Aircraft overall scheme generative design method based on PILO-GAN

Abstract

The invention relates to the technical field of aircraft design, in particular to an aircraft overall scheme generative design method based on PILO-GAN, and the method comprises the steps: converting an unstructured natural language demand into structured design parameters; then constructing a PILO-GAN network, generating initial design parameters by taking structured parameters as conditional input and combining with random noise, embedding aircraft design empirical criterion constraints, and driving physical optimization based on multidisciplinary performance evaluation; performing multidisciplinary physical performance verification on the initial design parameters to obtain quantitative performance indexes; constructing a gradient propagation path by adopting a REINFORCE algorithm, and feeding back a performance index gradient to the generator to realize parameter updating; and finally, dynamically adjusting the weight of the loss item of the constraint and physical optimization, and carrying out iterative optimization until a preset requirement is met. Through a physical closed-loop optimization mechanism, end-to-end intelligent generation of feasible schemes from demands to engineering is realized, and the physical consistency and engineering credibility of the design scheme are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of aircraft design technology, and in particular to a generative design method for overall aircraft schemes based on PILO-GAN. Background Technology

[0002] The overall design of an aircraft is the most critical and forward-looking stage in the entire lifecycle of aircraft development, directly determining the aircraft's core performance, lifecycle cost, and technological feasibility. Currently, the overall design of aircraft still relies heavily on expert experience, requiring repeated iterative verification during the design process, which presents problems such as long iteration cycles and difficulty in guaranteeing the engineering effectiveness of initial solutions.

[0003] With intensifying technological competition and diversified mission requirements, aircraft design is shifting towards unconventional layouts, novel propulsion systems, and intelligent design paradigms. Traditional design methods, developed based on empirical knowledge and iterative verification, struggle to handle complex nonlinear relationships involving multidisciplinary coupling, and rely excessively on sequential trial and error, limiting design space exploration and failing to meet the demands of efficient and intelligent aircraft design.

[0004] Generative artificial intelligence (GAN) technology, especially Generative Adversarial Networks (GANs), offers new solutions for aircraft design. While existing GAN technologies have been applied in fields such as aerodynamic design, significant shortcomings remain in generating overall aircraft designs: First, there is a lack of effective integration with the inherent physical constraints of multidisciplinary coupling within aircraft, making it difficult to meet engineering practice requirements; second, most GANs focus on two-dimensional airfoils or point cloud models, failing to cover complete overall aircraft designs and exhibiting insufficient adaptability to dynamic design needs; and third, their generalization ability is weak in scenarios with limited high-quality samples, failing to effectively integrate aerospace knowledge with physical mechanisms.

[0005] Therefore, there is an urgent need for a method for generating overall aircraft solutions that can integrate domain knowledge and physical constraints, support demand-driven generation, and balance the innovation of solutions with engineering effectiveness, in order to solve the problems of existing technologies such as reliance on expert experience, long iteration cycles, and poor physical consistency. Summary of the Invention

[0006] To address these issues, this invention provides a generative design method for overall aircraft schemes based on PILO-GAN, thereby resolving the aforementioned problems in the prior art.

[0007] To achieve the above objectives, this invention provides a generative design method for overall aircraft schemes based on PILO-GAN, comprising:

[0008] Step S1: Obtain aircraft design requirements information and convert its unstructured natural language requirements into structured design parameters;

[0009] Step S2: The structured design parameters are used as conditional inputs to construct a PILO-GAN network, which combines random noise to generate initial design parameters for the overall aircraft scheme. The parameters are embedded with aircraft design empirical criteria and driven by multidisciplinary performance evaluation to optimize physics.

[0010] Step S3: Perform multidisciplinary physical performance verification on the initial design parameters to obtain quantitative performance indicators;

[0011] Step S4: The REINFORCE algorithm is used to construct a gradient propagation path, and the gradient of the quantization performance index is fed back to the generator to realize parameter update and obtain update result;

[0012] Step S5: Dynamically adjust the weights of the loss terms in the constraints and physical optimization, and iteratively execute steps S2 to S4 based on the update results until the generated design parameters meet the preset requirements to obtain the target design parameters.

[0013] Furthermore, the process of step S2 includes:

[0014] Based on the structured design parameters and random noise, the initial design parameters are obtained by constructing a generator using a ResNet architecture and integrating a self-attention mechanism for forward propagation computation.

[0015] Based on aircraft design experience guidelines, the degree of constraint violation of the design parameters is calculated through a parameter constraint mechanism. Weighted hyperbolic tangent nonlinear scaling and batch averaging constraints are used to obtain compliance verification results.

[0016] Based on the requirements of multidisciplinary performance evaluation, a weighted hyperbolic tangent nonlinear optimization of the maximum takeoff weight, empty weight, and fuel consumption weight is applied to the initial design function through a physical optimization mechanism to obtain the physical loss value.

[0017] Based on the fusion features of the initial design parameters and conditional labels, a discriminator is constructed using a CNN architecture and an integrated self-attention mechanism to perform authenticity scoring, thereby obtaining adversarial training signals to complete the construction of the PILO-GAN network.

