An adaptive optimization method and system applied to the aerodynamic structure design of an aircraft

Abstract

This invention belongs to the field of intelligent optimization and aircraft technology, and discloses an adaptive optimization method and system applied to the aerodynamic structure design of aircraft. The method includes: using the aerodynamic structure parameters of the aircraft as design variables, determining the search space of the design variables, initializing sampling points, and evaluating the objective function value of each sampling point; entering the global search stage; constructing an evaluation support point set, using a surrogate model to predict the evaluation support points, calculating the predicted quality evaluation index based on the predicted values, and if the predicted quality evaluation index meets the switching conditions, entering the focused development stage; otherwise, returning to the global search stage. This invention significantly improves the solution efficiency and accuracy of expensive black-box optimization problems without relying on gradient information by designing a stage switching index based on surrogate model prediction of quality evaluation, a dimension-adaptive search strategy, and an iterative surrogate model refinement local development strategy.

Description

Technical Field

[0001] This invention belongs to the field of intelligent optimization and aircraft technology, specifically relating to an adaptive optimization method and system for aircraft aerodynamic structure design. Background Technology

[0002] In numerous engineering and scientific fields such as aerospace, new energy, new materials, pharmaceuticals, and artificial intelligence, the objective functions of many optimization problems exist in a black-box manner, meaning that the analytical expression and gradient information of the objective function are unavailable, and the computational cost of each function evaluation is extremely high. For example, in complex simulation optimization, a single simulation evaluation may require several hours or even days of computation. Therefore, how to efficiently find the global optimum within a limited function evaluation budget is a core problem of significant engineering value.

[0003] Surrogate-assisted optimization is the mainstream paradigm for solving the aforementioned problems. Its core idea is to construct a computationally inexpensive surrogate model to approximate the objective function using existing function evaluation data, and then guide new sampling directions by optimizing the surrogate model. The Gaussian process (GP) is one of the most commonly used surrogate models, assuming that the objective function follows a Gaussian distribution. Its advantage lies in providing not only a mean prediction of the objective function but also quantifying the uncertainty of the prediction. Gaussian-assisted optimization typically uses sampling functions (such as expected improvement, confidence bounds, etc.) to balance exploring regions of high uncertainty with developing regions of low predicted values.

[0004] However, existing agent-assisted optimization methods still face the following prominent problems in engineering applications:

[0005] (1) The problem of the uniformity of search strategy. Existing methods usually adopt a uniform search strategy throughout the optimization process (e.g., always use the expected improvement acquisition function), which lacks the ability to perceive and respond to the characteristics of different stages of optimization. In the early stage of optimization, global exploration is necessary; however, after the surrogate model has accurately identified the potential regions, continuing global exploration will waste valuable function evaluation budget. At this time, it is necessary to shift to more focused local development.

[0006] (2) Insufficient dimensionality adaptability. As the dimensionality of design variables increases, the search space grows rapidly. Most methods use a fixed-size set of candidate points and a search strategy with fixed parameters, which leads to insufficient coverage and decreased search efficiency in medium- and high-dimensional problems.

[0007] (3) Lack of adaptability in stage switching. Existing research has attempted to combine global search with local search. For example, the paper "Scalable Global Optimization via Local Bayesian Optimization" proposes a Trust-Region Bayesian Optimization (TuRBO) method, which combines global and local optimization by performing fine-grained searches in local regions and dynamically adjusting the size of the trust region. In addition, the paper "LABCAT: Locally Adaptive Bayesian Optimization with Trust Regions" proposes a locally adaptive trust region optimization method, which improves optimization efficiency through a local surrogate model and region adjustment mechanism. However, the above methods usually rely on preset rules or heuristic strategies (such as the number of iterations, convergence threshold, or trust region shrinkage rules) during stage switching, and lack explicit characterization and adaptive adjustment capabilities for the current optimization state. Specifically, switching too early may lead to insufficient global exploration and the omission of potential optimal regions; switching too late will consume a large amount of computational resources in inefficient global search, reducing overall optimization efficiency. In addition, existing methods often fail to fully utilize the predictive information and uncertainty measures of surrogate models to guide stage switching, making the balance between global exploration and local development still dependent on empirical settings, which makes it difficult to achieve stable performance in complex high-dimensional problems.

[0008] (4) Low efficiency of local development. After switching to local development, existing methods usually adopt a single-step surrogate model optimization or a fixed-step search strategy, failing to fully utilize the information gain brought by each new evaluation to update and iteratively refine the surrogate model in real time. In such methods, the information gain brought by each evaluation of the true function is not fully utilized, and the update of the surrogate model is disconnected from the local search process, causing the local development stage to often remain in the "single-point improvement" mode, making it difficult to quickly approximate the potential optimal region. This problem is particularly prominent in high-dimensional complex problems, easily causing slow local convergence speed, or even stagnation in the suboptimal region.

[0009] Evolutionary algorithms such as Particle Swarm Optimization (PSO) and Genetic Algorithm (GA) are widely used in engineering optimization because they do not rely on gradient information and have good global search capabilities. However, these methods are essentially population-based stochastic search strategies, and their convergence process depends on repeated evaluation of a large number of candidate solutions. This is too inefficient for expensive optimization problems where a single evaluation costs several hours.

[0010] In summary, there is an urgent need for a proxy-assisted optimization method that can adaptively perceive the optimization state, efficiently switch between global exploration and local development, and has good dimensional scalability. Summary of the Invention

[0011] The purpose of this invention is to provide an adaptive optimization method and system for aircraft aerodynamic structure design. This method is a surrogate-assisted global optimization with adaptive stage switching, suitable for global optimization scenarios where function evaluation is costly and gradient information is unavailable. This invention significantly improves the efficiency and accuracy of solving expensive black-box optimization problems without relying on gradient information by designing a stage switching index based on surrogate model prediction quality evaluation, a dimension-adaptive search strategy, and an iterative surrogate model refinement local development strategy.

[0012] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0013] Firstly, an adaptive optimization method for aircraft aerodynamic structure design is provided, comprising the following steps:

[0014] Using the aerodynamic structural parameters of the aircraft as design variables, the search space of the design variables is determined, the sampling points are initialized, and the objective function value of each sampling point is evaluated.

