Aircraft hierarchical collaborative optimization method based on data driving
Through a data-driven hierarchical collaborative optimization method, the design variables are decomposed using the low-fidelity data set and machine learning model, combined with high-fidelity CFD optimization, the problem of inefficiency of high-precision CFD simulation methods is solved, and the aerodynamic performance and computing efficiency of the aircraft are improved.
Patent Information
- Application Number
- CN202510601157.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-12
- Publication Date
- 2025-08-01
AI Technical Summary
The existing high-precision CFD simulation method is inefficient and difficult to obtain satisfactory optimization results when dealing with the problem of aircraft aerodynamic optimization of large-scale design variables.
The data-driven hierarchical collaborative optimization method is adopted to obtain the low-fidelity data set through Latin hypercube sampling, and regression prediction and feature importance analysis are used to analyze the LightGBM machine learning model, design variables are decomposed, and high-fidelity collaborative optimization is performed in combination with the CETLBO algorithm.
It significantly improves the aerodynamic performance of the aircraft, achieves a balance between computing efficiency and optimization accuracy, improves the lift-to-resistance ratio and reduces the calculation time.
Smart Images
Figure CN120408859A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of aircraft design, and particularly to a data-driven hierarchical collaborative optimization method for aircraft. Background Art
[0002] The performance and efficiency of an aircraft largely depend on its shape design. A good shape design can significantly improve the aerodynamic performance of the aircraft. However, aerodynamic analysis is characterized by high time consumption and complex calculations, which severely restricts the application and development of aerodynamic optimization algorithms. Although low-fidelity aerodynamic engineering estimation methods and rough geometric parameterization methods can quickly perform aerodynamic optimization, they are increasingly unable to meet the needs of aircraft shape design. Fine geometric parameterization methods and complex aircraft shapes will lead to an increase in decision variables, and the influence of decision variables on the objective is significantly different. If not distinguished and processed, it will significantly affect the performance of the algorithm.
[0003] The aircraft design optimization process in the prior art generally includes the following steps: First, parameterize the aircraft geometric model and extract the required design variables and constraints; Second, generate accurate CFD calculation grids according to the aircraft geometric shape; Then perform CFD simulation calculations to obtain the aerodynamic parameters of the aircraft; Finally, import the aircraft design parameters and aerodynamic parameters into a high-dimensional optimization algorithm for optimization to obtain the optimized aircraft design variables and complete the optimization. Improve the performance of low-fidelity aerodynamic engineering methods.
[0004] However, this high-precision CFD simulation method is inefficient and difficult to obtain satisfactory optimization results when dealing with aerodynamic optimization problems with a large number of design variables. Summary of the Invention
[0005] The purpose of the present invention is to provide a data-driven hierarchical collaborative optimization method for aircraft, which solves the problems that the existing high-precision CFD simulation method is inefficient and difficult to obtain satisfactory optimization results when dealing with aerodynamic optimization problems with a large number of design variables.
[0006] To achieve the above object, the present invention provides a data-driven hierarchical collaborative optimization method for aircraft, including the following steps:
[0007] Obtain a low-fidelity data set;
[0008] Based on the low-fidelity data set, use the LightGBM machine learning model for regression prediction and evaluate the model accuracy through the coefficient of determination;
[0009] Conduct feature importance analysis on the LightGBM machine learning model and group the design variables;
[0010] Perform collaborative optimization based on the grouped design variables to obtain the final optimization result.
[0011] Among them, to obtain the low-fidelity dataset, the specific steps include:
[0012] Generate uniformly distributed sampling points in the design space through the Latin hypercube sampling method, and perform aerodynamic analysis on the airfoil or wing using the low-fidelity flow analysis method to obtain the initial low-fidelity dataset.
[0013] Among them, for obtaining the low-fidelity dataset:
[0014] The number of initial sampling points is 10 - 15 times that of the design variables.
[0015] Among them, based on the low-fidelity dataset, use the LightGBM machine learning model for regression prediction, and evaluate the model accuracy through the coefficient of determination. The specific steps include:
[0016] Use the initial low-fidelity dataset to train the LightGBM machine learning model for regression prediction;
[0017] Calculate the coefficient of determination R of the LightGBM machine learning model 2 to evaluate the model accuracy;
[0018] When R 2 > 0.8, then perform low-fidelity optimization design using the CETLBO algorithm based on the trained LightGBM model;
[0019] When R 2 ≤ 0.8, then supplement the sampling points and retrain the LightGBM model until the accuracy requirement is met.