[0018] Furthermore, the process of constructing a generator using the ResNet architecture and integrating a self-attention mechanism for forward propagation computation includes:

[0019] ,

[0020] in, denoted as generator function, outputting initial design parameters, z as random noise, y as conditional label, ⊕ as feature concatenation, FC1 and FC2 as fully connected layers, ResBlocks as residual block sequence, and Attention as self-attention mechanism.

[0021] Furthermore, the process of calculating the degree of constraint violation of the initial design parameters through the parameter constraint mechanism includes:

[0022] ,

[0023] in, ωk is the parameter constraint loss value, m is the number of training batch samples, n is the total number of constraint terms, ωk is the weight coefficient of the kth constraint, sk is the non-linear scaling factor, and Lk(i) is the original loss value of the i-th sample under the kth constraint.

[0024] Furthermore, the process of applying a weighted hyperbolic tangent nonlinear optimization of the initial design function using the maximum takeoff weight, empty weight, and fuel consumption weight through a physical optimization mechanism to obtain the physical loss value includes:

[0025] ,

[0026] in, This represents the physical loss value. To optimize the principal loss weights in physics, These represent the losses corresponding to maximum takeoff weight, empty weight, and fuel consumption weight, respectively. This corresponds to the non-linear scaling factor.

[0027] Furthermore, the process of step S3 includes:

[0028] The initial design parameters were verified using the vortex lattice method to obtain normalized quantitative performance indicators, which include the ratio of fuel consumption weight to maximum takeoff weight, the ratio of empty weight to maximum takeoff weight, and the lift coefficient.

[0029] Furthermore, the process of step S4 includes:

[0030] By generating N independent noise samples and propagating them forward through the generator, the physical performance is evaluated to obtain the quantitative performance index of each sample.

[0031] The sample mean is calculated based on the quantitative performance index as a baseline, and the relative performance of each sample is obtained by calculating the dominance function.

[0032] Based on the relative performance advantages and disadvantages, the gradient of the generator output with respect to the input noise is calculated, and the REINFORCE algorithm is used to estimate the gradient of the physical performance with respect to the noise.

[0033] Based on the gradient, it is propagated to the generator parameters using the chain rule, and the parameters are updated by combining the gradients of the adversarial loss and the constraint loss to obtain the update result.

[0034] Furthermore, the process of propagating the gradient to the generator parameters using a chain rule, and updating the parameters by combining the gradients of the adversarial loss and the constraint loss to obtain the updated result includes:

[0035] The generator parameter gradient is calculated using the generator Jacobian matrix based on the gradient of the physical properties with respect to noise.

[0036] The total gradient is obtained by weighted fusion of the generator parameter gradient, the adversarial loss gradient, and the constraint loss gradient.

[0037] The magnitude of the total gradient is limited to a preset threshold range by gradient clipping, and the generator parameters are updated using the Adam optimizer to obtain the updated result.

[0038] Furthermore, the process of step S5 includes:

[0039] By setting a constraint loss threshold and a physical loss threshold, the update result is monitored for loss value to obtain a loss exceeding the limit judgment result;

[0040] Based on the loss exceeding the limit judgment result, the weight coefficient of the loss item is sampled within the range of 1.0-2.0 times to obtain the target weight coefficient;

[0041] The PILO-GAN network is retrained based on the target weight coefficients and steps S2 to S4 are executed iteratively to obtain the target design parameters, thereby obtaining the overall scheme of the target aircraft.

[0042] Further, the process of retraining the PILO-GAN network based on the target weight coefficients and iteratively executing steps S2 to S4 to obtain the target design parameters includes:

[0043] A multi-stage strategy is employed for training to obtain the target design parameters, wherein,

[0044] Phase 1 (0-25% iteration cycle): Only optimize the adversarial loss to enable the generator to learn the basic distribution characteristics of the data;

[0045] Phase 2 (25-50% of the iteration cycle): The adversarial loss weight decays linearly to 0, the constraint loss weight increases linearly to 1, and the generated parameters are guided to meet the empirical criterion constraints.

[0046] Phase 3 (50-75% of the iteration cycle): The physical loss weight is linearly increased to 1, the constraint loss weight is kept at 1, and the multidisciplinary performance of the generated scheme is optimized;

[0047] Phase 4 (75-100% iteration cycle): With the weights of physical loss and constraint loss fixed at 1, the comprehensive optimization and convergence of the generated scheme are achieved, and the target design parameters are obtained.

