CN117540489B - Airfoil pneumatic data calculation method and system based on multitask learning - Google Patents
Airfoil pneumatic data calculation method and system based on multitask learning Download PDFInfo
- Publication number
- CN117540489B CN117540489B CN202311506032.4A CN202311506032A CN117540489B CN 117540489 B CN117540489 B CN 117540489B CN 202311506032 A CN202311506032 A CN 202311506032A CN 117540489 B CN117540489 B CN 117540489B
- Authority
- CN
- China
- Prior art keywords
- airfoil
- network
- pressure coefficient
- input
- curve
- 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
- 238000004364 calculation method Methods 0.000 title claims abstract description 19
- 238000000034 method Methods 0.000 claims abstract description 24
- 238000012549 training Methods 0.000 claims abstract description 8
- 239000013598 vector Substances 0.000 claims description 40
- 230000006870 function Effects 0.000 claims description 32
- 238000013527 convolutional neural network Methods 0.000 claims description 24
- 230000004927 fusion Effects 0.000 claims description 16
- 230000007246 mechanism Effects 0.000 claims description 16
- 230000007613 environmental effect Effects 0.000 claims description 14
- 238000013528 artificial neural network Methods 0.000 claims description 12
- 239000011159 matrix material Substances 0.000 claims description 8
- 238000000605 extraction Methods 0.000 claims description 7
- 238000012545 processing Methods 0.000 claims description 7
- 238000005452 bending Methods 0.000 claims description 3
- 230000002708 enhancing effect Effects 0.000 claims description 2
- 238000007781 pre-processing Methods 0.000 claims description 2
- 238000013135 deep learning Methods 0.000 abstract description 11
- 239000012530 fluid Substances 0.000 abstract description 4
- 238000013136 deep learning model Methods 0.000 abstract description 3
- 230000004913 activation Effects 0.000 description 7
- 238000001994 activation Methods 0.000 description 7
- 238000013461 design Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 6
- 238000004458 analytical method Methods 0.000 description 5
- 230000007547 defect Effects 0.000 description 4
- 238000005457 optimization Methods 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 230000015556 catabolic process Effects 0.000 description 2
- 238000006731 degradation reaction Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 210000002569 neuron Anatomy 0.000 description 2
- 239000007787 solid Substances 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 238000013256 Gubra-Amylin NASH model Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 210000004027 cell Anatomy 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000001427 coherent effect Effects 0.000 description 1
- 239000003086 colorant Substances 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 230000004069 differentiation Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000003709 image segmentation Methods 0.000 description 1
- 230000009191 jumping Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 230000001052 transient effect Effects 0.000 description 1
Abstract
The invention relates to an airfoil pneumatic data calculation method and system based on multi-task learning, and belongs to the technical field of aerodynamics and deep learning. The scheme adopts a multitask learning method, an airfoil reconstruction network is constructed to extract airfoil geometric features, an airfoil flow field prediction network is designed to extract a coarse-granularity pressure coefficient curve of an airfoil in a global environment, an airfoil pressure coefficient generation network is designed to extract a fine-granularity pressure coefficient curve of the airfoil in a local environment, the airfoil flow field prediction network and the fine-granularity pressure coefficient curve are combined linearly through a cross-stitch network, and aerodynamic coefficients such as lift force, resistance and moment of the airfoil are calculated. The invention effectively combines the deep learning model AE, GAN, U-NET and the multi-task learning model with the fluid mechanics and the aerodynamics, digs the hidden characteristics of the deep learning model AE, GAN, U-NET and the multi-task learning model, accurately predicts the aerodynamic coefficient of the wing section, can well improve the comprehensiveness and the effectiveness of the training of the prediction model, can improve the accuracy of the calculation of the aerodynamic data, and has wide application prospect.
Description
Technical Field
The invention belongs to the technical field of aerodynamics and deep learning, and relates to an airfoil pneumatic data calculation method and system based on multi-task learning.
Background
The airfoil optimization design is a key step in the aircraft design, determines the performance index of the aircraft in actual flight, and accurately calculates aerodynamic data such as lift force, resistance and the like of the airfoil, which has extremely important influence on the design optimization of the airfoil and the performance, maneuverability and stability of the aircraft, but the airfoil aerodynamic analysis often has the defects of long period, high computing resource expenditure and the like, and is not suitable for complex and changeable aircraft flight environments.
The traditional airfoil pneumatic analysis method is generally CFD simulation, and has the advantages that the operational safety risk does not exist, meanwhile, an approximate control equation is calculated, the data precision is relatively high, and the cost is relatively low; the method has the defects that the calculation accuracy is easily affected by the grid density, and is a very time-consuming calculation process when large iterative flow solutions are needed for airfoil optimization, fluid-solid coupling and the like, and a part of complex control equations have no numerical solutions. The current popular solutions are ROM degradation models, deep learning methods, etc. The ROM degradation model reduces the order of the control equation, simplifies the control equation, improves the solving efficiency, but the high nonlinearity among the data makes the calculation of the pneumatic data more difficult, so that the method is difficult to be suitable for multi-scale, transient and discontinuous processes, and the deep learning method can exactly make up the defects of the traditional method, so that the method adopts the deep learning method to calculate the pneumatic data of the aircraft.
In recent decades, with the advent of the information age, data has been explosively increased, so that many excellent data analysis methods have been promoted, and deep learning is the most blazed new star. Deep learning, namely DNN, is realized by designing a multi-layer neural network, so that neurons of each layer are deepened and widened continuously, and the neurons can be mapped to any function theoretically, so that the problem of complexity is solved, and the method is very outstanding in the fields of image recognition, face recognition and the like. In the hydrodynamic and aerodynamic fields such as aeroplane aerodynamic performance analysis and flow field prediction, the deep learning method also has the remarkable potential, the solving speed of aerodynamic data is greatly improved, the problems of high nonlinearity and high dimensionality among the data are solved, and the deep learning is used for feature extraction, feature fusion and the like, which are commonly used in hydrodynamic analysis.
Disclosure of Invention
In view of the above, the present invention aims to provide a method and a system for calculating aerodynamic data of an airfoil based on multi-task learning, which adopts a multi-task learning method to construct an airfoil reconstruction network to extract geometrical characteristics of the airfoil, designs a coarse-granularity pressure coefficient curve of the airfoil flow field prediction network to extract the airfoil in a global environment, designs a fine-granularity pressure coefficient curve of the airfoil pressure coefficient generation network to extract the airfoil in a local environment, and combines the two in a linear manner through a cross-stitch network to calculate aerodynamic coefficients such as lift force, resistance, moment, etc. of the airfoil. Training data is reduced through a multitask learning method, meanwhile, the model has physical interpretability due to the fact that PINN is added, the result accuracy of the multitask learning model is higher due to the fact that the model is preprocessed through geometric features of the wing profile, training speed is faster, the model can be converged more quickly, and the prediction model is more comprehensive.
In order to achieve the above purpose, the present invention provides the following technical solutions:
an airfoil pneumatic data calculation method based on multitask learning, which comprises the following steps:
S1, data preprocessing: the method comprises the steps of extracting geometrical characteristics of an airfoil by using a convolutional neural network CNN, inputting calculated coordinates, curvature and airfoil pictures of the airfoil, reconstructing the airfoil and self-attention through an AE network, and extracting latent vector of the airfoil;
s2, task one: predicting a flow field of the wing profile by using a U-NET model, fusing the wing profile latent vector in the step S1 by using the model, and fusing the partial conductance of the pressure field and the speed field in PINN by using a loss function;
s3, task two: reconstructing a fine-grained airfoil pressure coefficient curve by using GAN, and carrying out feature fusion on airfoil environmental features and airfoils latent vector;
S4, extracting the characteristic of a rough-granularity airfoil pressure coefficient curve of the predicted flow field in the task I through a convolutional neural network, combining the rough-granularity airfoil pressure coefficient curve with the fine-granularity airfoil pressure coefficient curve obtained in the task II through a cross-stitch network, and outputting the aerodynamic coefficient of the airfoil.
Further, in step S1, specifically including:
S11、 curve extraction airfoil curvature characteristics: use/> The curve represents the two-dimensional coordinates of the airfoil as a plurality of polynomials, and the bending degree, namely the curvature of the curve is calculated from the tangent space of each polynomial;
s12, extracting airfoil geometric features by a convolutional neural network CNN: converting the two-dimensional coordinates of the wing profile, the wing profile picture and the curvature obtained in S11 into a matrix data form to be used as the input of CNN;
s13, extracting latent vector of an airfoil by an AE network: by encoder-decoder architecture, convolutional and deconvolution layers are constructed to extract latent vector of airfoil geometry in a manner that reconstructs the airfoil.
Further, in step S2, specifically including:
S21, openFoam software generates flow field data and aerodynamic data of the airfoil shape: under the real condition, the Mach number is 0.1-0.6, the airfoil attack angle is between plus or minus 10 degrees, the Reynolds number is between 6 times and 7 times, and parameters such as lift force, resistance, moment and the like of different airfoil flow fields and airfoil aerodynamic forces under each condition are calculated;
S22, predicting a flow field of the airfoil by a network structure of the U-NET: inputting the wing profile picture and the environmental characteristic picture, such as attack angle, flow speed and the like, into a model, outputting a corresponding pressure field and speed fields in x and y directions of the wing profile under the condition, wherein the LOSS function is L1 LOSS;
S23, enhancing the interpretability of the model using PINN: knowing that the loss function in S22 is MSE, the bias of the pressure field and the bias of the velocity field are added to the loss function to make it have a certain physical interpretability.
Further, in step S3, specifically including:
S31, full-connection-layer connection airfoil latent vector and environmental features: latent vector obtained by the S1-3 is linearly spliced with the environmental characteristics through a fully connected neural network, and the influence of each node on the result is corrected through the step S32;
s32, calculating latent vector a mutual influence relation between the self-attention mechanism and the environment characteristics: the self-attention mechanism is fused with a fully connected network, the main purpose of which is to let the network notice the correlation between the different parts of the whole input. Firstly, setting key, query, value according to different input vector matrixes, secondly, calculating the correlation between every two input vectors by using keys and query, namely calculating attention value alpha, performing softmax operation on the matrix A to obtain a new matrix A', and finally obtaining output B corresponding to input; the output obtained by the self-attention mechanism module contains the interrelationship among all the inputs, so that the model training effect is better;
S33, fully-connected network fusion airfoil profile: taking the airfoil hidden layer characteristics and the environmental characteristics which are obtained through the self-attention module processing in the step S32 as the input of a fusion network, and compressing the input characteristics to lower dimensionality through the fusion network to be used as the input of a subsequent GAN network;
s34, generating a pressure coefficient curve by a generator of the GAN: obtaining the input of a generator according to a fusion network formed by a self-attention mechanism in the step S33, and simultaneously taking a random variable z as part of the input of the generator according to the property of the GAN, and generating a new pressure coefficient curve as the input of a discriminator of the GAN through a deconvolution network;
S35, a discriminator of GAN discriminates the generated and true pressure coefficient curves: according to the step S34, the generated pressure coefficient curve and the false pressure coefficient curve are obtained, and meanwhile, the true pressure coefficient curve is input into the discriminator together, and the difference between the true pressure coefficient curve and the true pressure coefficient curve is calculated;
s36, calculating LOSS, and updating network parameters: and (3) according to the difference obtained in the step S35, simultaneously updating the whole network parameters by combining the loss function, and training.
Further, in step S4, specifically including:
s41, connecting output of a first task and output of a second task through a cross stitch network: linearly connecting the vector of the task I passing through the full connection layer and the vector of the task II passing through the convolution neural network, and calculating the mutual influence coefficient of the vector and the vector;
s42, outputting the aerodynamic coefficient of the airfoil, calculating LOSS, and updating the network: through the study of the cross stitch network, the aerodynamic coefficients of lift force, resistance, moment and the like of the wing profile are output, and the network is updated by using L1 LOSS.
The invention also provides an airfoil pneumatic data computing system based on the multi-task learning.
The invention has the beneficial effects that:
The invention designs a new model, effectively combines a deep learning model AE, GAN, U-NET and a multi-task learning model with fluid mechanics and aerodynamics, digs hidden characteristics of the model, and accurately predicts aerodynamic coefficients of wing sections. Meanwhile, the geometric shape of the wing profile can be better represented by the model through extracting the geometric information, curvature and other characteristics of the wing profile. And a self-attention mechanism is used for constructing a fusion network, so that the model can better fuse the relations among different features, and the prediction result is more accurate. Use PINN makes the model more physically interpretable. The use of a deep learning network in the model enables the model to converge faster and enables the predictive model to be trained more comprehensively.
Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention. The objects and other advantages of the invention may be realized and obtained by means of the instrumentalities and combinations particularly pointed out in the specification.
Drawings
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in the following preferred detail with reference to the accompanying drawings, in which:
FIG. 1 is a logic block diagram of an airfoil pneumatic data calculation method based on multitasking learning of the present invention;
FIG. 2 is a detailed network architecture diagram of pneumatic data computation;
fig. 3 is a diagram of a cross stitch network structure.
Detailed Description
The technical scheme of the invention is described in detail below with reference to the accompanying drawings.
In the present invention, the following concepts are mainly included:
(1) Curvature: the curvature of a curve is defined by differentiation for the rotation rate of the tangential angle to the arc length at a point on the curve, indicating the extent to which the curve deviates from a straight line. Mathematically, a numerical value indicating the degree of curve bending at a certain point.
(2) Airfoil geometry: refers to the chord length, thickness, etc. of the airfoil in Euclidean space.
