CN117150680A - Airfoil profile optimization design method based on deep learning and reinforcement learning - Google Patents
Airfoil profile optimization design method based on deep learning and reinforcement learning Download PDFInfo
- Publication number
- CN117150680A CN117150680A CN202311196497.4A CN202311196497A CN117150680A CN 117150680 A CN117150680 A CN 117150680A CN 202311196497 A CN202311196497 A CN 202311196497A CN 117150680 A CN117150680 A CN 117150680A
- Authority
- CN
- China
- Prior art keywords
- airfoil
- design
- lift
- reinforcement learning
- parameters
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000013461 design Methods 0.000 title claims abstract description 92
- 238000005457 optimization Methods 0.000 title claims abstract description 60
- 238000000034 method Methods 0.000 title claims abstract description 47
- 230000002787 reinforcement Effects 0.000 title claims abstract description 39
- 238000013135 deep learning Methods 0.000 title claims abstract description 26
- 238000012549 training Methods 0.000 claims abstract description 33
- 230000009471 action Effects 0.000 claims description 27
- 230000008569 process Effects 0.000 claims description 15
- 239000012530 fluid Substances 0.000 claims description 14
- 230000009467 reduction Effects 0.000 claims description 12
- 238000004364 calculation method Methods 0.000 claims description 6
- 238000013527 convolutional neural network Methods 0.000 claims description 6
- 230000008859 change Effects 0.000 claims description 3
- 230000036244 malformation Effects 0.000 claims description 3
- 230000003993 interaction Effects 0.000 abstract description 3
- 238000012986 modification Methods 0.000 abstract description 2
- 230000004048 modification Effects 0.000 abstract description 2
- 238000004422 calculation algorithm Methods 0.000 description 7
- 239000003795 chemical substances by application Substances 0.000 description 7
- 238000011160 research Methods 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 239000008358 core component Substances 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000012938 design process Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000005284 excitation Effects 0.000 description 1
- 230000002068 genetic effect Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000005728 strengthening Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/17—Mechanical parametric or variational 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
- G06F30/28—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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2113/00—Details relating to the application field
- G06F2113/08—Fluids
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Geometry (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Pure & Applied Mathematics (AREA)
- Software Systems (AREA)
- Mathematical Optimization (AREA)
- Computer Hardware Design (AREA)
- Mathematical Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Fluid Mechanics (AREA)
- Computational Mathematics (AREA)
- Algebra (AREA)
- Structures Of Non-Positive Displacement Pumps (AREA)
Abstract
The invention belongs to the technical field of mechanical optimization design, and discloses an airfoil optimization design method based on deep learning and reinforcement learning, which comprises the steps of constructing an airfoil parameterized design platform based on the deep learning, parameterizing and dimension-reducing a basic airfoil to be optimized into at least four fitting parameters based on the airfoil parameterized design platform, and completing geometric shape modification of the basic airfoil by adjusting the fitting parameters to obtain geometric coordinates of a first airfoil; acquiring an airfoil lift coefficient and a drag coefficient based on an airfoil performance prediction model; based on the wing section optimal design model, interaction with the wing section performance prediction model and the wing section parameterized design platform is realized, a termination condition for reinforcement learning training is set for the wing section optimal design model until a preset termination condition is met, and the optimized geometric coordinates and wing section lift-drag ratio of the first wing section are output.
Description
Technical Field
The invention relates to the technical field of mechanical optimization design, in particular to an airfoil optimization design method based on deep learning and reinforcement learning.
Background
Fluid machinery and equipment such as aircrafts, underwater thrusters, nuclear power station main pumps and the like face high-parameter design requirements. The design of the blade as a core component of such high-end fluid machinery and equipment is related to the performance of the equipment. The airfoil is the basis of blade design, and research on the optimization design of the airfoil has great significance for improving the performance of fluid machinery and equipment.
The optimal design method of the wing profile relates to the parameterized modeling technology of the wing profile and the optimal design method of the wing profile. At present, in the parametric modeling method of the airfoil, design parameters describing the first target are too many, the airfoil optimization design faces the problem of large design space, so that not only can the computational fluid dynamics prediction cost be increased, but also the searching difficulty of an optimization algorithm can be increased. Regarding the optimization design method of the airfoil, the application of some evolutionary algorithms, such as genetic algorithms, to the optimization design research of the airfoil becomes a current research hotspot.
With the development of artificial intelligence, deep learning and reinforcement learning methods are increasingly applied to the field of optimal design of mechanical equipment. Particularly, alpha-go has good performance in the field of go or games, and the application of reinforcement learning in the field of control is one of the most popular research fields in the current machine learning. Therefore, the application research of the deep learning and the reinforcement learning on the aspect of the optimal design of mechanical equipment is imperative, and the problems of too many airfoil parameters and difficult searching of an optimization algorithm can be relieved.
In view of the above, the invention provides an airfoil optimization design method based on deep learning and reinforcement learning.
Disclosure of Invention
In order to overcome the problems in the prior art, the invention provides an airfoil optimal design method based on deep learning and reinforcement learning, which solves the problems of overlarge design space, overlarge calculation cost and large algorithm search space caused by overlarge design parameter dimension in the existing airfoil design. Meanwhile, the problems that the application process is unknown, and the reward value and the training termination condition are difficult to determine due to the fact that the current reinforcement learning technology is applied less in the field of fluid mechanical optimization design are solved, so that the application potential of reinforcement learning is found to a greater extent.
According to one aspect of the present invention, there is provided an airfoil optimization design method based on deep learning and reinforcement learning, comprising the steps of:
selecting a basic airfoil profile to be optimized, and obtaining fitting parameters of the basic airfoil profile through an airfoil profile parameterized design platform based on deep learning: the fitting parameters of the basic wing profile are adjusted to obtain a large number of geometric coordinates of the first wing profile;
constructing an airfoil performance prediction model based on hydrodynamic software Xfoil; based on the geometric coordinates of the first airfoil and preset flow working condition parameters, a lift coefficient and a drag coefficient corresponding to the first airfoil are obtained, and the ratio of the lift coefficient to the drag coefficient is marked as an airfoil lift-drag ratio;
establishing an airfoil optimization design model by adopting an DQN reinforcement learning frame, marking the geometric coordinates of a first airfoil and the airfoil lift-drag ratio as an airfoil optimization state, and setting the rewarding value based on the change of the airfoil lift-drag ratio; the wing profile fitting parameters are used for continuously predicting action strategies based on the first adjustment interval, so that the wing profile fitting parameters reciprocate, and a large number of state-action-rewarding value samples can be obtained; the airfoil optimization design model based on the DQN is learned according to a large number of acquired samples, so that the maximization of the rewarding value is realized;
the network training process is carried out in 2 rounds, each round is preset with a fixed step number, after the fixed step number is reached, the basic wing profile is updated to the optimal wing profile geometry and the fitting parameters thereof obtained in the earlier training process, and the next round of optimization and network training are carried out continuously by taking the optimal wing profile geometry and the fitting parameters as starting points;
and setting a termination condition of reinforcement learning training on the airfoil optimization design model until a preset termination condition is met, and outputting the optimized geometric coordinates of the first airfoil and the airfoil lift-drag ratio.
Preferably, the specific application logic of the airfoil parameterized design platform is:
taking an airfoil image dataset in a public airfoil database as an input value of an airfoil parameterized design platform;
convolutional neural network for extracting airfoil image dataClassifying and compressing the airfoil image data features through a multi-layer perceptron, and describing the dimensionality reduction of the airfoil image data features into N fitting parameters; the fitting parameter is marked as P n ，P n ＝P 0 、P 1 、......、P N-1 N is more than or equal to 4;
selecting a basic airfoil profile to be optimized, and recovering the geometric coordinates of the basic airfoil profile corresponding to the N fitting parameters to obtain the geometric coordinates of the initialized basic airfoil profile;
and adjusting N fitting parameters of the dimension reduction description based on a first adjustment interval, recovering the geometric coordinates of the adjusted basic airfoil, setting constraint conditions for preventing airfoil deformity, marking the airfoil meeting the constraint conditions as a first airfoil, and outputting the geometric coordinates of the first airfoil. .
Preferably, the fitting parameters of the base airfoil are at least four fitting parameters.
Preferably, the specific logic of the constraint condition for preventing airfoil deformity is:
y up ＞y down ；
wherein y is up Is the ordinate value of the coordinate point of the upper surface of the airfoil; y is down Is the ordinate value of the airfoil lower surface coordinate point.
