CN113836634B - Deep neural network modeling method for large-difference pneumatic data - Google Patents

Deep neural network modeling method for large-difference pneumatic data Download PDF

Info

Publication number
CN113836634B
CN113836634B CN202110972718.7A CN202110972718A CN113836634B CN 113836634 B CN113836634 B CN 113836634B CN 202110972718 A CN202110972718 A CN 202110972718A CN 113836634 B CN113836634 B CN 113836634B
Authority
CN
China
Prior art keywords
aerodynamic
model
pneumatic
flight state
fcn
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
Application number
CN202110972718.7A
Other languages
Chinese (zh)
Other versions
CN113836634A (en
Inventor
向渝
王文正
胡力卫
汪文勇
张骏
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Electronic Science and Technology of China
Original Assignee
University of Electronic Science and Technology of China
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by University of Electronic Science and Technology of China filed Critical University of Electronic Science and Technology of China
Priority to CN202110972718.7A priority Critical patent/CN113836634B/en
Publication of CN113836634A publication Critical patent/CN113836634A/en
Application granted granted Critical
Publication of CN113836634B publication Critical patent/CN113836634B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • 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 discloses a deep neural network modeling method of large-difference pneumatic data, which belongs to the technical field of aircraft system modeling and is characterized by comprising the following steps: a. calculating a pneumatic data set; b. preprocessing a pneumatic data set; c. c, constructing a model, namely determining the dimension of input data and the dimension of output data and data quantity information through the pneumatic data set calculated in the step a, determining the number of nodes of an input layer and an output layer in FCN_1 and CNN, preliminarily constructing a deep neural network model, and determining the number of network layers and the number of nodes of each layer in FCN_2 according to the scale of the pneumatic data set; d. forward propagation; e. counter-propagating; f. model verification and optimization. The invention can meet the modeling requirement when the aerodynamic shape and the flight state are changed simultaneously, can give out the influence degree of the flight state change and the aerodynamic shape change on the aerodynamic characteristics of the aircraft, and can perform good optimization on the aerodynamic characteristics of the aircraft.

