A mach number adaptive multi-expert optimization design method for a wide-speed-range aircraft component

Abstract

The application relates to the technical field of intelligent optimization design of aircraft structures, and provides a Mach number adaptive multi-expert optimization design method for aircraft components in a wide speed range, which comprises the following steps: constructing a wide speed range sample data set; encoding the Mach number based on a preset Mach number center point and a radial basis function, and mapping the Mach number into a Mach number coding vector representing different speed range characteristics; constructing enhanced input features; constructing a Mach number adaptive multi-expert surrogate model, and training the model through the wide speed range sample data set; constructing a wide speed range normalized integral performance evaluation index; and obtaining a Pareto optimal solution set by using a multi-objective optimization algorithm according to the optimization requirements of the aircraft components. The wide speed range normalized integral performance evaluation index constructed by the application breaks through the limitation that the optimization target is only limited to a single Mach number or a small number of typical working conditions, and is more in line with the actual working requirements of the aircraft components in the wide speed range.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent optimization design of aircraft structures, and in particular to a Mach number adaptive multi-expert optimization design method for wide-speed-range aircraft components. Background Technology

[0002] In actual service, critical aerodynamic components of aircraft often do not operate at a single Mach number or under a single design condition, but rather need to operate continuously or intermittently across subsonic, transonic, supersonic, and even higher speed ranges. For example, the control wings of reusable launch vehicles, the airfoils of high-speed aircraft, and missile wings all need to maintain stable aerodynamic performance and reliable structural load-bearing capacity across a wide speed range. Designing only for a typical Mach number or a few discrete conditions will result in configurations that cannot meet the performance and safety requirements of the aircraft across its entire mission profile. Therefore, establishing surrogate models and optimization design methods for aircraft components across a wide speed range is a key technical requirement for meeting the needs of rapid performance evaluation, aerodynamic / structural co-optimization, and reliable design of aircraft components across the entire mission profile.

[0003] To reduce the computational costs of high-precision numerical simulations and experiments and improve the efficiency of aircraft component optimization design, existing technologies typically employ a combination of surrogate models and optimization algorithms. However, existing publicly available solutions still have significant shortcomings in meeting the requirements for full-mission profile and multi-performance collaborative optimization of aircraft components. On the one hand, existing surrogate models are mostly established for single typical operating conditions, local velocity domains, or a few discrete Mach number points. While such methods can improve prediction efficiency under specific operating conditions, they are difficult to characterize the continuous variation of aerodynamic and structural responses across the entire Mach number range. For example, the surrogate model constructed by Chinese patent CN117634042A is mainly aimed at aerodynamic shape optimization under transonic conditions; although Chinese patent CN113609596B considers the influence of the incoming Mach number on aerodynamic characteristics, it essentially involves interval modeling based on Mach number and angle of attack ranges. The lack of continuous adaptive correlation between different interval models makes it difficult to avoid the problem of insufficient prediction continuity at interval boundaries. On the other hand, while existing wide-speed-range optimization designs have attempted to consider performance requirements across different speed ranges, their objective functions mostly involve directly selecting performance indicators at a few Mach number conditions for parallel optimization or weighted combination. For example, Chinese patents CN115489751B and CN109484623B both select individual Mach number points in different speed ranges and use the performance indicators at those selected Mach number points to characterize wide-speed-range performance. While this method can address performance requirements at a few Mach number points, it lacks a comprehensive evaluation across the entire Mach number range, making it difficult to fully characterize performance changes in non-design Mach number regions, thus affecting the overall performance of the optimized configuration across the entire mission profile. Furthermore, existing wide-speed-range optimization schemes often prioritize aerodynamic indicators such as lift, drag, and lift-to-drag ratio as their main optimization objectives, neglecting structural response constraints and making it difficult to ensure that the optimized configuration simultaneously meets aerodynamic performance and structural safety requirements across the entire service speed range.