Preferably, the specific application logic of the airfoil performance prediction model is:
taking a first target and flow working condition parameters as input values of an airfoil performance prediction model, wherein the flow working condition parameters comprise airfoil attack angles and Reynolds numbers;
and taking the lift-drag ratio of the airfoil as an output value of the airfoil performance prediction model.
Preferably, the airfoil optimization design model specifically includes the following logic:
the optimization state of the wing profile corresponding to each round is Z= [ P, CL, CD ], wherein the fitting parameters of the first wing profile are P, and CL and CD respectively represent the lift coefficient and the drag coefficient of the first wing profile under the corresponding fitting parameters;
determining the next action by the airfoil optimization design model according to the airfoil optimization state corresponding to Z, so as to obtain a fitting parameter set P_next of the next first airfoil;
wherein: the fitting parameter set P of the first airfoil is an N-dimensional vector, namely: p= [ P ] 0 、P 1 、......、P N-1 ]N is the number of parameters after the dimension reduction of the airfoil parameters; the action is a vector with the same shape as the fitting parameter set P of the first airfoil, and represents the influence exerted by the action on each parameter, namely: action= [ a ] 0 、a 1 、......、a N-1 ]；a 0 -a N-1 Different values can be taken as required, and the fixed step number is uniformly sampled with step=0.001 in the design, namely a 0 -a N-1 The first adjustment interval of each variable is + -0.001; after determining the action, the next round of airfoil optimization state Z m+1 The method comprises the following steps:
Z m+1 ＝[P next ，CL next ，CD next ]
P next ＝P+action
wherein CL is next ，CD next From new parameters P next Calculated by computational fluid dynamics software.
Preferably, the prize value is set according to the optimization objective and various constraints, and the weight of each prize in the whole Reward is controlled by the parameter alpha, and the specific formula is as follows:
wherein, the optimization objective is: maximizing the lift-drag ratio; the constraint conditions are as follows: cl/Cd > Cl0/Cd0; the lift-drag ratio of the airfoil corresponding to the initial state is Cl0/Cd0; the lift-drag ratio of the wing profile corresponding to the current state is Cl (i)/Cd (i); the lift-drag ratio of the airfoil corresponding to the former state is Cl (i-1)/Cd (i-1); alpha 1 And alpha is 2 Is a weight coefficient. .
Preferably, the termination condition of the reinforcement learning training process includes:
if the first wing profile generated in each round has the wing profile malformation, ending the track and returning, and rewarding the value of reward=a;
if the calculation result of the first airfoil profile generated by each round by using the computational fluid dynamics software cannot be converged, ending the track and returning, wherein the reward value is found=b;
if the first wing profile generated in each round does not meet the additional design condition, ending the current track and returning, wherein the reward value is reorder=c;
if the number of steps of the reinforcement learning training exceeds the preset fixed number of steps, terminating the reinforcement learning training.
The wing section optimization design method based on deep learning and reinforcement learning has the technical effects and advantages that:
the invention provides a new thought and solving method for the optimal design of the wing section and even the fluid machinery, and the dimension reduction of the first target design parameter is realized based on a deep learning method, so that the calculation workload and the algorithm search space in the optimal design can be reduced; the real-time optimization design of the wing profile is further realized through the reinforcement learning technology, the real-time interaction between simulation software and actions can be realized, and the total sample size is reduced. Meanwhile, the application of the strengthening technology in the aspect of the optimal design of the fluid machinery belongs to the newer field, and the application process and how to set the rewarding value and the termination condition during the training are all the current difficulties.
Drawings
FIG. 1 is a flow chart of an airfoil optimization design method of the present invention;
FIG. 2 is a network training process of the airfoil optimization design method of the present invention;
FIG. 3 is a graph of network training effects of the airfoil optimization design method of the present invention;
fig. 4 is a window diagram of wing profile parameterized design software based on deep learning construction.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Example 1
Referring to fig. 1-2, the method for optimizing design of an airfoil based on deep learning and reinforcement learning according to the present embodiment includes the following steps:
selecting a basic airfoil profile to be optimized, and obtaining fitting parameters of the basic airfoil profile through an airfoil profile parameterized design platform based on deep learning: the fitting parameters of the basic wing profile are adjusted to obtain a large number of geometric coordinates of the first wing profile;
what needs to be explained here is: after fitting parameters of the basic airfoil are adjusted, design optimization of basic airfoil geometry is realized, and an optimized airfoil geometry image is marked as a first airfoil as shown in fig. 4; in practical application, the fitting parameters of the adjusting basic airfoil have an applicable interval, and the applicable interval is marked as a first adjusting interval.
The specific application logic of the airfoil parameterized design platform is as follows:
taking an airfoil image dataset in a public airfoil database as an input value of an airfoil parameterized design platform;
the method comprises the steps of extracting airfoil image data features by a convolutional neural network, classifying and compressing the airfoil image data features through a multi-layer perceptron, and describing the airfoil image data features into N fitting parameters through parameterization dimension reduction; the fitting parameter is marked as P n ，P n ＝P 0 、P 1 、......、P N-1 N is more than or equal to 4;
selecting a basic airfoil profile to be optimized, and recovering the geometric coordinates of the basic airfoil profile corresponding to the N fitting parameters to obtain the geometric coordinates of the initialized basic airfoil profile;
and adjusting N fitting parameters of the dimension reduction description based on a first adjustment interval, recovering the geometric coordinates of the adjusted basic airfoil, setting constraint conditions for preventing airfoil deformity, marking the airfoil meeting the constraint conditions as a first airfoil, and outputting the geometric coordinates of the first airfoil.
Further, the number of the fitting parameters is preferably 6.
What needs to be explained here is: the airfoil parameterized design comprises 6 fitting parameters for describing the geometrical characteristics of the airfoil, wherein the 6 parameters do not belong to the traditional geometrical characteristics of the airfoil such as chord length, thickness, half span, front edge sweepback angle, thickness distribution and the like, but can accurately express the airfoil geometry.
The six fitting parameters can carry out enough flexibility adjustment on the wing profile in the design process, and meanwhile, the design space is not excessively complicated, so that the efficiency and stability of an optimization algorithm can be maintained; the number of parameters may of course also be adjusted according to the requirements and complexity of the particular problem.
In particular exemplary, the present example uses NACA0010 as the base airfoil to be optimized, scaled down to 6 fitting parameters [0.1625, -0.0610,0.1478,0.0179, -0.2006, -0.1182] by a deep-learning based airfoil parametric design platform.
Further, the airfoil parameterized design platform comprises a deep learning model of a combination of a convolutional neural network and a multi-layer perceptron;
furthermore, the specific application logic of the airfoil parameterized design platform is as follows:
taking an airfoil image dataset in a public airfoil database as an input value of an airfoil parameterized design platform;
what needs to be explained here is: the disclosed airfoil image dataset in the airfoil database is airfoil two-dimensional image data obtained based on conversion of an airfoil coordinate point database disclosed by the university of Ill Baner-Champagne division.
Extracting data features of the airfoil image by using a convolution layer, classifying and compressing the image features by using a multi-layer perceptron, and compressing a first target into 6 fitting parameters;
and 6 fitting parameters described by dimension reduction are adjusted based on a first adjustment interval, the geometric coordinates of the adjusted basic wing profile are recovered, constraint conditions for preventing deformation of the wing profile are set, the wing profile meeting the constraint conditions is marked as a first wing profile, and the geometric coordinates of the first wing profile are output.
What needs to be explained here is: the wing profile parameterized design platform can realize two functions of obtaining 6 dimension reduction parameters of a basic wing profile and adjusting 6 fitting parameters to obtain a large number of geometric coordinates of a first wing profile.
The specific logic of the constraint condition for preventing airfoil profile deformity is as follows:
y up ＞y down ；
wherein y is up Is the ordinate value of the coordinate point of the upper surface of the airfoil; y is down Is the ordinate value of the airfoil lower surface coordinate point.
Constructing an airfoil performance prediction model based on hydrodynamic software Xfoil; based on the geometric coordinates of the first airfoil and preset flow working condition parameters, a lift coefficient and a drag coefficient corresponding to the first airfoil are obtained, and the ratio of the lift coefficient to the drag coefficient is marked as an airfoil lift-drag ratio;
further, the specific application logic of the airfoil performance prediction model is:
taking the geometric coordinates of the first airfoil and preset flow working condition parameters as input values of an airfoil performance prediction model, wherein the flow working condition parameters comprise attack angles and Reynolds numbers; illustratively, the airfoil angle of attack is selected here to be 3 ° and the reynolds number is 30000; the optimum design is iterated for about 600 rounds in total, and the effect is shown in fig. 3. In fig. 3, where the blue curve represents the airfoil with the greatest lift-to-drag ratio in each pass, the orange curve is the average of all the greatest lift-to-drag ratios for the past 100 passes, and passes less than 100 are not calculated. Through the optimal design, the lift-drag ratio 93 of the final NACA0010 airfoil under the working conditions of 30000 Reynolds number and 3 DEG attack angle is improved to 102, and the improvement ratio is as follows: 9.68%.
