CN115345064A - Integrated deep neural network aerodynamic modeling method integrated with physical constraints - Google Patents
Integrated deep neural network aerodynamic modeling method integrated with physical constraints Download PDFInfo
- Publication number
- CN115345064A CN115345064A CN202210684115.1A CN202210684115A CN115345064A CN 115345064 A CN115345064 A CN 115345064A CN 202210684115 A CN202210684115 A CN 202210684115A CN 115345064 A CN115345064 A CN 115345064A
- Authority
- CN
- China
- Prior art keywords
- neural network
- pod
- integrated
- aerodynamic
- deep neural
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/28—Design optimisation, verification or simulation using fluid dynamics, e.g. using Navier-Stokes equations or computational fluid dynamics [CFD]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2113/00—Details relating to the application field
- G06F2113/08—Fluids
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/14—Force analysis or force optimisation, e.g. static or dynamic forces
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation
Abstract
The invention relates to an integrated deep neural network aerodynamic modeling method integrated with physical constraints, which is used for calculating a pressure coefficient C of an airfoil surface under different flow parameter states p And corresponding aerodynamic force (C) L 、C M ) By intrinsic orthogonal decomposition technique on the pressure coefficient C p Two networks need to be constructed for the integrated deep neural network with the physical constraints reduced in dimension, and the weight and the bias item in the neural network model are adjusted through iteration for a certain number of times by adopting a gradient back propagation method, so that the difference degree between the model predicted value and the true value of the neural network is reduced under the constraint of the distribution characteristics, and the precision of the neural network is improved. Integration into physical constraintsThe deep neural network model is constructed based on a multilayer neural network, has stronger nonlinear feature learning capacity and stronger generalization capacity, and realizes high-precision prediction of aerodynamic force.
Description
Technical Field
The invention belongs to the technical field of aerodynamics, and relates to an integrated deep neural network aerodynamic modeling method integrated with physical constraints.
Background
In the field of aerodynamics, a pneumatic load is one of the most important loads in design and use of an aircraft, and is the key input of links such as design of the aerodynamic appearance of the aircraft, structural design, strength check, design of a control system and the like, so that the prediction precision of the pneumatic load is directly related to the performance and safety of the aircraft. Traditionally, at the beginning of aircraft design, wind tunnel tests or Computational Fluid Dynamics (CFD) simulations have been generally used to obtain the aerodynamic coefficients of airfoils. However, during model design, wind tunnel testing of a large number of vehicle passes, both from financial cost and time period, is not affordable. CFD simulation is widely used in the aviation industry to analyze the aerodynamic performance of different aircraft during the design process. These simulations may reduce time and cost compared to wind tunnel tests or flight tests. However, in the aerodynamic design phase of modern aircraft, high fidelity CFD simulation is typically a computationally intensive, time consuming iterative process. Data-driven surrogate models are efficient methods of estimating approximations of a priori unknown functional distributions, typically applied to practical engineering problems, and therefore, surrogate models are beginning to be considered as alternatives to CFD tools that are reasonably predictive at present.
The patent CN201610077146.5 adopts a support vector machine regression method to provide a biaxial unsteady aerodynamic modeling method, and the patent CN201610863848.6 provides an unsteady aerodynamic model parameter prediction method based on an extreme learning machine ELM. A polynomial response surface method, a Kriging model, support vector regression, a radial basis function, a neural network model and the like are widely applied to solving the black box problems of large calculated amount and high calculated cost as common proxy models to reduce the calculation burden. The core of the methods is a mapping model between input and output, and parameters, a framework and the like of the model are adjusted through learning of a sample set, so that the model has a good fitting effect.
The prior method has the following defects:
(1) In practical engineering application, a black box model is adopted for building the aerodynamic force agent model, the method is low in interpretability and usually needs a large data volume for learning, and the acquisition cost of a sample is high. Under a small sample, the traditional direct modeling method has insufficient prediction precision and low generalization capability.
(2) The CFD calculation or wind tunnel test can obtain not only the concentration force information but also the distribution force information, and in the aerodynamic force modeling process, researchers usually only pay attention to the integrated aerodynamic force and moment information, so that the utilization rate of the distribution force information is low.
The defects show that aerodynamic modeling is performed based on black box models such as a traditional deep neural network model and a Kriging model, the number of samples needed is large, the precision of the model is not enough under small samples, the generalization capability is low, and the utilization of distribution force information is low. At present, a proxy model for integrating the distribution force information into the aerodynamic modeling process still needs to be developed.