[0015] Global search phase: Train the surrogate model using all evaluated sampling points, generate a dimensional adaptive candidate point set based on the search space, select the optimal candidate point as the initial candidate point using the progressive focusing acquisition function, perform local gradient optimization on the progressive focusing acquisition function starting from the initial candidate point to obtain fine sampling points, evaluate the fine sampling points and add them to the evaluated point set, and update the global optimal solution;

[0016] Construct an evaluation support point set, use a surrogate model to predict the evaluation support points, calculate the predicted quality evaluation index based on the predicted values, and if the predicted quality evaluation index meets the switching conditions, enter the focused development stage; otherwise, return to the global search stage.

[0017] Focused development phase: Train a surrogate model using all evaluated sampling points, perform multi-starting point local optimization with the posterior mean of the surrogate model as the optimization objective, select the point with the best predicted value in the multi-starting point local optimization results as the focus candidate point, evaluate the focus candidate point and add it to the evaluated point set, update the global optimal solution, iterate the focused development phase until the number of iterations is reached, and output the global optimal solution as the optimal aerodynamic structure parameters of the aircraft.

[0018] Several alternative methods are provided below, but they are not intended as additional limitations on the overall solution above. They are merely further additions or optimizations. Provided there are no technical or logical contradictions, each alternative method can be combined individually with respect to the overall solution above, or multiple alternative methods can be combined with each other.

[0019] Preferably, the progressive focusing acquisition function has progressive weights that decrease exponentially with the current iteration number.

[0020] Preferably, the dimension-adaptive candidate point set is wherein the number of candidate points increases linearly with the dimension of the search space within a maximum number threshold.

[0021] Preferably, the evaluation support point set is constructed from the following three sources:

[0022] First source: Determine the number of local evaluation support points based on the dimension of the search space, and randomly select sampling points of the number of local evaluation support points from the set of evaluated points as evaluation support points;

[0023] Second source: Centered on the current global optimal solution, within a preset multiple of the evaluation neighborhood radius, generate the number of normal perturbation points of the local evaluation support points, which serve as evaluation support points;

[0024] Third source: Determine the number of global evaluation support points based on the dimensions of the search space, and uniformly and randomly generate sampling points of the number of global evaluation support points within the search space as evaluation support points.

[0025] Preferably, the step of calculating the predicted quality evaluation index based on the predicted value includes:

[0026] The predicted values ​​output by the proxy model are normalized to obtain normalized predicted values. If the normalization range is less than the range threshold, the prediction quality evaluation index is set to a preset constant.

[0027] Otherwise, calculate the Euclidean distance from each evaluation support point in the evaluation support point set to the global optimal solution, and classify the evaluation support points whose Euclidean distance is less than the evaluation neighborhood radius into the neighborhood group; otherwise, classify them into the neighborhood group.

[0028] If either the intra-neighbor group or the out-neighbor group is empty, the prediction quality evaluation index is set to a preset constant; otherwise, the minimum normalized prediction value in the intra-neighbor group is taken as the intra-neighbor prediction minimum, the minimum normalized prediction value in the out-neighbor group is taken as the out-neighbor prediction minimum, and the maximum value between the difference between the intra-neighbor prediction minimum and the out-neighbor prediction minimum and the preset constant is taken as the prediction quality evaluation index.

[0029] Preferably, the evaluation neighborhood radius is characterized by the following values:

[0030] Calculate the diagonal length of the search space, and set the diagonal length weighted by the length coefficient as the evaluation neighborhood radius, wherein the evaluation neighborhood radius is smaller than the diagonal length.

[0031] As a preferred method, the process for determining whether the predicted quality evaluation indicators meet the switching conditions is as follows:

[0032] If the current iteration number is less than or equal to the minimum iteration number in the first stage, or if the number of valid values ​​in the historical records of the prediction quality evaluation index is less than the number threshold, then it is determined that the prediction quality evaluation index has not met the switching conditions.

[0033] Otherwise, if the immediate or trend conditions are met, the prediction quality evaluation index is determined to have met the switching conditions; otherwise, the prediction quality evaluation index is determined not to have met the switching conditions.

[0034] The immediate condition is that the current predicted quality evaluation index is less than the basic evaluation threshold; the trend condition includes: taking the most recent preset number of valid values ​​to form a sliding window, the median of the sliding window is lower than a first preset multiple of the basic evaluation threshold, and the current predicted quality evaluation index is lower than a second preset multiple of the basic evaluation threshold, and the linear regression slope of the predicted quality evaluation index within the sliding window is negative, and the first preset multiple is less than the second preset multiple.

[0035] Preferably, the multi-starting point local optimization is performed in the following way:

[0036] Select the first few sampling points with the smallest objective function value from the set of evaluated points as the starting point for refinement;

[0037] Several random starting points are generated in the search space using the Latin hypercube sampling method;

[0038] The refined starting point and the random starting point are used as multiple starting points, and local optimization is performed at each starting point.

[0039] Preferably, the method further includes a deduplication check on the candidate focal points, wherein the deduplication check includes:

[0040] Calculate the minimum Euclidean distance between the candidate focus point and all sampled points in the evaluated point set;

[0041] If the minimum Euclidean distance is less than the distance threshold, a random perturbation is applied to the candidate points for focusing and they are clipped into the search space to obtain new candidate points for focusing.

[0042] Secondly, an adaptive optimization system for aircraft aerodynamic structure design is provided, comprising:

[0043] The initialization module is used to determine the search space of the design variables, initialize the sampling points, and evaluate the objective function value of each sampling point, using the aerodynamic structural parameters of the aircraft as design variables.

[0044] The global search module is used to train the surrogate model using all evaluated sampling points, generate a dimensional adaptive candidate point set based on the search space, select the optimal candidate point as the initial candidate point using the progressive focusing acquisition function, perform local gradient optimization on the progressive focusing acquisition function starting from the initial candidate point to obtain fine sampling points, evaluate the fine sampling points and add them to the evaluated point set, and update the global optimal solution.

[0045] The switching decision module is used to construct an evaluation support point set, use a surrogate model to predict the evaluation support points, and calculate the predicted quality evaluation index based on the predicted value. If the predicted quality evaluation index meets the switching conditions, it enters the focused development stage; otherwise, it returns to the global search stage.