[0048] Compared with existing technologies, the advantages of this invention are as follows: This invention couples the aerodynamic performance parameters and geometric shape parameters of an aircraft through fluid dynamics principles, ensuring that the generated wingspan, wing area, and lift coefficient are mutually matched; it constrains structural weight parameters with material properties and load-bearing layout through mechanical equilibrium relationships, ensuring that the empty weight and maximum takeoff weight meet strength design criteria; and it dynamically correlates fuel consumption indicators with propulsion efficiency and flight drag through an energy conservation mechanism, achieving coordinated optimization of fuel consumption and the power system. By embedding empirical criteria constraints, it ensures that the generated parameters meet engineering specifications such as component position, geometry, and static stability characteristics. The REINFORCE algorithm establishes a backpropagation channel from performance indicators to design parameters, enabling dynamic balance between physical losses and constraint losses during multi-stage training. Ultimately, it outputs an overall aircraft design that satisfies the physical laws of multiple disciplines such as aerodynamics, structure, and propulsion, and possesses engineering credibility. This significantly shortens the design iteration cycle while improving the feasibility of the design. Attached Figure Description

[0049] Figure 1 A flowchart illustrating the generative design method for overall aircraft schemes based on PILO-GAN provided by this invention;

[0050] Figure 2 This is a mission profile and key parameter diagram of a large transport aircraft generated by the large language model module in an embodiment of the present invention;

[0051] Figure 3 These are three views of a large transport aircraft generated in an embodiment of the present invention;

[0052] Figure 4 This is a profile and key parameter diagram of a medium-range passenger aircraft mission generated by the large language model module in an embodiment of the present invention;

[0053] Figure 5 These are three-view drawings of a medium-range passenger aircraft generated in an embodiment of the present invention;

[0054] Figure 6 This is a profile and key parameter diagram of a business jet mission generated by the large language model module in an embodiment of the present invention;

[0055] Figure 7 The three-view drawing of a business jet generated in the embodiments of the present invention. Detailed Implementation

[0056] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0057] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0058] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0059] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0060] Please see Figure 1 As shown, this invention provides a generative design method for overall aircraft schemes based on PILO-GAN, including:

[0061] Step S1: Obtain aircraft design requirements information and convert its unstructured natural language requirements into structured design parameters;

[0062] Specifically, the natural language conversion layer is built on a large language model and integrates retrieval-enhanced generation (RAG) technology. It calls upon authoritative aerospace knowledge bases containing NASA technical reports, ICAO guidelines, FAA airworthiness certification documents, etc., to ensure the rationality of parameters.

[0063] Specifically, the process involves acquiring aircraft design requirements information by receiving natural language descriptions of the aircraft type and mission requirements input by the user, resulting in unstructured requirement data to be processed. Based on this unstructured requirement data, a large language model guided by the CO-STAR framework prompts is used to parse the core requirements, yielding parsed requirement elements. Based on these parsed requirement elements, a search-enhanced generation technique is used to access an authoritative aerospace knowledge base to generate structured parameters covering basic characteristics and mission profiles. Finally, based on these structured parameters, a three-tiered iterative correction process involving format verification, rationality constraint checks, and logical consistency checks is employed to obtain structured design parameters that conform to engineering standards.

[0064] Step S2: The structured design parameters are used as conditional inputs to construct a PILO-GAN network, which combines random noise to generate initial design parameters for the overall aircraft scheme. The parameters are embedded with aircraft design empirical criteria and driven by multidisciplinary performance evaluation to optimize physics.

[0065] Specifically, step S2 includes the following process:

[0066] Based on the structured design parameters and random noise, the initial design parameters are obtained by constructing a generator using a ResNet architecture and integrating a self-attention mechanism for forward propagation computation.

[0067] Specifically, the process of constructing a generator using the ResNet architecture and integrating a self-attention mechanism for forward propagation computation includes:

[0068] ,

[0069] in, denoted as generator function, outputting initial design parameters, z as random noise, y as conditional label, ⊕ as feature concatenation, FC1 and FC2 as fully connected layers, ResBlocks as residual block sequence, and Attention as self-attention mechanism.

[0070] Specifically, based on structured design parameters and random noise, a generator is constructed using a ResNet architecture and integrating a self-attention mechanism. The generator's input layer receives 64-dimensional random noise and features concatenated with payload weight condition labels to obtain a fused feature vector. Based on the fused feature vector, a first fully connected layer FC1 is used for linear transformation to obtain a first intermediate feature. FC1 maps the concatenated features to the input dimension of a first residual block. Based on the first intermediate feature, deep feature extraction is performed using a residual block sequence ResBlocks, where each residual block contains two convolution operations and skip connections to obtain a second intermediate feature. Based on the second intermediate feature, the correlation weights between features are calculated using the self-attention mechanism layer Attention and weighted fusion is performed to obtain a third intermediate feature that enhances the contextual features. Based on the third intermediate feature, a second fully connected layer FC2 is used for linear transformation to obtain a 50-dimensional output feature, which is then non-linearly activated using the LeakyReLU activation function to obtain the initial overall design parameters of the aircraft, including weight parameters, geometric parameters, and configuration parameters. The negative slope parameter of LeakyReLU is set to 0.01 to avoid gradient vanishing.

[0071] Based on aircraft design experience guidelines, the degree of constraint violation of the design parameters is calculated through a parameter constraint mechanism. Weighted hyperbolic tangent nonlinear scaling and batch averaging constraints are used to obtain compliance verification results.