[0004] To overcome the aforementioned technical bottlenecks, there is an urgent need in this field to develop a Mach number adaptive multi-expert optimization design method for wide-speed-range aircraft components, in order to support the intelligent, reliable, and efficient optimization design of aircraft components across the entire mission profile. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a Mach number adaptive multi-expert optimization design method for wide-speed-range aircraft components, the optimization design method comprising: Step 1: Construct a wide-speed-range sample dataset using the operating Mach number range, geometric design variables, and aerodynamic and structural responses of the wide-speed-range aircraft components to be optimized. Step 2: Encode the Mach number based on the preset Mach number center point and radial basis function (RBF), and map it into a Mach number encoding vector that represents the characteristics of different velocity domains; Step 3: Construct enhanced input features based on geometric design variables, Mach number encoding vectors, interaction terms between Mach number and each geometric design variable, and interaction terms among geometric design variables. Step 4: Construct a Mach number adaptive multi-expert agent model based on enhanced input features, and train it using a wide-speed-domain sample dataset; Step 5: Construct a wide-range normalized integral performance evaluation index based on the trained Mach number adaptive multi-expert agent model; Step 6: Based on the optimization requirements of the aircraft components, a multi-objective optimization algorithm is used to obtain the Pareto optimal solution set.

[0006] Furthermore, in step one, the Mach number range covers at least two of the subsonic, transonic, supersonic, or higher speed ranges.

[0007] Furthermore, the aircraft components include airfoils, wings, control surfaces, tail fins, missile wings, grid rudders, air intakes, or partial shapes of high-speed aircraft.

[0008] The aerodynamic response parameters include at least one of the following: lift coefficient, drag coefficient, moment coefficient, lift-to-drag ratio, or other response parameters used to characterize aerodynamic performance. The structural response parameters include at least one of the following: maximum displacement, maximum stress, maximum temperature, natural frequency, buckling load, or other response parameters used to characterize structural integrity.

[0009] Furthermore, the specific steps of step one are as follows: The operating Mach number range, geometric design variables, and aerodynamic and structural responses of the wide-speed-range aircraft components to be optimized are determined. Within the range of Mach number and design variable values, sample points are generated using Latin hypercube, orthogonal, uniform, or adaptive sampling methods. Experiments or numerical simulations are conducted on each sample point to obtain the mapping relationship with Mach number and geometric design variables as inputs and aerodynamic and structural responses as outputs, forming a complete wide-speed-range sample dataset. The wide-speed-range sample dataset is divided into training, validation, and test sets. The Mach number is retained in its original value and directly input into the surrogate model. Geometric design variables and aerodynamic and structural responses are normalized based on the training set. The validation and test sets only reuse the normalized parameters fitted from the training set for transformation, avoiding data leakage issues.

[0010] Furthermore, the specific steps in step two are as follows: For any sample in the wide-speed-domain sample dataset, let its Mach number be... Preset based on the Mach number operating conditions experienced by the aircraft components One Mach number center point used for radial basis encoding: The width of the corresponding radial basis function is determined based on the distance between the centers of adjacent Mach numbers. Then the first Radial basis function encoded components for: In the formula, For the first The center point of Mach number, For the first The width of the radial basis function corresponding to each Mach number center point; This yields the Mach number encoding vector corresponding to the sample. : By performing the above encoding on all samples in the sample dataset, the Mach number encoding vector corresponding to each sample can be obtained.

[0011] Furthermore, for any sample in the wide-range sample dataset, the enhanced input feature vector in step three is represented as: In the formula, This represents the geometric design variables corresponding to this sample. This represents the interaction terms between the Mach number and the various geometric design variables. This represents the interaction terms between different geometric design variables. The number of geometric design variables is specified. For each sample in the sample dataset, enhanced input features are constructed as described above and used as input to the Mach number adaptive multi-expert agent model.

[0012] Furthermore, the Mach number adaptive multi-expert agent model described in step four includes: Multiple expert subnetworks are configured in parallel to learn features under different speed states or different response patterns. Each expert subnetwork takes enhanced input features as input and corresponding expert features as output. Each expert subnetwork adopts a multilayer perceptron, convolutional neural network, residual network, or other neural network structure or combination thereof suitable for learning data-driven parameter mapping relationships. Different expert subnetworks have independent network parameters to extract aerodynamic / structural response features of aircraft components under wide speed domain conditions from different perspectives. During training, each expert subnetwork does not pre-fix a specific Mach number interval, but automatically learns the response feature expression under different speed states based on Mach number gating weights, feature modulation parameters, and backpropagation results of prediction errors. This achieves adaptive division of labor for features in different speed domains, avoiding reliance on manually dividing Mach number intervals to build independent surrogate models.

