Transition Control Method for Vertical / Short Takeoff and Landing Aircraft Based on Intelligent Control Allocation

Abstract

This invention relates to a transition phase control method for vertical / short takeoff and landing (VTOL) aircraft based on intelligent control allocation. The method employs a nonlinear incremental dynamic inverse approach to design the control law, which is then used to calculate the acceleration control commands that meet flight requirements. To calculate the actual control quantity based on the acceleration control commands, an optimization objective function is designed. An improved gray wolf optimization algorithm is then used to solve for the control quantity. An initialization strategy based on a set of optimal points is adopted to avoid the common random initialization that leads to individuals concentrating at local optima, thus improving search efficiency. A nonlinear variation strategy based on the LeCunTanh activation function of a neural network is proposed to nonlinearly adjust global and local search capabilities, achieving better optimization results. Historical optimal information is preserved to guide gray wolf optimization. Fitness variance and gray wolf average distance are introduced. When the algorithm gets stuck in a local optimum, half of the gray wolf population with poor fitness is selected, and the Levy flight strategy is executed to escape local optima.

Description

Technical Field

[0001] This invention belongs to the field of aircraft control, specifically relating to a transition control method for vertical / short takeoff and landing aircraft based on intelligent control allocation. Background Technology

[0002] Compared to conventional fixed-wing aircraft, vertical / short takeoff and landing (V / STOL) aircraft can reduce or even eliminate their dependence on runways; compared to helicopters, they possess greater speed, range, and payload, and have attracted considerable attention in recent years. V / STOL aircraft can be mainly classified into thrust-vectoring, tiltrotor, and tail-seat types based on different power system configurations. Thrust-vectoring aircraft are widely used in fixed-wing fighter jets, with typical examples including the British Harrier and the American F-35B.

[0003] The thrust vectoring V / STOL aircraft differs significantly from the other two types in its power system layout and nonlinear characteristics. Its power system deflection angle, thrust, and flight speed are all factors contributing to strong nonlinearity. Using conventional gain pre-setting methods would not only result in an enormous computational burden but also struggle to handle nonlinear issues. Secondly, for tiltrotor aircraft, which also have thrust deflection capabilities, the bandwidth of its tilt mechanism is significantly lower than that of other actuators such as control surfaces. Therefore, during transitions, the tilt mechanism does not participate in the aircraft's attitude control to ensure a smoother transition. However, the thrust vectoring V / STOL aircraft has a higher bandwidth for its thrust vector control mechanism than its engine, thus requiring the use of thrust deflection to control attitude and improve control performance. This significantly increases the number of control variables, and problems such as coupling and nonlinearity between different control mechanisms arise, necessitating the comprehensive handling of numerous control variables. To overcome these shortcomings, intelligent algorithms can be used for allocation and solution.

[0004] Intelligent algorithms typically mimic the social phenomena of organisms in nature, searching and solving problems according to certain rules, and possess advantages such as high speed and strong robustness. The Grey Wolf Optimization Algorithm (GWO), proposed by Mirjalili in 2014, is a metaheuristic swarm intelligence optimization algorithm that simulates the predation behavior and social hierarchy mechanisms of grey wolf packs. The GWO algorithm features a simple model structure, few control parameters, fast convergence speed, and good optimization performance. Numerous studies have shown that compared to traditional heuristic optimization algorithms, such as particle swarm optimization and genetic algorithms, the GWO algorithm is more advantageous for solving high-dimensional optimization problems. However, this algorithm also suffers from problems such as being prone to getting trapped in local optima and slow convergence in later stages, requiring further improvement. Summary of the Invention

[0005] The purpose of this invention is to provide a transition control method for V / STOL aircraft. The method employs an incremental nonlinear dynamic inversion (INDI) approach to design the control law, which is then used to calculate the acceleration control command that meets flight requirements. To calculate the actual control quantity based on the acceleration control command, an optimization objective function is designed after considering actual physical constraints such as actuator position and speed. An improved grey wolf optimization algorithm is then used to solve for the control quantity that meets the conditions, avoiding getting trapped in local optima and accelerating the later convergence speed.

