A centrifugal impeller design verification method applied to a small-power civil aviation engine

Abstract

The application discloses a centrifugal impeller design verification method applied to a small-power civil aviation engine, which is characterized in that: a three-dimensional model of the centrifugal impeller is decomposed into two-dimensional air channel boundaries, blade thickness distribution, blade angle distribution and other elements, and continuous data points are reduced in dimension to extract characteristic values, and meanwhile, other overall performance characteristics such as rated rotating speed, inlet total temperature, inlet total pressure and outlet static pressure are combined to train a BP neural network, and then pressure ratio, flow and efficiency are outputted. The method can be used for the front quick verification calculation of the centrifugal impeller design of the small-power civil aviation engine, can quickly screen out the centrifugal impellers that cannot meet the design requirements, and can reduce the workload of three-dimensional flow field simulation, so that the development efficiency of the centrifugal impeller is improved.

Description

Technical field:

[0001] The invention relates to a centrifugal impeller design verification method applied to a low-power civil aviation engine, and belongs to the field of impeller machinery design such as civil aviation engines, gas turbines and fans. Background technology:

[0002] Centrifugal impellers offer high air compression ratios and are widely used in civil aviation engines and gas turbines. The design efficiency of a centrifugal impeller is highly correlated with the designer's engineering experience in designing elements such as the compressor flow path, blade thickness, and blade angle.

[0003] The current main approaches to the design of centrifugal compressors are as follows: one-dimensional and meridional flow channel design, and three-dimensional blade design. After years of development, one-dimensional and meridional flow channel design has accumulated rich design experience. Based on test results and design experience, the airflow loss in the centrifugal compressor is subdivided into various forms such as surface friction loss, blade load loss, leakage loss, backflow loss, separation loss, shock wave loss, etc., and corresponding empirical formulas are given to make the design process as accurate as possible. Three-dimensional blade design refers to the initial blade shape design based on one-dimensional and meridional flow channel design by giving the root tip blade mid-arc blade angle distribution and thickness distribution, and then performing iterative design based on the three-dimensional flow field simulation results. If the simulation results show that the one-dimensional design parameters are not selected reasonably, it is necessary to return to the one-dimensional simulation and re-given the basic design parameters. Otherwise, it is directly iterated by the three-dimensional blade shape parameters until the centrifugal impeller aerodynamic performance meets the design requirements and the flow field distribution is reasonable. The current approach has the following main disadvantages: (1) The design and verification process is long, which affects the design efficiency; (2) It is highly dependent on the engineering experience of the designer, and the design method is difficult to expand and promote.

[0004] Therefore, it is necessary to improve the existing technology to solve the shortcomings of the existing technology. Summary of the invention:

[0005] To address the problems of the prior art, the present invention provides a design verification method for centrifugal impellers used in low-power civil aircraft engines. This method deconstructs a three-dimensional centrifugal impeller model into elements such as the two-dimensional airway boundary, blade thickness distribution, and blade angle distribution. This method then reduces the dimensionality of continuous data points to extract eigenvalues. Furthermore, the method combines other overall performance characteristics, such as rated speed, inlet total temperature, inlet total pressure, and outlet static pressure, to train a BP neural network and output pressure ratio, flow rate, and efficiency. This method can be used for rapid pre-calculation verification of low-power civil aircraft engine centrifugal impeller designs, quickly eliminating impellers that fail to meet design requirements, reducing the workload of three-dimensional flow field simulation, and ultimately improving the efficiency of centrifugal impeller development.

[0006] The application adopts the following technical scheme: a centrifugal impeller design verification method applied to a small-power civil aviation engine, and the specific steps are as follows:

[0007] Step one: for a small-power civil aviation engine, a large number of centrifugal impeller three-dimensional design models are generated, the pressure ratio π, the flow rate m and the efficiency η numerical values are output through three-dimensional flow field simulation, and training sample data are generated;

[0008] Step two: for each centrifugal impeller three-dimensional design model, the centrifugal impeller two-dimensional air passage boundary point position, the 100% blade height continuous blade thickness distribution, the 0% blade height continuous blade thickness distribution, the 100% blade height continuous angle distribution and the 0% blade height continuous blade angle distribution are decomposed;