Disclosure of Invention
Technical problem to be solved
In order to avoid the defects of the prior art, the invention provides an integrated deep neural network aerodynamic modeling method integrated with physical constraints, which makes full use of aerodynamic distribution information generated in the sampling process, and guides the neural network aerodynamic modeling process by taking the distribution information characteristics as constraints in the modeling process. According to the modeling method, the distributed load information generated by CFD calculation or wind tunnel test is introduced into the modeling process of aerodynamic force, so that model training samples can be effectively reduced, the accuracy, robustness and generalization of the model are improved, and the model is used for guiding the development of subsequent work.
Technical scheme
An integrated deep neural network aerodynamic modeling method integrated with physical constraints is characterized by comprising the following steps:
step 1, sample acquisition: calculating the pressure coefficient C of the airfoil surface under different flow parameter states by using a numerical calculation method p And corresponding aerodynamic force (C) L 、C M )；
Step 2, extracting the distribution force information characteristics: by intrinsic orthogonal decomposition technique on pressure coefficient C p Reducing the dimension of the data set to obtain corresponding POD modal characteristics, and acquiring POD coefficients corresponding to different input parameters;
step 3, fusing the distribution information characteristics into a neural network to perform aerodynamic modeling:
constructing two networks Net1 and Net2 to form an integrated deep neural network integrated with physical constraints, and performing unified training on the two networks;
the first network Net1 is constructed as follows: inputting flow parameters Ma, re and AOA, if the airfoil shape changes, adding CST parameters corresponding to the airfoil shape into the input parameters, and outputting POD coefficient alpha of distribution force information POD ；
Constructing a second network Net 2: the input and output of the first network are simultaneously used as the input of the second network, and the output is aerodynamic force (C) L 、C M ) (ii) a Where Ma is the incoming flow Mach number, re is the Reynolds number, AOA is the incoming flow incidence angle, alpha POD POD coefficient, C, extracted for features L Is a coefficient of lift, C M Is a moment coefficient;
and 4, step 4: by adopting a gradient back propagation method, the weight and the bias item in the neural network model are adjusted through multiple iterations, so that the difference degree between the model predicted value and the true value of the neural network is finally reduced under the constraint of the distribution characteristics, the precision of the neural network is improved, and the construction of the integrated deep neural network aerodynamic model is completed;
wherein: loss function of said neural network consisting of
α POD -α POD Three-part error composition in which the POD coefficients obtained after the distribution information feature extraction are applied with a scaling factor γ, i.e., γ × (α) POD -α POD )。
Advantageous effects
The invention provides an integrated deep neural network aerodynamic modeling method integrated with physical constraints, which is used for calculating the pressure coefficient C of the airfoil surface under different flow parameter states p And corresponding aerodynamic force (C) L 、C M ) By intrinsic orthogonal decomposition technique on the pressure coefficient C p Integrated deep neural network structure for reducing dimension and integrating physical constraintsTwo networks are built, a gradient back propagation method is adopted, and weights and bias items in a neural network model are adjusted through iteration for a certain number of times, so that the difference degree between model predicted values and true values of the neural network is finally reduced under the constraint of distribution characteristics, and the precision of the neural network is improved. The integrated deep neural network model integrated with the physical constraints is constructed based on a multilayer neural network, and has stronger nonlinear feature learning capability and stronger generalization capability so as to realize high-precision prediction of aerodynamic force.
Compared with the traditional Kriging algorithm and the deep neural network direct modeling method, the integrated deep neural network aerodynamic modeling method integrated with physical constraints disclosed by the invention can effectively improve the accuracy, robustness and generalization of aerodynamic modeling and increase the interpretability of a neural network model. Meanwhile, the integrated deep neural network aerodynamic modeling method based on the physical constraint can effectively utilize distributed load information generated in the CFD calculation or wind tunnel test process, reduce the number of learning samples, reduce the cost of sample acquisition and have important significance for guiding the development of subsequent work.
