Conditional diffusion model-based aircraft layout inverse design method and system

Abstract

The invention discloses an aircraft layout inverse design method and system based on a conditional diffusion model, and relates to the technical field of aircraft design and deep learning. Down-sampling, coding, diffusion, denoising and decoding of an aircraft point cloud model are achieved through three main components of a hidden space codec, a diffusion generator and a condition encoder, and the aircraft point cloud three-dimensional model meeting specific aerodynamic performance and stealth performance can be rapidly generated according to input design requirements. Dependence on experience of designers is reduced, design efficiency is improved, cost is reduced, and innovative discovery of aircraft layout is promoted.

Description

Technical Field

[0001] The present invention relates to the field of aircraft design and deep learning technology, and more specifically, to an aircraft layout inverse design method and system based on a conditional diffusion model. Background Art

[0002] Aircraft layout design methods are divided into two categories: forward design and reverse design. Forward design is to iteratively design the design parameters using wind tunnel tests and optimization algorithms after the designer proposes multiple layouts. The entire design process is iterated repeatedly, whether it is wind tunnel testing or numerical optimization, the economic and computational overhead is huge. In addition, with the introduction of multiple disciplines, it becomes more difficult for designers to propose initial solutions, which further restricts the innovation of aircraft layout design. Traditional reverse design methods aim to change the aerodynamic shape of the configuration so that its surface pressure distribution is as consistent as possible with the given target pressure distribution, so as to obtain a shape that meets the expected performance. However, this type of method further increases the requirements for designers because it requires a reasonable and excellent target pressure distribution, and it is difficult to give a reasonable target pressure distribution for three-dimensional shapes. When designers need to consider multidisciplinary constraints such as aerodynamics, stealth, and structure, traditional reverse design methods are difficult to handle high-dimensional reverse design problems corresponding to multiple objectives.

[0003] In recent years, generative artificial intelligence has developed rapidly and has attracted attention and innovation in various fields. Generative artificial intelligence can learn and simulate complex sample distributions, can effectively handle the generation problem of high-dimensional design space, and has made great achievements in the field of image and video generation. Based on methods such as generative adversarial networks (GANs) and variational autoencoders (VAE), designers have achieved target characteristic-guided airfoil inverse design by giving target characteristics such as aerodynamic coefficients and radar cross-sectional area (RCS) mean. However, the GANs model has problems such as training convergence difficulties and mode collapse, while the VAE model is prone to transition smoothing of generated samples, resulting in poor diversity of generated samples. These problems restrict the application of such methods in the inverse design of aircraft three-dimensional layout. The diffusion model processes the generation task as a denoising process. By precisely controlling the addition and removal of noise, it has the advantages of high flexibility, high quality of generated samples, and stable training. It has become the mainstream of generative model research.

[0004] The present invention proposes a generative design method for aircraft layout based on the conditional diffusion model. By giving geometric constraints, aerodynamic performance under different flight conditions, and RCS mean values ​​under different incident wave frequencies, a three-dimensional aircraft layout that meets the conditions is generated, reducing the designer's reliance on experience, providing new ideas for the three-dimensional layout design of aircraft, and laying the foundation for innovative discoveries in aircraft layout. Summary of the invention

[0005] The purpose of the present invention is to provide an aircraft layout inverse design method and system based on a conditional diffusion model to solve the problems of difficulty in handling multi-objective high-dimensional inverse design problems in the prior art and the limitations of generative artificial intelligence models.