[0009] Step three: the principal component analysis method is adopted to reduce the continuous data points of the centrifugal impeller two-dimensional air passage boundary point position, the 100% blade height continuous blade thickness distribution, the 0% blade height continuous blade thickness distribution, the 100% blade height continuous angle distribution and the 0% blade height continuous blade angle distribution to two dimensions, and the single variable characteristic rated speed N, the inlet total temperature T, the inlet total pressure P0 and the outlet static pressure P1 are added;

[0010] Step four: a BP neural network model is trained;

[0011] Step five: for subsequent centrifugal impeller three-dimensional design model data, the feature vectors selected from the early data set are used for dimension reduction, the feature values of the dimension-reduced centrifugal impeller sample and the rated speed, the inlet total temperature, the inlet total pressure and the outlet static pressure are input into the trained neural network function to calculate the results, and the pressure ratio, the flow rate and the efficiency are output as the rapid design evaluation before three-dimensional flow field simulation.

[0012] Further, step one is specifically as follows:

[0013] (1.1) samples are obtained by using a conventional centrifugal impeller design method;

[0014] (1.2) the three-dimensional flow field simulation output value array of sample 1 is [π1, m1, η1], denoted as Op (1) , wherein π1 is the pressure ratio of sample 1, m1 is the flow rate of sample 1, η1 is the efficiency of sample 1, and the three-dimensional flow field simulation output results of 100 samples form a matrix [Op (1) ; Op (2) ; ……; Op (100) ] T , denoted as Op.

[0015] Further, step two is specifically as follows:

[0016] (2.1) The two-dimensional air channel boundary of the centrifugal impeller design sample includes the upper casing surface air channel boundary and the lower hub surface air channel boundary, the casing surface air channel boundary two-dimensional curve is composed of 100 two-dimensional coordinates, 100 point coordinate samples are evenly taken in the i-th casing surface air channel boundary two-dimensional curve as denoted as A i , the hub surface air channel boundary two-dimensional curve is also composed of 100 two-dimensional coordinates, 100 point coordinate samples are evenly taken in the i-th hub surface air channel boundary two-dimensional curve as denoted as B i ;

[0017] (2.2) 100 points in the continuous blade thickness distribution at 0% blade height of the i-th sample of the centrifugal impeller design are evenly taken and represented by an array denoted as Ha i ; 100 points in the continuous blade thickness distribution at 100% blade height of the i-th sample are evenly taken and represented by an array denoted as Hb i ;

[0018] (2.3) 100 points in the continuous blade angle distribution at 0% blade height of the i-th sample of the centrifugal impeller design are evenly taken and represented by an array denoted as Aa i ; 100 points in the continuous blade angle distribution at 100% blade height of the i-th sample are evenly taken and represented by an array denoted as Ab i .

[0019] Further, step three is specifically as follows:

[0020] (3.1) For the two-dimensional coordinate group matrix A i and B i of the air channel boundary sample, the principal component analysis method is used to reduce to one dimension;

[0021] (3.2) The covariance matrix C of the matrix A i is calculated, the eigenvalues λ1, λ2 of C are calculated, and the eigenvalue λ k with a covariance contribution rate greater than 90% is selected;

[0022] (3.3) The eigenvalue corresponding to λ k is selected to form the characteristic transformation matrix P k ∈R 1*2 ;

[0023] (3.4) Dimension reduction transformation of the matrix A i : The result after dimension reduction is denoted as As i , and the matrix B iTake the same operation, get the feature transformation matrix P m ∈R 1*2 , B i Dimensionality reduction results recorded as Bs i .

[0024] (3.5) all A i (i = 1, 2 … … 100) after dimensionality reduction matrix [As1, As2 … … As 100 ] recorded as As;

[0025] (3.6) calculate the covariance matrix C1 of the matrix As, calculate the eigenvalue λ a1 , λ a2 … λ a100 , select the sum of the covariance contribution rate of two eigenvalues λ i and λ j ;

[0026] (3.7) select λ i and λ j corresponding eigenvector composition feature transformation matrix P k2 ∈R 2*100 ;

[0027] (3.8) dimensionality reduction transformation of matrix As: Thus the i-th casing surface airway boundary two-dimensional curve sample dimensionality reduction to two eigenvalues: Take the same operation for B i , get the feature transformation matrix P m2 , hub surface airway boundary two-dimensional curve sample to two eigenvalues:

[0028] (3.9) for the i-th sample of blade thickness and blade angle feature matrix Ha i , Hb i , Aa i , Ab i Repeat steps (3.6) ~ (3.8), respectively, get the feature transformation matrix P x1 , P x2 , P x3 , P x4 , after dimensionality reduction, the i-th sample of 0% blade height blade thickness distribution, 100% blade height blade thickness distribution, 0% blade height blade angle distribution and 100% blade height blade angle distribution are respectively reduced to eigenvalues: (3.10) so the two-dimensional design continuous characteristics of the i-th sample can be extracted as 12 eigenvalues, combined with the engine overall performance parameters rated speed N (i) , import total temperature T (i) , import total pressure outlet static pressure The design features of the centrifugal impeller sample i can be expressed by the following 16 feature parameter groups:

[0029]

[0030] denoted as Ps (i) , i = 1, 2, ……100, the sample set matrix [Ps (1) ; Ps (2) ; ……; Ps (100) ] T , denoted as Ps.

[0031] Further, step four is specifically as follows:

[0032] The dimensionality reduction data sample set Ps is taken as an input matrix, and the simulation result matrix Op is taken as an output matrix. A one-layer hidden layer network structure is adopted, and the hidden layer nodes are determined according to an empirical formula , wherein H is the number of hidden layer nodes, I is the number of input layer nodes, O is the number of output layer nodes, and H = 6 is selected.

[0033] The weight matrix between the input layer and the hidden layer is denoted as

[0034] The weight matrix between the hidden layer and the output layer is denoted as

[0035] The bias vector between the input layer and the hidden layer is denoted as B = (B1, B2, ……, B6) T ;

[0036] The bias vector between the hidden layer and the output layer is denoted as C = (C1, C2, C3) T ;

[0037] The weight matrices W (1) , W (2) and the bias vectors B and C are initialized and valued, and the initialization values are random numbers in the interval (0.5, 1);

[0038] The transfer function is selected as a Sigmoid function, that is,

[0039] The jth hidden layer value of the ith sample is calculated as

[0040]

[0041] , wherein is the jth row of the weight matrix W (1) , Ps (i) is the ith sample in the sample set Ps, and B j is the jth value in the bias vector B.

[0042] Similarly, calculate the j-th output layer value of the i-th sample

[0043]

[0044] in is the weight matrix W (2) The jth row, Hid (i) is the hidden layer output value vector of the i-th sample in the sample set, C j is the jth value in the bias vector C;

[0045] Refer to the output sample Op in the training sample and calculate the i-th sample Op (i) and its Output (i) The error E between (i) :

[0046]

[0047] in is the i-th sample Op (i) The kth value in is the i-th sample Output (i) The kth value in ;

[0048] The total (average) cost E of all training samples total for:

[0049]

[0050] The weights and biases are updated using batch gradient descent until the system error is minimized:

[0051]

[0052] Where μ is the learning rate, and the selection range is (0.05, 0.8)

[0053] The neural network model after training is recorded as f BP (x),x∈R 16*1 .

[0054] Furthermore, step five is as follows:

[0055] For the design of a new centrifugal impeller 3D model, deconstruct the 2D geometric elements;

[0056] 100 points are evenly and equidistantly selected on the curve, and the airway boundary data of the casing surface is recorded as A0∈R 100*2 , the hub surface airway boundary data is recorded as B0∈R 100*2, the blade angle at 0% blade height of the centrifugal impeller is recorded as Aa0∈R 100*1 , the blade angle at 100% blade height of the centrifugal impeller is recorded as Ab0∈R 100*1 , the blade thickness at 0% blade height of the centrifugal impeller is recorded as Ha0∈R 100*1 , the blade thickness at 100% blade height of the centrifugal impeller is recorded as Hb0∈R 100*1 ;

[0057] Data dimension reduction is performed on the above data:

[0058]

[0059] The overall performance parameters rated speed N (0) , total temperature T (0) , total pressure outlet static pressure are added to form a new input sample Ps (0) ∈R 16*1 :

[0060] wherein, is the characteristic value of the two-dimensional air channel boundary of the new sample after dimension reduction, is the characteristic value of the blade thickness distribution at 0% blade height and 100% blade height of the new sample after dimension reduction, is the characteristic value of the blade angle distribution at 0% blade height and 100% blade height of the new sample after dimension reduction;

[0061] The neural network function is calculated, f BP (Ps (0) )=[π0,m0,η0];

[0062] Whether the output pressure ratio π0, flow rate m0 and efficiency η0 meet the design requirements is evaluated, if yes, the three-dimensional flow field simulation verification is continuously performed, and if not, the design is re-performed.

