CN114722498A - Wing section design method based on generation of countermeasure network - Google Patents
Wing section design method based on generation of countermeasure network Download PDFInfo
- Publication number
- CN114722498A CN114722498A CN202210350609.6A CN202210350609A CN114722498A CN 114722498 A CN114722498 A CN 114722498A CN 202210350609 A CN202210350609 A CN 202210350609A CN 114722498 A CN114722498 A CN 114722498A
- Authority
- CN
- China
- Prior art keywords
- network
- airfoil
- data
- output
- lift
- 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]
- G06F30/10—Geometric CAD
- G06F30/15—Vehicle, aircraft or watercraft design
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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 provides an airfoil design method based on a generation countermeasure network. The method learns by using the randomly generated profile data as training data. The generating network can generate the wing profile according with the lift-drag coefficient by inputting the target lift-drag coefficient. The problem that the lift-drag coefficient and the wing profile have one-to-many mapping relation is solved by adding a hidden code into a generating network. A covert code is a set of numbers that obey a certain distribution (gaussian in the present invention). The output of the generated network, i.e. the wing profile, can be varied by repeated random sampling of the covert code. Finally, diversified wing profiles meeting the target lift-drag coefficient can be obtained.
Description
Technical Field
The invention relates to a wing section design method based on a generation countermeasure network, which can efficiently design a wing section conforming to a lift-drag coefficient under a certain flight state according to the lift-drag coefficient, can generate diversified wing sections by repeatedly sampling hidden codes in input, meets the target lift-drag coefficient, and belongs to the field of aircraft appearance design.
Background
Aircraft design is a large and complex project that relies heavily on the design experience of engineers. The problem that the design cycle time is long exists in the design of the aircraft at the present stage, wherein each design stage needs a large amount of simulation verification, a large amount of data is generated, how to combine deep learning with the design process of the aircraft, self-learning is realized through data modeling, the design efficiency is improved, the cost is reduced, the design cycle is reduced, the design automation process is optimized, and the design method has important engineering and theoretical significance.
Disclosure of Invention
In order to realize the combination of deep learning and an aircraft design process, the problems that engineering data are few in the aircraft pneumatic design process, a one-to-many mapping relation exists in the aircraft design process and the like are solved through the deep learning. The invention constructs a wing profile design model based on a generation countermeasure network (GAN), and obtains a wing profile capable of meeting target lift and drag coefficients by inputting a lift coefficient and a drag coefficient; for a group of lift coefficient and drag coefficient, a plurality of airfoils possibly correspond to or approximate to the lift coefficient and the drag coefficient, so that hidden codes which obey Gaussian distribution are added into the input of the neural network, the shape of the airfoil is changed by randomly sampling the hidden codes, and the one-to-many mapping problem is solved.
In order to achieve the purpose, the technical idea adopted by the invention is as follows:
by randomly sampling hidden codes which obey Gaussian distribution and taking the hidden codes and the target aerodynamic coefficient as the input of the generation network G, the generation network outputs diversified airfoil geometric data, so that the airfoil design is realized. The training process of the encoder E and the generated network specifically includes:
at step S1, airfoil data is collected for training. The airfoil data comprises geometric data of an airfoil and target aerodynamic coefficient data, wherein the geometric data of the airfoil are coordinates x and y of discrete points of the airfoil, and x of different airfoils is fixed and unified to the same value. The target aerodynamic coefficient C is a lift coefficient ClAnd/or coefficient of resistance Cd。
And step S2, constructing an encoder E to input the airfoil profile data y and output the distribution z of the hidden codes. Expressed as:
whereinIs a standard Gaussian distribution; mu.szAnd σzAn average value and a standard deviation of z, respectively, indicate a dot product.
And step S3, random sampling of the hidden code. Wherein the hidden code z1Randomly sampling from z, latent code z2FromRandomly sampling;
step S4, converting the hidden code z1And z2Respectively forming an array with the target parameters as the input of a generation network G, and outputting predicted airfoil profile coordinates y 'by the generation network G'1、y′2. Generation of airfoil data y 'with encoder E input simultaneously'2Output hidden code z'2。
Step S5, constructing a discrimination network D to input the airfoil profile data y and y'1、y′2And a target aerodynamic coefficient, and judging the authenticity of the data;
step S6, updating the neural network parameters through the target function training, finishing the training when the generating network G reaches convergence, and obtaining the trained generating network and the hidden codes which obey Gaussian distribution, wherein the target function is expressed as:
wherein λ isb、λz、λKLIs a hyper-parameter;
wherein p isdataData representing a training data set;to expect, | × | purple1D (×) represents the output of the discriminator D, G (×) represents the output of the generation network G, E (×) represents the output of the encoder E,to measure the similarity of two distributions for divergence, p (y) represents the implicit function of y.
Further, the step S1 specifically includes the following sub-steps:
step S1.1, fixing x as
Wherein n ispThe number of discrete points on the upper surface of the airfoil profile is the same as the x-axis coordinate of the discrete points on the lower surface of the airfoil profile; i is 1 to npIs a positive integer of (1). Generating an airfoil shape by using a CST parameterization method to obtain a geometric shape represented by the sample point;
step S1.2, performing aerodynamic calculation on the airfoil shape generated in the step S1.1 under a certain inflow condition by adopting open source aerodynamic software XFOIL;
and S1.3, analyzing the aerodynamic calculation result in the step S1.2 to obtain a required aerodynamic coefficient and obtain airfoil profile data.
Further, in step S2, the network structure of the encoder E is 3 fully-connected layers, and each fully-connected layer is followed by a batch normalization layer and a LeakyReLU activation function.
Further, the network structure for generating the network G employs a Unet network.
Further, in step S5, the network structure of the network D is determined to be connected by multiple fully connected layers, and each fully connected layer is followed by the batch normalization layer and the ReLU activation function.
Further, the airfoil with the inconsistent lift-drag coefficient is removed from the airfoil geometric data of the diversity output by the generation network, and the airfoil with the largest lift-drag ratio is selected as a design result.
Further, in the step S6, the convergence criterion is
The invention has the beneficial effects that:
1. by inputting a target lift-drag coefficient and a hidden code which follows Gaussian distribution, the generating network can generate an airfoil profile which meets the target lift-drag coefficient under certain flight conditions.
2. By repeatedly and randomly sampling the hidden codes in the generated network input, the result of the generated network can be changed, and different wing profiles can be obtained, wherein the wing profiles can meet the target lift-drag coefficient under certain flight conditions. The problem that the lift drag coefficient and the wing profile have one-to-many mapping relation is solved to a certain extent.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a sampling range of airfoil training data and test data;
FIG. 3 is a schematic diagram of various neural network structures;
FIG. 4 is a schematic diagram of a Unet network structure in a generating network;
FIG. 5 is a cross-sectional view of a mission of an unmanned aerial vehicle;
FIG. 6 shows the design results of different airfoils obtained by the generation network according to the present invention after inputting the same lift coefficient and different hidden codes;
FIG. 7 is a result of overlaying the different airfoils of FIG. 6;
fig. 8 is a schematic diagram of the wing profile design result provided by the present invention applied to an unmanned aerial vehicle.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings.
FIG. 1 is a flow chart of an airfoil design method based on generation of a countermeasure network according to the present invention.
The present embodiment contemplates a drone aerofoil that has an engine start warm-up, a roll-off, a takeoff, a climb acceleration, a cruise, a descent, a landing, a roll-off, a shutdown (as shown in fig. 5). The unmanned aerial vehicle is mainly determined to have the flight performance of the unmanned aerial vehicle during stalling, takeoff, climbing rate, climbing speed and cruise. The cruise stage of the aircraft is selected as a design target, and the lift coefficient required by the aircraft in the cruise state is 0.2272 and is used as an input, so that the airfoil profile meeting the target is obtained.
Specifically, the method for training the encoder and generating the network to obtain the hidden code obeying the Gaussian distribution comprises the following steps:
step S1, airfoil data and corresponding lift coefficient data are collected for training. Specifically, the method comprises the following steps:
s1.1, randomly generating wing profile data by adopting a CST parameterization method, wherein the sampling range of the training data and the testing data of the wing profile is shown in figure 2. The geometrical data of the airfoil is that the coordinates x, y and x of discrete points of the airfoil are fixed
Wherein n ispTaking 99 out of the embodiment, the x-axis coordinate of the discrete points on the lower surface of the airfoil is the same as that of the upper surface; i is 1 to npIs a positive integer of (1).
Step S1.2, adopting open source aerodynamic software XFOIL to make the incoming flow speed of the airfoil shape generated in step S1.1 be 50m/S at an attack angle of 1.1319 degrees and the Reynolds number be 1.46 multiplied by 106Performing aerodynamic calculations under the conditions of (a);
and S1.3, analyzing the aerodynamic calculation result in the step S1.2 to obtain a lift coefficient.
In step S2, the encoder E inputs the airfoil data y and outputs a hidden code distribution z, denoted as E (y). The encoder E is a neural network for reducing the high-dimensional data samples into low-dimensional data, the input is the y value of the airfoil data, and the output is the distribution z of the hidden codes. The network structure of encoder E (fig. 3) is 3 layers of fully-connected layers, each layer of fully-connected layers followed by a bulk normalization layer and a leakyreu activation function. Since the output implicit code is approximately normal to the standard distribution, the output of the encoder E is the average value mu of zzAnd standard deviation σz。
and step S3, random sampling of the hidden code. Wherein the hidden code z1From E (y) random sampling, implicit code z2FromRandom sampling;
step S4, generating network G input lift coefficient ClAnd a hidden code z1Generating airfoil y'1. Wherein the generation network G is used for generating diversified airfoils with the input of lift coefficient ClAnd hidden codeThe output is the y value which generates the airfoil profile coordinate points. The network structure (fig. 3) of the generated network G adopts a net network model, and the network structure is relatively symmetrical and is often used for models with consistent input and output dimensions. The diversity of the wing profiles is realized by repeatedly sampling the hidden code z from the standard normal distribution;
step S5, generating a network G input lift coefficient ClAnd a hidden code z2Generating airfoil y'2. Wherein, the generation network G is the same as the generation network in step S4;
in step S6, encoder E inputs generation airfoil profile data y'2Output hidden code z'2. Wherein, the encoder E is the same as the encoder in step S2;
in step S7, airfoil profile data y and y 'are inputted to the discrimination network D'1、y′2And coefficient of lift ClTo obtain an output D (C)l,y)、D(Cl,y′1)、D(Cl,y′2). The judgment network D is used as a judge to distinguish whether the input data is from the real sample or generated by the generation network G, if the input data is from the real sample, the output tends to 1, and if the input data is from the generation network G, the output tends to 0. The network structure of discrimination network D (fig. 3) is connected by multiple fully-connected layers, each followed by a bulk normalization layer and a ReLU activation function.
Step S8, calculating an objective function:
wherein p isdataData representing a training data set; wherein D (_) denotes the output of the discriminator D, G (_) denotes the output of the generation network G, E (_) denotes the output of the encoder E,to measure the similarity of two distributions for divergence, p (y) represents the implicit function of y.
Step S9, updating the neural network parameters through the objective function:
wherein λ isb、λz、λKLFor hyper-parameters, here 10, 1, 10;
in step S10, the training is repeated from step two, and when the generated network G converges, the training is terminated. Wherein the convergence criterion is
After training is finished, the wing profile is designed by taking the lift coefficient of the wing to be 0.2272 in the cruise state as a final consideration in the embodiment. The hidden code is repeatedly sampled to obtain a design result shown in figure 6, 16 different wing profiles can be selected, the wing profiles are numbered from left to right and from top to bottom as numbers 1-16, and the lift-drag coefficient calculated by pneumatic software is arranged above the wing profiles. Overlapping these airfoils, see fig. 7, it can be seen that the resultant airfoils are relatively close in profile, differing in detail, with the major variation being at the leading edge of the airfoil lower surface.