[0006] The specific technical solution adopted in this invention is: a V / STOL aircraft transition control method based on intelligent control allocation, including a control law module and an allocation law module;

[0007] The control law module divides the state variables of the flight control system into multiple loops according to their response speed, and designs the control law for each loop from the inside out. A fixed-time disturbance observer is used to estimate the angular acceleration. The fixed-time disturbance observer takes the following form:

[0008]

[0009] In the formula, Let x and d be the estimated values ​​of state x and uncertainty d, respectively. To estimate the error, the function sig(y) r Defined as sig(y) r =sgn(y)|y| r The coefficients k1,k2,k3>0, 0<λ1<1, λ2>1, and the uncertainty d satisfies the following conditions:

[0010] The distributive law module employs an improved gray wolf optimization algorithm, simulating the hunting process of a gray wolf pack searching, surrounding, and attacking prey to achieve optimal search. The GWO algorithm divides the gray wolf population into four levels: α, β, δ, and ω, where α is the highest level, corresponding to the optimal solution in the search; β and δ correspond to the suboptimal and near-suboptimal solutions, respectively; and ω represents other solutions in the solution space. ω wolves update their positions based on the positions of α, β, and δ wolves until the search ends. Compared to traditional optimization algorithms, the GWO algorithm has advantages such as fewer adjustment parameters, faster convergence speed, and stronger optimization performance, but it also has disadvantages such as being prone to getting trapped in local optima and slow convergence in the later stages of iteration. Therefore, this invention makes corresponding improvements to the basic GWO algorithm: specifically, it initializes the population by superimposing a set of optimal points with random variables, and then calculates the fitness values ​​of individuals.

[0011] Furthermore, the initialization of the population using a set of optimal points superimposed with random variables includes:

[0012] The set of ideal points containing N points is represented as:

[0013] P n (i)=U L +{γ i *n}*(U H -U L i = 1, 2, ..., dn = 1, 2, ..., N

[0014]

[0015] In the formula, d represents the dimension of the problem to be solved, and U L U H Let represent the upper and lower bounds of the corresponding dimension, respectively; N be the number of gray wolves; q be the smallest prime number satisfying q-3≥2d; {γ i *n} represents taking γ i The decimal part of *n};

[0016] The superposition of random variables is expressed as:

[0017]

[0018] In the formula, ε is a random number between 0 and 1.

[0019] The initial position of each gray wolf is obtained by superimposing a set of optimal points onto random variables.

[0020] Furthermore, the calculation of individual fitness values ​​includes: using historical optimal information to guide the search process, and incorporating the individual's historical optimal position into the position update formula, i.e., the position update formula for α wolves, β wolves, and δ wolves in the wolf pack is:

[0021] X p (t+1)=X l (t)-A|CX l (t)-X(t)|+λ(X p,best -X l (t))p=α,β,δ

[0022] In the formula, t is the current iteration number, and X is... l X(t) represents the prey's position, X(t) and X(t+1) represent the positions of individual wolves in the pack during the two iterations, A is the convergence factor, C is the oscillation factor, and λ is the individual's optimal position memory parameter. p,best This represents the historically optimal position for the three wolf species.

[0023] Based on the method for calculating the resultant force in the Artificial Potential Field (APF) method, the position update formula for ωwolf is designed as follows:

[0024] X ω (t+1)=η α Xα (t+1)+η β X β (t+1)+η δ X δ (t+1)

[0025]

[0026] In the formula, η k These are the weighting coefficients for the three types of wolf leaders.

[0027] Furthermore, the convergence factor A and the oscillation factor C are calculated as follows:

[0028] A=(2r1-1)a

[0029] C = 2r²

[0030] r1 and r2 are both random coefficients between [0,1], and a is the distance control parameter;

[0031] The parameter 'a' employs a nonlinear variation strategy based on the LeCun Tanh activation function of a neural network, namely:

[0032]

[0033] In the formula, a init a end These are the initial and final distance control parameters, respectively, where t is the current iteration number, and T is the distance control parameter at the beginning and end of the iteration. max Let tanh(x) be the total number of iterations, and let tanh(x) = (e^(x-1 x -e -x ) / (e x +e -x ).

[0034] Furthermore, a init =2,a end =0.1.

