Intelligent aircraft turning regulation and control system

Abstract

The invention belongs to the technical field of aviation flight control, and particularly discloses an intelligent aircraft turning regulation and control system, which comprises a data acquisition and preprocessing module for collecting flight data and preprocessing the collected flight data; the DQN intelligent optimization module outputs an optimal turning action a based on the aircraft state s to realize an optimal turning strategy; and importing the DQN intelligent optimization module into airborne equipment, adjusting control parameters of the airplane by the DQN intelligent optimization module, and controlling the airplane to complete turning action. According to the method, a real-time optimization model is established by utilizing a Deep Q-Network (DQN) deep reinforcement learning algorithm, the model can adaptively calculate the optimal turning angle, gradient and turning rate according to multivariate information such as current state parameters, environmental factors and air route planning targets of the aircraft, and safer and more efficient autonomous turning operation of the aircraft in a complex environment is realized.

Description

Technical Field

[0001] The invention belongs to the technical field of aviation flight control, and in particular relates to an intelligent aircraft turning control system. Background Art

[0002] Aircraft turning operations involve the coordinated work of multiple components and systems. For example, when using ailerons to tilt and turn the aircraft, coordinated operations need to be performed through the rudder. Because when the aircraft tilts, a yaw moment will be generated. If the rudder is not used for coordination, the aircraft will not only turn, but also skid. In modern aircraft, the autopilot system can effectively assist the pilot in completing the turning operation. It can coordinate the ailerons, rudder, elevator and other control surfaces according to the set heading, turning angle and other parameters to perform precise turning operations. However, in complex weather conditions or special flight conditions, it is difficult to achieve accurate and safe turns, and individual differences in pilot control skills will also affect turning efficiency and safety. Summary of the invention

[0003] The purpose of the present invention is to overcome the defects in the prior art and provide an intelligent aircraft turning control system.

[0004] The present invention provides an intelligent aircraft turning control system, comprising:

[0005] The data acquisition and preprocessing module collects flight data from actual flight records, simulator simulations or wind tunnel experiments, and preprocesses the collected flight data; the flight data includes flight performance parameters under different speeds, altitudes, weight distributions, and meteorological conditions, as well as corresponding turning operation details. The preprocessing includes removing outliers, filling missing values, and converting the flight data format to make it suitable for deep learning models;

[0006] DQN intelligent optimization module, which outputs the best turning action a based on the aircraft state s to achieve the optimal turning strategy;

[0007] The DQN intelligent optimization module is introduced into the airborne equipment, and the DQN intelligent optimization module adjusts the control parameters of the aircraft to control the aircraft to complete the turning action.

[0008] A further solution is that the DQN intelligent optimization module includes:

[0009] Deep Q-Network model, adjusts the control parameters of the aircraft and controls the aircraft to complete the turning action;

[0010] The environment definition unit is used to define the flight environment state, and quantify and discretize the flight environment state into a finite state space S; S = {s_1, s_2, ..., s_n}, where s_n represents the state at time t; and define the action space A; A = {a_1, a_2, ..., a_m}, where a_m represents the action taken at time t;

[0011] The flight environment includes aircraft status and executable actions;

[0012] Reward function unit, constructs reward function R(s,a,s') to measure the effect of reaching the new state s' after executing action a in state s;

[0013] Q-value function unit, used to obtain the value evaluation Q(s,a) of action a in state s;

[0014] Experience replay buffer unit, used to store a large amount of experience (<s,a,r,s'> ), where s is the current state, a is the action taken, r is the immediate reward received, and s' is the next state after performing the action.

[0015] A further solution is that the Q-value function unit uses a deep neural network as an approximator of the Q-value function, the input of the deep neural network is the state s, and the output is the value evaluation Q(s,a) of the action a under the state s;

[0016] The network architecture of the deep neural network adopts a multi-layer perceptron, and the last layer corresponds to the value evaluation Q(s,a) of each action.

