Small sample learning-based axial flow compressor flow optimization method

Abstract

The invention provides an axial flow compressor flow optimization method based on small sample learning. The axial flow compressor flow optimization method comprises the following steps that an optimization algorithm based on AI / ML is designed; constructing a model based on an optimization algorithm; an optimization design program capable of flexibly adjusting optimization parameters is developed; and the through-flow optimization function of the axial flow compressor is achieved. According to the axial flow compressor through-flow optimization method based on small sample learning, the optimal algorithm or algorithm combination is screened and designed by researching and analyzing a multi-objective and multi-disciplinary meta-heuristic method and a modern AI optimization algorithm, a global optimization scheme is found with fewer iterations, the efficiency and effect of the optimization process are improved, and the method is suitable for large-scale popularization and application. A high-dimensional model, low-quality data modeling and a reduced-order model are utilized to quickly update the technology, a high-precision proxy model is established by utilizing limited sample points, the influence of noise interference and sample quality on model precision is reduced, a through-flow optimization function is realized, and the overall performance of the compressor design is improved by utilizing the optimization method.

Description

Technical Field

[0001] The present invention relates to the field of axial flow compressor flow calculation optimization, and in particular to an axial flow compressor flow optimization method based on small sample learning based on artificial intelligence (AI) and machine learning (ML). Background Art

[0002] At present, the development of compressor flow calculation programs has provided an effective tool for analyzing and evaluating the performance of axial flow compressors. However, the existing flow software lacks fast multi-objective optimization functions, making it difficult to achieve performance optimization and automatic optimization of loss models. For complex simulation-based optimization problems, facing the challenges of long simulation times and expensive constraints, CAE simulation methods require iterative simulation calls to find the optimal solution, which has high computational costs. By combining modern AI optimization algorithms with flow calculation methods, efficient solutions to complex optimization problems can be achieved. This is conducive to avoiding the situation in which excessive constraints lead to optimization loss and excessive freedom leads to significant changes in aerodynamic layout when performing full three-dimensional optimization through CFD, thereby ensuring the rationality and optimization efficiency of the axial flow compressor flow design scheme.

[0003] However, when faced with complex optimization problems such as large variables, black boxes, and strong constraints, we are usually faced with problems such as how to use limited sample points to establish a high-precision proxy model, reduce the impact of sample quality on model accuracy, and quickly update the model. These are key technologies that need to be solved in modern optimization. Summary of the invention

[0004] In view of the above problems, the present invention proposes an axial compressor flow optimization method based on small sample learning, which can significantly improve the optimization efficiency and overcome the limitations of the prior art.

[0005] To achieve the above object, the present invention provides an axial flow compressor flow optimization method based on small sample learning, comprising the following steps:

[0006] (1) Design optimization algorithms based on AI / ML;

[0007] (2) Build a model based on optimization algorithm;

[0008] (3) Develop an optimization design program with flexible adjustment of optimization parameters;

[0009] (4) Realize the flow optimization function of the axial flow compressor.

[0010] Further: the AI / ML-based optimization algorithm designed in step (1) includes:

[0011] Step (1.1), research meta-heuristic algorithms and modern AI optimization algorithms, and expand and adjust these algorithms to be applicable to multi-objective and multi-disciplinary optimization problems; the meta-heuristic methods include genetic algorithms, particle swarm algorithms and gray wolf algorithms, and the AI ​​optimization algorithm is a modern optimization algorithm that combines artificial neural networks and gray wolf optimization algorithms; the research content covers the basic theories and implementation methods of these algorithms, as well as their application effect evaluation in multi-objective and multi-disciplinary optimization problems;

[0012] Step (1.2): select appropriate metaheuristics and modern AI optimization algorithms, evaluate the effectiveness of the algorithms and screen out the best algorithm or algorithm combination, effectively combine it with the flow calculation method, and design an intelligent optimization algorithm based on artificial intelligence / machine learning to optimize and improve the flow performance of the axial flow compressor.