[0046] The focused development module is used to train a surrogate model using all evaluated sampling points. It performs multi-starting-point local optimization with the posterior mean of the surrogate model as the optimization objective. The point with the best predicted value among the multi-starting-point local optimization results is selected as a focused candidate point. The focused candidate point is evaluated and added to the evaluated point set. The global optimal solution is updated. The focused development phase is iterated until the number of iterations is reached. The global optimal solution is output as the optimal aerodynamic structure parameters of the aircraft.

[0047] The adaptive optimization method and system for aircraft aerodynamic structure design provided by this invention have the following advantages compared with the prior art:

[0048] (1) This invention proposes an adaptive phase-switching agent-assisted global optimization framework that does not rely on gradient information. By explicitly quantifying the confidence of the agent model in the current optimal region through predictive quality evaluation indicators, the optimization process is divided into a global search phase and a focused development phase, and adaptive seamless switching between the two phases is achieved, overcoming the defect of low efficiency of a single search strategy in the later stage of optimization.

[0049] (2) The present invention designs a dimension-adaptive search and evaluation mechanism. The number of candidate points, the number of evaluation support points, and the evaluation neighborhood radius are adaptively set according to the problem dimension and the search space scale, so that the method has stable performance on problems from 4 to 32 dimensions.

[0050] (3) This invention proposes an iterative agent refinement focused development strategy. In the second stage, the GP agent model is retrained and multi-starting point optimization is re-executed after each evaluation, making full use of the information gain brought by each evaluation to reduce the prediction bias of the agent model. Compared with the one-time agent optimization scheme, the accuracy and convergence speed of local search are significantly improved.

[0051] (4) Experimental results show that, on the 4-dimensional to 32-dimensional versions of the standard test function, the method of the present invention achieves better final optimization results than the traditional proxy-assisted optimization method within a total evaluation budget of 350 times (150 initial times and 200 iterations), verifying the effectiveness and versatility of the present invention. Attached Figure Description

[0052] Figure 1 This is a flowchart of the adaptive optimization method for aircraft aerodynamic structure design according to the present invention;

[0053] Figure 2 This is a schematic diagram of the adaptive stage switching determination logic of the present invention;

[0054] Figure 3 This is a comparison graph of the method of the present invention and the comparative method on the test function in the experiment of the present invention;

[0055] Figure 4 This is a comparison chart of fuel consumption between the method of this invention and the comparative method in the aerodynamic structure optimization problem during the experiment of this invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0058] like Figure 1 As shown, this embodiment provides an adaptive optimization method for aircraft aerodynamic structure design, including the following steps:

[0059] Step 1, Initialization Phase: Using the aerodynamic structural parameters of the aircraft as design variables, determine the search space of the design variables, initialize the sampling points, and evaluate the objective function value of each sampling point.

[0060] Step 11: Determine the search space.

[0061] Design variables can be aerodynamic shape design variables (shape geometry), such as airfoil parameters (relative airfoil thickness, airfoil camber, leading edge radius), wing plane parameters (root-to-tip ratio, sweep angle, dihedral angle, angle of attack), fuselage parameters (fuselage length-to-diameter ratio, nose cone angle, fuselage maximum cross-section position), and tail parameters (horizontal / vertical tail area, sweep angle, relative position); or structural design variables (dimensions / layouts / layouts), such as wing structure (skin thickness, upper and lower panel thickness, spars and web thickness, wing edge cross-sectional area), stiffeners / stringers (stiffener spacing, cross-sectional dimensions, stringer thickness), composite materials (layout angle, number of layup layers, layup ratio in each direction), and fuselage structure (frame spacing, panel thickness, stiffener dimensions); or a combined optimization of aerodynamic shape design variables and structural design variables, which is not limited in this invention.

[0062] The objective function can be an aerodynamic objective, such as minimizing the aerodynamic drag coefficient, maximizing the lift-to-drag ratio, or minimizing induced drag; it can also be a structural objective, such as minimizing the total mass of the aircraft / wing structure, minimizing the maximum equivalent stress of the structure, minimizing the maximum deformation of the wing, or maximizing the first-order natural frequency of the structure; or it can be a weighted composite single-objective function or a multi-objective function built on the basis of a single objective, and this invention does not impose any restrictions.

[0063] Take design variables The search space is A bounded rectangular region of dimension 1:

[0064]

[0065] in, It is a bounded rectangular region. and The first The lower and upper bounds of each dimension. For the design variable of the first Each dimension.

[0066] Step 12: Generate data in the search space using Latin Hypercube Sampling (LHS). 1 initial sampling point; LHS will divide the interval of each dimension Divide into the initial number of sampling points Each sub-interval is used to randomly select a sampling point, and the sampling order of each dimension is randomly arranged to ensure that the sampling points are uniformly distributed across all dimensions. This generates the initial sample set. for:

[0067]

[0068] in, For the first One initial sampling point, .

[0069] Step 13, for The objective function value is evaluated at each of the initial sampling points to obtain the initial objective function value set. :

[0070]

[0071] in, For the first An initial objective function value, The objective function is denoted as .

[0072] Step 2, First Stage: Global Search Based on Progressive Focus Strategy.

[0073] Step 21: Train a Gaussian process surrogate model using all currently evaluated sampling points and their objective function values. Take the current set of evaluated points as... Among them, the historical sampling point set Historical objective function value set , This represents the total number of sampled points currently evaluated. The Gaussian process surrogate model uses a constant kernel and... The product of kernels is used as a composite kernel function:

[0074]

[0075] in, This represents a composite kernel function that measures two sampling points in the search space. The similarity between them Represents the signal variance, with a constant kernel. for nuclear, These are smoothness parameters. Among them, the constant kernel parameter... The search scope is ; Nuclear Analytical form of time for:

[0076]

[0077] in, For Euclidean distance, For each dimension, there is an independent length scale parameter vector, and the search range for each component is... The upper bound of the search range for the kernel function parameters (including the length scale parameter and the signal variance parameter) of the Gaussian process surrogate model is set to a large value to adapt to objective functions with different scales and smoothness, and to avoid underfitting of GP due to an excessively narrow parameter range. It corresponds to a two-differentiable objective function and has good generalization ability in engineering optimization.