[0072] Specifically, the process of calculating the degree of constraint violation of the initial design parameters through the parameter constraint mechanism includes:

[0073] ,

[0074] in, ωk is the parameter constraint loss value, m is the number of training batch samples, n is the total number of constraint terms, ωk is the weight coefficient of the kth constraint, sk is the non-linear scaling factor, and Lk(i) is the original loss value of the i-th sample under the kth constraint.

[0075] Specifically, based on aircraft design experience criteria, n constraints are defined for the initial design parameters, including component position, geometry, aerodynamic performance, and static stability. Each constraint is assigned a corresponding weighting coefficient ωk and a nonlinear scaling factor sk, resulting in a constraint configuration parameter set. Based on this constraint configuration parameter set, the original loss value Lk(i) is calculated for the k-th constraint of the i-th sample in the training batch. This original loss value represents the deviation of the design parameters from the allowable range of the experience criteria. Based on the original loss value Lk(i) and the nonlinear scaling factor sk, the loss of each constraint is nonlinearly scaled using the hyperbolic tangent function tanh, mapping it to the interval [-1,1], resulting in the nonlinear scaling loss value tanh(sk·Lk(i)) of the k-th constraint for the i-th sample. Based on the nonlinear scaling loss value and the weighting coefficient ωk, all n constraints for the i-th sample are weighted and summed to obtain the constraint violation degree of the i-th sample. Based on the constraint violation degree of the i-th sample, the arithmetic mean is calculated for all m samples in the training batch to obtain the batch average constraint loss value.

[0076] Based on the requirements of multidisciplinary performance evaluation, a weighted hyperbolic tangent nonlinear optimization of the maximum takeoff weight, empty weight, and fuel consumption weight is applied to the initial design function through a physical optimization mechanism to obtain the physical loss value.

[0077] Specifically, the process of applying a weighted hyperbolic tangent nonlinear optimization of the initial design function using the physical optimization mechanism, based on the maximum takeoff weight, empty weight, and fuel consumption weight, to obtain the physical loss value includes:

[0078] ,

[0079] in, This represents the physical loss value. To optimize the principal loss weights in physics, These represent the losses corresponding to maximum takeoff weight, empty weight, and fuel consumption weight, respectively. This corresponds to the non-linear scaling factor.

[0080] Specifically, based on the requirements of multidisciplinary performance evaluation, the following parameters are set: physical optimization main loss weight coefficient ω1, maximum takeoff weight nonlinear scaling factor s_mtow, empty weight nonlinear scaling factor s_oew, fuel consumption weight nonlinear scaling factor s_fc, and the weight coefficient ωk and nonlinear scaling factor sk of the k-th constraint, to obtain the physical optimization configuration parameter set. Based on the physical optimization configuration parameter set, for the initial design parameters of the i-th sample in the training batch, the maximum takeoff weight loss value L_mtow, empty weight loss value L_oew, fuel consumption weight loss value L_fc, and the original loss value Lk(i) of the k-th constraint are calculated to obtain the original loss set of the sample. Based on the original loss set of the sample, a combined physical loss value is calculated for the i-th sample through weighted summation and hyperbolic tangent nonlinear scaling, to obtain the nonlinear optimization loss value of the i-th sample. Based on the nonlinear optimization loss value of the i-th sample, an arithmetic mean is performed on all m samples in the training batch to obtain the physical loss value.

[0081] Based on the fusion features of the initial design parameters and conditional labels, a discriminator is constructed using a CNN architecture and an integrated self-attention mechanism to perform authenticity scoring, thereby obtaining adversarial training signals to complete the construction of the PILO-GAN network.

[0082] Specifically, based on the initial design parameters and payload weight condition labels, the 50-dimensional design parameter vector and condition labels are fused through feature concatenation to obtain the discriminator input feature vector. Based on the discriminator input feature vector, a CNN architecture is used to construct the discriminator, with convolutional layers extracting local features and pooling layers reducing dimensionality to obtain a convolutional feature map. Based on the convolutional feature map, a self-attention mechanism is integrated to calculate the correlation weights between features and performs weighted fusion to obtain an enhanced feature representation. Based on the enhanced feature representation, a linear transformation is performed through a fully connected layer, and the loss function of WGAN-GP is used to constrain the sample authenticity score, which is then used as an adversarial training signal to complete the construction of the PILO-GAN network.

[0083] Step S3: Perform multidisciplinary physical performance verification on the initial design parameters to obtain quantitative performance indicators;

[0084] Specifically, step S3 includes the following process:

[0085] The initial design parameters were verified using the vortex lattice method to obtain normalized quantitative performance indicators, which include the ratio of fuel consumption weight to maximum takeoff weight, the ratio of empty weight to maximum takeoff weight, and the lift coefficient.