[0013] Further: the step (2) of constructing a model based on an optimization algorithm includes:

[0014] One or more surrogate models between design variables and optimization objectives and constraints are constructed and verified, and a meta-heuristic optimization algorithm is used as a design space exploration strategy to update the surrogate model when new samples are generated and continue the optimization.

[0015] Furthermore: the model based on the optimization algorithm constructed in step (2) includes a high-dimensional model for large-scale complex problems to solve the problem of establishing a high-precision proxy model for high-dimensional black box problems under finite samples; noise interference data modeling to reduce the impact of data noise interference and sample quality on model accuracy; and a reduced-order model to quickly update the model to adapt to new data and needs.

[0016] Further: the step (3) of developing an optimization design program with the ability to flexibly adjust optimization parameters includes:

[0017] a. The program can automatically add the required variables as design variables and set the variable type, initial value, upper and lower limits;

[0018] b. The throughflow calculation result can be used as a parameter through the calculator component, and then converted into a variable range to express the constraint condition;

[0019] c. Set the objective function to be optimized according to the target trend, guide the optimization process, and perform single-objective optimization or multi-objective optimization on the objective function;

[0020] d. Set Penalty as a penalty term to indicate the feasibility of the current solution, which is determined by comparing it with the previous solution and whether it violates the constraints;

[0021] The program allows designers to adjust design variables, customize related constraints, and set related optimization goals.

[0022] Further: the step (4) of achieving the axial flow compressor flow optimization function includes:

[0023] Optimize the flow aerodynamic model; the optimization program can call the flow solver to achieve the optimization of the axial flow compressor flow calculation.

[0024] Further: the step (4) specifically comprises:

[0025] Step (4.1), using high-precision data from experiments or simulations, optimizing the flow aerodynamic models including various loss models, lagging angle models and blockage models;

[0026] Step (4.2), using the optimization program to call the flow solver through the flow aerodynamic model to optimize the flow aerodynamic performance.

[0027] The beneficial effects of the present invention are as follows: the intelligent optimization algorithm described in the present invention can effectively solve multi-objective and multi-disciplinary optimization problems, find a global optimization solution with fewer iterations, support repeated verification and re-learning, and improve the efficiency and effect of the optimization process. The high-dimensional model, noise interference data modeling, and reduced-order model can quickly update information, effectively solve problems such as large variables, high-dimensional black boxes, and strong constraints, and improve the reliability of the optimization results and the flexibility of the optimization process. The present invention realizes the flow optimization of the axial compressor and uses the intelligent optimization method to improve the overall performance of the compressor design. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a flow chart of an axial compressor flow optimization method based on small sample learning according to the present invention;

[0029] Figure 2 The basic flow chart for intelligent optimization of the present invention;

[0030] Figure 3 The optimization flow chart of the functional throughflow model of the present invention is optimized;

[0031] Figure 4 This is a flow chart for optimizing the functional flow and aerodynamic layout of the present invention;

[0032] Figure 5 A flow chart of optimizing and improving compressor flow calculation results using an intelligent optimization algorithm based on artificial intelligence / machine learning in the present invention;

[0033] Figure 6 This is a chart showing the calculation results of the spanwise distribution of the total temperature, total pressure, meridian velocity and airflow angle along the blade height direction of the first row of stationary blades before and after optimization in Example 2 of the present invention.

[0034] Figure 7This is a chart showing the calculation results of the spanwise distribution of the total temperature, total pressure, meridian velocity and airflow angle along the blade height direction of the 4th row of stator blades before and after optimization in Example 2 of the present invention.

[0035] Figure 8 This is a comparison chart of the calculation results of the spanwise distribution of the total temperature, total pressure, meridian velocity and airflow angle along the blade height direction of the 9th row of stator blades before and after optimization in Example 2 of the present invention. DETAILED DESCRIPTION

[0036] To facilitate understanding of the objectives, technical means and advantages of the present invention, as well as to achieve the objectives and effects, the present invention will be further described in detail below in conjunction with specific implementation methods and with reference to the accompanying drawings.

