An active flow control method and system based on intelligent algorithms
By constructing a proxy model for the air intake flow field characteristics using an intelligent algorithm framework based on surrogate model-transfer learning-reinforcement learning, the problem of real-time control of the air intake of wide-speed-range aircraft is solved, thereby improving the aerodynamic efficiency and safety of the aircraft.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING AEROSPACE TECH INST
- Filing Date
- 2025-12-30
- Publication Date
- 2026-05-26
AI Technical Summary
Existing active flow control technologies cannot achieve real-time response to dynamic changes in the flow field in the air intake of wide-speed-range aircraft. Traditional control methods have poor generalization ability and are difficult to achieve precise control under Mach number and variable atmospheric conditions.
A three-level intelligent algorithm framework based on surrogate model, transfer learning, and reinforcement learning is adopted. A surrogate model of the air intake flow field characteristics is constructed through deep learning. Combined with transfer learning and reinforcement learning, real-time optimal control in a wide speed range is achieved.
It achieves real-time optimal control of the airflow field in the air intake of a wide-speed-range aircraft, improving aerodynamic efficiency and flight safety, and solving the problem of poor generalization ability of traditional methods across operating conditions.
Smart Images
Figure CN122085656A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of aircraft aerodynamics and intelligent control, specifically relating to an active flow control method and system based on intelligent algorithms. Background Technology
[0002] The performance of the air intake of a wide-speed-range aircraft is a core factor affecting its flight efficiency and mission reliability. Taking the throat's influence on the air intake as an example, as a key area for flow field compression and airflow regulation within the air intake, the flow field state directly determines the air intake's core performance parameters such as the total pressure recovery coefficient and pressure ratio. The throat's contraction ratio, a key parameter reflecting the throat's flow field stability and compression efficiency, varies drastically with external conditions such as flight Mach number, angle of attack, and atmospheric environment. At low Mach numbers, an excessively high contraction ratio can easily cause airflow obstruction, preventing the air intake from starting. At high Mach numbers, an insufficient contraction ratio may lead to an excessively high Mach number at the air intake exit, resulting in insufficient pressure and reduced resistance to back pressure.
[0003] In existing active flow control technologies, the regulation of the air intake mostly adopts an open-loop control mode, that is, the deformation of the air intake is fixed according to a preset flight profile, which cannot respond to dynamic changes in the flow field in real time. Some closed-loop control methods rely on traditional PID algorithms, but are limited by linear control logic and have difficulty dealing with strongly nonlinear physical phenomena in the flow field such as shock wave-boundary layer interference and flow separation. At the same time, traditional control methods require a large amount of pre-simulation data to build control models, which is not only time-consuming and labor-intensive, but also has poor generalization ability in complex scenarios such as Mach number transitions and varying atmospheric conditions, and cannot achieve precise and real-time regulation of the air intake flow field characteristics. Summary of the Invention
[0004] The purpose of this invention is to provide an active flow control method and system based on intelligent algorithms. Through a three-level intelligent algorithm framework of "surrogate model-transfer learning-reinforcement learning", it can stably adapt to flight scenarios with wide speed range and large airspace, realize real-time optimal control of the air intake flow field under wide speed range, and solve the problem of poor generalization ability of traditional methods under cross-operating conditions.
[0005] The technical solution adopted in this invention is as follows:
[0006] This invention provides an active flow control method based on intelligent algorithms, comprising the following steps:
[0007] For a typical flight state point, with simulation economy and sampling efficiency as constraints, the number of intake parameters used for numerical simulation of intake flow field characteristics is determined by sampling method.
[0008] Numerical simulation of the flow field is performed according to the number of sets of the inlet parameters to obtain dataset 1 with inlet parameters as independent variables and flow field characteristics as dependent variables;
[0009] For a certain set of inlet parameters, a data set 2 with flight state as the independent variable and flow field characteristics as the dependent variable was obtained through flow field numerical simulation.
[0010] Based on dataset 1, a proxy model of the air intake flow field features at typical flight state points is constructed using deep learning methods.
