considers its parameters as optimization variables.

[0009] For the existing method 1, a test was conducted using a long-range wide-body aircraft model as an example. Latin hypercube sampling was used across the entire range of optimization variables. In the generated samples, the flap positions that met the scissor-fork requirement in the takeoff configuration accounted for 3.2% of all sampling points, while the landing configuration accounted for only 1.7%. Figure 2 and Figure 3 The diagrams show the flap deflection angles of the inner and outer sections sampled for takeoff and landing configurations, respectively. If this method is used, the optimization algorithm will likely sample points that do not meet the scissor-fork requirements during aerodynamic optimization, thus generating negative feedback. A large amount of negative feedback can easily cause the optimization algorithm to diverge, making it impossible to obtain optimal flap position parameters.

[0010] For the existing method 2, the outer flap parameters are not considered as optimization variables, which actually leads to the compression of the optimization space. The optimal flap parameters are very likely not to fall within it. Summary of the Invention

[0011] To address the aforementioned problems, this invention proposes a cGAN-based method for resolving the scissor-cross problem in flap optimization. It employs a conditional generative adversarial neural network (cGAN), where a generator and discriminator engage in a game-like interaction. Using a large amount of data conforming to the scissor-cross condition as samples, the generator can output data that meets the scissor-cross requirements, facilitating aerodynamic optimization. This method allows for a greater focus on improving aerodynamic performance during flap optimization, preventing the optimization algorithm from diverging due to a large amount of negative feedback from non-scissor-cross data. Simultaneously, it avoids excessive compression of the sample space, which could lead to a decrease in the performance of the optimal solution.

[0012] The method of this invention, based on cGAN, is a way to solve the scissor-fork problem in flap optimization, and consists of the following steps:

[0013] A method based on cGAN to solve the scissor-fork problem in flap optimization is characterized by the following specific steps:

[0014] Step 1: Sample the given optimization variables using the Latin hypercube sampling method, including the inner flap optimization variables and the outer flap parameters; the inner flap parameters are in scalar form. The outer flap parameters are in scalar form. θ out ;in, and To determine two reference points for a spatial straight line, the X and Z coordinates are considered as optimization variables. The total number of optimization variables for the inner flap section is... The total number of optimization variables for the outer flap is The above The X and Z coordinates of reference point 1 of the inner flap rotation axis. The X and Z coordinates are the reference points for the rotation axis of the inner flap; The X and Z coordinates of reference point 1 for the rotation axis of the outer flap are given. The X and Z coordinates are the reference points for the rotation axis of the outer flap.

[0015] Step 2: Calculate whether the samples in sample library A meet the scissor-fork requirement in turn, and form sample library B with the samples that meet the requirement.

[0016] Step 3: For each sample in the sample library B, a set of virtual number vectors is randomly generated.

[0017] Step 4: Construct a generator neural network and a discriminator neural network, and initialize them.

[0018] The generator takes 10 inputs, 5 of which correspond to the optimization variables of the inner flap. The other five inputs correspond to a randomly generated set of virtual number vectors λ1, λ2, λ3, λ4, and λ5; the generator outputs five values, corresponding to the optimization variables of the outer flap. The discriminator has 15 inputs, of which 5 inputs correspond to The other 5 inputs correspond to λ1, λ2, λ3, λ4, λ5, and the remaining 5 inputs correspond to... The output is 1.

[0019] Step 5: Perform the first training of the discriminator using the Adam method, with the loss function as follows:

[0020] LossD1 = MES(y,cos(α))

[0021] In the formula, MES() is the mean squared error; y represents the actual output of the discriminator; cos(α) is the cosine similarity.

[0022] Step 6: Take the randomly selected sample from Step 5 The first 5 parameters and the last 5 parameters form a new sequence. Input is fed into the generator, which then generates the output. The superscript indicates that the data was generated by a generator; further, the sequence... The input is fed into the discriminator, and the Adam method is used to train the discriminator a second time. The loss function for the second training of the discriminator is:

[0023] LossD2 = MES(y, -1)

[0024] Where y represents the actual output of the discriminator.

[0025] Step 7: Using the samples extracted in Step 5, train a primary generator using the Adam method. The loss function is:

[0026] LossG = MES(y,1).

[0027] Step 8: Repeat steps 5 to 7 until the maximum number of training units is reached, and the trained generator is applied to a typical aerodynamic optimization framework.

[0028] The advantages of this invention are:

[0029] 1. This invention is a method based on cGAN to solve the scissor cross problem in flap optimization. It incorporates the scissor cross constraint into the aerodynamic optimization, avoiding invalid aerodynamic calculations caused by sampling a large number of samples that do not conform to the scissor cross constraint during the optimization process, and also avoiding the optimization algorithm from diverging due to receiving a large number of negative feedbacks.

[0030] 2. This invention is based on cGAN to solve the scissor-fork problem in flap optimization. It modifies the loss function of the discriminator during the first training in the traditional generative adversarial neural network and introduces vector similarity to make the generated outer flap data as similar as possible to the virtual number, thereby making the generated samples sufficiently rich and not compressing the value range of the optimization variables.