[0086] Specifically, based on the initial design parameters, a multidisciplinary integrated simulation environment encompassing aerodynamics, weight engineering, propulsion systems, and flight mechanics is constructed using the SUAVE tool to obtain a simulation configuration model. Based on the geometric and configuration parameters in the initial design parameters, the lift coefficient C_L, drag coefficient C_D, and pitching moment coefficient C_m are calculated in the simulation configuration model using the vortex lattice method to obtain aerodynamic performance data. Based on the weight and geometric parameters in the initial design parameters, the structural weight, system weight, and payload distribution are calculated using a weight engineering model to obtain the empty weight W_oew. The maximum takeoff weight W_mtow is calculated. Based on the mission profile and propulsion system configuration in the initial design parameters, the fuel consumption weight W_fc is calculated using the propulsion system model to obtain propulsion performance data. Based on the maximum takeoff weight W_mtow, the empty weight W_oew, and the fuel consumption weight W_fc, the ratios of fuel consumption weight to maximum takeoff weight W_fc / W_mtow and the ratios of empty weight to maximum takeoff weight W_oew / W_mtow are calculated through normalization. Combined with the lift coefficient C_L, the normalized quantitative performance index is obtained.

[0087] Step S4: The REINFORCE algorithm is used to construct a gradient propagation path, and the gradient of the quantization performance index is fed back to the generator to realize parameter update and obtain update result;

[0088] Specifically, step S4 includes the following process:

[0089] By generating N independent noise samples and propagating them forward through the generator, the physical performance is evaluated to obtain the quantitative performance index of each sample.

[0090] Specifically, N independent random noise samples are generated in the deep learning framework. Each noise sample is combined with the structured design parameters and then input into the generator for forward propagation calculation to obtain N different initial design parameter samples. For each sample, the physical simulation environment constructed in step S3 is called to perform multidisciplinary performance calculation to obtain the corresponding quantitative performance index value of each sample.

[0091] The sample mean is calculated based on the quantitative performance index as a baseline, and the relative performance of each sample is obtained by calculating the dominance function.

[0092] Specifically, the baseline value is obtained by calculating the arithmetic mean of the quantitative performance indicators of all samples in the batch. The baseline value is then subtracted from the quantitative performance indicator of each sample to obtain the performance advantage value of each sample relative to the average level of the batch.

[0093] Based on the relative performance advantages and disadvantages, the gradient of the generator output with respect to the input noise is calculated, and the REINFORCE algorithm is used to estimate the gradient of the physical performance with respect to the noise.

[0094] Based on the gradient, it is propagated to the generator parameters using the chain rule, and the parameters are updated by combining the gradients of the adversarial loss and the constraint loss to obtain the update result.

[0095] Specifically, the process of propagating the gradient to the generator parameters using the chain rule, and then updating the parameters by combining the gradients of the adversarial loss and the constraint loss to obtain the updated result includes:

[0096] The generator parameter gradient is calculated using the generator Jacobian matrix based on the gradient of the physical properties with respect to noise.

[0097] Specifically, based on the gradient of the physical properties with respect to noise Through the Jacobian matrix of the generator Calculate the gradient of generator parameters The Jacobian matrix is ​​calculated using automatic differentiation techniques.

[0098] The total gradient is obtained by weighted fusion of the generator parameter gradient, the adversarial loss gradient, and the constraint loss gradient.

[0099] Specifically, based on the gradient of the generator parameters Adversarial loss gradient and constrained loss gradient The total gradient is obtained by weighted fusion. The loss weight coefficients α, β, and γ are dynamically adjusted by step S5.

[0100] The magnitude of the total gradient is limited to a preset threshold range by gradient clipping, and the generator parameters are updated using the Adam optimizer to obtain the updated result.

[0101] Specifically, based on the total gradient The gradient magnitude is calculated using the L2 norm. When the magnitude exceeds a preset threshold τ, it is scaled proportionally to make the magnitude equal to τ, resulting in a clipped gradient vector. Based on the clipped gradient vector, the Adam optimizer is used with a learning rate η=0.0001, a first-order momentum coefficient β1=0.5, and a second-order momentum coefficient β2=0.9. Gradient descent is performed on the generator parameter θ to update θ←θ-η·Adam(∇_θL_total), resulting in the update result.

[0102] Step S5: Dynamically adjust the weights of the loss terms in the constraints and physical optimization, and iteratively execute steps S2 to S4 based on the update results until the generated design parameters meet the preset requirements to obtain the target design parameters.

[0103] Specifically, step S5 includes the following process:

[0104] By setting a constraint loss threshold and a physical loss threshold, the update result is monitored for loss value to obtain a loss exceeding the limit judgment result;

[0105] Specifically, constraint loss values ​​are calculated based on the update results and the currently generated design parameters. and physical loss value ,Will With preset constraint loss threshold τ_cons, The loss value is compared with a preset physical loss threshold τ_phy. If any loss value exceeds the corresponding threshold, the result is "needs adjustment"; otherwise, it is "meets requirements". τ_cons is set to 0.1-0.3 based on empirical criteria for constraint strictness, and τ_phy is set to 0.15-0.25 based on performance optimization objectives. The optimal threshold combination is determined from historical training data using a grid search method.