[0078] The posterior distribution of a Gaussian process is at any point posterior mean at and prediction variance for:

[0079]

[0080]

[0081] in, The kernel matrix corresponding to the training points. ; To observe the noise variance; The kernel is the identity matrix. The kernel parameters are optimized by maximizing the logarithmic marginal likelihood using a bounded gradient optimization method with three random restarts (Limited-memory Broyden-Fletcher-Goldfarb-Shanno with Bounds, L-BFGS-B). The objective function values ​​are standardized (zero mean, unit variance) before training.

[0082] Step 22: Construct the progressive focusing acquisition function ,in and For predicting the posterior mean and standard deviation. Asymptotic weighting coefficients. Adjusted exponentially with the current iteration number: ,in, This represents the current iteration number in the first phase (counting from 0). To optimize the total budget for the iteration phase (budget being the number of iterations), it equals the total evaluation budget minus the initial number of samples; where the initial number of samples is the number of initial sampling points generated in step 12 using Latin hypercube sampling. For example, in this embodiment, the total evaluation budget is set to 350 times, and the initial sampling number is set to 150 times. Therefore, the total budget for the optimization iteration phase is... It is 200 times.

[0083] The constants in the formula are preferably selected from those in the formula. , , The characteristic of this attenuation law is: in hour, , The weighting is high, and the search focuses on exploring regions of high uncertainty; with Increase to , Gradually decreasing to approximately 0.568, the search gradually focuses on the development region with the smallest predicted mean. This progressively focused acquisition function eliminates the need for forced switching between different acquisition function types. The smooth transition between the exploration and development strategies is achieved through the continuous decay of the progressive weight coefficients. This gradual adjustment of the strategy requires no manual intervention, realizing an automatic balance between exploration and development and effectively avoiding sudden changes in the search introduced by rigid, artificial rules.

[0084] Step 23: Generate a dimension-adaptive candidate point set in the search space. Number of candidate points. Determined using the following adaptive formula:

[0085]

[0086] This design ensures that: hour , hour The sampling size increases linearly with the dimension of the search space within a maximum threshold (2000 in this embodiment, which can be adjusted according to actual needs in other embodiments). This ensures that there are enough candidate points to cover the search space in problems of different dimensions, solving the problem of insufficient coverage of fixed-size candidate points in high-dimensional spaces. The LHS method is used to generate... After selecting candidate points, a dimension-adaptive candidate point set is formed. Calculate all candidate points The optimal value is selected as the initial candidate point. :

[0087]

[0088] Step 24: Using the initial candidate points Starting from this point, the local gradient optimization of the acquisition function is performed using the bounded constraint gradient optimization method to obtain fine sampling points. The optimization problem can be formalized as follows:

[0089]

[0090] Local refinement optimizes the gradient of the acquisition function itself, rather than the objective function. Since the posterior mean and posterior standard deviation of a Gaussian process are both analytically differentiable functions of the input, they can be directly calculated. about The gradient is calculated, and the local refinement process takes a maximum of 20 iterations. This local refinement does not consume the objective function evaluation budget and has extremely low computational cost (milliseconds), but it can significantly improve the quality of candidate points, thereby increasing the information gain of the next objective function evaluation.

[0091] Step 25: Evaluate the new sampling points. Evaluate the objective function values ​​of the refined sampling points. The fine sampling points and their corresponding objective function values ​​are then compared. Add to the evaluated point set and update. and And update the current global optimal solution. and the global optimal objective function value .

[0092] Step 3: Calculate the quality evaluation index of the surrogate model prediction and make a stage switching judgment.

[0093] Step 31, in each iteration of the first stage (iteration number) Afterwards, calculate the predicted quality evaluation index. This metric measures the surrogate model's confidence in predicting the current optimal region relative to the global search space, serving as a core driver for phase switching.

[0094] Step 311: Calculate the scale features of the search space. (Diagonal length of the search space):

[0095]

[0096] Set the evaluation neighborhood radius :

[0097]

[0098] The preferred ratio coefficient is... This means that 5% of the diagonal length is taken as the evaluation neighborhood radius. This setting adaptively adjusts with the search space scale, ensuring the geometric consistency of the evaluation index across problems of different scales.

[0099] Step 312: Construct a set of evaluation support points It contains three sources, and the quantity of each part is adaptively adjusted according to the dimension:

[0100] First source: Determine the number of local evaluation support points based on the dimensions of the search space. Randomly selected from the already evaluated point set Each sampling point serves as a support point for evaluation.

[0101] Second source: Centered on the current global optimal solution, within an evaluation neighborhood radius of a preset multiple (0.5 in this embodiment, which can be adjusted according to actual needs in other embodiments), generate... A normal perturbation point is used as the evaluation support point; the generation formula is:

[0102]

[0103] in, This represents the current globally optimal solution. Indicates the first Local disturbance sampling points, For the first part that follows a zero-mean multivariate Gaussian distribution A random perturbation vector, for 3D identity matrix This indicates that the mean vector is 0 and the covariance matrix is... The multivariate normal distribution is used to generate local candidate samples in the neighborhood of the current optimum, thus characterizing the search space characteristics near the preferred region.

[0104] Third source: Determine the number of global evaluation support points based on the dimensions of the search space. Uniformly and randomly generated within the search space Each sampling point serves as an evaluation support point. In this embodiment, with the same dimensionality of the search space, the number of global evaluation support points is greater than the number of local evaluation support points.

[0105] The evaluation support points generated from the above three sources are merged and trimmed into the search space to obtain the final set of evaluation support points. .

[0106] Step 313: Utilize the GP proxy model to... For each evaluation support point in the data, a prediction value vector is obtained. Minimum normalization is applied to the predicted values:

[0107]

[0108] in, For the normalized prediction vector, The minimum predicted value in the prediction vector. The maximum predicted value in the prediction vector. Used to prevent division by zero. If the normalization range (denominator) is less than... Then, the predicted quality evaluation index is directly set to a preset constant (the preset constant is 0 in this embodiment), that is, set... This indicates that the surrogate model's predictions for this region are highly concentrated, and there is no need for global exploration.

[0109] Otherwise according to From each evaluation support point to the current global optimal solution The Euclidean distance is used to classify evaluation support points whose Euclidean distance is less than the evaluation neighborhood radius into the in-neighbor group. Otherwise, the evaluation support points will be classified as outside the neighborhood group. .

[0110] If either the intra-neighbor group or the extra-neighbor group is empty, then the prediction quality evaluation index is set to a preset constant. Otherwise, calculate the minimum normalized predicted value for each of the two groups:

[0111]

[0112]

[0113] in, The predicted minimum value within the neighborhood. The minimum predicted value outside the neighborhood is calculated. The prediction quality evaluation index is then calculated. .