[0006] The above technical objectives of the present invention are achieved through the following technical solutions: an aircraft layout inverse design method and system based on a conditional diffusion model, comprising the following steps:

[0007] Step 1: Use a pre-established database containing aircraft point cloud models, each model containing three-dimensional point cloud coordinates and corresponding aerodynamic performance and stealth performance data;

[0008] Step 2: Use the latent space codec to downsample and encode the aircraft point cloud model used for training to obtain the point cloud representation in the latent space;

[0009] Step 3: Perform forward point cloud diffusion and reverse denoising on the point cloud representation in the latent space through the diffusion generator to generate a denoised point cloud representation;

[0010] Step 4: Use the conditional encoder to encode the input design performance requirements into the latent space and combine it with the denoised point cloud representation;

[0011] Step 5: Decode the point cloud representation in the latent space into three-dimensional space through the decoder to generate the corresponding aircraft point cloud three-dimensional model.

[0012] The present invention is further configured as follows: the latent space codec adopts an autoencoder architecture based on an attention mechanism to capture the global structural information of the point cloud.

[0013] The present invention is further configured as follows: the latent space codec includes an encoder and a decoder, the encoder is used to convert a point cloud in a three-dimensional space into a latent space representation, and the decoder is used to decode the latent space representation into a three-dimensional point cloud.

[0014] The present invention is further configured as follows: the diffusion generator includes a forward diffusion network and a reverse denoising network, the forward diffusion network is used to gradually add Gaussian noise to the point cloud representation, and the reverse denoising network is used to gradually remove the noise and restore the point cloud representation.

[0015] By adopting the above technical scheme, the present invention realizes the automation and efficiency improvement of the design process by adopting the aircraft layout inverse design method and system based on the conditional diffusion model. Compared with the traditional design method, it significantly reduces the number of iterations of wind tunnel tests and numerical optimization, reduces economic and computational costs, and improves the diversity and innovation of the design. In addition, the method reduces the dependence on the personal experience of designers, enhances the ability to handle multidisciplinary constraints, and improves the quality of generated samples, avoiding common problems in GANs and VAE models, such as mode collapse and poor sample diversity. The latent space codec adopts an autoencoder architecture based on the attention mechanism to effectively capture the global structural information of the point cloud, while the diffusion generator generates high-quality samples by accurately controlling the addition and removal of noise, and the training process is stable. The present invention also reduces the dependence on large-scale annotated data, improves the accuracy and reliability of the design, promotes multidisciplinary collaborative design, and brings innovation and progress to the field of aircraft design.

[0016] Another object of the present invention is to provide an aircraft layout generative design system based on a conditional diffusion model, comprising:

[0017] Latent space encoder, used to downsample and encode the training aircraft point cloud model;

[0018] Diffusion generator, used for forward point cloud diffusion and inverse denoising;

[0019] A conditional encoder for encoding design performance requirements into a latent space;

[0020] The latent space decoder is used to decode the point cloud representation in the latent space into three-dimensional space and generate a three-dimensional model of the aircraft point cloud.

[0021] The present invention is further configured as follows: the conditional encoder uses a self-trained language model to encode the input design performance requirements into a latent space, and the encoder is capable of processing the design performance requirements input by a natural language and generating a corresponding conditional vector in the latent space.

[0022] The present invention is further configured such that: the system can generate a point cloud three-dimensional model of an aircraft that meets specific aerodynamic performance and stealth performance according to input design requirements.

[0023] The present invention is further configured such that: the conditional encoder uses a self-trained language model to convert the design performance requirements into a conditional vector in a latent space.

[0024] By adopting the above technical solution, the invention provides a generative design system for aircraft layout based on conditional diffusion model, which downsamples and encodes the aircraft point cloud model through a latent space encoder, uses a diffusion generator to perform forward point cloud diffusion and reverse denoising, and a conditional encoder to encode the design performance requirements into the latent space, and finally decodes the point cloud representation in the latent space into three-dimensional space through a latent space decoder to generate an aircraft point cloud three-dimensional model that meets specific requirements. The system specially designs a conditional encoder, uses a self-trained language model to process the design performance requirements of natural language input, and generates corresponding conditional vectors in the latent space, thereby making the design process more intelligent and automated.

[0025] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-mentioned aircraft layout generative design method based on the conditional diffusion model.

