A method and system for optimal design of turbomachinery based on orthogonal coders

Abstract

The application discloses a kind of orthogonal encoder-based impeller mechanical optimization design method and system, by constructing the orthogonal self-encoder OAE with performance guide function, the depth compression and reconstruction of blade geometric characteristics are realized;Introduce singular value decomposition SVD to centralize and orthogonalize the hidden space characteristics, while eliminating feature redundancy, the search space dimension is effectively reduced by principal component screening;Finally, a high-precision surrogate model is constructed using generalized regression neural network GRNN, to realize the rapid prediction and global optimization of aerodynamic performance under small sample conditions.

Description

Technical Field

[0001] This invention relates to the field of turbomachinery technology, specifically to a turbomachinery optimization design method and system based on orthogonal encoders, involving deep learning, dimensionality reduction techniques, and surrogate models in turbomachinery optimization design. Background Technology

[0002] Modern high-speed turbomachinery includes centrifugal turbines, centrifugal compressors, axial flow fans, gas turbines, and high-speed centrifugal pumps. During their development, the coordinated operation of impellers, nozzles, guide vanes, volutes, diffusers, and flow channel aerodynamic components directly determines the overall flow efficiency, pressure ratio, and operational stability of the machine. Taking the core component—the three-dimensional impeller—as an example, the current design process faces significant challenges in high-dimensional nonlinear optimization: First, the geometric parameter space of the three-dimensional impeller is extremely complex, involving dozens of degrees of freedom such as blade inlet and outlet angles, wrap angles, thickness distribution, and flow channel curvature. Traditional design processes rely on designers' engineering intuition for initial selection, followed by iterative iterations through a closed loop of "design-mesh generation-CFD simulation-analysis-modification." This results in long development cycles, high costs, and is easily limited by the designer's personal experience. Second, using high-precision three-dimensional viscous flow field calculations (CFD) to improve aerodynamic performance takes hours or days for a single simulation. If genetic algorithms (GA) or particle swarm optimization (PSO) global optimization methods are used, hundreds or thousands of CFD evaluations are required, which is time-consuming. These heuristic algorithms are prone to getting trapped in local optima when dealing with small sample data and are difficult to handle complex geometric constraints, such as the continuity of blade thickness distribution and interference constraints between blades.

[0003] To accelerate optimization, the engineering community often introduces surrogate models such as Kriging and RBF to replace expensive CFD calculations. However, when faced with high-dimensional geometric parameters, the prediction accuracy of traditional surrogate models drops sharply with the increase in the number of geometric control points, making it difficult to effectively capture the highly nonlinear mapping relationship between blade geometry parameters and aerodynamic performance. Especially under small sample data conditions, traditional models lose their guiding significance due to overfitting or excessive smoothing. Heuristic optimization-based surrogate model search algorithms such as Kriging-PSO and RBF-GA are also limited by empirical parameter setting or search mechanism constraints, leading to fluctuating optimization results or difficulty in convergence.

[0004] Common dimensionality reduction surrogate model methods, such as linear dimensionality reduction surrogate models PCA-Kriging and POD-RBF, and nonlinear dimensionality reduction surrogate models KPCA-Kriging, have achieved significant results in areas such as turbomachinery aerodynamic optimization and airfoil design. However, for high-dimensional turbomachinery (… d>50) Design problems and multi-objective optimization still have difficulties: First, linear dimensionality reduction such as PCA and POD cannot effectively extract the nonlinear features of blade geometry, and the latent space features have insufficient discriminative power; Second, although nonlinear dimensionality reduction such as KPCA can improve the feature representation ability, it is accompanied by a serious loss of geometric reconstruction accuracy, which makes it impossible to accurately reconstruct the low-dimensional optimal latent space vector into the original dimension optimal design parameters.