[0020] Among them, perform feature importance analysis on the LightGBM machine learning model and group the design variables. The specific steps include:
[0021] Use the SHAP method to perform feature importance analysis on the LightGBM machine learning model;
[0022] Group the design variables according to the SHAP values and physical meanings, and decompose the high-dimensional optimization problem into multiple low-dimensional sub-optimization problems.
[0023] Among them, perform collaborative optimization based on the grouped design variables to obtain the final optimization result. The specific steps include:
[0024] Based on the grouped design variables, use the CETLBO algorithm combined with high-fidelity CFD to solve the N - S equations for collaborative optimization to obtain the final optimization result.
[0025] A data-driven hierarchical collaborative optimization method for aircraft, which obtains a low-fidelity data set through Latin hypercube sampling, uses the LightGBM machine learning model for regression prediction and optimization, and adopts the CETLBO algorithm for low-fidelity optimization when the model accuracy reaches the standard; then analyzes and groups the contribution degrees of design variables through the SHAP method, decomposes the high-dimensional problem into multiple low-dimensional sub-problems; finally, conducts high-fidelity collaborative optimization based on the grouping results. This method effectively solves the problems of low computational efficiency, difficult high-dimensional optimization, and insufficient accuracy in traditional aerodynamic optimization, realizes the balance between computational efficiency and optimization accuracy, and significantly improves the aerodynamic performance of the aircraft. Brief Description of the Drawings
[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art.
[0027] Figure 1 It is a flowchart of the data-driven hierarchical collaborative optimization method for aircraft of the present invention.
[0028] Figure 2 It is a schematic diagram of geometric parameterization modeling of the three-dimensional wing configuration of the present invention.
[0029] Figure 3 It is a schematic diagram of the coarse grid calculation polyhedral mesh of the present invention.
[0030] Figure 4 It is a schematic diagram of the fine grid calculation polyhedral mesh of the present invention.
[0031] Figure 5 It is a comparison diagram of the true value and the predicted value of the present invention.
[0032] Figure 6 It is an analysis diagram of the wing feature importance based on the SHAP method of the present invention.
[0033] Figure 7 It is a comparison diagram of the convergence process of the wing high-dimensional optimization of the present invention.
[0034] Figure 8 It is a cloud chart of the wing upper surface pressure coefficient of the benchmark model and the optimized model of the present invention.
[0035] Figure 9 It is a comparison diagram of the geometric shapes of the control sections of the benchmark model and the optimized wing of the present invention.
[0036] Figure 10 It is a comparison diagram of the geometric shapes of the control sections of the benchmark model and the optimized wing of the present invention.
[0037] Figure 11It is the flowchart of the steps of the data-driven hierarchical collaborative optimization method for aircraft of the present invention. Detailed implementation manners
[0038] The embodiments of the present invention will be described in detail below. The examples of the embodiments are shown in the accompanying drawings. The embodiments described below by referring to the accompanying drawings are exemplary and are intended to explain the present invention, but should not be construed as a limitation to the present invention.
[0039] Please refer to Figures 1 to 11 , the present invention provides a data-driven hierarchical collaborative optimization method for aircraft, including the following steps:
[0040] S101: Obtain a low-fidelity data set;
[0041] Specifically, uniform sampling points are generated in the design space by the Latin Hypercube Sampling (LHS) method, and a large amount of low-cost data is obtained by analyzing airfoils, wings, etc. using a low-fidelity flow analysis method. The initial number of sampling points is about 10-15 times the number of design variables.
[0042] S102: Based on the low-fidelity data set, use the LightGBM machine learning model for regression prediction, and evaluate the model accuracy through the coefficient of determination;
[0043] Specifically, a large amount of low-cost data is used for regression prediction by the machine learning method LightGBM (Light Gradient Boosting Machine). The coefficient of determination (Coefficient of Determination, R 2 ) is used to evaluate the accuracy of the prediction model. If R 2 > 0.8, then based on the LightGBM model, the CETLBO algorithm is used for rapid low-fidelity optimization design. Otherwise, new sample points are added to retrain the LightGBM model.
[0044] S103: Perform feature importance analysis on the LightGBM machine learning model and group the design variables;
[0045] Specifically, the SHAP (SHapley Additive ex Planations) method is used to perform feature importance analysis on the LightGBM model. The contribution degree and influence law of the design variables on the target are quantitatively analyzed. The design variables are grouped according to the SHAP values and physical meanings, so as to decompose a large-scale high-dimensional optimization problem into several simple low-dimensional sub-optimization problems. The design variables of this sub-optimization problem are subsets of the design variables of the original problem. This increases the possibility of the algorithm finding the optimal solution.
