Turbine blade profile parameterization method based on variational auto-encoder generation model

Abstract

The invention discloses a turbine blade profile parameterization method based on a variational auto-encoder generation model, and belongs to the field of turbine blade profile parameterization methods. According to the method, a free deformation-variational auto-encoder generative adversarial network method is utilized, key geometric features are found through the learning ability of a machine learning model to data features, and the geometric control parameter quantity is remarkably reduced. And an SHAP interpretable analysis tool is adopted to interpret and analyze the geometric process of the latent space generated blade profile of the generative adversarial network model of the variational auto-encoder, an intuitive mapping relation between the latent space characteristics and the real geometry is found, and the influence between the latent space and the aerodynamic performance is analyzed. According to the method, turbine blade profile parameterization is carried out by combining the high-degree-of-freedom fine regulation geometric advantage of the free deformation parameterization method and the efficient dimension reduction advantage and the high-quality blade profile generation advantage of the variational auto-encoder generation type model, the diversity and the high degree of freedom of the generated blade profile geometry are guaranteed, and the aerodynamic performance of the turbine blade profile is improved.

Description

Technical Field

[0001] The present invention belongs to the field of turbomachine blade profile parameterization, and relates to a turbomachine blade profile parameterization method based on a variational autoencoder generative adversarial network, which is used to efficiently generate geometric parameters of turbomachine blades with high precision and high aerodynamic performance. Background Technique

[0002] As the core aerodynamic component of aeroengines and gas turbines, the geometric design of turbomachines directly affects the overall aerodynamic performance and operating efficiency of the machine. With the rapid development of research in the field of computers, data-driven parameterization methods have gradually shown their advantages. The core of these methods lies in using data to guide the determination, optimization, or modeling process, reducing the dependence on prior knowledge or manual adjustment. Existing geometric parameterization methods have defects such as relying on empirical correction for controlling parameter values, high-dimensional aerodynamic optimization, and insufficient exploration of the design space. Generative models provide new ideas for breaking through traditional design paradigms. The core goal is to learn the distribution of data and generate new data similar to the data characteristics based on this. Although the blade profiles generated by the variational autoencoder generative adversarial network generation model conform to physical laws in terms of geometric shape, the correlation between its latent space characteristics, aerodynamic parameters, and blade profile geometry is difficult to explain. Summary of the Invention

[0003] The present invention provides a turbomachine blade profile parameterization method based on a variational autoencoder generation model. This method combines the high-degree-of-freedom fine-tuning geometric advantages of the free-form deformation parameterization method, as well as the efficient dimensionality reduction advantage and high-quality blade profile generation advantage of the variational autoencoder generative model to perform turbomachine blade profile parameterization, and can analyze the influence between space and aerodynamic performance, ensure the diversity and high degree of freedom of the generated blade profile geometry, and improve the aerodynamic performance of the turbomachine blade profile.

[0004] The object of the present invention is achieved through the following technical solutions:

[0005] A turbomachine blade profile parameterization method based on a variational autoencoder generation model disclosed by the present invention includes the following steps:

[0006] Step 1: Based on the original blade profile geometry, use the direct manipulation free-form deformation method to perform geometric deformation on it. The direct manipulation free-form deformation method, based on the free-form deformation method, can directly specify the displacement change amounts of one or more points on the target geometric surface and inversely calculate the change of the control volume framework, and recalculate the positions of other points to be deformed under the updated control volume framework. Use the direct manipulation free-form deformation method to set several control points on the blade profile surface. To reduce the probability of generating deformed blade profiles, the control points are set more than 5% - 20% of the axial chord length away from the leading and trailing edges of the blade profile. Obtain the displacement magnitudes of each control point through the Latin hypercube sampling method, and input the data into the direct manipulation free-form deformation method for batch processing to generate a sample library containing a large number of different blade profiles.

[0007] Step 2: Organize the x and y coordinates of the points on the blade profile surface in the sample library into a one-dimensional vector as the input vector of the variational autoencoder generative adversarial network model, and train the generative model of the variational autoencoder generative adversarial network. The variational autoencoder generative adversarial network model consists of three parts. The first part is the variational autoencoder part, which includes two structures: an encoder and a decoder. The encoder maps the input parameters to the latent space, and the decoder reconstructs the latent space into output data. The KL divergence is used as the encoder loss, and the binary cross-entropy loss is used as the decoder loss. The second part is the discriminator, which is used to distinguish whether the generated sample is a real sample or a generated sample, and the probability of correct judgment by the discriminator is used as the loss in training. The third part is the joint training part of the autoencoder and the discriminator, where the autoencoder is used as the generator to conduct adversarial training with the discriminator.

[0008] Step 3: Use the MISES numerical simulation method to perform mesh division and simulation calculations on each geometric sample in the blade profile sample library generated by the variational autoencoder generative adversarial network to obtain the aerodynamic performance of each sample. The initial pressure condition for the MISES calculation is input in the form of pressure ratio (P1 / P 01 ), where P1 is the inlet static pressure and P 01 is the inlet total pressure. Set a fixed inlet Mach number, and the MISES numerical simulation method automatically iteratively calculates the inlet total pressure according to the input Mach number and outlet parameters. To achieve an accurate description of the flow field, a turbulence model based on Drela correction and the AGS transition model are used to solve the flow field.

[0009] Step 4: Establish an association between the latent space and the blade profile geometry and aerodynamic performance through the SHAP attribution analysis method. The main aerodynamic performance parameter, the total pressure loss coefficient, is expressed as:

[0010]

[0011] where, P1 * and P1 are the inlet total pressure and the inlet static pressure, P 02is the local total pressure. Adjust the latent space vector according to this correlation to generate a blade profile with better aerodynamic performance. The SHAP importance calculation formula is as follows:

[0012]

[0013] where φ i represents the SHAP value of feature i, reflecting the contribution of feature i to the model prediction result. F represents the complete set of features, that is, the set of all input features. represents the feature subset S that does not contain feature i, and the formula sums over the subset S. |S| and |F| represent the number of features in subset S and the total number of features in the complete set F respectively. f S∪i (x S∪i ) represents the predicted value of the model trained on the feature set S∪i, that is, the model output considering feature i. f S (x S ) represents the predicted value of the model trained on the feature set S, that is, the model output without considering feature i. f S∪i (x S∪i ) - f S (x S ) represents the incremental contribution to the model predicted value after adding feature i. represents the weight term, indicating the proportion of the probability of the subset S appearing among all possible combinations.

[0014] It also includes Step Five: According to the correlation between the latent space established in Step Four and the blade profile geometry and aerodynamic performance, analyze the influence between the space and the aerodynamic performance, and through the directional regulation of the key latent space features, ensure the diversity and high degree of freedom of the generated blade profile geometry, and improve the aerodynamic performance of the turbomachine blade profile.

[0015] Beneficial effects:

[0016] 1. A parameterization method for turbomachine blade profiles based on a variational autoencoder generation model disclosed in the present invention uses the free form deformation-variational autoencoder generative adversarial network method to find key geometric features through the learning ability of the machine learning model for data features, significantly reducing the number of geometric control parameters and improving the parameterization efficiency of turbomachine blade profiles.

[0017] 2. A parameterization method for turbomachine blade profiles based on a variational autoencoder generation model disclosed in the present invention uses the SHAP interpretable analysis tool to interpret and analyze the process of generating blade profile geometry in the latent space of the variational autoencoder generative adversarial network model, find the intuitive mapping relationship between the latent space features and the real geometry, analyze the influence between the latent space and the aerodynamic performance, lay a foundation for blade profile optimization, and then improve the aerodynamic performance of the turbomachine blade profile according to the turbomachine blade profile parameterization result.

[0018] 3. A method for parameterizing the blade profile of a turbomachine based on a variational autoencoder generative model can ensure the diversity and high degree of freedom of the generated blade profile geometry through the directional regulation of key latent space features, and can effectively improve the aerodynamic performance of the blade profile.

[0019] 4. A method for parameterizing the blade profile of a turbomachine based on a variational autoencoder generative model, on the basis of achieving the above beneficial effects 1, 2, and 3, combines the high-degree-of-freedom and fine geometric regulation advantages of the free-form deformation parameterization method, as well as the efficient dimensionality reduction advantage and high-quality blade profile generation advantage of the variational autoencoder generative model to perform parameterization of the turbomachine blade profile, and can analyze the influence between space and aerodynamic performance, ensure the diversity and high degree of freedom of the generated blade profile geometry, and improve the aerodynamic performance of the turbomachine blade profile. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of the application of the free-form deformation - variational autoencoder generative adversarial network blade profile parameterization method in the traditional blade profile optimization process;

[0021] Figure 2 Schematic diagram of the network structure in the training process of the variational autoencoder generative adversarial network model;

[0022] Figure 3 Schematic diagram of the deformation of the blade profile geometry by moving the control volume framework using the free-form deformation method;

[0023] Figure 4 Schematic diagram of the blade profile generated by the variational autoencoder generative adversarial network model with the number of training steps;

[0024] Figure 5 Schematic diagram of the latent space vector of the variational autoencoder generative adversarial network model;

[0025] Figure 6 Schematic diagram of the training accuracy verification of the XGBoost model;

[0026] Figure 7 Schematic diagram of the mapping between latent space features and blade profile geometry;

[0027] Figure 8 Influence of the latent space of the variational autoencoder generative adversarial network model on aerodynamic performance.