[0005] In summary, existing design methods are insufficient to meet the R&D requirements of high-speed turbomachinery for high efficiency, high reliability, and rapid response under extreme operating conditions. Summary of the Invention

[0006] The purpose of this invention is to provide a turbomachinery optimization design method and system based on orthogonal encoders. Based on deep learning, dimensionality reduction technology, and surrogate models, it constructs a deeply coupled data-driven system of "design parameters and performance parameters" to achieve the inversion of optimal geometric parameters under a given target performance. This breaks through the efficiency bottleneck of traditional design methods and realizes the automated, intelligent, and precise design of aerodynamic components of high-speed turbomachinery. It features geometric feature dimensionality reduction, high generalization ability under small sample conditions, and an end-to-end intelligent optimization method that can directly embed complex physical constraints.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A turbomachinery optimization design method based on orthogonal encoders, comprising the following steps: S1. Pneumatic component parameterization: For the curved portion of the turbomachinery, a high-order Bezier curve is used to fit and establish a path mathematical model, and the coordinates of the corresponding curve control points are obtained; for the non-curved portion, direct parameterization is performed and organized into a design parameter array. x i, j , i For the sample size, j Dimensions of design parameters; S2. Based on the design parameter array from step S1 x i, j Define the upper and lower limits of sampling and the number of samples, and generate the design parameter sample matrix using a sampling algorithm (such as random sampling, orthogonal sampling, Latin hypercube sampling, Monte Carlo sampling, Sobol sequence sampling). X samp For the design parameter sample matrix X samp For each sample, 3D modeling, mesh generation, and high-fidelity CFD simulation are performed to obtain a sample matrix of performance parameters to be optimized. Y samp Design parameter sample matrix X samp , Sample matrix of performance parameters to be optimized Y sampCombined into a dataset sample matrix D set and the dataset sample matrix D set Clean and remove outliers, and use standardization methods to eliminate the impact of differences in the dimensions of different geometric parameters on model training; S3, Input dataset sample matrix D set Set up hidden layer neurons n hidden Latent space dimension d laten The number of epochs is used to train the improved orthogonal autoencoder (OAE); after training, the latent space vectors are obtained. z train The outputs of the decoder function (decoder()) are used to perform deep compression and reconstruction of the impeller's geometric features. S4. The latent space vector z train Decentralization, for latent space vectors z train Singular value decomposition is introduced for decomposition, and orthogonal basis coefficients are calculated. While eliminating feature redundancy, principal component screening is used to reduce the dimensionality of the search space. S5. Establish a nonlinear mapping relationship between orthogonal basis coefficients and performance parameters using a generalized regression neural network (GRNN), construct a high-precision surrogate model, input the target performance parameters into the high-precision surrogate model, invert the orthogonal basis coefficients, reconstruct the latent space vector of the optimal design parameters, decode the optimal design parameters, and then... x opt Draw a 3D model of the impeller, perform mesh generation, and conduct CFD simulation to obtain actual performance parameters; Performance parameters include: For compression media (centrifugal / axial compressors, blowers, fans): pressure ratio or boost pressure, isentropic efficiency, surge margin, flow rate; For expansion power (turbine expanders, steam turbines, gas turbines): expansion ratio, isentropic efficiency, output power, axial force; For liquid transport (centrifugal pumps, mixed-flow pumps, axial-flow pumps): head, hydraulic efficiency, required net positive suction head (NPSH), shaft power.

[0008] S6. Judgment and Sample Addition: Comparing Target Performance Parameters y opt Compared with actual performance parameters y new If the convergence condition is met, the optimization ends; otherwise, the new sample matrix is ​​used. D new =[ x new , y new [Incorporate into dataset sample matrix]D set Then, return to S2 to retrain and optimize.

[0009] Preferably, the curved portion of the impeller mechanism in step S1 includes the front and rear cover plate curves of the axial projection, the front and rear cover plate blade rib line curves in the meridional coordinate system, the blade load distribution curve, the impeller blade thickness distribution curve, the nozzle profile and thickness distribution curve, and the volute profile; the non-curved portion includes the number of blades, the inlet and outlet placement angles, and the volute tongue angle.

[0010] Preferably, the formula for designing the parameter array in step S1 is: x i, j = [ x (i,1) , x (i,2) , …, x (i,j) ] In the formula: i For the sample size, j The dimension of the design parameters.

[0011] Preferably, in step S2, the design parameter array x i, j The sampling method is Latin hypercube sampling; the resulting design parameter sample matrix X samp for:

[0012] Sample matrix of performance parameters to be optimized Y samp for:

[0013] In the formula k The dimension of the performance parameters; Dataset sample matrix D set for: D set = [ X samp , Y samp ].