[0063] The application has the following beneficial effects:

[0064] (1) The centrifugal impeller three-dimensional entity is expressed by continuous two-dimensional geometric data, the continuous two-dimensional data is uniformly and densely sampled, data dimension reduction is performed through principal component analysis, and the centrifugal impeller three-dimensional entity is finally expressed in a parameterized manner, the BP neural network model is trained based on the foregoing parameterized expression, the output result of the three-dimensional flow field simulation is fitted, and the design evaluation efficiency of the civil aviation engine compressor end designer is improved.

[0065] (2) The centrifugal impeller design verification method for small-power civil aviation engines can be applied to the future artificial intelligence large model participating in the auxiliary design of the compressor components of the civil aviation engine, and has great reference significance. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 A two-dimensional air passage boundary graph of the centrifugal impeller sample 1.

[0067] Figure 2 A blade thickness distribution graph of 0% blade height and 100% blade height of the centrifugal impeller sample 1.

[0068] Figure 3 A blade angle distribution graph of 0% blade height and 100% blade height of the centrifugal impeller sample 1.

[0069] Figure 4 A BP neural network training iteration flowchart.

[0070] Figure 5(a) 、 5(b) , 5(c), 5(d) are three-dimensional model and two-dimensional geometric element disassembly schematic diagram of centrifugal impeller.

[0071] Figure 6 A centrifugal impeller design verification method for small-power civil aviation engines is shown in the flowchart. DETAILED DESCRIPTION

[0072] The application will be further described below with reference to the drawings.

[0073] The centrifugal impeller design verification method for small-power civil aviation engines has the following specific steps:

[0074] Step one: for small-power civil aviation engines, especially 500kW below single-stage centrifugal compressor configuration small-power civil aviation engines, a large number of centrifugal impeller three-dimensional design models are generated, and the pressure ratio π, flow rate m and efficiency η values are output through three-dimensional flow field simulation to generate training sample data.

[0075] Step one is as follows:

[0076] (1.1) The sample is obtained by using the conventional centrifugal impeller design method, including: one-dimensional design of centrifugal impeller selected according to the key design parameter selection criterion, two-dimensional air passage profile selected according to the load distribution of the centrifugal impeller end wall area, three-dimensional blade aerodynamic shape selected according to the load distribution of the centrifugal impeller three-dimensional blade, and centrifugal impeller aerodynamic performance parameters obtained by means of three-dimensional flow field simulation.

[0077] (1.2) The three-dimensional flow field simulation output value array of sample 1 is [π1, m1, η1], denoted as Op (1), where π1 is the pressure ratio of sample 1, m1 is the flow rate of sample 1, and η1 is the efficiency of sample 1. The three-dimensional flow field simulation output results of 100 samples form the matrix [Op (1) ;Op (2) ;……;Op (100) ] T , denoted as Op.

[0078] Step 2: For each centrifugal impeller 3D design model, deconstruct the centrifugal impeller 2D airway boundary points, continuous blade thickness distribution at 100% blade height, continuous blade thickness distribution at 0% blade height, continuous angle distribution at 100% blade height and continuous blade angle distribution at 0% blade height. See the attached figure for details. Figure 1 , Attachment Figure 2 , Attachment Figure 3 For each deconstructed two-dimensional design element, 100 data points are evenly taken.

[0079] Step 2 is as follows:

[0080] (2.1) The two-dimensional airway boundary of the centrifugal impeller design sample is shown in the attached Figure 1 The above is the two-dimensional curve of the airway boundary of the casing surface, which consists of 100 two-dimensional coordinates. The sample of the two-dimensional curve of the airway boundary of the i-th casing surface (100 points are evenly taken) is A i The following is the two-dimensional curve of the hub surface airway boundary, which is also composed of 100 two-dimensional coordinates. The sample of the i-th hub surface airway boundary two-dimensional curve (100 points are evenly taken) is Denoted as B i .

[0081] (2.2) The blade thickness distribution of the centrifugal impeller design sample at 0% blade height and 100% blade height is shown in the attached figure. Figure 2 The upper curve is the continuous leaf thickness distribution at 0% leaf height (100 points are uniformly taken in the continuous distribution), and the i-th sample can be arrayed Indicated by Ha i The lower curve is the continuous leaf thickness distribution at 100% leaf height (100 points are uniformly taken in the continuous distribution). The i-th sample can be arrayed Indicated by Hb i .