[0028] Figure 9 Comparison of blade profiles after feature adjustment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0029] To better illustrate the purpose and advantages of the present invention, the following further describes the content of the invention in conjunction with the drawings and examples.

[0030] Select a transonic compressor airfoil (whose design parameters are shown in Table 1), with an incoming flow Mach number of 1.09 and a Reynolds number of 1.9×10 6 .

[0031] As Figure 1 shown, a method for parameterizing the airfoil of a turbomachine based on a variational autoencoder generation model disclosed in this embodiment is specifically implemented as follows:

[0032] Step 1: Perform direct manipulation free-form deformation parameterization on a two-dimensional transonic planar cascade. The suction surface and pressure surface of the airfoil are each described by 128 points, for a total of 256 points to be deformed. As Figure 3 shown, the points distributed around the airfoil are the control points on the control volume framework, and the position change of each control point will change the position of the points to be deformed on the airfoil surface near it.

[0033] Table 1 Design geometric parameters of the TSG97 planar cascade

[0034]

[0035] Step 2: Uniformly select 100 points to be deformed at the positions of the suction surface and pressure surface of the airfoil and apply displacements to them to achieve geometric deformation. The free-form deformation program can apply displacements in the x and y directions to the points to be deformed, calculate the normal direction vector of the airfoil geometry at each point to be deformed, and use Latin hypercube sampling to apply displacements in the x and y directions to the points to be deformed, which is changed to applying a displacement magnitude to the normal direction of the point, and the key geometric feature ranges of the airfoil database are shown in Table 2.

[0036] Table 2 Key geometric parameter ranges of the generated samples

[0037]

[0038] Step 3: Train the variational autoencoder generative adversarial network model. The training hyperparameters are shown in Table 3. The input dimension is 516-dimensional airfoil geometry data, and the latent space dimension is set to 40 to control the complexity of the latent space. The batch size of 32 means that 32 samples are used for each training to balance training efficiency and memory usage. Use an initial learning rate of 0.0002 and combine it with the momentum parameter of the Adam optimizer to promote the stable convergence of the model. The number of training epochs is set to 1200 to ensure that the model has enough training time to learn the latent structure of the data. As Figure 4As shown, as the number of training steps increases, the generated blade profile geometry gradually becomes smoother and more similar to the blade profile geometries in the input blade profile sample library. During the training process, first, the discriminator is trained, and its losses for real samples and generated samples are calculated, and its parameters are updated. Next, the generator is trained, and the parameters of the generator are optimized by generating fake samples and calculating their loss functions. Finally, the variational autoencoder part is trained, and the ReLU activation function is used to calculate its reconstruction loss and KL divergence loss to optimize the parameters of the variational autoencoder.

[0039] Table 3 Hyperparameters of the variational autoencoder generative adversarial network

[0040]

[0041] Step 4: Use the trained model to generate new blade profiles. 5000 new blade profile geometries are generated by randomly perturbing the latent space. The relationship between the latent space vector features and the generated blade profile geometries is as Figure 5 shown. The numerical simulation tool is used to calculate the aerodynamic performance of the generated new geometries. The inlet boundary conditions are set as shown in Table 4 during the simulation calculation. Each scatter point represents a different blade profile sample, and each sample point contains three aerodynamic information, namely the abscissa angle of attack range and the ordinate minimum loss.

[0042] Table 4 Simulation boundary condition settings

[0043]

[0044] Step 5: Use SHAP to quantitatively interpret the prediction results of the trained model as the contribution of each feature to the predicted value, which can provide an explanation for complex machine learning models. The SHAP method establishes the association between input and output parameters by using the XGBoost surrogate model. On this basis, the SHAP analysis method is used to find the relationship between input and output parameters. The comparison between the training prediction values and the true values of the XGBoost surrogate model is as Figure 6 shown. The average absolute value of the SHAP values between the 40-dimensional low-dimensional vector features and the 5 aerodynamic performance parameters in all samples is calculated according to the SHAP value calculation method as Figure 7 shown. The latent space vector features with strong influence on the aerodynamic performance are calculated by the SHAP method, and the mapping relationship between these features and the blade profile geometry points is found as Figure 8 shown.