[0014] Preferably, training the improved orthogonal autoencoder (OAE) in step S3 includes the following steps: (1) Structured feature extraction: Train the encoder to map the high-dimensional blade geometric parameters to low-dimensional vectors in the latent space. z ; (2) Geometric reconstruction and fidelity preservation: train the decoder to convert the low-dimensional vectors in the latent space into geometric reconstruction and fidelity preservation. z Restored to the original high-dimensional geometric parameters; (3) Performance supervision: using low-dimensional vectors in the latent space z As input, predict the performance indicators of the impeller; force the use of low-dimensional vectors in the latent space. z It should not only contain geometric information, but also features that are strongly related to the aerodynamic performance of the turbomachinery, so that the latent space becomes an aerodynamically sensitive feature space. During optimization, it can directly search in the direction of optimal performance, rather than blindly traversing within the geometric space. (4) Loss function: During the training of the improved orthogonal autoencoder (OAE), the model simultaneously optimizes the following three loss terms: Geometric reconstruction loss Loss recon To ensure the accuracy of the geometry; Performance prediction loss Loss guide : to ensure low-dimensional vectors in the latent space z The mapping relationship with aerodynamic performance indicators is accurate; Hidden space orthogonality Loss ortho To reduce the correlation between different dimensions of the latent space and achieve decoupling of geometric features; Combine the above three losses into a single overall loss, and assign a weight to each: Loss total= a·Loss recon + b·Loss guide + c· Loss ortho In the formula, a , b、c These are the weighting coefficients; a backpropagation mechanism is used to make... Loss total Minimize.

[0015] Preferably, the latent space vector in step S4 z train The decentralization formula is: z center = z train – mean( z train ) For latent space vectors z train The decomposition formula using singular value decomposition is as follows: [ Uz , S z , V z ]=SVD( z center ) The formula for calculating the orthogonal basis coefficients is: C z = U z · z center Preferably, the formula for the nonlinear mapping relationship between the orthogonal basis coefficients and performance parameters established by the generalized regression neural network (GRNN) in step S5 is as follows: NOT grnn =GRNN ( Y samp , C z , σ ) In the formula, σ The bandwidth of the Gaussian kernel function is obtained through optimization using cross-validation. The formula for inputting target performance parameters is: y opt =[ y 1-opt , y 2-opt , … , y j-opt ] The inversion orthogonal basis coefficients are: C z-opt = NOT grnn ( y opt , σ ) The latent space vector reconstruction of the optimal design parameters is expressed as: z pred = U z · C z-opt + mean( z train ) The optimal design parameters are decoded as follows: x new = decoder( z pred) The actual performance parameters are: y new =[ y 1-new , y 2- new , … , y j-new ].

[0016] Preferably, the convergence condition in step S6 is: c = .