[0114] This embodiment predicts quality evaluation indicators. The physical meaning is: if the surrogate model considers the current global optimal solution... If the predicted minimum value within the neighborhood is not better than the predicted minimum value outside the neighborhood, then... This indicates that the surrogate model is fully confident that the current optimal region is the globally optimal region, and the marginal benefit of continuing the global search is extremely low; conversely, if This indicates that the surrogate model believes there are still better candidate regions outside the neighborhood, and the global search still needs to continue. The smaller the size, the more opportune the time for a phase transition.

[0115] It should be noted that the specific numerical parameters in the above formulas for calculating the number of candidate points, the number of local evaluation support points, the number of global evaluation support points, and the evaluation neighborhood radius are all empirical parameters selected in combination with factors such as algorithm stability, search coverage, statistical reliability, and computational cost, and are not unique limiting values. The constant term is used to ensure the minimum sample size or candidate point size under low-dimensional problems, the dimensionality coefficient term is used to adaptively expand the corresponding parameters as the problem dimension increases, and the scaling factor of 0.05 is used to control the local evaluation radius within a certain range of the search space scale, so as to balance local discriminative ability and statistical stability.

[0116] The above parameters can be adjusted according to specific application scenarios to ensure the stability and comparability of the evaluation index calculation across problems of different dimensions and scales. For example, when the objective function evaluation cost is high, the number of candidate points or the number of support points evaluated can be appropriately reduced to decrease additional computational overhead; when the problem dimension is higher, the objective function terrain is more complex, or the stability requirements for switching decisions are higher, the corresponding parameters can be appropriately increased to improve sampling coverage and the robustness of statistical decisions. Therefore, this invention is not limited to the specific values ​​mentioned above. Any equivalent substitution or reasonable adjustment of the above empirical parameters under the same technical concept should fall within the protection scope of this invention.

[0117] Step 32, Adaptive Phase Switching Determination: If the predicted quality evaluation indicators meet the switching conditions, proceed to the focused development phase; otherwise, return to the global search phase. For example... Figure 2 As shown, the specific steps are as follows:

[0118] Step 321: Set the basic evaluation threshold Minimum number of iterations in the first stage Ensure that the first phase executes at least a sufficient number of iterations to prevent the proxy model from prematurely switching due to insufficient training data.

[0119] Step 322, Preconditions: If the current iteration count is less than or equal to the minimum iteration count in the first stage, or if the number of valid values ​​in the historical records of the prediction quality evaluation index is less than the threshold, then it is determined that the prediction quality evaluation index has not met the switching condition, and the process returns to the global search stage; otherwise, proceed to step 323. In this embodiment, when the iteration count... Afterwards, check Historical records Does the number of valid values ​​be no less than 5?

[0120] Step 323: If the immediate or trend conditions are met, the predicted quality evaluation index is determined to have met the switching conditions, and the process enters the focused development stage; otherwise, the predicted quality evaluation index is determined not to have met the switching conditions, and the process returns to the global search stage.

[0121] The first set of conditions (immediate conditions) is as follows:

[0122]

[0123] in, This is the current prediction quality evaluation metric. If the current evaluation metric immediately falls below the basic evaluation threshold, it indicates that the surrogate model is highly confident that the current optimal region is the global optimum, and a switch is immediately triggered.

[0124] The second set of conditions (trend condition, all three conditions must be met simultaneously): Take the most recent... Second effective Values ​​form a sliding window The switching is triggered when all three of the following conditions are met:

[0125] (a) Median of the sliding window The first preset multiple below the basic evaluation threshold:

[0126]

[0127] (b) Current The value is lower than the second preset multiple of the basic evaluation threshold:

[0128]

[0129] (c) The linear regression slope of the index value within the sliding window is negative. (Values ​​are trending downwards)

[0130]

[0131] in, The slope of the linear regression. As the independent variable, The mean of the independent variable. The dependent variable is the independent variable. This represents the mean of the dependent variable.

[0132] Condition (a) corresponds to the overall statistical level of the sliding window, and condition (b) corresponds to the current single-point evaluation level. To ensure that the overall trend judgment is stricter than the single-point judgment, the preset multiple in condition (a) should be less than the preset multiple in condition (b). In one embodiment, the preset multiple of condition (a) is 5, and the preset multiple of condition (b) is 10. The former is used to ensure that most indicator values ​​within the sliding window have entered a lower level range, and the latter is used to reserve fluctuation margin for the current single-point indicator value. Through the above settings, the switching criterion can simultaneously take into account overall stability and single-point sensitivity, avoid false switching caused by instantaneous abnormal fluctuations, and improve the reliability of stage switching timing judgment. The median statistics are robust to outliers. The purpose of the trend condition design is: for The value is declining slowly but has not yet broken through the critical threshold. In such cases, statistical analysis using a sliding window can be used to determine the continuous improvement trend of the surrogate model's prediction quality. If the indicator shows a continuous downward trend and is at a relatively low level, a switch can still be triggered to avoid unnecessary consumption of too much evaluation budget in the first stage.

[0133] Step 4, Second Phase: Focused Development of Iterative Agent Refinement.

[0134] Step 41: Once the stage switching determination is met, record the number of iterations in which the switching occurred. Calculate the remaining iteration budget :

[0135]

[0136] in, To optimize the total budget during the iteration phase; This represents the actual number of iterations consumed in the first phase (i.e., the iteration count when the switching decision is valid). This is the remaining iteration budget when entering the second phase.

[0137] It should be noted that when the stage switching condition is met, if the remaining iteration budget... Then proceed to step 42 of the second phase to execute iterative local reinforcement development; if the remaining iteration budget If the condition is met at the end of the budget, then a new second-stage iteration will not be started. Instead, the current global optimal solution will be directly output as the optimal aerodynamic structure parameters of the aircraft.

[0138] Step 42: In each iteration of the second phase, perform the following steps:

[0139] (a) Using all currently evaluated samples Retrain the Gaussian process surrogate model to ensure that the surrogate model incorporates the information gain from the latest evaluation results in real time.