[0013] Mach number gating networks employ fully connected networks to encode Mach number vectors. As input, the system adaptively outputs the expert weights and expert feature modulation parameters of each expert subnetwork at different Mach numbers. The output expert weights characterize the contribution of different expert subnetworks under the current Mach number condition, and the expert feature modulation parameters are used to perform dimensionality-level correction on the output features of the expert subnetworks. For any Mach number... The weights of each expert subnetwork output by the Mach number gating network satisfy the non-negativity constraint and the sum of the weights is 1. Through this Mach number gating network, the Mach number adaptive multi-expert surrogate model can adaptively adjust the contribution ratio and feature expression mode of different expert subnetworks according to the change of the input Mach number, thereby enhancing the surrogate model's adaptability to the response law of different velocity domains.

[0014] The expert feature modulation and fusion module modulates the expert features output by each expert subnetwork using the expert feature modulation parameters output by the Mach number-gated network. It then performs weighted fusion of the modulated expert features according to expert weights to obtain Mach number-adaptive fused features. The expert feature modulation parameters amplify, weaken, or shift different dimensions of each expert feature, thereby improving the feature adaptability under different Mach number conditions. Through the above expert feature modulation and weighted fusion methods, not only can the overall contribution of different expert subnetworks be controlled, but also the local dimensions of the expert output features can be adaptively corrected, thereby improving the surrogate model's expressive power and prediction stability in different velocity domains.

[0015] A shared backbone network is used to input Mach number adaptive fusion features into the shared backbone network and further extract shared features for aerodynamic and structural response prediction. The multi-task prediction module for aerodynamic and structural responses constructs aerodynamic prediction branches and structural prediction branches based on shared features. Each branch consists of an independent task-specific sub-network and an output head, enabling aerodynamic and structural responses to learn their respective output mapping rules based on shared features. A Cross-Stitch soft-sharing unit is set between the intermediate feature layers of the aerodynamic and structural prediction branches, and the soft-sharing processed aerodynamic and structural intermediate features are input to the corresponding output heads to obtain the predicted aerodynamic and structural response parameters. The Cross-Stitch soft-shared unit linearly combines intermediate features from different task branches using trainable cross-shared parameters to adaptively adjust the sharing ratio between aerodynamic and structural task features. It also automatically updates its parameters through error backpropagation, thereby improving the aerodynamic / structural multi-task prediction capability.

[0016] Furthermore, in step four, during the training of the Mach number adaptive multi-expert surrogate model, the Mach number is directly input with its original value retained, and the geometric design variables and aerodynamic and structural responses are normalized based on the training set. The surrogate model is trained using a wide-velocity sample dataset, and a loss function is constructed based on the neural network predictions and the actual response values ​​in the sample data. This allows the parameters of the expert sub-network, Mach number gating network module, expert feature modulation and fusion module, aerodynamic and structural branches, and Cross-Stitch soft shared units to be updated synchronously through error backpropagation. Training stops when the validation set loss meets a preset stopping condition or the number of training rounds reaches a preset upper limit, resulting in the trained Mach number adaptive multi-expert surrogate model. This model can quickly output the aerodynamic and structural response parameters corresponding to the aircraft component after inputting any given Mach number and geometric design variables.

[0017] Furthermore, in step five, based on the trained Mach number adaptive multi-expert surrogate model, the performance response of the aircraft components within a wide Mach number range is treated as an approximately continuous function varying with the Mach number, thus constructing a wide-speed-domain normalized integral performance evaluation index. In actual calculations, by selecting discrete Mach number sampling points and employing the trapezoidal integral approximation method, the prediction results of the surrogate model at multiple Mach number sampling points are transformed into an integral evaluation index characterizing the overall performance across the entire wide speed range.