[0037] The present invention provides an axial flow compressor flow optimization method based on small sample learning, comprising the following steps:

[0038] like Figure 1 As shown in the figure, an intelligent optimization algorithm based on artificial intelligence / machine learning (AI / ML) is designed, a model based on the optimization algorithm is constructed, and an optimization design program with flexible adjustment of optimization parameters is developed to finally realize the flow optimization function of the axial compressor. The design of an intelligent optimization algorithm based on AI / ML specifically includes the following steps:

[0039] (1) Research meta-heuristic algorithms and modern AI optimization algorithms, and expand and adjust these algorithms to apply to multi-objective and multi-disciplinary optimization problems. The meta-heuristic methods include genetic algorithms, particle swarm algorithms, gray wolf algorithms, etc. The AI ​​optimization algorithm is a modern optimization algorithm that combines artificial neural networks and gray wolf optimization algorithms.

[0040] (2) Select appropriate metaheuristics and modern AI optimization algorithms, evaluate the effectiveness of the algorithms and select the best algorithm or algorithm combination, effectively combine it with the flow calculation method, and design an intelligent optimization algorithm based on artificial intelligence / machine learning to optimize and improve the performance of compressor flow calculation. Use the metaheuristic optimization algorithm as a design space exploration strategy to continuously generate new samples, update the proxy model, and continue to optimize.

[0041] In the optimization program, the design variable is the independent variable, and the optimization result is achieved by changing the value of the design variable. The design variable is also the parameter of the optimization component, and its data type can be real number Real, integer Integer, etc. The program can automatically add the required variables as design variables, and set the variable type, initial value, upper and lower limits.

[0042] Constraints refer to the limit constraints on output parameters. Constraints are expressed in the range of parameter values. Constraints often appear in the form of inequalities. In the optimization design program, the calculator component can be used to convert the flow calculation results into a variable range as a parameter to express the constraints.

[0043] The optimized objective function is the optimization target. If the objective function needs to be continuously reduced during the optimization process, the objective trend is set to minimize, and the weighted sum of each objective is directly calculated. On the contrary, when the objective trend is set to maximize, the opposite values ​​of each objective function are weighted and summed to guide the entire optimization process. Penalty is used as a penalty term to indicate the feasibility of the current solution.

[0044] like Figure 2 As shown, firstly, random sampling is performed in the sample space, simulation analysis is performed based on the sampled data, and the results are recorded in the database. Problem identification and machine learning are performed on the results of the sampled analysis to perform knowledge discovery and refinement on the optimization problem. If the conditions are met, the optimization is stopped and the optimal solution is returned. If not, the knowledge obtained by machine learning is used to continue sampling in the sample space and repeat the above steps. In the process of optimization, machine learning is used to learn the different global sensitivities of variables, the correlations between variables, and the effective value ranges of variables. In the process of continuous learning, the accuracy of machine learning is continuously improved, and the learned knowledge is repeatedly verified and relearned. The final realization includes but is not limited to compressor flow optimization, and the overall performance of the compressor design is improved using intelligent optimization methods.

[0045] Example 1

[0046] like Figures 1 to 4 As shown, a flow optimization method for an axial compressor based on small sample learning includes the following steps:

[0047] (1) Build an optimization algorithm for AI / ML based on small samples;

[0048] (2) Build a model based on optimization algorithm;

[0049] (3) Develop an optimization design process with flexible adjustment of optimization parameters;

[0050] (4) Combine optimization algorithms and models to achieve compressor flow optimization.

