A Tracking and Control Method for Quadrotor UAVs Based on Parallel Optimization

By extending the control problem of quadrotor UAVs in virtual space and combining it with data-driven reinforcement learning, the problems of stability and conditional limitations in the tracking control of quadrotor UAVs are solved, and more efficient and accurate trajectory tracking is achieved.

CN117311164BActive Publication Date: 2026-11-10BEIJING INST OF TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202311465642.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-11-07
Publication Date
2026-11-10
Estimated Expiration
2043-11-07

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively apply parallel optimization methods to improve the tracking performance of quadcopter UAVs, especially under unknown conditions. Furthermore, solving the Hamilton-Jacobi-Bellman equations is difficult, leading to limitations in the stability of the closed-loop system and initial control conditions.

Method used

By employing a parallel optimization learning method, the control problem of quadcopter UAVs is extended to virtual space. Combined with data-driven reinforcement learning, and through virtual-reality interaction, control inputs are designed to relax excitation and initial control conditions, satisfying Lyapunov stability and reducing usage restrictions.

Benefits of technology

It improves the accuracy and efficiency of trajectory tracking for quadcopter UAVs, reduces experimental costs, shortens tracking stabilization time, and enhances the stability and control precision of the closed-loop system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117311164B_ABST
    Figure CN117311164B_ABST
Patent Text Reader

Abstract

This invention discloses a parallel optimization-based quadrotor UAV tracking control method, belonging to the field of UAV trajectory tracking technology. This invention employs a parallel optimization control method, expanding the dimension of the control input under a parallel system by differentiating the input, extending the actual physical problem to a virtual space. Through a virtual-reality interactive approach, it achieves force and torque control of the quadrotor UAV, stabilizing the tracking error system to minimize the error between the actual flight trajectory and the reference trajectory, thereby controlling the quadrotor UAV's trajectory tracking. Simultaneously, a data-driven approach reduces the difficulty of solving the optimal problem, better assigns values ​​to the weight vectors of the neural network, relaxes the continuous excitation condition to a finite excitation condition, and reduces the constraints of the initial control conditions. The virtual-reality interactive control completed in a virtual system can improve the tracking performance of the quadrotor UAV's flight trajectory and reduce experimental costs.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention provides a tracking and control method for a quadrotor unmanned aerial vehicle (UAV) based on parallel optimization, belonging to the field of UAV trajectory tracking technology. Background Technology

[0002] In recent years, with the rapid development of the aerospace field, drones have been widely used in both military and civilian applications. Quadrotor drones, in particular, have attracted much attention due to their compact structure, flexible operation, and economical cost, and have proven effective in disaster relief, environmental monitoring, and military reconnaissance.

[0003] Prior technology [1] (see Lu J, Wei Q, Wang F Y. Parallel control for optimal tracking via adaptive dynamic programming [J]. IEEE / CAA Journal of Automatica Sinica, 2020, 7(6): 1662-1674.) proposed the concept and specific proof of parallel optimization, but it is only applicable to general control systems and has not been practically applied to quadcopter UAV systems.

[0004] To improve the tracking performance of UAVs and achieve optimal properties, the Hamilton-Jacobi-Bellman (HJB) equations are typically solved. However, solving the HJB equations for underactuated UAV dynamics systems is quite difficult, and current methods mainly employ optimization learning and reinforcement learning. Simultaneously, to ensure the stability of the closed-loop system and the accuracy of the estimator weights, it is also necessary to ensure the allowable finite excitation and initial control conditions. From an application perspective, this can only be achieved by collecting a sufficient amount of input signal information to meet the parameter estimation requirements. Therefore, there is a need to propose a UAV tracking control method with fewer constraints and better performance. Summary of the Invention

[0005] The purpose of this invention is to provide a parallel optimization-based tracking control method for quadrotor UAVs. For cases where the UAV can achieve ideal acceleration, a parallel optimization learning method is used to construct a tracking error system with stable enhancement capabilities, obtaining the control targets for the force and torque of the quadrotor UAV. This extends the physical problem to a virtual space, achieving virtual-reality interactive control. Through a data-driven reinforcement learning method, the amount of data is effectively increased, and the finite excitation and initial control conditions are relaxed, satisfying the Lyapunov stability of the closed-loop system and reducing the limitations on the application conditions for quadrotor UAV tracking control. Based on this, this invention enables UAVs to perform trajectory tracking more efficiently and accurately.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] This invention discloses a tracking and control method for a quadrotor unmanned aerial vehicle (UAV) based on parallel optimization, comprising the following steps:

[0008] Step 1: Establish the body coordinate system and inertial coordinate system of the quadcopter UAV, simplifying the quadcopter UAV as a rigid body. The coordinate center is the geometric center point of the quadcopter UAV. Define F... B ={O B ,X B ,Y B Z B Let} be the coordinate system of the quadcopter UAV, and define F. I ={O I ,X I ,Y I Z I} represents the inertial coordinate system. A dynamic model of the quadrotor UAV is established in the body coordinate system to describe its position and attitude.

