CN-121980948-A - Blade pneumatic robustness assessment method based on dual fidelity cooperative neural network

CN121980948ACN 121980948 ACN121980948 ACN 121980948ACN-121980948-A

Abstract

The invention discloses a blade pneumatic robustness assessment method based on a dual-fidelity cooperative neural network, and relates to the technical field of pneumatic analysis of turbomachine blades. The method comprises the steps of obtaining a plurality of groups of uncertainty parameters affecting aerodynamic characteristics of a blade, inputting the uncertainty parameters into a dual-fidelity collaborative neural network model, extracting feature vectors of the uncertainty parameters through a basic trend prediction module, predicting low-fidelity aerodynamic performance, weighting the feature vectors through a cross-fidelity attention fusion module, fusing the uncertainty parameters with the weighted feature vectors, inputting the fused vectors into a residual error correction module to obtain residual error prediction values, adding the residual error prediction values and the low-fidelity aerodynamic performance prediction values to obtain high-fidelity aerodynamic performance prediction values, and determining the robustness index of the blade according to the high-fidelity aerodynamic performance prediction values corresponding to the plurality of groups of uncertainty parameters. The method improves the pneumatic performance precision of the blade and ensures the calculation efficiency.

Inventors

XIE YONGHUI
LI GUOJIA
ZHANG DI
Li Chaokao

Assignees

西安交通大学

Dates

Publication Date: 20260505
Application Date: 20260129

Claims (7)

1. The blade pneumatic robustness assessment method based on the dual fidelity cooperative neural network is characterized by comprising the following steps of: the method comprises the steps of obtaining a dual-fidelity collaborative neural network model, wherein the dual-fidelity collaborative neural network model comprises a basic trend prediction module, a cross-fidelity attention fusion module and a residual error correction module, wherein the basic trend prediction module is used for rapidly capturing basic physical rules of pneumatic performance of a blade by utilizing low-fidelity data; Generating a plurality of groups of uncertainty parameters affecting the aerodynamic characteristics of the blade by adopting a Monte Carlo simulation method, wherein each group of uncertainty parameters comprises boundary condition parameters, material property parameters and geometric deviation of the blade; Aiming at any group of uncertainty parameters, the uncertainty parameters are input into a dual-fidelity collaborative neural network model, feature vectors of the uncertainty parameters are extracted through a basic trend prediction module, and low-fidelity aerodynamic performance is predicted; the method comprises the steps of weighting a feature vector through a cross-fidelity attention fusion module, fusing uncertainty parameters with the weighted feature vector, and inputting the fused vector into a residual error correction module to obtain a residual error predicted value; and determining a robustness index of the blade according to the high-fidelity aerodynamic performance predicted values corresponding to the multiple groups of uncertainty parameters, wherein the robustness index is used for reflecting aerodynamic characteristics of the blade.
2. The method of claim 1, wherein the base trend prediction module and the residual correction module are each constructed from a fully connected deep neural network, the cross-fidelity attention fusion module comprises a multi-layer perceptron, an activation function and a multiplication unit connected in sequence, and the weighting of the feature vectors by the cross-fidelity attention fusion module comprises: The uncertainty parameters and the feature vectors are spliced and then sequentially pass through a multi-layer perceptron and an activation function to obtain attention weight vectors; and multiplying the attention weight vector by the feature vector to obtain a weighted feature vector.
3. The method of claim 2, wherein the process of constructing the dual-fidelity collaborative neural network model comprises: The method comprises the steps of obtaining a low-fidelity training sample set and a high-fidelity training sample set, wherein the low-fidelity training sample set comprises low-fidelity sample uncertain parameters and corresponding low-fidelity pneumatic performance real values; Training the basic trend prediction module through a low-fidelity training sample set; after the basic trend prediction module is trained, parameters of the basic trend prediction module are frozen, and the cross-fidelity attention fusion module and the residual error correction module are trained through the high-fidelity training sample set, so that the dual-fidelity collaborative neural network model is obtained.
4. A method according to claim 3, wherein the loss function employed in the training of the underlying trend prediction module is a mean square error; the loss function adopted in the training process of the cross-fidelity attention fusion module and the residual error correction module is as follows: Wherein, the In order to achieve a value of the loss function, For high fidelity training of the sample number in the sample set, Is the first High fidelity pneumatic performance realism values corresponding to the high fidelity samples, Is the first Low fidelity pneumatic performance realism values corresponding to the high fidelity samples, Is the first Residual prediction values corresponding to the high fidelity samples, In order for the attention deficit regularization term, In order for the regularization factor to be a good, Is the first Attention weight vectors corresponding to the high fidelity samples.
5. A method according to claim 3, characterized in that the method further comprises: and normalizing the input data and the output data of the dual-fidelity collaborative neural network model to [ -1,1] by adopting a linear normalization method.
6. The method of claim 3, wherein obtaining a low fidelity training sample set and a high fidelity training sample set comprises: Respectively establishing a high-fidelity pneumatic analysis model and a low-fidelity pneumatic analysis model by constructing grids with different scales, wherein the high-fidelity pneumatic analysis model adopts fine grids, and the low-fidelity pneumatic analysis model adopts coarse grids; Generating N L low-fidelity sample uncertainty parameters by using a Latin hypercube sampling mode, and selecting N H samples from the N L low-fidelity sample uncertainty parameters according to a Max-Min criterion to serve as high-fidelity sample uncertainty parameters; Evaluating uncertainty parameters of N L low-fidelity samples through a low-fidelity pneumatic analysis model, and calculating the low-fidelity pneumatic performance true value of the uncertainty parameters to obtain a low-fidelity training sample set; And evaluating uncertainty parameters of N H high-fidelity samples through a high-fidelity pneumatic analysis model, and calculating the high-fidelity pneumatic performance true value of the uncertainty parameters to obtain a high-fidelity training sample set.
7. The method of claim 1, wherein determining the robustness index of the blade based on the high fidelity aerodynamic performance predictions for the plurality of sets of uncertainty parameters comprises: Calculating the mean value and standard deviation of all high-fidelity aerodynamic performance predicted values; And determining the ratio of the mean value to the standard deviation as the robustness index of the blade, wherein the larger the robustness index is, the more robust the blade performance is, and the better the robustness is.