Description

Deep neural network modeling method for large-difference pneumatic data
Technical Field
The invention relates to the technical field of aircraft system modeling, in particular to a deep neural network modeling method of large-difference pneumatic data.
Background
The flight state parameters and the aerodynamic layout appearance parameters are two different types of parameters with larger difference, and the physical meaning, the value range and the association degree of aerodynamic characteristic change of the two types of parameters are different. Because of the large difference between the two different types of parameters, the aerodynamic characteristics of the aircraft are different along with the change of the flight state and the sensitivity of the change of the key layout characteristic parameters, so that the prediction results of the model are affected by different orders of magnitude. Neglecting the variability and relevance of these two types of parameters in the modeling process tends to lead to difficulty in dividing the variable range, and unreasonable parameter space can cause loss of modeling accuracy. Thus, the impact of the variability and relevance of the two on modeling accuracy requires further discussion and study to be developed for different classes of model input parameters.
How to correlate two types of parameters at the same time, and constructing a pneumatic agent model is always a key problem in the field of pneumatic modeling.
The Chinese patent document with publication number of CN 111597698A and publication date of 2020, 08 and 28 discloses a method for realizing pneumatic optimization design based on a deep learning multi-precision optimization algorithm, which is characterized by comprising the following steps:
step S1: generating sample points from the wing design space by Latin Hypercube Sampling (LHS) and carrying out normalization processing, wherein a high-precision training data set Xhi and a low-precision training data set Xlo are respectively formed after normalization; the method comprises the steps that a wing design space comprises geometrical parameters of a wing section shape and geometrical parameters of a wing plane shape, boundary values of all parameters are standard values + -disturbance values of a wing standard model to be optimized, which corresponds to the parameters, the disturbance values are not more than 10 percent of the standard values, and each generated sample point comprises the geometrical parameters of the wing section shape and the geometrical parameters of the wing plane shape and jointly represents the geometrical appearance of an aircraft wing;
step S2, aerodynamic coefficients of the high-precision data set and the low-precision data set are calculated in parallel, wherein the aerodynamic coefficients respectively represent one or more aerodynamic coefficients of the high-precision data set and the low-precision data set which are required to be optimized;
step S3, an initial multi-precision deep neural network proxy model is built and trained, the multi-precision deep neural network proxy model is composed of three fully connected neural networks, wherein a first neural network NNL (x, theta) is used for fitting low-precision data, the input is a low-precision data set Xlo, the output is ylo, a second neural network and a third neural network are respectively used for fitting linear Fl and nonlinear FNl relations between low-precision aerodynamic coefficients and high-precision aerodynamic coefficients, the output ylo of the first neural network NNL (x, theta) is used as a first input, a high-precision data set Xhi is used as a second input, and finally, the high-precision aerodynamic coefficients are obtained through alpha fitting, and the multi-precision deep neural network proxy model is represented by the following formula:
yhi＝αFl(x,ylo)+(1-α)Fnl(x,ylo),α∈[0,1]
wherein, alpha is the super parameter of the model, x is the input variable of the model, θ, βi, i=1, 2 are the super parameters of three neural networks respectively;
step S4, global optimization is carried out on a current multi-precision deep neural network proxy model through a Particle Swarm Optimization (PSO), and an optimal solution Xop of the current proxy model is searched, wherein an input variable x to be predicted during optimization is simultaneously input into three trained neural networks;
step S5, adding Xop as an updated sample point of the high-precision data set into the high-precision data set;
step S6, obtaining low-precision update sample points Xlo and update through a low-precision point adding criterion, and adding the updated low-precision sample points into a low-precision data set;
step S7, calculating aerodynamic coefficients of high-precision updated sample points Xop and low-precision updated sample points Xlo and update in parallel;
s8, retraining a multi-precision deep neural network proxy model by using the updated high-low precision data set;
step S9, performing global optimization on the new multi-precision deep neural network by using a particle swarm optimization algorithm, and searching an optimal solution Xop of the current agent model;
step S10, taking the optimal solution in the high-precision data set as the optimal solution of the iteration; checking whether the optimal solution of the iteration meets the convergence criterion of the algorithm, if so, outputting the optimal solution as an optimization result, and terminating the iteration; if not, returning to the step S5 to continue iterative optimization.
According to the method for realizing pneumatic optimization design by the multi-precision optimization algorithm based on deep learning disclosed by the patent document, for a low-precision data set, the distance between sample points is measured by solving the Euclidean distance between the sample points, and the position of the lacking sample points is updated, so that the sample points which are uniformly distributed in the whole design domain are generated. However, the modeling requirements when the aerodynamic profile and the flight state are simultaneously changed cannot be met, so that the degree of influence of the flight state change and the aerodynamic profile change on the aerodynamic characteristics of the aircraft cannot be given, and the aerodynamic characteristics of the aircraft cannot be well optimized.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention provides a deep neural network modeling method of large-difference pneumatic data, which can meet modeling requirements when the pneumatic appearance and the flight state are changed simultaneously, can give out the influence degree of the flight state change and the pneumatic appearance change on the pneumatic characteristics of the aircraft, and can perform good optimization on the pneumatic characteristics of the aircraft.
The invention is realized by the following technical scheme:
the deep neural network modeling method of the large-difference pneumatic data is characterized by comprising the following steps of:
a. calculating a pneumatic data set, wherein the pneumatic data set comprises flight state parameters, pneumatic layout appearance parameters and aerodynamic coefficient true values, the pneumatic layout appearance parameters and the flight state parameters are used as input data, and the aerodynamic coefficient true values are used as output data;
b. the method comprises the steps of preprocessing a pneumatic data set, filtering and screening abnormal values and missing values existing in the pneumatic data set, normalizing all data in the pneumatic data set, and dividing a training set, a verification set and a test set according to a preset proportion;
c. c, constructing a model, namely determining the dimension of input data and the dimension of output data and data quantity information through the pneumatic data set calculated in the step a, determining the number of nodes of an input layer and an output layer in FCN_1 and CNN, preliminarily constructing a deep neural network model, and determining the number of network layers and the number of nodes of each layer in FCN_2 according to the scale of the pneumatic data set;
