CN114154250B - DCCGAN-based hypersonic aircraft flow thermosetting coupling physical field solving method - Google Patents
DCCGAN-based hypersonic aircraft flow thermosetting coupling physical field solving method Download PDFInfo
- Publication number
- CN114154250B CN114154250B CN202111564724.5A CN202111564724A CN114154250B CN 114154250 B CN114154250 B CN 114154250B CN 202111564724 A CN202111564724 A CN 202111564724A CN 114154250 B CN114154250 B CN 114154250B
- Authority
- CN
- China
- Prior art keywords
- data
- model
- output
- layer
- input
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 30
- 230000008878 coupling Effects 0.000 title claims abstract description 24
- 238000010168 coupling process Methods 0.000 title claims abstract description 24
- 238000005859 coupling reaction Methods 0.000 title claims abstract description 24
- 229920001187 thermosetting polymer Polymers 0.000 title claims abstract description 20
- 230000008569 process Effects 0.000 claims abstract description 10
- 230000006870 function Effects 0.000 claims description 64
- 230000004913 activation Effects 0.000 claims description 28
- 238000006073 displacement reaction Methods 0.000 claims description 25
- 238000012549 training Methods 0.000 claims description 24
- 238000012545 processing Methods 0.000 claims description 11
- 230000014509 gene expression Effects 0.000 claims description 6
- 238000005070 sampling Methods 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 5
- 238000013528 artificial neural network Methods 0.000 claims description 3
- 230000008094 contradictory effect Effects 0.000 abstract description 2
- 238000009826 distribution Methods 0.000 description 9
- 238000004088 simulation Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 4
- 239000012530 fluid Substances 0.000 description 3
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000008034 disappearance Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000000691 measurement method Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 238000009827 uniform distribution Methods 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000001808 coupling effect Effects 0.000 description 1
- 238000012407 engineering method Methods 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 238000003475 lamination Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000013515 script Methods 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 230000001502 supplementing effect Effects 0.000 description 1
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
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Geometry (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Computer Hardware Design (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Mathematical Analysis (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Evolutionary Biology (AREA)
- Fluid Mechanics (AREA)
- Algebra (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Automation & Control Theory (AREA)
- Aviation & Aerospace Engineering (AREA)
- Computational Mathematics (AREA)
- Medical Informatics (AREA)
- Aerodynamic Tests, Hydrodynamic Tests, Wind Tunnels, And Water Tanks (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
Aiming at the problem that solving precision and solving efficiency are contradictory in the hypersonic aircraft flow thermosetting coupling physical field solving process, the invention provides a hypersonic aircraft flow thermosetting coupling physical field solving method based on DCCGAN.
Description
Technical Field
The invention belongs to the field of computer-aided analysis and artificial intelligence, and particularly relates to a method for solving a hypersonic aircraft flow thermosetting coupling physical field based on DCCGAN, wherein DCCGAN (Deep Convolution Conditional GENERATIVE ADVERSARIAL Networks) refers to a deep convolution condition generation countermeasure network.
Background
Hypersonic aircrafts are complex in aerodynamic appearance, the flying speed exceeds 5Ma, the aerodynamic and thermodynamic environments are very severe, and the problems of aerodynamic, thermal and structural deformation multi-field coupling are very easy to occur. The flow heat set physical field coupling effect can have a great influence on aerodynamic characteristics and control efficiency of the aircraft, and is a serious difficulty which must be solved in hypersonic aircraft modeling simulation.
The primary problem faced by complex multi-physical field coupling solution is modeling of each sub-discipline, and the traditional engineering method such as a face element method, a modal analysis method and the like has the problem of insufficient solution precision. The adoption of a numerical calculation method based on computational fluid dynamics (Computational Fluid Dynamics, CFD) as a representative can effectively improve the solving precision and the reliability of the solving result. However, CFD has low calculation efficiency, and it takes several hours or even longer to complete calculation of an example, so that it is difficult to meet the requirements of fast simulation, and a new modeling simulation thought is required to match the task requirements of multi-field coupling fast simulation.
GAN is a model of generation of challenge learning proposed by Goodfellow in 2014, and consists of a generator model and a discriminant model. By fully learning the complex distribution of the original data samples, capturing the high-order correlation of the data samples, generating realistic data samples, the problem of insufficient data samples in the aerodynamic field can be well solved. The conditional generation type countermeasure network model (CGAN) can be regarded as an extension to GAN, and the content generated by the generator is controlled by giving the generator and the arbiter a condition information. The deep convolution condition generation countermeasure network (Deep Convolution Conditional GENERATIVE ADVERSARIAL Networks, abbreviated as DCCGAN) is to change the full connection layer between the generator and the arbiter in the CGAN model into a convolution layer, so that the full connection layer can be stably trained on a generation model with higher resolution and deeper network, and therefore, how to simulate and model the flow thermosetting coupling physical field of the hypersonic aircraft based on DCCGAN is a problem worthy of research.
Disclosure of Invention
Aiming at the problem that solving precision and solving efficiency are contradictory in the hypersonic aircraft flow thermosetting coupling physical field solving process, the invention provides a hypersonic aircraft flow thermosetting coupling physical field solving method based on DCCGAN.
The specific technical scheme of the invention is as follows:
a DCCGAN-based hypersonic aircraft flow thermosetting coupling physical field solving method comprises the following steps:
S1: and carrying out batch processing solution on the flow thermosetting coupling physical fields under different working conditions and different initial grids by utilizing CFD software, and deriving the deformation displacement data and the real distributed aerodynamic data obtained by calculation as an offline data set.
S2: and constructing DCCGAN0 models taking the working conditions as condition information, and generating distributed aerodynamic force data corresponding to the working conditions.
S3: the DCCGAN model is trained, and errors of the distributed aerodynamic data and the real distributed aerodynamic data generated by the generator model G0 and results output by the discriminator model D0 are used for guiding the network model to learn.
S4: and constructing DCCGAN models taking different initial grids as condition information, and generating deformation displacement data and distributed aerodynamic data corresponding to the different initial grids.
Preferably, the deformation displacement data is stored in a three-dimensional tensor of (N, 3) according to a node sequence number, and the distributed aerodynamic data is stored in a three-dimensional tensor of (N, 1) according to a node sequence number, wherein N 2 is larger than the grid node number.
Preferably, the working condition is subjected to One-hot coding (One-hot coding), then is subjected to dimension splicing with noise and is used as input of a generator model, and the output of the generator model G0 is a three-dimensional tensor of (N, N, 1), namely the generated distributed aerodynamic data; and performing dimension splicing on the distributed aerodynamic data generated by the real or generator model G0 and working conditions, and taking the distributed aerodynamic data and the working conditions as input of a discriminator model D0, wherein the discriminator model D0 outputs a true and false discrimination result.
Preferably, in the generator model G0, input data passes through a fully connected neural network and 5 deconvolution modules with output channel numbers of 256, 128, 64, 32 and 1 respectively, wherein BN layers are added after deconvolution layers of the first 4 deconvolution modules, a ReLU activation function layer is added after the BN layers, and a Tanh activation function layer is added after deconvolution layers of the 5 th deconvolution module;
wherein, the expression of ReLU and Tanh activation function is as follows:
Wherein ReLU (x) refers to the output value of the ReLU function, and Tanh (x) refers to the output value of the Tanh function; max (0, x) is the larger number of 0 and input x, e x is the exponential function operation of the input value, e -x is the exponential function operation of the input value after taking the negative value.
Preferably, the input data in the discriminator model D0 passes through five convolution modules with the output channel numbers of 64, 128, 256, 512 and 1 output layer convolution module with the output channel number of 1, wherein BN layers are added after the convolution layers of the 5 convolution modules, leakyRelu activation function layers are added after the BN layers, and Sigmoid activation function layers are added after the convolution layers of the output layer convolution modules;
The expressions of LeakyRelu and Sigmoid activation functions are as follows:
wherein, leakyRelu (x) refers to the output value of LeakyRelu functions, sigmoid (x) refers to the output value of Sigmoid functions; alpha refers to a leakage value, generally taking 0.01; max (alpha x, x) is the larger number in alpha x obtained by multiplying the input x and the input by the leakage value, and e -x is the exponential function operation performed after the input value takes the negative value.
Preferably, firstly, a loss function based on cross entropy of a network model is established, and in the training process, the generator model G0 hopes to make the discriminator D0 discriminate the generated distributed aerodynamic data as true, and strives to minimize the loss function; the discriminant model D0 is intended to promote the ability to discriminate that the generated distributed aerodynamic data is false, striving to maximize the loss function, and training is stopped when the generator model G0 and the discriminant model D0 both achieve the optimal solution.
Preferably, the deformation displacement data obtained in the step S1 and the distributed aerodynamic data generated in the step S2 are subjected to dimension splicing and serve as input of a generator model G1, and the output of the generator model G1 is a three-dimensional tensor of (N, 4), namely the generated deformation displacement and the distributed aerodynamic data; the data generated by the real or generator model G1 and the condition information are subjected to dimension splicing and serve as input of a discriminator model D1, and the discriminator model D1 outputs a true and false discrimination result; and the error of the data generated by the generator model G1 and the real data and the result output by the discriminant model D1 are used for guiding the network model to learn, so as to generate distributed aerodynamic data and deformation displacement data under different working conditions and different initial grids.
Preferably, the generator model G1 performs downsampling and upsampling on the input data sequentially, and during the downsampling, the input data passes through 3 convolution modules with output channel numbers of 64, 128 and 256 and 4 convolution modules with output channel numbers of 512, and a BN layer is added after the convolution layer of each convolution module, and a ReLU activation function layer is added after the BN layer; during up-sampling processing, the down-sampled data passes through deconvolution layers with the number of 3 output channels being 512 and the number of 4 output channels being 256, 128, 64 and 32 in sequence, and is output by an output convolution module with the number of 4 output channels, except the first deconvolution layer, a convolution module and a dimension splicing layer are sequentially added before each deconvolution layer, the convolution module consists of a convolution layer, a BN layer and a ReLU activation function layer, and the output convolution module consists of a convolution layer and a Tanh activation function layer.
Preferably, the structure of the discriminant model D1 is the same as the discriminant model D0.
Compared with the prior art, the invention has the following beneficial effects:
1. The invention provides a DCCGAN-based hypersonic aircraft flow thermosetting coupling physical field solving method, which utilizes a DCCGAN model to automatically learn data distribution and quickly generate a realistic sample by constructing a high-precision offline data set, solves the problem of contradiction between solving precision and solving efficiency, and provides effective technical support for hypersonic aircraft real-time simulation.
2. The invention designs a cascade DCCGAN model with two generators and two discriminants for high-dimensional nonlinear data, has higher precision compared with a numerical fitting model, and can realize the fitting of real data sample distribution in any input form. The distributed aerodynamic force sample and deformation displacement sample under different working conditions and different initial grids can be generated by controlling the condition information of the model, and the problem of insufficient data sample in the aerodynamic field can be solved.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the prior art, the drawings that are required in the description of the embodiments or the prior art will be briefly described, it will be apparent that the drawings in the description below are embodiments and details of the present invention, and other drawings may be obtained according to the drawings without inventive effort for a person skilled in the art.
FIG. 1 is an overall flow chart of the DCCGAN-based hypersonic aircraft flow thermoset coupling physical field solution method of the present invention;
FIG. 2 is a flow chart of a CFD solution in the solution method of the present invention;
FIG. 3 is a CFD computational domain grid of the present invention;
FIG. 4 is a flow thermoset coupling solution model of the present invention;
FIG. 5 is a residual convergence graph of the present invention;
FIG. 6 is a DCCGAN network block diagram of the present invention;
FIG. 7 is a DCCGAN model training flow chart of the present invention
FIG. 8 is a diagram of the network structure of DCCGAN of the present invention;
Detailed Description
In order that the above-recited objects, features and advantages of the present invention will be more clearly understood, a more particular description of the invention will be rendered by reference to the appended drawings and appended detailed description. It should be noted that, without conflict, the embodiments of the present invention and features in the embodiments may be combined with each other.
In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention, but the present invention may be practiced in other ways than those described herein, and therefore the scope of the present invention is not limited to the specific embodiments disclosed below.
In order to facilitate understanding of the above technical solutions of the present invention, the following detailed description of the above technical solutions of the present invention is provided by specific embodiments.
According to the method, a DCCGAN-based hypersonic aircraft flow thermosetting coupling physical field solving method is established, and the hypersonic aircraft flow thermosetting coupling physical field is subjected to multi-working-condition batch processing solving through CFD software, so that the problem of insufficient solving precision is solved. By constructing an offline data set, the distribution of deformation displacement and aerodynamic data is learned by utilizing DCCGAN models, a realistic sample is quickly generated by utilizing the trained generator model, and the problem of low solving efficiency is solved. The flow chart of the whole solving method is shown in fig. 1, and specifically comprises the following steps:
S1: and carrying out batch processing solution on the flow thermosetting coupling physical fields under different working conditions and different initial grids by utilizing CFD software, deriving calculated deformation displacement data and distributed aerodynamic data as an offline data set, and providing the offline data set and the offline data set to a DCCGAN network for training. Wherein aerodynamic data is stored in node number to a three-dimensional tensor of (128, 1) and deformation displacement data is stored in node number to a three-dimensional tensor of (128, 128,3). The flow of CFD solution is shown in fig. 2.
In some embodiments, step S1 specifically includes:
S1-1: importing the geometric model of the aircraft into ANSYS SCDM for river basin division, encrypting key parts of the aircraft, setting global parameters to generate polyhedral grids for CFD solving, and generating 200-ten-thousand-magnitude computational domain grids as shown in figure 3.
S1-2: the flow thermosetting coupling solving model is constructed by utilizing ANSYS Workbench software, as shown in fig. 4, the temperature distribution and aerodynamic force on the coupling surface are solved by utilizing a Fluent solver, and the stress and deformation displacement distribution on the structural grid is calculated by utilizing a Static Structural solver.
S1-3: the solution parameters and conditions are set in Fluent: the solver is set as a density solver; solving an opening energy equation in the model and selecting STANDARD K-e turbulence model; the model material part sets the fluid as ideal gas; the boundary condition part sets the outermost boundary as a pressure far field, and the aircraft surface as a solid wall boundary and a thermal boundary; setting temperature, pressure, incoming flow Mach number and airflow velocity components in all directions according to working condition parameters; a monitor is set, given a solution termination condition.
S1-4: initializing the solution. And solving convergence and then deriving aerodynamic force and deformation displacement data on the grid nodes of the surface of the aircraft. The residual convergence curve obtained by iterative solution is shown in fig. 5.
S1-5: the derived data is stored in a three-dimensional tensor as a set of solutions by node number. The aircraft has 16128 nodes, aerodynamic data can be stored into three-dimensional tensors with the size of (128, 1), and redundant positions are subjected to 0 supplementing processing; the deformation displacement data are stored in the three-dimensional tensors of (128, 128,3) according to the node serial numbers in the same way, wherein the channels 1, 2 and 3 sequentially store the data of X, Y, Z axes in the deformation displacement.
S1-6: according to the flight state of the hypersonic aircraft, 3 variables are selected as the multiple operating mode parameters, as shown in table 1. The aerodynamic force and deformation displacement data under different working conditions are automatically solved in batches by writing Journal scripts, and therefore multi-working-condition offline solution set construction is completed.
TABLE 1 Multi-operating parameter Table
S2: and constructing DCCGAN models taking the working conditions as condition information to generate distributed aerodynamic force samples corresponding to the working conditions. And performing dimension splicing on the working conditions by adopting One-hot coding and noise as input of a generator, wherein the output of the generator is a three-dimensional tensor of (128, 1) and represents a generated aerodynamic sample. And carrying out dimension splicing on the real or generated aerodynamic force data and the working condition to serve as input of a discriminator, and outputting a true and false discriminating result by the discriminator. A specific structure diagram of the generator model G0 and the discrimination model D0 is shown in fig. 6.
In some embodiments, step S2 specifically includes:
S2-1: and adopting single-heat coding for working conditions. Mach numbers are 5 feature values in total, and [6.6,6.8,7.0,7.2,7.4] can be expressed as [10000,01000,00100,00010,00001] by using one-time thermal coding, and similarly, the attack angle can be expressed as [10000,01000,00100,00010,00001] and the sideslip angle as [100,010,001]. The condition code can be represented by a 13-dimensional column vector.
S2-2: and performing dimension splicing (Concat) on the working condition codes and the noise to serve as input of a generator G0, and converting 113-dimensional input data into a three-dimensional tensor of (4, 512) by utilizing a fully-connected neural network (Dense) to serve as input of a next deconvolution layer.
S2-3: : this 3-dimensional tensor is converted to a three-dimensional tensor of (128, 1) by a deconvolution module of 5 output channels 256, 128, 64, 32 and 1, respectively, representing the generated aerodynamic sample. The deconvolution layer in the convolution module has a size 4*4, a step size of 2, and a fill of 1.
S2-4: BN (Batch Normalization) layers were added after the first 4 deconvolution layers. The BN layer is used for calling a batch normalization regularization method, and the specific processing mode is that after the convolution layer, the calculated average value is subtracted from batch data (batch size) channel by channel and divided by standard deviation, so that the value range of input data can be unified into intervals [ -1,1 ]. The BN layer can solve the problem of difficult learning of the network model, and is beneficial to gradient update of back propagation and quickens the speed of network convergence.
S2-5: adding a ReLU activation function layer behind the first 4 BN layers to improve the learning speed of the model; and adding a Tanh activation function layer after the 5 th deconvolution layer, so as to reduce the problem of gradient disappearance. The expressions for Relu and the Tanh function are shown as follows:
S2-6: and splicing the generated aerodynamic force sample or the real aerodynamic force sample with the working condition codes after dimension conversion, inputting the obtained aerodynamic force sample or the real aerodynamic force sample into convolution modules with five layers of output channels of 64, 128, 256, 512 and 512 respectively, and outputting a probability value indicating the judgment result of the discriminator by using the output layer convolution module with the output channel number of 1. The first 5 convolutional layers in the convolutional module are 4*4 in size, 2 in step size, 1 in padding, and the last convolutional layer is 1*1 in size, 1 in step size, 0 in padding.