[0045] Step 6: According to Step 5, it is found that spatial feature 1 has a greater impact on the total pressure loss. By adjusting the size of feature 1 and through multiple simulation iterations, try to select better-performing features to optimize the blade profile. The optimized blade profile geometry is obtained as Figure 9 shown.

[0046] The specific description above further elaborates on the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above is only a specific embodiment of the present invention and is not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A parameterization method for the blade profile of a turbine based on a variational autoencoder generation model, characterized in that: Including the following steps: Step 1: Based on the original blade profile geometry, use the direct manipulation free-form deformation method to perform geometric deformation on it. The direct manipulation free-form deformation method, based on the free-form deformation method, can directly specify the displacement change amount of one or more points on the target geometric surface and inversely calculate the change of the control volume framework, and recalculate the positions of other points to be deformed under the updated control volume framework. Use the direct manipulation free-form deformation method to set a number of control points on the blade profile surface. In order to reduce the probability of generating deformed blade profiles, the control points are set more than 5% - 20% of the axial chord length away from the leading and trailing edges of the blade profile. Obtain the displacement magnitude of each control point through the Latin hypercube sampling method, and input the data into the direct manipulation free-form deformation method for batch processing to generate a sample library containing a large number of different blade profiles. Step 2: Organize the x and y coordinates of the points on the blade profile surface in the sample library into a one-dimensional vector as the input vector of the variational autoencoder generative adversarial network model, and train the generative model of the variational autoencoder generative adversarial network. The variational autoencoder generative adversarial network model consists of three parts. The first part is the variational autoencoder part, which includes two structures: an encoder and a decoder. The encoder maps the input parameters to the latent space, and the decoder reconstructs the latent space into output data. The KL divergence is used as the encoder loss, and the binary cross-entropy loss is used as the decoder loss. The second part is the discriminator, which is used to distinguish whether the generated sample is a real sample or a generated sample, and the probability of correct judgment by the discriminator is used as the loss in training. The third part is the joint training part of the autoencoder and the discriminator, where the autoencoder is used as the generator to perform adversarial training with the discriminator. Step 3: Use the MISES numerical simulation method to perform mesh division on each geometric sample in the leaf profile sample library generated by the variational autoencoder generative adversarial network and conduct simulation calculations to obtain the aerodynamic performance of each sample; the initial pressure condition for MISES calculation is input in the form of pressure ratio (P1 / P 01 ), where P1 is the inlet static pressure and P 01 is the inlet total pressure; set a fixed inlet Mach number, and the MISES numerical simulation method automatically iteratively calculates the inlet total pressure according to the input Mach number and outlet parameters; in order to accurately describe the flow field, use the turbulence model based on Drela correction and the AGS transition model to solve the flow field; Step 4: Establish an association between the latent space and the blade profile geometry and aerodynamic performance through the SHAP attribution analysis method. The main aerodynamic performance parameter, the total pressure loss coefficient, is expressed as: where P1* and P1 are the total inlet pressure and the static inlet pressure, and P 02 is the local total pressure; adjusting the potential space vector according to this correlation to generate a blade profile with better aerodynamic performance; the SHAP importance calculation formula is as follows: where φ i represents the SHAP value of feature i, reflecting the contribution of feature i to the model prediction result; F represents the complete set of features, that is, the set of all input features; represents the feature subset S that does not contain feature i, and the sum over subset S is taken in the formula; |S| and |F| represent the number of features in subset S and the total number of features in the complete set F, respectively; f S∪i (x S∪i ) represents the predicted value of the model trained on the feature set S∪i, that is, the model output after considering feature i; f S (x S ) represents the predicted value of the model trained on the feature set S, that is, the model output without considering feature i; f S∪i (x S∪i ) - f S (x S ) represents the incremental contribution to the model predicted value after adding feature i; represents the weight term, indicating the proportion of the probability of the occurrence of subset S among all possible combinations.

2. The parameterization method of the turbine blade profile based on the variational autoencoder generation model according to claim 1, characterized in that: It also includes Step 5: According to the association established between the latent space and the blade profile geometry and aerodynamic performance in Step 4, analyze the influence between the space and the aerodynamic performance. By directionally regulating the key latent space features, ensure the diversity and high degree of freedom of the generated blade profile geometry, and improve the aerodynamic performance of the turbine blade profile.

Images