[0017] A further solution is that the Deep Q-Network model training process includes:

[0018] Explore the environment through the ε-greedy strategy, perform random actions or choose the action with the highest current Q value;

[0019] After executing the action, a new state s' and a corresponding reward r are obtained, and the experience is stored in the experience playback buffer unit;

[0020] A batch of data is sampled from the experience replay buffer unit for learning and the DQN network parameters are updated to narrow the gap between the value assessment Q(s,a) and the actual return.

[0021] A further solution is that the control system also includes: a dynamic bank turning control module for continuously and smoothly adjusting the bank of the aircraft body to reduce the turning radius and reduce overload;

[0022] The dynamic bank turning control module comprises:

[0023] The motion model building unit builds the motion model of the aircraft based on the principles of aircraft dynamics and aerodynamics. The motion model includes a six-degree-of-freedom motion state and physical quantities related to the six-degree-of-freedom motion state. The motion model equation is:

[0024] m*d^2x / dt^2=F_x+m*g*sin(θ)

[0025] m*d^2y / dt^2=F_y

[0026] I_z*d^2ψ / dt^2=L_z

[0027] Where m is the mass, I_z is the moment of inertia, (x,y) are the position coordinates, ψ is the roll angle, (F_x, F_y) are the components of the resultant force in the xy plane, L_z is the moment about the z axis, and θ is the pitch angle;

[0028] The indicator definition unit is used to define the performance indicator, and the quadratic form of the performance indicator J is:

[0029] J=∫[Q(x,u)+R(u)]dt

[0030] Where x is the state vector, u is the control input vector, Q(x,u) is the state cost function, and R(u) is the control input cost function;

[0031] A turning dynamic adjustment unit dynamically adjusts the aircraft's attitude and thrust distribution during the turning process based on the calculated control rate to achieve a banked turning action based on a preset performance index J;

[0032] The closed-loop control unit obtains the output result of the Deep Q-Network model, generates action instructions based on the output structure, and the airborne device receives and executes the instructions to form a feedback loop.

[0033] A further solution is that the control rate is a mathematical model for controlling input calculated based on the error, and the mathematical model is a PID controller; the error is a result of comparing the current aircraft state parameter with the target setting value.

[0034] A further solution is that the control system also includes a model real-time optimization module. Based on the closed-loop control principle, a closed-loop control system including multi-source sensor data fusion, error detection, real-time calculation and command output is constructed. The closed-loop control is used to monitor the deviation of the actual aircraft from the predetermined trajectory during the turning process and take compensation measures immediately;

[0035] The model real-time optimization module includes:

[0036] The data fusion unit integrates multi-source information through sensor fusion technology to generate flight state estimation;

[0037] A feedback unit obtains an error signal in real time, and immediately generates and sends a control instruction after obtaining the error signal to control the input of the Deep Q-Network model, thereby continuously adjusting the output of the Deep Q-Network model until it approaches or reaches the expected state;

[0038] A performance optimization unit records key performance indicators in real time, and uses the key indicators to evaluate the effect of the PID controller. Based on the evaluation results, an online self-correction algorithm or an offline parameter tuning method is used to continuously optimize the PID controller parameters to improve the control performance;

[0039] The security mechanism unit sets a security threshold. When the state variable exceeds the security threshold, an alarm is triggered or the system enters a safe mode.

[0040] A further solution is that the control system also includes a fault detection module, which is connected to the security mechanism unit. After the alarm is triggered or the safety mode is entered, the security mechanism unit sends a fault code to the fault detection module. The fault detection module identifies the fault code and switches to a backup system or takes emergency measures to ensure flight safety.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] The present invention uses the Deep Q-Network (DQN) deep reinforcement learning algorithm to establish a real-time optimization model, which can adaptively calculate the optimal turning angle, slope and turning rate based on multiple information such as the aircraft's current state parameters, environmental factors and route planning goals, thereby achieving safer and more efficient autonomous turning operations of the aircraft in complex environments.