As can be seen from the lift-drag coefficients of the airfoils shown in fig. 6, the lift coefficients of the airfoils nos. 2, 3, 7 and 14 do not reach 0.2272, which is not satisfactory for the intended purpose, and thus are left out. In order to compare the advantages and disadvantages of other airfoils, the lift-drag ratio is calculated to obtain the following table. The following table shows that the lift-drag ratio of the 13 # airfoil is the largest, the lift coefficient is 0.251, the lift requirement of the aircraft in the cruise state is met, and the 13 # airfoil can be considered as a design result. Fig. 8 is a schematic view of a drone using a No. 13 airfoil.
It should be understood that the above examples are only for clarity of illustration and are not intended to limit the embodiments. Other variations and modifications will be apparent to persons skilled in the art in light of the above description. This need not be, nor should all embodiments be exhaustive. And obvious variations or modifications of the invention may be made without departing from the scope of the invention.
Claims (7)
1. An airfoil design method based on a generation countermeasure network is characterized in that hidden codes which obey Gaussian distribution are randomly sampled, and a target aerodynamic coefficient is used as the input of a generation network G, so that airfoil geometric data of the generation network output diversity is realized, and airfoil design is realized; the training process for generating the network G specifically includes:
step S1, collecting airfoil profile data for training; the airfoil data comprises geometric data of an airfoilThe target aerodynamic coefficient data are obtained, wherein the geometrical data of the airfoil are coordinates x and y of discrete points of the airfoil, and the x of different airfoils are fixed and respectively unified into the same value; the target aerodynamic coefficient C is a lift coefficient ClAnd/or coefficient of resistance Cd;
Step S2, constructing an encoder E to input airfoil data y and output the distribution z of the hidden codes; expressed as:
whereinIs a standard Gaussian distribution; mu.szAnd σzThe mean and standard deviation of z, respectively; an indication of a dot product;
step S3, sampling the hidden code randomly; wherein the hidden code z1Randomly sampling from z, latent code z2FromRandom sampling;
step S4, hiding the code z1And z2Respectively forming an array with the target parameters as the input of a generation network G, and outputting predicted airfoil profile coordinates y 'by the generation network G'1、y′2(ii) a Generation of airfoil data y 'with encoder E input simultaneously'2Output hidden code z'2;
Step S5, constructing a discrimination network D to input the airfoil profile data y and y'1、y′2And the target aerodynamic coefficient, and judging the authenticity of the data;
step S6, updating the neural network parameters through the target function training, finishing the training when the generating network G reaches convergence, and obtaining the trained generating network and the hidden codes which obey Gaussian distribution, wherein the target function is expressed as:
wherein λ isb、λz、λKLIs a hyper-parameter;
wherein p isdataData representing a training data set;to expect, | | × | cage1D (×) represents the output of the discriminator D, G (×) represents the output of the generation network G, E (×) represents the output of the encoder E,to measure the similarity of two distributions for divergence, p (y) represents the implicit function of y.
2. The method as claimed in claim 1, wherein the step S1 comprises the following sub-steps:
step S1.1, fixing x as
Wherein n ispThe number of discrete points on the upper surface of the airfoil profile is the same as the x-axis coordinate of the discrete points on the lower surface of the airfoil profile; i is 1 to npA positive integer of (d); generating an airfoil shape by using a CST parameterization method to obtain a geometric shape represented by the sample point;
step S1.2, performing aerodynamic calculation on the airfoil shape generated in the step S1.1 under a certain inflow condition by adopting open source aerodynamic software XFOIL;
and S1.3, analyzing the aerodynamic calculation result in the step S1.2 to obtain a required aerodynamic coefficient and obtain airfoil profile data.
3. The method of claim 1, wherein in the step S2, the network structure of the encoder E is 3 layers of fully-connected layers, and each layer of fully-connected layers is followed by a batch normalization layer and a leakyreu activation function.
4. The method of claim 1, wherein the network fabric that generates network G employs a Unet network.
5. The method according to claim 1, wherein in step S5, the network structure of the discrimination network D is connected by a plurality of fully-connected layers, each layer being followed by a bulk normalization layer and a ReLU activation function.
6. The method of claim 1, wherein airfoils with non-compliant lift-drag coefficients are removed from the diverse airfoil geometry data that generate the net output, and the airfoil with the largest lift-drag ratio is selected as the design result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210350609.6A CN114722498A (en) | 2022-04-02 | 2022-04-02 | Wing section design method based on generation of countermeasure network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210350609.6A CN114722498A (en) | 2022-04-02 | 2022-04-02 | Wing section design method based on generation of countermeasure network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114722498A true CN114722498A (en) | 2022-07-08 |
Family
ID=82241977
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210350609.6A Pending CN114722498A (en) | 2022-04-02 | 2022-04-02 | Wing section design method based on generation of countermeasure network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114722498A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116628854A (en) * | 2023-05-26 | 2023-08-22 | 上海大学 | Wing section aerodynamic characteristic prediction method, system, electronic equipment and storage medium |
-
2022
- 2022-04-02 CN CN202210350609.6A patent/CN114722498A/en active Pending
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116628854A (en) * | 2023-05-26 | 2023-08-22 | 上海大学 | Wing section aerodynamic characteristic prediction method, system, electronic equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106529093B (en) | A kind of pneumatic/structure/pneumostatic bullet for high aspect ratio wing couples optimization method | |
Wickramasinghe et al. | Designing airfoils using a reference point based evolutionary many-objective particle swarm optimization algorithm | |
Thirumalainambi et al. | Training data requirement for a neural network to predict aerodynamic coefficients | |
CN110516318B (en) | Airfoil design method based on radial basis function neural network proxy model | |
CN110110384A (en) | A kind of application method based on the non-dominant genetic algorithm of concentration strategy in the optimization of aircraft wing-type design | |
CN113777931B (en) | Icing wing type pneumatic model construction method, device, equipment and medium | |
CN112699622B (en) | Close formation pneumatic coupling effect modeling method for fixed-wing unmanned aerial vehicle | |
CN114722498A (en) | Wing section design method based on generation of countermeasure network | |
CN111598435A (en) | Quality trend prediction method based on adaptive feature selection and improved thought evolution algorithm | |
CN115438584A (en) | Wing profile aerodynamic force prediction method based on deep learning | |
Karali et al. | Design of a deep learning based nonlinear aerodynamic surrogate model for UAVs | |
CN113408044A (en) | Multi-rotor unmanned aerial vehicle blade optimization design method | |
Lane et al. | A surface parameterization method for airfoil optimization and high lift 2D geometries utilizing the CST methodology | |
CN109783858A (en) | A method of optimize for low Reynolds number airfoil | |
Johnson et al. | Using artificial neural networks and self-organizing maps for detection of airframe icing | |
CN108009320A (en) | A kind of multisystem association modeling method of hypersonic aircraft Control-oriented | |
CN113110558B (en) | Hybrid propulsion unmanned aerial vehicle demand power prediction method | |
CN115796021A (en) | Adaptive PID control parameter setting method for variant aircraft based on neural network | |
Khurana et al. | Development and validation of an efficient direct numerical optimisation approach for aerofoil shape design | |
CN111625901B (en) | Intelligent pressure coefficient curve generation method for wing profile | |
Ivanov et al. | Influence of selected turbulence model on the optimization of a class-shape transformation parameterized airfoil | |
Kody et al. | An efficient methodology for using a multi-objective evolutionary algorithm for winglet design | |
CN114676639A (en) | Aircraft aerodynamic shape optimization method, device and medium based on neural network | |
Safikhani et al. | Pareto optimization of two-element wing models with morphing flap using computational fluid dynamics, grouped method of data handling artificial neural networks and genetic algorithms | |
Fernando et al. | Bio-Inspired Aircraft Wing Modification Analysis in ANSYS Fluent |
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 |