[0031] 1. This invention is based on cGAN to solve the scissor-fork problem in flap optimization. The introduction of vector similarity makes the data generated by the generator similar to the virtual number vector. The relationship between the data generated by the generator and the virtual number remains basically monotonic, without increasing the difficulty of the optimization algorithm to find the optimal solution. Attached Figure Description

[0032] Figure 1 a refers to the scissor-fork phenomenon of forward and backward misalignment between the inner and outer flaps during the flap design process;

[0033] Figure 1 b represents the cross-misalignment scissor fork phenomenon that occurs between the inner and outer flaps during the flap design process;

[0034] Figure 1 c represents the vertical misalignment scissor-fork phenomenon that occurs between the inner and outer flaps during the flap design process;

[0035] Figure 2 The results are based on sampling of the takeoff configuration of a wide-body passenger aircraft according to existing method 1;

[0036] Figure 3 The results are based on sampling of the landing configuration of a wide-body passenger aircraft according to existing method 1;

[0037] Figure 4This provides the implementation framework for the method of the present invention based on cGAN to solve the scissor fork problem in flap optimization;

[0038] Figure 5 This invention provides a framework for optimizing flap aerodynamics using a generator trained by the method of this invention.

[0039] Figure 6 The result is the sampling of the takeoff configuration of a wide-body passenger aircraft by the generator trained using the method of this invention;

[0040] Figure 7 The result is a sample of the landing configuration of a wide-body passenger aircraft using a generator trained by the method of this invention. Detailed Implementation

[0041] This invention provides a method for addressing the dynamic changes in optimization variables caused by flap scissor braces, based on conditional generative adversarial neural networks. Figure 4 As shown, the specific steps are as follows:

[0042] Step 1: Using the Latin hypercube sampling method, for the given optimization variables, including the inner flap parameters... (scalar form is) ), outer flap parameters (scalar form is) Samples are taken and stored in sample library A. and, Two reference points are used to determine a spatial straight line. These two reference points contain only four variables. In this invention, the X and Z coordinates are considered as optimization variables. Therefore, the total number of optimization variables for the inner section flap is... The total number of optimization variables for the outer flap is The above The X and Z coordinates of reference point 1 of the inner flap rotation axis. The X and Z coordinates are the reference points for the rotation axis of the inner flap; The X and Z coordinates of reference point 1 for the rotation axis of the outer flap are given. The X and Z coordinates are the reference points for the rotation axis of the outer flap.

[0043] Step 2: Calculate whether the samples in sample library A meet the scissor-fork requirement in turn, and form sample library B with the samples that meet the requirement.

[0044] The scissor fork calculation process is as follows: select any point Q on the outer side of the inner flap. in A point Q on the inner section of the outer flap out Furthermore, when the flaps are in the cruise position, the positions of these two points are completely coincident. With the flaps open, the finite helical displacement matrix [R]... H The coordinates of the flap opening positions at these two points can be calculated. and

[0045]

[0046] in, Represented as the finite helical displacement matrix of the inner flap. For the finite helical displacement matrix of the outer flap; and Q in With Q out The coordinates of the two cruise positions. Further calculations. and Does the distance between the points meet the requirements?

[0047] Step 3: For each sample in sample library B, randomly generate a set of virtual number vectors (λ1, λ2, λ3, λ4, λ5) corresponding to it, where the value of each element λ is between 0 and 1. At this point, each sample in sample library B can be represented as...

[0048] Step 4: Construct a generator neural network and a discriminator neural network, and initialize them.

[0049] like Figure 4 As shown, the generator has 10 inputs, 5 of which correspond to the optimization variables of the inner flap. The other five inputs correspond to a randomly generated set of virtual number vectors λ1, λ2, λ3, λ4, and λ5; the generator outputs five values, corresponding to the optimization variables of the outer flap. The discriminator has 15 inputs, of which 5 inputs correspond to The other 5 inputs correspond to λ1, λ2, λ3, λ4, λ5, and the remaining 5 inputs correspond to... The output is 1.

[0050] Step 5: Use the Adam method to train the discriminator for the first time.

[0051] A sample is randomly drawn from sample library B. The cosine similarity cos(α) is calculated using equation (2).

[0052]

[0053] Where, n i For the i-th optimization variable m i Let (λ1, λ2, λ3, λ4, λ5) be the i-th randomly generated virtual number element.

[0054] The loss function for the first training of the discriminator, LossD1, differs from that of classic generative adversarial neural networks, being:

[0055] LossD1=MES(y,cos(α)) (4)

[0056] In the formula, MES() is used to calculate the mean squared error; y represents the actual output of the discriminator.

[0057] The loss function takes into account the similarity between the virtual number vector and the generated outer flap parameters.

[0058] In classic generative adversarial neural network (GAN) design, only the generated data needs to meet certain requirements. However, for the flap position optimization problem, the generated data also needs to be sufficiently rich to minimize excessive compression of the variable space. This invention uses cosine similarity to evaluate the similarity between the virtual number and the generated data, aiming to make the generated data as close as possible to the virtual number vector, thereby maximizing data richness.