[0018] Specifically, for any response parameter output by the Mach number adaptive multi-expert agent model... Its wide-range normalized integral performance evaluation index is defined as: In the formula, Indicates the first Normalized integral performance evaluation index of each response parameter over a wide speed range; For the geometric design variables of aircraft components; It is the Mach number; For the wide speed range of Mach numbers; This indicates that the Mach number adaptive multi-expert agent model is in Mach number... and geometric design variables The predicted first One response parameter; The performance evaluation index of the wide-range normalized integral is approximated by the trapezoidal discrete integral, and is calculated by selecting within the Mach number range of the wide-range domain. Mach number sampling points Then we can get: In the formula, This is the index of the Mach number sampling point; Indicates the first Mach number sampling points, Indicates the first Mach number sampling points; equal ; equal .

[0019] Furthermore, for the restrictive response parameters, a wide-speed-domain performance constraint index is constructed simultaneously, which requires that the corresponding response values ​​of the candidate design variables at all Mach number sampling points meet the preset constraint conditions, thereby ensuring that the optimized configuration meets the performance constraint requirements throughout the entire wide-speed-domain.

[0020] Furthermore, in step six, based on the specific optimization requirements of the aircraft components, a multi-objective optimization model is constructed using the aforementioned surrogate model and wide-speed-domain normalized integral performance evaluation indicators or other requirements. The optimization objectives and constraints can be combined and set according to requirements such as aerodynamic performance improvement, structural lightweighting, aerodynamic efficiency maintenance, aerodynamic performance constraints, and structural safety constraints. Subsequently, NSGA-II, NSGA-III, multi-objective particle swarm optimization, multi-objective Bayesian optimization, or other multi-objective optimization algorithms are used to solve the problem. During the optimization process, the multi-objective optimization algorithm iteratively updates the candidate geometric design variables and calls the surrogate model to quickly predict the performance response at multiple Mach number sampling points. It then calculates the wide-speed-domain normalized integral performance evaluation indicators and constraint satisfaction, ultimately obtaining the Pareto optimal solution set that satisfies the constraints. The comprehensive optimal solution can be further selected based on engineering requirements.

[0021] The present invention has the following beneficial effects: (1) The present invention encodes the Mach number based on a preset Mach number center point and radial basis function (RBF), maps it into a Mach number encoding vector that represents the characteristics of different velocity ranges, and constructs enhanced input features based on geometric design variables, Mach number encoding vectors, interaction terms between Mach number and each geometric design variable, and interaction terms between geometric design variables, thereby enhancing the surrogate model’s ability to perceive the differences in aerodynamic / structural response in different Mach number ranges and improving the surrogate model’s ability to express wide velocity range nonlinear response.

[0022] (2) This invention constructs a Mach number adaptive multi-expert agent model, which adaptively adjusts the contribution weights of different expert networks and automatically integrates expert information through the Mach number gating mechanism. This avoids the prediction discontinuity problem at the velocity interval boundary of the traditional segmented agent model, which is conducive to improving the prediction accuracy and generalization ability in the wide velocity domain.

[0023] (3) The wide-speed-range normalized integral performance evaluation index constructed by the present invention breaks through the limitation that the optimization target is limited to a single Mach number or a few typical working conditions, and is more in line with the actual working requirements of wide-speed-range aircraft components.

[0024] (4) The present invention embeds the Mach number adaptive multi-expert agent model into a multi-objective optimization algorithm, which can quickly evaluate a large number of candidate design schemes, significantly reduce computational costs, and improve the efficiency of wide-speed-range aircraft component optimization design.

[0025] (5) This invention is applicable to various aircraft components that have experienced wide Mach number service environments, such as wing surfaces, control surfaces, tail fins, missile wings, grid fins, air intakes and local shapes of high-speed aircraft, and has strong engineering applicability. Attached Figure Description

[0026] Figure 1 This is an optimized design flowchart of an embodiment.

[0027] Figure 2 This is a structural diagram of the Mach number adaptive multi-expert agent model in an embodiment.

[0028] Figure 3 This is a graph showing the changes in the weights of each expert as a function of the Mach number, output by the Mach number-gated network in this embodiment.