And taking the lift-drag ratio of the airfoil as an output value of the airfoil performance prediction model.
Establishing an airfoil optimization design model by adopting an DQN reinforcement learning frame, marking the geometric coordinates of a first airfoil and the airfoil lift-drag ratio as an airfoil optimization state, and setting the rewarding value based on the change of the airfoil lift-drag ratio; the wing profile fitting parameters are used for continuously predicting action strategies based on the first adjustment interval, so that the wing profile fitting parameters reciprocate, and a large number of state-action-rewarding value samples can be obtained; the airfoil optimization design model based on the DQN is learned according to a large number of acquired samples, so that the maximization of the rewarding value is realized;
further, the airfoil optimization design model specifically includes the following logic:
the optimization state of the wing profile corresponding to each round is Z= [ P, CL, CD ], wherein the fitting parameters of the first wing profile are P, and CL and CD respectively represent the lift coefficient and the drag coefficient of the first wing profile under the corresponding fitting parameters;
determining a next action by the airfoil optimization design model according to the airfoil optimization state corresponding to Z, so as to obtain a fitting parameter set P_next of a next first airfoil, as shown in FIG. 3;
wherein: the fitting parameter set P of the first airfoil is an N-dimensional vector, namely: p= [ P ] 0 、P 1 、......、P N-1 ]N is the number of parameters after the dimension reduction of the airfoil parameters; the action is a vector with the same shape as the fitting parameter set P of the first airfoil, and represents the influence exerted by the action on each parameter, namely: action= [ a ] 0 、a 1 、......、a N-1 ]；a 0 -a N-1 Different values can be taken as required, and the fixed step number is uniformly sampled with step=0.001 in the design, namely a 0 -a N-1 The first adjustment interval of each variable is + -0.001; after determining the action, the next round of airfoil optimization state Z m+1 The method comprises the following steps:
Z m+1 ＝[P next ，CL next ，CD next ]
P next ＝P+action
wherein CL is next ，CD next From new parameters P next Calculated by computational fluid dynamics software.
Specific examples: the specific application logic of the airfoil optimization design model is as follows:
the wing section optimal design model is a reinforced learning wing section optimal design framework based on DQN, the DQN is an agent built by a convolutional neural network, and the actions of adjusting 6 fitting parameters and the interactions with computational fluid dynamics software xfoil are realized based on the agent;
when training the neural network, the intelligent agent selects the next action according to the corresponding rewarding value of the wing-shaped lift-drag ratio output by the wing-shaped performance prediction model, and interacts with the wing-shaped performance prediction model again to give the next wing-shaped lift-drag ratio after the action and rewards corresponding to the wing-shaped lift-drag ratio;
in this way, a large number of airfoil fitting parameter-action-rewarding value samples are generated in the training process, the neural network is trained by using the samples, network super-parameters are updated continuously, estimation accuracy of the network to rewards is improved, and training is finished when the training meets the termination condition.
What needs to be explained here is: the convolutional neural network is used as an agent and is responsible for selecting actions for adjusting six fitting parameters of the airfoil.
The training of the agent is based on rewards, and the rewards function is generally based on the airfoil lift-drag ratio output by the airfoil performance prediction model. In each training step, the intelligent agent interacts with the wing profile performance prediction model according to the current wing profile parameter configuration, obtains the lift-drag ratio and takes the lift-drag ratio as a rewarding value. This prize value is used to instruct the agent to select the next action.
The prize value is set according to the optimization target and various constraint conditions, and the weight of each prize in the whole Reward is controlled by the parameter alpha, wherein the specific formula is as follows:
wherein, the optimization objective is: maximizing the lift-drag ratio; the constraint conditions are as follows: cl/Cd is more than Cl0/CdO; the lift-drag ratio of the wing profile corresponding to the initial state is ClO/Cd0; the lift-drag ratio of the wing profile corresponding to the current state is Cl (i)/Cd (i); the lift-drag ratio of the airfoil corresponding to the former state is Cl (i-1)/Cd (i-1); alpha 1 And alpha is 2 As a weight coefficient, specifically exemplified, alpha 1 And alpha is 2 Preferably 0.5, respectively, this value and the specific prizeThe excitation value sub-term can be changed according to the application scene, the optimization target and the constraint condition of the wing profile to be optimized.
And obtaining a large number of airfoil geometric parameter-action-rewarding value samples through an airfoil optimization design model, and continuously adjusting airfoil parameters to obtain new airfoil lift-drag ratio and rewarding value. These experience samples are accumulated and used to train the reinforcement learning network to evaluate the difference between the actual rewards and the estimated rewards. The network weights are updated continuously during the training process to improve the accuracy of estimating the rewards.
The network training process is carried out in 2 rounds, each round is preset with a fixed step number, after the fixed step number is reached, the basic wing profile is updated to the optimal wing profile geometry and the fitting parameters thereof obtained in the earlier training process, and the next round of optimization and network training are carried out continuously by taking the optimal wing profile geometry and the fitting parameters as starting points; .
Specific exemplary, the termination conditions for the reinforcement learning training process include:
if the first wing profile generated in each round has the wing profile malformation condition, ending the current track and returning, and rewarding the value of reward= -100;
if the calculation result of the first airfoil profile generated by each round can not be converged by using the computational fluid dynamics software, ending the current track and returning, and then rewarding the value of the reward= -100;
if the first wing profile generated in each round does not meet the additional design conditions, such as the maximum thickness of the wing profile, the moment coefficient of the wing profile and the like, ending the track and returning, and then rewarding the value of reward= -1;
if the number of steps of the reinforcement learning training exceeds the preset fixed number of steps, terminating the reinforcement learning training.
What needs to be explained here is: such termination conditions may avoid wasting time on improper airfoil configuration and learn an effective optimization strategy faster. Meanwhile, the device can also process the situation that the pneumatic parameters cannot be calculated so as to exit and enter the next round of training in time when a problem occurs. These measures help to improve the efficiency and stability of the model.
The above formulas are all formulas with dimensionality removed and numerical calculation, the formulas are formulas with the latest real situation obtained by software simulation through collecting a large amount of data, and preset parameters and threshold selection in the formulas are set by those skilled in the art according to the actual situation.
The foregoing is merely illustrative of the present invention, and the present invention is not limited thereto, and any person skilled in the art will readily recognize that variations or substitutions are within the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Finally: the foregoing description of the preferred embodiments of the invention is not intended to limit the invention to the precise form disclosed, and any such modifications, equivalents, and alternatives falling within the spirit and principles of the invention are intended to be included within the scope of the invention.
Claims (8)
1. The wing section optimal design method based on deep learning and reinforcement learning is characterized by comprising the following steps of:
selecting a basic airfoil profile to be optimized, and obtaining fitting parameters of the basic airfoil profile through an airfoil profile parameterized design platform based on deep learning: the fitting parameters of the basic wing profile are adjusted to obtain a large number of geometric coordinates of the first wing profile;
constructing an airfoil performance prediction model based on hydrodynamic software Xfoil; based on the geometric coordinates of the first airfoil and preset flow working condition parameters, a lift coefficient and a drag coefficient corresponding to the first airfoil are obtained, and the ratio of the lift coefficient to the drag coefficient is marked as an airfoil lift-drag ratio;
establishing an airfoil optimization design model by adopting a DQN reinforcement learning frame, and marking the geometric coordinates of a first airfoil and the airfoil lift-drag ratio as an airfoil optimization state; and setting the prize value based on a change in airfoil lift-drag ratio; the wing profile fitting parameters are used for continuously predicting action strategies based on the first adjustment interval, so that the wing profile fitting parameters reciprocate, and a large number of state-action-rewarding value samples can be obtained; the airfoil optimization design model based on the DQN is learned according to a large number of acquired samples, so that the maximization of the rewarding value is realized;
the network training process is carried out in 2 rounds, each round is preset with a fixed step number, after the fixed step number is reached, the basic wing profile is updated to the optimal wing profile geometry and the fitting parameters thereof obtained in the earlier training process, and the next round of optimization and network training are carried out continuously by taking the optimal wing profile geometry and the fitting parameters as starting points;
and setting a termination condition of reinforcement learning training on the airfoil optimization design model until a preset termination condition is met, and outputting the optimized geometric coordinates of the first airfoil and the airfoil lift-drag ratio.
2. The method for optimizing design of an airfoil based on deep learning and reinforcement learning according to claim 1, wherein the specific application logic of the airfoil parameterized design platform is as follows:
taking an airfoil image dataset in a public airfoil database as an input value of an airfoil parameterized design platform;
the method comprises the steps of extracting airfoil image data features by a convolutional neural network, classifying and compressing the airfoil image data features through a multi-layer perceptron, and describing the airfoil image data feature dimension reduction as N fitting parameters; the fitting parameter is marked as P n ，P n ＝P 0 、P 1 、……、P N-1 N is more than or equal to 4;
selecting a basic airfoil profile to be optimized, and recovering the geometric coordinates of the basic airfoil profile corresponding to the N fitting parameters to obtain the geometric coordinates of the initialized basic airfoil profile;
and adjusting N fitting parameters of the dimension reduction description based on a first adjustment interval, recovering the geometric coordinates of the adjusted basic airfoil, setting constraint conditions for preventing airfoil deformity, marking the airfoil meeting the constraint conditions as a first airfoil, and outputting the geometric coordinates of the first airfoil.
3. The method for optimizing design of an airfoil based on deep learning and reinforcement learning according to claim 2, wherein the fitting parameters of the basic airfoil are at least four fitting parameters.
4. The method for optimizing design of airfoil based on deep learning and reinforcement learning according to claim 2, wherein the specific logic of the constraint condition for preventing airfoil deformity is:
y up ＞y down ；
wherein y is up Is the ordinate value of the coordinate point of the upper surface of the airfoil; y is down Is the ordinate value of the airfoil lower surface coordinate point.
5. The method for optimizing design of an airfoil based on deep learning and reinforcement learning according to claim 4, wherein the specific application logic of the airfoil performance prediction model is as follows:
taking a first target and flow working condition parameters as input values of an airfoil performance prediction model, wherein the flow working condition parameters comprise airfoil attack angles and Reynolds numbers;
and taking the lift-drag ratio of the airfoil as an output value of the airfoil performance prediction model.
6. The method for optimizing design of an airfoil based on deep learning and reinforcement learning according to claim 5, wherein the airfoil optimizing design model specifically comprises the following logic:
the optimization state of the wing profile corresponding to each round is Z= [ P, CL, CD ], wherein the fitting parameters of the first wing profile are P, CL and CD respectively represent the lift coefficient and the drag coefficient of the first wing profile under the corresponding fitting parameters;
determining the next action by the airfoil optimization design model according to the airfoil optimization state corresponding to Z, so as to obtain a fitting parameter set P_next of the next first airfoil;
wherein: the fitting parameter set P of the first airfoil is an N-dimensional vector, namely: p= [ P ] 0 、P 1 、......、P N-1 ]N is the number of parameters after the dimension reduction of the airfoil parameters; the action is a vector with the same shape as the fitting parameter set P of the first airfoil, and represents the influence exerted by the action on each parameter, namely: action= [ a ] 0 、a 1 、......、a N-1 ]；a 0 -a N-1 Different values can be taken as required, and the fixed step number is uniformly sampled with step=0.001 in the design, namely a 0 -a N-1 The first adjustment interval of each variable is + -0.001; after determining the action, the next round of airfoil optimization state Z m+1 The method comprises the following steps:
Z m+1 ＝[P next ，CL next ，CD next ]
P next ＝P+action
wherein CL is next ，CD next From new parameters P next Calculated by computational fluid dynamics software.
7. The method for optimizing design of airfoil based on deep learning and reinforcement learning according to claim 6, wherein the prize value is set according to the optimization objective and various constraints, and the weight of each prize in the whole report is controlled by the parameter α, and the specific formula is:
wherein, the optimization objective is: maximizing the lift-drag ratio; the constraint conditions are as follows: CI/Cd>CI0/Cd0; the lift-drag ratio of the airfoil corresponding to the initial state is Cl0/Cd0; the lift-drag ratio of the wing profile corresponding to the current state is Cl (i)/Cd (i); the lift-drag ratio of the airfoil corresponding to the former state is CI (i-1)/Cd (i-1); alpha 1 And alpha is 2 Is a weight coefficient.
8. The method for optimizing design of an airfoil based on deep learning and reinforcement learning according to claim 7, wherein the termination condition of the reinforcement learning training process comprises:
if the first wing profile generated in each round has wing profile malformation, ending the track and returning, and rewarding the value of reward=a;
if the calculation result of the first airfoil profile generated by each round by using the computational fluid dynamics software cannot be converged, ending the track and returning, wherein the reward value is found=b;
if the first wing profile generated in each round does not meet the additional design condition, ending the current track and returning, wherein the reward value is reorder=c;
if the number of steps of the reinforcement learning training exceeds the preset fixed number of steps, terminating the reinforcement learning training.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311196497.4A CN117150680A (en) | 2023-09-15 | 2023-09-15 | Airfoil profile optimization design method based on deep learning and reinforcement learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311196497.4A CN117150680A (en) | 2023-09-15 | 2023-09-15 | Airfoil profile optimization design method based on deep learning and reinforcement learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117150680A true CN117150680A (en) | 2023-12-01 |
Family
ID=88906016
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311196497.4A Pending CN117150680A (en) | 2023-09-15 | 2023-09-15 | Airfoil profile optimization design method based on deep learning and reinforcement learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117150680A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117350178A (en) * | 2023-12-05 | 2024-01-05 | 深圳十沣科技有限公司 | Airfoil lift resistance prediction method, apparatus, device and storage medium |
-
2023
- 2023-09-15 CN CN202311196497.4A patent/CN117150680A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117350178A (en) * | 2023-12-05 | 2024-01-05 | 深圳十沣科技有限公司 | Airfoil lift resistance prediction method, apparatus, device and storage medium |
| CN117350178B (en) * | 2023-12-05 | 2024-04-02 | 深圳十沣科技有限公司 | Airfoil lift resistance prediction method, apparatus, device and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106600059B (en) | Intelligent power grid short-term load prediction method based on improved RBF neural network | |
| CN111104522A (en) | Regional industry association effect trend prediction method based on knowledge graph | |
| CN110751318A (en) | IPSO-LSTM-based ultra-short-term power load prediction method | |
| CN117150680A (en) | Airfoil profile optimization design method based on deep learning and reinforcement learning | |
| CN109492748B (en) | Method for establishing medium-and-long-term load prediction model of power system based on convolutional neural network | |
| CN111079343B (en) | Wind turbine generator effective wind speed estimation method based on width learning | |
| CN112784140B (en) | Search method of high-energy-efficiency neural network architecture | |
| CN112633584B (en) | River sudden water pollution accident water quality prediction method based on improved LSTM-seq2seq model | |
| CN112183742B (en) | Neural network hybrid quantization method based on progressive quantization and Hessian information | |
| CN116503676B (en) | Picture classification method and system based on knowledge distillation small sample increment learning | |
| CN111355633A (en) | Mobile phone internet traffic prediction method in competition venue based on PSO-DELM algorithm | |
| CN111126549A (en) | Double-star spectrum fitting method based on strategy improved goblet and sea squirt intelligent algorithm | |
| CN110837915A (en) | Low-voltage load point prediction and probability prediction method for power system based on hybrid integrated deep learning | |
| CN112766603A (en) | Traffic flow prediction method, system, computer device and storage medium | |
| CN107578101B (en) | Data stream load prediction method | |
| CN110930514B (en) | Three-dimensional grid model self-adaptive simplification method based on local region characteristics | |
| CN116432697A (en) | Time sequence prediction method integrating long-term memory network and attention mechanism | |
| CN117290721A (en) | Digital twin modeling method, device, equipment and medium | |
| CN109840335A (en) | Based on the radial forging pit prediction optimization method for strengthening T-S fuzzy neural network | |
| CN112149896A (en) | Attention mechanism-based mechanical equipment multi-working-condition fault prediction method | |
| CN117554818A (en) | Method for predicting residual service life of lithium ion battery | |
| CN117034060A (en) | AE-RCNN-based flood classification intelligent forecasting method | |
| CN108734116B (en) | Face recognition method based on variable speed learning deep self-coding network | |
| CN116663745A (en) | LSTM drainage basin water flow prediction method based on PCA_DWT | |
| CN116523877A (en) | Brain MRI image tumor block segmentation method based on convolutional neural network |
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 |