d. forward propagation, inputting aerodynamic layout appearance parameters in the same aerodynamic data in a training set into CNN, and outputting as vector f 1 Inputting the corresponding flight state parameters into FCN_1, and outputting the flight state parameters as a vector f 2 And then inputting the aerodynamic layout appearance parameter and the flight state parameter into the FCN_2 together, outputting the aerodynamic layout appearance parameter and the flight state parameter as a vector c, and outputting the aerodynamic layout appearance parameter and the flight state parameter as a neural network:
Figure BDA0003226484300000031
wherein y is z Is the output value of the model when the z-th sample is taken as input; q is the dimension of the aerodynamic coefficient true value, and is equal to the output dimension of CNN and FCN_1; f (f) 1i,z Is the ith output component when the z-th input sample is input to the CNN; c i,z Weights for the ith output component when the z-th input sample is input to the CNN; f (f) 2j,z Is the jth output component when the jth input sample is input to fcn_1; c q+j,z A weight for a jth output component when a zth input sample is input to fcn_1;
e. back propagation, the loss function of the model is as follows:
Figure BDA0003226484300000032
wherein L is MSE A loss function for the model; n is the number of samples during model training; y is Z Is the predicted aerodynamic coefficient for the z-th sample;
Figure BDA0003226484300000033
for the z-th pneumatic sampleA true aerodynamic coefficient in the pneumatic dataset;
f. model verification and optimization, the number of layers and node numbers of FCN_1, FCN_2 and CNN networks are continuously adjusted, and the model is verified and optimized through the constructed verification set and test set.
In the step a, the calculation of the aerodynamic data set refers to the calculation of aerodynamic coefficient true values of the aircraft through CFD software according to two-dimensional coordinates and flight state parameters of the aircraft wing shape.
In the step a, the coordinate information of the aerodynamic layout appearance parameters is obtained through profile software according to the wing shape of the real aircraft, and the aerodynamic coefficient true value is calculated through CFD software.
In the step a, flight state parameters comprise Mach numbers, attack angles and Reynolds numbers, and aerodynamic coefficient true values comprise lift coefficients and drag coefficients.
In the step b, the training set, the verification set and the test set are divided according to a preset proportion, namely, the training set, the verification set and the test set are divided according to a proportion of 8:1:1.
In the step d, the dimension of the vector c is 2q, the first q values express the weight of the aerodynamic force coefficient of the aerodynamic layout appearance parameter in the current flight state, the last q values express the weight of the aerodynamic force coefficient of the specific aerodynamic appearance in the current flight state, and the difference and the similarity between the first q values and the last q values of the vector c express the difference and the relevance between the aerodynamic layout appearance parameter and the flight state parameter.
In the step e, the back propagation specifically means by solving L MSE Regarding the gradient of the model parameter theta, updating the model parameter theta according to the gradient and the learning rate eta;
Figure BDA0003226484300000041
wherein θ is a model parameter, η is a learning rate, L MSE Is a loss function of the model.
In the step f, the verification and optimization of the model through the constructed verification set and the test set specifically means that K-fold cross verification is adopted, the model is verified on the constructed verification set, the structure of the model is adjusted and optimized according to the under-fitting and over-fitting phenomena in the verification process, and finally, the optimized model is evaluated on the test set.
The K-fold cross validation specifically refers to that an initial sample is divided into K sub-samples, one single sub-sample is reserved as data of a validation model, the other K-1 samples are used for training, the cross validation is repeated for K times, each sub-sample is validated once, and the results of the K times are averaged, so that a single estimation is finally obtained.
The CNN is used for processing the aerodynamic layout appearance parameters, FCN_1 is used for processing the flight state parameters and FCN_2 is used for learning the differences and the correlations of the flight state parameters and the aerodynamic layout appearance parameters in predicting aerodynamic coefficients.
The fcn_1 refers to a first fully connected neural network.
The fcn_2 refers to a second fully connected neural network.
The CNN refers to a convolutional neural network.
CFD software according to the present invention refers to computational fluid dynamics software.
The profile software refers to aircraft wing profile design software.
The beneficial effects of the invention are mainly shown in the following aspects:
1. the method comprises the steps of a, calculating a pneumatic data set, wherein the pneumatic data set comprises flight state parameters, pneumatic layout appearance parameters and pneumatic power coefficient true values, taking the pneumatic layout appearance parameters and the flight state parameters as input data, and taking the pneumatic power coefficient true values as output data; b. the method comprises the steps of preprocessing a pneumatic data set, filtering and screening abnormal values and missing values existing in the pneumatic data set, normalizing all data in the pneumatic data set, and dividing a training set, a verification set and a test set according to a preset proportion; c. model construction, namely determining the dimension of input data and the dimension of output data through the pneumatic data set calculated in the step a, determining the node number of an input layer and an output layer in FCN_1 and CNN through data quantity information, and preliminarily constructing a deep nerveThe network model is used for determining the number of network layers and the number of nodes of each layer in the FCN_2 according to the scale of the pneumatic data set; d. forward propagation, inputting aerodynamic layout appearance parameters in the same aerodynamic data in a training set into CNN, and outputting as vector f 1 Inputting the corresponding flight state parameters into FCN_1, and outputting the flight state parameters as a vector f 2 And then inputting the aerodynamic layout appearance parameter and the flight state parameter into the FCN_2 together, outputting the aerodynamic layout appearance parameter and the flight state parameter as a vector c, and outputting the aerodynamic layout appearance parameter and the flight state parameter as a neural network: e. counter-propagating; f. the number of layers and node numbers of the FCN_1, FCN_2 and CNN networks are continuously adjusted through model verification and optimization, and the model verification and optimization are performed through a constructed verification set and a test set.
2. Aiming at the problem that the pneumatic layout appearance parameters and the flight state parameters cannot be modeled jointly in the pneumatic modeling field, the invention provides a deep neural network modeling method for modeling the flight state parameters and the pneumatic layout appearance parameters jointly, which combines two deep neural networks of FCNs and CNNs, can respectively process the flight state parameters and the pneumatic layout appearance parameters at the same time, and additionally adopts one FCN to evaluate the difference and the relevance of the two parameters.