[0029] Figure 4 This is a schematic diagram illustrating the construction of the wide-range normalized integral performance evaluation index in this embodiment.

[0030] Figure 5 This is a diagram showing the Pareto optimization results of an example. Detailed Implementation

[0031] The technical solution of the present invention will be further described in detail below with reference to specific embodiments. However, these embodiments are not intended to limit the present invention. Any similar structures and similar variations of the present invention should be included in the protection scope of the present invention. The commas in the present invention all indicate the relationship between and. The English letters in the present invention are case-sensitive.

[0032] like Figure 1 As shown, this embodiment uses the reusable launch vehicle return segment grid fin as an example to illustrate the present invention, and its optimized design method is as follows: Step 1: Construct a wide-speed-range sample dataset using the operating Mach number range, geometric design variables, and aerodynamic and structural responses of the wide-speed-range aircraft components to be optimized. In this embodiment, the grid rudder geometry design variables are: ,in , , , , These represent the grid width, chord length, inner grid thickness, outer grid thickness, and local sweep angle of the grid fin; the wide-range Mach number range of the grid fin is defined as follows: This range covers subsonic, transonic, and supersonic speeds; the lift coefficient is selected. and drag coefficient As an aerodynamic response parameter, maximum displacement and maximum stress As a structural response parameter.

[0033] Within the Mach number range and geometric design variable range, 500 sample points were generated using the Latin hypercube sampling method. For each sample point, a corresponding grid rudder geometry model was generated through parametric modeling, and aerodynamic and structural response parameters were obtained using fluid-thermal-structure interaction simulation to obtain the input. and output The mapping relationship between them constitutes a complete wide-range sample dataset. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio. The Mach number is retained and directly input into the model. Geometric design variables and output responses are normalized based on the training set. The validation and test sets only reuse the normalized parameters fitted from the training set for transformation, thus avoiding data leakage issues.

[0034] Step 2: Encode the Mach number based on the preset Mach number center point and radial basis function (RBF), and map it into a Mach number encoding vector representing different velocity domain characteristics; enable the neural network to perceive the velocity domain state of the current sample, which is used for the subsequent Mach number gating network to generate expert weights and feature modulation parameters.

[0035] For 0.5 For a Mach number range of 3.0, eight Mach number center points are set for radial basis encoding, as follows: The width of the corresponding radial basis function is determined based on the distance between the centers of adjacent Mach numbers: Then the first k Radial basis function encoded components for: This yields the Mach number encoding vector corresponding to the sample. : In this embodiment, if the input Mach number of a certain sample is Based on the Mach number center point, radial basis function width, and radial basis function coding components mentioned above, the Mach number coding vector corresponding to this sample can be approximately: The above results show that when the input Mach number is 1.2, the radial basis coding component corresponding to the center point 1.2 is the largest, and the coding components corresponding to the adjacent Mach number centers also have a certain response, thus reflecting the continuous correlation between the Mach number sample and the adjacent velocity state.

[0036] Step 3: Construct enhanced input features based on geometric design variables, Mach number encoding vectors, interaction terms between Mach number and each geometric design variable, and interaction terms among geometric design variables. For any sample in the wide-speed-domain sample dataset, it is represented as: Step 4: Construct a Mach number adaptive multi-expert agent model based on enhanced input features, and train it using a wide-speed-domain sample dataset; like Figure 2 As shown, the Mach number adaptive multi-expert agent model includes three expert sub-networks, a Mach number gating network, an expert feature modulation and fusion module, a shared backbone network, and an aerodynamic / structural multi-task prediction module.

[0037] First, three parallel expert subnetworks are constructed using Residual Multilayer Perceptron (ResMLP), each receiving enhanced input features. and output the corresponding expert features. , ,in Indicates the first A network of experts.

[0038] Secondly, the Mach number encoding vector Input the Mach number-gated network, and output the weights of different expert subnetworks from the Mach number-gated network. and characteristic modulation parameters, where, , and These represent the weights of the three expert subnetworks; in this embodiment, the curves showing the variation of each expert weight with Mach number output by the Mach number-gated network are shown below. Figure 3 As shown.