[0051] Furthermore, in step (1), an optimization algorithm based on AI / ML is designed, which specifically includes the following two aspects:

[0052] Step (1.1), meta-heuristic algorithms and modern AI optimization algorithms are studied, and these algorithms are expanded and adjusted to be applicable to multi-objective and multi-disciplinary optimization problems. The meta-heuristic methods include genetic algorithms, particle swarm algorithms and gray wolf algorithms, and the AI ​​optimization algorithm is a modern optimization algorithm that combines artificial neural networks and gray wolf optimization algorithms; the research content covers the basic theories, implementation methods, and evaluation of their application effects in multi-objective and multi-disciplinary optimization problems. Based on the current optimization methods, a novel meta-model-based optimization method is proposed. This method directly aims to reduce the number of objective function and constraint evaluations. This method is based on the existing mode pursuit sampling method and enhances performance through two strategies: (1) generating more sample points near promising areas; (2) using constraint information to bias the generation of sample points towards feasible areas. A new optimization sampling technique is proposed, which systematically uses incremental information about constraints to bias the generation of sample points towards the feasible area of ​​the problem domain, and is named constraint importance mode pursuit optimization technique. The main implementation algorithm flow of this method is as follows:

[0053] (a) Input function:

[0054] Input: f(x) black box function; G = {gk(x)≤0, k = 1, ..., K} a set of constraints; Problem domain.

[0055] (b) Output:

[0056] xmin The global minimum of f(x), or NULL on failure.

[0057] (c) Calculation process:

[0058]

[0059] The algorithm contains the following process:

[0060] Process 1. Initial sampling: Generate a set of m sample points X, which are randomly generated in the feasible domain D of the problem, but no longer perform constraint checks on each sample point. These points are called expensive points because their function values ​​are evaluated by the black box function f(x).

[0061] Process 2. Function approximation and guide function definition: Use linear spline to fit the expensive points in X and create the metamodel f^(x). Define a sampling guide function h(x) = c0-f^(x), where c0 is a constant, ensuring that c0 ≥ max{f^(x)}. This function will guide the generation of sample points to concentrate them in areas with lower function values.

[0062] Process 3. Pattern pursuit sampling and constrained importance sampling: Generate N cheap sample points XN, which are randomly generated in the problem domain without constraint checking, and then use the bootstrap function h(x) to evaluate their function values. Use discriminative sampling techniques to extract m sample points Xm from XN, which tend to be concentrated near the mode of h(x). Then perform constraint checks on the sample points in Xm, and assign a penalty value vpenalty to sample points that do not meet the constraints. This value can be static or dynamic.

[0063] Process 4. Quadratic regression and local optimization: Perform quadratic regression on the sample points in Xm. If the regression is accurate enough, perform local optimization in this subregion to obtain the minimum value xmin. If xmin is in a certain subregion around the pattern of h(x), return it as the global minimum of f(x). Otherwise, add xmin to the expensive point set X and restart the algorithm from process 2.

[0064] Furthermore, in step (1.2), a compressor flow analysis model based on the optimization algorithm is constructed, and the implementation steps are as follows:

[0065] The above-mentioned AI optimization algorithm based on small samples is selected, the effect of the algorithm is evaluated and the best algorithm or algorithm combination is screened out, which is effectively combined with the compressor flow calculation method to design an intelligent optimization algorithm based on artificial intelligence / machine learning to achieve the purpose of optimizing and improving the compressor flow calculation results. The specific construction process is as follows: Figure 5 As shown, the algorithm is described as follows:

[0066] Operation 1: Initial random sampling:

[0067] Generate N uniform random points X_wolf.

[0068] Evaluate the samples through simulation, [F_wolf, G_wolf]←Simulation(X_wolf). Store the initial samples in the sample library, [X,F,G]←[X_wolf,F_wolf,G_wolf]. Operation 2: Pareto set determination:

[0069] If a feasible solution exists, determine the current Pareto set, [XP, FP, Gp].

[0070] Otherwise, constraint processing operation 1 is performed to generate a Pareto set.

[0071] Operation 3: Dimension selection:

[0072] Use LASSO to split the variables into two subsets: one for the objective, X_var_obj, and the other for the constraints, X_var_con.

