Wing large-incidence-angle flow separation combination intelligent control method based on wind tunnel experiment

Abstract

The invention provides a wind tunnel experiment-based wing high-attack-angle flow separation combined intelligent control method, which comprises the following steps of: arranging active flow control devices at 10% chord length and trailing edge flap of a wing to inhibit wing high-attack-angle flow separation; the active flow control device at the 10% chord length of the wing adopts direct jet flow, the jet flow direction is perpendicular to the chord line of the wing, three jet flow exciters are uniformly arranged along the spanwise direction, the active flow control device at the trailing edge flap adopts sweeping jet flow, the jet flow direction is parallel to the surface of the wing at the position, and 18 sweeping jet flow actuators are uniformly arranged along the spanwise direction; the jet flow of the two sets of flow control devices is adjusted through corresponding mass flow controllers, jet flow momentum distribution is dynamically optimized in combination with a reinforcement learning algorithm, the lift coefficient is remarkably increased, and energy consumption is reduced. The flow field state is fed back in real time through a pressure measuring system and a force measuring system, a closed-loop control system is constructed, and an execution action is output according to the flow field information around the wing to control an active flow control device arranged on the wing so as to inhibit flow separation on the surface of the wing, so that the wing gets rid of the stall state and the lift force is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of active flow control, and particularly relates to a wing large angle of attack flow separation combined intelligent control method based on a wind tunnel experiment. BACKGROUND

[0002] Due to the high dimension and strong nonlinearity of fluid mechanics, it is often very tedious to design an active flow control strategy using a traditional control method, and a large amount of trial and error and experience accumulation are required.

[0003] The traditional flow control method has many significant limitations in dealing with complex flow fields, especially the flow separation problem at a large angle of attack. First, in the control strategy layer, the open-loop control mode (such as a fixed frequency plasma excitation) lacks a real-time feedback mechanism and cannot adaptively respond to the dynamically changing flow field state, resulting in rigid control, poor robustness, and being easily affected by external disturbances. Secondly, relying on a single type of actuator (such as using only synthetic jets), its control ability is often limited, it is difficult to effectively suppress complex flow structures such as strong separation vortices, and it may face the problem of high energy consumption. Thirdly, the controller has too strong dependence on numerical models, and the design based on computational fluid dynamics (CFD) highly depends on accurate turbulence models (such as RANS), however, under complex working conditions such as a large angle of attack, these models often have significant errors, which seriously affect the reliability of the controller. In addition, the existing intelligent control methods also face severe challenges: the high-precision wind tunnel experiment is costly, resulting in a lack of sample data available for training intelligent algorithms such as deep learning, and the model generalization ability is insufficient; the closed-loop control combining traditional CFD and machine learning has a millisecond-level delay, which is difficult to match the real-time change requirements of high-speed flow. In summary, the single actuator and open-loop control strategy commonly used by traditional methods are difficult to achieve efficient, stable and adaptive flow control effect due to their insufficient regulation ability, lack of adaptability and strong model dependence. SUMMARY

[0004] The purpose of the present application is to solve the above technical problems, and provide a wing large angle of attack flow separation combined control method based on deep reinforcement learning.

[0005] The technical solution adopted by the present application is as follows: The present application provides a wing large angle of attack flow separation combined intelligent control method based on a wind tunnel experiment, active flow control devices are arranged at the leading edge and trailing edge flaps of the wing, and jet excitation is output according to the instantaneous system state of the wing and the surrounding flow field; the excitation intensity of the flow control device is controlled by an artificial neural network, and is obtained by training as follows: Step S1, constructing an agent containing an active flow control device and a wind tunnel experiment environment; Step S2, taking the flat wing equipped with 30% chord trailing edge flap as the training object, the agent is trained in the form of shallow stall condition of the wing without control opportunity in each new round; 150 time steps are set in each round; Step S3, training the agent by using the proximal policy optimization algorithm, including training of the artificial neural network of the decision maker Actor of the agent and training of the artificial neural network of the evaluator Critic; Step S4, when the agent performs active flow control in the wind tunnel experiment environment, the control trajectory is stored in the experience pool of the deep reinforcement learning algorithm; Step S5, if the current time reaches the maximum time step set in step S2, the current round is ended, otherwise, the step S3 is returned to continue training and perform the current flow control process; Step S6, if the control strategy represented by the agent has converged, the deep reinforcement learning training process is ended, that is, the optimal flow control strategy is obtained; otherwise, the step S2 is returned to start a new training round until the optimal flow control strategy is obtained.