[0039] Then, an expert feature modulation and fusion module is constructed. Based on the feature modulation parameters output by the Mach number-gated network, the expert features output by each expert subnetwork are modulated and processed, and then weighted and fused according to the expert weights to obtain the Mach number adaptive fusion features. : In the formula, A value of 3 indicates an expert subnetwork; For the first The weights of each expert subnetwork at the current Mach number; The first after feature modulation A characteristic of an expert.

[0040] Then, the Mach number adaptively fuses the features. Input a shared backbone network to extract shared features; Finally, based on the fused shared features, aerodynamic prediction branches and structural prediction branches are constructed separately. The aerodynamic prediction branch is used to output the grid rudder aerodynamic response parameters. The structural prediction branch is used to output the grid rudder structural response parameters. To enhance the feature interaction capabilities between aerodynamic and structural prediction tasks, a Cross-Stitch soft-shared unit is set between the intermediate feature layers of these two branches. The features after passing through this unit can be represented as: in, and These represent the initial intermediate features of the aerodynamic prediction branch and the structural prediction branch, respectively; and These represent the aerodynamic and structural branching characteristics after soft sharing; , , and This is a Cross-Stitch soft-sharing parameter used to adjust the sharing ratio of aerodynamic and structural task characteristics and the cross-characteristics between them. The characteristics after soft-sharing... and By inputting the corresponding pneumatic output head and structural output head respectively, any Mach number can be obtained. and geometric design variables Predicted results of lift coefficient, drag coefficient, maximum displacement and maximum stress under the following conditions. , , , .

[0041] A Mach number adaptive multi-expert agent model is trained based on a wide-speed-domain sample dataset. During training, the Mach number and geometric design variables from the training set are input into the Mach number adaptive multi-expert agent model to obtain the predicted output. and compared with the actual response values ​​obtained from the simulation. A comparison was conducted to construct a multi-output mean squared error loss function. Using the error backpropagation algorithm, the parameters of the Mach number gating network, expert sub-network, feature modulation and fusion module, and aerodynamic / structural prediction branch were updated synchronously. After training, the model's prediction performance was evaluated using validation and test sets, resulting in a Mach number adaptive multi-expert surrogate model that can be used for subsequent optimization design.

[0042] Step 5: Construct a wide-range normalized integral performance evaluation index based on the trained Mach number adaptive multi-expert agent model; like Figure 4 As shown, selecting within the wide speed range of Mach numbers A number of Mach number sampling points are used for evaluation of wide-range normalized integral performance. In this embodiment, [number of sampling points are used]. ,Right now: The Mach number sampling points here are independent of the Mach number center point of the radial basis function. They can be set according to the prediction accuracy, integration accuracy and computational cost. In this embodiment, the same value is used for both, which is only a preferred implementation method.

[0043] Based on the functional characteristics of grid fins, which are mainly used for attitude control and aerodynamic regulation during the aircraft's return process, this embodiment selects the lift coefficient as the main performance evaluation index over a wide speed range. For any candidate geometric design variable... The trained Mach number adaptive multi-expert proxy model was invoked at each of the eight Mach number sampling points to obtain the predicted lift coefficients at each Mach number. , ; And a trapezoidal integral approximation is used to construct the wide-velocity normalized integral lift coefficient: in, Used to characterize the grid rudder at 0.5 The overall lift level within the Mach 3.0 range. Simultaneously, the drag coefficient is used as a wide-speed-range constraint, requiring that the drag coefficient of the candidate design variables at all eight Mach number sampling points does not exceed the drag coefficient of the initial structure. .

[0044] Step 6: Based on the optimization requirements of the aircraft components, a multi-objective optimization algorithm is used to obtain the Pareto optimal solution set; Based on the lightweight design and lift coefficient improvement requirements, and considering structural integrity requirements, the maximum Mach number operating condition is... Using the maximum stress and maximum displacement as representative constraints, the following bi-objective optimization model is constructed: The constraints are as follows: in, The volume of the grid rudder structure; , and These represent the drag coefficient, maximum displacement, and maximum stress of the initial grid fin structure at the corresponding Mach number.