[0011] Based on the inlet flow field feature proxy model at the flight state point, and using the transfer learning algorithm and dataset 2, an inlet flow field feature proxy model under the wide speed domain flight state is constructed.
[0012] An environment-agent interaction model is constructed based on a deep reinforcement learning algorithm. The inlet flow field feature proxy model under the wide speed domain flight state is used as the environment model for deep reinforcement learning. The state of the environment model is the flight state and flow field features, and the action of the environment model is the inlet parameters.
[0013] Based on the environment-agent interaction model, optimization termination conditions are set, and through deep reinforcement learning, a control sequence that continuously optimizes the inlet performance under different flight conditions is obtained.
[0014] Furthermore, the inlet parameters include at least the forebody deformation and the inlet deformation; the flow field characteristics include at least the inlet flow coefficient, the total pressure recovery coefficient, and the pressure ratio; and the flight state includes at least the flight motion pressure and the flight angle of attack.
[0015] Furthermore, when determining the number of groups of the air intake parameters, multiple typical flight state points are set, and the number of groups of the air intake parameters is determined under different typical flight state points.
[0016] Furthermore, the step of constructing a proxy model of the inlet flow field feature for typical flight state points using deep learning based on dataset 1 specifically involves: constructing a proxy model of the inlet flow field feature for each dependent variable for typical flight state points using deep learning based on all independent variables in dataset 1.
[0017] Furthermore, the sampling method is one of the following: Latin hypercube sampling method, quasi-random sequence sampling method, and adaptive sampling method.
[0018] Furthermore, the deep learning method can be determined by combining the characteristics of dataset 1, and can be one of the following methods: convolutional neural network, recurrent neural network, long short-term memory network, etc.
[0019] Furthermore, the deep reinforcement learning algorithm is one of PPO, DDPG, and TD3.
[0020] Furthermore, the reward function of the environment-agent interaction model is:
[0021]
[0022] Where Φ is the intake flow coefficient and σ is the total pressure recovery coefficient. This refers to the boost ratio.
[0023] This invention also provides an active flow control system based on intelligent algorithms, comprising:
[0024] The sampling and processing module is used to determine the number of intake parameters for numerical simulation of intake flow field characteristics based on typical flight state points and under the constraints of simulation economy and sampling efficiency.
[0025] The dataset module is used to perform flow field numerical simulation according to the number of inlet parameters determined by the sampling and processing module, and to construct dataset 1 with inlet parameters as independent variables and flow field characteristics as dependent variables; and to construct dataset 2 with flight state as independent variables and flow field characteristics as dependent variables based on a certain set of inlet parameters through flow field numerical simulation.
[0026] An initial proxy model is used to construct a proxy model of the air intake flow field characteristics at typical flight state points based on dataset 1 using deep learning methods;
[0027] A wide-speed-domain proxy model is used to construct a proxy model of the air intake flow field characteristics under wide-speed-domain flight conditions based on the transfer learning algorithm and dataset 2.
[0028] An environment-agent interaction model is used to construct an environment-agent interaction model based on a deep reinforcement learning algorithm. The inlet flow field feature proxy model under the wide speed domain flight state is used as the environment model for deep reinforcement learning. The state of the environment model is the flight state and flow field features, and the action of the environment model is the inlet parameters.
[0029] The air intake control module is used to set optimization termination conditions based on the environment-agent interaction model, and optimize parameters through deep reinforcement learning self-learning and self-training to obtain the control sequence that continuously optimizes the air intake performance under different flight conditions.
[0030] The beneficial effects of this invention compared to the prior art are as follows:
[0031] This invention provides an active flow control method based on intelligent algorithms, which uses flight state and flow field characteristics as control state variables (including but not limited to parameters such as flight pressure, flight angle of attack, inlet flow coefficient, total pressure recovery coefficient, and pressure ratio), and inlet parameters as control variables (including but not limited to parameters such as forebody deformation and inlet deformation). It is suitable for real-time flow field regulation of inlet inlets of wide-speed-range aircraft, and can especially solve the problem that inlet performance is difficult to maintain at the optimal level under different flight states, thereby improving the aerodynamic efficiency and flight safety of aircraft over a wide Mach number range.
[0032] 1. This invention replaces the real CFD environment with a proxy model, which greatly improves the interaction speed between the agent and the environment; at the same time, the experience replay mechanism of reinforcement learning can make efficient use of historical data and reduce the number of repeated interactions. Combined with transfer learning, it further reduces the amount of training data in complex scenarios and realizes rapid policy adaptation in wide speed domain scenarios.
[0033] 2. This invention, through a three-level intelligent algorithm framework of "surrogate model - transfer learning - reinforcement learning", can stably adapt to flight scenarios with wide speed range and large airspace, solving the problem of poor generalization ability of traditional methods under cross-operating conditions. Attached Figure Description
[0034] The accompanying drawings, which form part of this specification, are provided to further illustrate embodiments of the invention and, together with the textual description, explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.
[0035] Figure 1 A flowchart of an active flow control method based on intelligent algorithms provided for a specific embodiment of the present invention;
[0036] Figure 2 A schematic diagram of an active flow control environment-agent interaction model based on deep reinforcement learning provided in a specific embodiment of the present invention. Detailed Implementation
[0037] Specific embodiments of the present invention will now be described in detail. In the following description, specific details are set forth for purposes of explanation and not limitation, in order to aid in a thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention may be practiced in other embodiments departing from these specific details.
[0038] It should be noted that, in order to avoid obscuring the invention with unnecessary details, only the device structure and / or processing steps closely related to the solution of the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.
[0039] As one aspect of the present invention, an active flow control method based on intelligent algorithms is provided, comprising the following steps:
[0040] For a typical flight state point, with simulation economy and sampling efficiency as constraints, the number of intake parameters used for numerical simulation of intake flow field characteristics is determined by sampling method.
[0041] Numerical simulation of the flow field is performed according to the number of sets of the inlet parameters to obtain dataset 1 with inlet parameters as independent variables and flow field characteristics as dependent variables;
[0042] For a certain set of inlet parameters, a data set 2 with flight state as the independent variable and flow field characteristics as the dependent variable was obtained through flow field numerical simulation.
[0043] Based on dataset 1, a proxy model of the air intake flow field features at typical flight state points is constructed using deep learning methods.
[0044] Based on the surrogate model of the inlet flow field characteristics at typical flight states, and using the transfer learning algorithm, a surrogate model of the inlet flow field characteristics under wide speed domain flight states is constructed based on dataset 2.
[0045] An environment-agent interaction model is constructed based on a deep reinforcement learning algorithm. The inlet flow field feature proxy model under the wide speed domain flight state is used as the environment model for deep reinforcement learning. The state of the environment model is the flight state and flow field features, and the action of the environment model is the inlet parameters.
[0046] Based on the environment-agent interaction model, optimization termination conditions are set, and through deep reinforcement learning, a control sequence that continuously optimizes the inlet performance under different flight conditions is obtained.
[0047] Furthermore, the inlet parameters include at least the forebody deformation and the inlet deformation; the flow field characteristics include at least the inlet flow coefficient, the total pressure recovery coefficient, and the pressure ratio; and the flight state includes at least the flight pressure and the flight angle of attack. In practical applications, the inlet parameters, flow field characteristics, and flight state parameters are adjusted as needed.
[0048] Furthermore, when determining the number of sets of the inlet parameters, multiple typical flight state points are set, and the number of sets of the inlet parameters is determined separately for each typical flight state point. This method increases the number of sets of inlet parameters used for numerical simulation of the inlet flow field characteristics, thereby increasing the amount of data in Data Set 1 and making the construction of the inlet flow field characteristic proxy model for typical flight state points more accurate.
[0049] Furthermore, based on dataset 1, the method of constructing a surrogate model of the inlet flow field characteristics at typical flight state points using deep learning specifically involves: based on all independent variables in dataset 1, constructing a surrogate model of the inlet flow field characteristics at typical flight state points for each dependent variable using deep learning; subsequently, constructing a surrogate model of the inlet flow field characteristics under wide-speed-domain flight states using a transfer learning algorithm. This approach improves the accuracy of the surrogate model.
[0050] Furthermore, the reward function R of the environment-agent interaction model needs to be analyzed specifically based on the specific implementation case. This invention provides a method for constructing the reward function R, as follows:
[0051] The magnitude of the reward function R is determined by the characteristic quantities of the current state (determining the inlet parameters and the flight state). The flow field characteristic parameters include the inlet flow coefficient Φ, the total pressure recovery coefficient σ, and the pressure ratio p / p. ∞ Intake duct activation is a prerequisite for evaluating intake duct performance. The intake duct flow coefficient can characterize whether the intake duct is activated. Simply put, the 50% flow coefficient can be used as the dividing line between intake duct activation and non-activation. Furthermore, using the tanh function to reflect the impact of intake duct activation on intake duct performance, the reward function R(Φ) can be obtained as follows:
[0052]
[0053] Based on intake manifold start-up, the total pressure recovery coefficient σ and the boost ratio p / p ∞ The larger the value, the better the intake performance; therefore, the reward function R is designed as follows:
[0054]
[0055] In another aspect, the present invention also provides an active flow control system based on intelligent algorithms, comprising:
[0056] The sampling and processing module is used to determine the number of intake parameters for numerical simulation of intake flow field characteristics based on typical flight state points and under the constraints of simulation economy and sampling efficiency.
[0057] The dataset module is used to perform flow field numerical simulation according to the number of inlet parameters determined by the sampling and processing module, and to construct dataset 1 with inlet parameters as independent variables and flow field characteristics as dependent variables; and to construct dataset 2 with flight state as independent variables and flow field characteristics as dependent variables based on a certain set of inlet parameters through flow field numerical simulation.
[0058] An initial proxy model is used to construct a proxy model of the air intake flow field characteristics at typical flight state points based on dataset 1 using deep learning methods;
[0059] A wide-speed-domain proxy model is used to construct a proxy model of the air intake flow field characteristics under wide-speed-domain flight conditions based on the transfer learning algorithm and dataset 2.
[0060] An environment-agent interaction model is used to construct an environment-agent interaction model based on a deep reinforcement learning algorithm. The inlet flow field feature proxy model under the wide speed domain flight state is used as the environment model for deep reinforcement learning. The state of the environment model is the flight state and flow field features, and the action of the environment model is the inlet parameters.
[0061] The air intake control module is used to set optimization termination conditions based on the environment-agent interaction model, and optimize parameters through deep reinforcement learning self-learning and self-training to obtain the control sequence that continuously optimizes the air intake performance under different flight conditions.
[0062] As another aspect of the present invention, such as Figures 1-2 As shown, an active flow control method based on intelligent algorithms is provided. First, a surrogate model of the inlet flow field characteristics at typical flight state points is established using deep learning algorithms. Then, a surrogate model of the inlet flow field characteristics under wide-speed-domain flight states is established using transfer learning methods. Finally, an interactive framework of "agent-surrogate model" is constructed using reinforcement learning methods. With flight state and flow field characteristics as core state variables and inlet parameters as control variables, real-time optimal control of the inlet throat flow field under wide-speed-domain conditions is achieved. The specific technical solution includes the following steps:
[0063] Step S1: For a typical flight state point, the Latin hypercube sampling method is used to determine the number of sets of numerical simulations of the inlet flow field characteristics required for multi-element inlet parameters (including but not limited to forebody deformation, inlet deformation, etc.). In particular, the reason for using the Latin hypercube sampling method to determine the number of simulation sets in this step is to balance the economy of simulation and sampling efficiency. In addition, quasi-random sequence, adaptive sampling, and other sampling methods can be selected for sampling. When economic conditions permit, numerical simulation under full combination can be performed to improve the accuracy of the surrogate model.
[0064] Step S2: Based on CFD numerical simulation software such as ANSYS Fluent, perform flow field simulation calculations on the number of sets determined in Step S1 to obtain Dataset 1, with inlet parameters (including but not limited to forebody deformation and inlet deformation) as independent variables and flow field characteristics (including but not limited to inlet flow coefficient, total pressure recovery coefficient, pressure ratio, etc.) as dependent variables; Based on CFD numerical simulation software such as ANSYS Fluent, for a certain set of fixed inlet parameters, obtain Dataset 2, with flight state (including but not limited to flight pressure, flight angle of attack, etc.) as independent variables and flow field characteristics (including but not limited to inlet flow coefficient, total pressure recovery coefficient, pressure ratio, etc.) as dependent variables.
[0065] Step S3: Using a convolutional neural network deep learning method and combining it with dataset 1, construct a proxy model of the air intake flow field characteristics at typical flight state points. In particular, a convolutional neural network is chosen here to better learn the flow field characteristics. If learning is required for the non-steady time-domain characteristics of the flow field, a recurrent neural network is preferred. Furthermore, this step is not limited to these two neural network methods. Equivalent substitutions and improvements of other advanced deep learning methods are all within the scope of protection of this invention.
[0066] Step S4: Based on the transfer learning algorithm, and combined with dataset 2 of wide-speed-domain flight state under fixed inlet parameters, construct a proxy model of inlet flow field features under wide-speed-domain flight state;
[0067] Step S5: Construct an environment-agent interaction model based on a deep reinforcement learning algorithm. The inlet flow field feature proxy model under wide-speed-domain flight conditions is used as the environment model for reinforcement learning. The state includes flight state and flow field features (including but not limited to parameters such as flight pressure, flight angle of attack, inlet flow coefficient, total pressure recovery coefficient, and pressure ratio). The action includes inlet parameters (including but not limited to parameters such as forebody deformation and inlet deformation). A reward function is established by comprehensively considering constraints such as inlet parameters, flight state, and flow field features, and a reward value is output in each round. Specifically, to meet the optimization conditions of continuous state space and continuous action space, deep reinforcement learning algorithms such as PPO, DDPG, and TD3 are preferred. Furthermore, equivalent substitutions and improvements of other advanced deep reinforcement learning methods are all within the scope of protection of this invention.
[0068] Step S6: Based on the environment-agent interaction model and reward function in step S5, set the optimization termination condition, optimize the parameters through deep reinforcement learning self-learning and self-training, and finally obtain the control sequence that continuously optimizes the inlet performance under different flight conditions.
[0069] This invention employs intelligent algorithms to achieve efficient and precise control of the inlet flow field across a wide speed range. The main applications of the intelligent algorithm are as follows: 1. Based on flow field feature datasets under different inlet parameter conditions at typical flight states, a proxy model of the inlet flow field features at typical flight states is constructed using deep learning intelligent algorithms; 2. Combining the flow field feature datasets for wide-speed-range flight states, a proxy model of the inlet flow field features under wide-speed-range flight states is constructed using transfer learning methods and the proxy model of the inlet flow field features at typical flight states; 3. Using the proxy model of the inlet flow field features under wide-speed-range flight states as the environment, a reinforcement learning algorithm is used to enable the agent to interact with the environment and autonomously learn the optimal strategy, thereby achieving efficient and precise control of the inlet flow field across a wide speed range.
[0070] The features described and / or illustrated above for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or in combination with or in lieu of features in other embodiments.
[0071] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, whole, step, or component, but does not exclude the presence or addition of one or more other features, wholes, steps, components, or combinations thereof.
[0072] Many features and advantages of these embodiments are apparent from this detailed description, and therefore the appended claims are intended to cover all such features and advantages of these embodiments that fall within their true spirit and scope. Furthermore, since many modifications and alterations will readily occur to those skilled in the art, the embodiments of the invention are not intended to be limited to the precise structures and operations illustrated and described, but rather to encompass all suitable modifications and equivalents falling within their scope.
[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
[0074] The parts of this invention not described in detail are techniques known to those skilled in the art.
Claims
1. An active flow control method based on intelligent algorithms, characterized in that, Includes the following steps: For a typical flight state point, with simulation economy and sampling efficiency as constraints, the number of intake parameters used for numerical simulation of intake flow field characteristics is determined by sampling method. Numerical simulations were performed according to the number of sets of the inlet parameters to obtain dataset 1 with inlet parameters as independent variables and flow field characteristics as dependent variables; For a certain set of inlet parameters, a dataset 2 was obtained by numerical simulation, with flight state as the independent variable and flow field characteristics as the dependent variable; Based on dataset 1, a proxy model of the air intake flow field characteristics at typical flight state points is constructed using deep learning methods. Based on the inlet flow field feature proxy model at the flight state point, and using the transfer learning algorithm and dataset 2, an inlet flow field feature proxy model under the wide speed domain flight state is constructed. An environment-agent interaction model is constructed based on a deep reinforcement learning algorithm. The inlet flow field feature proxy model under the wide speed domain flight state is used as the environment model for deep reinforcement learning. The state of the environment model is the flight state and flow field features, and the action of the environment model is the inlet parameters. Based on the environment-agent interaction model, optimization termination conditions are set, and through deep reinforcement learning, a control sequence that continuously optimizes the inlet performance under different flight conditions is obtained.
2. The method according to claim 1, characterized in that, The inlet parameters include at least the forebody deformation and the inlet deformation; the flow field characteristics include at least the inlet flow coefficient, the total pressure recovery coefficient, and the pressure ratio; and the flight state includes at least the flight motion pressure and the flight angle of attack.
3. The method according to claim 1, characterized in that, When determining the number of groups of the air intake parameters, multiple typical flight state points are set, and the number of groups of the air intake parameters is determined under different typical flight state points.
4. The method according to claim 1, characterized in that, The step of constructing a surrogate model of the inlet flow field characteristics for typical flight state points using deep learning based on dataset 1 is as follows: based on all independent variables in dataset 1, a surrogate model of the inlet flow field characteristics for typical flight state points for each dependent variable is constructed using deep learning.
5. The method according to claim 1, characterized in that, The sampling method is one of the following: Latin hypercube sampling method, quasi-random sequence sampling method, and adaptive sampling method.
6. The method according to claim 1, characterized in that, The deep learning method mentioned is one of the following: convolutional neural network, recurrent neural network, or long short-term memory network.
7. The method according to claim 1, characterized in that, The deep reinforcement learning algorithm is one of PPO, DDPG, or TD3.
8. The method according to claim 1, characterized in that, The reward function of the environment-agent interaction model is: Where Φ is the intake flow coefficient and σ is the total pressure recovery coefficient. This refers to the boost ratio.
9. An active flow control system based on intelligent algorithms, characterized in that, include: The sampling and processing module is used to determine the number of intake parameters for numerical simulation of intake flow field characteristics based on typical flight state points and under the constraints of simulation economy and sampling efficiency. The dataset module is used to perform flow field numerical simulation according to the number of inlet parameters determined by the sampling and processing module, and to construct dataset 1 with inlet parameters as independent variables and flow field characteristics as dependent variables; and to construct dataset 2 with flight state as independent variables and flow field characteristics as dependent variables based on a certain set of inlet parameters through flow field numerical simulation. An initial proxy model is used to construct a proxy model of the air intake flow field characteristics at typical flight state points based on dataset 1 using deep learning methods; A wide-speed-domain proxy model is used to construct a proxy model of the air intake flow field characteristics under wide-speed-domain flight conditions based on the transfer learning algorithm and dataset 2. An environment-agent interaction model is used to construct an environment-agent interaction model based on a deep reinforcement learning algorithm. The inlet flow field feature proxy model under the wide speed domain flight state is used as the environment model for deep reinforcement learning. The state of the environment model is the flight state and flow field features, and the action of the environment model is the inlet parameters. The air intake control module is used to set optimization termination conditions based on the environment-agent interaction model, and optimize parameters through deep reinforcement learning self-learning and self-training to obtain the control sequence that continuously optimizes the air intake performance under different flight conditions.
10. The system according to claim 9, characterized in that, The inlet parameters include at least the forebody deformation and the inlet deformation; the flow field characteristics include at least the inlet flow coefficient, the total pressure recovery coefficient, and the pressure ratio; and the flight state includes at least the flight motion pressure and the flight angle of attack.