[0006] Specifically, the experimental model used in the wind tunnel experiment is a NACA0015 flat wing, the trailing edge flap accounts for 30% of the chord length, the flap upper end is fixed with a rudder arm, the flap deflection angle is adjusted by adjusting the fixed hole position of the rudder arm on the front wing, there are 0°, 5°, 10°, 15°, 20°, 25°, 30°, a total of 7 flap deflection angle gears; and a transition band is arranged on the upper surface of the model about 5% chord length of the front edge to promote transition; the transition band is arranged by double-sided tape and green silicon carbide abrasive.

[0007] Specifically, the agent in step S1 includes a decision maker Actor and an evaluator Critic, wherein the decision maker is used to output a control action according to the instantaneous system state of the wing and the surrounding flow field at time t Give flow control action The evaluator is used to output a value estimate according to the wing and the surrounding flow field, that is, the system state Take the current action Good or bad evaluation to guide the decision maker to optimize.

[0008] Specifically, the step S1 active flow control device is two groups, one group is arranged at 10% chord length of the wing leading edge, for direct jet, jet outlet slit width 1mm, jet direction perpendicular to the wing chord, along the span uniform distribution of three jet actuators, the total length of the slit 570mm; another group is arranged upstream of the flap, for sweeping jet, jet outlet position at 70% chord length, jet outlet height 1mm, jet direction parallel to the wing surface at the position, along the span uniform distribution of 18 sweeping jet actuators, actuator spacing 33mm; two groups of flow control device are adjusted by corresponding mass flow controller jet flow; the wing lift and surface pressure information are obtained by using force sensor and pressure sensor, and are transmitted to the intelligent agent in real time.

[0009] Specifically, the artificial neural network of the decision maker Actor of the intelligent agent in the step S3, the network structure contains two layers of fully connected hidden layers, each layer has 256 neural units, the activation function is ReLU function, the input flow field state , the average value of the output action and the standard deviation , and the calculation expression of the control action output by the intelligent agent is: ; Wherein, ; the artificial neural network structure of the evaluator Critic contains two layers of fully connected hidden layers, each layer has 256 neural units, the activation function is ReLU function, the input flow field state and the execution action, and the output is the goodness of the current state . The neural network is optimized and learned based on the Adaptive Moment Estimation algorithm.

[0010] Specifically, the objective function of the artificial neural network training of the decision maker Actor in the step S3 is as follows: ; Wherein, represents the parameters of the artificial neural network of the decision maker, represents the advantage function at time t, clip is a clipping function, after exceeding the specified upper and lower limits, the function will return the corresponding upper limit or lower limit, is a hyperparameter, taking 0.2.

[0011] Specifically, the objective function of the artificial neural network training of the evaluator Critic is as follows: ; Wherein, represents the current control strategy of the decision maker, represents the transition probability distribution corresponding to the current control policy, represents the parameters of the evaluator's artificial neural network, is the expectation, represents the state-action evaluation value given by the current evaluator artificial neural network, represents a relatively more accurate evaluation value: ; wherein, is the policy corresponding to the entropy at the current observation state, is the weight of the action entropy in the evaluation function, is the influence factor of the future evaluation on the current evaluation; represents all the actions that can be used in the current state.

[0012] Further, the step S4 is specifically: the wind tunnel experiment environment is used for carrying out the active flow control process of the agent, the jet flow of the active flow control device on the wing is changed according to the control action, the flow field change is observed by using the pressure sensor on the wing surface and the force sensor at the bottom of the wing to obtain the flow field instantaneous state at the moment, the control trajectory is stored into the experience pool of the deep reinforcement learning algorithm; the flow field instantaneous state includes the pressure coefficients at 20% and 90% of the chord length of the wing surface and the jet flow of the two groups of flow control devices; the reward and punishment function is represented as follows: ; wherein, represents the wing lift coefficient at t moment, represents the initial state lift coefficient of the wing, represents the jet flow of the leading edge flow control device, represents the jet flow of the flow control device at the flap, and are corresponding weights.

[0013] The beneficial effects of the present application are as follows: 1. The active flow control device of the present application solves the problem of complicated control strategy design in the traditional control mode, directly establishes an end-to-end control strategy, and realizes the suppression of wing flow separation.

[0014] 2. The present application proposes the idea of combining the leading edge jet and the sweeping jet for wing flow control, and builds a closed-loop control system based on deep reinforcement learning, which can achieve better flow control effect and less energy consumption compared with single control means.

[0015] 3. The active flow control method of the present application can be applied to active flow control of a wing. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 Flow chart for a round of reinforcement learning of the active flow control of the present application; Figure 2 Schematic diagram of a NACA0015 flat wing model; Figure 3 Schematic diagram of a cross section of a NACA0015 flat wing model; Figure 4 Final control strategy of the reinforcement learning agent and corresponding flow field information change graph; Figure 5 Wing lift coefficient change graph when the leading edge jet excitation timing of different momentum coefficients; Figure 6 Fluorescent microfilament state change graph of the wing surface during the flow control process of the agent. DETAILED DESCRIPTION

[0017] The present application will be described in detail below with reference to the accompanying drawings.

[0018] The present application provides a wing high angle of attack flow separation combined intelligent control method based on wind tunnel experiment, active flow control devices are arranged at the wing leading edge and flaps, a closed-loop control system is constructed based on deep reinforcement learning, as shown in Figure 1 The active flow control devices arranged on the wing are controlled to suppress wing surface flow separation, make the wing get out of the stall state and improve the lift according to the flow field information around the wing. The excitation strength of the flow control device is controlled by an artificial neural network, which is obtained by training as follows: An agent containing active flow control devices and a wind tunnel experiment environment are constructed; wherein the wind tunnel experiment environment is used for the active flow control process of the agent: the jet flow of the active flow control device on the wing is changed according to the control action, and the flow field change is observed by using a sensor to obtain the flow field instantaneous state at time t The control trajectory is stored in the experience pool of the deep reinforcement learning algorithm; the agent contains a decision maker Actor and an evaluator Critic, wherein the decision maker is used to give flow control action according to the flow field instantaneous system state of the wing and its surroundings at time t The evaluator is used to output the system state of the wing and its surroundings to evaluate the pros and cons of the current action to guide the decision maker to optimize; the flow field instantaneous state ​The pressure coefficient at 20% and 90% chord length of the airfoil surface and the jet flow rate of the two sets of flow control devices; the reward-punishment function is expressed as follows: ; wherein, represents the lift coefficient of the airfoil at time t, represents the initial state lift coefficient of the airfoil, represents the jet flow rate of the leading edge flow control device, represents the jet flow rate of the flow control device at the flap, and are the corresponding weights; As an embodiment, the artificial neural network structure of the decision maker in the present embodiment comprises two layers of fully connected hidden layers, each layer having 256 neural units, and the activation function is a ReLU function, the input flow field state , and the average value and the standard deviation of the action are outputted, and the control action outputted by the intelligent agent is calculated as: ; wherein, ; The artificial neural network structure of the evaluator comprises two layers of fully connected hidden layers, each layer having 256 neural units, and the activation function is a ReLU function, the input flow field state and the executed action are inputted, and the goodness of the current state is outputted. .

[0019] The experimental model used is a NACA0015 flat airfoil, as shown in Figure 2 , the chord length l = 305 mm, the span length b = 620 mm, the trailing edge flap accounts for 30% of the chord length, the flap upper end is fixed with a rudder arm, the flap deflection angle is adjusted by adjusting the fixed hole position of the rudder arm on the front airfoil, there are 0°, 5°, 10°, 15°, 20°, 25°, 30°, a total of 7 flap deflection angle gears. A transition strip with a height of about 0.2 mm is arranged on the upper surface of the model leading edge about 5% of the chord length to promote transition. The transition strip is arranged by double-sided tape with a width of about 2 mm and green silicon carbide abrasive.

[0020] There are two sets of active flow control devices, as shown in Figure 3As shown, one group is arranged at 10% chord length of the wing leading edge, as direct jet, jet outlet slit width 1 mm, jet direction perpendicular to the wing chord line, along the span direction uniform arrangement of three jet actuators, the total length of the slit is 570 mm; another group is arranged upstream of the flap, as a sweeping jet, the jet outlet position is located at 70% chord length, the jet outlet height is 1 mm, the jet direction is parallel to the wing surface at the position, 18 sweeping jet actuators are uniformly arranged along the span direction, and the actuator spacing is 33 mm. Both groups of flow control devices are adjusted by corresponding mass flow controllers. In the experiment, the force sensor and the pressure sensor are used to obtain the wing lift and surface pressure information, and the information is transmitted to the agent in real time.

[0021] The training process of the agent starts in the form of rounds, as shown in Figure 1 , each new round starts from the shallow stall condition of the wing without control. The NACA0015 flat wing with a 30% chord length trailing edge flap is used as the research object, the initial wing angle of attack is 15°, the flap deflection angle is 10°, the Reynolds number , and the wing is in a stall state.

[0022] Start training, use the proximal policy optimization (PPO) algorithm to train the agent. The agent includes decision maker Actor and evaluator Critic, both of which are represented by neural networks. Among them, the objective function of the decision maker Actor artificial neural network is as follows: ; Where, represents the parameters of the decision maker's artificial neural network, represents the advantage function at time t, and clip is the clipping function, which will return the corresponding upper limit or lower limit when the function exceeds the specified upper and lower limits, is a hyperparameter, generally taken as 0.2; the objective function of the evaluator's artificial neural network is as follows: ; Where, represents the current control strategy of the decision maker, represents the transition probability distribution corresponding to the current control strategy, represents the parameters of the evaluator's artificial neural network, is the expectation, represents the state-action evaluation value given by the current evaluator artificial neural network, represents a relatively more accurate evaluation value: ; Where, is the entropy corresponding to the strategy under the current observation state, The weight of the action entropy in the evaluation function. This refers to the impact factors of future evaluation items on the current evaluation; This represents all possible actions that can be taken in the current state; The active flow control process of the agent is conducted in a wind tunnel experimental environment, and the control trajectory is stored in the experience pool of a deep reinforcement learning algorithm. The agent discretizes the continuous control process into a Markov decision process and observes the instantaneous state of the flow field at time t. The decision-maker, the Actor, bases its decisions on the instantaneous state at time t. Give control actions The mass flow controller (MFC) changes the jet flow rate of the two sets of flow control devices according to the control action, and obtains the flow rate using pressure sensors on the wing surface and force sensors on the wing bottom. Instantaneous state of the flow field at time t and the reward / penalty functions corresponding to the control actions. The evaluator, Critic, provides a value estimate and uses it to estimate the advantage function. Finally, the AdaptiveMoment Estimation algorithm is used to update the artificial neural networks of the decision-maker and evaluator.

[0023] If the current time reaches the set value, that is, the maximum time step of the round set in step S2, then the current round ends; otherwise, training continues and the current flow control process is carried out.

[0024] If the control policy of the agent's representation has converged, the deep reinforcement learning training process ends; otherwise, it returns to start a new training round.

[0025] Figure 4 To reinforce the learning agent's final control strategy and the corresponding flow field information change diagram. Figure 4 Figure (a) shows the jet flow rate changes of the two sets of flow control devices; Figure 4 Figure (b) shows the variation of the pressure coefficient on the wing surface. The pressure coefficient at 20% of the chord length. The pressure coefficient is 90% of the chord length; Figure 4 Figure (c) shows the wing lift coefficient increment. The changes. In the first 20 steps, It rapidly increased, reaching approximately 160 SLM, corresponding to a jet momentum coefficient of 0.068%, while the number of agents also slowly increased. It reaches approximately 80 SLM. Around step 20, due to the excitation of the leading-edge jet, the pressure coefficient changes significantly, such as... Figure 4 Figure (b) shows the characteristics before stalling. Correspondingly, There has also been a significant increase, such as Figure 4As shown in Figure (c), it is approximately 0.33. Following this, the agent adjusted its blowing strategy, significantly increasing... gradually decrease The pressure coefficient fluctuated somewhat, but eventually stabilized. ≈-1.55、 ≈0. It also increased further to approximately 0.5, with the lift coefficient increasing by 72.6%. To demonstrate the advantages of the combined control method, tests were conducted on individual leading-edge jets, such as... Figure 5 The figure shows the changes in the lift coefficient of the wing under different momentum coefficients of leading-edge jet excitation. The single leading-edge jet can only increase the lift coefficient by up to 38%, indicating that by using two sets of AFC devices in combination with reinforcement learning agents, greater lift benefits can be obtained.

[0026] Figure 6 This image shows the state changes of fluorescent microfilaments on the wing surface during intelligent agent flow control. Figure 6 Figure (a) shows the initial stall state. Figure 6 Figure (b) shows the reattachment state of the airflow at the front section of the wing. Figure 6 Figure (c) shows the final stable state. Initially, the wing is in a stall state, and all the fluorescent microfilaments on the upper surface of the wing exhibit significant oscillation, with some even appearing "hooked," indicating that the flow separation point on the upper surface of the wing is located at the leading edge (within the first 20% of the chord length). As the leading edge jet excitation intensifies, as... Figure 6 As shown in Figure (b), the three fluorescent microfilaments on the leading edge of the wing are attached to the wing surface without significant oscillation, indicating that the airflow has reattached to the wing surface and the wing has escaped the stall state. However, some fluorescent microfilaments on the flaps are still oscillating, indicating that flow separation still exists on the flaps. As the agent increases the flow rate of the swept jet upstream of the flaps, the oscillation amplitude of the fluorescent microfilaments on the flaps continuously weakens, eventually reaching a stable state, as shown in Figure (b). Figure 6 Figure 6 As shown in Figure (c), all fluorescent microfilaments are completely attached to the wing surface, thus suppressing flow separation.

[0027] This invention is a combined flow control method that can also be applied to other models, but the agent needs to be retrained using the method in this invention, and the control effect may also change.

[0028] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A combined intelligent control method for flow separation at high angles of attack on an airfoil based on wind tunnel experiments, characterized in that, Active flow control devices are installed at the leading and trailing edge flaps of the wing, respectively, to output jet excitation based on the instantaneous system state of the flow field in and around the wing; the excitation intensity of the flow control devices is controlled by an artificial neural network, which is trained and obtained through the following method: Step S1: Construct an intelligent agent and wind tunnel experimental environment that includes an active flow control device; Step S2: Using a straight wing with 30% trailing edge flaps as the training object, the agent is trained in each new round under the condition of shallow stall of the wing without control; each round is set to contain 150 time steps. Step S3: Train the agent using a proximal policy optimization algorithm, including training the artificial neural network of the agent's decision-maker (Actor) and the artificial neural network of the evaluator (Critic). Step S4: When performing active flow control of the agent in the wind tunnel experimental environment, the control trajectory is stored in the experience pool of the deep reinforcement learning algorithm. Step S5: If the current time reaches the maximum time step of the round set in step S2, then end the current round; otherwise, return to step S3 to continue training and perform the current flow control process. Step S6: If the control strategy represented by the agent has converged, end the deep reinforcement learning training process to obtain the optimal flow control strategy. Otherwise, return to step S2 to begin deriving a new training round until the optimal flow control strategy is obtained.

2. The intelligent control method for airfoil high angle-of-attack flow separation based on wind tunnel experiments as described in claim 1, characterized in that, The wind tunnel experiment used a NACA0015 straight wing model with trailing edge flaps accounting for 30% of the chord length. A rudder arm is fixed to the upper end of the flap, and the flap deflection angle is adjusted by adjusting the fixing hole position of the rudder arm on the front section of the wing. There are 7 flap deflection angle positions: 0°, 5°, 10°, 15°, 20°, 25°, and 30°. A transition zone is set on the upper surface of the model at about 5% of the chord length at the leading edge to promote transition. The transition zone is made of double-sided tape and green silicon carbide diamond grit.

3. The intelligent control method for airfoil high angle-of-attack flow separation based on wind tunnel experiments as described in claim 1, characterized in that, In step S1, the intelligent agent includes a decision-maker (Actor) and an evaluator (Critic), wherein the decision-maker is used to determine the instantaneous system state of the wing and surrounding flow field at time t. Give flow control actions The evaluator uses it to estimate the output value based on the flow field in and around the wing, i.e., the system state. Take the current action The evaluation of its merits and demerits is used to guide decision-makers in making optimizations.

4. The intelligent control method for airfoil high angle-of-attack flow separation based on wind tunnel experiments as described in claim 1, characterized in that, In step S1, there are two sets of active flow control devices. One set is located at 10% chord length of the wing leading edge, providing a direct jet with a 1mm wide jet outlet slit and a jet direction perpendicular to the wing chord. Three jet actuators are evenly arranged along the spanwise direction, with a total slit length of 570mm. The other set is located upstream of the flap, providing a swept jet with a jet outlet at 70% chord length and a 1mm high outlet. The jet direction is parallel to the wing surface at its location, and 18 swept jet actuators are evenly arranged along the spanwise direction with a spacing of 33mm between the actuators. Both sets of flow control devices are regulated by corresponding mass flow controllers to adjust the jet flow rate. Force and pressure sensors are used to acquire wing lift and surface pressure information, which is transmitted to the intelligent agent in real time.

5. The intelligent control method for airfoil high angle-of-attack flow separation based on wind tunnel experiments as described in claim 1, characterized in that, The artificial neural network of the decision-maker (Actor) in step S3 has a network structure consisting of two fully connected hidden layers, each with 256 neurons, and the activation function is the ReLU function. The input is the flow field state. The average value of the output action and standard deviation The final control action output by the intelligent agent The calculation expression is: ； in, The Critic's artificial neural network structure consists of two fully connected hidden layers, each with 256 neurons, and uses the ReLU activation function. The input is the flow field state. And execute actions, outputting the current state's condition (good or bad). The neural network is optimized based on the Adaptive Moment Estimation algorithm.

6. The intelligent control method for airfoil high angle-of-attack flow separation based on wind tunnel experiments as described in claim 1, characterized in that, The objective function for training the artificial neural network by the decision-maker (Actor) in step S3 is as follows: ； in, The parameters of the artificial neural network representing the decision-maker, Let represent the dominance function at time t, and 'clip' be the cutoff function. Once the specified upper or lower limit is exceeded, the function will return the corresponding upper or lower limit. This is a hyperparameter, set to 0.

2.

7. The intelligent control method for airfoil high angle-of-attack flow separation based on wind tunnel experiments as described in claim 1, characterized in that, The objective function for training the artificial neural network by the evaluator Critic is as follows: ； in, This indicates the current decision-maker's control strategy. This represents the transition probability distribution corresponding to the current control strategy. The parameters of the artificial neural network representing the evaluator, It is an expectation. This represents the state-action evaluation value given by the current evaluator's artificial neural network. This represents a more precise assessment value: ； in, The strategy under the current observation state The corresponding entropy, The weight of the action entropy in the evaluation function. This refers to the impact factors of future evaluation items on the current evaluation; This represents all possible actions that can be taken in the current state.

8. The intelligent control method for airfoil high angle-of-attack flow separation based on wind tunnel experiments as described in claim 1, characterized in that, Step S4 specifically involves: using a wind tunnel environment to conduct the active flow control process of the intelligent agent. This involves changing the jet flow rate of the active flow control device on the wing based on control actions, and observing the flow field changes using pressure sensors on the wing surface and force sensors at the wing bottom. Instantaneous state of the flow field at time 1 The control trajectory is stored in the experience pool of the deep reinforcement learning algorithm; Instantaneous state of the flow field Includes pressure coefficients at 20% and 90% chord length of the wing surface and jet flow rates from two sets of flow control devices; reward / penalty function. It is expressed as follows: ； in, This represents the wing lift coefficient at time t. This represents the initial lift coefficient of the wing. This indicates the jet flow rate of the leading edge flow control device. This indicates the jet flow rate of the flow control device at the flap. and For the corresponding weights.

Images