[0045] The non-dominated sorting genetic algorithm NSGA-II was used to solve the above multi-objective optimization model to obtain the Pareto optimal solution set that satisfies the constraints, such as... Figure 5 As shown; subsequently, based on the engineering design requirements, candidate solutions that meet the requirements of increasing lift coefficient and reducing structural weight are selected from the Pareto optimal solution set, and the final engineering optimal solution is determined by using the maximization of lift-to-drag ratio as a comprehensive selection criterion.

[0046] This invention encodes the Mach number based on a preset Mach number center point and radial basis functions (RBF), mapping it to Mach number encoding vectors representing different velocity range characteristics. Enhanced input features are constructed based on geometric design variables, the Mach number encoding vectors, interaction terms between the Mach number and various geometric design variables, and interaction terms among the geometric design variables themselves. This enhances the surrogate model's ability to perceive differences in aerodynamic / structural responses across different Mach number ranges and improves its ability to express wide-velocity-range nonlinear responses. By constructing a Mach number-adaptive multi-expert surrogate model, the contribution weights of different expert networks are adaptively adjusted through a Mach number gating mechanism, and expert information is automatically fused. This avoids the prediction discontinuity problem at velocity range boundaries in traditional piecewise surrogate models, thus improving wide-velocity-range prediction accuracy and generalization ability. This invention is applicable to various aircraft components that have experienced wide Mach number service environments, such as wings, control surfaces, tail fins, missile wings, grid fins, air intakes, and local shapes of high-speed aircraft, demonstrating strong engineering applicability.

[0047] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

Claims

1. A Mach number adaptive multi-expert optimization design method for wide-speed-range aircraft components, characterized in that, The optimization design method includes: Step 1: Construct a wide-speed-range sample dataset using the operating Mach number range, geometric design variables, and aerodynamic and structural responses of the wide-speed-range aircraft components to be optimized. Step 2: Encode the Mach number based on the preset Mach number center point and radial basis function, and map it into a Mach number encoding vector that represents the characteristics of different velocity domains; Step 3: Construct enhanced input features based on geometric design variables, Mach number encoding vectors, interaction terms between Mach number and each geometric design variable, and interaction terms among geometric design variables. Step 4: Construct a Mach number adaptive multi-expert agent model based on enhanced input features, and train it using a wide-speed-domain sample dataset; Step 5: Construct a wide-range normalized integral performance evaluation index based on the trained Mach number adaptive multi-expert agent model; For any response parameter output by the Mach number adaptive multi-expert agent model Its wide-range normalized integral performance evaluation index is defined as: In the formula, Indicates the first Normalized integral performance evaluation index of each response parameter over a wide speed range; For the geometric design variables of aircraft components; It is the Mach number; For the wide speed range of Mach numbers; This indicates that the Mach number adaptive multi-expert agent model is in Mach number... and geometric design variables The predicted first One response parameter; The performance evaluation index of the wide-range normalized integral is approximated by the trapezoidal discrete integral, and is calculated by selecting within the Mach number range of the wide-range domain. Mach number sampling points Then we can get: In the formula, This is the index of the Mach number sampling point; Indicates the first Mach number sampling points, Indicates the first Mach number sampling points; equal ; equal ; Step 6: Based on the optimization requirements of the aircraft components, a multi-objective optimization algorithm is used to obtain the Pareto optimal solution set.

2. The Mach number adaptive multi-expert optimization design method for a wide-speed-range aircraft component according to claim 1, characterized in that, In step one, the Mach number range covers at least two of the subsonic, transonic, supersonic, or higher speed ranges.

3. The Mach number adaptive multi-expert optimization design method for a wide-speed-range aircraft component according to claim 1, characterized in that, The aircraft components include airfoils, wings, control surfaces, tail fins, missile wings, grid rudders, air intakes, or partial shapes of high-speed aircraft.

4. The Mach number adaptive multi-expert optimization design method for a wide-speed-range aircraft component according to claim 1, characterized in that, The specific steps for step one are as follows: Determine the operating Mach number range, geometric design variables, and aerodynamic and structural responses of the wide-speed-range aircraft components to be optimized. Within the range of Mach number and design variable values, generate sample points using Latin hypercube, orthogonal, uniform, or adaptive sampling methods. Conduct experiments or numerical simulations on each sample point to obtain the mapping relationship with Mach number and geometric design variables as inputs and aerodynamic and structural responses as outputs, thus forming a complete wide-speed-range sample dataset.

5. The Mach number adaptive multi-expert optimization design method for a wide-speed-range aircraft component according to claim 1, characterized in that, The specific steps in step two are as follows: For any sample in the wide-speed-domain sample dataset, let its Mach number be... Preset based on the Mach number operating conditions experienced by the aircraft components One Mach number center point used for radial basis encoding: The width of the corresponding radial basis function is determined based on the distance between the centers of adjacent Mach numbers. Then the first Radial basis function encoded components for: In the formula, For the first Mach number center point, For the first The width of the radial basis function corresponding to each Mach number center point; This yields the Mach number encoding vector corresponding to the sample. : The above encoding is performed on all samples in the sample dataset to obtain the Mach number encoding vector corresponding to each sample.

6. The Mach number adaptive multi-expert optimization design method for a wide-speed-range aircraft component according to claim 5, characterized in that, In step three, for any sample in the wide-range sample dataset, the enhanced input feature is represented as: In the formula, This represents the geometric design variables corresponding to this sample. This represents the interaction terms between the Mach number and the various geometric design variables. This represents the interaction terms between different geometric design variables. The number of geometric design variables is specified; for each sample in the sample dataset, enhanced input features are constructed in the manner described above and used as input to the Mach number adaptive multi-expert agent model.

7. The Mach number adaptive multi-expert optimization design method for a wide-speed-range aircraft component according to claim 1, characterized in that, The Mach number adaptive multi-expert agent model described in step four includes: Multiple expert subnetworks are set up in parallel to learn features under different speed states or different response patterns. Each expert subnetwork takes the enhanced input features as input and the corresponding expert features as output. Each expert subnetwork adopts a multilayer perceptron, convolutional neural network, residual network or other neural network structure or combination thereof suitable for learning data-type parameter mapping relationships. Mach number gated networks employ fully connected networks, taking the Mach number encoding vector as input and adaptively outputting expert weights and expert feature modulation parameters for each expert subnetwork at different Mach numbers. The output expert weights characterize the contribution of different expert subnetworks under the current Mach number condition, while the expert feature modulation parameters are used to perform dimensionality-level correction on the output features of the expert subnetworks. For any Mach number... The weights of each expert subnetwork output by the Mach number gated network satisfy the non-negativity constraint and the sum of the weights is 1; The expert feature modulation and fusion module modulates the expert features output by each expert subnetwork using the expert feature modulation parameters output by the Mach number gating network, and then performs weighted fusion of the modulated expert features according to the expert weights to obtain Mach number adaptive fusion features. A shared backbone network is used to input Mach number adaptive fusion features into the shared backbone network and further extract shared features for aerodynamic and structural response prediction. The multi-task prediction module for aerodynamic and structural responses constructs aerodynamic prediction branches and structural prediction branches based on shared features. Each branch consists of an independent task-specific sub-network and an output head. A Cross-Stitch soft-shared unit is set between the intermediate feature layers of the aerodynamic and structural prediction branches. The aerodynamic and structural intermediate features processed by the soft-shared unit are input into the corresponding output heads to obtain the predicted aerodynamic and structural response parameters.

8. The Mach number adaptive multi-expert optimization design method for a wide-speed-range aircraft component according to claim 1, characterized in that, In step four, during the training of the Mach number adaptive multi-expert agent model, the Mach number is directly input with its original value retained, while the geometric design variables and aerodynamic and structural responses are normalized based on the training set.

9. The Mach number adaptive multi-expert optimization design method for a wide-speed-range aircraft component according to claim 1, characterized in that, For the restrictive response parameters, a wide-speed-domain performance constraint index is constructed simultaneously, which requires that the corresponding response values ​​of the candidate design variables at all Mach number sampling points meet the preset constraint conditions, thereby ensuring that the optimized configuration meets the performance constraint requirements throughout the entire wide-speed-domain.

Images