[0046] S104: Perform collaborative optimization based on the grouped design variables to obtain the final optimization result.
[0047] Specifically, perform high-fidelity collaborative optimization on the decomposed sub-problems based on the low-fidelity optimization result. The dimensionality of the sub-optimization problem variables is reduced, and with the guidance of prior knowledge, the algorithm converges easily. The CETLBO algorithm is also used in the solution process, and high-fidelity flow analysis is achieved by solving the N-S equations.
[0048] It can be seen from this that aerodynamic optimization is a process of iterative solution and requires repeated calls to the CFD solver. Compared with CFD simulation, the machine learning model accelerates the iteration by reducing the simulation time. In addition, the hierarchical modeling method with variable fidelity can reduce the simulation time. Fernández-Godino pointed out that there are usually three types of different fidelities. The first is to simplify the physical model by changing the solution method of the differential equation; the second is to change the resolution of spatial or temporal discretization; the third is to directly use experimental data as the highest-fidelity data. Accordingly, the present invention establishes a multi-fidelity model for compressible flow. When there is a correlation between low-fidelity and high-fidelity data, the low-fidelity model provides the trend of high-fidelity data and achieves accelerated calculation. Specifically, the high-fidelity model corresponds to solving the RANS equations on a fine grid, which describes the behavior of viscous fluids. The low-fidelity model corresponds to solving the Euler equations by ignoring viscosity on a coarse grid.
[0049] To verify the reliability and superiority of the data-driven hierarchical collaborative aerodynamic optimization method, perform hierarchical collaborative optimization design on a supersonic wing and compare the results with those of the full-parameter optimization method based on GA. The local geometry of the entire wing of the supersonic wing is controlled by six typical cross-sections. The fifth-order CST parameterization method is used to parameterize each cross-section, and each cross-section has 11 design variables There are a total of 66 design variables. The parameterized airfoils are sequentially installed along the wing span, and the cross-sectional shapes at other positions can be obtained by interpolation from adjacent control airfoils, as Figure 2 shown.
[0050] The design state is: Ma = 1.6, α = 2.3 。 , Re = 5.89×10 6 , and the mathematical model of this high-dimensional optimization problem is as follows:
[0051] maximize: j(x) = CL / CD
[0052] subjict to: C1 = 31 - Vol ≤ 0
[0053] C2 = CL0 - CL ≤ 0
[0054]
[0055] In the formula, \(CL\) and \(CD\) respectively represent the lift coefficient and drag coefficient of the wing, \(Vol\) represents the volume of the wing, and \(x\) is a vector composed of design parameters. Directly solving the above high-dimensional optimization model is very inefficient and it is difficult to obtain satisfactory optimization results. Use the optimization framework proposed in the above steps S101 to S104 for data-driven hierarchical collaborative optimization. The computational grids are as Figure 3 and Figure 4 shown, and two sets of polyhedral grids with different levels of refinement are adopted.
[0056] The fine grid takes into account the viscous effect of the fluid and sets 20 layers of boundary layers on the wing near-wall surface. The optimized design parameters are shown in Table 1. Specifically, in the first stage, an inviscid solver based on the Euler equation is used for aerodynamic analysis on a rough grid. In contrast, in the second stage, high-fidelity flow analysis uses a fine grid and the k-ω SST turbulence model to solve the N-S equations. The CETLBO algorithm is used for optimization in both stages.
[0057] Table 1 Multi-fidelity hierarchical optimization design parameters for wing shape design
[0058] Parameter Stage1 Stage2 Solution method Euler N-S equation Turbulence model - k-ω SST Number of grids 300,000 1,100,000 Sampling point 1000 - Population size 66×2 50+82 Number of iterations 100 (10+10)×3 Number of design variables 66 25+41 CPU model EPYC 9554 EPYC 9554
[0059] The present invention uses the SHAP method to perform interpretive analysis on the LightGBM machine learning model. By extracting design knowledge, it can provide a basis for the decomposition of variables. First, in the first stage, low-fidelity data-driven optimization design is carried out on all parameters of the wing. The prediction ability of LightGBM is related to the accuracy of optimization. If the accuracy is too low, it will mislead the search direction of the algorithm. Therefore, the accuracy of the trained LightGBM model is verified. Figure 5 shows the variation of predicted values and true values on the test set, and the variation trend of the predicted values is similar to that of the true values. The R 2 and RMSE calculation results of the prediction model are 0.861 and 0.153 respectively, meeting the requirements of engineering applications.