[0106] Based on the loss exceeding the limit judgment result, the weight coefficient of the loss item is sampled within the range of 1.0-2.0 times to obtain the target weight coefficient;

[0107] Specifically, when the loss exceeds the limit and the judgment result is "adjustment is needed", the weight coefficients corresponding to the loss terms that caused the exceedance are identified. The TPE sampler of the Optuna hyperparameter optimization framework is used to perform Bayesian optimization sampling within the range of 1.0-2.0 times the current weight coefficients to obtain the target weight coefficient ωk^new. The constraint loss weight ω_cons and the physical loss weight ω_phy are sampled independently, with a sampling search space of ω_current × [1.0, 2.0]. The sampling objective function is set to the expected loss reduction in the next iteration cycle.

[0108] The PILO-GAN network is retrained based on the target weight coefficients and steps S2 to S4 are executed iteratively to obtain the target design parameters, thereby obtaining the overall scheme of the target aircraft.

[0109] Specifically, the process of retraining the PILO-GAN network based on the target weight coefficients and iteratively executing steps S2 to S4 to obtain the target design parameters includes:

[0110] A multi-stage strategy is employed for training to obtain the target design parameters, wherein,

[0111] Phase 1 (0-25% iteration cycle): Only optimize the adversarial loss to enable the generator to learn the basic distribution characteristics of the data;

[0112] Specifically, based on the target weight coefficients, the adversarial loss weight α1=1.0, the constraint loss weight α2=0, and the physical loss weight α3=0 are set. Only the discriminator loss and the generator adversarial loss are optimized, so that the generator can quickly learn the basic distribution characteristics of the aircraft parameters in the training data. Every 10 training cycles, the generator weight checkpoint is saved once.

[0113] Phase 2 (25-50% of the iteration cycle): The adversarial loss weight decays linearly to 0, the constraint loss weight increases linearly to 1, and the generated parameters are guided to meet the empirical criterion constraints.

[0114] Specifically, the generator weights are loaded based on the checkpoint saved in Phase 1. The adversarial loss weight is set to decay to 0 according to the linear function α1(γ)=1-γ, and the constraint loss weight is linearly increased to 1 according to α2(γ)=γ, where γ=(current cycle-starting cycle of Phase 2) / total number of cycles in Phase 2. The physical loss weight is kept at α3=0, guiding the generator to prioritize satisfying the empirical criterion constraints. The constraint loss value is calculated once every 20 training cycles. If the constraint loss decrease rate is less than 0.5% / cycle, the next stage is entered early.

[0115] Phase 3 (50-75% of the iteration cycle): The physical loss weight is linearly increased to 1, the constraint loss weight is kept at 1, and the multidisciplinary performance of the generated scheme is optimized;

[0116] Specifically, based on the final weight loading generator in Phase 2, the adversarial loss weight α1=0, the constraint loss weight α2=1, and the physical loss weight are linearly increased to 1 by the linear function α3(γ)=γ. At the same time, the constraint loss and physical loss are optimized so that the generated scheme can improve the multidisciplinary performance while meeting the empirical criteria. The physical loss value is calculated once every 30 training cycles. If the physical loss does not decrease for 3 consecutive times, the early stop mechanism is triggered to enter the next phase.

[0117] Phase 4 (75-100% iteration cycle): With the weights of physical loss and constraint loss fixed at 1, the comprehensive optimization and convergence of the generated scheme are achieved, and the target design parameters are obtained.

[0118] Specifically, based on the final weight loading generator in Phase 3, the adversarial loss weight α1=0, the constraint loss weight α2=1, and the physical loss weight α3=1 are fixed. Fine-tuning is performed using constant weights, and the loss value of S51 is monitored every 50 training cycles. When the constraint loss... < 0.2 and physical loss If the condition < 0.25 is met simultaneously, the training is deemed to have converged and the target design parameters are output; if convergence is not achieved after 100% iteration cycle, the historical optimal parameters are output as the target design parameters.

[0119] Example 1: Generation of Overall Scheme for Large Transport Aircraft

[0120] 1. Input natural language requirement: "Large transport aircraft (payload: 120 tons; range: 4260km; can take off and land on a simple runway)".

[0121] 2. Natural Language Transformation Layer Processing: Structured parameters are generated through LLM parsing, including maximum fuselage width of 5.8m, engine nacelle entrance width of 2.2m, payload weight of 120,000kg, maximum load / maximum load capacity of 2.5 / 3.5, subsonic turbofan engine configuration, cargo support system configuration, and a flight mission profile comprising 3 climb phases, 1 cruise phase, and 5 descent phases, such as... Figure 2 As shown.

[0122] 3. PILO-GAN model training: The structured parameters are input into the PILO-GAN main module, and the model is trained for 2000 iterations according to a multi-stage training strategy. The final constraint loss is 3.11 and the physical loss is 1.86.

[0123] 4. Performance Verification and Solution Output: The fuel consumption weight of the generated solution was calculated to be 62355 kg using the SUAVE tool. A 3D model of the aircraft was then generated using the OpenVSP tool, as shown below. Figure 3 As shown; the generated configuration has an empty weight of 140,029 kg, a maximum takeoff weight of 338,239 kg, a wingspan of 57.93 m, and a wing area of ​​497.4 m². 2 With an aspect ratio of 6.7, compared with the key parameters of the real large transport aircraft C-5M, the empty weight is reduced by 18.76% (in line with the weight optimization trend of composite material applications), and the remaining parameters are all within a reasonable engineering range.