[0073] Operation 4: ANN approximation:

[0074] Train the target ANN on X_var_obj.

[0075] Train a constrained ANN for all variables X.

[0076] Operation 5: Location Update:

[0077] The positions of selected variables are updated through the Grey Wolf Algorithm (GWO) update mechanism,

[0078] X_new_wolf_var_obj←GWO(X_wolf_var_obj,XP).

[0079] Operation 6: Constraint Operation 2:

[0080] Perform constraint operation 2 to generate a feasible sample set,

[0081] X_new_c_wolf=[X_newc_wolf_var_obj,X_new_c_wolf_var_con].

[0082] Output the feasible solution set X_newc_fe.

[0083] Operation 7: Determination of non-dominated solutions:

[0084] if Find the non-dominated solution based on the target prediction value of ANN, X_non-dom←findNonDom(X_newc_fe,ANN_obj(X_newc_fe)).

[0085] Otherwise, randomly select a sample from X_new_c_wolf, X_non-dom←randSelect(X_new_c_wolf).

[0086] Evaluate the non-dominated solution through simulation, [X_non-dom,F_non-dom,G_non-dom]←simulation(X_non-dom).

[0087] Store the samples in the sample library, [X,F,G]←[[X,F,G];[X_non-dom,F_non-dom,G_non-dom]].

[0088] Operation 8: Pareto set update:

[0089] Check the feasibility of X_non-dom through the actual constraint value G_non-dom, X_non-dom_fe←checkFeasible(X_non-dom,G_non-dom).

[0090] if Update the Pareto set, [XP, FP, Gp]←updatePareto([XP, FP, Gp], [X_non-dom_fe, F_non-dom_fe, G_non-dom_fe]).

[0091] Otherwise, the Pareto set remains unchanged.

[0092] Operation 9: Constraint Operation 3:

[0093] Constraint operation 3 is performed to check the current Pareto set.

[0094] Action 10: Stop criteria checking:

[0095] If the number of simulation evaluations is greater than the preset maximum number of function evaluations, the optimization is terminated; otherwise, return to operation 3.

[0096] One or more surrogate models between design variables and optimization objectives and constraints are constructed and verified, and a meta-heuristic optimization algorithm is used as a design space exploration strategy to update the surrogate model when new samples are generated and continue the optimization.

[0097] The improvements of the present invention are: using high-dimensional models of large-scale complex problems to solve the problem of establishing a high-precision proxy model for high-dimensional black box problems using limited sample points; modeling noise interference data to reduce the impact of data noise interference and sample quality on model accuracy; and reducing the order model to quickly update the model to adapt to new data and needs.

[0098] Furthermore, in step (3), an optimization design program with the ability to flexibly adjust optimization parameters is developed, wherein the optimization parameters include:

[0099] Design variables, which refer to variable parameters used as input parameters in optimization problems. The types of design variables can be real numbers (Real), integers (Integer), etc.;

[0100] Optimization constraints, which refer to the limits of the output parameters and the range of variation of the design variables, and are often in the form of inequalities;

[0101] The target variable represents the parameter of the optimization target, that is, the performance parameter that needs to be maximized or minimized during the optimization process. The target variable can be an output parameter or an input parameter.

[0102] The flexible adjustment of optimization parameters means that the program allows designers to flexibly adjust design variables, customize related constraints, and set related optimization goals. The optimization program specifically includes:

[0103] a. The program can automatically add the required variables as design variables, and set the variable type, initial value, upper and lower limits;

[0104] b. The throughflow calculation result can be used as a parameter through the calculator component, and then converted into a variable range to express the constraint condition;

[0105] c. Set the objective function according to the target trend to guide the optimization process. It should be pointed out that the essential difference between single-objective optimization and multi-objective optimization problems is the number of objective functions.

[0106] d. Set Penalty as a penalty term to indicate the feasibility of the current solution, which is determined by comparing it with the previous solution and whether it violates the constraints.

[0107] During the optimization process, machine learning is used to learn the different global sensitivities of variables, the correlations between variables, and the effective value ranges of variables. In the process of continuous learning, the accuracy of machine learning is continuously improved, and the learned knowledge is repeatedly verified and re-learned.

[0108] Furthermore, in step (4), the flow optimization function of the axial compressor is realized, which specifically includes:

[0109] Step (4.1), using high-precision data from experiments or simulations, optimizing the flow aerodynamic models including various loss models, lagging angle models and blockage models;

[0110] Step (4.2), using the optimization program to call the flow solver through the flow aerodynamic model to optimize the flow aerodynamic performance, such as the pressure ratio, aerodynamic efficiency, etc.

[0111] Example 2

[0112] Flow optimization includes two parts. One is to use high-precision data from experiments or simulations to optimize models including various loss models, lagging angle models, blockage models, radial mixing models, etc.; the other is to use intelligent optimization methods to optimize the flow aerodynamic layout.

[0113] like Figure 3As shown in the figure, various coefficients including various loss models, lag angle models, blockage models, radial mixing models and other models are used as design variables of the optimization problem, and relevant constraints are defined. The root mean square error between the compressor performance curve obtained by flow calculation and the performance curve obtained by experiment or simulation is used as the optimization target, and then the intelligent optimization algorithm engine is called to find its minimum value. After the optimization is completed, the coefficients of the model are output as the optimized model for flow calculation.

[0114] Step 1: Add the corresponding correction coefficient to each item of the loss calculation formula of the axial flow compressor, as follows:

[0115]

[0116] Where m total , m1, m2, m3, m4 are correction coefficients, is the minimum loss of the basic model, is the non-design condition loss of the basic model, As the basic model shock wave loss, Add loss to the base model.

[0117] Step 2: Add corresponding correction coefficients to each item of the axial compressor backward angle calculation formula, as follows:

[0118] δ=m total (m1·δ*+m2·δ correct +m3·δ i )

[0119] Among them, m1, m2, m3 are correction coefficients, δ * Design the lag angle for the base model, δ correct is the corrected lag angle for the basic model, δ i Non-designed lag angle for the base model.

[0120] Step 3: Divide all components (IGV, rotor and stator) of a 9-stage axial compressor into two groups, the first group is the transonic group (IGV and the rotor and stator of the first two stages), and the other group is the subsonic group (rotor and stator of the last seven stages), where the components in the same group use the same loss model and lagging angle model coefficient;

[0121] Step 4: Use the loss model and lag angle model coefficients of different groups as optimization input variables, and set their respective upper and lower limits;

[0122] Step 5: The absolute value of the difference between the total pressure ratio and the total isentropic efficiency results of the flow calculation and the high-precision data is used as the optimization target variable, that is, abs(PRtt_2d-PRtt_standard) and abs(ETAtt_2d-ETAtt_standard), and the minimum value of the optimization target variable is set, where the high-precision data can come from the three-dimensional CFD calculation results;

[0123] Step 6: Run the optimizer, call the throughflow program calculation during the optimization process, return the optimization target result, predict the next optimization input based on the optimization target result, and repeat this process until the optimization ends or the maximum number of optimizer runs is reached. The optimizer can determine whether the loss model and lag angle model coefficient calculation of the optimizer input failed by obtaining the return value of the throughflow program calculation. If the calculation fails, the invalid point will be automatically discarded.

[0124] Step 7: After the optimization is completed, check the Pareto chart to select the optimal point for multi-objective optimization.