[0009] The dynamic model of the quadcopter UAV is expressed by equations (1) to (4).

[0010]

[0011]

[0012]

[0013]

[0014] In the formula, p is the center of gravity of the UAV, v and ω are the velocity and angular velocity of the UAV, respectively, and T and τ are the applied control force and torque, respectively. g is the acceleration due to gravity. The attitude of the UAV can be represented by a rotation matrix R, which is defined according to Euler's equations as follows: Where θ∈(-π,π), These correspond to Euler angles and coordinate axes, respectively.

[0015] Step 2: Determine the reference trajectory for the quadcopter drone's flight and the error between the actual flight trajectory and the reference trajectory. Define the parameters such as position, velocity, and acceleration required to control the quadcopter drone's flight trajectory. This will facilitate the subsequent use of parallel optimization learning control methods and data-driven reinforcement learning methods when solving Hamiltonian functions in nonlinear systems.

[0016] Step 2.1: Determine the reference trajectory for the quadcopter drone. The reference trajectory for the quadcopter drone is defined with an index of 0, i.e., a reference position p0, a reference velocity v0, and a reference acceleration a0 are defined, all of which are bounded. The derivative of the reference acceleration is: In the formula, d is a known positive constant.

[0017] At this point, the method of differentiating the acceleration is the parallel optimization learning control method. Since the control process of a quadcopter UAV only involves a second-order integrator, while this invention controls a virtual control input in a parallel system, it expands the dimension of the quadcopter UAV's control input. By integrating the derivative of the acceleration as control and applying it to the actual quadcopter UAV system, the virtual-reality interaction is achieved. Analysis of the quadcopter UAV's dynamic model shows that by controlling the applied force T and torque τ, the tracking error between the actual flight trajectory and the reference trajectory of the quadcopter UAV can be kept consistent and bounded.

[0018] Step 2.2: Determine the error between the actual flight trajectory of the quadcopter UAV and the reference trajectory. For the actual flight trajectory of the quadcopter UAV, define the position and velocity tracking errors as follows: and The dynamic equations for the position and velocity of a quadcopter UAV are defined as follows:

[0019]

[0020]

[0021] In the formula, the control of the force applied to the quadcopter drone is: like The rotation matrix is ​​obtained using the minimum rotation theorem.

[0022]

[0023] Therefore, the force exerted by the quadcopter drone is T = m||u||. Similarly, the attitude definition for the quadcopter drone during flight is... This represents the attitude tracking error. Differentiation has

[0024]

[0025] In the formula, the angular velocity error is definition Depend on The conclusion is Depend on Derivation definition Then when have And when and Sometimes, The attitude dynamics error is then...

[0026]

[0027]

[0028] In the formula,

[0029] Step 3: Since the control optimization process requires solving the partial differential equations of the nonlinear system, namely the Hamilton-Jacobi-Bellman (HJB) equations, which cannot be directly obtained, a parallel optimization learning control method is adopted. This method expands the dimension of the control input under the parallel system by differentiating the input, extending the actual physical problem into a virtual space. Through virtual-real interaction, the force and torque applied to the quadrotor UAV are controlled, stabilizing the tracking error system to minimize the error between the actual flight trajectory and the reference trajectory of the quadrotor UAV, thereby controlling the quadrotor UAV to track the trajectory. Simultaneously, a data-driven reinforcement learning method is used to solve the Hamiltonian function in the nonlinear system, enriching the data and relaxing the control conditions to finite excitation conditions, expanding the range of initial control conditions, satisfying the Lyapunov stability of the closed-loop system, and improving the tolerance of the operating conditions for quadrotor UAV tracking control.

[0030] Step 3.1: Design the control of the forces applied by the quadcopter drone. Definition of force control for the quadcopter drone. And introduce In the formula k u >0 and Choose an appropriate k u <g-k ξ to satisfy The conditions are as follows. The nominal dynamic equation is defined as...

[0031]

[0032] In the formula, s = [p eT ,v eT ,δ T ] T , A = [0 3×3 0 3×3 I3] T Introducing performance index equations for the actual flight trajectory of quadcopter UAVs. Where the nonnegative continuous consumption function is: In the formula rs >0, r ζ >0. Define the optimal cost function for the actual flight trajectory of the quadcopter UAV as: The Hamiltonian equation for a quadcopter unmanned aerial vehicle (UAV) system is defined as follows:

[0033]

[0034] In the formula, For the optimal cost function L * The partial derivatives. The optimal cost function L is obtained by... * To find the minimum value, a parallel optimization control method is used to optimize the derivative of the control input with respect to the parallel system. That is, the designed approximate parallel optimization control method is:

[0035]

[0036] Since the analytical solution of the Hamiltonian function in a nonlinear system cannot be found, an approximate optimal control strategy is derived using a reinforcement learning algorithm. The learning weights are updated using a data-driven reinforcement learning algorithm, effectively relaxing the continuous excitation condition to a finite excitation condition, allowing for restrictions on the initial control conditions. Specifically, the neurons in the weighted neural network are assigned weights... p for

[0037]

[0038] In the formula, For the ideal weight vector, For the activation function vector, To estimate the error, calculate the partial derivative of the weighted neural network.