[0082] (2.3) The blade angle distribution of the centrifugal impeller design sample at 0% blade height and 100% blade height is shown in the attached figure. Figure 3 The upper curve is the continuous leaf angle distribution at 0% leaf height (100 points are uniformly taken in the continuous distribution). The i-th sample can be arrayed as Indicated by Aa i; the lower curve is the continuous blade angle distribution at 100% span (100 points evenly taken in the continuous distribution), the i-th sample can be represented by an array denoted as Ab i .

[0083] Step three: using principal component analysis method, reduce the continuous data points of the centrifugal impeller gas passage boundary, 100% span and 0% span blade thickness distribution, 100% span and 0% span blade angle distribution to two dimensions, plus single variable characteristics rated speed N, import total temperature T, import total pressure P0, outlet static pressure P1, after dimension reduction, the characteristic value of a single centrifugal impeller sample has 16.

[0084] Step three is as follows:

[0085] (3.1) for the two-dimensional coordinate group matrix A i and B i of the gas passage boundary sample, use principal component analysis method to reduce to one dimension.

[0086] (3.2) specifically, calculate the covariance matrix C of matrix A i , calculate the eigenvalues λ1, λ2, select the eigenvalue λ k with covariance contribution rate greater than 90%.

[0087] (3.3) select the eigenvector corresponding to λ k , form the characteristic transformation matrix P k ∈R 1*2 .

[0088] (3.4) dimension reduction transformation of matrix A i : The result after dimension reduction is denoted as As i . The same operation is taken on matrix B i , the characteristic transformation matrix P m ∈R 1*2 is obtained, and the result after dimension reduction is denoted as Bs i .

[0089] (3.5) all A i (i = 1, 2 … … 100) after dimension reduction form the matrix [As1, As2 … … As 100 ], denoted as As.

[0090] (3.6) calculate the covariance matrix C1 of all sample matrix As, calculate the eigenvalues λ a1 , λ a2 …… λ a100 , select two eigenvalues λ i and λ j with the sum of covariance contribution rate greater than 90%.

[0091] (3.7) Selecting λ i and λ j The corresponding eigenvectors, constitute the feature transformation matrix P k2 ∈R 2*100 .

[0092] (3.8) Dimensionality reduction transformation on the data set As: Thus the i-th two-dimensional curve sample of the casing surface air passage boundary can be reduced to two eigenvalues: For B i Take the same operation, get the feature transformation matrix P m2 , the hub surface air passage boundary two-dimensional curve sample reduced to two eigenvalues:

[0093] (3.9) The matrix Ha i , Hb i , Aa i , Ab i of the blade thickness and blade angle of the i-th sample is repeated steps (3.6)-(3.8), respectively, to get the feature transformation matrix P x1 , P x2 , P x3 , P x4 . After dimensionality reduction, the blade thickness distribution at 0% blade height, the blade thickness distribution at 100% blade height, the blade angle distribution at 0% blade height and the blade angle distribution at 100% blade height of the i-th sample are respectively:

[0094] (3.10) Thus the two-dimensional design continuous features of the i-th sample can be extracted as 12 eigenvalues, combined with the engine overall performance parameters rated speed N (i) , inlet total temperature T (i) , inlet total pressure outlet static pressure The design features of the centrifugal impeller sample i can be expressed by the following 16 feature parameter groups:

[0095]

[0096] Denoted as Ps (i) , i = 1, 2, ……100. The sample set matrix [Ps (1) ; Ps (2) ; ……; Ps (100) ] T , denoted as Ps.

[0097] Step four: training BP neural network model.

[0098] Step four is as follows:

[0099] The dimensionality reduction data sample set Ps is used as the input matrix, and the simulation result matrix Op is used as the output matrix. Figure 4 .

[0100] Specifically, a hidden layer network structure is adopted, and the hidden layer nodes are calculated according to the empirical formula Determine, where H is the number of hidden layer nodes, I is the number of input layer nodes, and O is the number of output layer nodes. In this example, H = 6.

[0101] The weight matrix between the input layer and the hidden layer is recorded as

[0102] The weight matrix between the hidden layer and the output layer is recorded as

[0103] The bias vector between the input layer and the hidden layer is denoted as B = (B1, B2, ..., B6) T .

[0104] The bias vector between the hidden layer and the output layer is denoted as C = (C1, C2, C3) T .

[0105] For the weight matrix W (1) 、W (2) And the bias vectors B and C are initialized and assigned values, and the initialization values ​​are random numbers in the range of (0.5, 1).