(3) Pressure coefficient: the pressure coefficient is a dimensionless number that describes the relative pressure throughout the flow field in fluid dynamics. The pressure coefficient is applied to aerodynamics and hydrodynamics, and the pressure coefficient is different at each point in the fluid flow field.
(4) Pneumatic coefficient: the aerodynamic coefficient is a dimensionless number, referred to in this patent as the lift, drag, and moment of the airfoil.
(5) GAN: the false data is generated by the generator, and the false data and the true data are identified by the generator and are mutually game, so that the purpose of model convergence is achieved.
(6) Self-attention mechanism: the self-attention mechanism mainly solves the correlation problem between different parts in the input without considering the time and latitude, and the problem is solved by key, value, query.
(7) CNN: convolutional neural networks are a type of feedforward neural network that includes convolutional computation and has a deep structure, and are one of representative algorithms for deep learning. Convolutional neural networks have a characteristic learning capability and can perform translation-invariant classification on input information according to a hierarchical structure of the convolutional neural networks, so the convolutional neural networks are also called as 'translation-invariant artificial neural networks'.
(8) U-NET: in the network model which is used for image processing such as image segmentation, image generation and the like, compared with CNN, the network model is more focused on classification of image pixel levels, has better image processing effect and is suitable for generating airfoil flow field prediction images. The network architecture employs an encoder-decoder architecture, including downsampling, upsampling, and jumping layers.
The invention provides an airfoil pneumatic data calculation method based on multitask learning, which can accurately predict airfoil pneumatic data, wherein an overall logic block diagram is shown in figure 1, and comprises an airfoil characteristic extraction module, an airfoil flow field prediction module, an airfoil pressure coefficient calculation module and an airfoil pneumatic data calculation module, an overall network structure diagram is shown in figure 2, and an overall network structure diagram of a multitask model is shown.
The embodiment specifically comprises the following steps:
a first part comprising the following three steps:
Step S11: The curve provides airfoil curvature characteristics and codes. Use/> The curve represents the airfoil two-dimensional coordinates as a plurality of polynomials, and the degree of curvature, i.e., the curvature, of the curve is calculated from the tangential space of each polynomial. Given a two-dimensional airfoil coordinate set d= { P i＝(xi,yi) } i=1, 2, …, M }, M representing the number of coordinate points, P representing coordinate points, (x i,yi) representing abscissa and ordinate, D is composed of 192 two-dimensional coordinates from which 3 degrees/> can be constructed by each airfoil coordinate set, according to the actual coordinate setCurve, formula is
R (D; t) representsThe function of the curve, D is the sample space in which the airfoil coordinates are located, t is/>Parameters of the curve, n is/>The order of the curve, B i,n (t), needs to satisfy the following formula
According to 3 degrees of constructionThe curve can be represented as a smooth curve with 4 points, then the entire two-dimensional airfoil coordinate set can be represented as multiple segments of 3 degrees/>The linear combination of curves characterizes the entire airfoil, which can be derived/>The wing profile curve is perceived through the control parameter t in the curve, and the curvature at each t can be conveniently calculated.
The smooth continuity of the airfoil surface is expressed in terms of curvature as follows
Y 'is the first derivative of the airfoil ordinate y, y″ is the second derivative of the airfoil ordinate y, y' "is the third derivative of the airfoil ordinate y, and each coordinate point can take a curvature, so each airfoil has 192 curvatures which can be represented.
The fully connected neural network encodes airfoil curvature characteristics. The generated curvature is encoded using a fully connected neural network, the result of which is input to an AE network encoder.
Step S12: CNN encodes airfoil geometry. The two-dimensional coordinate of the wing profile and the wing profile picture are encoded by using a convolutional neural network CNN, geometrical characteristics of the wing profile, namely the chord length, the thickness and the like of the wing profile are extracted, and the result is used as input of a convolutional layer. The two-dimensional coordinates of the wing profile and the wing profile pictures are from a UIUC wing profile database, after data processing, about 1000 wing profiles can be used, and each wing profile two-dimensional coordinate passes throughAfter curve processing, 192 two-dimensional coordinate points are taken to represent the geometric shape of each airfoil, namely, the input characteristic size of CNN is 192 x2 x1, and the airfoil picture is similarly preprocessed, wherein the input characteristic size is 256 x 3. The CNN encoded airfoil geometry feature consists of a plurality of convolution layers, each followed by a BN layer, relu functions and dropout functions, the kernel size of all convolution layers being set to 4*2, the size of the last layer being 512-dimensional data, meaning that the airfoil geometry feature is represented using 512-dimensional data.
Step S13: the AE network extracts latent vector of the airfoil. The convolution layer and the deconvolution layer are constructed through the architecture of the encoder-decoder, the model is input into the curvature codes obtained in the step S11 and the wing profile geometric features obtained in the step S12, the model is output into the wing profile, and latent vector of the wing profile geometric features are extracted in a wing profile reconstruction mode. The structure of the convolution layer is referred to in steps S11 and S12, and the deconvolution layer structure is opposite to the convolution layer structure. The LOSS function uses a classical mean square error function MSE, the formula is as follows:
Y i represents the original airfoil coordinates or airfoil picture, The wing profile coordinates or wing profile pictures generated through the AE model are represented, and n represents the number of wing profiles. In summary, when the model is trained to converge, an encoder is obtained that can arbitrarily generate the airfoil latent vector.
A second part comprising the following three steps:
Step S21: openFoam software generates flow field data and aerodynamic data for the airfoil. And writing a Python script file of OpenFoam software, and performing flow field analysis and aerodynamic computation on more than 1000 airfoils in the UICC database. Under the real condition, the Mach number has the value range of Ma epsilon {0.1,0.2,0.3,0.4,0.5,0.6}, the airfoil attack angle has the value range of AOA epsilon [ -10 degrees, 10 degrees ], the Reynolds number has the value range of Re epsilon {1e6,6e6,1e7}, and the parameters of different airfoil flow fields, airfoil aerodynamic forces, such as lift force, resistance force, moment and the like, under each condition are calculated.
Step S22: the network structure of the U-NET predicts the flow field of the airfoil. The input of the model is an airfoil picture, an airfoil attack angle picture and an airfoil Mach number picture respectively, and the pictures are respectively represented by two different solid colors, so that the influence of the pictures on different prediction results is distinguished; the output of the model is three pictures, namely a pressure field picture of the wing profile and a speed field picture in the X and Y directions, so as to represent the change condition of the current wing profile in the current condition. The model is constructed as a U-NET, also an encoder-decoder architecture, and in the encoded section, cross-row convolution is used to progressively reduce the image size by a factor of 2. This allows the network to extract larger and larger scale and abstract information in more and more feature channels. The decoding part of the network reflects this behavior and reduces the number of feature layers by averaging the pooling layers to increase spatial resolution. The skip connection connects all channels of the coding branch to the corresponding branch of the decoding section, effectively doubling the number of channels per decoding block. These skipped connections help the network take low-level input information into account when decoding the layer reconstruction solution. Each part of the network comprises, in addition to the nonlinear activation function, a convolution layer, a normalization layer. In this embodiment 7 convolution blocks are used to convert 1282 by 3 inputs into a single data point with 512 features, typically using a convolution kernel of size 42. As an activation function, a ReLU function with a slope of 0.2 is used at the encoding layer and regular ReLU activation is used at the decoding layer. The decoder section reconstructs the objective function using 7 further symmetrical layers, with the desired dimension being 1282 x 3. The jump connection of the encoder to the decoder ensures that the original information can be used in the decoder to ensure that the information loses excessive accuracy during downsampling, upsampling.
Step S23: the loss function of U-NET and its use PINN enhance the interpretability of the model. L1 LOSS is generally used as a LOSS function for U-NET, and the formula is as follows:
The difference between the MSE LOSS function and the generated picture is calculated by subtracting the pixel of the generated picture from the pixel of the original picture, and the main defects are that the generated picture is not accurate enough, the curve of the wing profile is not coherent enough, the whole experiment is driven by data and lacks a certain physical meaning, so that the PINN study of the forefront edge is added into the LOSS function to lead the LOSS to have a certain physical meaning, and the LOSS function is as follows:
The latter two terms are the partial derivatives of the velocity field and the pressure field, so as to update the parameters of the neural network.
A third part comprising the following six steps:
Step S31: the full tie layer connects the resulting airfoil latent vector to the airfoil environmental characteristics of S13. The airfoil environmental features are Mach number, reynolds number and attack angle, the range of values is the same as S21, the environmental features with dimension 3 are expanded into hidden layer variables with latitude 256 through a full-connection layer, and the hidden layer variables are connected with latent vector of the airfoil to serve as input of a feature fusion module.