[0039]

[0040] In the formula, Let φ(s) be the partial derivative with respect to s, and Therefore, the approximate Hamiltonian function is obtained as follows:

[0041]

[0042] In the formula, ∈ h For N, it is approximately bounded The residual, For an unknown positive constant. Definition for The estimate. Make The parallel optimization control method is then implemented based on equation (17).

[0043]

[0044] The control of the applied force is then...

[0045]

[0046] The ideal Hamiltonian function satisfies The approximate Hamiltonian function is derived

[0047]

[0048] The above formula is correct. Find the partial derivative.

[0049]

[0050] and define This facilitates the subsequent auxiliary dynamic system calculation and solution process.

[0051] Introducing auxiliary dynamic systems

[0052]

[0053] In the formula, and There is M p (0) = 0, The forgetting factor satisfies the equation In the formula, κ1, κ2, All are positive constants, and κ1 < κ2. The auxiliary dynamical system solution is:

[0054]

[0055] Using all information over a complete time interval may degrade the estimation performance quality under finite stimulus conditions. Therefore, a data selection criterion is introduced.

[0056]

[0057] In the formula,

[0058] Furthermore, completely overcoming the influence of initial control is quite difficult; therefore, robust control is introduced to reduce its impact. The design should satisfy the nominal dynamic equations. Optimized control, i.e., the existence of an unknown normal number. satisfy Define a switch function

[0059]

[0060] In the formula, To ensure the stability of the closed-loop system, when At that time, gradient descent is designed as

[0061]

[0062] In the formula, β1 is a positive constant. The weights of the neural network are updated accordingly, designed as follows:

[0063]

[0064] In the formula, the learning rate β² > 0. The estimation error is defined as follows. Combining equations (16), (19), and (22), we get

[0065]

[0066] In the formula, satisfy

[0067]

[0068] The enhanced closed-loop system position error is

[0069]

[0070] Step 3.2: Design the control of the torque applied by the quadcopter UAV. An auxiliary dynamic parameter is introduced to control the torque applied by the quadcopter UAV. Define variables The enhanced attitude error dynamics equation is expressed by the derivative with respect to time.

[0071]

[0072] In the formula, The performance index equation for the actual flight trajectory of a quadcopter UAV is defined as follows: Where the continuous consumption function is: In the formula r z , All are positive numbers. The optimal cost function for designing the actual flight trajectory of the quadcopter UAV is: Define the Hamiltonian function of the system as follows:

[0073]

[0074] In the formula, To obtain the minimum optimal cost function J * An approximate parallel optimization control method is designed as follows:

[0075]

[0076] By designing an approximate optimization value function to avoid directly solving the Hamiltonian function, neurons with assigned weights are applied in the neural network.a

[0077] J * =ν T ψ(z)+ε (33)

[0078] In the formula, For the ideal weight vector, For the activation function vector, To estimate the error, its partial derivative is:

[0079]

[0080] In the formula, Therefore, the approximate Hamiltonian function is:

[0081]

[0082] In the formula, ε h Let be the residual of the N approximation. For an estimate of ν, the approximate optimal consumption function is: Its gradient is The designed parallel optimization control method is as follows

[0083]

[0084] The control of the applied torque is then...

[0085]

[0086] The ideal Hamiltonian function satisfies The approximate Hamiltonian function is derived

[0087]

[0088] right Finding the gradient gives us

[0089]

[0090] Introduce two auxiliary variables And satisfy M a (0) = 0 And expressed by the following formula

[0091]

[0092] In the formula, κ3, κ4, All are positive constants, κ3 < κ4. It satisfies

[0093]

[0094] To satisfy the data richness requirement of finite incentive conditions, M a , Select as

[0095]

[0096]

[0097] In the formula,

[0098] Similar to the control of forces applied to a quadcopter drone, defining the existence of an optimal control scheme that keeps the dynamic system stable can be expressed as the existence of an unknown constant. satisfy Define switch function

[0099]

[0100] In the formula, Based on gradient descent theory, when At that time, it was launched

[0101]

[0102] In the formula, β3 is a positive constant. The learning weights are updated using a data-driven reinforcement learning algorithm, effectively relaxing the continuous activation condition to a finite activation condition, allowing for restrictions on the initial control conditions, and then the weight update of the neural network is designed.

[0103]

[0104] In the formula, the learning rate β4 > 0. Definition To estimate the error, the derivative is obtained based on the design steps.