Drawings
FIG. 1 is a flow chart of integrated deep neural network aerodynamic modeling
FIG. 2 is a schematic structural diagram of aerodynamic modeling of an integrated deep neural network
Detailed Description
The invention will now be further described with reference to the following examples and drawings:
in this embodiment, specifically, for the calculation data obtained by solving the Reynolds-Averaged Navier-Stokes (RANS) equation, the characteristics of the distribution force information are merged into the aerodynamic modeling process as constraints, and the method specifically includes:
s1: determining the parameter space of the flow states Ma, re and AOA calculated by CFD to obtain the pressure coefficient distribution C in each state p Coefficient of lift C L Coefficient of moment C M Wherein aerodynamic force data (C) L 、C M ) Training samples as part of the output in the second network;
s2: by intrinsic positivePerforming dimension reduction on pressure distribution data by the cross decomposition technology, performing feature extraction, and acquiring POD coefficients alpha corresponding to different input parameters POD ；
S3: method for integrating distribution information characteristics into neural network to perform aerodynamic modeling
S301: the first network Net1 is constructed as follows: inputting flow parameters Ma, re and AOA and outputting POD coefficient alpha of distribution force information POD (ii) a A second network Net2 is constructed: the input parameters of the first network are connected to the input layer of the second network in a jumping mode, and the characteristic coefficient alpha of the force information is distributed POD At the same time as the input of the second network, the output is aerodynamic (C) L 、C M ). And finally, uniformly training the two networks so as to construct an integrated deep neural network model. The training samples of the neural network are Ma, re, AOA and alpha POD 、C L 、C M . Compared with the existing deep neural network direct modeling method, the method can comprehensively utilize the distribution information generated in the sampling process, and greatly improves the modeling precision and robustness of the existing aerodynamic modeling method.
S302: constructing a loss function of the integrated deep neural network through the training samples in the step S301, wherein the loss function is formed by
α POD -α POD Three-part error composition in which a POD coefficient obtained after the feature extraction of the distribution force information is applied with a scaling factor γ, i.e., γ × (α) POD -α POD )。
S303: training the integrated deep neural network model through the training samples in the step S301 and the loss functions in the step S302, and continuously adjusting the weights and bias terms of the multilayer neural network by adopting a gradient back propagation method when training the neural network so as to minimize the loss functions under the constraint of the distribution information characteristic coefficients and simultaneously reduce the difference degree between the predicted value and the true value of the model and improve the modeling precision of the neural network;
s4: after the training is finished, aerodynamic force test is carried out for the state out of the design parameter spaceThe state adopts an integrated deep neural network aerodynamic model integrated with physical constraints, and can obtain a lift coefficient with high efficiency and high precision
Coefficient of moment
Compared with the prior art, the integrated deep neural network aerodynamic modeling method integrated with the physical constraints directly performs aerodynamic modeling based on the deep neural network method, can comprehensively utilize distribution force data generated in the sampling process, introduces the distribution characteristic coefficients as the physical constraints into the neural network modeling process, improves the interpretability of the model, effectively improves the modeling precision, robustness and generalization, reduces the number of learning samples, and reduces the cost for obtaining the samples.
Claims (1)
1. An integrated deep neural network aerodynamic modeling method integrated with physical constraints is characterized by comprising the following steps:
step 1, sample acquisition: calculating the pressure coefficient C of the airfoil surface under different flow parameter states by using a numerical calculation method p And corresponding aerodynamic force (C) L 、C M )；
Step 2, extraction of distribution force information characteristics: by intrinsic orthogonal decomposition technique on pressure coefficient C p Reducing the dimension of the data set to obtain corresponding POD modal characteristics, and acquiring POD coefficients corresponding to different input parameters;
step 3, fusing the distribution information characteristics into a neural network to perform aerodynamic modeling:
constructing two networks Net1 and Net2 to form an integrated deep neural network integrated with physical constraints, and performing unified training on the two networks;
the first network Net1 is constructed as follows: inputting flow parameters Ma, re and AOA, if the airfoil shape changes, adding CST parameters corresponding to the airfoil shape into the input parameters, and outputting POD coefficient alpha of distribution force information POD ；