S2-7: BN layers and LeakyRelu activation function layers are added after the first 5 convolution layers, and Sigmoid activation function layers are added after the last 1 convolution layers. LeakyRelu the activation function is a modified version of the ReLU activation function, which performs better in the arbiter. The Sigmoid activation function is a two-classification problem to deal with whether the input sample is true or false. The expressions for LeakyRelu and Sigmoid functions are shown as follows:
S3: and training the model, and guiding the network model to learn by utilizing the generated errors of aerodynamic data and the true values and the results output by the discriminator. The model training flow is shown in fig. 7. The whole training is a process of game with the generator G0 and the arbiter D0, and the generator G0 hopes to make the arbiter D0 identify the generated sample as true, so as to try to minimize the loss function; the arbiter D0 wants to improve the ability to discriminate that the generated sample is false, so strives to maximize the loss function. The training process stops when the generator G0 and the arbiter D0 both obtain the optimal solution, theoretically reaching the nash balance.
In some embodiments, step S3 specifically includes:
s3-1: designing a cross entropy-based contrast loss function of a network, as shown in the following formula:
Wherein G represents the generator, D represents the discriminator is the reactive loss function that generates the reactive network, and L CGAN (D, G) is the reactive loss function that conditionally generates the reactive network. Representing fixed arbiter parameters, the generator minimizing a loss function; The generator parameters are fixed and the arbiter maximizes the loss function. x is a true data sample vector following the target distribution P o, y is condition information, i.e., the condition code, and z is a noise sample vector following the uniform distribution P z. D (x|y) represents the decision of the arbiter when the condition code is input to the arbiter together with the true data sample vector, 1 represents true, and 0 represents false. G (z|y) represents the sample vector generated by the generator when the condition code is input to the generator along with the noise sample vector. D (G (z|y)) represents the judgment of the discriminator when the sample vector generated by the generator is input to the discriminator. E represents the mathematical expectation of the real data sample vector x and the noise sample vector z, and log represents a logarithmic function.
The first item represents a loss value when the real data is discriminated as the specified class y under the given condition information y; the second term represents the loss value calculated by the sample data generated by the generation model after the condition information y is combined. Updating the parameter information of the judging module by the judging device through maximizing the loss function, and accurately identifying real data and generating data; the generator minimizes the loss function so that the data generated by the generator can be judged as the y category by the discriminator, thereby achieving the effect of confusion of the discriminator.
The loss constraint target of the countermeasure network is to make the reconstructed data sample approach to the original real data to the maximum extent, only the countermeasure loss can not obtain the wanted effect, the traditional loss constraint is needed to be added, and an L1 distance measurement method is selected, wherein the L1 distance measurement method is shown in the following formula:
Wherein L L1 (G) is the L1 penalty between the reconstructed data and the real data, Meaning that x meets the expectations of the true value distribution P o and z meets the uniform distribution P z. G (z|y) represents the sample vector generated by the generator when the condition code is input to the generator along with the noise sample vector. II 1 denotes 1 norm
To balance CGAN loss terms and L1 loss terms, a super parameter λ was added: the DCCGAN final loss function G * used in the algorithm is therefore shown as follows:
Where L CGAN (D, G) is the reactive loss function of the conditional generation reactive network, L L1 (G) is the L1 loss between reconstructed data and real data, and λ is the hyper-parameter. Representing the parameters of a fixed generator, maximizing the counterloss function by the arbiter, fixing the parameters of the arbiter, and taking the values of the parameters of the generator G and the arbiter D when the generator minimizes the counterloss function.
S3-2: and initializing weight parameters of the network. The weight initialization can enable the network model to have a better initial position, better and faster convergence when the global optimal solution is sought. The weight of the convolution layer is initialized by adopting random normal distribution with the mean value of 0 and the variance of 0.02.
S3-3: and randomly extracting a batch of real data from the multi-condition offline solution set, training the discriminator and updating the parameters of the discriminator. The first training sends the real data to the discriminator, and the real data is used as parameter update of the discriminator, so that the capability of the discriminator for judging the real data is improved. The second training uses the generator to generate false data with the same scale, and the parameters of the discriminator are updated, so that the capacity of the discriminator for judging the false data is improved. The method adopted for parameter updating is an Adam method modified from the gradient descent method.