[0124] Example 2: Generation of Overall Scheme for Medium-Range Passenger Aircraft

[0125] 1. Input natural language requirement: "Medium-range passenger aircraft (200 passengers; range 5765km)".

[0126] 2. Natural Language Transformation Layer Processing: Generates structured parameters including maximum fuselage width of 3.8m, engine nacelle entrance width of 1.8m, payload weight of 20,000kg, ultimate load / ultimate load of 2.5 / 3.75, subsonic turbofan engine, mid-range flight assistance system configuration, and corresponding flight mission profiles, such as... Figure 4 As shown.

[0127] 3. PILO-GAN model training: After 2000 iterations, the constraint loss was 4.42, the physical loss was 4.66, and the fuel consumption weight was 31175 kg.

[0128] 4. Performance Verification and Solution Output: The generated solution has an empty weight of 34,100 kg, a maximum takeoff weight of 86,198 kg, a wingspan of 36.63 m, and a wing area of ​​134.6 m². 2 With an aspect ratio of 9.96, its parameter distribution is consistent with that of the actual medium-range passenger aircraft B737-800, and it can meet the specified range requirements under full load. The generated scheme is as follows... Figure 5 As shown.

[0129] Example 3: Generation of Overall Business Jet Solution

[0130] 1. Input natural language requirement: "Business jet (10 passengers; range 6667km)".

[0131] 2. Natural Language Transformation Layer Processing: Generates structured parameters including maximum fuselage width of 2.8m, engine nacelle entrance width of 0.9m, payload weight of 1800kg, maximum load / maximum load capacity of 2.5 / 3.5, subsonic turbofan engine, configuration of business jet-specific auxiliary systems, and corresponding flight mission profiles, such as... Figure 6 As shown.

[0132] 3. PILO-GAN model training: After 2000 iterations, the constraint loss was 4.53, the physical loss was 4.91, and the fuel consumption weight was 6544 kg.

[0133] 4. Performance Verification and Solution Output: The generated solution has an empty weight of 8823 kg, a maximum takeoff weight of 19125 kg, a wingspan of 21.38 m, and a wing area of ​​48.8 m². 2 With an aspect ratio of 9.3, it closely matches the key parameters of the actual business jet Gulfstream G280, meeting the design requirements of long range and high comfort. The generated solution is as follows... Figure 7 As shown.

[0134] Specifically, this invention couples the aerodynamic performance parameters and geometric parameters of an aircraft through fluid dynamics principles, ensuring that the generated wingspan, wing area, and lift coefficient are mutually matched. It also constrains structural weight parameters with material properties and load-bearing layout through mechanical equilibrium relationships, ensuring that the empty weight and maximum takeoff weight meet strength design criteria. Furthermore, it dynamically correlates fuel consumption indicators with propulsion efficiency and flight drag through an energy conservation mechanism, achieving coordinated optimization of fuel consumption and the power system. By embedding empirical criteria constraints, it ensures that the generated parameters meet engineering specifications such as component position, geometry, and static stability characteristics. The REINFORCE algorithm establishes a backpropagation channel from performance indicators to design parameters, enabling dynamic balance between physical losses and constraint losses during multi-stage training. Ultimately, it outputs an overall aircraft design that satisfies the physical laws of aerodynamics, structure, propulsion, and other disciplines, possessing engineering credibility. This significantly shortens the design iteration cycle while improving the feasibility of the design.

[0135] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0136] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A generative design method for overall aircraft schemes based on PILO-GAN, characterized in that, include: Step S1: Obtain aircraft design requirements information and convert its unstructured natural language requirements into structured design parameters; Step S2: The structured design parameters are used as conditional inputs to construct a PILO-GAN network, which combines random noise to generate initial design parameters for the overall aircraft scheme. The parameters are embedded with aircraft design empirical criteria and driven by multidisciplinary performance evaluation to optimize physics. Step S3: Perform multidisciplinary physical performance verification on the initial design parameters to obtain quantitative performance indicators; Step S4: The REINFORCE algorithm is used to construct a gradient propagation path, and the gradient of the quantization performance index is fed back to the generator to realize parameter update and obtain update result; Step S5: Dynamically adjust the weights of the loss terms in the constraints and physical optimization, and iteratively execute steps S2 to S4 based on the update results until the generated design parameters meet the preset requirements to obtain the target design parameters.

2. The generative design method for overall aircraft scheme based on PILO-GAN according to claim 1, characterized in that, The process of step S2 includes: Based on the structured design parameters and random noise, the initial design parameters are obtained by constructing a generator using a ResNet architecture and integrating a self-attention mechanism for forward propagation computation. Based on aircraft design experience guidelines, the degree of constraint violation of the design parameters is calculated through a parameter constraint mechanism. Weighted hyperbolic tangent nonlinear scaling and batch averaging constraints are used to obtain compliance verification results. Based on the requirements of multidisciplinary performance evaluation, a weighted hyperbolic tangent nonlinear optimization of the maximum takeoff weight, empty weight, and fuel consumption weight is applied to the initial design function through a physical optimization mechanism to obtain the physical loss value. Based on the fusion features of the initial design parameters and conditional labels, a discriminator is constructed using a CNN architecture and an integrated self-attention mechanism to perform authenticity scoring, thereby obtaining adversarial training signals to complete the construction of the PILO-GAN network.

3. The generative design method for overall aircraft schemes based on PILO-GAN according to claim 2, characterized in that, The process of constructing a generator using the ResNet architecture and integrating a self-attention mechanism for forward propagation computation includes: ， in, denoted as generator function, outputting initial design parameters, z as random noise, y as conditional label, ⊕ as feature concatenation, FC1 and FC2 as fully connected layers, ResBlocks as residual block sequence, and Attention as self-attention mechanism.

4. The generative design method for overall aircraft schemes based on PILO-GAN according to claim 3, characterized in that, The process of calculating the degree of constraint violation of the initial design parameters through the parameter constraint mechanism includes: ， in, ωk is the parameter constraint loss value, m is the number of training batch samples, n is the total number of constraint terms, ωk is the weight coefficient of the kth constraint, sk is the non-linear scaling factor, and Lk(i) is the original loss value of the i-th sample under the kth constraint.

5. The generative design method for overall aircraft scheme based on PILO-GAN according to claim 4, characterized in that, The process of applying a weighted hyperbolic tangent nonlinear optimization of the initial design function using a physical optimization mechanism, based on the maximum takeoff weight, empty weight, and fuel consumption weight, to obtain the physical loss value includes: ， in, This represents the physical loss value. To optimize the principal loss weights in physics, These represent the losses corresponding to maximum takeoff weight, empty weight, and fuel consumption weight, respectively. This corresponds to the non-linear scaling factor.

6. The generative design method for overall aircraft scheme based on PILO-GAN according to claim 5, characterized in that, The process of step S3 includes: The initial design parameters were verified using the vortex lattice method to obtain normalized quantitative performance indicators, which include the ratio of fuel consumption weight to maximum takeoff weight, the ratio of empty weight to maximum takeoff weight, and the lift coefficient.

7. The generative design method for overall aircraft schemes based on PILO-GAN according to claim 6, characterized in that, The process of step S4 includes: By generating N independent noise samples and propagating them forward through the generator, the physical performance is evaluated to obtain the quantitative performance index of each sample. The sample mean is calculated based on the quantitative performance index as a baseline, and the relative performance of each sample is obtained by calculating the dominance function. Based on the relative performance advantages and disadvantages, the gradient of the generator output with respect to the input noise is calculated, and the REINFORCE algorithm is used to estimate the gradient of the physical performance with respect to the noise. Based on the gradient, it is propagated to the generator parameters using the chain rule, and the parameters are updated by combining the gradients of the adversarial loss and the constraint loss to obtain the update result.

8. The generative design method for overall aircraft schemes based on PILO-GAN according to claim 7, characterized in that, The process of propagating the gradient to the generator parameters using the chain rule, and updating the parameters by combining the gradients of the adversarial loss and the constraint loss to obtain the updated result includes: The generator parameter gradient is calculated using the generator Jacobian matrix based on the gradient of the physical properties with respect to noise. The total gradient is obtained by weighted fusion of the generator parameter gradient, the adversarial loss gradient, and the constraint loss gradient. The magnitude of the total gradient is limited to a preset threshold range by gradient clipping, and the generator parameters are updated using the Adam optimizer to obtain the updated result.

9. The generative design method for overall aircraft schemes based on PILO-GAN according to claim 8, characterized in that, The process of step S5 includes: By setting a constraint loss threshold and a physical loss threshold, the update result is monitored for loss value to obtain a loss exceeding the limit judgment result; Based on the loss exceeding the limit judgment result, the weight coefficient of the loss item is sampled within the range of 1.0-2.0 times to obtain the target weight coefficient; The PILO-GAN network is retrained based on the target weight coefficients and steps S2 to S4 are executed iteratively to obtain the target design parameters, thereby obtaining the overall scheme of the target aircraft.

10. The generative design method for overall aircraft schemes based on PILO-GAN according to claim 9, characterized in that, The process of retraining the PILO-GAN network based on the target weight coefficients and iteratively executing steps S2 to S4 to obtain the target design parameters includes: A multi-stage strategy is employed for training to obtain the target design parameters, wherein, Phase 1 (0-25% iteration cycle): Only optimize the adversarial loss to enable the generator to learn the basic distribution characteristics of the data; Phase 2 (25-50% of the iteration cycle): The adversarial loss weight decays linearly to 0, the constraint loss weight increases linearly to 1, and the generated parameters are guided to meet the empirical criterion constraints. Phase 3 (50-75% of the iteration cycle): The physical loss weight is linearly increased to 1, the constraint loss weight is kept at 1, and the multidisciplinary performance of the generated scheme is optimized; Phase 4 (75-100% iteration cycle): With the weights of physical loss and constraint loss fixed at 1, the comprehensive optimization and convergence of the generated scheme are achieved, and the target design parameters are obtained.

Images