3. The method is used for carrying out joint modeling on the flight state parameters and the aerodynamic layout appearance parameters, and can analyze the differences and the correlations between the flight state parameters and the aerodynamic layout appearance parameters in the modeling process.
Drawings
The invention will be further specifically described with reference to the drawings and detailed description below:
FIG. 1 is a flow chart of the present invention.
Detailed Description
Example 1
Referring to fig. 1, a deep neural network modeling method of large-difference pneumatic data includes the following steps:
a. calculating a pneumatic data set, wherein the pneumatic data set comprises flight state parameters, pneumatic layout appearance parameters and aerodynamic coefficient true values, the pneumatic layout appearance parameters and the flight state parameters are used as input data, and the aerodynamic coefficient true values are used as output data;
b. the method comprises the steps of preprocessing a pneumatic data set, filtering and screening abnormal values and missing values existing in the pneumatic data set, normalizing all data in the pneumatic data set, and dividing a training set, a verification set and a test set according to a preset proportion;
c. c, constructing a model, namely determining the dimension of input data and the dimension of output data and data quantity information through the pneumatic data set calculated in the step a, determining the number of nodes of an input layer and an output layer in FCN_1 and CNN, preliminarily constructing a deep neural network model, and determining the number of network layers and the number of nodes of each layer in FCN_2 according to the scale of the pneumatic data set;
d. forward propagation, inputting aerodynamic layout appearance parameters in the same aerodynamic data in a training set into CNN, and outputting as vector f 1 Inputting the corresponding flight state parameters into FCN_1, and outputting the flight state parameters as a vector f 2 And then inputting the aerodynamic layout appearance parameter and the flight state parameter into the FCN_2 together, outputting the aerodynamic layout appearance parameter and the flight state parameter as a vector c, and outputting the aerodynamic layout appearance parameter and the flight state parameter as a neural network:
Figure BDA0003226484300000061
wherein y is z Is the output value of the model when the z-th sample is taken as input; q is the dimension of the aerodynamic coefficient true value, and is equal to the output dimension of CNN and FCN_1; f (f) 1i,z Is the ith output component when the z-th input sample is input to the CNN; c i,z Weights for the ith output component when the z-th input sample is input to the CNN; f (f) 2j,z Is the jth output component when the jth input sample is input to fcn_1; c q+j,z A weight for a jth output component when a zth input sample is input to fcn_1;
e. back propagation, the loss function of the model is as follows:
Figure BDA0003226484300000062
wherein L is MSE A loss function for the model; n is the number of samples during model training; y is Z Is the predicted aerodynamic coefficient for the z-th sample;
Figure BDA0003226484300000063
the true aerodynamic coefficient of the z-th aerodynamic sample in the aerodynamic data set;
f. model verification and optimization, the number of layers and node numbers of FCN_1, FCN_2 and CNN networks are continuously adjusted, and the model is verified and optimized through the constructed verification set and test set.
As a complete technical scheme, compared with the prior art, the method can meet modeling requirements when the aerodynamic shape and the flight state are changed simultaneously, can give out the influence degree of the flight state change and the aerodynamic shape change on the aerodynamic characteristics of the aircraft, and can well optimize the aerodynamic characteristics of the aircraft.
Example 2
Referring to fig. 1, a deep neural network modeling method of large-difference pneumatic data includes the following steps:
a. calculating a pneumatic data set, wherein the pneumatic data set comprises flight state parameters, pneumatic layout appearance parameters and aerodynamic coefficient true values, the pneumatic layout appearance parameters and the flight state parameters are used as input data, and the aerodynamic coefficient true values are used as output data;
b. the method comprises the steps of preprocessing a pneumatic data set, filtering and screening abnormal values and missing values existing in the pneumatic data set, normalizing all data in the pneumatic data set, and dividing a training set, a verification set and a test set according to a preset proportion;
c. c, constructing a model, namely determining the dimension of input data and the dimension of output data and data quantity information through the pneumatic data set calculated in the step a, determining the number of nodes of an input layer and an output layer in FCN_1 and CNN, preliminarily constructing a deep neural network model, and determining the number of network layers and the number of nodes of each layer in FCN_2 according to the scale of the pneumatic data set;
d. forward propagation, inputting aerodynamic layout appearance parameters in the same aerodynamic data in a training set into CNN, and outputting as vector f 1 Inputting the corresponding flight state parameters into FCN_1, and outputting the flight state parameters as a vector f 2 And then inputting the aerodynamic layout appearance parameter and the flight state parameter into the FCN_2 together, outputting the aerodynamic layout appearance parameter and the flight state parameter as a vector c, and outputting the aerodynamic layout appearance parameter and the flight state parameter as a neural network:
Figure BDA0003226484300000071
wherein y is z Is the output value of the model when the z-th sample is taken as input; q is the dimension of the aerodynamic coefficient true value, and is equal to the output dimension of CNN and FCN_1; f (f) 1i,z Is the ith output component when the z-th input sample is input to the CNN; c i,z Weights for the ith output component when the z-th input sample is input to the CNN; f (f) 2j,z Is the jth output component when the jth input sample is input to fcn_1; c q+j,z A weight for a jth output component when a zth input sample is input to fcn_1;
e. back propagation, the loss function of the model is as follows:
Figure BDA0003226484300000072
wherein L is MSE A loss function for the model; n is the number of samples during model training; y is Z Is the predicted aerodynamic coefficient for the z-th sample;
Figure BDA0003226484300000073
the true aerodynamic coefficient of the z-th aerodynamic sample in the aerodynamic data set;
f. model verification and optimization, the number of layers and node numbers of FCN_1, FCN_2 and CNN networks are continuously adjusted, and the model is verified and optimized through the constructed verification set and test set.
In the step a, the calculation of the aerodynamic data set refers to the calculation of aerodynamic coefficient true values of the aircraft through CFD software according to two-dimensional coordinates and flight state parameters of the aircraft wing shape.
In the step a, the coordinate information of the aerodynamic layout appearance parameters is obtained through profile software according to the wing shape of the real aircraft, and the aerodynamic coefficient true value is calculated through CFD software.
In the step a, flight state parameters comprise Mach numbers, attack angles and Reynolds numbers, and aerodynamic coefficient true values comprise lift coefficients and drag coefficients.
Example 3
Referring to fig. 1, a deep neural network modeling method of large-difference pneumatic data includes the following steps:
a. calculating a pneumatic data set, wherein the pneumatic data set comprises flight state parameters, pneumatic layout appearance parameters and aerodynamic coefficient true values, the pneumatic layout appearance parameters and the flight state parameters are used as input data, and the aerodynamic coefficient true values are used as output data;
b. the method comprises the steps of preprocessing a pneumatic data set, filtering and screening abnormal values and missing values existing in the pneumatic data set, normalizing all data in the pneumatic data set, and dividing a training set, a verification set and a test set according to a preset proportion;
c. c, constructing a model, namely determining the dimension of input data and the dimension of output data and data quantity information through the pneumatic data set calculated in the step a, determining the number of nodes of an input layer and an output layer in FCN_1 and CNN, preliminarily constructing a deep neural network model, and determining the number of network layers and the number of nodes of each layer in FCN_2 according to the scale of the pneumatic data set;
d. forward propagation, inputting aerodynamic layout appearance parameters in the same aerodynamic data in a training set into CNN, and outputting as vector f 1 Inputting the corresponding flight state parameters into FCN_1, and outputting the flight state parameters as a vector f 2 Inputting the aerodynamic layout shape parameter and the flight state parameter into FCN_2, and outputting asThe output of the vector c, neural network is:
Figure BDA0003226484300000081
wherein y is z Is the output value of the model when the z-th sample is taken as input; q is the dimension of the aerodynamic coefficient true value, and is equal to the output dimension of CNN and FCN_1; f (f) 1i,z Is the ith output component when the z-th input sample is input to the CNN; c i,z Weights for the ith output component when the z-th input sample is input to the CNN; f (f) 2j,z Is the jth output component when the jth input sample is input to fcn_1; c q+j,z A weight for a jth output component when a zth input sample is input to fcn_1;
e. back propagation, the loss function of the model is as follows:
Figure BDA0003226484300000082
wherein L is MSE A loss function for the model; n is the number of samples during model training; y is Z Is the predicted aerodynamic coefficient for the z-th sample;
Figure BDA0003226484300000083
the true aerodynamic coefficient of the z-th aerodynamic sample in the aerodynamic data set;
f. model verification and optimization, the number of layers and node numbers of FCN_1, FCN_2 and CNN networks are continuously adjusted, and the model is verified and optimized through the constructed verification set and test set.
In the step a, the calculation of the aerodynamic data set refers to the calculation of aerodynamic coefficient true values of the aircraft through CFD software according to two-dimensional coordinates and flight state parameters of the aircraft wing shape.
In the step a, the coordinate information of the aerodynamic layout appearance parameters is obtained through profile software according to the wing shape of the real aircraft, and the aerodynamic coefficient true value is calculated through CFD software.
In the step a, flight state parameters comprise Mach numbers, attack angles and Reynolds numbers, and aerodynamic coefficient true values comprise lift coefficients and drag coefficients.
In the step b, the training set, the verification set and the test set are divided according to a preset proportion, namely, the training set, the verification set and the test set are divided according to a proportion of 8:1:1.
In the step d, the dimension of the vector c is 2q, the first q values express the weight of the aerodynamic force coefficient of the aerodynamic layout appearance parameter in the current flight state, the last q values express the weight of the aerodynamic force coefficient of the specific aerodynamic appearance in the current flight state, and the difference and the similarity between the first q values and the last q values of the vector c express the difference and the relevance between the aerodynamic layout appearance parameter and the flight state parameter.
Aiming at the problem that the pneumatic layout appearance parameters and the flight state parameters cannot be modeled jointly in the pneumatic modeling field, the method for modeling the flight state parameters and the pneumatic layout appearance parameters by using the deep neural network is provided, two deep neural networks of FCN and CNN are combined, the flight state parameters and the pneumatic layout appearance parameters can be processed respectively and simultaneously, and the difference and the relevance of the two parameters are evaluated by additionally adopting one FCN.
Example 4
Referring to fig. 1, a deep neural network modeling method of large-difference pneumatic data includes the following steps:
a. calculating a pneumatic data set, wherein the pneumatic data set comprises flight state parameters, pneumatic layout appearance parameters and aerodynamic coefficient true values, the pneumatic layout appearance parameters and the flight state parameters are used as input data, and the aerodynamic coefficient true values are used as output data;
b. the method comprises the steps of preprocessing a pneumatic data set, filtering and screening abnormal values and missing values existing in the pneumatic data set, normalizing all data in the pneumatic data set, and dividing a training set, a verification set and a test set according to a preset proportion;
c. c, constructing a model, namely determining the dimension of input data and the dimension of output data and data quantity information through the pneumatic data set calculated in the step a, determining the number of nodes of an input layer and an output layer in FCN_1 and CNN, preliminarily constructing a deep neural network model, and determining the number of network layers and the number of nodes of each layer in FCN_2 according to the scale of the pneumatic data set;
d. forward propagation, inputting aerodynamic layout appearance parameters in the same aerodynamic data in a training set into CNN, and outputting as vector f 1 Inputting the corresponding flight state parameters into FCN_1, and outputting the flight state parameters as a vector f 2 And then inputting the aerodynamic layout appearance parameter and the flight state parameter into the FCN_2 together, outputting the aerodynamic layout appearance parameter and the flight state parameter as a vector c, and outputting the aerodynamic layout appearance parameter and the flight state parameter as a neural network:
Figure BDA0003226484300000101
wherein y is z Is the output value of the model when the z-th sample is taken as input; q is the dimension of the aerodynamic coefficient true value, and is equal to the output dimension of CNN and FCN_1; f (f) 1i,z Is the ith output component when the z-th input sample is input to the CNN; c i,z Weights for the ith output component when the z-th input sample is input to the CNN; f (f) 2j,z Is the jth output component when the jth input sample is input to fcn_1; c q+j,z A weight for a jth output component when a zth input sample is input to fcn_1;
e. back propagation, the loss function of the model is as follows:
Figure BDA0003226484300000102
wherein L is MSE A loss function for the model; n is the number of samples during model training; y is Z Is the predicted aerodynamic coefficient for the z-th sample;
Figure BDA0003226484300000103
the true aerodynamic coefficient of the z-th aerodynamic sample in the aerodynamic data set;
f. model verification and optimization, the number of layers and node numbers of FCN_1, FCN_2 and CNN networks are continuously adjusted, and the model is verified and optimized through the constructed verification set and test set.
In the step a, the calculation of the aerodynamic data set refers to the calculation of aerodynamic coefficient true values of the aircraft through CFD software according to two-dimensional coordinates and flight state parameters of the aircraft wing shape.
In the step a, the coordinate information of the aerodynamic layout appearance parameters is obtained through profile software according to the wing shape of the real aircraft, and the aerodynamic coefficient true value is calculated through CFD software.
In the step a, flight state parameters comprise Mach numbers, attack angles and Reynolds numbers, and aerodynamic coefficient true values comprise lift coefficients and drag coefficients.
In the step b, the training set, the verification set and the test set are divided according to a preset proportion, namely, the training set, the verification set and the test set are divided according to a proportion of 8:1:1.
In the step d, the dimension of the vector c is 2q, the first q values express the weight of the aerodynamic force coefficient of the aerodynamic layout appearance parameter in the current flight state, the last q values express the weight of the aerodynamic force coefficient of the specific aerodynamic appearance in the current flight state, and the difference and the similarity between the first q values and the last q values of the vector c express the difference and the relevance between the aerodynamic layout appearance parameter and the flight state parameter.
In the step e, the back propagation specifically means by solving L MSE Regarding the gradient of the model parameter theta, updating the model parameter theta according to the gradient and the learning rate eta;
Figure BDA0003226484300000111
/>
wherein θ is a model parameter, η is a learning rate, L MSE Is a loss function of the model.
Example 5
Referring to fig. 1, a deep neural network modeling method of large-difference pneumatic data includes the following steps:
a. calculating a pneumatic data set, wherein the pneumatic data set comprises flight state parameters, pneumatic layout appearance parameters and aerodynamic coefficient true values, the pneumatic layout appearance parameters and the flight state parameters are used as input data, and the aerodynamic coefficient true values are used as output data;
b. the method comprises the steps of preprocessing a pneumatic data set, filtering and screening abnormal values and missing values existing in the pneumatic data set, normalizing all data in the pneumatic data set, and dividing a training set, a verification set and a test set according to a preset proportion;
c. c, constructing a model, namely determining the dimension of input data and the dimension of output data and data quantity information through the pneumatic data set calculated in the step a, determining the number of nodes of an input layer and an output layer in FCN_1 and CNN, preliminarily constructing a deep neural network model, and determining the number of network layers and the number of nodes of each layer in FCN_2 according to the scale of the pneumatic data set;
d. forward propagation, inputting aerodynamic layout appearance parameters in the same aerodynamic data in a training set into CNN, and outputting as vector f 1 Inputting the corresponding flight state parameters into FCN_1, and outputting the flight state parameters as a vector f 2 And then inputting the aerodynamic layout appearance parameter and the flight state parameter into the FCN_2 together, outputting the aerodynamic layout appearance parameter and the flight state parameter as a vector c, and outputting the aerodynamic layout appearance parameter and the flight state parameter as a neural network:
Figure BDA0003226484300000112
wherein y is z Is the output value of the model when the z-th sample is taken as input; q is the dimension of the aerodynamic coefficient true value, and is equal to the output dimension of CNN and FCN_1; f (f) 1i,z Is the ith output component when the z-th input sample is input to the CNN; c i,z Weights for the ith output component when the z-th input sample is input to the CNN; f (f) 2j,z Is the jth output component when the jth input sample is input to fcn_1; c q+j,z A weight for a jth output component when a zth input sample is input to fcn_1;
e. back propagation, the loss function of the model is as follows:
Figure BDA0003226484300000113
wherein L is MSE A loss function for the model; n is the number of samples during model training; y is Z Is the predicted aerodynamic coefficient for the z-th sample;
Figure BDA0003226484300000114
the true aerodynamic coefficient of the z-th aerodynamic sample in the aerodynamic data set;
f. model verification and optimization, the number of layers and node numbers of FCN_1, FCN_2 and CNN networks are continuously adjusted, and the model is verified and optimized through the constructed verification set and test set.
In the step a, the calculation of the aerodynamic data set refers to the calculation of aerodynamic coefficient true values of the aircraft through CFD software according to two-dimensional coordinates and flight state parameters of the aircraft wing shape.
In the step a, the coordinate information of the aerodynamic layout appearance parameters is obtained through profile software according to the wing shape of the real aircraft, and the aerodynamic coefficient true value is calculated through CFD software.
In the step a, flight state parameters comprise Mach numbers, attack angles and Reynolds numbers, and aerodynamic coefficient true values comprise lift coefficients and drag coefficients.
In the step b, the training set, the verification set and the test set are divided according to a preset proportion, namely, the training set, the verification set and the test set are divided according to a proportion of 8:1:1.
In the step d, the dimension of the vector c is 2q, the first q values express the weight of the aerodynamic force coefficient of the aerodynamic layout appearance parameter in the current flight state, the last q values express the weight of the aerodynamic force coefficient of the specific aerodynamic appearance in the current flight state, and the difference and the similarity between the first q values and the last q values of the vector c express the difference and the relevance between the aerodynamic layout appearance parameter and the flight state parameter.
In the step e, the back propagation specifically means by solving L MSE Regarding the gradient of the model parameter theta, updating the model parameter theta according to the gradient and the learning rate eta;
Figure BDA0003226484300000121
wherein θ is a model parameter, η is a learning rate, L MSE Is a loss function of the model.
In the step f, the verification and optimization of the model through the constructed verification set and the test set specifically means that K-fold cross verification is adopted, the model is verified on the constructed verification set, the structure of the model is adjusted and optimized according to the under-fitting and over-fitting phenomena in the verification process, and finally, the optimized model is evaluated on the test set.
The K-fold cross validation specifically refers to that an initial sample is divided into K sub-samples, one single sub-sample is reserved as data of a validation model, the other K-1 samples are used for training, the cross validation is repeated for K times, each sub-sample is validated once, and the results of the K times are averaged, so that a single estimation is finally obtained.
The CNN is used for processing the aerodynamic layout appearance parameters, FCN_1 is used for processing the flight state parameters and FCN_2 is used for learning the differences and the correlations of the flight state parameters and the aerodynamic layout appearance parameters in predicting aerodynamic coefficients.
On one hand, the method is used for carrying out joint modeling on the flight state parameters and the aerodynamic layout appearance parameters, and on the other hand, the method can analyze the variability and the relevance between the flight state parameters and the aerodynamic layout appearance parameters in the modeling process.

Claims (10)

1. The deep neural network modeling method of the large-difference pneumatic data is characterized by comprising the following steps of:
a. calculating a pneumatic data set, wherein the pneumatic data set comprises flight state parameters, pneumatic layout appearance parameters and aerodynamic coefficient true values, the pneumatic layout appearance parameters and the flight state parameters are used as input data, and the aerodynamic coefficient true values are used as output data;
b. the method comprises the steps of preprocessing a pneumatic data set, filtering and screening abnormal values and missing values existing in the pneumatic data set, normalizing all data in the pneumatic data set, and dividing a training set, a verification set and a test set according to a preset proportion;
c. c, constructing a model, namely determining the dimension of input data and the dimension of output data and data quantity information through the pneumatic data set calculated in the step a, determining the number of nodes of an input layer and an output layer in FCN_1 and CNN, preliminarily constructing a deep neural network model, and determining the number of network layers and the number of nodes of each layer in FCN_2 according to the scale of the pneumatic data set;
d. forward propagation, inputting aerodynamic layout appearance parameters in the same aerodynamic data in a training set into CNN, and outputting as vector f 1 Inputting the corresponding flight state parameters into FCN_1, and outputting the flight state parameters as a vector f 2 And then inputting the aerodynamic layout appearance parameter and the flight state parameter into the FCN_2 together, outputting the aerodynamic layout appearance parameter and the flight state parameter as a vector c, and outputting the aerodynamic layout appearance parameter and the flight state parameter as a neural network:
Figure FDA0003226484290000011
wherein y is z Is the output value of the model when the z-th sample is taken as input; q is the dimension of the aerodynamic coefficient true value, and is equal to the output dimension of CNN and FCN_1; f (f) 1i,z Is the ith output component when the z-th input sample is input to the CNN; c i,z Weights for the ith output component when the z-th input sample is input to the CNN; f (f) 2j,z Is the jth output component when the jth input sample is input to fcn_1; c q+j,z A weight for a jth output component when a zth input sample is input to fcn_1;
e. back propagation, the loss function of the model is as follows:
Figure FDA0003226484290000012
wherein L is MSE A loss function for the model; n is the number of samples during model training; y is Z To the z-th sampleThe predicted aerodynamic coefficient of the model;
Figure FDA0003226484290000013
the true aerodynamic coefficient of the z-th aerodynamic sample in the aerodynamic data set;
f. model verification and optimization, the number of layers and node numbers of FCN_1, FCN_2 and CNN networks are continuously adjusted, and the model is verified and optimized through the constructed verification set and test set.
2. The deep neural network modeling method of widely-differentiated pneumatic data according to claim 1, wherein: in the step a, the calculation of the aerodynamic data set refers to the calculation of aerodynamic coefficient true values of the aircraft through CFD software according to two-dimensional coordinates and flight state parameters of the aircraft wing shape.
3. The deep neural network modeling method of widely-differentiated pneumatic data according to claim 1, wherein: in the step a, the coordinate information of the aerodynamic layout appearance parameters is obtained through profile software according to the wing shape of the real aircraft, and the aerodynamic coefficient true value is calculated through CFD software.
4. The deep neural network modeling method of widely-differentiated pneumatic data according to claim 1, wherein: in the step a, flight state parameters comprise Mach numbers, attack angles and Reynolds numbers, and aerodynamic coefficient true values comprise lift coefficients and drag coefficients.
5. The deep neural network modeling method of widely-differentiated pneumatic data according to claim 1, wherein: in the step b, the training set, the verification set and the test set are divided according to a preset proportion, namely, the training set, the verification set and the test set are divided according to a proportion of 8:1:1.
6. The deep neural network modeling method of widely-differentiated pneumatic data according to claim 1, wherein: in the step d, the dimension of the vector c is 2q, the first q values express the weight of the aerodynamic force coefficient of the aerodynamic layout appearance parameter in the current flight state, the last q values express the weight of the aerodynamic force coefficient of the specific aerodynamic appearance in the current flight state, and the difference and the similarity between the first q values and the last q values of the vector c express the difference and the relevance between the aerodynamic layout appearance parameter and the flight state parameter.
7. The deep neural network modeling method of widely-differentiated pneumatic data according to claim 1, wherein: in the step e, the back propagation specifically means by solving L MSE Regarding the gradient of the model parameter theta, updating the model parameter theta according to the gradient and the learning rate eta;
Figure FDA0003226484290000021
wherein θ is a model parameter, η is a learning rate, L MSE Is a loss function of the model.
8. The deep neural network modeling method of widely-differentiated pneumatic data according to claim 1, wherein: in the step f, the verification and optimization of the model through the constructed verification set and the test set specifically means that K-fold cross verification is adopted, the model is verified on the constructed verification set, the structure of the model is adjusted and optimized according to the under-fitting and over-fitting phenomena in the verification process, and finally, the optimized model is evaluated on the test set.
9. The deep neural network modeling method of widely-differentiated pneumatic data according to claim 8, wherein: the K-fold cross validation specifically refers to that an initial sample is divided into K sub-samples, one single sub-sample is reserved as data of a validation model, the other K-1 samples are used for training, the cross validation is repeated for K times, each sub-sample is validated once, and the results of the K times are averaged, so that a single estimation is finally obtained.
10. The deep neural network modeling method of widely-differentiated pneumatic data according to claim 1, wherein: the CNN is used for processing the aerodynamic layout appearance parameters, FCN_1 is used for processing the flight state parameters and FCN_2 is used for learning the differences and the correlations of the flight state parameters and the aerodynamic layout appearance parameters in predicting aerodynamic coefficients.
CN202110972718.7A 2021-08-24 2021-08-24 Deep neural network modeling method for large-difference pneumatic data Active CN113836634B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110972718.7A CN113836634B (en) 2021-08-24 2021-08-24 Deep neural network modeling method for large-difference pneumatic data

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110972718.7A CN113836634B (en) 2021-08-24 2021-08-24 Deep neural network modeling method for large-difference pneumatic data

Publications (2)

Publication Number Publication Date
CN113836634A CN113836634A (en) 2021-12-24
CN113836634B true CN113836634B (en) 2023-05-05

Family

ID=78961014

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110972718.7A Active CN113836634B (en) 2021-08-24 2021-08-24 Deep neural network modeling method for large-difference pneumatic data

Country Status (1)

Country Link
CN (1) CN113836634B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114297779A (en) * 2021-12-28 2022-04-08 中国航天空气动力技术研究院 Method, system and equipment for quickly predicting aerodynamic coefficient of aircraft
CN114861292B (en) * 2022-03-22 2024-01-26 电子科技大学 Aircraft geometric feature and parameter joint modeling method based on deep learning
CN117634365B (en) * 2024-01-25 2024-04-26 中国航空工业集团公司沈阳空气动力研究所 Airplane aerodynamic force prediction method, electronic equipment and storage medium

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110730099A (en) * 2019-10-21 2020-01-24 电子科技大学 Flow prediction method based on historical flow data time sequence
CN111597698A (en) * 2020-05-08 2020-08-28 浙江大学 Method for realizing pneumatic optimization design based on deep learning multi-precision optimization algorithm
CN112597702A (en) * 2020-12-21 2021-04-02 电子科技大学 Pneumatic modeling generation type confrontation network model training method based on radial basis function

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3772709A1 (en) * 2019-08-06 2021-02-10 Robert Bosch GmbH Deep neural network with equilibrium solver

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110730099A (en) * 2019-10-21 2020-01-24 电子科技大学 Flow prediction method based on historical flow data time sequence
CN111597698A (en) * 2020-05-08 2020-08-28 浙江大学 Method for realizing pneumatic optimization design based on deep learning multi-precision optimization algorithm
CN112597702A (en) * 2020-12-21 2021-04-02 电子科技大学 Pneumatic modeling generation type confrontation network model training method based on radial basis function

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
邵会兵 ; 崔乃刚 ; 詹韬 ; .基于神经网络的飞行器控制方法及仿真研究.计算机仿真.2018,(第10期),104-108. *
黄煌 ; .一种基于深度神经网络的非线性控制分配方法.战术导弹技术.2017,(第04期),96-100. *

Also Published As

Publication number Publication date
CN113836634A (en) 2021-12-24

Similar Documents

Publication Publication Date Title
CN113836634B (en) Deep neural network modeling method for large-difference pneumatic data
Han et al. Surrogate-based optimization
CN111079228B (en) Pneumatic shape optimization method based on flow field prediction
CN105718634A (en) Airfoil robust optimization design method based on non-probability interval analysis model
CN112084727A (en) Transition prediction method based on neural network
CN104048676B (en) MEMS (Micro Electro Mechanical System) gyroscope random error compensating method based on improved particle filter
Kyriacou et al. Efficient PCA-driven EAs and metamodel-assisted EAs, with applications in turbomachinery
CN114235330A (en) Multi-source pneumatic load model construction method for correlated wind tunnel test and calculation data
Lee et al. Development of an optimized trend kriging model using regression analysis and selection process for optimal subset of basis functions
Amrit et al. Applications of surrogate-assisted and multi-fidelity multi-objective optimization algorithms to simulation-based aerodynamic design
Durasov et al. Debosh: Deep bayesian shape optimization
Amrit et al. Design strategies for multi-objective optimization of aerodynamic surfaces
CN114564787A (en) Bayesian optimization method, device and storage medium for target-related airfoil design
CN111625901B (en) Intelligent pressure coefficient curve generation method for wing profile
CN108984864A (en) Higher-dimension complex aircraft model characterizing method based on Gaussian process variance evaluation
Pehlivanoglu Direct and indirect design prediction in genetic algorithm for inverse design problems
CN115345064A (en) Integrated deep neural network aerodynamic modeling method integrated with physical constraints
Pisaroni et al. A multilevel monte carlo evolutionary algorithm for robust aerodynamic shape design
Amrit et al. Efficient multi-objective aerodynamic optimization by design space dimension reduction and co-kriging
CN113657029A (en) Efficient approximate optimization method for aircraft driven by heterogeneous data
Shloof et al. A highly accurate artificial neural networks scheme for solving higher multi‐order fractal‐fractional differential equations based on generalized Caputo derivative
CN114861292B (en) Aircraft geometric feature and parameter joint modeling method based on deep learning
Chakraborty et al. Multigoal-oriented dual-weighted-residual error estimation using deep neural networks
CN112182739A (en) Aircraft structure non-probability credible reliability topological optimization design method
Liu et al. A new data assimilation method of recovering turbulent flow field at high-Reynolds numbers for turbulence machine learning

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