Description

Blade pneumatic robustness assessment method based on dual fidelity cooperative neural network Technical Field The application relates to the technical field of pneumatic analysis of turbine mechanical blades, in particular to a blade pneumatic robustness assessment method based on a dual-fidelity cooperative neural network. Background The blade is used as a core component of power machines such as a gas turbine, a steam turbine and the like, pneumatic analysis is a key link of design optimization and reliable operation, and the economy of the whole machine is directly determined. However, the blades are subjected to various loads in operation, resulting in an abnormally complex aerodynamic characteristic thereof, and various uncertainty factors exist during operation, such as inlet gas temperature, pressure, random fluctuation of the incident angle of the gas flow, geometric deviation of the blade profile due to machining and manufacturing, and the like. These multi-source uncertainty factors can significantly affect the aerodynamic properties of the blade, thereby causing performance fluctuations of the turbomachine that deviate from intended. Therefore, pneumatic evaluation of the blade is extremely necessary. Most of the existing pneumatic evaluation methods are based on Shan Baozhen-degree proxy models for evaluation, however, when facing high-fidelity data with high single sampling cost, the accuracy and the calculation efficiency are often mutually restricted, and the simultaneous consideration of the accuracy and the calculation efficiency is difficult. Disclosure of Invention Based on the above, it is necessary to provide a blade aerodynamic robustness assessment method based on a dual fidelity cooperative neural network. The technical scheme adopted in the specification is as follows: the specification provides a blade pneumatic robustness assessment method based on a dual fidelity cooperative neural network, which comprises the following steps: the method comprises the steps of obtaining a dual-fidelity collaborative neural network model, wherein the dual-fidelity collaborative neural network model comprises a basic trend prediction module, a cross-fidelity attention fusion module and a residual error correction module, wherein the basic trend prediction module is used for rapidly capturing basic physical rules of pneumatic performance of a blade by utilizing low-fidelity data; Generating a plurality of groups of uncertainty parameters affecting the aerodynamic characteristics of the blade by adopting a Monte Carlo simulation method, wherein each group of uncertainty parameters comprises boundary condition parameters, material property parameters and geometric deviation of the blade; Aiming at any group of uncertainty parameters, the uncertainty parameters are input into a dual-fidelity collaborative neural network model, feature vectors of the uncertainty parameters are extracted through a basic trend prediction module, and low-fidelity aerodynamic performance is predicted; the method comprises the steps of weighting a feature vector through a cross-fidelity attention fusion module, fusing uncertainty parameters with the weighted feature vector, and inputting the fused vector into a residual error correction module to obtain a residual error predicted value; and determining a robustness index of the blade according to the high-fidelity aerodynamic performance predicted values corresponding to the multiple groups of uncertainty parameters, wherein the robustness index is used for reflecting aerodynamic characteristics of the blade. Optionally, the basic trend prediction module and the residual correction module are both constructed according to a fully-connected deep neural network, and the cross-fidelity attention fusion module comprises a multi-layer perceptron, an activation function and a multiplication unit which are sequentially connected, wherein the cross-fidelity attention fusion module weights the feature vectors and comprises: The uncertainty parameters and the feature vectors are spliced and then sequentially pass through a multi-layer perceptron and an activation function to obtain attention weight vectors; and multiplying the attention weight vector by the feature vector to obtain a weighted feature vector. Optionally, the construction process of the dual-fidelity collaborative neural network model comprises the following steps: The method comprises the steps of obtaining a low-fidelity training sample set and a high-fidelity training sample set, wherein the low-fidelity training sample set comprises low-fidelity sample uncertain parameters and corresponding low-fidelity pneumatic performance real values; Training the basic trend prediction module through a low-fidelity training sample set; after the basic trend prediction module is trained, parameters of the basic trend prediction module are frozen, and the cross-fidelity attention fusion module and the residual error correction mo