[0140] (b) Constructing a set of starting points for multi-starting-point optimization :from Selecting the objective function value The smallest front The sampling points corresponding to each observation are used as the starting point for refinement in this embodiment. At the same time, a random number is generated using the LHS method. The random starting points are merged into:

[0141]

[0142] in, The first one that minimizes the objective function value A refined starting point This is the 30th random starting point.

[0143] (c) For each starting point in the GP, the posterior mean is used. To optimize the objective, a bounded-constraint gradient method (maximum 30 iterations) is used for multi-starting-point local optimization to obtain the optimization result. The optimization problem is set as follows:

[0144]

[0145] in, For the first The optimization results corresponding to each starting point are analyzed. The point with the best predicted value from the surrogate model is selected as the focus candidate point from all the multi-starting-point optimization results. , To optimize the results The posterior mean of the corresponding surrogate model's predicted output.

[0146] (d) Perform a deduplication test on the candidate focal points and calculate... arrive Minimum Euclidean distance of all historical sampling points :

[0147]

[0148] like Distance threshold Then to Apply random perturbations and crop the data to the search space to obtain new candidate focal points, thus avoiding repeated evaluation of the same location:

[0149]

[0150] in, This is a truncation function. The vector is Gaussian noise. The mean vector is 0, and the covariance matrix is... The Gaussian distribution.

[0151] Step 43: Evaluate the objective function value of the candidate focus points. ,Will Add to the evaluated point set and update. and And update the current global optimal solution. and the global optimal objective function value .

[0152] Step 44, repeat steps 42 and 43. Repeat this process until the remaining iteration budget is exhausted. Output the globally optimal solution. and its corresponding objective function value .

[0153] In each iteration of the second phase, the surrogate model is retrained, and the results of each evaluation of the true objective function are immediately incorporated into the surrogate model, ensuring that the surrogate model's characterization of the local optimum is continuously refined. Compared to a one-time optimization of the surrogate model's surface, the iterative refinement strategy can correct the bias of the surrogate model after each round of evaluation, gradually narrowing the gap between the surrogate model's predicted minimum and the true optimum, and achieving rapid approximation of the true global optimum.

[0154] To visually demonstrate the advantages of this invention, the following verification was performed.

[0155] (1) System verification of the adaptive optimization method of the present invention applied to the aerodynamic structure design of aircraft using standard test functions.

[0156] (11) The Rosenbrock function was selected as a representative standard test function to verify the method of the present invention: Rosenbrock function: Search domain , For function values, for 3D decision variable vector, For the first Dimensional decision variables, For the first Dimensional decision variables.

[0157] (12) Experimental setup: , Total budget assessment: 350 times. , The comparison method is standard proxy-assisted optimization (BO). The evaluation metrics are as follows:

[0158]

[0159] in To normalize the improvement, the closer it is to 1, the better the optimization effect. Among these, This represents the optimal objective function value in the initial sample set. This represents the optimal objective function value obtained up to the current time step in the current iteration process. This represents the theoretical global optimum of the corresponding test function. The above metric measures the degree of normalization improvement of the current optimization result relative to the initial optimum, and its value range is... The closer the value is to 1, the closer the current result is to the theoretical optimal value, and the better the optimization effect.

[0160] (13) The experimental results summarize the optimization gaps obtained by the method of this invention (hereinafter referred to as adaptive switching agent-assisted optimization) and the standard Bayesian optimization (standard BO) method on the 16-dimensional Rosenbrock test function at the end of a total evaluation budget of 350 runs. Value and the number of stage switching iterations of the method of the present invention .

[0161] The method of this invention achieved optimal results in 14 out of 16 tests. , among which Figure 3As shown, the most representative example is the 16-dimensional Rosenbrock function: the final version of this invention The optimal function value is 0.9768 (optimal function value 4262.43), while the standard BO is only 0.2019 (optimal value 146917.66), meaning the method in this invention improves upon the BO by approximately 3.84 times. This result fully verifies the superior performance of the adaptive stage switching mechanism and iterative proxy refinement strategy on high-dimensional complex optimization problems.

[0162] The Rosenbrock function has a narrow, winding optimal valley with a global optimum of 0. In high dimensions, it demands extremely high local search accuracy, making it a typical case for validating the iterative proxy refinement strategy of this invention. The experiment uses a 16-dimensional Rosenbrock function (search domain...). The total evaluation budget is 350 times (150 initial samplings in the first stage and 200 iterations in the second stage), with a random seed of 100.

[0163] The advantages of this invention are particularly pronounced in 16 dimensions: In 16 dimensions, this invention... It is 0.9768, the standard BO. Only 0.2019. It's worth noting that the phase switch in the 16-dimensional Rosenbrock only occurs on the 148th iteration. (A total of 200 iterations) This demonstrates that the problem requires a relatively long global exploration time to locate the optimal valley region in high dimensions. The adaptive switching mechanism of this invention can perceive this requirement, delaying the switch until the surrogate model is fully confident before entering focused development, ultimately achieving results far superior to the comparative methods. The absolute advantage of this invention is as high as 34 times that of the BO results.

[0164] The experimental results above demonstrate that the method of this invention optimizes the gap on the 16-dimensional Rosenbrock function. Compared to standard BO ( The algorithm achieves a high-precision convergence, improving efficiency by approximately 383.8%, thus verifying the effectiveness of the adaptive stage switching architecture and iterative proxy refinement strategy of this invention in high-dimensional, expensive black-box optimization problems.

[0165] In summary, using the 16-dimensional Rosenbrock function as a typical verification case, under the constraint of a total evaluation budget of 350 iterations, the method of this invention achieves a significantly better final optimization gap than standard Bayesian optimization. This invention, through an adaptive stage switching architecture driven by prediction quality evaluation metrics, a dimension-adaptive search and evaluation mechanism, and a focused development strategy of iterative agent refinement, achieves efficient solutions to high-dimensional, expensive black-box optimization problems, thus verifying the effectiveness of this invention.