[0105]

[0106] In the formula, According to equations (30) and (36), the position error of the enhanced closed-loop system can be obtained as follows:

[0107]

[0108] This method employs a parallel optimization learning control approach, which expands the dimension of the control input under a parallel system by differentiating it, extending the actual physical problem into a virtual space. Through virtual-reality interaction, it achieves control over the force and torque applied to the quadcopter UAV, stabilizes the tracking error system to minimize the error between the actual flight trajectory and the reference trajectory of the quadcopter UAV, and then controls the quadcopter UAV to track the trajectory.

[0109] Beneficial effects:

[0110] 1. The parallel optimization-based quadrotor UAV tracking control method disclosed in this invention employs a parallel optimization control approach. It expands the dimension of the control input under a parallel system by differentiating the input, extending the actual physical problem into a virtual space. Through virtual-reality interaction, it achieves control over the force and torque applied to the quadrotor UAV, stabilizing the tracking error system to minimize the error between the actual flight trajectory and the reference trajectory of the quadrotor UAV, thereby controlling the quadrotor UAV's tracking trajectory. Compared to directly controlling the quadrotor UAV, the virtual-reality interactive control design completed in a virtual system improves the tracking performance of the quadrotor UAV's flight trajectory and reduces experimental costs.

[0111] 2. The parallel optimization-based quadrotor UAV tracking control method disclosed in this invention avoids the difficulty of directly solving the Hamiltonian function by combining it with an RL neural network. It reduces the difficulty of solving the optimization problem through a data-driven approach, allowing for better assignment of weight vectors to the neural network. Simultaneously, it relaxes the continuous excitation conditions to finite excitation conditions and initial control conditions, satisfying the Lyapunov stability of the closed-loop system. This reduces the limitations on the application conditions for quadrotor UAV tracking control, and effectively controls the tracking error between the actual flight trajectory and the reference trajectory of the quadrotor UAV to a smaller value, reduces the time required to achieve basic tracking stability, and improves the tracking control accuracy of the quadrotor UAV. Attached Figure Description

[0112] Figure 1 This is a schematic diagram of the coordinate system setting for the quadcopter UAV of the present invention.

[0113] Figure 2 This is a flowchart of the parallel optimization-based tracking control process for a quadcopter UAV according to the present invention. Detailed Implementation

[0114] To overcome the shortcomings of existing technologies, this invention aims to propose a parallel optimization-based tracking control method for quadrotor UAVs, optimizing and improving UAV tracking control methods to ensure the accuracy and performance of UAV tracking trajectories. The technical solution adopted in this invention mainly addresses two aspects: first, using parallel optimization technology to reduce control errors; and second, designing the weight values ​​of the neural network through a data-driven reinforcement learning algorithm.

[0115] like Figure 2 As shown in the figure, the specific steps of the parallel optimization-based quadcopter UAV tracking and control method disclosed in this embodiment are as follows:

[0116] Step 1: Establish the body coordinate system and inertial coordinate system of the quadcopter UAV, simplifying the quadcopter UAV as a rigid body. The coordinate center is the geometric center point of the quadcopter UAV. Define F...B ={O B ,X B ,Y B Z B Let} be the coordinate system of the quadcopter UAV, and define F. I ={O I ,X I ,Y I Z I} represents the inertial coordinate system. A dynamic model of the quadrotor UAV is established in the body coordinate system to describe its position and attitude.

[0117] The dynamic model of the quadcopter UAV is expressed by equations (1) to (4).

[0118]

[0119]

[0120]

[0121]

[0122] In the formula, p is the center of gravity of the UAV, v and ω are the velocity and angular velocity of the UAV, respectively, and T and τ are the applied control force and torque, respectively. g is the acceleration due to gravity. The attitude of the UAV can be represented by a rotation matrix R, which is defined according to Euler's equations as follows: in These correspond to Euler angles and coordinate axes, respectively.

[0123] The mass and moment of inertia of the UAV are determined as follows:

[0124] m = 1.5 kg

[0125] J = diag(0.01, 0.01, 0.02) kgm 2

[0126] Substitute the above physical quantities into the dynamic model of the quadcopter UAV.

[0127] Step 2: Determine the reference trajectory for the quadcopter drone's flight and the error between the actual flight trajectory and the reference trajectory. Define the parameters such as position, velocity, and acceleration required to control the quadcopter drone's flight trajectory. This will facilitate the subsequent use of parallel optimization learning control methods and data-driven reinforcement learning methods when solving Hamiltonian functions in nonlinear systems.

[0128] Step 2.1: Determine the reference trajectory for the quadcopter drone. The reference trajectory for the quadcopter drone is defined with an index of 0, i.e., a reference position p0, a reference velocity v0, and a reference acceleration a0 are defined, all of which are bounded. The derivative of the reference acceleration is: ≤d, where d is a known positive constant.

[0129] At this point, the method of differentiating the acceleration is the parallel optimization learning control method. Since the control process of a quadcopter UAV only involves a second-order integrator, while this invention controls a virtual control input in a parallel system, it expands the dimension of the quadcopter UAV's control input. By integrating the derivative of the acceleration as control and applying it to the actual quadcopter UAV system, the virtual-reality interaction is achieved. Analysis of the quadcopter UAV's dynamic model shows that by controlling the applied force T and torque τ, the tracking error between the actual flight trajectory and the reference trajectory of the quadcopter UAV can be kept consistent and bounded.

[0130] Step 2.2: Determine the error between the actual flight trajectory of the quadcopter UAV and the reference trajectory. For the actual flight trajectory of the quadcopter UAV, define the position and velocity tracking errors as follows: and

[0131] Its reference trajectory is set as

[0132] p0=[6cos(0.1t),6sin(0.1t),10(1-e -0.05t )] T

[0133] Set the initial position and attitude of the drone as follows:

[0134] p(0) = [6, 3, -0.5] T

[0135] The dynamic equations for the position and velocity of a quadcopter UAV are defined as follows:

[0136]

[0137]

[0138] In the formula, the control of the force applied to the quadcopter drone is: like The rotation matrix is ​​obtained using the minimum rotation theorem.

[0139]

[0140] Therefore, the force exerted by the quadcopter drone is T = m||u||. Similarly, the attitude definition for the quadcopter drone during flight is... This represents the attitude tracking error. Differentiation has

[0141]

[0142] In the formula, the angular velocity error is definition Depend on The conclusion is Depend on Derivation definition Then when have And when and Sometimes, The attitude dynamics error is then...

[0143]

[0144]

[0145] In the formula,

[0146] Step 3: Since the control optimization process requires solving the partial differential equations of the nonlinear system, namely the Hamilton-Jacobi-Bellman (HJB) equations, which cannot be directly obtained, a parallel optimization learning control method is adopted. This method expands the dimension of the control input under the parallel system by differentiating the input, extending the actual physical problem into a virtual space. Through virtual-real interaction, the force and torque applied to the quadrotor UAV are controlled, stabilizing the tracking error system to minimize the error between the actual flight trajectory and the reference trajectory of the quadrotor UAV, thereby controlling the quadrotor UAV to track the trajectory. Simultaneously, a data-driven reinforcement learning method is used to solve the Hamiltonian function in the nonlinear system, enriching the data and relaxing the control conditions to finite excitation conditions, expanding the range of initial control conditions, satisfying the Lyapunov stability of the closed-loop system, and improving the tolerance of the operating conditions for quadrotor UAV tracking control.

[0147] Step 3.1: Design the control of the forces applied by the quadcopter drone. Definition of force control for the quadcopter drone. And introduce In the formula k u >0 and Choose an appropriate k u <g-k ξ to satisfy The conditions are as follows. The nominal dynamic equation is defined as...

[0148]

[0149] In the formula, s = [p eT ,v eT ,δ T ] T , A = [0 3×3 0 3×3 I3] T Introducing performance index equations for the actual flight trajectory of quadcopter UAVs. Where the nonnegative continuous consumption function is: In the formula r s >0, r ζ >0. Define the optimal cost function for the actual flight trajectory of the quadcopter UAV as: The Hamiltonian equation for a quadcopter unmanned aerial vehicle (UAV) system is defined as follows:

[0150]

[0151] In the formula, For the optimal cost function L * The partial derivatives. The optimal cost function L is obtained by... * To find the minimum value, a parallel optimization control method is used to optimize the derivative of the control input with respect to the parallel system. That is, the designed approximate parallel optimization control method is:

[0152]

[0153] Since the analytical solution of the Hamiltonian function in a nonlinear system cannot be found, an approximate optimal control strategy is derived using a reinforcement learning algorithm. The learning weights are updated using a data-driven reinforcement learning algorithm, effectively relaxing the continuous excitation condition to a finite excitation condition, allowing for restrictions on the initial control conditions. Specifically, the neurons in the weighted neural network are assigned weights... p for

[0154]

[0155] In the formula, For the ideal weight vector, For the activation function vector, To estimate the error, calculate the partial derivative of the weighted neural network.

[0156]

[0157] In the formula, Let φ(s) be the partial derivative with respect to s, and Therefore, the approximate Hamiltonian function is obtained as follows:

[0158]

[0159] In the formula, ∈ h For N, it is approximately bounded The residual, For an unknown positive constant. Definition for The estimate. Make The parallel optimization control method is then implemented based on equation (17).

[0160]

[0161] The control of the applied force is then...

[0162]

[0163] The ideal Hamiltonian function satisfies The approximate Hamiltonian function is derived

[0164]

[0165] The above formula is correct. Find the partial derivative.

[0166]

[0167] and define This facilitates the subsequent auxiliary dynamic system calculation and solution process.

[0168] Introducing auxiliary dynamic systems

[0169]

[0170] In the formula, and There is M p (0) = 0, The forgetting factor satisfies the equation In the formula, κ1, κ2, All are positive constants, and κ1 < κ2. The auxiliary dynamical system solution is:

[0171]

[0172] Using all information over a complete time interval may degrade the estimation performance quality under finite stimulus conditions. Therefore, a data selection criterion is introduced.

[0173]

[0174] In the formula,

[0175] Furthermore, completely overcoming the influence of initial control is quite difficult; therefore, robust control is introduced to reduce its impact. The design should satisfy the nominal dynamic equations. Optimized control, i.e., the existence of an unknown normal number. satisfy Define a switch function

[0176]

[0177] In the formula, To ensure the stability of the closed-loop system, when At that time, gradient descent is designed as

[0178]

[0179] In the formula, β1 is a positive constant.

[0180] Suppose that the neural network with activation function has 18 neurons, and its weights are expressed as follows:

[0181]

[0182] The initial estimate was set as follows: δ = 13, τ = 0.5 × 13.

[0183] This updates the weights of the neural network, designed as follows:

[0184]

[0185] In the formula, the learning rate β² > 0. The estimation error is defined as follows. Combining equations (16), (19), and (22), we get

[0186]

[0187] In the formula, satisfy

[0188]

[0189] The enhanced closed-loop system position error is

[0190]

[0191] Step 3.2: Design the control of the torque applied by the quadcopter UAV. An auxiliary dynamic parameter is introduced to control the torque applied by the quadcopter UAV. Define variables The enhanced attitude error dynamics equation is expressed by the derivative with respect to time.

[0192]

[0193] In the formula, The performance index equation for the actual flight trajectory of a quadcopter UAV is defined as follows: Where the continuous consumption function is: In the formula r z , All are positive numbers. The optimal cost function for designing the actual flight trajectory of the quadcopter UAV is: Define the Hamiltonian function of the system as follows:

[0194]

[0195] In the formula, To obtain the minimum optimal cost function J * An approximate parallel optimization control method is designed as follows:

[0196]

[0197] By designing an approximate optimization value function to avoid directly solving the Hamiltonian function, neurons with assigned weights are applied in the neural network. a

[0198] J * =v T ψ(z)+ε (33)

[0199] In the formula, For the ideal weight vector, For the activation function vector, To estimate the error, its partial derivative is:

[0200]

[0201] In the formula, Therefore, the approximate Hamiltonian function is:

[0202]

[0203] In the formula, ε h Let be the residual of the N approximation. For an estimate of ν, the approximate optimal consumption function is: Its gradient is The designed parallel optimization control method is as follows

[0204]

[0205] The control of the applied torque is then...

[0206]

[0207] The ideal Hamiltonian function satisfies The approximate Hamiltonian function is derived

[0208]

[0209] right Finding the gradient gives us

[0210]

[0211] Introduce two auxiliary variables And satisfy M a (0) = 0 And expressed by the following formula

[0212]

[0213] In the formula, k3, k4 All are positive numbers, k3 <k4, It satisfies

[0214]

[0215] To satisfy the data richness requirement of finite incentive conditions, M a , Select as

[0216]

[0217]

[0218] In the formula,

[0219] Similar to the control of forces applied to a quadcopter drone, defining the existence of an optimal control scheme that keeps the dynamic system stable can be expressed as the existence of an unknown constant. satisfy Define switch function

[0220]

[0221] In the formula, Based on gradient descent theory, when At that time, it was launched

[0222]

[0223] In the formula, β3 is a positive constant. Suppose the neural network with activation function contains 18 neurons, and its weights are expressed as follows:

[0224]

[0225]

[0226] The initial estimate was set as follows: δ = 13, τ = 0.5 × 13.

[0227] The learning weights are updated using a data-driven reinforcement learning algorithm, effectively relaxing the continuous incentive condition to a finite incentive condition, allowing for restrictions on the initial control conditions, and then the weight update of the neural network is designed.

[0228]

[0229] In the formula, the learning rate β4 > 0. Definition To estimate the error, the derivative is obtained based on the design steps.

[0230]

[0231] In the formula, According to equations (30) and (36), the position error of the enhanced closed-loop system can be obtained as follows:

[0232]

[0233] The control strategy and reinforcement learning parameters are selected as follows:

[0234] r s =1

[0235] r ζ =1

[0236] r z =1

[0237]

[0238] k u =4

[0239] β1=1

[0240] β2=10

[0241] β3=1

[0242] β4 = 10

[0243] κ1=0.5

[0244] κ2=80

[0245]

[0246] κ3=1

[0247] κ4 = 10

[0248]

[0249] The above parameters were substituted into the corresponding formulas for solution, and flight verification was conducted. From this specific implementation, it can be seen that the quadcopter UAV can adjust its flight state in a short time, fly well along the reference trajectory, and has small position and attitude errors. The applied forces and torques can also be adjusted well within the control range, with a short convergence time and good tracking performance.

[0250] This method employs a parallel optimization learning control approach, which expands the dimension of the control input under a parallel system by differentiating it, extending the actual physical problem into a virtual space. Through virtual-reality interaction, it achieves control over the force and torque applied to the quadcopter UAV, stabilizes the tracking error system to minimize the error between the actual flight trajectory and the reference trajectory of the quadcopter UAV, and then controls the quadcopter UAV to track the trajectory.

Claims

1. A tracking and control method for a quadrotor unmanned aerial vehicle based on parallel optimization, characterized in that: Includes the following steps, Step 1: Establish the body coordinate system and inertial coordinate system of the quadcopter UAV, simplifying the quadcopter UAV as a rigid body, with the coordinate center being the geometric center point of the quadcopter UAV, and define... Define the coordinate system for the quadcopter UAV and define... Using an inertial coordinate system, a dynamic model of the quadrotor UAV is established in the body coordinate system to describe its position and attitude. Step 2: Determine the reference trajectory of the quadcopter drone and the error between the actual flight trajectory of the quadcopter drone and the reference trajectory. Define the position, velocity, acceleration and other parameters required to control the flight trajectory of the quadcopter drone, so as to facilitate the subsequent use of parallel optimization learning control method and data-driven reinforcement learning when solving Hamiltonian functions in nonlinear systems. In step 2, Step 2.1: Determine the reference trajectory for the quadcopter drone's flight; define the reference trajectory of the quadcopter drone with the index 0, i.e., define the reference position. Reference speed Reference acceleration And all are bounded; the derivative of the reference acceleration has In the formula For known positive constants; At this point, the method of finding the derivative of acceleration is the parallel optimization learning control method. Since the control process of the quadcopter only has a second-order integrator, and what is being controlled is the virtual control input under the parallel system, the dimension of the quadcopter's control input is expanded. By integrating the derivative of acceleration as control and applying it to the actual quadcopter system, the virtual-reality interaction is completed. Analysis of the dynamic model of the quadcopter drone shows that by controlling the applied force... and torque This ensures that the tracking error between the actual flight trajectory and the reference trajectory of the quadcopter drone is consistent and bounded. Step 2.2: Determine the error between the actual flight trajectory of the quadcopter UAV and the reference trajectory; for the actual flight trajectory of the quadcopter UAV, define the position and velocity tracking errors as follows: and The dynamic equations for the position and velocity of a quadcopter UAV are defined as follows: In the formula, the control of the force applied to the quadcopter drone is: ;like The rotation matrix is ​​obtained using the minimum rotation theorem. Therefore, the force exerted by the quadcopter drone is Similarly, the attitude definition for quadcopter drones during flight... For attitude tracking error; Differentiation has In the formula, the angular velocity error is ;definition ,Depend on The conclusion is ,Depend on Derivation ;definition Then when have , and when and Sometimes, The attitude dynamics error is then... In the formula, , , ; Step 3: Since the control optimization process requires solving the partial differential equations of the nonlinear system, namely the Hamilton-Jacobi-Bellman (HJB) equations, which cannot be directly obtained, a parallel optimization learning control method is adopted. This method expands the dimension of the control input under the parallel system by differentiating the input, extending the actual physical problem into a virtual space. Through virtual-reality interaction, the force and torque applied to the quadrotor UAV are controlled, stabilizing the tracking error system to minimize the error between the actual flight trajectory and the reference trajectory of the quadrotor UAV, thereby controlling the quadrotor UAV to track the trajectory. Simultaneously, a data-driven reinforcement learning method is used to solve the Hamiltonian function in the nonlinear system, enriching the data and relaxing the control conditions to finite excitation conditions, expanding the range of initial control conditions, satisfying the Lyapunov stability of the closed-loop system, and improving the tolerance of the operating conditions for quadrotor UAV tracking control.

2. The quadrotor UAV tracking and control method based on parallel optimization as described in claim 1, characterized in that: In step 1, The dynamic model of the quadcopter UAV is expressed by equations (1) to (4). In the formula, The center of gravity of the drone , These are the drone's speed and angular velocity, respectively. , These are the applied control force and torque, respectively; It is the acceleration due to gravity. The attitude of the drone can be determined by a rotation matrix. To represent, according to Euler's equations Defined as ,in , These correspond to Euler angles and coordinate axes, respectively.

3. The quadrotor UAV tracking and control method based on parallel optimization as described in claim 1, characterized in that: Step 3 is implemented as follows: Step 3.1: Design the force control for the quadcopter drone; define the force control for the quadcopter drone. and introduce In the formula and Choose the appropriate to satisfy The conditions; the nominal dynamic equation is defined as In the formula, , , Introducing performance index equations for the actual flight trajectory of quadcopter UAVs. The non-negative continuous consumption function is In the formula , The optimal cost function for the actual flight trajectory of a quadcopter UAV is defined as follows: The Hamiltonian equations for a quadcopter unmanned aerial vehicle (UAV) system are defined as follows: In the formula, For the optimal cost function The partial derivatives; by obtaining the optimal cost function. To find the minimum value, a parallel optimization control method is used to optimize the derivative of the control input with respect to the parallel system. That is, the designed approximate parallel optimization control method is: Since the analytical solution of the Hamiltonian function in a nonlinear system cannot be found, an approximate optimal control strategy is obtained through reinforcement learning algorithms. The learning weights are updated using a data-driven reinforcement learning algorithm, effectively relaxing the continuous excitation conditions to finite excitation conditions, allowing for restrictions on the initial control conditions, and assigning weights to neurons in the neural network. for In the formula, For the ideal weight vector, For the activation function vector, To estimate the error; calculate the partial derivative of the weighted neural network as follows: In the formula, for right The partial derivative, and Therefore, the approximate Hamiltonian function is obtained as follows: In the formula, For N, it is approximately bounded The residual, Let be an unknown positive constant; define for The estimate; to make , The parallel optimization control method is then implemented based on equation (17). The control of the applied force is then... The ideal Hamiltonian function satisfies The approximate Hamiltonian function is obtained. The above formula is correct. Find the partial derivative. and define , This is to facilitate the subsequent auxiliary dynamic system calculation and solution process; Introducing auxiliary dynamic systems In the formula, and ,have , The forgetting factor satisfies the equation In the formula , , All are positive numbers, and The auxiliary dynamic system solution is Using all information over a complete time interval may degrade the estimation performance quality under finite stimulus conditions; therefore, data selection criteria are introduced. In the formula, ; Furthermore, completely overcoming the influence of initial control is quite difficult; therefore, robust control is introduced to reduce its impact. The design should satisfy the nominal dynamic equations. Optimized control, i.e., the existence of an unknown normal number. satisfy Define a switch function. In the formula, To ensure the stability of the closed-loop system, when At that time, gradient descent is designed as In the formula, It is a positive constant; therefore, the weights of the neural network are updated, designed as follows: In the formula, the learning rate Define estimation error Combining equations (16), (19), and (22), we get In the formula, ;satisfy The enhanced closed-loop system position error is Step 3.2: Design the control of the torque applied by the quadcopter UAV; for the torque applied by the quadcopter UAV, introduce an auxiliary dynamic parameter. , ; Define variables The enhanced attitude error dynamics equation is expressed by the derivative with respect to time. In the formula, The performance index equation for the actual flight trajectory of a quadcopter UAV is defined as follows: The continuous consumption function is In the formula , All are positive numbers; the optimal cost function for designing the actual flight trajectory of the quadcopter UAV is: The Hamiltonian function of the system is defined as follows: In the formula, To obtain the minimum optimal cost function An approximate parallel optimization control method is designed as follows: By designing an approximate optimization value function to avoid directly solving the Hamiltonian function, neurons with assigned weights are applied in the neural network. In the formula, For the ideal weight vector, For the activation function vector, To estimate the error, its partial derivative is: In the formula, , Therefore, the approximate Hamiltonian function is: In the formula, Let the residuals be the N approximation; for The approximate optimal consumption function is then estimated as follows: Its gradient is The designed parallel optimization control method is as follows: The control of the applied torque is then... The ideal Hamiltonian function satisfies The approximate Hamiltonian function is obtained. right Finding the gradient gives us Introduce two auxiliary variables , And satisfy , And expressed by the following formula In the formula, , , , All are positive numbers. , , It satisfies To satisfy the data richness requirement of finite incentive conditions, , Select as In the formula, ; Similar to the control of forces applied to a quadcopter drone, defining the existence of an optimal control scheme that keeps the dynamic system stable can be expressed as the existence of an unknown constant. satisfy Define the switching function In the formula, Based on gradient descent theory, when At that time, it was launched In the formula, Given a positive constant, the learning weights are updated using a data-driven reinforcement learning algorithm, effectively relaxing the continuous activation condition to a finite activation condition, allowing for restrictions on the initial control conditions, and then designing the weight update mechanism for the neural network. In the formula, the learning rate ;definition To estimate the error, its derivative is obtained based on the design steps. In the formula, According to equations (30) and (36), the position error of the enhanced closed-loop system can be obtained as follows: By employing a parallel optimization learning control method, the control input under the parallel system is differentiated and expanded in dimension, extending the actual physical problem into a virtual space. Through virtual-reality interaction, the force and torque control of the quadcopter UAV is realized. The stable tracking error system minimizes the error between the actual flight trajectory and the reference trajectory of the quadcopter UAV, thereby controlling the quadcopter UAV to track the trajectory.

Citation Information

Patent Citations

  • Control system for unmanned 4-rotor-helicopter

    EP1901153A1

  • User interaction with an autonomous unmanned aerial vehicle

    US20190378423A1