Step S32: the self-attention mechanism calculates the relationship of the different input features. The self-attention mechanism is fused with the fully-connected network, and the main purpose of the self-attention mechanism is to make the network notice the correlation between different parts in the whole input, so that the network can converge more quickly, the calculated amount is reduced, and the model accuracy is increased. As can be seen from S31, the input to the self-attention module is the airfoil latent coding and the environmental features, let each feature be assumed to be a i, and each feature is resized to maintain a 512-dimensional data size as the input to the self-attention module, as shown in FIG. 2. The self-attention mechanism has the following steps:
Step one, for each input vector a i, multiplying three coefficients w q、wk、wv, respectively, with the formula
qi＝wq·ai
ki＝wk·ai
vi＝wv·ai
Q, k and v obtained respectively represent query, key and value;
Calculating the correlation between two input vectors by using the obtained Q and K, namely calculating the value alpha of Attention, wherein the calculation is usually performed by using a dot product mode, the formula is alpha i,j＝qi·kj, the written vector form is A=K T.Q, and each numerical value in the matrix A records the Attention size alpha of the corresponding two input vectors;
Step three, performing softmax operation or relu operation on the matrix A to obtain A';
Step four, calculating an output vector b i of self-attention layers corresponding to each input vector a i by using the obtained A' and V, wherein the formula is that The corresponding vector form is o=v·a';
Through the calculation steps of the autonomous force mechanism, the interrelation between the input features can be known, each feature a i is input, and the output feature b i is obtained, and the dimension size of the output feature b i is not changed.
Step S33: the fully connected network merges airfoil features. The airfoil geometry and the environmental characteristics obtained through the self-attention module processing in the step S32 are used as the input of the fusion network, and the input characteristics are compressed to a lower latitude through the fusion network and used as the input of the subsequent GAN network.
Step S34: as can be seen from fig. 2, the pressure coefficient curve is generated by a generator of GAN. The real data of the pressure coefficient curve of the airfoil is set as x and is used as the input of the discriminator. The input z of the generator is obtained according to the fusion network formed by the self-attention mechanism in S32, and the newly generated pressure coefficient curve is obtained through the decoder formed by the deconvolution networkAs input to the discriminator. Setting random noise/>Also as input to the discriminator. The encoder is composed of deconvolution layers, and outputs as synthetic and spurious airfoil pressure coefficient curve data.
Step S35: the discriminator D discriminates between the generated and true pressure coefficient curves. And (3) obtaining a generated pressure coefficient curve and a false pressure coefficient curve according to the step (S34), inputting the true pressure coefficient curves into a discriminator together, and respectively calculating the difference of the two pressure coefficient curves. The purpose of the discriminator is to discriminate, as far as possible, that the data generated by the generator is false.
Step S36: and calculating LOSS and updating network parameters. And according to the difference obtained in the step S32, simultaneously updating the whole network parameters by combining with the LOSS function, and training. The AE has poor capability of generating samples, so that a discriminator introducing GAN accelerates model convergence and improves the quality of model generated data. The generator attempts to generate spurious data to fool the discriminator D, which continually sharpens its decision boundary to determine the resultant spurious samples from the true samples, with the max-min game cross entropy loss function as follows:
GAN generates samples purely from random noise, and it is difficult to obtain an explicit mapping from the data domain to the feature domain. Therefore, the input data is subjected to feature extraction through the AE model, and an AE-GAN model is constructed. Note that in AE-GAN, the model generates a reconstructed sample Given the real sample x, a false sample/>, is generated simultaneouslyDirectly from random noise/>Reconstructed sample/>And false samples/>Should be classified as false by discriminator D, only x is considered as true by the discriminator, so the GAN loss function can be modified to the following equation:
It takes into account real, reconstructed and false samples, and in addition, to stabilize the training process and sharpen the decision boundary of D, another loss function L layer is introduced, formulated as follows:
In summary, the GAN network loss function is a weighted combination of the above loss functions, and the formula is as follows:
L＝λ0Llayer+λ1LGAN
a fourth part comprising the following three steps:
Step S41: the output of S2 and the output of S3 are linked by a "cross stitch" network. The structure of a "cross-stitch" network is shown in fig. 3, which attempts to find the best shared representations for multitasking learning, they model these shared representations using linear combinations, and learn the best linear combinations for a given set of tasks. These cross stitch units are integrated into convolutional neural networks and provide an end-to-end learning framework. Consider a case of multitasking learning, where there are two tasks a and B on the same input image. For ease of explanation, consider two networks trained separately for these tasks. The present invention proposes a new unit, the "cross stitch" unit, which combines the two networks into a multi-tasking network so that the task can oversee how much sharing is needed. At each layer of the network, the token sharing is modeled by learning a linear combination of activation maps using cross-stitch cells. Given two activation maps x A,xB for layer 1 of two tasks, learn a linear combination of two input activations And feed these combinations as inputs to the filters of the next layer. This linear combination is parameterized with α. Specifically, at the position (i, j) of the activation map, the formula is as follows:
Step S42: and outputting the aerodynamic coefficient of the airfoil, calculating LOSS, and updating the network. The output of the first task carries out the characteristic extraction of the fine granularity pressure coefficient curve through the fully connected neural network, the input of the second task carries out the characteristic extraction of the coarse granularity pressure coefficient curve through the convolution neural network, the two tasks directly carry out cross learning through the cross stitch network, the aerodynamic coefficients of the wing profile such as lift force, resistance, moment and the like are output, and finally the L1 LOSS is used for updating network parameters.
Finally, it is noted that the above embodiments are only for illustrating the technical solution of the present invention and not for limiting the same, and although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solution of the present invention may be modified without departing from the spirit and scope of the technical solution, and all such modifications are included in the scope of the claims of the present invention.
Claims (4)
1. A wing section pneumatic data calculation method based on multitask learning is characterized in that: the method comprises the following steps:
S1, data preprocessing: the method comprises the steps of extracting geometrical characteristics of an airfoil by using a convolutional neural network CNN, inputting calculated coordinates, curvature and airfoil pictures of the airfoil, reconstructing the airfoil and self-attention through an AE network, and extracting latent vector of the airfoil;
s2, task one: predicting a flow field of the wing profile by using a U-NET model, fusing the wing profile latent vector in the step S1 by using the model, and fusing the partial conductance of the pressure field and the speed field in PINN by using a loss function;
s3, task two: reconstructing a fine-grained airfoil pressure coefficient curve by using GAN, and carrying out feature fusion on airfoil environmental features and airfoils latent vector;
s4, extracting the characteristic of a rough-granularity airfoil pressure coefficient curve of a predicted flow field in the task I through a convolutional neural network, combining the rough-granularity airfoil pressure coefficient curve with a fine-granularity airfoil pressure coefficient curve obtained in the task II through a cross-stitch network, and outputting the pneumatic coefficient of an airfoil;
In the step S2, specifically, the method includes:
S21, openFoam software generates flow field data and aerodynamic data of the airfoil shape: under the real condition, the Mach number is 0.1-0.6, the airfoil attack angle is between plus or minus 10 degrees, the Reynolds number is 10 to the power of 6 to 7, and different airfoil flow fields and airfoil aerodynamic forces under each condition are calculated;
s22, predicting a flow field of the airfoil by a network structure of the U-NET: inputting the wing profile picture and the environment characteristic picture into a model, outputting a corresponding pressure field and speed fields in the x and y directions of the wing profile under the condition, wherein the LOSS function is L1 LOSS;
S23, enhancing the interpretability of the model using PINN: knowing that the loss function in S22 is MSE, adding the bias of the pressure field and the bias of the velocity field into the loss function to make the loss function have a certain physical interpretability;
in the step S3, the method specifically includes:
S31, full-connection-layer connection airfoil latent vector and environmental features: latent vector obtained by the S1-3 is linearly spliced with the environmental characteristics through a fully connected neural network, and the influence of each node on the result is corrected through the step S32;
S32, calculating latent vector a mutual influence relation between the self-attention mechanism and the environment characteristics: fusing a self-attention mechanism with a fully-connected network, firstly setting key, query, value according to different input vector matrixes, secondly calculating the correlation between every two input vectors by using keys and queries, namely calculating attention value alpha, performing softmax operation on the matrix A to obtain a new matrix A', and finally obtaining output B corresponding to the input;
S33, fully-connected network fusion airfoil profile: taking the airfoil hidden layer characteristics and the environmental characteristics which are obtained through the self-attention module processing in the step S32 as the input of a fusion network, and compressing the input characteristics to lower dimensionality through the fusion network to be used as the input of a subsequent GAN network;
s34, generating a pressure coefficient curve by a generator of the GAN: obtaining the input of a generator according to a fusion network formed by a self-attention mechanism in the step S33, and simultaneously taking a random variable z as part of the input of the generator according to the property of the GAN, and generating a new pressure coefficient curve as the input of a discriminator of the GAN through a deconvolution network;
S35, a discriminator of GAN discriminates the generated and true pressure coefficient curves: according to the step S34, the generated pressure coefficient curve and the false pressure coefficient curve are obtained, and meanwhile, the true pressure coefficient curve is input into the discriminator together, and the difference between the true pressure coefficient curve and the true pressure coefficient curve is calculated;
s36, calculating LOSS, and updating network parameters: and (3) according to the difference obtained in the step S35, simultaneously updating the whole network parameters by combining the loss function, and training.
2. The airfoil aerodynamic data computing method based on multitasking learning of claim 1, wherein: in step S1, specifically, the method includes:
S11、 curve extraction airfoil curvature characteristics: use/> The curve represents the two-dimensional coordinates of the airfoil as a plurality of polynomials, and the bending degree, namely the curvature of the curve is calculated from the tangent space of each polynomial;
s12, extracting airfoil geometric features by a convolutional neural network CNN: converting the two-dimensional coordinates of the wing profile, the wing profile picture and the curvature obtained in S11 into a matrix data form to be used as the input of CNN;
s13, extracting latent vector of an airfoil by an AE network: by encoder-decoder architecture, convolutional and deconvolution layers are constructed to extract latent vector of airfoil geometry in a manner that reconstructs the airfoil.
3. The airfoil aerodynamic data computing method based on multitasking learning of claim 2, wherein: in step S4, specifically, the method includes:
S41, connecting output of a first task and output of a second task through a cross stitch network: linearly connecting the vector of the task I passing through the full connection layer and the vector of the task II passing through the convolution neural network, and calculating the mutual influence coefficient of the vector and the vector;
S42, outputting the aerodynamic coefficient of the airfoil, calculating LOSS, and updating the network: through the learning of the cross stitch network, the aerodynamic coefficient of the airfoil is output, and the network is updated using the L1 LOSS.
4. An airfoil pneumatic data computing system based on multitasking learning, characterized in that: the system adopts the airfoil pneumatic data calculation method based on the multi-task learning as claimed in any one of claims 1 to 3.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311506032.4A CN117540489B (en) | 2023-11-13 | Airfoil pneumatic data calculation method and system based on multitask learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311506032.4A CN117540489B (en) | 2023-11-13 | Airfoil pneumatic data calculation method and system based on multitask learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117540489A CN117540489A (en) | 2024-02-09 |
| CN117540489B true CN117540489B (en) | 2024-06-21 |
Family
ID=
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112668696A (en) * | 2020-12-25 | 2021-04-16 | 杭州中科先进技术研究院有限公司 | Unmanned aerial vehicle power grid inspection method and system based on embedded deep learning |
| CN113160375A (en) * | 2021-05-26 | 2021-07-23 | 郑健青 | Three-dimensional reconstruction and camera pose estimation method based on multi-task learning algorithm |
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112668696A (en) * | 2020-12-25 | 2021-04-16 | 杭州中科先进技术研究院有限公司 | Unmanned aerial vehicle power grid inspection method and system based on embedded deep learning |
| CN113160375A (en) * | 2021-05-26 | 2021-07-23 | 郑健青 | Three-dimensional reconstruction and camera pose estimation method based on multi-task learning algorithm |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110490946B (en) | Text image generation method based on cross-modal similarity and antagonism network generation | |
| Mahajan et al. | Categorical image classification based on representational deep network (RESNET) | |
| CN111127146B (en) | Information recommendation method and system based on convolutional neural network and noise reduction self-encoder | |
| CN108805157B (en) | Remote sensing image classification method based on partial random supervision discrete hash | |
| Yang et al. | Extracting and composing robust features with broad learning system | |
| Wu et al. | Optimized deep learning framework for water distribution data-driven modeling | |
| CN114022372B (en) | Mask image patching method for introducing semantic loss context encoder | |
| CN112734913B (en) | Three-dimensional model sphere expression calculation method based on multi-stage deformation reconstruction | |
| CN114493014A (en) | Multivariate time series prediction method, multivariate time series prediction system, computer product and storage medium | |
| CN115661550A (en) | Graph data class imbalance classification method and device based on generation countermeasure network | |
| Zeng et al. | Dual swin-transformer based mutual interactive network for RGB-D salient object detection | |
| Shi et al. | EBNAS: Efficient binary network design for image classification via neural architecture search | |
| Xue et al. | Self-distribution binary neural networks | |
| Hu et al. | M-gcn: Multi-scale graph convolutional network for 3d point cloud classification | |
| Zheng et al. | Multi-class indoor semantic segmentation with deep structured model | |
| CN117540489B (en) | Airfoil pneumatic data calculation method and system based on multitask learning | |
| Guo et al. | Efficient convolutional networks learning through irregular convolutional kernels | |
| CN117540489A (en) | Airfoil pneumatic data calculation method and system based on multitask learning | |
| Nista et al. | The influence of adversarial training on turbulence closure modeling | |
| CN112487231B (en) | Automatic image labeling method based on double-image regularization constraint and dictionary learning | |
| CN113065336B (en) | Text automatic generation method and device based on deep learning and content planning | |
| Hu et al. | IMMAT: Mesh reconstruction from single view images by medial axis transform prediction | |
| CN115438575A (en) | Analysis method for high-precision airfoil flow field prediction | |
| Ahn et al. | Multi-branch neural architecture search for lightweight image super-resolution | |
| Nandal et al. | A Synergistic Framework Leveraging Autoencoders and Generative Adversarial Networks for the Synthesis of Computational Fluid Dynamics Results in Aerofoil Aerodynamics |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant |