Search

CN-122020858-A - Modeling method of intelligent pneumatic prediction model based on physical knowledge constraint

CN122020858ACN 122020858 ACN122020858 ACN 122020858ACN-122020858-A

Abstract

The invention provides an intelligent pneumatic prediction model modeling method based on physical knowledge constraint, which comprises the following five steps of (1) obtaining aerodynamic physical knowledge through data mining, (2) obtaining a function expression of aerodynamic force, (3) adopting a depth operator neural network to approximately replace the function expression of aerodynamic force, (4) training the depth operator neural network, and (5) developing full-section aerodynamic modeling based on the trained depth operator neural network.

Inventors

  • Peng Xuhao
  • WANG DING
  • LIU YUANCHUN
  • GAO YUAN
  • TIAN CHUAN
  • SUN JINGYANG
  • DUAN YI
  • LI SIYI
  • LI WENHAO
  • WANG XIAOFENG
  • TAN YANG
  • MIAO MENG
  • LI HAOGE
  • ZHOU NAIZHEN

Assignees

  • 北京临近空间飞行器系统工程研究所

Dates

Publication Date
20260512
Application Date
20260211

Claims (11)

  1. 1. A modeling method of an intelligent pneumatic prediction model based on physical knowledge constraint is characterized by comprising the following five steps of (1) obtaining aerodynamic physical knowledge through data mining, (2) obtaining a function expression of aerodynamic force, (3) adopting a depth operator neural network to approximately replace the function expression of aerodynamic force, (4) training the depth operator neural network, and (5) developing full-section aerodynamic modeling based on the trained depth operator neural network.
  2. 2. The method of modeling a physical knowledge constrained intelligent aerodynamic prediction model of claim 1, wherein the step (2) comprises the steps of: obtaining a function expression of aerodynamic force according to the analysis result of data mining: , wherein, For the mach number of the flight, In order to be able to take the altitude of the flight, In order to achieve an angle of attack, As the slip angle of the slide-in plate, 、 And All representing rudder deflection.
  3. 3. The method for modeling an intelligent aerodynamic prediction model based on physical knowledge constraints according to claim 1, wherein the step (3) comprises the following steps: Approximating substitution functions using two neural networks with n-dimensional outputs, respectively And , wherein, 。
  4. 4. The method of modeling a physical knowledge constrained intelligent aerodynamic prediction model of claim 1, wherein the step (4) comprises the steps of: 1) Forward propagation; 2) Minimizing a loss function; 3) And updating the weight parameters of the depth operator neural network.
  5. 5. The method for modeling an intelligent aerodynamic prediction model based on physical knowledge constraints according to claim 4, wherein the step 1) comprises the following steps: Forward propagation of depth operator neural networks involves activation of all neurons from the input layer to the output layer, item i Of layers of The neuron receives the first -1 Layer The input of the individual neurons, the vector form of the activation function is: 。
  6. 6. the method for modeling a physical knowledge constrained intelligent aerodynamic prediction model according to claim 4, wherein the step 2) comprises the steps of: the loss function is used for measuring the difference between the output of the depth operator neural network and the real output, and guiding the training process by minimizing the loss function, wherein the loss function expression is as follows: , wherein, Is the number of training samples to be used, Is the predicted output of the device, Is the true output.
  7. 7. The method for modeling a physical knowledge constrained intelligent aerodynamic prediction model according to claim 4, wherein the step 3) comprises the steps of: 7) Calculating a gradient; 8) Updating the first moment estimate and the second moment estimate; 9) Correcting the first moment estimate and the second moment estimate; 10 Updating the weight parameters using the corrected first moment estimate and second moment estimate.
  8. 8. The method for modeling a physical knowledge constrained intelligent aerodynamic prediction model according to claim 7, wherein the step a) comprises the steps of: calculating the gradient of the loss function with respect to the weight parameters: Wherein, the method comprises the steps of, The weight parameter is represented by a number of weight parameters, Representing the current time step.
  9. 9. The method of modeling a physical knowledge constrained intelligent aerodynamic prediction model of claim 7, wherein said step b) comprises the steps of: Updating the first moment estimate: Updating the second moment estimate: Wherein, the method comprises the steps of, Is the exponential decay rate of the first moment estimate, Is the exponential decay rate of the second moment estimate, Representation of Is the element-wise square of (c).
  10. 10. The method for modeling a physical knowledge constrained intelligent aerodynamic prediction model according to claim 7, wherein said step c) comprises the steps of: And Initialized to a zero vector, and, at an initial time step, And Will bias to zero vector, calculate offset corrected first moment estimate Sum and second moment estimation 。
  11. 11. The method for modeling a physical knowledge constrained intelligent aerodynamic prediction model according to claim 7, wherein said step d) comprises the steps of: using corrected first moment estimates Sum and second moment estimation Updating weight parameters: Wherein, the method comprises the steps of, Is the initial learning rate of the device, Is constant.