[0026] By adopting the above technical solution, the present invention provides a computer-readable storage medium, in which the stored computer program can realize the aircraft layout generative design method based on the conditional diffusion model when executed by the processor, so that the design process is automated and intelligent, greatly improving the design efficiency and quality, reducing the dependence on the experience of designers, and reducing the economic and computational costs of wind tunnel tests and numerical optimization. At the same time, it can handle multidisciplinary constraints and generate aircraft models that meet specific aerodynamic performance and stealth performance requirements, thereby enhancing the diversity and innovation of the design. In addition, the present invention understands and converts natural language input through a self-trained language model, making the encoding of design performance requirements more precise, improving the accuracy and reliability of the design, and providing a novel, flexible and efficient solution for aircraft layout design.

[0027] Another object of the present invention is to provide a method for using the above system, comprising the following steps:

[0028] S1: Build an aircraft point cloud model dataset;

[0029] S2: Conduct aerodynamic and stealth performance evaluation;

[0030] S3: Establish a latent space point cloud generation diffusion model based on dynamic attention mechanism;

[0031] S4: Training conditional diffusion model;

[0032] S5: Generate and evaluate the aircraft point cloud model.

[0033] By adopting the above technical solution, the present invention provides a method for using the above system. The method of the present invention not only improves the automation level of the design process, reduces the dependence on traditional wind tunnel tests and numerical optimization methods, reduces design costs and time, but also enhances the model's ability to capture the complex structure of point cloud data through a dynamic attention mechanism, and improves the accuracy and reliability of the generated model. In addition, the method can generate a variety of aircraft layout schemes according to specific aerodynamic and stealth performance requirements by training a conditional diffusion model, greatly improving the innovation and flexibility of the design, and bringing a new method that is efficient, accurate and has broad application prospects to the field of aircraft design.

[0034] In summary, the present invention has the following beneficial effects:

[0035] 1. The present invention significantly reduces the need for repeated iterations in the traditional design process, including wind tunnel testing and numerical optimization, through an automated design method based on a conditional diffusion model. This method reduces economic and computational overhead while improving the efficiency of the design process. By reducing the reliance on the experience of the designer, the present invention can quickly generate a point cloud 3D model of an aircraft that meets specific aerodynamic and stealth performance requirements, thereby accelerating the design cycle and reducing design costs.

[0036] 2. The present invention improves the efficiency of aircraft reverse design by combining the advantages of deep learning and diffusion models. It is applicable to single and multi-disciplines, has rich generation model diversity, and has strong method scalability. By encoding the design performance requirements into the latent space through the conditional encoder and combining it with the denoised point cloud representation, the present invention can generate multiple shapes, improving design efficiency and model quality. This method helps to realize innovative discovery of aircraft layout, reduces dependence on large-scale annotated data, and helps to realize innovative discovery of aircraft layout. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a schematic diagram of an aircraft layout inverse design framework based on a conditional diffusion model in an embodiment of the present invention;

[0038] Figure 2 is a point cloud distribution diagram of the sample shape in an embodiment of the present invention;

[0039] Figure 3 is a schematic diagram of a generation process in an embodiment of the present invention;

[0040] Figure 4 It is a schematic diagram of the appearance generated in the embodiment of the present invention. DETAILED DESCRIPTION

[0041] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solution of the present invention will be further described in detail below in conjunction with the embodiments of the present invention and the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0042] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will be described in detail below in conjunction with the embodiments.

[0043] Embodiment: An aircraft layout inverse design method based on a conditional diffusion model, such as Figure 1 As shown, the following steps are included:

[0044] Step 1: Build an aircraft point cloud model dataset

[0045] Use computer-aided design software to create a parametric three-dimensional geometric model of the aircraft (taking the total area and total length as design parameters), and discretize the geometric model into point cloud data P = {x1, y1, z1, x2, y2, z2, …, x N ,y N ,z N}, where N is the number of points. Figure 2 For the shape shown, N=40000.

[0046] Step 2: Aerodynamic and stealth performance evaluation

[0047] Computational fluid dynamics (CFD) software is used to evaluate the aerodynamic performance of the aircraft shape. By simulating the flow of the aircraft under different flight conditions (the incoming flow Mach number is 0.5, 0.85, and 1.5, and the angle of attack is 2°, 4°, and 6° respectively), the aerodynamic performance, such as lift and drag coefficients, lift-to-drag ratio, and moment coefficient, is calculated.

[0048] Evaluate the stealth performance of the aircraft shape and calculate the mean circumferential scattering cross section (RCS) in different radar bands: Where σ(θ,φ) is the direction-dependent scattering cross section, and θ and φ are angular parameters. In this example, the mean circumferential RCS of the aircraft for vertical and horizontal polarizations is evaluated for incident wave frequencies of 0.5 GHz and 9 GHz.

[0049] Step 3: Establish a latent space point cloud generation diffusion model based on dynamic attention mechanism

[0050] Step 3.1: Construct a latent space encoder-decoder: Design an autoencoder, including an encoder E and a decoder D. The encoder maps the point cloud P to the latent space point cloud representation X: X = E(P). The decoder maps the latent space point cloud representation X back to the point cloud

[0051] Step 3.2, build diffusion generator: build forward diffusion process and reverse denoising process. The forward process is to gradually add t times of Gaussian noise Z to the point cloud coordinate set U, and obtain the vertex coordinate set P after t times of adding noise t : in, is the parameter after adding noise t times; and α x Represents the parameters after adding noise for the xth time. The reverse denoising model iteratively predicts the noise introduced at each step and recovers clean data from the noise. This process can be expressed as:

[0052]

[0053]

[0054] in is the latent space point cloud data generated by the final denoising, Z t is the noisy point cloud data at time step t. At each step t, the model uses a denoising network ∈ θ (P t ,t) to predict the noise Z introduced in the forward process. This network is based on the dynamic Transformer architecture and can capture the complex structure of point cloud data. The self-attention mechanism of the dynamic Transformer multiplies the distribution of the point cloud latent space representation by the Q (query), K (key), and V (value) matrices, calculates the similarity between the query and the key by dot product operation, converts it into weights using softmax, and then passes through AdaLN (adaptive layer normalization). A MLP (multi-layer perceptron) is used to dynamically adjust the mean and variance of each layer, and the mean and variance of each layer are generated by MLP according to the position code P. Adaptive layer normalization enables the model to dynamically adjust its parameters according to time information and adapt to changes in point cloud data. It can then be expressed as:

[0055]

[0056] in is the latent space representation of the lth layer, D (l) is the dimension of the feature, P t is the positional encoding determined by time step t.

[0057] In the condition generation process, when using external conditions (such as text), and It is calculated from the external conditions through the conditional network τ, that is

[0058]

[0059] Dynamic Transformer fuses time step information by introducing dynamic position encoding, embeds it into the attention mechanism of Transformer, and uses dynamic position encoding to adjust the attention weight. The purpose of this is to allow the model to take time information into account in each diffusion process and capture the time dependency of point cloud data.

[0060] Step 3.3, construct a conditional encoder: Use a self-trained language model to convert the design performance requirement R into a conditional vector C in the latent space, expressed as C = Enc (R).

[0061] Step 3.4, training conditional diffusion model: use Adam optimization algorithm to accelerate gradient descent, initial learning rate is 0.0001, time step is 1000 steps, training round is 2000 rounds, batch training sample number is 64. The generation process diagram is as follows Figure 3 shown.

[0062] Step 4: Generate and evaluate the aircraft point cloud model

[0063] The above model is used to generate a latent space aircraft point cloud model, and the latent space point cloud representation is decoded into three-dimensional space through a decoder and evaluated. The evaluation indicators include aerodynamic performance and stealth performance. Given the Mach number of 1.5, the lift coefficient at an angle of attack of 2°, the lift-to-drag ratio, the circumferential RCS mean of vertical polarization at 9GHz, the circumferential RCS mean of horizontal polarization at 9GHz, and geometric constraints, three flying wing layouts are generated, such as Figure 4 The generated samples are numerically evaluated and the results are shown in Table 1.

[0064] Table 1 Comparison of generated results and target values

[0065]

[0066] Through the above steps, the present invention realizes the generative design of aircraft layout based on the conditional diffusion model, and can generate the corresponding aircraft point cloud model according to the input design requirements, thereby improving the design efficiency and reducing the dependence on the design experience of designers, which has important engineering application value.

[0067] This specific embodiment is merely an explanation of the present invention and is not a limitation of the present invention. After reading this specification, those skilled in the art may make non-creative modifications to the present embodiment as needed. However, as long as they are within the scope of the claims of the present invention, they are protected by the patent law.

Claims

1. An aircraft layout inverse design method based on a conditional diffusion model, characterized by: The following steps are involved: Step 1: Use a pre-established database containing aircraft point cloud models, each model containing three-dimensional point cloud coordinates and corresponding aerodynamic performance and stealth performance data; Step 2: Use the latent space encoder and decoder to downsample and encode the training aircraft point cloud model to obtain the point cloud representation in the latent space; Step 3: Perform forward point cloud diffusion and reverse denoising on the point cloud representation in the latent space through the diffusion generator to generate a denoised point cloud representation; Step 4: Use the conditional encoder to encode the input design performance requirements into the latent space and combine it with the denoised point cloud representation; Step 5: Decode the point cloud representation in the latent space into three-dimensional space through the decoder to generate the corresponding aircraft point cloud three-dimensional model.

2. The method according to claim 1, characterized in that: The latent space encoder-decoder adopts an attention-based autoencoder architecture to capture the global structural information of the point cloud.

3. The method according to claim 1, characterized in that: The latent space codec includes an encoder and a decoder, the encoder is used to convert a point cloud in a three-dimensional space into a latent space representation, and the decoder is used to decode the latent space representation into a three-dimensional point cloud.

4. The method according to claim 1, characterized in that: The diffusion generator includes a forward diffusion network and a reverse denoising network, the forward diffusion network is used to gradually add Gaussian noise to the point cloud representation, and the reverse denoising network is used to gradually remove the noise and restore the point cloud representation.

5. A generative design system for aircraft layout based on a conditional diffusion model, characterized by: include: Latent space encoder, used to downsample and encode the training aircraft point cloud model; Diffusion generator, used for forward point cloud diffusion and inverse denoising; A conditional encoder for encoding design performance requirements into a latent space; The latent space decoder is used to decode the point cloud representation in the latent space into three-dimensional space and generate a three-dimensional model of the aircraft point cloud.

6. The system according to claim 5, characterized in that: The conditional encoder uses a self-trained language model to encode the input design performance requirements into a latent space, and the encoder is capable of processing the design performance requirements inputted in a natural language and generating a corresponding conditional vector in the latent space.

7. The system according to claim 5, characterized in that: The system can generate a point cloud three-dimensional model of an aircraft that meets specific aerodynamic performance and stealth performance according to input design requirements.

8. The system according to claim 5, characterized in that: The conditional encoder uses a self-trained language model to convert design performance requirements into conditional vectors in a latent space.

9. A computer-readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the steps of the aircraft layout generative design method based on the conditional diffusion model described in any one of claims 1 to 4 are implemented.

10. A method of using the system of claim 5, characterized in that: The following steps are involved: S1: Build an aircraft point cloud model dataset; S2: Conduct aerodynamic and stealth performance evaluation; S3: Establish a latent space point cloud generation diffusion model based on dynamic attention mechanism; S4: Training conditional diffusion model; S5: Generate and evaluate the aircraft point cloud model.

Images