Quadrotor unmanned aerial vehicle near-ground flight control method based on physical information neural network

Abstract

The application discloses a quad-rotor unmanned aerial vehicle near-ground flight control method based on a physical information neural network, and belongs to the technical field of unmanned aerial vehicle automatic control; a dynamic model of a quad-rotor unmanned aerial vehicle is established, and a motion equation containing ground effect is established; a geometric controller is designed as a benchmark controller to realize basic flight control; in a simulation environment, the benchmark controller is used to collect unmanned aerial vehicle near-ground flight state parameters; a physical information neural network structure is constructed, a loss function is designed, network training is realized by adjusting the weight of an explicit physical information loss term; a composite anti-interference controller integrating the physical information neural network and an environmental disturbance observer is designed to realize unmanned aerial vehicle near-ground flight control. The application adopts a model and data hybrid driving method, accurately estimates the ground effect through the physical information neural network, and realizes composite anti-interference control in combination with the environmental disturbance observer, so that the control precision and system stability of the unmanned aerial vehicle in near-ground flight are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) automatic control technology, and in particular to a near-ground flight control method for a quadcopter UAV based on a physical information neural network. Background Technology

[0002] Multirotor drones are significantly affected by the ground effect when flying near the ground. The ground effect refers to the phenomenon where the downwash airflow from the rotor interacts with the ground as the multirotor drone approaches boundaries such as the ground or walls, resulting in increased lift. This affects the accuracy of the drone's near-ground flight control. Current research on methods to address this phenomenon mainly includes:

[0003] 1) Physical model-based approach. This approach describes the disturbance characteristics by establishing a mathematical analytical model of the ground effect. However, due to the simplified nature of the model and the complex nonlinear properties of the ground effect disturbance itself, establishing an accurate physical model is quite difficult.

[0004] 2) Interference observer-based method. Ground effects and other interferences are treated as lumped interferences and estimated using interference observers. However, due to the different characteristics of each interference, the estimation accuracy is difficult to guarantee.

[0005] 3) Deep learning-based methods. These methods use neural networks to fit varying aerodynamic coefficients, but purely data-driven approaches lack physical constraints, have limited generalization capabilities, and only consider a limited range of interference scenarios.

[0006] Therefore, there is an urgent need for a method that can combine prior physical knowledge and deep learning to achieve near-ground flight control of multi-rotor UAVs. Summary of the Invention

[0007] The purpose of this invention is to propose a near-ground flight control method for quadrotor unmanned aerial vehicles (UAVs) based on a physical information neural network to solve the problems mentioned in the background art. This invention improves the accuracy of ground effect prediction and the interpretability of the neural network model by introducing physical model constraints of ground effect interference into the neural network training process, and achieves composite anti-interference control by combining it with an environmental interference observer.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A near-ground flight control method for quadrotor UAVs based on physical information neural networks includes the following steps:

[0010] S1. Perform dynamic modeling on the quadcopter drone;

[0011] S2. Design a geometric controller as a reference controller;

[0012] S3. Data acquisition is performed in a simulation environment using a reference controller;

[0013] S4. Design the physical information neural network structure and loss function, and train it.

[0014] S5. Design a composite anti-interference controller that integrates a physical information neural network and an environmental interference observer.

[0015] Preferably, S1 specifically includes the following:

[0016] A translational dynamics model of a quadrotor UAV under near-ground flight conditions is established based on the Newton-Euler method:

[0017]

[0018] Where m represents the mass of the drone; The vector represents the linear acceleration of the drone; g represents the gravitational acceleration; e3 represents the vector [0,0,1]. T ;F d d represents the expected thrust of the drone when it is far from the ground; ge Δ represents the ground effect interference force experienced by the drone; Δ represents other unknown environmental interference.

[0019] Preferably, the reference controller in S2 includes a position loop controller and an attitude loop controller, wherein:

[0020] The function of the position loop controller is expressed as follows:

[0021]

[0022] Among them, e x e v These are the error terms for position and velocity, respectively; k x k v and k int All are positive gain coefficients; The desired acceleration; sat κ (·) is the saturation function; taking the input ζ as an example, specifically:

[0023]

[0024] Where κ is the saturation value; when the input ζ is a vector, the saturation function is applied to each element of ζ, and the final result is F. d For the desired thrust;

[0025] The function representation of the attitude loop controller is as follows:

[0026]

[0027] Among them, eR e ω These are the attitude tracking error term and the angular velocity tracking error term, respectively, where ω represents the body angular velocity; k R k ω All are positive gain coefficients; J represents the inertial matrix of the UAV; [·] × R represents an antisymmetric matrix; R represents a rotation matrix, and M represents an antisymmetric matrix. d This represents the desired triaxial torque.

[0028] Preferably, the data collected in S3 are the state parameters of the UAV's near-ground trajectory tracking flight, and the state parameters include:

[0029] ① The relative ground height of the four propellers; ② The three-axis velocity of the UAV; ③ The three-axis angular velocity of the UAV; ④ The attitude quaternion of the UAV; ⑤ The input PWM signals of the four motors.

[0030] Preferably, S4 specifically includes the following:

[0031] S4.1 Construct a neural network structure, wherein the neural network structure includes:

[0032] ① Input layer: has 18 neurons, used to receive the state parameters;

[0033] ② Three hidden layers: containing 30, 35, and 20 neurons respectively, with ReLU activation function;

[0034] ③ Output layer: has 4 neurons, used to output the estimated ground effect disturbance force on the four propellers.

[0035] S4.2 Design the loss function, which includes:

[0036] ① Implicit physical information loss term:

[0037]

[0038] Where, N s The number of samples is represented by ; i represents the sample number; j = {1, 2, 3, 4} represents the four propellers; This represents the theoretical value of the ground effect disturbance force on the j-th propeller of the i-th sample. This represents the neural network's predicted output variable for the ground effect interference force on the j-th propeller of the i-th sample;

[0039] In a single sample i The calculation process is as follows: Based on the propeller model of a multi-rotor UAV, the thrust generated by propeller j in the region without ground effect is:

[0040]

[0041] Among them, C T It is a dimensionless constant, ρ represents air density, and N ij J represents the rotational speed of the propeller in revolutions per minute, and D represents the diameter of the propeller.

[0042] The thrust generated by propeller j near the ground is:

[0043]

[0044] in, This represents the actual lift coefficient of propeller j; Represents rotational speed in radians per second;

[0045] and There is a proportional relationship:

[0046]

[0047] Combining the above equations, we obtain:

[0048]

[0049] Choose a physical empirical formula:

[0050]

[0051] Among them, C a and C b The coefficients are determined based on the rotor blade geometry of the quadcopter UAV; R is the rotor radius; z ij This indicates the height of each propeller above the ground.

[0052] The ground effect disturbance force on propeller j was calculated as follows:

[0053]

[0054] ② Explicit physical information loss term:

[0055]

[0056] in, This represents the thrust generated by the j-th propeller of the i-th sample in the region without ground effect. For predicting the proportion; This is an empirical proportion term;

[0057] ③ Overall network loss function:

[0058] Loss = Loss1 + λ ca Loss2

[0059] Where, λ ca This represents the weight of the explicit physical information loss term Loss2;

[0060] S4.3 Network training includes the following steps:

[0061] S4.3.1 Design a cyclic annealing scheduling strategy and set... Where mod(·) represents the modulo operation, n represents the index of the current training epoch of the neural network; T represents the maximum number of training epochs of the neural network; and M represents the period of cyclic annealing.

[0062] Set λ within each annealing cycle ca Maintain the maximum value λ max The proportion is μ;

[0063] When β > μ,

[0064] When β≤μ, λ ca =λ max ;

[0065] This adjusts the weight λ of the explicit physical information loss term Loss2. ca ;

[0066] S4.3.2 Perform singular value decomposition on the weight matrix W of each layer in the neural network:

[0067] W=UΣV T

[0068] Where U is the left singular vector matrix; Σ is the singular value diagonal matrix, whose diagonal elements σ1, σ2, ..., σ3 are... r Satisfying σ1≥σ2≥…≥σ r ≥0, r is the rank of matrix Σ; V T It is the transpose of the right singular vector matrix;

[0069] The Lipschitz constant γ of the weight matrix W is equal to the maximum singular value σ1, thus constraining the maximum singular value; a singular value threshold σ is set. max If σ1>σ max Then W new =W*(σ max / σ1), thus the constraint guarantees the Lipschitz continuity of the network.

[0070] Preferably, the composite disturbance observer described in S5 is used to estimate other unknown environmental disturbances Δ besides the ground effect. Based on the quadcopter dynamics model constructed in S1, the state equation of the disturbance observer is designed as follows:

[0071]

[0072] Where ξ represents the intermediate observed variable; l = α / m, α is a positive gain coefficient; p(v) = α*v, used to estimate other unknown disturbances:

[0073] Composite anti-interference control is achieved by integrating a physical information neural network and an environmental interference observer. The position controller is designed as follows:

[0074]

[0075] in, This represents the ground effect interference force predicted by the physical information neural network; This represents other environmental disturbances estimated by the disturbance observer.

[0076] Compared with existing technologies, this invention provides a near-ground flight control method for quadrotor UAVs based on physical information neural networks, which has the following advantages:

[0077] (1) This invention improves the prediction accuracy of the neural network for ground effects and its generalization ability to datasets outside the training data by introducing physical model constraints.

[0078] (2) The present invention adopts a cyclic annealing scheduling strategy to effectively adjust the weight of the explicit physical information loss term;

[0079] (3) This invention combines an environmental interference observer to achieve synergistic compensation for ground effects and other unknown environmental interferences;

[0080] (4) Compared with the anti-interference control strategy based on interference observer, the composite anti-interference control method proposed in this invention has better control accuracy and robustness. Attached Figure Description

[0081] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings involved in the embodiments are now briefly described. Obviously, the drawings in the following description are merely illustrative of some embodiments of the present invention. For those skilled in the art, other forms of drawings can be constructed based on these drawings without creative effort.

[0082] Figure 1 This is a flowchart illustrating the UAV near-ground flight control method based on physical information neural network proposed in this invention.

[0083] Figure 2 This is a schematic diagram of the system architecture of the physical information neural network mentioned in Embodiment 1 of the present invention;

[0084] Figure 3This is a schematic diagram of the training process of the physical information neural network mentioned in Embodiment 1 of the present invention;

[0085] Figure 4 This is a schematic diagram of the near-ground flight closed-loop control system framework for the UAV mentioned in Embodiment 1 of the present invention. Detailed Implementation

[0086] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0087] Example 1:

[0088] This embodiment proposes a near-ground flight control method for quadrotor UAVs based on physical information neural networks, such as... Figure 1 As shown, the specific implementation process is as follows:

[0089] Step 1: Quadrotor Dynamics Modeling

[0090] Based on the Newton-Euler method, a translational dynamics model of a quadcopter UAV under near-ground multi-source interference is established:

[0091]

[0092] Where m represents the mass of the drone; Let represent the linear acceleration of the drone, g represent the gravitational acceleration, and e3 represent the vector [0,0,1]. T F d d represents the expected thrust of the drone. ge This represents the ground effect interference force experienced by the drone, while Δ represents other unknown environmental interference.

[0093] Step 2, Geometric Controller Design:

[0094] 1) Position loop controller:

[0095]

[0096] Among them, e x e v These are the error terms for position and velocity, respectively; k x k v and k int All are positive gain coefficients; The desired acceleration; sat κ (·) is the saturation function. Taking the input ζ as an example, it is as follows:

[0097]

[0098] Where κ is the saturation value; when the input ζ is a vector, the saturation function is applied to each element of ζ, and the final result is F. d For the desired thrust;

[0099] 2) Attitude loop controller:

[0100]

[0101] Among them, e R e ω These are the attitude tracking error term and the angular velocity tracking error term, respectively; ω represents the body angular velocity; k R k ω All are positive gain coefficients; J represents the inertial matrix of the UAV; [·] × R represents the antisymmetric matrix; R represents the rotation matrix, and finally, the desired triaxial torque M is obtained. d .

[0102] Step 3, Data Collection:

[0103] In this embodiment, the simulation environment uses Simulink in Matlab. A dynamic model of the quadcopter UAV is built in Simulink, and a ground effect parameter model is obtained by polynomial fitting of experimental data. Communication with FlightGear is established in Simulink for co-simulation. The UAV is controlled to fly in near-Earth orbit by the designed geometric controller, and flight state parameters are collected, including the ground clearance of each propeller, flight speed, angular velocity, attitude quaternion, and PWM signals of the four motors.

[0104] Step 4: Design and training of the physical information neural network:

[0105] Design a ground effect estimation method based on a physical neural network model, wherein the design and training of the network model specifically includes:

[0106] Overall network architecture:

[0107] like Figure 2 As shown, the physical information neural network includes:

[0108] 1) Input layer (18 neurons):

[0109] a) Propeller height input unit (4-dimensional): z1, z2, z3, z4;

[0110] b) Linear velocity input unit (3D): v x v y v z ;

[0111] c) Angular velocity input unit (3D): ω x ω y ω z ;

[0112] d) Attitude quaternion input units (4-dimensional): q0, q1, q2, q3;

[0113] e) Normalized motor PWM signal input unit (4-dimensional): n1, n2, n3, n4;

[0114] 2) Hidden layer:

[0115] a) First hidden layer (30 neurons), using ReLU activation function;

[0116] b) The second hidden layer (35 neurons) uses ReLU as the activation function;

[0117] c) The third hidden layer (20 neurons) uses ReLU as the activation function;

[0118] 3) Output layer (4 neurons):

[0119] Corresponding ground effect interference forces on the four propellers:

[0120] Loss function design:

[0121] 1) Implicit physical information loss term:

[0122]

[0123] Where, N s This represents the number of samples, where i represents the sample number, and j = {1, 2, 3, 4} represents the four propellers. This represents the theoretical value of the ground effect disturbance force on the j-th propeller of the i-th sample. The output variable representing the neural network's prediction of the ground effect disturbance force on the j-th propeller of the i-th sample;

[0124] In a single sample i The calculation process is as follows, based on the propeller model of a multi-rotor UAV, the thrust generated by propeller j in the region without ground effect:

[0125]

[0126] Among them, C T It is a dimensionless constant, where ρ represents air density and N is... ij The value of propeller j is expressed as revolutions per minute (rpm), and D represents the diameter of the propeller. The thrust generated by propeller j near the ground is:

[0127]

[0128] in, This represents the actual lift coefficient of propeller j. Represents rotational speed in radians per second; and There is a proportional relationship:

[0129]

[0130] Combining the above equations, we obtain:

[0131]

[0132] Choose a physical empirical formula:

[0133]

[0134] Among them, C a and C b These are coefficients determined based on the geometry of the quadcopter blades, where R is the rotor radius and z is the rotor radius. ij This represents the height of each propeller above the ground, and finally, the ground effect disturbance force on propeller j is calculated:

[0135]

[0136] 2) Explicit physical information loss term design. Taking a single sample i as an example, firstly, due to the actual thrust generated by each propeller near the ground... and the thrust they generate in the region without ground effect The following relationship exists between them:

[0137]

[0138] in, It is the ground effect interference force experienced by propeller j.

[0139] Divide both sides of equation (2) have to:

[0140]

[0141] Therefore, the prediction scaling term of the neural network is designed as follows:

[0142]

[0143] Empirical proportions containing physical information:

[0144]

[0145] In formula (4) Let represent the ground effect disturbance force experienced by propeller j as predicted by the network. Therefore, the physical information loss function for propeller j across all samples is as follows:

[0146]

[0147] in, This represents the thrust generated by the j-th propeller of the i-th sample in the region without ground effect.

[0148] 3) The overall network loss function is designed as follows:

[0149] Loss = Loss1 + λ ca Loss2 (7)

[0150] Where, λ ca This represents the weight of the explicit physical information loss term Loss2. The specific method for determining its value is explained in the cyclic annealing scheduling strategy.

[0151] Training strategy design:

[0152] like Figure 3 As shown, the strategy design during network training includes:

[0153] 1) Cyclic annealing scheduling strategy:

[0154] In this invention, a cyclic annealing strategy is used to adjust the weights of the explicit physical information loss term Loss2, guiding the network to fully explore the optimization space. During training, the hyperparameter λ is periodically adjusted. ca This causes the network to vary its emphasis on the two loss terms at different training stages. Specifically, within one annealing cycle, λ ca Starting from the maximum value, the value is gradually reduced to 0 before starting the next annealing cycle. This helps the network, when processing quadcopter UAV flight data containing complex nonlinear ground effects, to utilize the latent encoding learned in the previous annealing cycle as a warm restart, progressively improving performance and thus enabling it to better learn complex relationships in the data. ca The values ​​are shown in equation (8):

[0155]

[0156] Where, λ max It is the setting of λ ca The maximum value determines the maximum impact of the explicit physical information loss term Loss2 on the network during training; μ is used to determine λ within each annealing cycle. ca Maintain the proportion of the maximum value. The value of β is given by equation (9):

[0157]

[0158] Where n represents the index of the current training epoch of the neural network, T represents the maximum number of training epochs of the neural network, and M is the cycle of cyclic annealing. Therefore, the entire training process of the neural network is divided into M cycles, with λ for each cycle. ca Annealing from the maximum value to 0, the network's loss function is restored to a single implicit physical information loss term, Loss1.

[0159] 2) Weight constraint strategy:

[0160] In this invention, a weight constraint strategy is used to constrain the Lipschitz constant of the network layer weight matrix, enhancing the network model's generalization ability to unknown data. Specifically, singular value decomposition is performed on the weight matrix W of each layer in the neural network:

[0161] W=UΣV T

[0162] Where U is the left singular vector matrix, Σ is the singular value diagonal matrix, and its diagonal elements σ1, σ2, ..., σ3 are... r Satisfying σ1≥σ2≥…≥σ r ≥0 (r is the rank of matrix Σ), V T As the transpose of the right singular vector matrix, the Lipschitz constant γ of the weight matrix W is equal to the maximum singular value σ1. To ensure network stability, a constraint is imposed on the maximum singular value by setting a singular value threshold σ. max If σ1>σ max Then W new =W*(σ max / σ1), thus the constraint guarantees the Lipschitz continuity of the network.

[0163] Step 5: Composite anti-interference control

[0164] This embodiment proposes a composite anti-interference control system integrating a physical information neural network and an environmental interference observer. The framework of the entire UAV near-ground flight closed-loop control system is as follows: Figure 4 As shown, it includes:

[0165] Ground effect prediction module based on physical information neural network:

[0166] 1) Input the current flight status parameters;

[0167] 2) Real-time estimation of ground effect disturbance force.

[0168] Other unknown interference estimation modules based on environmental interference observers:

[0169] 1) Nonlinear disturbance observer structure;

[0170] 2) Estimate other unknown disturbances.

[0171] Composite anti-interference controller:

[0172] 1) Basic geometric controller;

[0173] 2) Ground effect feedforward compensation;

[0174] 3) Environmental interference feedforward compensation.

[0175] Signal flow direction as follows Figure 4 As indicated by the middle arrow, the whole system constitutes a closed-loop control system.

[0176] Specifically, the ground effect prediction module based on physical information neural network adopts a method based on... Figure 2 The physical information network structure shown has an 18-dimensional state vector as its real-time input and a 4-dimensional ground effect interference force as its output, with a sampling frequency of 200Hz.

[0177] Specifically, taking a certain type of quadcopter UAV as the application object, and adopting a northeast-northeast coordinate system, the aforementioned environmental interference observer is designed:

[0178] Position dynamics model of a quadcopter UAV under near-ground multi-source interference:

[0179]

[0180] The estimation error for environmental disturbances is set as follows:

[0181]

[0182] in, This represents the perturbation estimate. Taking the derivative of equation (10), we get:

[0183]

[0184] It should be noted that, due to the lack of prior knowledge about the actual perturbations, the following general assumptions are made here.

[0185] make:

[0186]

[0187] Combining equations (10) and (12), we can obtain:

[0188]

[0189] Since acceleration information is difficult to obtain in practice, both sides of equation (13) are subtracted simultaneously. get:

[0190]

[0191] make:

[0192]

[0193] Therefore:

[0194]

[0195] Let intermediate variables be defined. Differentiating with respect to ξ, we get:

[0196]

[0197] Equation (17) is for the observer of the intermediate variable ξ. After obtaining ξ, the unknown disturbance is estimated. It can be calculated relatively easily:

[0198]

[0199] Therefore, it is possible to observe the unknown disturbance Δ.

[0200] Based on the above, a composite anti-interference controller is designed, specifically:

[0201] Establish a full dynamic model of a quadcopter UAV under near-ground multi-source interference:

[0202]

[0203] Where R represents the rotation matrix from the body coordinate system to the Earth coordinate system:

[0204]

[0205] In equation (20), c g s g These are abbreviated expressions for the cosine and sinine functions, respectively, [φ,θ,ψ]. T This represents three Euler angles. Additionally, in equation (19), ω = [p, q, r]. T Let J represent the angular velocity of the aircraft, J represent the inertial matrix of the UAV, and τ represent the total torque across the three axes. × Represents an antisymmetric matrix, [ω] × It can be represented as:

[0206]

[0207] [Jω] × The representation method is similar.

[0208] For the position loop, the ground effect disturbance force predicted by the physical information neural network and the unknown disturbance estimated by the environmental observer are used as feedforward compensation. The desired thrust f is used as the control input, and the control law is designed as follows:

[0209]

[0210] f = (-F d ) T Re3 (24)

[0211] The error terms are as follows:

[0212] e x =pp d (25)

[0213]

[0214] And the saturation function is set as follows:

[0215]

[0216] Gain coefficient k x k v k int Both c1 and c2 are greater than 0. In equation (23) The predictions come from the physical information neural network and the estimates from the environmental disturbance observer, respectively.

[0217] For the attitude loop, a geometric controller based on a special Euclidean group is used to achieve the desired torque M. d To control the input, the control law is designed as follows:

[0218]

[0219] The error terms are as follows:

[0220]

[0221] e ω =ω-R T R d ω d (31)

[0222] In equation (30), the parentheses (·) ∨ This is called a vee mapping, and it is [·]. × The inverse transform. Gain coefficient k R k ω All are greater than 0.

[0223] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A near-ground flight control method for a quadrotor UAV based on a physical information neural network, characterized in that, Includes the following steps: S1. Perform dynamic modeling on the quadcopter drone; S2. Design a geometric controller as a reference controller; S3. Data acquisition is performed in a simulation environment using a reference controller; The collected data are the state parameters of the UAV's near-ground trajectory tracking flight, and the state parameters include: ① The relative ground height of the four propellers; ② The three-axis velocity of the UAV; ③ The three-axis angular velocity of the UAV; ④ The attitude quaternion of the UAV; ⑤ The input PWM signals of the four motors; S4. Design the physical information neural network structure and loss function, and train it; specifically including the following: S4.1 Construct a neural network structure, wherein the neural network structure includes: ① Input layer: has 18 neurons, used to receive the state parameters; ② Three hidden layers: containing 30, 35, and 20 neurons respectively, with ReLU activation function; ③ Output layer: has 4 neurons, used to output the estimated ground effect disturbance force on the four propellers. ; S4.2 Design the loss function, which includes: ① Implicit physical information loss term: in, N s Indicates the number of samples; i This indicates the sample number; j={1,2,3,4} represents the four propellers; Indicates the first i The first sample j The theoretical value of the ground effect interference force on a propeller Indicates the neural network for the first... i The first sample j Predicted output variables for ground effect disturbance force on a propeller; single sample i middle The calculation process is as follows, based on the propeller model of a multi-rotor UAV, the propeller... j The thrust generated in the region without ground effect is: in, C T It is a dimensionless constant. Indicates air density, N ij Indicates propeller j Rotational speed, measured in revolutions per minute. D Indicates the diameter of the propeller; propeller j The thrust generated near the ground is: in, Indicates propeller j The actual lift coefficient; Represents rotational speed in radians per second; and There is a proportional relationship: Combining the above equations, we obtain: Choose a physical empirical formula: in, C a and C b These are coefficients determined based on the geometry of the quadcopter blades; R The rotor radius; z ij This indicates the height of each propeller above the ground. Calculations yielded the propeller j Ground effect interference: ； ② Explicit physical information loss term: in, Indicates the first i The first sample j The thrust generated by a propeller in the area without ground effect; For predicting the proportion; This is an empirical proportion term; ③ Overall network loss function: in, λ ca Represents the explicit physical information loss term Loss A weight of 2; S4.3 Network training includes the following steps: S4.3.1 Design a cyclic annealing scheduling strategy and set... ,in, This indicates a modulo operation. n Indicates the index of the current training epoch of the neural network; T This indicates the maximum number of training iterations for the neural network. M Indicates the cycle of cyclic annealing; Set within each annealing cycle λ ca Keep the maximum value λ max The proportion is μ ; when β > μ hour, ; when β ≤ μ hour, ; This adjusts the explicit physical information loss term. Loss Weight of 2 λ ca ; S4.3.2, For the weight matrix of each layer in the neural network W Perform singular value decomposition: in, U It is a left singular vector matrix; For a singular value diagonal matrix, its diagonal elements σ 1. σ 2、…、 σ r satisfy σ 1≥ σ 2≥…≥ σ r ≥0, r For matrix rank; V T It is the transpose of the right singular vector matrix; weight matrix W Lipschitz constant γ Equal to the maximum singular value σ 1. Constrain the maximum singular value; set a singular value threshold. σ max ,like σ 1> σ max Then W new = W *( σ max / σ 1) This constraint ensures the Lipschitz continuity of the network; S5. Design a composite anti-interference controller integrating a physical information neural network and an environmental interference observer; the environmental interference observer is used to estimate other unknown environmental interferences besides ground effects. Based on the quadrotor dynamics model constructed in S1, the state equations for the environmental disturbance observer are designed as follows: in, Indicates intermediate observed variables; m Indicates the mass of the drone; g Represents gravitational acceleration; e 3 represents the vector [0,0,1] T ; This indicates the linear velocity of the drone; l = α / m , α A positive gain coefficient; p ( v )= α* v Used to estimate other unknown disturbances: ; Composite anti-interference control is achieved by integrating a physical information neural network and an environmental interference observer. The composite anti-interference controller is designed as follows: in, , as well as All are positive gain coefficients; , These are the error terms for position and velocity, respectively; The desired acceleration; It is a saturation function; This represents the ground effect interference force predicted by the physical information neural network; This represents other environmental disturbances estimated by the environmental disturbance observer.

2. The near-ground flight control method for a quadrotor UAV based on a physical information neural network according to claim 1, characterized in that, S1 specifically includes the following: A translational dynamics model of a quadrotor UAV under near-ground flight conditions is established based on the Newton-Euler method: in, m Indicates the mass of the drone; This represents the linear acceleration of the drone; g Represents gravitational acceleration; e 3 represents the vector [0,0,1] T ; F d This represents the expected thrust of the drone when it is far from the ground. d ge This indicates the ground effect interference force experienced by the drone; This indicates other unknown environmental interference.

3. The near-ground flight control method for a quadrotor UAV based on a physical information neural network according to claim 1, characterized in that, The reference controller described in S2 includes a position loop controller and an attitude loop controller, wherein: The function of the position loop controller is expressed as follows: in, e x , e v These are the error terms for position and velocity, respectively; ; k x , k v as well as k int All are positive gain coefficients; The desired acceleration; For saturation functions; with input For example, specifically: in, This is the saturation value; when input... When the vector is used, the saturation function acts on the vector respectively. On each element, the final result is F d For the desired thrust; The function representation of the attitude loop controller is as follows: in, , e These are the attitude tracking error term and the angular velocity tracking error term, respectively. Indicates the angular velocity of the aircraft; k R , k All are positive gain coefficients; J The inertial matrix of the UAV is represented; Indicates an antisymmetric matrix; R Represents the rotation matrix. M d This represents the desired triaxial torque.

Images