[0043] The present invention adopts a control system using three-dimensional dynamic tilt technology, which can continuously and smoothly adjust the tilt of the aircraft body, reduce the turning radius and overload while ensuring flight stability and passenger comfort, and reduce the fuel consumption of the aircraft during the turning process, thereby reducing operating costs. A closed-loop control system including multi-source sensor data fusion, error detection, real-time calculation and command output is also constructed, which can monitor the deviation of the actual turning process of the aircraft from the predetermined trajectory in real time, and immediately take compensation measures to ensure that the aircraft can maintain the predetermined turning performance indicators at any stage of flight, thereby enhancing the intelligence and market competitiveness of the aircraft. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The following drawings are only used to illustrate and explain the present invention, and are not used to limit the scope of the present invention, wherein:

[0045] Figure 1 : Principle block diagram of the control system of the present invention;

[0046] In the figure: 1. Airborne equipment; 2. Data acquisition and preprocessing module; 3. DQN intelligent optimization module; 4. Dynamic bank turning control module; 5. Model real-time optimization module; 6. Fault detection module; 7. Environment definition unit; 8. Deep Q-Network model; 9. Reward function unit; 10. Q-value function unit; 11. Experience recovery buffer unit; 12. Deep neural network; 13. Motion model construction unit; 14. Closed-loop control unit; 15. Index definition unit; 16. Turning dynamic adjustment unit; 17. Data fusion unit; 18. Feedback unit; 19. Performance optimization unit; 20. Security mechanism unit. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution, design method and advantages of the present invention clearer, the present invention is further described in detail by specific embodiments in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0048] like Figure 1 As shown, the present invention provides an intelligent aircraft turning control system, which improves the aircraft turning efficiency, accuracy and fuel economy through innovative design, specifically including:

[0049] The data acquisition and preprocessing module collects flight data from actual flight records, simulator simulations or wind tunnel experiments, and preprocesses the collected flight data; the flight data includes flight performance parameters under different speeds, altitudes, weight distributions, and meteorological conditions, as well as corresponding turning operation details. The preprocessing includes removing outliers, filling missing values, and converting the flight data format to make it suitable for deep learning models;

[0050] DQN intelligent optimization module, which outputs the best turning action a based on the aircraft state s to achieve the optimal turning strategy;

[0051] The DQN intelligent optimization module is introduced into the airborne equipment, and the DQN intelligent optimization module adjusts the control parameters of the aircraft to control the aircraft to complete the turning action, wherein the DQN intelligent optimization module includes:

[0052] Deep Q-Network model, adjusts the control parameters of the aircraft and controls the aircraft to complete the turning action;

[0053] The environment definition unit is used to define the flight environment state, and quantify and discretize the flight environment state into a finite state space S; S = {s_1, s_2, ..., s_n}, where s_n represents the state at time t; and define the action space A; A = {a_1, a_2, ..., a_m}, where a_m represents the action taken at time t;

[0054] The flight environment includes aircraft status and executable actions, the aircraft status includes speed, altitude, attitude, etc., and the executable operations include control surface deflection angle, thrust adjustment amount, etc.

[0055] Reward function unit, constructs reward function R(s,a,s') to measure the effect of reaching the new state s' after executing action a in state s;

[0056] Q-value function unit, used to obtain the value evaluation Q(s,a) of action a in state s;

[0057] Experience replay buffer unit, used to store a large amount of experience (<s,a,r,s'> ), where s is the current state, a is the action taken, r is the immediate reward received, and s' is the next state after the action is performed. Some core pseudocodes are as follows:

[0058] #pseudocode

[0059] replay_buffer=deque(maxlen=replay_buffer_capacity)