[0166] It is particularly noteworthy that the method of this invention achieves the aforementioned optimization performance even with an extremely limited number of calls to the real objective function. In this experiment, the total evaluation budget was only 350 calls, with 150 calls used for initial sampling and only 200 calls for real function evaluation during the optimization iteration phase. In contrast, standard Bayesian optimization loses its advantage in high dimensions due to the degradation of the sampling function. This invention, through an adaptive phase switching mechanism, rationally allocates the limited total budget for the optimization iteration phase between global exploration and focused development. On the 16-dimensional Rosenbrock function, after switching, focused development with only 52 calls to the real function significantly reduced the function value from 184075 to 4262, demonstrating significantly better evaluation budget utilization efficiency than the comparative methods. For high-cost engineering simulation scenarios (such as aerodynamic optimization, finite element analysis, drug molecule simulation, etc.) where a single function evaluation requires several hours or even days of computation time, the ability of this invention to achieve high-precision convergence with extremely low evaluation budget has outstanding practical engineering value.

[0167] (2) Verification of aerodynamic coupling simulation tools.

[0168] To further verify the practicality of the method of this invention in real engineering simulation scenarios, this embodiment applies the two-stage quality-aware Bayesian optimization method (two-stage quality-aware BO) of this invention to the airfoil aerodynamic optimization problem of the aerodynamic structure coupling analysis tool OpenAeroStruct. OpenAeroStruct is an aerodynamic structure coupling analysis framework based on the vortex lattice method and the finite element beam model. It is widely used in the integrated aerodynamic structure optimization during the conceptual design stage of aircraft. Its single simulation evaluation requires the comprehensive solution of the coupling equations of aerodynamic forces and structural deformation, which has a high computational cost. It is a typical engineering benchmark for evaluating the practicality of expensive black-box optimization methods.

[0169] The optimization objective is to minimize the fuel consumption (in kg) of the aircraft under a given flight mission profile, using airfoil geometric parameters (including wing aspect ratio, tip-to-root ratio, twist angle distribution, and structural thickness) as design variables. Based on the initial configuration before optimization, the initial fuel consumption is 281,197.83 kg. Comparative experiments were conducted using the method of this invention and standard Bayesian optimization (standard BO) under the same total function evaluation budget, comparing the optimal fuel consumption achieved by the two methods.

[0170] like Figure 4As shown, experimental results indicate that compared to the initial configuration fuel consumption of 281,197.83 kg, the method of this invention optimizes fuel consumption to 123,983.68 kg, a reduction of 55.9%; while the standard BO only optimizes fuel consumption to 152,721.94 kg, a reduction of 45.7%. The method of this invention further reduces fuel consumption by 28,738.26 kg compared to the standard BO, representing a relative optimization improvement of 18.8 percentage points. This fully verifies the significant superiority of the adaptive stage switching mechanism and iterative proxy refinement strategy of this invention in real aerodynamic structure coupled optimization engineering problems.

[0171] The above results demonstrate that the performance advantages of the method of this invention are not limited to artificially synthesized test functions, but also have significant practical value in expensive black-box optimization problems in real engineering, such as aerodynamic structure coupling simulation. By using adaptive stage switching driven by predictive quality evaluation indicators, this invention can more efficiently focus on high-potential regions of the design space within a limited simulation budget, achieving high-precision solutions to complex multiphysics coupling optimization problems, and providing an efficient and reliable optimization tool for integrated aerodynamic structure design of aircraft.

[0172] In another embodiment, an adaptive optimization system for aircraft aerodynamic structure design is provided, comprising:

[0173] The initialization module is used to determine the search space of the design variables, initialize the sampling points, and evaluate the objective function value of each sampling point, using the aerodynamic structural parameters of the aircraft as design variables.

[0174] The global search module is used to train the surrogate model using all evaluated sampling points, generate a dimensional adaptive candidate point set based on the search space, select the optimal candidate point as the initial candidate point using the progressive focusing acquisition function, perform local gradient optimization on the progressive focusing acquisition function starting from the initial candidate point to obtain fine sampling points, evaluate the fine sampling points and add them to the evaluated point set, and update the global optimal solution.

[0175] The switching decision module is used to construct an evaluation support point set, use a surrogate model to predict the evaluation support points, and calculate the predicted quality evaluation index based on the predicted value. If the predicted quality evaluation index meets the switching conditions, it enters the focused development stage; otherwise, it returns to the global search stage.

[0176] The focused development module is used to train a surrogate model using all evaluated sampling points. It performs multi-starting-point local optimization with the posterior mean of the surrogate model as the optimization objective. The point with the best predicted value among the multi-starting-point local optimization results is selected as a focused candidate point. The focused candidate point is evaluated and added to the evaluated point set. The global optimal solution is updated. The focused development phase is iterated until the number of iterations is reached. The global optimal solution is output as the optimal aerodynamic structure parameters of the aircraft.

[0177] For limitations on adaptive optimization systems applied to aircraft aerodynamic structure design, please refer to the limitations on adaptive optimization methods applied to aircraft aerodynamic structure design; these will not be repeated in this embodiment.

[0178] In another embodiment, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of an adaptive optimization method applied to the aerodynamic structure design of an aircraft.

[0179] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory.

[0180] In another embodiment, the present invention provides a computer device including a processor and a memory storing a plurality of computer instructions, which, when executed by the processor, implement steps of an adaptive optimization method applied to the aerodynamic structure design of an aircraft.

[0181] The memory and processor are electrically connected directly or indirectly to enable data transmission or interaction. For example, these components can be electrically connected to each other via one or more communication buses or signal lines. The memory stores a computer program that can run on the processor, and the processor implements the method of the present invention by running the computer program stored in the memory.

[0182] The memory can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), and Electrically Erasable Programmable Read-Only Memory (EEPROM). The memory stores the program, which the processor executes upon receiving execution instructions.

[0183] A processor can be an integrated circuit chip with data processing capabilities. The processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention.

[0184] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

[0185] The above embodiments merely illustrate several implementation methods of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the appended claims.

Claims

1. An adaptive optimization method for aircraft aerodynamic structure design, characterized in that, Includes the following steps: Using the aerodynamic structural parameters of the aircraft as design variables, the search space of the design variables is determined, the sampling points are initialized, and the objective function value of each sampling point is evaluated. Global search phase: Train the surrogate model using all evaluated sampling points, generate a dimensional adaptive candidate point set based on the search space, select the optimal candidate point as the initial candidate point using the progressive focusing acquisition function, perform local gradient optimization on the progressive focusing acquisition function starting from the initial candidate point to obtain fine sampling points, evaluate the fine sampling points and add them to the evaluated point set, and update the global optimal solution; Construct an evaluation support point set, use a surrogate model to predict the evaluation support points, calculate the predicted quality evaluation index based on the predicted values, and if the predicted quality evaluation index meets the switching conditions, enter the focused development stage; otherwise, return to the global search stage. Focused development phase: Train a surrogate model using all evaluated sampling points, perform multi-starting point local optimization with the posterior mean of the surrogate model as the optimization objective, select the point with the best predicted value in the multi-starting point local optimization results as the focus candidate point, evaluate the focus candidate point and add it to the evaluated point set, update the global optimal solution, iterate the focused development phase until the number of iterations is reached, and output the global optimal solution as the optimal aerodynamic structure parameters of the aircraft.