[0060] The design knowledge provided by the LightGBM algorithm mainly includes the importance degree and variation law of design variables. The present invention ranks all design variables according to the average SHAP value, and this value reflects the features that have the greatest influence on the objective function. In the optimization process, this type of design knowledge is the aspect that we focus on and are interested in. <( Figure 6(a) shows the contribution ranking and relative influence degree of each variable on the lift-drag ratio of the wing. Each airfoil section is controlled by 5th-order CST parameters, but their contributions are significantly different. "Sec2_l1" is a feature with the largest contribution to the lift-drag ratio. The top ten design variables are mainly distributed in sections 0 to 4, and the variables near the leading edge of the airfoil have a greater contribution than other parts of the airfoil. Therefore, in the case of supersonic oncoming flow, the aerodynamic force and shock wave drag received by the wing root and the middle region of the wing are relatively large, which have an important impact on the performance and flight characteristics of the aircraft and are the key areas of concern in the optimization design.
[0061] In contrast, the overall contribution of the latter 41 variables is 0.07, and their influence on the objective function is relatively small. Figure 6 (b) further reveals the actual variation law between the variables and the prediction results. For example, a larger value of "Sec2_l1" has a positive SHAP value, that is, the points extending to the right side of the coordinate axis are getting redder, which indicates that the larger "Sec2_l1" is, the larger the lift-drag ratio is, showing a positive correlation. The relationship between "Sec3_u0" and the lift-drag ratio is exactly the opposite, showing a negative correlation. This helps designers extract the design rules, have an intuitive judgment on the design variables, and thus guide the optimization design.
[0062] In high-dimensional aerodynamic optimization, optimizing all variables as a whole is very inefficient. To improve the efficiency of wing aerodynamic optimization, considering the contribution information of the design variables comprehensively, the design variables are decomposed. The present invention divides the first 25 variables with the absolute value of SHAP greater than 0.01 and the remaining 41 variables into two groups, decomposing a high-dimensional optimization problem into two sub-optimization problems, laying a foundation for subsequent collaborative co-optimization.
[0063] The convergence process of the two-stage hierarchical collaborative optimization is as Figure 7 shown. The black line represents the result of high-fidelity aerodynamic optimization based on the genetic algorithm, and the colored curves represent the results obtained through multi-fidelity hierarchical collaborative optimization.
[0064] Obviously, in the first stage, the CETLBO algorithm converges rapidly on the constructed approximate model. The data-driven optimization method can quickly search for favorable regions, but the low-fidelity solution method results in a slightly larger lift-drag ratio. Subsequently, the high-fidelity cooperative co-optimization method is used to make minor adjustments to the wing shape to obtain more accurate results. Due to the introduction of prior knowledge and cooperative strategies, the optimization in the second stage converges with fewer iterations. The transition between the two stages causes the lift-drag ratio to instantaneously adjust from 12.76 to 8.68. In addition, the black line represents the traditional optimization method based on GA, and its convergence speed is significantly slower compared with the method proposed in this paper. The detailed comparison results are shown in Table 2. Thanks to the data-driven low-fidelity optimization in the first stage, the calculation time is significantly reduced, and the efficiency is improved by approximately 38.4%. At the same time, it generates a better aerodynamic shape, and the lift-drag ratio is increased by 8.3% compared with the optimization method based on the genetic algorithm.
[0065] Table 2 Comparison of Optimization Results of Different Aerodynamic Optimization Methods
[0066]
[0067] Under the supersonic cruise condition, the contour maps of the pressure coefficient on the upper surface of the wing of the optimized configuration and the initial configuration are as Figure 8 shown. By comparison, it is found that in the optimized wing, the low-pressure area in the outer wing section decreases, and the pressure peak in the inner wing section increases, resulting in a new low-pressure area (red dotted line area). This helps to increase the pressure difference between the upper and lower wing surfaces, thereby increasing the lift.
[0068] Figure 9 and Figure 10 compare the airfoil geometric shapes and pressure coefficient distributions of the typical cross-sections of the baseline wing and the optimized wing. The airfoil shapes of the optimized wing change significantly at different spanwise positions. Under the condition of meeting the volume constraint, the optimized wing has a strong motivation to reduce the supersonic drag by reducing the airfoil thickness. Figure 9 (a) shows a "negative camber" at the rear section of the airfoil, and the leading-edge radius is relatively large, which may be caused by meeting the volume constraint and lift characteristics during the optimization process. Except for the root airfoil, the leading edges of the airfoil shapes of other cross-sections become sharp, and the thickness is smaller, with obvious supersonic airfoil characteristics, which helps to reduce the shear effect of the airflow and reduce the wave drag of the wing. Supersonic wings usually adopt slender and thin airfoils, which can reduce the shock wave drag and drag peak, and improve the flight efficiency. Compared with subsonic airfoils, supersonic airfoils are very thin, with a smaller cross-sectional area and lower drag during supersonic flight. Generally speaking, the lift-drag ratio of the optimized wing shape is 9.21, and the aerodynamic performance in the supersonic state is improved by approximately 101.1% compared with the baseline shape.
[0069] A multi-fidelity two-stage hierarchical optimization model is constructed by combining a cooperative co-evolution optimization strategy. In the first stage, a low-fidelity dataset is first established, and the LightGBM machine learning algorithm is used to train and extract knowledge from the low-fidelity data, thereby grouping high-dimensional design variables. Secondly, the trained LightGBM model is used for data-driven low-fidelity optimization design. This can significantly reduce the number of high-fidelity CFD calculations and narrow the search space. In the second stage, cooperative co-evolution optimization is performed on each group of optimized design variables. Function tests show that the performance of the CETLBO algorithm is significantly better than that of traditional heuristic algorithms (GA, DE, PSO). By performing high-dimensional variable decomposition and hierarchical optimization design on the wing, compared with the traditional full-parameterization optimization design, the proposed method can reduce the difficulty of high-dimensional optimization search and accelerate the convergence speed.
[0070] The above disclosure is only one or more preferred embodiments of the present application, and cannot be used to limit the scope of rights of the present application. Those of ordinary skill in the art can understand all or part of the processes of implementing the above embodiments, and the equivalent changes made according to the claims of the present application still fall within the scope covered by the present application.
Claims
1. A data-driven hierarchical collaborative optimization method for aircraft, characterized in that It includes the following steps: Obtain a low-fidelity dataset; Based on the low-fidelity dataset, use the LightGBM machine learning model for regression prediction and evaluate the model accuracy through the coefficient of determination; Conduct feature importance analysis on the LightGBM machine learning model and group the design variables; Based on the grouped design variables, perform collaborative optimization to obtain the final optimization result.
2. The data-driven hierarchical collaborative optimization method for aircraft according to claim 1, wherein To obtain a low-fidelity dataset, the specific steps include: Generate uniformly distributed sampling points in the design space through the Latin hypercube sampling method, and use the low-fidelity flow analysis method to conduct aerodynamic analysis on the airfoil or wing to obtain the initial low-fidelity dataset.
3. The data-driven hierarchical collaborative optimization method for aircraft according to claim 2, characterized in that When obtaining the low-fidelity dataset: The number of initial sampling points is 10 - 15 times the number of design variables.
4. The data-driven hierarchical collaborative optimization method for aircraft according to claim 2, wherein Based on the low-fidelity dataset, use the LightGBM machine learning model for regression prediction and evaluate the model accuracy through the coefficient of determination. The specific steps include: Use the initial low-fidelity dataset to train the LightGBM machine learning model for regression prediction; Calculate the coefficient of determination R of the LightGBM machine learning model 2 to evaluate the model accuracy; When R 2 > 0.8, the low-fidelity optimization design is carried out using the CETLBO algorithm based on the trained LightGBM model; When R 2 ≤ 0.8, supplementary sampling points are added and the LightGBM model is retrained until the accuracy requirement is met.
5. The data-driven hierarchical collaborative optimization method for aircraft according to claim 4, wherein Conduct feature importance analysis on the LightGBM machine learning model and group the design variables. The specific steps include: Use the SHAP method to conduct feature importance analysis on the LightGBM machine learning model; Group the design variables according to the SHAP values and physical meanings, and decompose the high-dimensional optimization problem into multiple low-dimensional sub-optimization problems.
6. The data-driven hierarchical collaborative optimization method for aircraft according to claim 5, characterized in that Based on the grouped design variables, perform collaborative optimization to obtain the final optimization result. The specific steps include: Based on the grouped design variables, use the CETLBO algorithm combined with high-fidelity CFD to solve the N - S equations for collaborative optimization to obtain the final optimization result.