[0060] In the above, the Q-value function unit uses a deep neural network as an approximator of the Q-value function, wherein the input of the deep neural network is the state s, and the output is the value evaluation Q(s,a) of the action a under the state s;

[0061] The network architecture of the deep neural network adopts a multi-layer perceptron, and the last layer corresponds to the value assessment Q(s,a) of each action; wherein, the main difference between the multi-layer perceptron and the single-layer neural network is that the multi-layer perceptron has an additional hidden layer. A single-layer neural network only has an input layer and an output layer, while a multi-layer neural network has one or more hidden layers located between the input layer and the output layer, so the multi-layer perceptron is also more complex. Multi-layer perceptron (MLP) is a feedforward neural network, a fully connected, forward-propagating neural network, consisting of an input layer, one or more hidden layers, and an output layer.

[0062] Some core pseudo codes are as follows:

[0063] #pseudocode

[0064] def Q_network(input_shape,output_dim,learning_rate):

[0065] model=Sequential()

[0066] #Build a deep neural network structure (such as fully connected layer, convolutional layer, etc.)

[0067] #...

[0068] model.compile(optimizer=Adam(lr=learning_rate),loss='mse')

[0069] return model

[0070] In the above, the Deep Q-Network model training process includes:

[0071] Explore the environment through the ε-greedy strategy, perform random actions or choose the action with the highest current Q value;

[0072] After executing the action, a new state s' and a corresponding reward r are obtained, and the experience is stored in the experience playback buffer unit;

[0073] A batch of data is sampled from the experience replay buffer unit for learning and the DQN network parameters are updated to narrow the gap between the value assessment Q(s,a) and the actual return.

[0074] Some core pseudo codes are as follows:

[0075] ·ε-greedy strategy selects actions:

[0076] epsilon = get_epsilon(timestep) # Gradually reduce the ε value as training progresses ifnp.random.rand() <epsilon:

[0077] action = np.random.choice(A) # Random exploration

[0078] else:

[0079] action = np.argmax(Q_network.predict(state)) #Use Q-network to select the best action

[0080] Execute actions and get results:

[0081] new_state, reward, done, _ = environment.step(action)

[0082] replay_buffer.append((state, action, reward, new_state, done))

[0083] · Update Q-network parameters:

[0084] if len(replay_buffer) > batch_size:

[0085] mini_batch = random.sample(replay_buffer, batch_size)

[0086] states_mb, actions_mb, rewards_mb, next_states_mb, dones_mb = zip(*mini_batch)

[0087] target_Q_values = Q_target(next_states_mb).detach().numpy()

[0088] max_Q_next = np.max(target_Q_values, axis = 1)

[0089] targets = rewards_mb + gamma * max_Q_next * (1 - dones_mb)

[0090] with torch.no_grad():

[0091] predicted_Q_values = Q_network(states_mb).gather(1, actions_mb.unsqueeze(-1)).squeeze(-1)

[0092] loss = F.mse_loss(predicted_Q_values, torch.tensor(targets))

[0093] optimizer.zero_grad()

[0094] loss.backward()

[0095] optimizer.step()

[0096] In order to enable the control system to continuously and smoothly adjust the tilt of the aircraft body, reduce the turning radius and reduce overload while ensuring flight stability and passenger comfort, the control system also includes: a dynamic tilt turning control module, which is used to continuously and smoothly adjust the tilt of the aircraft body, reduce the turning radius and reduce overload;

[0097] The dynamic bank turning control module comprises:

[0098] The motion model building unit builds the motion model of the aircraft based on the principles of aircraft dynamics and aerodynamics. The motion model includes a six-degree-of-freedom motion state and physical quantities related to the six-degree-of-freedom motion state. The motion model equation is:

[0099] m*d^2x / dt^2=F_x+m*g*sin(θ)

[0100] m*d^2y / dt^2=F_y

[0101] I_z*d^2ψ / dt^2=L_z

[0102] Where m is the mass, I_z is the moment of inertia, (x,y) are the position coordinates, ψ is the roll angle, (F_x, F_y) are the components of the resultant force in the xy plane, L_z is the moment about the z axis, and θ is the pitch angle;

[0103] The indicator definition unit is used to define the performance indicator, and the quadratic form of the performance indicator J is:

[0104] J=∫[Q(x,u)+R(u)]dt

[0105] Where x is the state vector, u is the control input vector, Q(x,u) is the state cost function, and R(u) is the control input cost function;

[0106] In this embodiment, an optimization objective function may also be set, which should reflect the goal of minimizing or maximizing the desired performance index. Constraints are established, which may involve aircraft operating limitations, such as speed range, overload limit, sideslip angle limit, control surface and engine operating limits, etc. The Hamilton-Jacobi-Bellman (HJB) equation or the Lagrange multiplier method is constructed in the form of:

[0107] H(x,u,λ)=Q(x,u)+λ'f(x,u)

[0108] Selecting a Linear Quadratic Regulator (LQR):

[0109] u ^u*=-K*x

[0110] K is obtained by solving the continuous-time Riccati equation:

[0111] 0 = A'P + PA - PBR^{-1}B'P + Q

[0112] Get the gain matrix K = R^{-1}B'P

[0113] The turning dynamic adjustment unit, based on the calculated control rate, calculates and updates the aircraft's attitude (roll, pitch) and thrust distribution in real time according to the data fed back after the execution of the action during the turning process, so as to achieve a fast, efficient and safe bank turning action based on the preset performance index J; some core pseudo codes are as follows:

[0114] def compute_control_inputs(state,desired_heading):

[0115] # Calculate the required roll rate and thrust distribution

[0116] roll_rate_command=calculate_optimal_roll_rate(state,desired_heading)

[0117] thrust_distribution=allocate_thrust_for_turn(state)

[0118] return roll_rate_command,thrust_distribution

[0119] The closed-loop control unit obtains the output results of the Deep Q-Network model, generates action instructions based on the output structure, and the airborne equipment (servo, engine, etc.) receives and executes the instructions to form a feedback loop. Some core pseudo codes are as follows:

[0120] error=setpoint-current_state

[0121] control_signal=K*error+optimal_control_input(state)

[0122] apply_control(control_signal)

[0123] In this embodiment, the control rate is a mathematical model for controlling input calculated based on the error. The mathematical model is a PID controller (proportional-integral-differential controller), which is used to select a suitable controller structure, set controller parameters, and use dynamic system modeling technology to analyze the stability, response speed and steady-state error characteristics of the system. Some core pseudo codes are as follows:

[0124]

[0125] The error is the result of comparing the current aircraft state parameters with the target set values. For state variables of multiple dimensions, such as attitude angle, position coordinates, speed, etc., the corresponding error values ​​need to be calculated separately. Some core pseudo codes are as follows:

[0126] def calculate_error(target_value,current_value):

[0127] return target_value-current_value

[0128] desired_heading = 90 # Assume the desired heading is 90 degrees

[0129] actual_heading=get_heading_from_imu()#Get the actual heading

[0130] heading_error=calculate_error(desired_heading,actual_heading)

[0131] In this embodiment, the control system also includes a model real-time optimization module. Based on the closed-loop control principle, a closed-loop control system including multi-source sensor data fusion, error detection, real-time calculation and command output is constructed. The closed-loop control is used to monitor the deviation of the actual turning process of the aircraft from the predetermined trajectory and take compensation measures immediately; the model real-time optimization module includes:

[0132] The data fusion unit uses a variety of sensors (such as gyroscopes, accelerometers, magnetometers, GPS navigation systems, etc.) to monitor the aircraft's attitude, position, speed, heading, and environmental parameters in real time. Sensor fusion technology is used to integrate multi-source information to generate accurate and complete flight state estimates. Some core pseudo codes are as follows:

[0133] #Example pseudo code

[0134] import sensor_library

[0135] imu = sensor_library.IMU() # Initialize the IMU sensor object

[0136] gps = sensor_library.GPS() # Initialize GPS sensor object

[0137] while True:

[0138] gyro_data,accel_data=imu.read()#Read and store the current attitude angular velocity and acceleration data

[0139] The feedback unit obtains the error signal in real time, and immediately generates and sends control instructions after obtaining the error signal to control the input of the Deep Q-Network model, so as to continuously adjust the output of the Deep Q-Network model until it approaches or reaches the expected state; some core codes are as follows:

[0140] def set_servo_angle(angle):

[0141] servo_device.set_angle(angle)

[0142] #Convert PID output to servo angle and execute

[0143] corrected_servo_angle=map_control_signal_to_angle(control_signal)

[0144] set_servo_angle(corrected_servo_angle)

[0145] The performance optimization unit records the key performance indicators in real time, and uses the key indicators to evaluate the effect of the PID controller. According to the evaluation results, the online self-correction algorithm or offline parameter tuning method is used to continuously optimize the PID controller parameters to improve the control performance. Some core codes are as follows:

[0146] record_error_history(heading_error)

[0147] optimize_pid_parameters(error_history)

[0148] The security mechanism unit sets the security threshold. When the state variable exceeds the security threshold, an alarm is triggered or the security mode is entered. Some core codes are as follows:

[0149] if abs(control_signal)>MAX_CONTROL_LIMIT:

[0150] trigger_safety_protocol()

[0151] The control system also includes a fault detection module, which is connected to the security mechanism unit. After triggering an alarm or entering the safety mode, the security mechanism unit sends a fault code to the fault detection module. The fault detection module identifies the fault code and switches to a backup system or takes emergency measures to ensure flight safety.

[0152] The embodiments of the present invention have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or technical improvements in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. An intelligent aircraft turning control system, characterized in that: include: Data acquisition, preprocessing module and DQN intelligent optimization module; The data acquisition includes collecting flight data from actual flight records, simulator simulations or wind tunnel experiments; the flight data includes flight performance parameters under different speeds, altitudes, weight distributions, and meteorological conditions, as well as corresponding turning operation details; The preprocessing module is connected to the data acquisition module to preprocess the collected flight data; The DQN intelligent optimization module is connected to the preprocessing module, and uses a deep learning model to obtain the aircraft state s based on the preprocessed flight data, and outputs the best turning action a based on the aircraft state s to obtain the optimal turning strategy.

2. The intelligent aircraft turning control system according to claim 1, characterized in that: The preprocessing includes removing outliers, filling missing values ​​and converting the flight data format to make it suitable for deep learning models.

3. The intelligent aircraft turning control system according to claim 2, characterized in that: The DQN intelligent optimization module includes: Deep Q-Network model, used to adjust the aircraft's control parameters and control the aircraft to complete the turning action; The environment definition unit is used to define the flight environment state, and quantify and discretize the flight environment state into a finite state space S; S = {s_1, s_2, ..., s_n}, where s_n represents the state at time t; and define the action space A; A = {a_1, a_2, ..., a_m}, where a_m represents the action taken at time t; The flight environment includes aircraft status and executable actions; A reward function unit, used to construct a reward function R(s,a,s'), based on which the effect of performing action a in state s to reach a new state s' is measured; Q-value function unit, used to obtain the value evaluation Q(s,a) of action a in state s; Experience replay buffer unit, used to store large amounts of experience data<s,a,r,s'> , where s is the current state, a is the action taken, r is the immediate reward received, and s' is the next state after performing the action.

4. The intelligent aircraft turning control system according to claim 3, characterized in that: The Q-value function unit uses a deep neural network as an approximator of the Q-value function, wherein the input of the deep neural network is the state s, and the output is the value evaluation Q(s,a) of the action a under the state s; The network architecture of the deep neural network adopts a multi-layer perceptron, and the last layer of the deep neural network corresponds to the value evaluation Q(s, a) of each action.

5. The intelligent aircraft turning control system according to claim 4, characterized in that: The Deep Q-Network model needs to be trained before being used to adjust the control parameters of the aircraft. The training process includes: Explore the environment through the ε-greedy strategy, perform random actions or choose the action with the highest current Q value; After executing the action, a new state s' and a corresponding reward r are obtained, and the experience is stored in the experience playback buffer unit; A batch of data is sampled from the experience replay buffer unit for learning, and the DQN network parameters are updated to narrow the gap between the value assessment Q(s, a) and the actual turning flight execution action of the aircraft.

6. The intelligent aircraft turning control system according to claim 5, characterized in that: The control system also includes: a dynamic bank turn control module, which is used to continuously and smoothly adjust the bank of the aircraft body, reduce the turning radius and reduce overload; The dynamic bank turning control module comprises: The motion model building unit builds the motion model of the aircraft based on the principles of aircraft dynamics and aerodynamics. The motion model includes a six-degree-of-freedom motion state and physical quantities related to the six-degree-of-freedom motion state. The motion model equation is: m*d^2x / dt^2=F_x+m*g*sin(θ) m*d^2y / dt^2=F_y I_z*d^2ψ / dt^2=L_z Where m is the mass, I_z is the moment of inertia, (x,y) are the position coordinates, ψ is the roll angle, (F_x, F_y) are the components of the resultant force in the xy plane, L_z is the moment about the z axis, and θ is the pitch angle; The indicator definition unit is used to define the performance indicator, and the quadratic form of the performance indicator J is: J=∫[Q(x,u)+R(u)]dt Where x is the state vector, u is the control input vector, Q(x,u) is the state cost function, and R(u) is the control input cost function; A turning dynamic adjustment unit dynamically adjusts the aircraft's attitude and thrust distribution during the turning process based on the calculated control rate to achieve a banked turning action based on a preset performance index J; The closed-loop control unit obtains the output result of the Deep Q-Network model, generates action instructions based on the output structure, and the airborne device receives and executes the instructions to form a feedback loop.

7. The intelligent aircraft turning control system according to claim 6, characterized in that: The control rate is a mathematical model for controlling input calculated based on the error, and the mathematical model is a PID controller; the error is the result of comparing the current aircraft state parameter with the target setting value.

8. The intelligent aircraft turning control system according to claim 7, characterized in that: The control system also includes a model real-time optimization module. Based on the closed-loop control principle, a closed-loop control system including multi-source sensor data fusion, error detection, real-time calculation and command output is constructed. The closed-loop control is used to monitor the deviation of the actual aircraft from the predetermined trajectory during the turning process and take compensation measures immediately; The model real-time optimization module includes: The data fusion unit integrates multi-source information through sensor fusion technology to generate flight state estimation; A feedback unit obtains an error signal in real time, and immediately generates and sends a control instruction after obtaining the error signal to control the input of the Deep Q-Network model, thereby continuously adjusting the output of the Deep Q-Network model until it approaches or reaches the expected state; A performance optimization unit records key performance indicators in real time, and uses the key indicators to evaluate the effect of the PID controller. Based on the evaluation results, an online self-correction algorithm or an offline parameter tuning method is used to continuously optimize the PID controller parameters to improve the control performance; The security mechanism unit sets a security threshold. When the state variable exceeds the security threshold, an alarm is triggered or the system enters a safe mode.

9. The intelligent aircraft turning control system according to claim 8, characterized in that: The control system also includes a fault detection module, which is connected to the security mechanism unit. After triggering an alarm or entering the safety mode, the security mechanism unit sends a fault code to the fault detection module. The fault detection module identifies the fault code and switches to a backup system or takes emergency measures to ensure flight safety.

Images