A second network Net2 is constructed: the input and output of the first network are simultaneously used as the input of the second network, and the output is aerodynamic force (C) L 、C M ) (ii) a Where Ma is the incoming flow Mach number, re is the Reynolds number, AOA is the incoming flow incidence angle, alpha POD POD coefficients, C, obtained for feature extraction L Is a coefficient of lift, C M Is a moment coefficient;
and 4, step 4: by adopting a gradient back propagation method, the weight and the bias item in the neural network model are adjusted through multiple iterations, so that the difference degree between the model predicted value and the true value of the neural network is finally reduced under the constraint of the distribution characteristics, the precision of the neural network is improved, and the construction of the integrated deep neural network aerodynamic model is completed;
wherein: loss function of said neural network consisting of
α POD -α POD Three-part error composition in which the POD coefficients obtained after the distribution information feature extraction are applied with a scaling factor γ, i.e., γ × (α) POD -α POD )。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210684115.1A CN115345064A (en) | 2022-06-17 | 2022-06-17 | Integrated deep neural network aerodynamic modeling method integrated with physical constraints |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210684115.1A CN115345064A (en) | 2022-06-17 | 2022-06-17 | Integrated deep neural network aerodynamic modeling method integrated with physical constraints |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115345064A true CN115345064A (en) | 2022-11-15 |
Family
ID=83948164
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210684115.1A Pending CN115345064A (en) | 2022-06-17 | 2022-06-17 | Integrated deep neural network aerodynamic modeling method integrated with physical constraints |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115345064A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117556713A (en) * | 2024-01-11 | 2024-02-13 | 中国空气动力研究与发展中心计算空气动力研究所 | Uncertainty quantization method for CFD multi-credibility high-dimensional correlation flow field |
-
2022
- 2022-06-17 CN CN202210684115.1A patent/CN115345064A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117556713A (en) * | 2024-01-11 | 2024-02-13 | 中国空气动力研究与发展中心计算空气动力研究所 | Uncertainty quantization method for CFD multi-credibility high-dimensional correlation flow field |
| CN117556713B (en) * | 2024-01-11 | 2024-04-02 | 中国空气动力研究与发展中心计算空气动力研究所 | Uncertainty quantization method for CFD multi-credibility high-dimensional correlation flow field |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Huang et al. | Research on multi-fidelity aerodynamic optimization methods | |
| Duru et al. | A deep learning approach for the transonic flow field predictions around airfoils | |
| Biancolini et al. | Static aeroelastic analysis of an aircraft wind-tunnel model by means of modal RBF mesh updating | |
| CN114235330A (en) | Multi-source pneumatic load model construction method for correlated wind tunnel test and calculation data | |
| CN115345064A (en) | Integrated deep neural network aerodynamic modeling method integrated with physical constraints | |
| Loeven et al. | Airfoil analysis with uncertain geometry using the probabilistic collocation method | |
| Xing et al. | Fusing sensor data with CFD results using gappy POD | |
| Li et al. | Aerodynamic shape optimization by continually moving ROM | |
| Wang et al. | A nacelle inlet design approach with more three-dimensional geometric consideration | |
| CN113849910B (en) | Dropout-based BiLSTM network wing resistance coefficient prediction method | |
| Chen et al. | Approximate Bayesian assisted inverse method for identification of parameters of variable stiffness composite laminates | |
| Wang et al. | Implicit large eddy simulation of the NASA CRM high-lift configuration near stall | |
| CN111177855B (en) | Pneumatic structure solving method and system in global aeroelasticity optimization | |
| Pisaroni et al. | A multilevel monte carlo evolutionary algorithm for robust aerodynamic shape design | |
| Akbarzadeh et al. | A Cell-Elimination method for solving steady and unsteady Navier–Stokes equations | |
| Liu et al. | Data augmented prediction of Reynolds stresses for flows around an axisymmetric body of revolution | |
| CN117291109B (en) | Modelica fluid model intelligent prediction method | |
| Alswaitti et al. | Dimensionality Reduction, Modelling, and Optimization of Multivariate Problems Based on Machine Learning. Symmetry 2022, 14, 1282 | |
| CN117874925A (en) | Gradient type pneumatic shape optimization method, system and equipment based on deep learning | |
| CN115495854B (en) | Parameter calibration method, device, equipment and medium of computer aided engineering model | |
| CN110309544B (en) | Stress point prediction method for large deformation of beam based on neural network | |
| CN116628854A (en) | Wing section aerodynamic characteristic prediction method, system, electronic equipment and storage medium | |
| CN107463734A (en) | Motor train unit body modal calculation method and device | |
| Deng et al. | A novel quasi-smooth tetrahedral numerical manifold method and its application in topology optimization based on parameterized level-set method | |
| Stanford et al. | Ongoing Aeroelastic Prediction and Validation Activities at NASA Langley Research Center |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |