A method for suppressing transonic buffeting load and response based on agent cooperation
By training the bulge and camber trailing edge collaboratively using agent-based deep reinforcement learning, the shortcomings of traditional flutter control methods in aerodynamic efficiency and nonlinear flutter suppression are solved, achieving effective suppression of wing flutter load and response and maintenance of aerodynamic performance.
Patent Information
- Application Number
- CN202610631300.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-09
- Publication Date
- 2026-07-03
AI Technical Summary
Traditional buffeting control struggles to suppress transonic buffeting in three-dimensional wings while maintaining aerodynamic efficiency, and existing control methods are ineffective under nonlinear and multi-physics coupling characteristics.
A method based on agent-based cooperative suppression is adopted. The control laws for the bulge and variable camber trailing edge are trained by deep reinforcement learning and combined with unidirectional fluid-structure interaction vibration response calculation to achieve cooperative suppression of wing flutter load and response.
While suppressing the structural vibration response caused by buffeting, the aerodynamic efficiency of the wing is maintained, and active intelligent intervention in the nonlinear buffeting flow field is achieved, which is better than the traditional control law.
Smart Images
Figure CN122331650A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of aeroelasticity and active flow control of aircraft, and specifically relates to a method for suppressing transonic chattering loads and responses based on intelligent agents in a collaborative manner. Background Technology
[0002] With the development of aircraft technology, the design of modern large aircraft strives to improve cruise speed and altitude. Currently, the cruise flight of large aircraft mostly occurs in the transonic region. During transonic flight, the interaction between shock waves and the boundary layer on the wings can easily induce wing flutter. In the fluttering state, the amplitude of aerodynamic forces and moments of the aircraft will change significantly, causing structural vibrations, thereby reducing flight quality, and even affecting the airframe strength and fatigue life. In severe cases, it can cause structural damage and threaten flight safety.
[0003] Currently, the main challenges in suppressing transonic buffeting in three-dimensional elastic wings are as follows:
[0004] First, most traditional control laws for buffeting suppression focus on reducing aerodynamic loads (such as lift coefficient pulsation). However, in practical engineering, the broadband characteristics of three-dimensional wing buffeting can trigger nonlinear aerodynamic-structural coupling, meaning that mitigating aerodynamic loads does not necessarily equate to reducing the aeroelastic response. Considering that practical engineering focuses more on the actual vibration response experienced by the wing, it is essential to incorporate the structural vibration response caused by buffeting into the control law design and evaluation system.
[0005] Secondly, regarding buffeting suppression schemes, existing buffeting suppression schemes are mainly divided into two categories: intervening in the shock wave-boundary layer region or the trailing edge flow region.
[0006] The main control methods for the shock wave-boundary layer region include jet control, vortex generators, and shock wave control bumps. Among them, shock wave control bumps can suppress the diffusion of the separation zone, reduce the amplitude of shock wave motion, and suppress the spanwise flow of three-dimensional wings, but they also introduce some lift fluctuations, making it difficult to guarantee the aerodynamic efficiency of the wing.
[0007] The main control measures for trailing edge flow include wake deflection devices and trailing edge flaps; these devices can regulate trailing edge flow, suppress airflow separation, and thus suppress flutter; however, trailing edge deflection alone is insufficient to regulate the spanwise flow of a three-dimensional wing.
[0008] Therefore, traditional single control devices are difficult to effectively suppress flutter while also ensuring aerodynamic efficiency.
[0009] Finally, in terms of control law design, the response characteristics of transonic chattering structures are complex. Relying solely on passive control or open-loop control makes it difficult to achieve good suppression effects under transonic chattering conditions. Moreover, cooperative control involves strong nonlinearity and multi-physics coupling characteristics, making it difficult to apply traditional model-based closed-loop control methods to this problem. Summary of the Invention
[0010] To address the challenge of traditional flutter control methods that fail to consider wing structural vibration response and aerodynamic performance, this invention provides a method for suppressing transonic flutter loads and responses based on intelligent agents. It introduces a unidirectional fluid-structure interaction vibration response computational architecture into the control law design and evaluation system, and obtains the optimal collaborative control law for the bulge and variable camber trailing edge through training a deep reinforcement learning model. This method suppresses structural vibration response caused by flutter while ensuring wing aerodynamic efficiency, possessing significant engineering and economic value for the development of modern large aircraft.
[0011] The method for suppressing transonic chattering load and response based on agent-based collaborative suppression comprises the following steps:
[0012] Step S1: For the wing of the aircraft under test, calculate the wingtip acceleration response of the elastic wing based on the one-way fluid-structure interaction method;
[0013] Specifically:
[0014] First, the required wing front diameter of the aircraft under test is determined. First mode, constructing the structural mode space of the wing ;
[0015] ,in, For the structural degrees of freedom of the wing.
[0016] Then, acquire the flow field and aerodynamic load state data of the wing of the aircraft under test under transonic flutter;
[0017] The flow field and aerodynamic load state data mentioned therein are derived from at least one or a combination of data from numerical simulation, wind tunnel experiments, or flight tests.
[0018] Next, the real-time extracted aerodynamic load state data is interpolated into the structural modal space using a thin-plate spline interpolation algorithm. Obtain aerodynamic load state data in structural modal space Thus, the real-time generalized aerodynamic forces of the wing are obtained;
[0019] No. Generalized aerodynamics The calculation is as follows:
[0020] Finally, the generalized aerodynamic forces of each order are substituted into the generalized structural dynamics equations, and the generalized displacements are solved by time-domain propagation using the fourth-order Runge-Kutta method, combined with the structural modal space. The modal values at the wingtip of the mid-wing were obtained by modal superposition to obtain the wingtip acceleration response of the wing structure.
[0021] Step S2: Construct an agent structure consisting of a policy network and an evaluation network based on deep reinforcement learning.
[0022] Deep reinforcement learning includes an observation space, an action space, and a multi-objective reward function;
[0023] Among them, observation space Includes load characteristics such as real-time lift-to-drag ratio and shock wave oscillation intensity, as well as wing structural response; action space It includes height commands for the telescopic micro-bulge and deflection angle commands for the seamless variable curvature trailing edge; the multi-objective weighted reward function includes vibration response, lift coefficient pulsation, lift-to-drag ratio, and control action penalty terms;
[0024] The formula is:
[0025]
[0026] in, , , and The weights of each term in the reward function, The amplitude of the wingtip acceleration response under uncontrolled conditions. To control the amplitude of the wingtip acceleration response after deployment, This represents the lift coefficient pulsation. For real-time rise-to-drag ratio, The maximum height of the bulge. The angle of deformation at the trailing edge of the variable curvature.
[0027] The policy network consists of an input layer, a first hidden layer, a second hidden layer, and an output layer. The number of neurons in each layer is "state dimension - 128 - 128 - action dimension". The hidden layers use ReLU as the post-activation function, and the output layer uses the Softmax function for normalization.
[0028] The evaluation network also consists of an input layer, a first hidden layer, a second hidden layer, and an output layer. The number of neurons in each layer is "state dimension - 128 - 280 - action dimension". The hidden layers use ReLU as the post-activation function, and the output of the evaluation network is the expected cumulative report prediction value for the current state.
[0029] The intelligent agent structure uses the cambered trailing edge and the telescoping microbulge of the wing as control mechanisms. Based on the wing's flow field state, it outputs coordinated action commands for the telescoping microbulge and the cambered trailing edge to suppress flutter, aiming to maintain the aerodynamic lift-to-drag ratio and suppress the wingtip acceleration response of the structure. The working principle is as follows:
[0030] First, input state values into the policy network. : , , , ; and the actions of the previous moment. : and According to the state Take action below probability distribution Output the action at the current moment: maximum height of the drum. and the angle of deformation at the trailing edge of the bend Boom suppression was implemented, further altering the wing's state. The evaluation network then predicted the state value of the wing's current state based on a multi-objective weighted reward function. .
[0031] Step S3: Train the agent structure and update the parameters of the evaluation network using the proximal policy optimization algorithm. .
[0032] The loss function is expressed as:
[0033]
[0034] Among them, the probability ratio Advantage function , It's a hyperparameter;
[0035] Represents the expected value in mathematics. This represents the updated policy network parameters. This represents the policy network parameters before the update. This represents the total reward obtained after the action at time t is performed. Indicates the evaluation network parameters as State value estimation.
[0036] After the agent is trained, a nonlinear mapping relationship is established between action commands and multiple targets such as lift-to-drag ratio, shock wave motion, and vibration response. This allows us to learn control strategies that can collaboratively drive the telescopic microbulge and the variable curvature trailing edge.
[0037] Step S4: Deploy the trained agent structure on the wing and observe the suppression effect of wing flutter by quantitatively evaluating the mitigation rate;
[0038] Specifically:
[0039] First, the wingtip acceleration response is calculated in real time based on the aerodynamic load data of the wing.
[0040] Then, the strategy network outputs coordinated action commands for the telescoping microbulge and the variable camber trailing edge based on the real-time flow field state and wingtip acceleration response of the wing, thereby coordinating the suppression of flutter load and wingtip acceleration response.
[0041] Next, the target characteristics such as the wingtip acceleration response amplitude and aerodynamic lift-to-drag ratio of the wing under cooperative suppression are calculated in real time.
[0042] Finally, the control effect is quantitatively evaluated by controlling the mitigation rates of various target characteristics before and after activation.
[0043] The advantages of this invention include:
[0044] 1) This invention is based on the unidirectional fluid-structure interaction method to calculate the wingtip acceleration response and introduces it as one of the indicators for flutter suppression into the control law design process, which has strong engineering practical value.
[0045] 2) This invention achieves the adaptive maintenance of the wing's reference lift-to-drag ratio performance through the coordinated control of the bulge and trailing edge while suppressing the structural vibration response caused by buffeting, thus solving the problem of aerodynamic efficiency reduction caused by a single control method.
[0046] 3) The reinforcement learning agent used in this invention can discover the phase compensation relationship between the bulge and the trailing edge, realizing active intelligent intervention in the nonlinear chattering flow field. Its suppression effect and robustness are better than traditional control laws. Attached Figure Description
[0047] Figure 1 This is a flowchart of a method for suppressing transonic chattering load and response based on intelligent agent collaboration according to the present invention;
[0048] Figure 2 This is a schematic diagram of the wing used in the embodiments of the present invention;
[0049] Figure 3 This is a comparison diagram of the wingtip acceleration response under uncontrolled and controlled conditions according to the present invention. Detailed Implementation
[0050] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings.
[0051] This invention proposes a method for suppressing transonic chattering loads and responses based on intelligent agents. It introduces a unidirectional fluid-structure interaction vibration response calculation architecture into the control law design and evaluation system. During chattering control, it takes into account both lift-to-drag ratio characteristics and vibration response. Based on reinforcement learning training, it guides the coordinated control of variable camber trailing edge and three-dimensional stretching bulge to suppress chattering. Finally, the control effect is evaluated based on the vibration response results under controlled and uncontrolled states.
[0052] like Figure 1 As shown, the specific steps are as follows:
[0053] Step S1: For the wing of the aircraft under test, calculate the wingtip acceleration response of the elastic wing based on the one-way fluid-structure interaction method;
[0054] First, the required wing front diameter of the aircraft under test is determined. First mode, constructing the structural mode space of the wing ;
[0055] ,in, For the structural degrees of freedom of the wing.
[0056] Then, acquire the flow field and aerodynamic load state data of the wing of the aircraft under test under transonic flutter;
[0057] The flow field and aerodynamic load state data mentioned therein are derived from at least one or a combination of data from numerical simulation, wind tunnel experiments, or flight tests.
[0058] Next, by extracting aerodynamic loads in real time, the load reference data is interpolated into the structural modal space using a thin-plate spline interpolation algorithm. Obtain aerodynamic load state data in structural modal space By combining the structural modal space of the wing, the real-time generalized aerodynamic forces of the wing at various orders can be obtained;
[0059] No. Generalized aerodynamics The calculation is as follows:
[0060] Finally, the generalized aerodynamic forces of each order are substituted into the right-hand side of the generalized structural dynamics equations, and the generalized displacements are solved by time-domain propagation using the fourth-order Runge-Kutta method, combined with the structural modal space. The modal values at the wingtip of the mid-wing are obtained by modal superposition to obtain the wingtip acceleration response of the wing structure; and the local load increment is updated in real time according to the action command of the actuator to achieve state output with action feedback characteristics.
[0061] Step S2: Construct an agent structure consisting of a policy network and an evaluation network based on deep reinforcement learning.
[0062] Deep reinforcement learning includes an observation space, an action space, and a multi-objective reward function;
[0063] Among them, observation space Includes load characteristics such as real-time lift-to-drag ratio and shock wave oscillation intensity, as well as wing structural response;
[0064]
[0065] in, For real-time rise-to-drag ratio, The intensity of shock wave oscillation. For wing structure response;
[0066] Action space Includes height instructions for the telescopic micro-bulge and deflection angle instructions for the seamless variable curvature trailing edge: .
[0067] The multi-objective weighted reward function balances maintaining lift-to-drag ratio characteristics and suppressing vibration response, and includes vibration response, lift coefficient pulsation, lift-to-drag ratio, and control action penalty term;
[0068] The formula is:
[0069]
[0070] in, , , and The weights of each term in the reward function, The amplitude of the wingtip acceleration response under uncontrolled conditions. To control the amplitude of the wingtip acceleration response after deployment, This represents the lift coefficient pulsation. For real-time rise-to-drag ratio, The maximum height of the bulge. The deformation angle of the variable curvature trailing edge is defined. A dual-drive system, employing both piezoelectric and electric motors, is used to drive the telescopic micro-bulge and the variable curvature trailing edge respectively.
[0071] The variable camber trailing edge achieves geometric deformation through a smooth change in airfoil camber, rather than rigid deflection. The telescopic micro-bulge is positioned at the shock wave foot of the fluttering event, and its shape is described using a Hicks-Henne type function.
[0072] The agent's structure consists of two parameterized neural networks: a policy network and an evaluation network. The policy network outputs the probability distribution of taking action a in state s. Evaluation network is used to predict the state value of the current state. It is used to assist in policy updates.
[0073] The policy network outputs control action commands based on the observed state. It consists of an input layer, a first hidden layer, a second hidden layer, and an output layer. The number of neurons in each layer is "state dimension - 128 - 128 - action dimension". The hidden layer uses ReLU as the activation function, and the output layer is normalized using the Softmax function.
[0074] The evaluation network assesses the value of the state after the current action is performed. It also includes an input layer, a first hidden layer, a second hidden layer, and an output layer. The number of neurons in each layer is "state dimension - 128 - 280 - action dimension". The hidden layers use ReLU as the activation function, and the output of the evaluation network is the expected cumulative report prediction value of the current state.
[0075] The correlation features between lift-to-drag ratio changes, load changes, and structural vibration amplitude are extracted through the hidden layer of a neural network, establishing a nonlinear mapping relationship between action commands, shock wave motion, and aerodynamic load distribution. The intelligent agent structure uses the variable camber trailing edge and telescopic microbulge of the wing as control mechanisms. Based on the flow field state of the wing, it outputs coordinated action commands of the telescopic microbulge and the variable camber trailing edge to suppress flutter, aiming to maintain the aerodynamic lift-to-drag ratio and suppress the wingtip acceleration response of the structure. The working principle is as follows:
[0076] First, input state values into the policy network. : , , , ; and the actions of the previous moment. : and According to the state Take action below probability distribution Output the action at the current moment: maximum height of the drum. and the angle of deformation at the trailing edge of the bend Boom suppression was implemented, further altering the wing's state. The evaluation network then predicted the state value of the wing's current state based on a multi-objective weighted reward function. .
[0077] Step S3: Train the agent structure and update the parameters of the evaluation network using the proximal policy optimization algorithm. .
[0078] Based on a dual-network architecture consisting of a policy network and an evaluation network composed of neural networks, the agent performs adaptive policy iteration by interacting with the environment and obtaining feedback states, aiming to maximize the cumulative reward R. The evaluation network parameters are updated using a proximal policy optimization algorithm, and the loss function is expressed as:
[0079]
[0080] Among them, the probability ratio Advantage function , It is a hyperparameter used to limit the update range.
[0081] Represents the expected value in mathematics. This represents the updated policy network parameters. This represents the policy network parameters before the update. This represents the total reward obtained after the action at time t is performed. Indicates the evaluation network parameters as State value estimation.
[0082] After the agent is trained, a nonlinear mapping relationship is established between action commands and multiple targets such as lift-to-drag ratio, shock wave motion, and vibration response. This allows us to learn control strategies that can collaboratively drive the telescopic microbulge and the variable curvature trailing edge.
[0083] Step S4: Deploy the trained agent structure on the wing and observe the suppression effect of wing flutter by quantitatively evaluating the mitigation rate;
[0084] Specifically:
[0085] First, the wingtip acceleration response is calculated in real time based on the aerodynamic load data of the wing.
[0086] Then, the strategy network outputs coordinated action commands for the telescoping microbulge and the variable camber trailing edge based on the real-time flow field state and wingtip acceleration response of the wing, thereby coordinating the suppression of flutter load and wingtip acceleration response.
[0087] Next, the target characteristics such as the wingtip acceleration response amplitude and aerodynamic lift-to-drag ratio of the wing under cooperative suppression are calculated in real time.
[0088] Finally, the control effect is quantitatively evaluated by controlling the mitigation rates of various target characteristics before and after activation.
[0089] This invention trains an intelligent agent using deep reinforcement learning to guide the expansion and contraction of micro-bulges to suppress separation zone diffusion, reduce shock wave amplitude, drive variable camber trailing edge to regulate trailing edge flow, and compensate for overall lift fluctuations caused by bulge expansion and contraction. This intelligent cooperative control law can maintain the wing's aerodynamic lift-to-drag ratio while suppressing vibration response caused by buffeting.
[0090] Example:
[0091] The research object of this example is a rectangular straight airfoil with a span of 0.8128m and a chord length of 0.4064m. The airfoil adopts the SC(2)0414 supercritical airfoil with a relative thickness of 14%. Figure 2 As shown. To suppress transonic flutter on the wing, the location and size parameters of the variable camber trailing edge and the three-dimensional bulge deformation region need to be determined beforehand on the wing. In this example, the three-dimensional bulge is set at 50% along the chord, with three bulges evenly spaced from the wing root to the wingtip. The shape of the bulge is described by a Hicks-Henne function. The variable camber trailing edge is designed in the middle section of the wing, with the chord starting position at 70% of the chord length, and the trailing edge length is 0.3 times the span. The trailing edge deflection function uses a quadratic curve deflection.
[0092] Step S1: Construct a simulation environment that simultaneously has the ability to calculate buffeting load and vibration response, as well as motion feedback.
[0093] To obtain the flow field and load state of the wing under transonic flutter, a unidirectional fluid-structure interaction method is developed to establish a simulation model that includes vibration response calculation and actuator action feedback.
[0094] The specific steps are as follows:
[0095] The wing is meshed, a structural mesh is generated on the wing surface and in the computational domain and boundary conditions are set. The diffusion smoothing method is used to realize the dynamic mesh deformation, so as to ensure that the mesh can adapt to the deformation of the bulge and the trailing edge.
[0096] User-defined functions were written to determine the deformation location, size, and deformation mode of the 3D bulge and the trailing edge with varying curvature. These functions were integrated into a unidirectional fluid-structure interaction vibration response calculation architecture. Specific steps included:
[0097] 1) Extract the first two elastic modes of the wing, and interpolate the structural mode shapes from the structural mesh to the aerodynamic center of the fluid mesh using thin plate spline interpolation.
[0098] 2) The generalized aerodynamic forces are calculated by integrating the structural modes and pressure distribution under the fluid grid.
[0099] 3) The generalized aerodynamic force is applied to the structural dynamics equations and solved using the Runge-Kutta method to obtain the wingtip acceleration response of the wing structure;
[0100] High-precision CFD calculations are performed using URANS or DDES turbulence models, with real-time mesh updates at each time step to calculate local load increments at the actuators, enabling real-time feedback of environmental actions to control movements and real-time calculation of vibration responses. After constructing the simulation environment, the simulation results of buffeting under uncontrolled conditions can be compared with experimental results from the same conditions in the literature to ensure the accuracy of the numerical simulation.
[0101] Step S2: Configure the multi-objective reinforcement learning agent. Specific steps include:
[0102] A deep reinforcement learning agent is deployed in a simulation environment, and the observation space, action space, and multi-objective reward function of the deep reinforcement learning algorithm are set. The variable curvature trailing edge and the telescopic microbulge are used as control mechanisms. The goal is to maintain the aerodynamic lift-to-drag ratio and suppress the vibration response of the structure. The agent is trained in the simulation environment using a proximal policy optimization algorithm to obtain the optimal control law that can collaboratively drive the telescopic microbulge and the variable curvature trailing edge.
[0103] In this example, the weighting coefficients are set as follows: ω1=10, ω2=5, ω3=5, ω4=5.
[0104] Step S3: Intelligent Collaboration Strategy Training. Specific steps include:
[0105] In the simulation environment, based on the dual-network architecture of a policy network and an evaluation network composed of neural networks, the agent performs interactive actions with the environment and obtains feedback states. With the goal of maximizing cumulative rewards, it performs adaptive policy iteration and updates the neural network weights using a proximal policy optimization algorithm. This enables the agent to establish a nonlinear mapping relationship between action commands and multiple objectives such as shock wave motion, aerodynamic load distribution, and vibration response, thereby learning a control strategy that can collaboratively drive the telescopic microbulge and the variable curvature trailing edge.
[0106] In the simulated environment, the agent is initialized to obtain an initialized cooperative control law model. Then, the agent performs actions with the simulation environment and obtains feedback states. Based on the feedback states, it outputs the next moment's control device action through the action network and inputs it into the simulation model established in step S1 to obtain the next moment's wing flow field state, completing the closed-loop interaction.
[0107] Meanwhile, during the training loop, the agent puts the interaction data into the experience pool. When the amount of data reaches a certain batch, the system performs random sampling and uses the Proximal Policy Optimization (PPO) algorithm to update the neural network weights.
[0108] This training process enables the agent to perform adaptive policy iteration with the goal of maximizing cumulative rewards, helping the agent to establish a nonlinear mapping relationship between action commands and shock wave motion and aerodynamic load distribution, thereby training an optimal control law that can collaboratively drive the telescopic bulge and the variable curvature trailing edge.
[0109] Step S4: Online Suppression and Evaluation. Specific steps include:
[0110] The trained cooperative control law is deployed in different buffeting environments of the target wing. Based on the real-time flow field state of the wing, cooperative action commands for the telescopic bulge and the variable camber trailing edge are output to suppress buffeting loads. The telescopic bulge suppresses the diffusion of the separation zone, reduces the shock wave amplitude, drives the variable camber trailing edge to regulate the trailing edge flow, and compensates for the overall lift fluctuation caused by the bulge telescopic movement. Finally, the unidirectional fluid-structure interaction module is used to calculate and quantitatively evaluate the amplitude of the suppressed structural vibration response and the aerodynamic lift-to-drag ratio characteristics in real time.
[0111] like Figure 3 As shown, after the control is activated, the wingtip acceleration response decreases from a peak of 4.60g in the uncontrolled state to 0.9g in the uncontrolled state, with a mitigation efficiency of 80.4%, indicating that the cooperative control has a significant effect on suppressing the wing vibration response caused by flutter.
[0112] Simulation results show that after the control is activated, the lift-to-drag ratio of the wing differs from that of the steady-state condition by less than 5%, ensuring no significant loss in the wing's aerodynamic efficiency. This demonstrates that the intelligent collaborative control law for the bulge and variable camber trailing edge designed based on deep reinforcement learning in this invention performs well, effectively maintaining the wing's lift-to-drag ratio characteristics while suppressing vibration response.
Claims
1. A method for suppressing transonic chattering loads and responses based on agent-based collaborative methods, characterized in that, The specific steps are as follows: Step S1: For the wing of the aircraft under test, calculate the wingtip acceleration response of the elastic wing based on the one-way fluid-structure interaction method; Step S2: Construct an agent structure consisting of a policy network and an evaluation network based on deep reinforcement learning; Deep reinforcement learning includes an observation space, an action space, and a multi-objective reward function; The multi-objective weighted reward function includes vibration response, lift coefficient pulsation, lift-to-drag ratio, and control action penalty term. The formula is: ; , , and The weights of each term in the reward function, The amplitude of the wingtip acceleration response under uncontrolled conditions. To control the amplitude of the wingtip acceleration response after deployment, This represents the lift coefficient pulsation. For real-time rise-to-drag ratio, The maximum height of the bulge. The angle of deformation at the trailing edge of the variable curvature; The intelligent agent structure uses the variable camber trailing edge and telescopic micro-bulge of the wing as the control mechanism. Based on the flow field state of the wing, it outputs the coordinated action command of the telescopic micro-bulge and the variable camber trailing edge to suppress flutter, with the goal of maintaining the aerodynamic lift-to-drag ratio and suppressing the wingtip acceleration response of the structure. Step S3: Train the agent structure and update the parameters of the evaluation network using the proximal policy optimization algorithm; After the agent training is completed, a nonlinear mapping relationship between the action instruction and the lift-drag ratio, shock motion, and wingtip acceleration response multi-objective is established, that is , so as to learn the control strategy capable of cooperatively driving the telescopic micro-bulge and the variable-camber trailing edge. Step S4: Deploy the trained agent structure on the wing and observe the suppression effect of wing flutter by quantitatively evaluating the mitigation rate.
2. The method of claim 1, wherein, Step S1 specifically involves: First, the wing front order mode required by the aircraft to be tested is solved ; wherein, is the structural freedom of the wing; Then, acquire the flow field and aerodynamic load state data of the wing of the aircraft under test under transonic flutter; Next, the real-time extracted aerodynamic load state data is interpolated into the structural modal space using a thin-plate spline interpolation algorithm. Obtain aerodynamic load state data in structural modal space Combining structural modal space Obtain the real-time generalized aerodynamic forces of the wing at all orders; No. Generalized aerodynamics The calculation is as follows: ; Finally, the generalized aerodynamic forces of each order are substituted into the generalized structural dynamics equations, and the generalized displacements are solved by time-domain propagation using the fourth-order Runge-Kutta method, combined with the structural modal space. The modal values at the wingtip of the mid-wing were obtained by modal superposition to obtain the wingtip acceleration response of the wing structure.
3. The method of claim 2, wherein, The flow field and aerodynamic load state data are derived from at least one or a combination of data from numerical simulation, wind tunnel experiments, or flight tests.
4. The method of claim 1, wherein, The step S2, observing space The load characteristics include real-time lift-drag ratio, shock oscillation intensity and wing structure response; action space The height instruction of the telescopic micro-bulge and the deflection angle instruction of the seamless variable-camber trailing edge.
5. The method of claim 4, wherein, In step S2, the policy network includes an input layer, a first hidden layer, a second hidden layer, and an output layer. The number of neurons in each layer is "state dimension - 128 - 128 - action dimension". The hidden layer uses ReLU as the activation function, and the output layer is normalized using the Softmax function. The evaluation network also consists of an input layer, a first hidden layer, a second hidden layer, and an output layer. The number of neurons in each layer is "state dimension - 128 - 280 - action dimension". The hidden layer uses ReLU as the post-activation function, and the output of the evaluation layer network is the expected cumulative report prediction value in the current state.
6. The method of claim 5, wherein, In step S2, the working principle of the intelligent agent structure is as follows: First, input state values into the policy network. : , , , ; and the actions of the previous moment. : and According to the state Take action below probability distribution Output the action at the current moment: maximum height of the drum. and the angle of deformation at the trailing edge of the bend Boom suppression was implemented, further altering the wing's state. The evaluation network then predicted the state value of the wing's current state based on a multi-objective weighted reward function. .
7. The method of claim 5, wherein, In step S2, the variable camber trailing edge achieves geometric deformation by smoothly changing the airfoil camber, rather than rigid deflection; the telescopic micro-bulge is set at the shock wave foot position of the fluttering, and the shape of the bulge is described by a Hicks-Henne type function.
8. The method as described in claim 5, characterized in that, In step S3, the loss function is expressed as: ； where the probability ratio , the advantage function , is a hyperparameter; Represents the expected value in mathematics. This represents the updated policy network parameters. This represents the policy network parameters before the update. This represents the total reward obtained after the action at time t is performed. Indicates the evaluation network parameters as State value estimation.
9. The method of claim 1, wherein, Step S4 specifically involves: First, the wingtip acceleration response is calculated in real time based on the aerodynamic load data of the wing. Then, the strategy network outputs coordinated action commands for the telescoping microbulge and the variable camber trailing edge based on the real-time flow field state and wingtip acceleration response of the wing, thereby coordinating the suppression of flutter load and wingtip acceleration response. Next, the target characteristics such as the wingtip acceleration response amplitude and aerodynamic lift-to-drag ratio of the wing under cooperative suppression are calculated in real time. Finally, the control effect is quantitatively evaluated by controlling the mitigation rates of various target characteristics before and after activation.