[0106] The transfer function is the Sigmoid function, that is

[0107] Calculate the jth hidden layer value of the i-th sample

[0108]

[0109] in is the weight matrix W (1) The jth row of Ps (i) is the i-th sample in the sample set Ps, B j is the jth value in the bias vector B;

[0110] Similarly, calculate the j-th output layer value of the i-th sample

[0111]

[0112] in is the weight matrix W (2) The jth row, Hid (i) is the hidden layer output value vector of the i-th sample in the sample set, C j is the jth value in the bias vector C;

[0113] Referring to the output sample Op in the training sample, the error E between the i-th sample Op (i) and its Output (i) is calculated (i) :

[0114]

[0115] wherein is the k-th value in the i-th sample Op (i) , is the k-th value in the i-th sample Output (i) ;

[0116] The total (average) cost E of all training samples is: total

[0117]

[0118] The weights and biases are updated by using batch gradient descent method until the system error is minimized:

[0119]

[0120]

[0121] wherein μ is the learning rate, selected in the range interval (0.05, 0.8).

[0122] The neural network model after training is denoted as f BP (x), x∈R 16*1 .

[0123] Step five: for the subsequent centrifugal impeller three-dimensional design model data, the feature vectors selected in the early data set are used for dimension reduction, and the feature values of the reduced centrifugal impeller sample, the rated speed, the inlet total temperature, the inlet total pressure and the outlet static pressure are input into the trained neural network function to calculate the results, and the pressure ratio, the flow rate and the efficiency are output as the rapid design evaluation before three-dimensional flow field simulation.

[0124] Step five is as follows:

[0125] For the new centrifugal impeller three-dimensional model design (Fig. 5(a)), the two-dimensional geometric elements are deconstructed: the air passage (Fig. 5(b)), the blade angle (Fig. 5(c)), and the blade thickness (Fig. 5(d)).

[0126] 100 points are selected on the curve at equal intervals, the casing surface air passage boundary data is denoted as A0∈R 100*2 , and the hub surface air passage boundary data is denoted as B0∈R 100*2 ​, the blade angle at 0% blade height of the centrifugal impeller is denoted as Aa0∈R 100*1 , the blade angle at 100% blade height of the centrifugal impeller is denoted as Ab0∈R 100*1 , the blade thickness at 0% blade height of the centrifugal impeller is denoted as Ha0∈R 100*1 , the blade thickness at 100% blade height of the centrifugal impeller is denoted as Hb0∈R 100*1 .

[0127] Data dimension reduction is performed on the above data:

[0128]

[0129] The overall performance parameters rated speed N (0) , total temperature T (0) , total pressure , outlet static pressure are added to constitute the input sample Ps (0) ∈R 16*1 :

[0130] wherein, is the feature value of the two-dimensional airway boundary of the new sample after dimension reduction, is the feature value of the blade thickness distribution at 0% blade height and 100% blade height of the new sample after dimension reduction, is the feature value of the blade angle distribution at 0% blade height and 100% blade height of the new sample after dimension reduction;

[0131] The neural network function is calculated, f BP (Ps (0) )=[π0,m0,η0]。

[0132] It is determined whether the output pressure ratio π0, flow rate m0 and efficiency η0 meet the design requirements. If yes, the three-dimensional flow field simulation verification is continued; if not, the design is re-designed.

[0133] The flow is shown in the accompanying Figure 6 .

[0134] The above only describes the preferred embodiments of the present application, and it should be noted that, for those skilled in the art, several improvements can be made without departing from the principles of the present application, and these improvements should also be considered as the protection scope of the present application.

