Flutter suppression design and verification method based on reinforcement learning and variable camber trailing edge
Through the vibration suppression design based on reinforcement learning and variable curvature trailing edge, a fast response stall flutter model was established, and closed-loop control was achieved using NARX neural network and DQN algorithm, which solved the problem of aircraft stall flutter control, and improved flight performance and calculation efficiency.
Patent Information
- Application Number
- CN202410447237.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-15
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-04-15
AI Technical Summary
The existing technology is difficult to effectively control the stall flutter that occurs in aircraft under large flight angles of attack, mainly due to the difficulty in modeling the strong nonlinear and high-complex aerodynamics of stall flutter, and the traditional control methods lack adaptability and design cycle. long.
Using the flutter suppression design method based on reinforcement learning and variable curvature trailing edge, the DQN algorithm is designed to achieve closed-loop active control and suppress wing by establishing a fast-responsive stall flutter model, using the NARX neural network to reduce the order aerodynamic torque, combined with the aeroelastic control equation, and designing the DQN algorithm to achieve closed-loop active control and suppress wing type stall flutter.
Effective suppression of airfoil stall flutter is achieved, aerodynamic noise is reduced, flight performance is improved, computational cost and model complexity is reduced, and the low complexity and output continuity of the control law is ensured.
Smart Images

Figure CN118133429B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intelligent aircraft design and fluid-solid coupling active control, and in particular to a flutter suppression design and verification method based on reinforcement learning and a variable-camber trailing edge. Background Art
[0002] The seamless variable camber trailing edge is a form of intelligent morphing wing and one of the innovative concepts in the aviation field. Compared with traditional rigid wings, the seamless variable camber trailing edge wing introduces a dynamically adaptable shape structure. By adopting advanced materials and structural design types, the wing is allowed to adjust its shape in real time during flight according to conditions such as Mach number, angle of attack, maneuvering, disturbance, etc., so as to change the aerodynamic characteristics such as lift and drag of the wing and optimize aerodynamic efficiency. Studies have shown that this flexibility not only improves the maneuverability of the aircraft, but also has the potential to reduce aerodynamic drag and improve fuel efficiency. It is one of the key innovations to promote performance improvement in the aviation field.
[0003] The interaction between aerodynamic force and structural elasticity may cause an unstable vibration phenomenon, namely stall flutter, at a large flight angle of attack. Stall flutter is a safety risk that must be avoided in aircraft design. It affects the maneuverability of the aircraft at large angles of attack and reduces the life of the structure. Therefore, effective control of the stall flutter of the aircraft is of great significance to ensure the safety of the aircraft and increase its life.
[0004] At present, there are few studies on the active control of the stall flutter process. The main reason is that it is difficult to model the strong nonlinear and highly complex aerodynamic forces of stall flutter. The existing control methods are mainly based on traditional linear feedback control. The parameters are highly dependent on empirical adjustment and lack parameter adaptability to various flight conditions. In addition, the design of control laws using high-precision computational fluid dynamics (CFD) calculations results in a long design cycle and high difficulty.
[0005] In general, there is a lack of fast, effective and adaptable control law design methods. Summary of the invention
[0006] In view of the above problems, the present invention proposes a flutter suppression design and verification method based on reinforcement learning and variable camber trailing edge, which realizes the stall flutter suppression of variable camber trailing edge through the steps of aerodynamic reduction of variable camber trailing edge, modeling of fast response stall flutter model, and design and verification of reinforcement learning control scheme.
[0007] The flutter suppression design and verification method based on reinforcement learning and variable camber trailing edge has the following specific steps:
[0008] Step 1: For the airfoil model with pitch freedom and active continuous trailing edge deflection, perform a composite motion of pitch and trailing edge deflection in the actively controlled stall flutter motion.
[0009] The compound motion system sets up an absolute coordinate system and a local coordinate system. The absolute coordinate system takes the leading edge of the stationary airfoil as the origin, the incoming flow direction (horizontally to the right) as the x direction, and the vertical upward direction as the y direction; the local coordinate system takes the leading edge of the airfoil as the origin, and the chord direction of the airfoil as the x direction. b Direction, the vertical chord is upward as y b direction.
[0010] The thickness of the airfoil trailing edge remains unchanged during deformation, and its governing equation is:
[0011]
[0012] Among them, x, y are the absolute coordinates of the control points on the airfoil surface, x b ,y b is the local coordinate of the control point on the airfoil surface, and d is the deflection of the control point on the chord line.
[0013] Step 2: Establish a fast response stall flutter model of a variable camber trailing edge airfoil including an unsteady aerodynamic moment and an aeroelastic system;
[0014] The modeling process is:
[0015] Firstly, the pitching motion angle θ and trailing edge deflection β of the airfoil model are established by means of sine superposition;
[0016]
[0017] Among them, A i With ω i is the randomly generated sine function amplitude and angular velocity;
[0018] Then, using CFD calculation, the unsteady aerodynamic moments corresponding to the pitch motion and the trailing edge deflection are obtained, namely:
[0019]
[0020] Where M is the aerodynamic moment, the subscripts pit and tem refer to pitch motion and trailing edge deflection, respectively, and CFD refers to computational fluid dynamics simulation.
[0021] After that, the incoming flow velocity U ∞ (t), pitch angle, trailing edge deflection and their respective unsteady aerodynamic moments as training sets; specifically:
[0022]
[0023] Next, two NARX neural networks are used to fit the unsteady aerodynamic moments under the two motions by error back propagation, and two reduced-order models are obtained:
[0024]
[0025] The input of the reduced-order model 1 is the angle of attack θ=[θ t ,θ t-1 ,...,θ t-10 ] and the incoming flow velocity U ∞ =[U ∞,t ,U ∞,t-1 ,...,U ∞,t-10 ], the output is the unsteady aerodynamic moment M of the large pitch motion at the current time step pit ;
[0026] The input of the reduced-order model 2 is the trailing edge deflection β = [β t ,β t-1 ,...,β t-10 ] and the incoming flow velocity U ∞ =[U ∞,t ,U ∞,t-1 ,...,U ∞,t-10 ], the output is the unsteady aerodynamic moment M of the airfoil trailing edge deflection motion at the current time step tem ;
[0027] Finally, the sum of the unsteady aerodynamic moments of the trailing edge deflection and large pitch compound motion of the two reduced-order models is calculated and combined with the ordinary differential equations of aeroelastic control:
[0028]
[0029] Where, I is the airfoil moment of inertia, C is the airfoil pitch motion damping, K is the airfoil pitch motion stiffness, is the angular acceleration of pitch motion, for
[0030] Pitch motion angular velocity.
[0031] The fourth-order Runge-Kutta method is used to solve formula (6) and the pitch angle θ is obtained. t and pitch angular velocity The stall flutter model f is obtained by combining the reduced-order model to achieve rapid aerodynamic and airfoil pitch motion response to random trailing edge deflection signals. SF :
[0032]
[0033] That is, by inputting any time sequence of trailing edge deflection, the airfoil can respond quickly to pitch motion and obtain the corresponding unsteady aerodynamic moment, pitch angle and pitch angular velocity.
[0034] Step 3: Design a fast response stall flutter model based on reinforcement learning DQN SFOffline closed-loop control scheme;
[0035] The offline control scheme refers to: using the established fast response stall flutter model f SF For the aeroelastic simulation environment, a fully connected neural network is trained using reinforcement learning algorithm as a control agent to form a closed-loop active control system. By observing the motion parameters under stall flutter such as the airfoil pitch angle and pitch angular velocity, the corresponding effective trailing edge deflection is given.
[0036] The reward function of the DQN algorithm consists of a reward term and a penalty term;
[0037] The rewards are set as:
[0038]
[0039] Among them, R t It is the reward for the previous action based on the current observation, which is divided into three categories:
[0040] (i) The trailing edge deflection is within the limit (0.32 rad), and the angle of attack and angular velocity are both very small ( θ t <0.1rad), it is considered close to stillness and rewards the previous action;
[0041] (ii) The trailing edge deflection is within the limit (0.32 rad), but does not reach a near-stationary suppression effect ( or θ t ≥0.1rad), the penalty for the previous action is the pitch angle of attack With angular velocity θ t Linear combinations of
[0042] (iii) When the trailing edge deflection exceeds the allowed amplitude (0.32 rad), a large penalty is imposed on the previous action regardless of whether the suppression effect is achieved.
[0043] The reinforcement learning agent is a fully connected neural network with 2 hidden layers, and the input is the environment information (angle of attack θ t , pitch angle
[0044] Speed θ t , trailing edge deflection β t ), the output is five optional actions (the trailing edge of the variable camber is greatly deflected upward Δβ=Δβ l , large downward deviation Δβ=-Δβ l , a small upward deviation Δβ=Δβ s , a slight downward deviation Δβ = -Δβ s , keep the Q value unchanged Δβ=0), select the action corresponding to the maximum Q value, and the trailing edge deflection β at the current moment t+1 =β t+Δβ is used as the controller output to control the next trailing edge deflection, forming a closed-loop active control in an offline environment.
[0045] Step 4: Online verification of the closed-loop active control scheme for high-precision fluid-solid control based on the combination of reinforcement learning and CFD;
[0046] Online verification of closed-loop active control solutions refers to:
[0047] First, the fully connected neural network trained offline is integrated into CFD to obtain the stall flutter model f SF In the aeroelastic simulation environment, the fully connected neural network is used as the control agent to form a closed-loop active control system. By observing the motion parameters of the airfoil under stall flutter, such as the pitch angle and pitch angular velocity, the corresponding effective trailing edge deflection is given.
[0048] The main steps are:
[0049] (i) CFD performs transient flow field calculations and outputs the aerodynamic force M on the airfoil surface corresponding to the current time t, which is substituted into the aeroelastic control equation (6);
[0050] (ii) Write the fourth-order Runge-Kutta method in the UDF to solve the aeroelastic control equation (6) corresponding to the current time t, and output
[0051] Pitching angle of attack θ t With angular velocity θ t ;
[0052] (iii) A fully connected neural network agent is embedded using a custom function UDF; the neural network agent is based on the current time t.
[0053] Environmental information [θ t ,θ t ,β t ] is the feedback signal, outputting five optional actions Δβ=[Δβ l ,-Δβ l ,-Δβ s ,Δβ s ,0] corresponding to the Q value;
[0054] (iv) Filter the maximum Q value as the next step trailing edge deflection increment, and add the trailing edge deflection at the current moment as the controller output β t+1 =β t +Δβ;
[0055] (v) for pitch θ t With trailing edge deflection β t+1 The composite motion airfoil surface control points perform coordinate calculations to control the structural shape and fluid grid deformation, forming a closed-loop active control in a high-fidelity environment.
[0056] The advantages of the present invention are:
[0057] 1) The present invention realizes effective suppression of airfoil stall flutter by continuously bending the trailing edge. Compared with the traditional rigid trailing edge method, it can achieve the same control effect, while greatly reducing aerodynamic noise and improving flight performance.
[0058] 2) The present invention adopts a nonlinear autoregressive neural network with exogenous input to reduce the order of complex aerodynamics. Compared with the recurrent neural network commonly used for time series reduction, the model complexity and training time are reduced; the reduced-order model is combined with differential equations to establish a stall flutter fast response model to identify the pitch response under various trailing edge deflection inputs. Compared with the fluid-solid coupling method commonly used for stall flutter calculation, the calculation cost is greatly reduced.
[0059] 3) The present invention adopts the discrete increment method to convert the airfoil trailing edge deflection in infinite continuous space into limited action selection, and imposes a large penalty on the trailing edge deflection that deviates from the limited conditions, thereby ensuring the low complexity, output continuity and rationality of the algorithm, preventing it from producing large mutations that cannot be completed by the actual mechanical structure, and avoiding the negative volume problem in computational fluid dynamics simulation;
[0060] 4) The controller designed by the reinforcement learning DQN algorithm in the present invention has good generalization performance for the model and is not limited by the fluid grid, calculation time step and other model parameters. BRIEF DESCRIPTION OF THE DRAWINGS
[0061] Figure 1 It is a flow chart of the flutter suppression design and verification method based on reinforcement learning and variable camber trailing edge of the present invention;
[0062] Figure 2 is a schematic diagram of the pitching and trailing edge deformation of the airfoil of the present invention;
[0063] Figure 3 is a schematic diagram of the aerodynamic modeling process of the present invention;
[0064] Figure 4 is a schematic diagram of a stall flutter rapid response model of the present invention;
[0065] Figure 5 is a schematic diagram of the stall flutter control scheme of the present invention;
[0066] Figure 6 It is a schematic diagram of the high-fidelity fluid-solid control coupling calculation process of the present invention. DETAILED DESCRIPTION
[0067] In order to facilitate those skilled in the art to understand and implement the present invention, the present invention is further described in detail and in depth below with reference to the accompanying drawings.
[0068] The present invention discloses a flutter suppression design and verification method based on reinforcement learning and variable camber trailing edge. A trailing edge continuous variable camber form is designed for an airfoil. A NARX neural network is used to reduce the order of nonlinearity and unsteady aerodynamic forces, and a stall flutter fast response model is established by mixing ordinary differential equations. A trailing edge deflection controller based on a DQN reinforcement learning algorithm is trained using the fast response model, and continuous trailing edge camber is achieved, thereby suppressing the stall flutter of the airfoil.
[0069] like Figure 1 As shown in the figure, the flutter suppression design and verification method based on reinforcement learning and variable camber trailing edge has the following specific steps:
[0070] Step 1: For the airfoil model with pitch freedom and active continuous trailing edge deflection, perform a composite motion of pitch and trailing edge deflection in the actively controlled stall flutter motion.
[0071] The compound motion system sets up an absolute coordinate system and a local coordinate system. The absolute coordinate system takes the leading edge of the stationary airfoil as the origin, the incoming flow direction (horizontally to the right) as the x direction, and the vertical upward direction as the y direction; the local coordinate system takes the leading edge of the airfoil as the origin, and the chord direction of the airfoil as the x direction. b Direction, the vertical chord is upward as y b direction.
[0072] The thickness of the airfoil trailing edge remains unchanged during deformation, and its governing equation is:
[0073]
[0074] Among them, x, y are the absolute coordinates of the control points on the airfoil surface, x b ,y b is the local coordinate of the control point on the airfoil surface, and d is the deflection of the control point on the chord line.
[0075] Step 2: Establish a fast response stall flutter model of variable camber trailing edge airfoil based on aerodynamic reduction model;
[0076] The stall flutter model includes two parts: unsteady aerodynamic moment modeling and aeroelastic system modeling. The unsteady aerodynamic moment of the combined motion process of airfoil trailing edge deflection and large pitching is approximated as the sum of the unsteady aerodynamic moments of two separate motions. Both types of unsteady aerodynamic moments are quickly calculated by establishing reduced-order models based on the high-precision CFD calculation results.
[0077] The modeling process is:
[0078] Firstly, a composite sinusoidal motion function is designed in the computational fluid dynamics software, and the pitching motion angle θ and the trailing edge deflection β of the airfoil model are established as time series inputs in a sinusoidal superposition manner.
[0079]
[0080] Among them, A i With ω i is the randomly generated sine function amplitude and angular velocity;
[0081] Then, using CFD calculation, the unsteady aerodynamic moments corresponding to the pitch motion and the trailing edge deflection are obtained, namely:
[0082]
[0083] After that, the incoming flow velocity U ∞ (t), pitch angle, trailing edge deflection and their respective unsteady aerodynamic moments as training sets; specifically:
[0084]
[0085] Then, two nonlinear autoregressive (NARX) neural networks with external input are used to fit the unsteady aerodynamic moments under the two motions by error back propagation, and two reduced-order models are obtained:
[0086]
[0087] The input of the reduced-order model 1 is the angle of attack θ=[θ t ,θ t-1 ,...,θ t-10 ] and the incoming flow velocity U ∞ =[U ∞,t ,U ∞,t-1 ,...,U ∞,t-10 ], the output is the unsteady aerodynamic moment M of the large pitch motion at the current time step pit ;
[0088] The input of the reduced-order model 2 is the trailing edge deflection β = [β t ,β t-1 ,...,β t-10 ] and the incoming flow velocity U ∞ =[U ∞,t ,U ∞,t-1 ,...,U ∞,t-10 ], the output is the unsteady aerodynamic moment M of the airfoil trailing edge deflection motion at the current time step tem ;
[0089] Finally, the sum of the unsteady aerodynamic moments of the trailing edge deflection and large pitch compound motion of the two reduced-order models is calculated, and combined with the ordinary differential equation of aeroelastic control of the pitch degree of freedom:
[0090]
[0091] Where, I is the airfoil moment of inertia, C is the airfoil pitch motion damping, K is the airfoil pitch motion stiffness, is the angular acceleration of pitch motion, is the pitch motion angular velocity.
[0092] The fourth-order Runge-Kutta method is used to solve formula (6) and the pitch angle θ is obtained. t and pitch angular velocity The stall flutter model f is obtained by combining the reduced-order model to achieve rapid aerodynamic and airfoil pitch motion response to random trailing edge deflection signals. SF :
[0093]
[0094] That is, by inputting any time sequence of trailing edge deflection, the airfoil can respond quickly to pitch motion and obtain the corresponding unsteady aerodynamic moment, pitch angle and pitch angular velocity.
[0095] Step 3: Design a fast-response stall flutter model based on reinforcement learning Deep-Q-Network (DQN) SF Offline closed-loop control scheme;
[0096] The offline control scheme refers to: using the established fast response stall flutter model f SF For the aeroelastic simulation environment, the reinforcement learning algorithm DQN is used to train a fully connected neural network as a control agent, and the agent observes the environmental information As the controller feedback signal, the agent receives the feedback signal and outputs the Q value of the optional action Δβ. The action corresponding to the maximum Q value is selected and the trailing edge deflection β at the current moment is added. t+1 =β t +Δβ is used as the controller output to control the next trailing edge deflection, forming a closed-loop active control in an offline environment. By observing the motion parameters of the airfoil under stall flutter, such as the pitch angle and pitch angular velocity, the corresponding effective trailing edge deflection is given.
[0097] The reward function of the DQN reinforcement learning algorithm consists of a reward term and a penalty term. The penalty value is positively correlated with the angle of attack and the angular velocity amplitude. It rewards situations that are close to complete suppression and imposes a large penalty on situations that exceed the trailing edge deflection and pitch motion limit values.
[0098] The rewards are set as:
[0099]
[0100] Among them, R t It is the reward for the previous action based on the current observation, which is divided into three categories:
[0101] (iv) The trailing edge deflection is within the limit (0.32 rad), and the angle of attack and angular velocity are both very small ( θ t <0.1rad), it is considered close to stillness and rewards the previous action;
[0102] (v) The trailing edge deflection is within the limit (0.32 rad), but does not reach a near-stationary suppression effect ( or θ t ≥0.1rad), the penalty amplitude for the previous action is the pitch angle ω t With angular velocity θ t Linear combinations of
[0103] (vi) If the trailing edge deflection exceeds the allowed amplitude (0.32 rad), a large penalty will be imposed on the previous action regardless of whether the suppression effect is achieved.
[0104] The reinforcement learning agent is a fully connected neural network with 2 hidden layers, and the input is the environment information (angle of attack θ t , pitch angular velocity Trailing edge deflection β t ), the output is five optional actions (the trailing edge of the variable camber is greatly deflected upward Δβ=Δβ l , large downward deviation Δβ=-Δβ l , a small upward deviation Δβ=Δβ s , slightly downward deviation Δβ = -Δβ s , keep still Δβ=0), the larger the Q value is, the more conducive the corresponding action is to suppress the pitch motion.
[0105] Select the action corresponding to the maximum Q value and the trailing edge deflection β at the current moment t+1 =β t +Δβ is used as the controller output to control the next trailing edge deflection, forming a closed-loop active control in an offline environment.
[0106] Step 4: Use the fully connected neural network agent designed by DQN to conduct online verification of the closed-loop active control scheme of high-precision fluid-solid control based on the combination of reinforcement learning and CFD;
[0107] Online verification of closed-loop active control solutions refers to:
[0108] Firstly, using CFD high-fidelity fluid-structure interaction calculation as the environment, a fully connected neural network agent is embedded using a user-defined function to simulate the stall flutter model f. SF For the aeroelastic simulation environment, the fully connected neural network is used as the control agent. The neural network agent uses the current time environment information The Q value corresponding to the five optional actions is output as the feedback signal, and the maximum value is selected as the trailing edge deflection increment for the next step, and the sum of the trailing edge deflection at the current moment is used as the controller output β t+1 =β t +Δβ, controls the deformation of the structural shape and the fluid grid, and forms a closed-loop active control in a high-fidelity environment; by observing the motion parameters of the airfoil under stall flutter, such as the pitch angle and pitch angular velocity, the corresponding effective trailing edge deflection is given.
[0109] The main steps are:
[0110] (vi) CFD performs transient flow field calculations and outputs the aerodynamic force M on the airfoil surface corresponding to the current time t, which is substituted into the aeroelastic control equation;
[0111] (vii) Write the fourth-order Runge-Kutta method in the UDF to solve the aeroelastic control equation corresponding to the current time t and output the pitch motion angle θ t With angular velocity
[0112] (viii) A fully connected neural network agent is embedded using a custom function UDF; the neural network agent uses the environmental information at the current time t is the feedback signal, outputting five optional actions Δβ=[Δβ l ,-Δβ l ,-Δβ s ,Δβ s ,0] corresponding to the Q value;
[0113] (ix) Filter the maximum Q value as the next step trailing edge deflection increment, and add the trailing edge deflection at the current moment as the controller output β t+1 =β t +Δβ;
[0114] (x) for pitch θ t With trailing edge deflection β t+1 The composite motion airfoil surface control points perform coordinate calculations to control the structural shape and fluid grid deformation, forming a closed-loop active control in a high-fidelity environment.
[0115] Example:
[0116] Step 1: Establish a fast response stall flutter model of a variable camber trailing edge airfoil based on an aerodynamic reduction model;
[0117] The geometric model used in this embodiment is a NACA0012 airfoil with pitch freedom and active continuous trailing edge deflection, such as Figure 2As shown, the airfoil trailing edge position is set at 0.7 to 1 times the chord length, the maximum deflection is 0.1 times the chord length, the airfoil pitch motion center is 0.35 times the chord length, and in the actively controlled stall flutter motion, a compound motion of pitching and trailing edge deflection is performed.
[0118] The absolute coordinate system of the composite motion takes the leading edge of the stationary airfoil as the origin, the incoming flow direction (horizontally to the right) as the x direction, and the vertical upward direction as the y direction; the local coordinate system takes the leading edge of the airfoil as the origin, and the chord direction of the airfoil as the x direction. b Direction, the vertical chord is upward as y b The airfoil trailing edge deflection is driven by the chord line, and its governing equation is:
[0119]
[0120] Where c is the chord length, x b is the chord length position of the control point on the chord, and d is the deflection of the control point on the chord.
[0121] The thickness of the airfoil trailing edge remains unchanged during deformation, and its governing equation is:
[0122]
[0123] Among them, x, y are the absolute coordinates of the control points on the airfoil surface, x b ,y b are the local coordinates of the control points on the airfoil surface.
[0124] The establishment of the stall flutter model of the variable camber trailing edge of the airfoil mainly includes two parts: unsteady aerodynamic modeling and aeroelastic system modeling; the aerodynamic force of the stall flutter process of the airfoil with a variable camber trailing edge is composed of the unsteady aerodynamic force caused by the deflection of the flexible trailing edge and the unsteady aerodynamic force of the pitch motion during the stall flutter process of the airfoil. The composite motion aerodynamic force containing pitch and trailing edge deflection is simplified to the linear sum of the two actions through linear assumptions:
[0125] f=f p +f t (3)
[0126] Among them, f p is the aerodynamic force during pitching motion, f t It is the aerodynamic force of the trailing edge deflection process.
[0127] The aerodynamic modeling process is as follows Figure 3 As shown in the figure, the aerodynamic forces of the two types of motion processes are calculated with high precision by CFD as the reduced-order model training set, considering the following composite sinusoidal motion:
[0128]
[0129] Among them, A iWith ω i The amplitude and angular velocity of the randomly generated sinusoidal function are used to establish the pitch motion angle of attack and trailing edge deflection signal input in a sinusoidal superposition manner, and the corresponding unsteady aerodynamic torque data are calculated in CFD.
[0130] After that, the reduced-order model is trained with the incoming flow velocity and pitch angle (trailing edge deflection) as input. The NARX neural network effectively simulates the time series process by taking the output results of multiple historical time steps as part of the network input. In the present invention, the number of neural network layers used for both types of processes is 10, the activation function is tanh, and the input time delay is 10 time steps.
[0131] Finally, the two aerodynamic reduction models can be used to quickly obtain the sum of the unsteady aerodynamic forces in the combined motion process of the airfoil trailing edge deflection and large pitching. By combining the ordinary differential equations of aeroelastic control, the aeroelastic equations are solved by the fourth-order Runge-Kutta method, and a simplified model for realizing rapid aerodynamic force and airfoil pitching motion response to random trailing edge deflection signals is obtained, such as Figure 4 This simplified model can achieve a fast response of the airfoil pitching motion under arbitrary trailing edge deflection input, and obtain the corresponding unsteady aerodynamic moment, pitch angle and pitch angular velocity.
[0132] Step 2: Offline design of active control scheme based on reinforcement learning and fast response stall flutter model;
[0133] The pitching motion of the airfoil under stall flutter can be effectively suppressed by the corresponding trailing edge deflection, and a closed-loop control system is established for the airfoil with a variable trailing edge structure, such as Figure 5 The control system gives the corresponding effective trailing edge deflection by observing the motion parameters of the airfoil under stall flutter, such as the pitch angle and pitch angular velocity.
[0134] The control system is designed using a reinforcement learning algorithm. An environment is established based on a fast-response aeroelastic model, which can quickly output the pitch motion response of the airfoil for any trailing edge deflection input; the reinforcement learning agent is a fully connected neural network with two hidden layers, and the agent's optional actions are the flexible trailing edge deflection directions, which are respectively large upward deflection Δβ=Δβ l , large downward deviation Δβ=-Δβ l , a small upward deviation Δβ=Δβ s , slightly downward deviation Δβ = -Δβ s , keep still Δβ=0, where the large deflection is 0.05rad and the small deflection is 0.03rad, both corresponding to the deflection increment under the current deflection amplitude;
[0135] The environmental variables observed by the agent are pitch angle, pitch angular velocity and current trailing edge deflection The agent determines the next action based on the current observation and uses it as a control signal to control the trailing edge deflection, thus forming a closed-loop active control.
[0136] Considering the control capability and pitch limitation of the actual mechanical structure, this embodiment sets 0.32 rad as the maximum amplitude of the trailing edge deflection. During the training process, the reward item is set as:
[0137]
[0138] Among them, R t It is the reward for the previous action based on the current observation, which is divided into three categories:
[0139] (xi) The trailing edge deflection is within the limit (0.32 rad), and the angle of attack and angular velocity are both very small ( θ t <0.1rad), it is considered close to stillness and rewards the previous action;
[0140] (xii) The trailing edge deflection is within the limit (0.32 rad), but does not reach a near-static suppression effect ( or θ t ≥0.1rad), the penalty amplitude for the previous action is the pitch angle ω t With angular velocity θ t Linear combinations of
[0141] (xiii) If the trailing edge deflection exceeds the allowed amplitude (0.32 rad), a large penalty will be imposed on the previous action regardless of whether the suppression effect is achieved.
[0142] Step 3: Online verification of active control scheme for high-precision fluid-solid control based on the combination of reinforcement learning and CFD;
[0143] like Figure 6 As shown in the figure, a fluid-solid control coupling calculation framework is established in the high-precision CFD environment, and the effect of the closed-loop control system designed by DQN is verified through high-fidelity fluid-solid coupling calculation.
[0144] The fourth-order Runge-Kutta method is written through the user-defined function UDF to solve the aeroelastic control equations. In the UDF, the coordinate calculation of the control points on the airfoil surface of the compound motion is performed to control the mesh deformation. The overlapping mesh technology is used to improve the calculation accuracy and efficiency and prevent the negative volume of the mesh. The trained agent is inserted into the UDF to form a closed-loop control system to verify the stall flutter suppression effect.
[0145] The invention is a cyclic process based on time sequence. Through the trailing edge camber at the current moment, the NARX neural network is used to perform aerodynamic reduction on the pitch and trailing edge deflection motion under a given composite sinusoidal signal, and the wing is deformed by composite motion. Based on the aerodynamic reduction model, a controlled rapid response model of the airfoil stall flutter process is established, and the flutter response (i.e., the stall flutter model f in an offline environment) is calculated. SF , the CFD in the online environment performs transient flow field calculation, outputs the aerodynamic force M on the airfoil surface corresponding to the current time t, and substitutes it into the aeroelastic control equation), and the DQN reinforcement learning method is used to train the neural network agent to join the closed loop. The pitch angle, angular velocity and trailing edge deflection signal are used as input feedback signals. The trailing edge deformation signal at the next moment is given through the designed neural network agent, thereby forming an airfoil stall flutter closed-loop active control; after several moments, the pitch angle and angular velocity are close to 0, that is, the flutter is suppressed.
Claims
1. A flutter suppression design and verification method based on reinforcement learning and variable camber trailing edge, characterized in that: The specific steps are as follows: Step 1: For the airfoil model with pitch freedom and active continuous trailing edge deflection, perform a compound motion of pitch and trailing edge deflection in the stall flutter motion under active control; Step 2: Establish a fast response stall flutter model of a variable camber trailing edge airfoil including an unsteady aerodynamic moment and an aeroelastic system; The modeling process is: Firstly, for the current time t, the pitching motion angle θ and the trailing edge deflection β of the airfoil model are established by means of sine superposition; Among them, A i With ω i is the randomly generated sine function amplitude and angular velocity; Then, using CFD calculation, the unsteady aerodynamic moments corresponding to the pitch motion and the trailing edge deflection are obtained, namely: M pit =f CFD (θ,t) (2) M tem =f CFD (β,t) After that, the time domain incoming flow velocity U ∞ (t), pitch angle, trailing edge deflection and their respective unsteady aerodynamic moments as training sets; specifically: Training set 1: M pit (t),θ(t),U ∞ (t)(3) Training set 2: M tem (t),β(t),U ∞ (t) Next, two NARX neural networks are used to fit the unsteady aerodynamic moments under the two motions by error back propagation, and two reduced-order models are obtained: Reduced-order model 1: M pit =f ROM1 (U ∞ ,θ,t)(4) Reduced-order model 2: M tem =f ROM2 (U ∞ ,β,t) Finally, the sum of the unsteady aerodynamic moments of the trailing edge deflection and large pitch compound motion of the two reduced-order models is calculated and combined with the ordinary differential equations of aeroelastic control: Where, I is the airfoil moment of inertia, C is the airfoil pitch motion damping, K is the airfoil pitch motion stiffness, is the angular acceleration of pitch motion, is the pitch motion angular velocity; The fourth-order Runge-Kutta method is used to solve formula (6) and the pitch angle θ is obtained. t and pitch angular velocity Combined with the reduced-order model, the stall flutter model f at the current time t is obtained SF : That is, input any time sequence of trailing edge deflection to achieve rapid response of the airfoil pitch motion and obtain the corresponding unsteady aerodynamic moment, pitch angle and pitch angular velocity; Step 3: Design a fast response stall flutter model based on reinforcement learning DQN SF Offline closed-loop control scheme; The reward function of the DQN algorithm consists of a reward term and a penalty term; The rewards are set as: Among them, R t It is the reward for the previous action based on the current observation, which is divided into three categories: (i) The trailing edge deflection is within the limit, and the angle of attack and angular velocity meet the requirements. θ t <0.1rad, it is considered close to stillness and rewards the previous action; (ii) The trailing edge deflection is within the limit but does not reach rad / s or θ t ≥0.1rad, the penalty for the previous action is the pitch angle ω t With angular velocity θ t Linear combinations of (iii) If the trailing edge deflection exceeds the allowed amplitude, a large penalty will be imposed on the previous action regardless of whether the suppression effect is achieved; The reinforcement learning agent is a fully connected neural network with 2 hidden layers, and the input is the environment information: the angle of attack θ t , pitch angle. Speed θ t , trailing edge deflection β t , the output is five optional actions Δβ=[Δβ l ,-Δβ l ,-Δβ s ,Δβ s ,0], select the action corresponding to the maximum Q value, and the trailing edge deflection β at the current moment t+1 =β t +Δβ is used as the controller output to control the next trailing edge deflection, forming a closed-loop active control in an offline environment; Step 4: Online verification of the closed-loop active control scheme for high-precision fluid-solid control based on the combination of reinforcement learning and CFD.
2. The flutter suppression design and verification method based on reinforcement learning and variable camber trailing edge according to claim 1, characterized in that: In the step 1, the compound motion sets the absolute coordinate system and the local coordinate system; The absolute coordinate system takes the leading edge of the static airfoil as the origin, the incoming flow direction as the x direction, and the vertical upward direction as the y direction; the local coordinate system takes the leading edge of the airfoil as the origin, and the chord direction of the airfoil as the x direction. b Direction, the vertical chord is upward as y b direction; The thickness of the airfoil trailing edge remains unchanged during deformation, and its governing equation is: x=x b (8) y=y b +d Among them, x, y are the absolute coordinates of the control points on the airfoil surface, x b ,y b is the local coordinate of the control point on the airfoil surface, and d is the deflection of the control point on the chord line.
3. The flutter suppression design and verification method based on reinforcement learning and variable camber trailing edge according to claim 1, characterized in that: In step 2, the input of the reduced-order model 1 is the angle of attack θ=[θ t ,θ t-1 ,...,θ t-10 ] and the incoming flow velocity U ∞ =[U ∞,t ,U ∞,t-1 ,...,U ∞,t-10 ], the output is the unsteady aerodynamic moment M of the large pitch motion at the current time step pit ; The input of the reduced-order model 2 is the trailing edge deflection β = [β t ,β t-1 ,...,β t-10 ] and the incoming flow velocity U ∞ =[U ∞,t ,U ∞,t-1 ,...,U ∞,t-10 ], the output is the unsteady aerodynamic moment M of the airfoil trailing edge deflection motion at the current time step tem .
4. The flutter suppression design and verification method based on reinforcement learning and variable camber trailing edge according to claim 1, characterized in that: In step 3, the offline control scheme refers to: using the established fast response stall flutter model f SF For the aeroelastic simulation environment, a fully connected neural network is trained using reinforcement learning algorithm as a control agent to form a closed-loop active control system. By observing the motion parameters of the airfoil pitch angle and pitch angular velocity under stall flutter, the corresponding effective trailing edge deflection is given.
5. The flutter suppression design and verification method based on reinforcement learning and variable camber trailing edge according to claim 1, characterized in that: In step 4, online verification of the closed-loop active control scheme refers to: Firstly, the fully connected neural network trained offline is integrated into CFD, and the aeroelastic simulation environment is set up with high-fidelity fluid-structure interaction calculation. The fully connected neural network is used as the control agent to form a closed-loop active control system. By observing the motion parameters of the airfoil under stall flutter, such as the pitch angle and pitch angular velocity, the corresponding effective trailing edge deflection is given. The main steps are: (i) CFD performs transient flow field calculations and outputs the aerodynamic force M on the airfoil surface corresponding to the current time t, which is substituted into the aeroelastic control equation (5); (ii) Write the fourth-order Runge-Kutta method in the UDF to solve the aeroelasticity governing equation (5) corresponding to the current time t, and input. Pitch motion angle θ t With angular velocity θ t ; (iii) A fully connected neural network agent is embedded using a custom function UDF; the neural network agent is based on the loop at the current time t. Environmental information [θ t ,θ t ,β t ] is the feedback signal, outputting five optional actions Δβ=[Δβ l ,-Δβ l ,-Δβ s ,Δβ s ,0] corresponding to the Q value; The trailing edge of the variable camber deviates greatly upward Δβ=Δβ l , large downward deviation Δβ=-Δβ l , a small upward deviation Δβ=Δβ s , slightly downward deviation Δβ = -Δβ s , keep still Δβ=0; (iv) Filter the maximum Q value as the next step trailing edge deflection increment, and add the trailing edge deflection at the current moment as the controller output β t+1 =β t +Δβ; (v) for pitch θ t With trailing edge deflection β t+1 The composite motion airfoil surface control points perform coordinate calculations to control the structural shape and fluid grid deformation, forming a closed-loop active control in a high-fidelity environment.
Citation Information
Patent Citations
Nonlinear large-attack-angle aerodynamic force order reduction model applicable to stall flutter
CN108182328A
Aerofoil flutter loss reduction method of hypersonic flight vehicle based on prediction control, and loss reduction controller
CN108846246A
Wing stall flutter closed-loop control method for continuous variable camber trailing edge
CN112948973A
Large-attack-angle flutter analysis method based on reduced-order modeling
CN113221237A
Active control surface modal system for aircraft buffet and gust load alleviation and flutter suppression
US6375127B1