S3-4: training the generator and updating its parameters. In training the generator, it is necessary to keep the parameters of the arbiter constant, i.e. to set the arbiter in a non-trainable mode. After the training of the wheel set generator is completed, the discriminator is set to be in a trainable mode, and the next iteration is waited for. The Adam method is also adopted for generator parameter updating.
S3-5: and training is performed in a reciprocating manner in a training batch, so that the alternating training and enhancement of the CGAN network discriminator and the generator are realized, and the training batch is finally close to an ideal Nash equilibrium state.
In some embodiments, step S4 specifically includes:
S4: and constructing DCCGAN models taking different initial grids as conditional information, and generating distributed aerodynamic samples corresponding to the different initial grids. And performing dimension stitching on the deformation displacement data and aerodynamic force sample data generated in the step S2 model to serve as input of a generator, wherein the output of the generator is a three-dimensional tensor of (128, 128,4) and represents the generated deformation displacement and aerodynamic force sample. And performing dimension splicing on the real or generated data and the condition information to serve as input of a discriminator, and outputting a true and false discriminating result by the discriminator. A specific structure diagram of the generator model G1 and the discrimination model D1 is shown in fig. 8. Training the model by using the same training method as in the step S3 to generate distributed aerodynamic force samples and deformation displacement samples under different working conditions and different initial grids.
S4-1: and generating aerodynamic force sample data by using the trained generator G0, and performing dimension splicing on the aerodynamic force sample data and initial grid node coordinate data subjected to stream thermosetting coupling simulation to obtain a three-dimensional tensor (128, 128,4) serving as input of the generator G1.
S4-2: and carrying out downsampling processing on the input data. The input data will pass through 3 convolution modules with output channel numbers of 64, 128, 256 and 4 convolution modules with output channel number of 512, respectively. In the downsampling process, the length and width of the input vector are reduced by half each time the input vector passes through a convolution module, and the size of 128 x 128 from the initial size becomes 1*1 at the end of the downsampling process. The convolution module consists of a convolution layer with the size 4*4, the step size 2 and the padding 1, a BN layer and a ReLU activation function layer.
S4-3: and carrying out up-sampling processing on the down-sampled data, and adjusting the size of the output sample data. In the up-sampling process, the data is subjected to deconvolution with the sizes of 256, 128, 64 and 32 of 4*4, the step length of 2 and the filling of 1 in sequence through the number of 3 layers of output channels of 512 and the number of 4 layers of output channels, the length and the width of the data are doubled through each deconvolution lamination block, and the size of 128 x 128 is gradually adjusted from the size 1*1 to the size required by outputting sample data. In the processing process, the output results of the corresponding layers in the upper and lower sampling are subjected to channel dimension superposition by using the layer jump connection (Concat), so that the integrity and feature richness of the output information are ensured. Then input into the convolution module to carry out dimension reduction operation and control the dimension of the channel. The convolution module consists of a convolution layer with the size 3*3, the step length of 1 and the filling of 1, a BN layer and a ReLU activation function layer. The last layer of up-sampling is composed of a convolution layer with the channel number of 4 and then a Tanh activation function layer is added, so that the problem of gradient disappearance is solved, and the size of output data is controlled.
S4-4: the generated aerodynamic force sample or the real aerodynamic force sample is subjected to dimension splicing with the condition information and then is input into a discriminator D1, and the discriminator discriminates whether the input sample is the generated sample or the real value. Wherein the arbiter D1 is identical in structure to D0.
S4-5: training the model according to the training methods in the steps S3-1 to S3-5, and guiding the network model to learn by utilizing the errors of the generated data and the real data and the output result of the discriminator. Therefore, distributed aerodynamic force samples and deformation displacement samples under different working conditions and different initial grids can be generated.
The above-described embodiments are merely preferred embodiments of the present invention, and are not intended to limit the present invention in any way. Any person skilled in the art, using the disclosure above, may make many more possible variations and modifications of the technical solution of the present invention, or make many more modifications of the equivalent embodiments of the present invention without departing from the scope of the technical solution of the present invention. Therefore, the equivalent changes according to the inventive concept should be covered in the protection scope of the present invention without departing from the technical scheme of the present invention.
Claims (1)
1. A DCCGAN-based hypersonic aircraft flow thermosetting coupling physical field solving method comprises the following steps:
S1: carrying out batch processing solution on the flow thermosetting coupling physical fields under different working conditions and different initial grids by utilizing CFD software, and deriving deformation displacement data and real distributed aerodynamic data obtained by calculation to serve as an offline data set;
S2: constructing DCCGAN0 models taking working conditions as condition information, and generating distributed aerodynamic force data corresponding to the working conditions;
s3: training DCCGAN a model, and guiding the network model to learn by utilizing errors of the distributed aerodynamic data and the real distributed aerodynamic data generated by the generator model G0 and the output result of the discriminator model D0;
S4: constructing DCCGAN models taking different initial grids as condition information, and generating deformation displacement data and distributed aerodynamic data corresponding to the different initial grids;
the deformation displacement data are stored into a three-dimensional tensor of (N, N, 3) according to node serial numbers, and the distributed aerodynamic data are stored into a three-dimensional tensor of (N, N, 1) according to node serial numbers, wherein N 2 is larger than the number of grid nodes;
the step S2 includes: performing dimension splicing on the working conditions and noise after single thermal coding, and taking the working conditions and the noise as input of a generator model, wherein the output of the generator model G0 is the three-dimensional tensor of (N, N, 1), namely the generated distributed aerodynamic data; carrying out dimension splicing on the distributed aerodynamic data generated by the real or generator model G0 and working conditions, and taking the data as the input of a discriminator model D0, wherein the discriminator model D0 outputs a true and false discrimination result;
In the generator model G0, input data passes through a fully-connected neural network and 5 deconvolution modules with the output channel numbers of 256, 128, 64, 32 and 1 respectively, wherein BN layers are added after deconvolution layers of the first 4 deconvolution modules, a ReLU activation function layer is added after the BN layers, and a Tanh activation function layer is added after deconvolution layers of the 5 th deconvolution module;
wherein, the expression of ReLU and Tanh activation function is as follows:
Wherein ReLU (x) refers to the output value of the ReLU function, and Tanh (x) refers to the output value of the Tanh function; max (0, x) is the larger number in 0 and input x, e x is the input value to perform exponential function operation, e -x is the input value to perform exponential function operation after taking the negative value;
the input data in the discriminator model D0 passes through five convolution modules with the number of output channels of 64, 128, 256, 512 and 1 output layer convolution module with the number of output channels of 1, wherein BN layers are added after convolution layers of 5 convolution modules, leakyRelu activation function layers are added after the BN layers, and Sigmoid activation function layers are added after convolution layers of the output layer convolution modules;
The expressions of LeakyRelu and Sigmoid activation functions are as follows:
Wherein, leakyRelu (x) refers to the output value of LeakyRelu functions, sigmoid (x) refers to the output value of Sigmoid functions; alpha refers to a leakage value, and is 0.01; max (ax, x) is that a larger number is obtained in ax after the input x and the input are multiplied by the leakage value, e -x is that the input value takes a negative value and then exponential function operation is carried out;
the step S3 includes:
Firstly, establishing a loss function of a network model based on cross entropy, wherein the generator model G0 hopes to make a discriminator D0 discriminate generated distributed aerodynamic data as true in the training process, and strives to minimize the loss function; the discriminant model D0 hopes to promote and distinguish the ability that the distributed aerodynamic data generated is false, strives to maximize the loss function, train and stop when generator model G0 and discriminant model D0 both obtain the optimal solution;
The step S4 includes:
Performing dimension splicing on the deformation displacement data obtained in the step S1 and the distributed aerodynamic data generated in the step S2, and taking the dimension spliced data as input of a generator model G1, wherein the output of the generator model G1 is a three-dimensional tensor of (N, N, 4), namely the generated deformation displacement and the distributed aerodynamic data; the data generated by the real or generator model G1 and the condition information are subjected to dimension splicing and serve as input of a discriminator model D1, and the discriminator model D1 outputs a true and false discrimination result; the error of the data generated by the generator model G1 and the real data and the result output by the discriminant model D1 are used for guiding the network model to learn, and distributed aerodynamic data and deformation displacement data under different working conditions and different initial grids are generated;
the generator model G1 sequentially performs downsampling and upsampling on input data, and during downsampling, the input data passes through 3 convolution modules with output channel numbers of 64, 128 and 256 and 4 convolution modules with output channel numbers of 512 respectively, a BN layer is added after the convolution layer of each convolution module, and a ReLU activation function layer is added after the BN layer; during up-sampling processing, the down-sampled data passes through deconvolution layers with the number of 3 output channels of 512 and the number of 4 output channels of 256, 128, 64 and 32 in sequence, and is output by an output convolution module with the number of 4 output channels, wherein a convolution module and a dimension splicing layer are sequentially added before each deconvolution layer except the first deconvolution layer, the convolution module consists of a convolution layer, a BN layer and a ReLU activation function layer, and the output convolution module consists of a convolution layer and a Tanh activation function layer;
The structure of the discriminator model D1 is the same as that of the discriminator model D0.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111564724.5A CN114154250B (en) | 2021-12-20 | 2021-12-20 | DCCGAN-based hypersonic aircraft flow thermosetting coupling physical field solving method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111564724.5A CN114154250B (en) | 2021-12-20 | 2021-12-20 | DCCGAN-based hypersonic aircraft flow thermosetting coupling physical field solving method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114154250A CN114154250A (en) | 2022-03-08 |
| CN114154250B true CN114154250B (en) | 2024-06-14 |
Family
ID=80451763
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111564724.5A Active CN114154250B (en) | 2021-12-20 | 2021-12-20 | DCCGAN-based hypersonic aircraft flow thermosetting coupling physical field solving method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114154250B (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109934924A (en) * | 2019-02-21 | 2019-06-25 | 北京航空航天大学 | A kind of hypersonic aircraft infrared signature Fast rendering method |
| CN111814246A (en) * | 2020-06-11 | 2020-10-23 | 中国空气动力研究与发展中心计算空气动力研究所 | Wing section inverse design method based on generation countermeasure network |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111401138B (en) * | 2020-02-24 | 2023-11-07 | 上海理工大学 | Countermeasure optimization method for generating countermeasure neural network training process |
| CN113534837B (en) * | 2021-07-06 | 2024-03-22 | 深圳市科卫泰实业发展有限公司 | Unmanned aerial vehicle flight state fusion control system and control method based on generation countermeasure network |
| CN113640203B (en) * | 2021-09-17 | 2023-11-24 | 中北大学 | Multi-parameter complex extreme environment simulation device |
-
2021
- 2021-12-20 CN CN202111564724.5A patent/CN114154250B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109934924A (en) * | 2019-02-21 | 2019-06-25 | 北京航空航天大学 | A kind of hypersonic aircraft infrared signature Fast rendering method |
| CN111814246A (en) * | 2020-06-11 | 2020-10-23 | 中国空气动力研究与发展中心计算空气动力研究所 | Wing section inverse design method based on generation countermeasure network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114154250A (en) | 2022-03-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Xiao et al. | Flows over periodic hills of parameterized geometries: A dataset for data-driven turbulence modeling from direct simulations | |
| Lu et al. | DeepXDE: A deep learning library for solving differential equations | |
| Zhou et al. | Combining global and local surrogate models to accelerate evolutionary optimization | |
| Giannakoglou et al. | Aerodynamic shape design using evolutionary algorithms and new gradient-assisted metamodels | |
| An et al. | A deep learning framework for hydrogen-fueled turbulent combustion simulation | |
| Sun et al. | High Reynolds number airfoil turbulence modeling method based on machine learning technique | |
| Zafar et al. | Convolutional neural network for transition modeling based on linear stability theory | |
| Liu et al. | GAN-MDF: An enabling method for multifidelity data fusion | |
| Illarramendi et al. | Performance and accuracy assessments of an incompressible fluid solver coupled with a deep convolutional neural network | |
| Ma et al. | A combined data-driven and physics-driven method for steady heat conduction prediction using deep convolutional neural networks | |
| White et al. | Neural networks predict fluid dynamics solutions from tiny datasets | |
| Hacioglu | Fast evolutionary algorithm for airfoil design via neural network | |
| Pierret et al. | Designing turbomachinery blades with the function approximation concept and the navier-stokes equations | |
| CN117854643B (en) | MEMS membrane simulation method and system based on graphic neural network | |
| Nauleau et al. | Topological Analysis of Ensembles of Hydrodynamic Turbulent Flows An Experimental Study | |
| Iserte et al. | Accelerating urban scale simulations leveraging local spatial 3D structure | |
| Gautier et al. | Multi-objective optimization algorithm assisted by metamodels with applications in aerodynamics problems | |
| Zhang et al. | Inverse method of centrifugal pump blade based on Gaussian process regression | |
| CN114154250B (en) | DCCGAN-based hypersonic aircraft flow thermosetting coupling physical field solving method | |
| Arabnia | Aerodynamic shape optimization of axial turbines in three dimensional flow | |
| Pu et al. | Solving Schrodinger equations using a physically constrained neural network | |
| Hiwa et al. | Design Mode Analysis of Pareto Solution Set for Decision‐Making Support | |
| Rai | Three-dimensional aerodynamic design using artificial neural networks | |
| CN114491790A (en) | MAML-based pneumatic modeling method and system | |
| Wauters et al. | SAMURAI: A new asynchronous bayesian optimization technique for optimization-under-uncertainty |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |