An incremental local weighting-based unmanned aerial vehicle angular motion control method

Abstract

The application discloses a kind of unmanned plane angular motion control methods based on incremental local weighting, first based on inverse system idea, the local approximate inverse model of unmanned plane inner loop is constructed, and the control law of unmanned plane inner loop is obtained;Then based on local weighting learning method, the parameter adaptive updating law of unmanned plane inner loop is designed, and the regression parameter is updated, the real-time approximation of local approximate inverse model is realized;Then incremental local weighting method is used to design structure adaptive updating law, and new local approximate inverse model is automatically added when facing unknown characteristic data;Finally, the outer loop controller of the angular motion control of unmanned plane is designed, and the attitude angular velocity command is input to the unmanned plane inner loop;The application utilizes the method that local weighting learning and dynamic inverse are combined to carry out feedback linearization to system, improves the angular motion control precision of unmanned plane, and utilizes incremental local weighting method, and local model is automatically optimized, and the covering of new learning result to original result is avoided.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) flight control technology, and mainly to a UAV angular motion control method based on incremental local weighting. Background Technology

[0002] Because the nonlinearity and coupling effects during UAV maneuvering cannot be ignored, the state parameters of sample points differ greatly under different flight conditions, resulting in poor performance when learning new unknown feature data.

[0003] In existing technologies, the neural network compensation dynamic inverse design method is designed for landing processes with relatively stable attitudes. However, it does not consider the problem that the characteristic parameters of the UAV vary greatly under different flight states during maneuvering flight. Learning new unknown data will overwrite the network's existing learning results, and repeated learning increases the computation time and reduces the control effect. Summary of the Invention

[0004] Purpose of the invention: To address the problems existing in the background technology, the present invention provides a UAV angular motion control method based on incremental local weighting. By using the incremental local weighting method, the local model can be automatically optimized, avoiding the overwriting of the original results by the new learning results, realizing incremental updating of the model structure, and improving the angular motion control accuracy of the UAV.

[0005] Technical solution: To achieve the above objectives, the technical solution adopted by this invention is as follows:

[0006] A method for angular motion control of a UAV based on incremental local weighting includes the following steps:

[0007] Step S1: Based on the concept of inverse systems, construct a local approximate inverse model of the UAV's internal loop and obtain the control law of the UAV's internal loop;

[0008] Step S2: Design the parameter adaptive update law of the UAV inner loop based on the local weighted learning method, update the regression parameters, and realize the real-time approximation of the local approximate inverse model;

[0009] Step S3: Based on the parameter adaptive update law described in step S2, an incremental local weighting method is used to design the structural adaptive update law, which automatically adds a new local approximate inverse model when facing unknown feature data.

[0010] Step S4: Design an external loop controller for UAV angular motion control based on the dynamic inverse principle, and input attitude angular velocity commands to the UAV internal loop.

[0011] Furthermore, the specific method for establishing the local approximate inverse model in step S1 is as follows:

[0012] Step S1.1, the internal loop expression of the UAV is as follows:

[0013]

[0014] Where x = [pqr] T Let p, q, and r represent the roll rate, pitch rate, and yaw rate of the UAV, respectively; f(x) = [F1 F2 F3] T Let g(x) be a nonlinear function matrix and g(x) be an instruction allocation matrix; u = [δ a δ e δ r ] T F1, F2, F3, and g(x) represent the deflection values ​​of the aileron, elevator, and rudder, respectively; their specific expressions are as follows:

[0015]

[0016]

[0017]

[0018]

[0019] F1, F2, and F3 all include the torque terms generated by the airflow angle and angular velocity during the UAV's flight, as well as the inertial coupling torque term; I x I y and I z In order, around x b y b and z b Moment of inertia of the shaft, Q, S, c A b represents dynamic pressure, wing reference area, wing mean aerodynamic chord length, and wing span, respectively. All are aerodynamic moment coefficients, where α represents the angle of attack and β represents the sideslip angle;

[0020] The expression for the internal loop control law of the UAV is as follows:

[0021]

[0022] Step S1.2: Based on the concept of inverse systems, construct a local approximate inverse model as follows:

[0023]

[0024] where u=[δ a δ e δ r ] T f(x) = [F1 F2 F3] T , K is the gain matrix, x c The reference value for the triaxial angular velocity, x r Indicates the output value of the external loop;

[0025] Step S1.3: Based on the local weighted learning algorithm, construct multiple local approximation models for f(x) and g(x) using a Gaussian kernel function, linearize the parameters, and the normalized weighted sum of all local approximation inverse models is: and That is, the approximation values ​​of f(x) and g(x); considering the approximation error, it is expressed in the following form:

[0026]

[0027]

[0028] Where k represents the construction of k local approximation models, y f，k and y g，k Let w represent the k-th local approximate inverse model. k (x) is the Gaussian kernel function, representing the weighting coefficients, specifically expressed as follows:

[0029]

[0030] Where c k D is the center of the Gaussian function of the k-th local approximate inverse model. k Distance measure representing the Gaussian kernel function;

[0031] Represents the normalized weighting coefficient; Δ f (x) and Δ g (x) represents the approximation error of f(x) and g(x), η f and η g The actual regression parameter matrix of the local approximate inverse model is represented by η. f，k and η g，k Let f(x) and g(x) represent the regression parameters corresponding to f(x) and g(x) in each local approximate inverse model, respectively; φ(x) is the simplified and normalized weighted basis function matrix, denoted by φ. k This represents each weighted basis function in the matrix; specifically,

[0032]

[0033]

[0034] for The local approximation model takes triaxial angular velocity, angle of attack α, sideslip angle β, and approximation deviation as inputs. The local approximation model takes the angle of attack α and the approximation deviation as inputs; from the above formula, it can be seen that... This approximation is equal to the local approximate inverse model. The weighted output, similarly This approximation is equal to the local approximate inverse model. Weighted output;

[0035] Step S1.4: According to the local weighted algorithm, the dynamic equation of the system error is as follows:

[0036]

[0037] Where K is the gain term, let and The error between the regression parameters used to approximate the local approximate inverse model and the actual required regression parameters can be expressed as follows:

[0038]

[0039] Furthermore, the parameter adaptive update law design method for the UAV internal loop in step S2 is as follows:

[0040] Based on the system error dynamic equation obtained in step S1, and considering the real-time approximation performance of the local approximate inverse model, the regression parameter η is... f，k and η g，k The regression parameter matrix is ​​updated in real time; the expression is as follows:

[0041]

[0042] The weighted squared prediction error function J for each local approximate inverse model in a continuous-time system k The expression is as follows:

[0043]

[0044] in From the original model The expression can be obtained as follows:

[0045]

[0046] in It relates to the center value of each local approximate inverse model;

[0047] Find the relationship between the two sides of the prediction error function. The derivative, i.e. The cost function J can be obtained. k The minimum regression parameter update law is expressed as follows:

[0048]

[0049]

[0050] By adding a projection operator to improve the convergence speed, the adaptive update law for local parameters can be obtained as follows:

[0051]

[0052] in, For Θ k The approximate derivative, Let λ represent the approximation error of the k-th local approximate inverse model, where λ is a constant greater than 0.

[0053] Furthermore, the structural adaptive update law design method in step S3 is as follows:

[0054] When w k <w d When the weight coefficients of the local approximate inverse model are less than the preset critical weights, none of the existing local approximate inverse models can be activated, and a new local model needs to be added. The center point values ​​of the basis functions of the new local model remain unchanged, and the distance measure D of the basis functions is changed. N+1 To update, use the following expression:

[0055]

[0056] D N+1 Using Cholesky decomposition, we construct the product of an upper triangular matrix ξ and its transpose. The adaptive update law for ξ is as follows:

[0057]

[0058] in, The weighted squared prediction error function J mentioned above represents... k The cross-validation cost function after adding a penalty term, where W represents the weight coefficient w of each local model. k The integral over time t, where γ represents the coefficient of the penalty term and a represents the learning rate; by updating D N+1 Add a local model and initialize it, then adaptively adjust the effective range of the new local model.

[0059] Furthermore, the specific method for designing the external loop controller for UAV angular motion control based on the dynamic inverse principle in step S4 is as follows:

[0060] Based on the dynamic inverse principle, the relationship between the roll angle φ, pitch angle θ, sideslip angle β, and attitude angular velocities p, q, and r of the angular motion command signals is linearized through feedback, resulting in the following expression for the relationship between the two:

[0061]

[0062]

[0063] Among them, G ya = mg(cosα sinβsinθ+cosβsinφcosθ-sinαsinβcosφcosθ), where T is the engine thrust, α is the angle of attack, Y is the side force, m is the UAV mass, V is the UAV airspeed, and K is the UAV airspeed. φ K θ and K β These are the bandwidths of the roll angle loop, pitch angle loop, and sideslip angle loop, respectively.

[0064] Beneficial effects:

[0065] The incremental local weighted UAV angular motion control method provided by this invention addresses the problems of nonlinearity and non-negligible coupling effects during UAV maneuvering flight. It utilizes a combination of local weighted learning and dynamic inverse to linearize the system through feedback, thereby improving the angular motion control accuracy of the UAV.

[0066] Furthermore, to address the issue of large differences in feature parameters of UAVs under different flight states during maneuvering flight and poor performance in learning new unknown feature data, an incremental local weighting method is used to automatically optimize the local model, thus avoiding the overwriting of existing results by new learning results. Attached Figure Description

[0067] Figure 1 This is the flowchart 56FE of the incremental locally weighted UAV angular motion control method provided by the present invention;

[0068] Figure 2 This is a schematic diagram of the three-channel step response of the incremental local weighted controller under parameter perturbation in an embodiment of the present invention;

[0069] Figure 3 This is a response diagram of the horizontal S-curve motion loop under crosswind interference in an embodiment of the present invention;

[0070] Figure 4 This is a block diagram of the incremental locally weighted UAV angular motion controller provided by the present invention. Detailed Implementation

[0071] The present invention will be further described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0072] This invention provides a method for controlling the angular motion of a UAV based on incremental local weighting, such as... Figure 1As shown, it includes the following steps:

[0073] Step S1: To address the issues of significant dynamic coupling effects and poor learning performance of new, unknown feature data during UAV maneuvering, a local approximate inverse model of the UAV's internal loop is constructed based on the inverse system concept to obtain the UAV's internal loop control law. Specifically,

[0074] Step S1.1, the internal loop expression of the UAV is as follows:

[0075]

[0076] Where x = [pqr] T Let p, q, and r represent the roll rate, pitch rate, and yaw rate of the UAV, respectively; f(x) = [F1 F2 F3] T Let g(x) be a nonlinear function matrix and g(x) be an instruction allocation matrix; u = [δ a δ e δ r ] T F1, F2, F3, and g(x) represent the deflection values ​​of the aileron, elevator, and rudder, respectively; their specific expressions are as follows:

[0077]

[0078]

[0079]

[0080]

[0081] F1, F2, and F3 all include the torque terms generated by the airflow angle and angular velocity during the UAV's flight, as well as the inertial coupling torque term. The expressions for F1, F2, and F3 are similar in structure: the first term is the additional torque generated by inertial coupling, the second term is the aerodynamic torque term, and the third term is the damping torque term. x I y and I z In order, around x b y b and z b Moment of inertia of the shaft, Q, S, c A b represents dynamic pressure, wing reference area, wing mean aerodynamic chord length, and wing span, respectively. All are aerodynamic moment coefficients, where α represents the angle of attack and β represents the sideslip angle;

[0082] The expression for the internal loop control law of the UAV is as follows:

[0083]

[0084] Step S1.2: Based on the concept of inverse systems, construct a local approximate inverse model as follows:

[0085]

[0086] where u=[δ a δ e δ r ] T f(x) = [F1 F2 F3] T , K is the gain matrix, x c The reference value for the triaxial angular velocity, x r Indicates the output value of the external loop;

[0087] Step S1.3: Based on the local weighted learning algorithm, construct multiple local approximation models for f(x) and g(x) using a Gaussian kernel function, linearize the parameters, and the normalized weighted sum of all local approximation inverse models is: and That is, the approximation values ​​of f(x) and g(x); considering the approximation error, it is expressed in the following form:

[0088]

[0089]

[0090] Where k represents the construction of k local approximation models, y f，k and y g，k Let w represent the k-th local approximate inverse model. k (x) is the Gaussian kernel function, representing the weighting coefficients, specifically expressed as follows:

[0091]

[0092] Where c k D is the center of the Gaussian function of the k-th local approximate inverse model. k Distance measure representing the Gaussian kernel function;

[0093] Δ represents the normalized weight coefficients, ensuring that each local model has a significant impact only within a portion of its receptive field; f (x) and Δ g (x) decibels represent the approximation error of f(x) and g(x), η f and η g The actual regression parameter matrix of the local approximate inverse model is represented by η. f，k and η g，kLet f(x) and g(x) represent the regression parameters corresponding to f(x) and g(x) in each local approximate inverse model, respectively; φ(x) is the simplified and normalized weighted basis function matrix, denoted by φ. k This represents each weighted basis function in the matrix; specifically,

[0094]

[0095]

[0096] for The local approximation model takes triaxial angular velocity, angle of attack α, sideslip angle β, and approximation deviation as inputs. The local approximation model takes the angle of attack α and the approximation deviation as inputs; from the above formula, it can be seen that... This approximation is equal to the local approximate inverse model. The weighted output, similarly This approximation is equal to the local approximate inverse model. Weighted output;

[0097] Step S1.4: According to the local weighted algorithm, the dynamic equation of the system error is as follows:

[0098]

[0099] Where K is the gain term, let and The error between the regression parameters used to approximate the local approximate inverse model and the actual required regression parameters can be expressed as follows:

[0100]

[0101] Step S2: After obtaining the error dynamic equation, considering the real-time approximation performance of the local model, an adaptive update law for the parameters of the inner loop controller is designed to achieve online approximation of the local model.

[0102] Based on the system error dynamic equation obtained in step S1, and considering the real-time approximation performance of the local approximate inverse model, the regression parameter η is... f，k and η g，k The regression parameter matrix is ​​updated in real time; the expression is as follows:

[0103]

[0104] The weighted squared prediction error function J for each local approximate inverse model in a continuous-time system k The expression is as follows:

[0105]

[0106] in From the original model The expression can be obtained as follows:

[0107]

[0108] in It relates to the center value of each local approximate inverse model;

[0109] Find the relationship between the two sides of the weighted squared prediction error function. The derivative, i.e. The cost function J can be obtained. k The minimum regression parameter update law is expressed as follows:

[0110]

[0111]

[0112] By adding a projection operator to improve the convergence speed, the adaptive update law for local parameters can be obtained as follows:

[0113]

[0114] in, For Θ k The approximate derivative, Let λ represent the approximation error of the k-th local approximate inverse model, where λ is a constant greater than 0.

[0115] Step S3: Based on the parameter adaptive update law described in step S2, an incremental local weighting method is used to design the structural adaptive update law, which automatically adds a new local approximate inverse model when facing unknown feature data.

[0116] Based on step S2, an incremental local weighting method is used to design the structural adaptive update law, and a given attitude angular velocity command is input into the inner loop. When w k <w d In other words, when the weight coefficients of the local approximate inverse model are less than the set critical weights, the existing local approximate inverse models cannot be activated, thus requiring the addition of a new local model. The center point values ​​of the basis functions in the new local model remain unchanged, while the distance measure D of the basis functions... N+1 Update. Distance measure D N+1 The index representing the proportion of data point x within the validity range of each linear model determines the size and shape of the receptive field of the local model, and its expression is as follows:

[0117]

[0118] D N+1Using Cholesky decomposition, we construct the product of an upper triangular matrix ξ and its transpose. The adaptive update law for ξ is as follows:

[0119]

[0120] in, This represents the cross-validation cost function after adding a penalty term to the weighted squared prediction error function mentioned above, where W represents the weight coefficient w of each local model. k The integral over time t, where γ represents the coefficient of the penalty term and a represents the learning rate; by updating D N+1 Add a local model and initialize it, then adaptively adjust the effective range of the new local model.

[0121] Step S4: Design an external loop controller for UAV angular motion control based on the dynamic inverse principle, and input attitude angular velocity commands to the UAV internal loop.

[0122] Based on the dynamic inverse principle, the relationship between the roll angle φ, pitch angle θ, sideslip angle β, and attitude angular velocities p, q, and r of the angular motion command signals is linearized through feedback, resulting in the following expression for the relationship between the two:

[0123]

[0124]

[0125] Among them, G ya = mg(cosαsinβsinθ+cosβsinφcosθ-sinαsinβcosφcosθ), where T is the engine thrust, α is the angle of attack, Y is the side force, m is the UAV mass, V is the UAV airspeed, and K is the engine thrust. φ K θ and K β These represent the bandwidths of the roll angle loop, pitch angle loop, and sideslip angle loop, respectively. The overall UAV angular motion controller block diagram is shown below. Figure 4 As shown.

[0126] The UAV angular motion control method provided by this invention not only decouples the nonlinear and coupling effects generated during UAV maneuvering, but also exhibits strong robustness to internal model uncertainties and different flight states. It avoids new learning results overwriting existing results, allows for rapid adaptive adjustment of state parameters, and possesses good control accuracy. When parameter perturbations exist, a square wave signal with an amplitude of 1 is applied to the roll angle, pitch angle, and sideslip angle, resulting in the following... Figure 2 As shown, when the drone performs a horizontal S-curve maneuver under crosswind interference after flying for a period of time, the three-channel response is as follows: Figure 3As shown, it is clear that the drone can stably complete flight maneuvers after rapid adaptive adjustments.

[0127] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for angular motion control of a UAV based on incremental local weighting, characterized in that, Includes the following steps: Step S1: Based on the concept of inverse systems, construct a local approximate inverse model of the UAV's internal loop and obtain the control law of the UAV's internal loop; Step S2: Design the parameter adaptive update law of the UAV inner loop based on the local weighted learning method, update the regression parameters, and realize the real-time approximation of the local approximate inverse model; Step S3: Based on the parameter adaptive update law described in step S2, an incremental local weighting method is used to design the structural adaptive update law, which automatically adds a new local approximate inverse model when facing unknown feature data. Step S4: Design an external loop controller for UAV angular motion control based on the dynamic inverse principle, and input attitude angular velocity commands to the UAV internal loop.

2. The UAV angular motion control method based on incremental local weighting according to claim 1, characterized in that, The specific method for establishing the local approximate inverse model in step S1 is as follows: Step S1.1, the internal loop expression of the UAV is as follows: Where x = [pqr] T Let p, q, and r represent the roll rate, pitch rate, and yaw rate of the UAV, respectively; f(x) = [F1 F2 F3] T Let g(x) be a nonlinear function matrix and g(x) be an instruction allocation matrix; u = [δ a δ e δ r ] T F1, F2, F3, and g(x) represent the deflection values ​​of the aileron, elevator, and rudder, respectively; their specific expressions are as follows: F1, F2, and F3 all include the torque terms generated by the airflow angle and angular velocity during the UAV's flight, as well as the inertial coupling torque term; I x I y and I z In order, around x b y b and z b Moment of inertia of the shaft, Q, S, c A b represents, in order, dynamic pressure, wing reference area, wing mean aerodynamic chord length, and wing span. All are aerodynamic moment coefficients, where α represents the angle of attack and β represents the sideslip angle; The expression for the internal loop control law of the UAV is as follows: Step S1.2: Based on the concept of inverse systems, construct a local approximate inverse model as follows: where u=[δ a δ e δ r ] T f(x) = [F1 F2 F3] T , K is the gain matrix, x c The reference value for the triaxial angular velocity, x r Indicates the output value of the external circuit; Step S1.3: Based on the local weighted learning algorithm, construct multiple local approximation models for f(x) and g(x) using a Gaussian kernel function, linearize the parameters, and the normalized weighted sum of all local approximation inverse models is: and That is, the approximation values ​​of f(x) and g(x); considering the approximation error, it is expressed in the following form: Where k represents the construction of k local approximation models, y f，k and y g，k Let w represent the k-th local approximate inverse model. k (x) is the Gaussian kernel function, representing the weighting coefficients, specifically expressed as follows: Where c k D is the center of the Gaussian function of the k-th local approximate inverse model. k Distance measure representing the Gaussian kernel function; Represents the normalized weighting coefficient; Δ f (x) and Δ g (x) represents the approximation error of f(x) and g(x), η f and η g The actual regression parameter matrix of the local approximate inverse model is represented by η. f，k and η g，k Let f(x) and g(x) represent the regression parameters corresponding to f(x) and g(x) in each local approximate inverse model, respectively; φ(x) is the simplified and normalized weighted basis function matrix, denoted by φ. k This represents each weighted basis function in the matrix; specifically, for The local approximation model takes triaxial angular velocity, angle of attack α, sideslip angle β, and approximation deviation as inputs. The local approximation model takes the angle of attack α and the approximation deviation as inputs; from the formula, we know that... This approximation is equal to the local approximate inverse model. The weighted output, similarly This approximation is equal to the local approximate inverse model. Weighted output; Step S1.4: The dynamic equation of the system error is obtained according to the local weighting algorithm as follows: Where K is the gain term, let and The error between the regression parameters used to approximate the local approximate inverse model and the actual required regression parameters is expressed as follows:

3. The UAV angular motion control method based on incremental local weighting according to claim 2, characterized in that, The parameter adaptive update law design method for the UAV internal loop in step S2 is as follows: Based on the system error dynamic equation obtained in step S1, and considering the real-time approximation performance of the local approximate inverse model, the regression parameter η is... f，k and η g，k The regression parameter matrix is ​​updated in real time; the expression is as follows: The weighted squared prediction error function J for each local approximate inverse model in a continuous-time system k The expression is as follows: in From the original model The expression is as follows: in It relates to the center value of each local approximate inverse model; Find the relationship between the two sides of the weighted squared prediction error function. The derivative, i.e. Find the cost function J k The minimum regression parameter update law is expressed as follows: Adding a projection operator to improve the convergence speed yields the local parameter adaptive update law as follows: in, For Θ k The approximate derivative, Let λ represent the approximation error of the k-th local approximate inverse model, where λ is a constant greater than 0.

4. The UAV angular motion control method based on incremental local weighting according to claim 3, characterized in that, The structural adaptive update law design method in step S3 is as follows: When w k <w d When the weight coefficients of the local approximate inverse model are less than the preset critical weights, none of the existing local approximate inverse models can be activated, and a new local model needs to be added. The center point values ​​of the basis functions of the new local model remain unchanged, and the distance measure D of the basis functions is changed. N+1 To update, use the following expression: D N+1 Using Cholesky decomposition, we construct the product of an upper triangular matrix ξ and its transpose. The adaptive update law for ξ is as follows: in, J represents the weighted squared prediction error function. k The cross-validation cost function after adding a penalty term, where W represents the weight coefficient w of each local model. k The integral over time t, where γ represents the coefficient of the penalty term and a represents the learning rate; by updating D N+1 Add a local model and initialize it, then adaptively adjust the effective range of the new local model.

5. The UAV angular motion control method based on incremental local weighting according to claim 4, characterized in that, The specific method for designing the external loop controller for UAV angular motion control based on the dynamic inverse principle in step S4 is as follows: Based on the dynamic inverse principle, the relationship between the roll angle φ, pitch angle θ, sideslip angle β, and attitude angular velocities p, q, and r of the angular motion command signals is linearized through feedback, resulting in the following expression for the relationship between the two: Among them, G ya = mg(cosαsinβsinθ+cosβsinφcosθ-sinαsinβcosφcosθ), where T is the engine thrust, α is the angle of attack, Y is the side force, m is the UAV mass, V is the UAV airspeed, and K is the engine thrust. φ K θ and K β These are the bandwidths of the roll angle loop, pitch angle loop, and sideslip angle loop, respectively.

Images