2. The adaptive optimization method for aircraft aerodynamic structure design according to claim 1, characterized in that, The progressive focusing acquisition function has progressive weights that decrease exponentially with the current iteration number.

3. The adaptive optimization method for aircraft aerodynamic structure design according to claim 1, characterized in that, The dimension-adaptive candidate point set is wherein the number of candidate points increases linearly with the dimension of the search space within a maximum threshold.

4. The adaptive optimization method for aircraft aerodynamic structure design according to claim 1, characterized in that, The evaluation support point set is constructed, and the evaluation support points have the following three sources: First source: Determine the number of local evaluation support points based on the dimension of the search space, and randomly select sampling points of the number of local evaluation support points from the set of evaluated points as evaluation support points; Second source: Centered on the current global optimal solution, within a preset multiple of the evaluation neighborhood radius, generate the number of normal perturbation points of the local evaluation support points, which serve as evaluation support points; Third source: Determine the number of global evaluation support points based on the dimensions of the search space, and uniformly and randomly generate sampling points of the number of global evaluation support points within the search space as evaluation support points.

5. The adaptive optimization method for aircraft aerodynamic structure design according to claim 1, characterized in that, The calculation of the predicted quality evaluation index based on the predicted value includes: The predicted values ​​output by the proxy model are normalized to obtain normalized predicted values. If the normalization range is less than the range threshold, the prediction quality evaluation index is set to a preset constant. Otherwise, calculate the Euclidean distance from each evaluation support point in the evaluation support point set to the global optimal solution, and classify the evaluation support points whose Euclidean distance is less than the evaluation neighborhood radius into the neighborhood group; otherwise, classify them into the neighborhood group. If either the intra-neighbor group or the out-neighbor group is empty, the prediction quality evaluation index is set to a preset constant; otherwise, the minimum normalized prediction value in the intra-neighbor group is taken as the intra-neighbor prediction minimum, the minimum normalized prediction value in the out-neighbor group is taken as the out-neighbor prediction minimum, and the maximum value between the difference between the intra-neighbor prediction minimum and the out-neighbor prediction minimum and the preset constant is taken as the prediction quality evaluation index.

6. The adaptive optimization method for aircraft aerodynamic structure design according to claim 4 or 5, characterized in that, The evaluation neighborhood radius is taken as follows: Calculate the diagonal length of the search space, and set the diagonal length weighted by the length coefficient as the evaluation neighborhood radius, wherein the evaluation neighborhood radius is smaller than the diagonal length.

7. The adaptive optimization method for aircraft aerodynamic structure design according to claim 1, characterized in that, The process for determining whether the predicted quality evaluation indicators meet the switching conditions is as follows: If the current iteration number is less than or equal to the minimum iteration number in the first stage, or if the number of valid values ​​in the historical records of the prediction quality evaluation index is less than the number threshold, then it is determined that the prediction quality evaluation index has not met the switching conditions. Otherwise, if the immediate or trend conditions are met, the prediction quality evaluation index is determined to have met the switching conditions; otherwise, the prediction quality evaluation index is determined not to have met the switching conditions. The immediate condition is that the current predicted quality evaluation index is less than the basic evaluation threshold; the trend condition includes: taking the most recent preset number of valid values ​​to form a sliding window, the median of the sliding window is lower than a first preset multiple of the basic evaluation threshold, and the current predicted quality evaluation index is lower than a second preset multiple of the basic evaluation threshold, and the linear regression slope of the predicted quality evaluation index within the sliding window is negative, and the first preset multiple is less than the second preset multiple.

8. The adaptive optimization method for aircraft aerodynamic structure design according to claim 1, characterized in that, The multi-starting-point local optimization is determined as follows: Select the first few sampling points with the smallest objective function value from the set of evaluated points as the starting point for refinement; Several random starting points are generated in the search space using the Latin hypercube sampling method; The refined starting point and the random starting point are used as multiple starting points, and local optimization is performed at each starting point.

9. The adaptive optimization method for aircraft aerodynamic structure design according to claim 1, characterized in that, It also includes deduplication checks on the candidate focal points, wherein the deduplication check includes: Calculate the minimum Euclidean distance between the candidate focus point and all sampled points in the evaluated point set; If the minimum Euclidean distance is less than the distance threshold, a random perturbation is applied to the candidate points for focusing and they are clipped into the search space to obtain new candidate points for focusing.

10. An adaptive optimization system for aircraft aerodynamic structure design, characterized in that, include: The initialization module is used to determine the search space of the design variables, initialize the sampling points, and evaluate the objective function value of each sampling point, using the aerodynamic structural parameters of the aircraft as design variables. The global search module is used to train the surrogate model using all evaluated sampling points, generate a dimensional adaptive candidate point set based on the search space, select the optimal candidate point as the initial candidate point using the progressive focusing acquisition function, perform local gradient optimization on the progressive focusing acquisition function starting from the initial candidate point to obtain fine sampling points, evaluate the fine sampling points and add them to the evaluated point set, and update the global optimal solution. The switching decision module is used to construct an evaluation support point set, use a surrogate model to predict the evaluation support points, and calculate the predicted quality evaluation index based on the predicted value. If the predicted quality evaluation index meets the switching conditions, it enters the focused development stage; otherwise, it returns to the global search stage. The focused development module is used to train a surrogate model using all evaluated sampling points. It performs multi-starting-point local optimization with the posterior mean of the surrogate model as the optimization objective. The point with the best predicted value among the multi-starting-point local optimization results is selected as a focused candidate point. The focused candidate point is evaluated and added to the evaluated point set. The global optimal solution is updated. The focused development phase is iterated until the number of iterations is reached. The global optimal solution is output as the optimal aerodynamic structure parameters of the aircraft.

Images