Claims

1. A centrifugal impeller design verification method for a low-power civil aviation engine, characterized by: The specific steps are as follows: Step 1: Generate a large number of 3D design models of centrifugal impellers for low-power civil aviation engines. Output the pressure ratio π, flow rate m, and efficiency η values ​​through 3D flow field simulation to generate training sample data. Step 2: For each centrifugal impeller 3D design model, deconstruct the 2D airway boundary points of the centrifugal impeller, the continuous blade thickness distribution at 100% blade height, the continuous blade thickness distribution at 0% blade height, the continuous angle distribution at 100% blade height, and the continuous blade angle distribution at 0% blade height; Step 3: Using principal component analysis, the continuous data points of the two-dimensional airway boundary points of the centrifugal impeller, the continuous blade thickness distribution at 100% blade height, the continuous blade thickness distribution at 0% blade height, the continuous angle distribution at 100% blade height, and the continuous blade angle distribution at 0% blade height are reduced to two dimensions, and the single variable characteristics of rated speed N, inlet total temperature T, inlet total pressure P0, and outlet static pressure P1 are added; Step 4: Train the BP neural network model; Step 5: For the subsequent three-dimensional design model data of the centrifugal impeller, the eigenvectors selected from the previous data set are used for dimensionality reduction. The eigenvalues ​​of the centrifugal impeller samples after dimensionality reduction, as well as the rated speed, inlet total temperature, inlet total pressure, and outlet static pressure, are input into the trained neural network function calculation results, and the pressure ratio, flow rate, and efficiency are output as a rapid design evaluation before the three-dimensional flow field simulation. Step 5 is as follows: For the design of a new centrifugal impeller 3D model, deconstruct the 2D geometric elements; 100 points are evenly and equidistantly selected on the curve, and the airway boundary data of the casing surface is recorded as A0∈R 100*2 , the hub surface airway boundary data is recorded as B0∈R 100*2 The blade angle of the centrifugal impeller at 0% blade height is recorded as Aa0∈R 100*1 , the blade angle of the centrifugal impeller at 100% blade height is recorded as Ab0∈R 100*1 , the blade thickness of the centrifugal impeller at 0% blade height is recorded as Ha0∈R 100*1 , the blade thickness of the centrifugal impeller at 100% blade height is recorded as Hb0∈R 100*1 ; Perform data dimensionality reduction on the above data: Add overall performance parameters Rated speed N (0) , total inlet temperature T (0) , total inlet pressure Outlet static pressure Construct a new input sample Ps (0) ∈R 16*1 : in, is the eigenvalue of the new sample's two-dimensional airway boundary after dimensionality reduction, is the eigenvalue of the leaf thickness distribution at 0% leaf height and 100% leaf height of the new sample after dimensionality reduction, is the eigenvalue of the leaf angle distribution at 0% leaf height and 100% leaf height of the new sample after dimensionality reduction; Substitute into the neural network function calculation, f BP (P s (0) )=[π0,m0,η0]; Evaluate whether the output pressure ratio π0, flow rate m0, and efficiency η0 values ​​meet the design requirements. If so, continue with the three-dimensional flow field simulation verification; if not, redesign.

2. The centrifugal impeller design verification method for a low-power civil aircraft engine according to claim 1, characterized in that: Step 1 is as follows: (1.1) Obtain samples using conventional centrifugal impeller design methods; (1.2) The output value array of the three-dimensional flow field simulation of sample 1 is [π1, m1, η1], denoted as Op (1) , where π1 is the pressure ratio of sample 1, m1 is the flow rate of sample 1, η1 is the efficiency of sample 1, and the three-dimensional flow field simulation output results of 100 samples form the matrix [Op (1) ;Op (2) ;……;Op (100) ] T , denoted as Op.

3. The centrifugal impeller design verification method for a low-power civil aircraft engine according to claim 2, characterized in that: Step 2 is as follows: (2.1) The two-dimensional airway boundary of the centrifugal impeller design sample includes the airway boundary of the upper casing surface and the airway boundary of the lower hub surface. The two-dimensional curve of the airway boundary of the casing surface consists of 100 two-dimensional coordinates. In the i-th two-dimensional curve of the airway boundary of the casing surface, 100 point coordinate samples are uniformly taken as A i , the hub surface airway boundary two-dimensional curve is also composed of 100 two-dimensional coordinates. In the i-th hub surface airway boundary two-dimensional curve, 100 point coordinate samples are uniformly taken as Denoted as B i ; (2.2) In the centrifugal impeller design, 100 points are uniformly selected from the continuous blade thickness distribution at 0% blade height of the i-th sample using the array Indicated by Ha i ; At 100% leaf height of the i-th sample, the continuous leaf thickness distribution is uniform and 100 points are taken using the array Indicated by Hb i ; (2.3) In the centrifugal impeller design, 100 points are uniformly selected from the continuous blade angle distribution at 0% blade height of the i-th sample using the array Indicated by Aa i ; uniformly select 100 points from the continuous leaf angle distribution at 100% leaf height of the i-th sample and use the array Indicates, denoted as Ab i .

4. The centrifugal impeller design verification method for a low-power civil aircraft engine according to claim 3, characterized in that: Step 3 is as follows: (3.1) Two-dimensional coordinate matrix A for airway boundary samples i and B i , reduced to one dimension using principal component analysis; (3.2) Calculate the matrix A i The covariance matrix C of C is calculated, and the eigenvalues ​​λ1 and λ2 of C are calculated. The eigenvalue λ with a covariance contribution rate greater than 90% is selected. k ; (3.3) Select λ k The corresponding eigenvectors form the eigentransformation matrix P k ∈R 1*2 ; (3.4) Matrix A i Dimensionality reduction transformation: The result after dimensionality reduction is recorded as As i , for the matrix B i Take the same operation to obtain the feature transformation matrix P m ∈R 1*2 , B i The result after dimensionality reduction is recorded as Bs i ; (3.5) A of all 100 samples i After dimensionality reduction, the matrix [As1, As2...As 100 ], recorded as As; (3.6) Calculate the covariance matrix C1 of the matrix As and calculate the eigenvalue λ of C1 a1 ,λ a2 ...λ a100 , select two eigenvalues ​​λ whose sum of covariance contribution is greater than 90% i and λ j ; (3.7) Select λ i and λ j The corresponding eigenvectors form the eigentransformation matrix P k2 ∈R 2*100 ; (3.8) Dimensionality reduction transformation of matrix As: Therefore, the dimensionality of the two-dimensional curve sample of the airway boundary of the i-th casing surface is reduced to two eigenvalues: To B i Take the same operation to get the feature transformation matrix P m2 , the two-dimensional curve sample of the hub surface airway boundary is reduced to two eigenvalues: (3.9) The blade thickness and blade angle feature matrix Ha for the i-th sample i , Hb i 、Aa i 、Ab i Repeat steps (3.6) to (3.8) to obtain the feature transformation matrix P x1 、P x2 、P x3 、P x4 After dimensionality reduction, the leaf thickness distribution at 0% leaf height, the leaf thickness distribution at 100% leaf height, the leaf angle distribution at 0% leaf height, and the leaf angle distribution at 100% leaf height of the i-th sample are obtained as the reduced dimensionality eigenvalues: (3.10) Therefore, the two-dimensional design continuous features of the i-th sample can be extracted into 12 eigenvalues, combined with the overall performance parameters of the engine, rated speed N (i) , total inlet temperature T (i) , total inlet pressure Outlet static pressure The design characteristics of centrifugal impeller sample i can be expressed by the following 16 characteristic parameter groups: Denoted as Ps (i) ,i=1,2,……100, sample set matrix [Ps (1) ;Ps (2) ;……;Ps (100) ] T , denoted as Ps.

5. The centrifugal impeller design verification method for a low-power civil aircraft engine according to claim 4, characterized in that: Step 4 is as follows: The reduced-dimensional data sample set Ps is used as the input matrix, the simulation result matrix Op is used as the output matrix, and a hidden layer network structure is adopted. The hidden layer nodes are calculated according to the empirical formula Determine, where H is the number of hidden layer nodes, I is the number of input layer nodes, and O is the number of output layer nodes. Select H = 6; The weight matrix between the input layer and the hidden layer is recorded as The weight matrix between the hidden layer and the output layer is recorded as The bias vector between the input layer and the hidden layer is denoted as B = (B1, B2, ..., B6) T ; The bias vector between the hidden layer and the output layer is denoted as C = (C1, C2, C3) T ; For the weight matrix W (1) 、W (2) And the bias vectors B and C are initialized and assigned values, and the initialization values ​​are random numbers in the range of (0.5, 1); The transfer function is the Sigmoid function, that is Calculate the jth hidden layer value of the i-th sample Where W j (1) is the weight matrix W (1) The jth row of P s (i) is the i-th sample in the sample set Ps, B j is the jth value in the bias vector B; Similarly, calculate the j-th output layer value of the i-th sample Where W j (2) is the weight matrix W (2) The jth row, Hid (i) is the hidden layer output value vector of the i-th sample in the sample set, C j is the jth value in the bias vector C; Refer to the output sample Op in the training sample and calculate the i-th sample Op (i) and its Output (i) The error E between (i) : in is the i-th sample Op (i) The kth value in is the i-th sample Output (i) The kth value in ; The total cost E of all training samples total for: The weights and biases are updated using batch gradient descent until the system error is minimized: Where μ is the learning rate, and the selection range is (0.05, 0.8) The neural network model after training is recorded as f BP (x),x∈R 16*1 .

Images