[0059] Step 6: Take the randomly selected sample from Step 5 The first 5 parameters and the last 5 parameters form a new sequence. Input is fed into the generator, which then generates the output. The superscript indicates that the data was generated by a generator. Further, the sequence... The input is fed into the discriminator, and the Adam method is used to train the discriminator a second time. The loss function for the second training of the discriminator is LossD2:

[0060] LossD2=MES(y,-1) (5)

[0061] Where y represents the actual output of the discriminator.

[0062] Step 7: Using the samples extracted in Step 5, train the generator once using the Adam method, with the loss function being LossG, where y represents the actual output of the discriminator.

[0063] LossG = MES(y,1) (6)

[0064] Step 8: Repeat steps 5 to 7 until the maximum number of training iterations is reached; in this invention, the maximum number of training iterations is set to 20,000. The trained generator is thus obtained, capable of... Figure 5 It is used in the typical aerodynamic optimization framework shown.

[0065] The test was also conducted using a model of a long-range wide-body aircraft as an example. Figure 6 and 7 As shown, the results indicate that the probability of samples generated from the takeoff configuration meeting the scissor-fork requirement increased to 93.3%, while the probability for the landing configuration increased to 82.0%. This demonstrates that the well-trained generator can generate samples that meet the requirements. Meanwhile, from... Figure 6 and Figure 7 As can be seen, the flap deflection parameters still cover a wide range, and the richness of the data has not decreased. Furthermore, the outer flap angle clearly increases with the inner flap angle, which is consistent with common understanding, indicating that a well-trained generator can produce appropriate data.

Claims

1. A method based on cGAN to solve the scissor-fork problem in flap optimization, characterized in that: The specific steps are as follows: Step 1: Sample the given optimization variables using the Latin hypercube sampling method, including the inner flap optimization variables and the outer flap parameters; the inner flap parameters are in scalar form. The outer flap parameters are in scalar form. ; in, Two reference points are used to determine a straight line in space, which are used to represent the rotation axis of the inner flap; and Two reference points are used to define a straight line in space, which are used to represent the rotation axis of the outer flap; and These represent the deflection angles of the inner and outer flaps, respectively; with the X and Z coordinates considered as optimization variables, the total number of optimization variables for the inner flap is... The total number of optimization variables for the outer flap is The above , The X and Z coordinates of reference point 1 for the inner flap rotation axis are given. , The X and Z coordinates are the reference points for the rotation axis of the inner flap; , The X and Z coordinates of reference point 1 for the rotation axis of the outer flap are given. , The X and Z coordinates are the reference points for the rotation axis of the outer flap; Step 2: Calculate the sample library sequentially Determine whether the samples in the database meet the scissor-cross test requirements, and compile a sample library from the samples that meet the requirements. The scissor-fork calculation process involves selecting any point on the outer side of the inner flap. and a point on the inner section of the outer flap Furthermore, when the flaps are in the cruise position, the positions of these two points are completely coincident; when the flaps are in the open position, through a finite helical displacement matrix... The coordinates of the flap opening positions at these two points can be calculated. and ; in, Represented as the finite helical displacement matrix of the inner flap. For the finite helical displacement matrix of the outer flap; and They are respectively and Coordinates of the two cruise positions; further calculations and Does the distance between the points meet the requirements? Step 3: For the sample library For each sample, a set of virtual number vectors is randomly generated; Step 4: Construct a generator neural network and a discriminator neural network, and initialize them; The generator takes 10 inputs, 5 of which correspond to the optimization variables of the inner flap. The other 5 inputs correspond to a randomly generated set of virtual number vectors. The generator outputs five variables, corresponding to the optimization variables of the outer flap section. The discriminator has 15 inputs, of which 5 inputs correspond to The other 5 inputs correspond to The remaining 5 inputs correspond to The output is 1. Step 5: Train the discriminator for the first time using the Adam method. The loss function is: In the formula, To calculate the mean square error; This represents the actual output of the discriminator; Cosine similarity; Step 6: Take the randomly selected sample from Step 5 The first 5 parameters and the last 5 parameters form a new sequence. The input is fed into the generator, which then generates the output. The superscript indicates that the data was generated by a generator; further, the sequence... The input is fed into the discriminator, and the Adam method is used to train the discriminator a second time. The loss function for the second training of the discriminator is: in, This represents the actual output of the discriminator; Step 7: Using the samples extracted in Step 5, train a primary generator using the Adam method. The loss function is: Step 8: Repeat steps 5 to 7 until the maximum number of training units is reached, and the trained generator is applied to a typical aerodynamic optimization framework.

2. The method for solving the scissor-fork problem in flap optimization based on cGAN as described in claim 1, characterized in that: Cosine similarity The method for obtaining it is as follows: From the sample library Randomly select a sample The cosine similarity is calculated using the following formula. : in, For the first One optimization variable ; For the first A randomly generated virtual number element .

Citation Information

Patent Citations

  • Supercritical airfoil profile generation method and device, electronic equipment and storage medium

    CN114692488A