[0125] The performance chart of the whole machine before and after optimization is as follows Figure 6 , Figure 7 and Figure 8 As shown, Figure 6 , Figure 7 and Figure 8 The following are the comparison charts of the calculation results of the total temperature, total pressure, meridian velocity and airflow angle along the blade height direction of the 1st, 4th and 9th rows of stator blades before and after optimization. In the chart: the solid line represents the input high-precision data results, and the dotted line represents the calculation results after optimization. The comparison of the calculation results is shown in the following table:

[0126]

[0127] It can be seen that the present invention effectively solves the problems of large variables, high-dimensional black boxes, strong constraints, etc., and improves the reliability of the optimization results and the flexibility of the optimization process. The present invention realizes the flow optimization of the axial compressor, and the error after optimization is greatly reduced, thereby improving the overall performance of the compressor design.

Claims

1. A method for optimizing flow through an axial compressor based on small sample learning, characterized in that: The following steps are involved: (1) Design optimization algorithms based on AI / ML; (2) Build a model based on optimization algorithm; (3) Develop an optimization design program with flexible adjustment of optimization parameters; (4) Realize the flow optimization function of the axial flow compressor.

2. The method for optimizing flow through an axial compressor based on small sample learning according to claim 1, characterized in that: The AI / ML-based optimization algorithm designed in step (1) includes: Step (1.1), studying meta-heuristic algorithms and modern AI optimization algorithms, and expanding and adjusting these algorithms to be applicable to multi-objective and multi-disciplinary optimization problems; the meta-heuristic methods include genetic algorithms, particle swarm algorithms and gray wolf algorithms, and the AI ​​optimization algorithm is a modern optimization algorithm that combines artificial neural networks and gray wolf optimization algorithms; Step (1.2): select appropriate metaheuristics and modern AI optimization algorithms, evaluate the effectiveness of the algorithms and screen out the best algorithm or algorithm combination, effectively combine it with the flow calculation method, and design an intelligent optimization algorithm based on artificial intelligence / machine learning to optimize and improve the flow performance of the axial flow compressor.

3. The method for optimizing flow through an axial compressor based on small sample learning according to claim 1, characterized in that: The model based on the optimization algorithm in step (2) includes: One or more surrogate models between design variables and optimization objectives and constraints are constructed and verified, and a meta-heuristic optimization algorithm is used as a design space exploration strategy to update the surrogate model when new samples are generated and continue the optimization.

4. The method for optimizing flow through an axial compressor based on small sample learning according to claim 3, characterized in that: It includes high-dimensional models for large-scale complex problems to solve the problem of establishing high-precision proxy models for high-dimensional black box problems under finite samples; noise interference data modeling to reduce the impact of data noise interference and sample quality on model accuracy; reduced-order models to quickly update models to adapt to new data and needs.

5. The method for optimizing flow through an axial compressor based on small sample learning according to claim 1, characterized in that: The optimization design program developed in step (3) with the ability to flexibly adjust optimization parameters includes: a. The program can automatically add the required variables as design variables and set the variable type, initial value, upper and lower limits; b. The throughflow calculation result can be used as a parameter through the calculator component, and then converted into a variable range to express the constraint condition; c. Set the objective function to be optimized according to the target trend, guide the optimization process, and perform single-objective optimization or multi-objective optimization on the objective function; d. Set Penalty as a penalty term to indicate the feasibility of the current solution, which is determined by comparing it with the previous solution and whether it violates the constraints; The program allows designers to adjust design variables, customize related constraints, and set related optimization goals.

6. The method for optimizing flow through an axial compressor based on small sample learning according to claim 1, characterized in that: The function of optimizing the flow through the axial compressor in step (4) includes: Optimize the flow aerodynamic model; the optimization program can call the flow solver to achieve the optimization of the axial flow compressor flow calculation.

7. The method for optimizing flow through an axial compressor based on small sample learning according to claim 6, characterized in that: Step (4) specifically includes: Step (4.1), using high-precision data from experiments or simulations, optimizing the flow aerodynamic models including various loss models, lagging angle models and blockage models; Step (4.2), using the optimization program to call the flow solver through the flow aerodynamic model to optimize the flow aerodynamic performance.

Images