[0017] A design optimization system for turbomachinery, used to implement the above method, wherein the data interfaces for each step of the process include: (1) Input and Average Line Module Functions: Performs the preliminary work of steps S1 and S2, supporting one-dimensional thermodynamic estimation and two-dimensional average line design for turbomachinery; performs Latin hypercube sampling based on physical boundary constraints; automatically calls external 3D modeling and mesh generation scripts; Output workpieces: Generate parameter layer workpieces and configuration layer workpieces; The parameter layer workpiece contains geometric parameters such as impeller profile control points and blade angle distribution; the configuration layer workpiece contains fluid medium properties, boundary conditions, and mesh topology control parameters. (2) Geometric parameterization module Function: Execute step S3; build a deep learning framework and deploy an improved orthogonal autoencoder; perform nonlinear dimensionality reduction on the input parameter layer artifact, extract low-dimensional manifold features, and save the decoder network weights; Output workpiece: Generate geometric layer workpiece; The geometry layer artifact includes latent space manifold vectors, reconstruction error evaluation curves, and decoder network model files; (3) Optimize scheduling and evaluation module Functions: Execute steps S4 and S6; perform centering and singular value decomposition (SVD) orthogonalization on the latent space vectors; manage the iterative strategy of active learning; evaluate the error between the actual feedback results of CFD simulation and the predicted values ​​of the surrogate model, and decide whether to trigger sample appending and retraining; Output artifacts: Establish a sample-level catalog, dynamically maintain the historical sample library and result-level artifacts; The result layer artifact includes the convergence history curve, Pareto front solution set, and optimal geometric scheme; (4) Results Analysis Module Functions: Execute step S5; build and train a GRNN-based performance prediction proxy model; provide interfaces for multi-objective optimization algorithms (such as NSGA-III) or inversion algorithms; provide 3D visualization interaction for design variable sensitivity, geometric deformation trends, and flow field cloud maps, and support local interactive re-optimization; Output of workpieces: Generate visual reports, proxy model accuracy verification curves, and final CAD / mesh production-grade files; (5) Data bus and workpiece management module Function: As the core skeleton of the system, it is responsible for the unified maintenance of the four serializable intermediate artifacts of the parameter layer, configuration layer, geometry layer and result layer. Control mechanism: Strictly manage the version number and upstream and downstream dependencies of each workpiece; provide a "playback control" function to support one-click retracing of any historical optimization node, ensuring the traceability of the design process and data consistency; (6) External interface module Functionality: Provides cross-language (C++ / Python / Fortran) dynamic link library interfaces; Calling objects: Real-time calling of high-precision multi-component thermal property libraries (such as NIST REFPROP, CoolProp) to obtain real gas / liquid properties; seamless integration with mainstream external physics solvers (such as ANSYS CFX, Fluent, NUMECA or OpenFOAM) to achieve fully automatic closed-loop calling of mesh generation and CFD simulation.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention establishes a method and system for the aerodynamic optimization design of turbomachinery using deep learning, dimensionality reduction techniques, and surrogate models. Building upon autoencoder neural network feature extraction, to further eliminate redundancy and correlation among latent space features, an orthogonalization mechanism based on Singular Value Decomposition (SVD) is introduced. By centering and orthogonally decomposing the latent space feature matrix, high-dimensional coupled geometric features are mapped to a set of mutually orthogonal basis vectors. This process not only achieves essential decoupling of geometric features but also filters out the principal feature components with the most significant impact on aerodynamic performance based on the magnitude of singular values. This significantly reduces the search space dimensionality of subsequent optimization algorithms, improving optimization efficiency and convergence stability under small sample conditions. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method described in this invention; Figure 2 This is a schematic diagram of the optimized compressor structure in the embodiment; Figure 3 This is a schematic diagram of the data interface process in the embodiment. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0021] like Figure 1 The impeller mechanical optimization design method based on orthogonal encoders, as shown, includes the following steps: S1. Parameterization of pneumatic component design: Input compressor flow m Rotation speed n Pressure ratio Π t Inlet temperature T t Import pressure P t The parameters are as follows, and the medium is nitrogen. The specific speed can be obtained from the aerodynamic thermodynamics calculations of the impeller. n q Shaft power P Q It can also determine the initial impeller geometry parameters and impeller outer diameter. d 2. Import radius d s Export width b 2. Wheel hub diameter d H The above aerodynamic and thermodynamic parameters are shown in Table 1.

[0022] Table 1

[0023] A high-order Bezier curve was used to fit the front and rear cover plate curves of the axial projection, the front and rear cover plate blade rib line curves of the meridional plane, and the volute profile to establish a path mathematical model, and the coordinates of the corresponding curve control points were obtained.

[0024] There are a total of 8 geometric parameters for the impeller axial projection: the upper and lower coordinates of the blade inlet edge. x 1. x 2. x 3. x 4. Coordinates of Bezier curve control points on the front hood x 5. x 6. Coordinates of Bezier curve control points on the rear cover plate x 7. x 8; Number of leaves x 9; There are a total of 8 control points for the meridional blade rib curve: namely, the inlet angle of the meridional front cover plate curve. x 10 Relative location of import control points x 11 Export perspective x12 Relative location of export control points x 13 Meridian plane rear cover plate curve inlet angle x 14 Relative location of import control points x 15 Export perspective x 16 Relative location of export control points x 17 The volute profile control points have five coordinates: the absolute height of the volute tongue, etc. x 18 relative height x 19 Relative angle x 20 Export relative height x 21 Relative angle x 22 Length of the expansion section at the volute exit section x 23 ; diameter of the volute exit section x 24 ; Cochlear tongue x 25 .

[0025] The initial design parameter samples and sampling range are shown in Table 2.

[0026] Table 2

[0027] Based on Latin hypercube sampling, the number of samples is set to 5. An initial design parameter sample matrix is ​​generated. X samp (5×25). There are two optimization objectives: the isentropic efficiency at 420 K with the inlet temperature. y 1 and isentropic efficiency at an inlet temperature of 120 K y 2. After reviewing... X samp After modeling, network partitioning, and CFD simulation, we can obtain Y samp (5×2), the initial performance parameter sample set is shown in Table 3.

[0028] Table 3

[0029] S2, Generate sample set D set =[ X samp , Y samp After data cleaning, the following can be obtained:D set (5×27); S3, Input D set Set up hidden layer neurons n hidden Latent space dimension d laten Number of rounds (epochs) a , b , c The improved orthogonal autoencoder (OAE) was trained; the training settings for the improved orthogonal autoencoder (OAE) are shown in Table 4.

[0030] Table 4

[0031] The latent space vector can be obtained z train The parameters are shown in Table 5.

[0032] Table 5

[0033] S4. The latent space vector z train Decentralization, for latent space vectors z train Singular value decomposition is introduced for decomposition, and orthogonal basis coefficients are calculated. While eliminating feature redundancy, principal component screening is used to reduce the dimensionality of the search space. Latent space vectors z train Decentralization, can be obtained z center Its parameters are shown in Table 6.

[0034] Table 6

[0035] Latent space vectors z train Perform SVD decomposition: [ U z , S z , V z ]=SVD( z center ) Calculate the orthogonal basis coefficients: C z = U z · z center achievable C z The parameters are shown in Table 7.

[0036] Table 7

[0037] S5. The nonlinear mapping relationship between orthogonal bases and performance parameters is established through a GRNN neural network, which can be expressed as: NOT grnn =GRNN( Y samp , C z , σ ) achievable σ= 0.016.

[0038] Optimal design parameter reconstruction and verification, with the target performance parameters as input, can be expressed as: y opt =[ y 1-opt , y 2-opt , … , y j-opt ] in y 1-opt =0.85, y 2-opt =0.85; The inversion orthogonal basis coefficients can be expressed as: C z-opt = NOT grnn ( y opt , σ ) achievable C z-opt =[0, 4.76837e-08, 8.34465e-08] T .

[0039] The reconstruction of the latent space vectors of the optimal design parameters can be expressed as: z pred = U z · C z-opt + mean( z train ) achievable zpred =[-1.32791, -2.12381, 0.869404] T .

[0040] The optimal design parameters can be decoded as follows: x new = decoder( z pred ) achievable x new The results of the first design parameter optimization are shown in Table 8.

[0041] Table 8

[0042] Performance verification: based on design parameters x opt Create a 3D model of the impeller, perform mesh generation, conduct CFD simulation, and obtain actual performance parameters. y new =[ y 1-new , y 2- new , … , y j-new ] achievable y new =[0.6885, 0.6356] S6. Judgment and Sample Addition: Calculate the CFD simulation results in accordance with the target. y opt deviation c： c =

[0043] achievable c =0.2414. x new , y new Merge into a new sample D new =[ x new , y new ], sample D new Merge into D set , D set =[ D set ;D new (6×27), return to S2, until... c Optimization is completed after convergence.

[0044] Table 9 shows the maintenance y 1-opt =0.85, y 2-opt =0.85 remains unchanged, and the performance parameters after 2-8 iterations convergence are as follows: c value.

[0045] Table 9

[0046] The above are merely preferred embodiments of the present invention. It should be noted that, for those skilled in the art, other equivalent modifications and improvements can be made based on the technical teachings provided by the present invention, and these should also be considered within the scope of protection of the present invention.

Claims

1. A turbomachinery optimization design method based on orthogonal encoders, characterized in that: The method includes the following steps: S1. Pneumatic component parameterization: For the curved portion of the turbomachinery, a high-order Bezier curve is used to fit and establish a path mathematical model, and the coordinates of the corresponding curve control points are obtained; for the non-curved portion, direct parameterization is performed and organized into a design parameter array. x i, j , i For the sample size, j Dimensions of design parameters; S2. Based on the design parameter array from step S1 x i, j Define the upper and lower limits of sampling and the number of samples to generate a sample matrix of design parameters. X samp For the design parameter sample matrix X samp For each sample, 3D modeling, mesh generation, and high-fidelity CFD simulation are performed to obtain a sample matrix of performance parameters to be optimized. Y samp Design parameter sample matrix X samp , Sample matrix of performance parameters to be optimized Y samp Combined into a dataset sample matrix D set and the dataset sample matrix D set Clean and remove outliers, and use standardization methods to eliminate the impact of differences in the dimensions of different geometric parameters on model training; S3, Input dataset sample matrix D set Set up hidden layer neurons n hidden Latent space dimension d laten The number of epochs is used to train the improved orthogonal autoencoder (OAE); after training, the latent space vectors are obtained. z train The outputs of the decoder function (decoder()) are used to perform deep compression and reconstruction of the impeller's geometric features. S4, implicit space vector z train Decentralization, for latent space vectors z train Singular value decomposition is introduced for decomposition, and orthogonal basis coefficients are calculated. While eliminating feature redundancy, principal component screening is used to reduce the dimensionality of the search space. S5. Establish a nonlinear mapping relationship between orthogonal basis coefficients and performance parameters using a generalized regression neural network (GRNN), construct a high-precision surrogate model, input the target performance parameters into the high-precision surrogate model, invert the orthogonal basis coefficients, reconstruct the latent space vector of the optimal design parameters, decode the optimal design parameters, and then... x opt Draw a 3D model of the impeller, perform mesh generation, and conduct CFD simulation to obtain actual performance parameters; S6. Judgment and Sample Addition: Comparing Target Performance Parameters y opt Compared with actual performance parameters y new If the convergence condition is met, the optimization ends; otherwise, the new sample matrix is ​​used. D new =[ x new , y new [Incorporate into dataset sample matrix] D set Then, return to S2 to retrain and optimize.

2. The turbomachinery optimization design method according to claim 1, characterized in that: The curved part of the impeller in step S1 includes the front and rear cover plate curves of the axial projection, the front and rear cover plate blade rib line curves of the meridional coordinate system, the blade load distribution curve, the impeller blade thickness distribution curve, the nozzle profile and thickness distribution curve, and the volute profile; the non-curved part includes the number of blades, the inlet and outlet placement angles, and the volute tongue angle.

3. The turbomachinery optimization design method according to claim 1, characterized in that: The formula for designing the parameter array in step S1 is: x i, j = [ x (i,1) , x (i,2) , …, x (i,j) ] In the formula: i For the number of samples, j The dimension of the design parameters.

4. The turbomachinery optimization design method according to claim 1, characterized in that: In step S2, the parameter array is designed. x i, j The sampling method is Latin hypercube sampling; the resulting design parameter sample matrix X samp for: Sample matrix of performance parameters to be optimized Y samp for: In the formula k The dimension of the performance parameters; Dataset sample matrix D set for: D set = [ X samp , Y samp ]。 5. The turbomachinery optimization design method according to claim 1, characterized in that: The training of the improved orthogonal autoencoder (OAE) in step S3 includes the following steps: (1) Structured feature extraction: Train the encoder to map the high-dimensional blade geometric parameters to low-dimensional vectors in the latent space. z ; (2) Geometric reconstruction and fidelity preservation: train the decoder to convert the low-dimensional vectors in the latent space into geometric reconstruction and fidelity preservation. z Restored to the original high-dimensional geometric parameters; (3) Performance supervision: using low-dimensional vectors in the latent space z As input, predict the performance indicators of the impeller; force the use of low-dimensional vectors in the latent space. z It should not only contain geometric information, but also features that are strongly related to the aerodynamic performance of the turbomachinery, so that the latent space becomes an aerodynamically sensitive feature space. During optimization, it can directly search in the direction of optimal performance, rather than blindly traversing within the geometric space. (4) Loss function: During the training of the improved orthogonal autoencoder (OAE), the model simultaneously optimizes the following three loss terms: Geometric reconstruction loss Loss recon To ensure the accuracy of the geometry; Performance prediction loss Loss guide : to ensure low-dimensional vectors in the latent space z The mapping relationship with aerodynamic performance indicators is accurate; Hidden space orthogonality Loss ortho To reduce the correlation between different dimensions of the latent space and achieve decoupling of geometric features; Combine the above three losses into a single overall loss, and assign a weight to each: Loss total= a·Loss recon + b·Loss guide + c· Loss ortho In the formula, a , b, c These are the weighting coefficients; a backpropagation mechanism is used to make... Loss total Minimize.

6. The turbomachinery optimization design method according to claim 1, characterized in that: The latent space vector in step S4 z train The decentralization formula is: z center = z train – mean( z train ) In the formula, mean() represents taking the average of the matrix and the latent space vector. z train We introduce Singular Value Decomposition (SVD) for decomposition: [ U z , S z , V z ]=SVD( z center ) The formula for calculating the orthogonal basis coefficients is: C z = U z ·z center 。 7. The turbomachinery optimization design method according to claim 1, characterized in that: The formula for the nonlinear mapping relationship between the orthogonal basis coefficients and performance parameters established by the generalized regression neural network (GRNN) in step S5 is as follows: NET grnn =GRNN(Y samp , C z , σ) In the formula, σ The bandwidth of the Gaussian kernel function is obtained through optimization using cross-validation. The formula for inputting target performance parameters is: y opt =[ y 1-opt , y 2-opt , … , y j-opt ] The inversion orthogonal basis coefficients are: C z-opt = NET grnn ( y opt , σ ) The latent space vector reconstruction of the optimal design parameters is expressed as: z pred = U z ·C z-opt + mean( z train ) The optimal design parameters are decoded as follows: x new = decoder( z pred ) The actual performance parameters are: y new =[ y 1-new , y 2- new , … , y j- new ]。 8. The turbomachinery optimization design method according to claim 1, characterized in that: The convergence condition in step S6 is: c = 。 9. A design optimization system for turbomachinery, characterized in that: This system is used to implement the method described in any one of claims 1-8, wherein the data interface for each step in the process includes: (1) Input and Average Line Module Functions: Performs the preliminary work of steps S1 and S2, supporting one-dimensional thermodynamic estimation and two-dimensional average line design for turbomachinery; performs Latin hypercube sampling based on physical boundary constraints; automatically calls external 3D modeling and mesh generation scripts; Output workpieces: Generate parameter layer workpieces and configuration layer workpieces; (2) Geometric parameterization module Function: Execute step S3; build a deep learning framework and deploy an improved orthogonal autoencoder; perform nonlinear dimensionality reduction on the input parameter layer artifact, extract low-dimensional manifold features, and save the decoder network weights; Output workpiece: Generate geometric layer workpiece; (3) Optimize scheduling and evaluation module Functions: Execute steps S4 and S6; perform centering and singular value decomposition (SVD) orthogonalization on the latent space vectors; manage the iterative strategy of active learning; evaluate the error between the actual feedback results of CFD simulation and the predicted values ​​of the surrogate model, and decide whether to trigger sample appending and retraining; Output workpieces: Establish sample-level catalogs and result-level workpieces; (4) Results Analysis Module Functions: Execute step S5; build and train a GRNN-based performance prediction proxy model; provide interfaces for multi-objective optimization algorithms or inversion algorithms; provide 3D visualization and interaction of design variable sensitivity, geometric deformation trends, and flow field cloud maps, and support local interactive re-optimization; Output of workpieces: Generate visual reports, proxy model accuracy verification curves, and final CAD / mesh production-grade files; (5) Data bus and workpiece management module Function: As the core skeleton of the system, it is responsible for the unified maintenance of the four serializable intermediate workpieces of the parameter layer, configuration layer, geometry layer and result layer. Control mechanism: Strictly manage the version number and upstream and downstream dependencies of each workpiece; provide "playback control" function, support one-click retracing of any historical optimization node, and ensure the traceability and data consistency of the design process; (6) External interface module Function: Provides cross-language dynamic link library interfaces; Calling objects: Real-time calling of high-precision multi-component thermal property libraries (such as NIST REFPROP, CoolProp) to obtain real gas / liquid properties; seamless integration with mainstream external physics solvers (such as ANSYS CFX, Fluent, NUMECA or OpenFOAM) to achieve fully automatic closed-loop calling of mesh generation and CFD simulation.

Images