Description

Modeling method of intelligent pneumatic prediction model based on physical knowledge constraint Technical Field The invention belongs to the technical field of intelligent algorithms, and is used for aerodynamic force prediction and modeling of a high-speed aircraft. Background In the field of aircraft design, pneumatic optimization design is a core link for improving the performance of an aircraft, reducing energy consumption and guaranteeing flight safety. Because aerodynamic problems typically involve multivariable coupling of altitude, mach number, angle of attack, etc., and have complex nonlinear characteristics, the optimization process is extremely challenging. The existing aerodynamic prediction mainly depends on a data interpolation method or a traditional neural network model, but the methods lack of embedding physical knowledge, and the prediction precision is difficult to further improve when facing unknown parameters. In addition, although the existing deep neural network model shows better fitting capability in the pneumatic problem, the model structure is generally complex, the set parameters are more, the dependence on the physical rule is insufficient, and the interpretability and generalization capability of the model are difficult to improve. Therefore, developing a aerodynamic modeling method that has high prediction accuracy and can be guided by combining physical principles is an important point of research in this field. The depth operator neural network is established on the basis of the function approximation convergence theory of the math scholars of the complex denier university in the last century, has been verified in the aspect of partial differential equation solving, but the model is not applied to aerodynamic modeling. Disclosure of Invention The invention aims to provide an intelligent pneumatic prediction model modeling method based on physical knowledge constraint, which is used for constructing a depth operator neural network applicable to complex nonlinear aerodynamic modeling by combining aerodynamic physical knowledge with the depth neural network, so that the prediction precision and physical interpretability of aerodynamic force under unknown conditions are improved. The invention provides an intelligent pneumatic prediction model modeling method based on physical knowledge constraint, which comprises the following five steps of (1) obtaining aerodynamic physical knowledge through data mining, (2) obtaining a function expression of aerodynamic force, (3) adopting a depth operator neural network to approximately replace the function expression of aerodynamic force, (4) training the depth operator neural network, and (5) developing full-section aerodynamic modeling based on the trained depth operator neural network. Further, the step (2) includes the steps of: obtaining a function expression of aerodynamic force according to the analysis result of data mining: , wherein, For the mach number of the flight,In order to be able to take the altitude of the flight,In order to achieve an angle of attack,As the slip angle of the slide-in plate,、AndAll representing rudder deflection. Further, the step (3) includes the steps of: Approximating substitution functions using two neural networks with n-dimensional outputs, respectively And, wherein,。 Further, the step (4) includes the steps of: 1) Forward propagation; 2) Minimizing a loss function; 3) And updating the weight parameters of the depth operator neural network. Further, the step 1) includes the steps of: Forward propagation of depth operator neural networks involves activation of all neurons from the input layer to the output layer, item i Of layers ofThe neuron receives the first-1 LayerThe input of the individual neurons, the vector form of the activation function is:。 further, the step 2) includes the steps of: the loss function is used for measuring the difference between the output of the depth operator neural network and the real output, and guiding the training process by minimizing the loss function, wherein the loss function expression is as follows: , wherein, Is the number of training samples to be used,Is the predicted output of the device,Is the true output. Further, the step 3) includes the steps of: a) Calculating a gradient; b) Updating the first moment estimate and the second moment estimate; c) Correcting the first moment estimate and the second moment estimate; d) The weight parameters are updated using the corrected first and second moment estimates. Further, the step a) includes the steps of: calculating the gradient of the loss function with respect to the weight parameters: Wherein, the method comprises the steps of, The weight parameter is represented by a number of weight parameters,Representing the current time step. Further, the step b) includes the steps of: Updating the first moment estimate: Updating the second moment estimate: Wherein, the method comprises the steps of, Is the exponen