[0035] Furthermore, it also includes calculating the fitness variance σ. 2 Average distance from the gray wolf When both are less than the critical value, the algorithm is determined to be trapped in a local optimum. The half of the gray wolves with the worst fitness value ω is selected, and the Levy flight strategy is executed.

[0036] Furthermore, the calculation of the fitness variance σ 2 Average distance from the gray wolf for:

[0037]

[0038]

[0039] In the formula, fit i and fit avg Here, X represents the fitness value of each gray wolf and the average fitness value of the wolf pack, where N is the number of gray wolves and X is the number of gray wolves. i Let Xα be the position of the gray wolf, ||X||2 represents the 2-norm; i represents any one of the N gray wolves.

[0040] Furthermore, the step of selecting the half of the gray wolves ω with the worst fitness value and executing the Levy flight strategy includes:

[0041]

[0042] In the formula, X ω,half To adapt the half of the gray wolves with poorer fitness values, μ is the step scaling factor, L is the step size coefficient, and L follows a Levy distribution, which is obtained by the following formula:

[0043]

[0044]

[0045]

[0046] In the formula, θ is a constant between 0 and 2, and N norm This indicates a normal distribution.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] To address the design challenges of V / STOL aircraft control systems, this invention proposes a hierarchical control system structure that designs control laws and allocation laws independently, reducing overall complexity. The modular design of control and allocation laws allows for independent operation, facilitating future updates or replacements. When control surfaces malfunction, multiple control laws are not required; system reconfiguration can be achieved simply by reallocating control surfaces. Furthermore, the allocation laws not only directly address the physical constraints of control surfaces but also facilitate mission-oriented management.

[0049] To address the problem of angular acceleration estimation, this invention proposes to use a fixed-time disturbance observer (FTDO) for estimation. Compared with the traditional sliding mode disturbance observer, since the sign function is in the integral term, the estimated value is continuous, eliminating the chattering problem. Furthermore, the upper bound of the convergence time does not depend on the initial state, but is determined by the control parameters.

[0050] To solve the objective function optimization problem, an improved Grey Wolf Optimization Algorithm (GWO) is adopted. Conventional GWOs use random initialization of the population, which may lead to initial individuals clustering in small areas or starting far from the optimal solution, making the algorithm prone to getting trapped in local optima, especially when solving high-dimensional optimization problems. Therefore, this invention adopts several improvements: an initialization strategy based on a set of optimal points is used to avoid the common random initialization that may cause individuals to concentrate in local optima, thus improving search efficiency; a nonlinear variation strategy based on the LeCun-Tanh activation function of a neural network is proposed to nonlinearly adjust global and local search capabilities, achieving better optimization results; referencing the Particle Swarm Optimization (PSO) algorithm, historical optimal information is preserved to guide the Grey Wolf optimization; fitness variance and average Grey Wolf distance are introduced, and when the algorithm gets trapped in a local optimum, half of the Grey Wolf population with poor fitness is selected to execute the Levy flight strategy, which can help escape local optima. Attached Figure Description

[0051] Figure 1 This is a flowchart of the gray wolf optimization algorithm of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. The specific embodiments described herein are merely illustrative and are not intended to limit the scope of the invention.

[0053] The present invention provides a V / STOL aircraft transition phase control method based on intelligent control allocation, comprising a control law module and an allocation law module;

[0054] The control law module divides the state variables of the flight control system into multiple loops according to their response speed, and designs the control law for each loop from the inside out. Specifically:

[0055] The angular velocity loop is a fast loop of the system. According to the INDI method, a first-order Taylor expansion is performed at the state point (ω0, δ0) to obtain an approximate expression for the neighborhood of the state point:

[0056]

[0057] In the formula, ω=[p,q,r] T For triaxial angular velocity, The corresponding angular acceleration is obtained through angular acceleration estimation or direct measurement, where J is the moment of inertia matrix, and M is the moment of inertia matrix. a M is the aerodynamic torque generated by the aircraft's motion. c The control torque generated for the control quantity δ (including aerodynamic control surfaces and thrust vector).

[0058] According to the principle of the INDI method, some components can be ignored in extremely short control steps. Therefore, after simplifying the above equation, the control law of the angular rate loop can be obtained as follows:

[0059]

[0060] In the formula, K ω This is the bandwidth matrix of the angular velocity loop.

[0061] The attitude angle loop is the slower loop in the system, providing commands to the angular rate loop. This loop does not require aerodynamic force or aerodynamic torque estimation, therefore conventional NDI control can be used, and the corresponding control law is:

[0062]

[0063] In the formula, Ω = [φ, θ, ψ] represents the three-axis attitude angles, and the subscript c indicates that it is a command signal. c For attitude angle commands, K Ω This is the attitude angle loop bandwidth matrix;

[0064] The slow loop is used to control the aircraft's speed and airflow angle. This loop includes aerodynamic calculations, therefore the INDI method is used for control. Similarly, a first-order Taylor expansion at the state point (ω0, δ0) yields:

[0065]

[0066] In the formula, V b =[u b v b w b ] T The three-axis velocities are in the body coordinate system. The corresponding acceleration is given by F, where m is the mass of the machine body. b =[F x F y F z ] T The resultant force acts on the three axes of the engine system, δ represents the aerodynamic control surfaces and thrust vector control quantities. The subscript 0 represents the state point in Taylor expansion.

[0067] Therefore, the control law of this loop is obtained after simplification:

[0068]

[0069] In the formula, F ac For the control force generated by the aerodynamic control surfaces, T b The control force generated by the thrust vector.

[0070] To address the angular acceleration estimation problem, this invention proposes a fixed-time disturbance observer (FTDO) for estimation. Compared to the traditional sliding mode disturbance observer, since the sign function is in the integral term, the estimated value is continuous, eliminating the chattering problem. Furthermore, the upper bound of the convergence time does not depend on the initial state but is determined by the control parameters. The observer has the following form:

[0071]

[0072] In the formula, Let x and d be the estimated values ​​of state x and uncertainty d, respectively. To estimate the error, the function sig(y) r Defined as sig(y) r =sgn(y)|y| r The coefficients k1,k2,k3 > 0, 0 < λ1 < 1, λ2 > 1, and the uncertainty has an upper bound satisfying the following condition.

[0073] By combining equations (2) and (5), the actual control quantity δ can be obtained, and then the improved Grey Wolf optimization method needs to be used to solve it.

[0074] The main factors to be considered during the transition process include:

[0075] 1) Can the aircraft accurately track control commands?

[0076] 2) The calculated control variables satisfy physical constraints, such as rate constraints, saturation constraints, and bandwidth constraints;

[0077] 3) Prioritize the use of aerodynamic control surfaces.

[0078] Therefore, the corresponding optimization objective function is set as follows:

[0079]

[0080]

[0081] dδ=δ-δ0

[0082]

[0083]

[0084]

[0085] In the formula, B eHere, v represents the virtual control command, dδ represents the control increment, δ represents the control variables (including 11 items: aileron, elevator, rudder, lift fan longitudinal deflection angle, lift fan lateral deflection angle, lift fan thrust, exhaust nozzle longitudinal deflection angle, exhaust nozzle lateral deflection angle, exhaust nozzle thrust, left roll nozzle thrust, right roll nozzle thrust, etc.), and δ0 represents the current control vector. δ is the position penalty parameter. i,min δ represents the minimum position of the i-th control variable. i,max Let ΔT be the maximum position of the i-th control variable, and ΔT be the control period. For rate penalty parameters, The minimum rate of the i-th control variable. λ is the maximum rate of the i-th control variable. i and η i These are the position proportional coefficient and the rate proportional coefficient for the i-th control variable, respectively.

[0086] The first item is the command tracking accuracy index, the second item is the actuator position constraint index, the third item is the actuator rate constraint index, and the last item is the thrust vector constraint index.

[0087] To solve the above-mentioned objective function, this invention employs the Improved Grey Wolf Optimizer (IGWO) algorithm, the algorithm flow of which is as follows: Figure 1 As shown below, the algorithm will be explained.

[0088] Population initialization: Generally, GWO uses random initialization, which may lead to initial individuals clustering in a small area or being located far from the optimal solution, making the algorithm prone to getting trapped in local optima, especially when solving high-dimensional optimization problems (11 dimensions in this invention). Therefore, this invention proposes a method of initializing the population by superimposing a set of optimal points with random variables. The set of optimal points containing N points is represented as follows:

[0089] P n (i)=U L +{γ i *n}*(U H -U L ) i=1,2,…,dn=1,2,…N (8)

[0090]

[0091] In the formula, d represents the dimension of the problem to be solved, and U L U H Let represent the upper and lower bounds of the corresponding dimension, respectively; N be the number of gray wolves; q be the smallest prime number satisfying q-3≥2d; {γ i *n} represents taking γi The decimal part of *n};

[0092] The superposition of random variables is expressed as:

[0093]

[0094] In the formula, ε is a random number between 0 and 1.

[0095] Calculating individual fitness: In the general GWO algorithm, the position update formulas for α, β, δ, and ω wolves in the wolf pack are as follows:

[0096]

[0097] It can be seen that the basic GWO algorithm updates the position of ω wolf based on the weights of the positions of the three leader wolves. The disadvantages are: 1. It directly accepts the new position without evaluating its merits, which reduces the search efficiency of the algorithm; 2. It does not use the optimal value to guide the search.

[0098] To utilize historical optimal information to guide the search process, this invention, referencing the Particle Swarm Optimization (PSO) algorithm, incorporates the individual's historical optimal position into the position update formula. Therefore, the position update strategies for α wolf, β wolf, and δ wolf are as follows:

[0099] X p (t+1)=X p (t)-A|CX p (t)-X(t)|+λ(X p,best -X p (t))p=α,β,δ(12)

[0100] In the formula, t is the current iteration number, and X is... p X(t) represents the prey's position, X(t) and X(t+1) represent the positions of individual wolves in the two iterations, A is the convergence factor, C is the oscillation factor, and λ is the individual's optimal position memory parameter.

[0101] This invention, based on the method for calculating the resultant force in the Artificial Potential Field (APF) method, defines the update strategy for ωwolf as follows:

[0102] X ω (t+1)=η α X α (t+1)+η β X β (t+1)+η δ X δ (t+1) (13)

[0103]

[0104] In the formula, η kThe weight coefficients for the three types of leader wolves are determined by weighting their respective fitness values ​​after normalization, thereby enhancing the guiding role of the optimal value in the search.

[0105] The fit function is the control part, that is, the optimization objective function of equation (7).

[0106] The convergence factor A and oscillation factor C in the basic GWO algorithm are calculated as follows:

[0107] A=(2r1-1)a (14)

[0108] C = 2r² (15)

[0109]

[0110] In the formula, r1 and r2 are both random coefficients between [0,1], and the distance control parameter a decreases linearly from 2 to 0 with the number of iterations.

[0111] The parameter 'a' plays a role in regulating the global and local search capabilities during the iteration process. Studies have found that using a nonlinear variation strategy for parameter 'a' can achieve better optimization results than a linear strategy.

[0112] Therefore, this invention proposes a nonlinear variation strategy based on the LeCun Tanh activation function of a neural network, namely:

[0113]

[0114] In the formula, a init a end These are the initial and final distance control parameters, respectively, where t is the current iteration number, and T is the distance control parameter at the beginning and end of the iteration. max Let tanh(x) be the total number of iterations, and let tanh(x) = (e^(x-1 x -e -x ) / (e x +e -x ).

[0115] a init a end The principle for choosing the value is to balance global and local search capabilities. If it's too large, the algorithm will initially jump too much within the feasible region; if it's too small, it will only jump locally. Based on these principles, and to ensure the algorithm retains a certain level of local search capability in later stages, this embodiment chooses 'a'. init =2,a end =0.1.

[0116] To address the issue of the basic GWO algorithm easily getting trapped in local optima in the later stages of the search, this invention introduces the fitness variance σ. 2 Average distance from the gray wolf When both are less than the critical value, it is considered that the algorithm may be trapped in a local optimum. Therefore, the half of the gray wolves with the worst fitness value ω is selected, and the Levy flight strategy is executed, i.e.:

[0117]

[0118]

[0119] In the formula, fit i and fit avg Here, X represents the fitness value of each gray wolf and the average fitness value of the wolf pack, where N is the number of gray wolves and X is the number of gray wolves. i Let ||X||2 represent the position of the gray wolf, and ||X||2 represent the 2-norm.

[0120]

[0121] In the formula, X ω,half To adapt the half of the gray wolves with poorer fitness values, μ is the step scaling factor, L is the step size coefficient, and L follows a Levy distribution, which is obtained by the following formula:

[0122]

[0123]

[0124]

[0125] In the formula, θ is a constant between 0 and 2; in this embodiment, θ is taken as 1.5, and N... norm This indicates a normal distribution.

[0126] The critical value is determined based on the desired accuracy and the properties of the objective function being solved, and is set manually during use. This invention, through experiments, found that the minimum value of the optimized objective function is generally around level 102. Therefore, the variance critical value is set to κ1 = 100, allowing the solutions to be distributed within ±20 of the minimum value. This preserves a certain degree of diversity, and the obtained solutions are acceptable to the control system. Similarly, the average distance... The critical value is set to κ2 = 10.

[0127] The method of this invention has a clear physical meaning, is easily understood by those skilled in the art, and has good application prospects. The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A transition control method for a vertical / short take-off and landing aircraft based on intelligent control distribution, characterized in that, The control law module and the distribution law module are included; The control law module divides state variables of the flight control system into multiple loops according to response speed, and designs the control law of each loop from inside to outside, wherein a fixed-time disturbance observer is used to estimate angular acceleration; The distribution law module uses an improved grey wolf optimization algorithm, including: initializing a population by using a method of superimposing random variables on a good point set, and then calculating individual fitness values; The method of superimposing random variables on the good point set to initialize the population includes: The good point set containing N points is expressed as: ， In the formula, For the dimension of the problem to be solved, These are the upper and lower limits for the corresponding dimensions. For the number of gray wolves, To meet The smallest prime number, Indicates taking The decimal part; After superimposing the random variables, it is expressed as: ， wherein is a random number between 0 and 1 ; The calculation of the individual fitness value includes: in a wolf pack wolf, wolf, wolf and The position update formula for the wolf is ， wherein, is the current iteration number, is the prey position, and is the position of the wolf pack individual in the previous iteration, is the convergence factor, is the swing factor, is the individual best position memory parameter; The three wolves are the historical optimum positions; ， In the formula, are the weight coefficients of the three leading wolves.

2. The transition control method for a vertical / short take-off and landing aircraft based on intelligent control distribution according to claim 1, characterized in that, The fixed-time disturbance observer is used to estimate the angular acceleration, and the fixed-time disturbance observer has the form of: ， wherein , are the estimates of the states and the uncertainty , respectively, is the estimation error, and the function is defined as , with the coefficients , , , the uncertainty satisfies .

3. The transition control method for a vertical / short take-off and landing aircraft based on intelligent control distribution according to claim 1, characterized in that, Convergence factor and the wobble factor The calculation method is as follows: ， with are both a random coefficient between is a distance control parameter; ， wherein , are the distance control parameters at the beginning and end, respectively, is the current iteration number, is the total number of iterations, the function .

4. The transition control method for a vertical / short take-off and landing aircraft based on intelligent control distribution according to claim 3, characterized in that, 。 5. The transition control method for a vertical / short takeoff and landing aircraft based on intelligent control distribution according to claim 1, characterized in that, Also include the calculation of fitness variance And the average distance of gray wolves When both are less than the critical value, the algorithm is trapped in local optimum, and the half of the gray wolves with the worst fitness value are selected Levy flight strategy is executed.

6. The transition control method for a vertical / short take-off and landing aircraft based on intelligent control distribution according to claim 5, characterized in that, the calculated fitness variance and gray wolf average distance is: ， wherein and is the fitness value for each wolf and the average fitness value of the wolf pack, is the number of wolves, is the position of the wolf, denotes the 2-norm.

7. The transition control method for a smart control distribution based vertical / short takeoff and landing aircraft according to claim 6, wherein, said selected half of the grey wolves having the worst fitness values performing a levy flight strategy comprises: ， wherein for the half of the gray wolves with worse values, is a step size scaling factor, is a step size coefficient, obeys a Levy distribution, the Levy distribution being obtained from the following equation: ， ， wherein is a constant between 0 and 2, denotes a normal distribution.

Images