Unmanned aerial vehicle anti-jamming attitude control method based on adaptive weight neural network

Abstract

This invention belongs to the field of UAV flight control technology, specifically a UAV anti-interference attitude control method based on adaptive weighted neural networks. Addressing the problem that existing controller designs mostly focus on the steady-state accuracy of the system, lacking a systematic and quantifiable design framework, leading to low efficiency in the controller design process and a final performance heavily reliant on the designer's experience, this invention proposes the following solution, comprising the following steps: S1: Establishing an attitude control model for a quadrotor UAV; S2: Decoupling the attitude control of the quadrotor UAV into a roll subsystem, a pitch subsystem, and a yaw subsystem. This invention uses RBFNN to estimate lumped disturbances online, combines this with a preset performance function to constrain tracking errors, and designs a novel adaptive law. The system is proven to be uniformly eventually bounded using the Lyapunov method, simultaneously ensuring transient and steady-state performance in complex environments, overcoming the problems of insufficient robustness of PID control, sliding mode chattering, and the lack of transient guarantees in traditional adaptive control.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) flight control technology, and in particular to an anti-interference attitude control method for UAVs based on an adaptive weighted neural network. Background Technology

[0002] As a typical underactuated, nonlinear, and strongly coupled system, the performance of attitude control is fundamental to achieving stable flight and accurate trajectory tracking for quadrotor UAVs. However, in practical applications, UAVs are inevitably affected by uncertainties such as model parameter perturbations, unmodeled internal dynamics, and external wind disturbances. These factors severely limit the performance of traditional control methods.

[0003] Currently, PID controllers, widely used in quadcopter attitude control, are simple in structure but cumbersome in parameter tuning, and often struggle to balance robustness and dynamic performance when facing complex disturbances and model changes. Sliding mode control, while known for its strong robustness, suffers from chattering that is difficult to completely eliminate, and its controller structure is relatively complex. Adaptive control can adjust parameters online, but its algorithms are typically complex, placing a heavy burden on real-time calculations, and it lacks direct and effective constraints on the system's transient performance.

[0004] Existing controller designs mostly focus on the steady-state accuracy of the system. However, key transient performance indicators such as rise time, overshoot, and convergence speed can usually only be indirectly met through repeated simulations and trial and error. There is a lack of a systematic and quantifiable design framework, which leads to low efficiency in the controller design process and the final performance is heavily dependent on the designer's experience. To address these issues, we propose an anti-interference attitude control method for UAVs based on adaptive weighted neural networks. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies, where controller design often focuses on the steady-state accuracy of the system, while key transient performance indicators such as rise time, overshoot, and convergence speed can usually only be indirectly satisfied through repeated simulations and trial and error. This lack of a systematic and quantifiable design framework leads to inefficient controller design processes and a heavy reliance on the designer's experience for final performance. Therefore, this invention proposes an anti-interference attitude control method for unmanned aerial vehicles based on an adaptive weighted neural network.

[0006] The UAV anti-interference attitude control method based on adaptive weighted neural network provided in this application adopts the following technical solution: The anti-interference attitude control method for UAVs based on adaptive weighted neural networks includes the following steps: S1: Establish an attitude control model for a quadcopter UAV; S2: Decouple the attitude control of the quadcopter UAV into a roll subsystem, a pitch subsystem, and a yaw subsystem; S3: RBFNN estimator design; S4: Error transformation based on a preset performance function (PPF); S5: Prove the stability of the system.

[0007] Furthermore, in S1, an attitude control model for the quadcopter UAV is established: (1); in, , , These represent roll angle, pitch angle, and yaw angle, respectively. , , These represent the angular velocities of roll, pitch, and yaw, respectively. , , These represent the moments for roll, pitch, and yaw, respectively. , , It is a quadcopter drone axis, axis, Moment of inertia of the shaft; The moment of inertia of a single motor rotor and propeller assembly about its axis of rotation; , , This is the drag coefficient; This refers to the total residual yaw moment caused by the speed imbalance of the quadcopter.

[0008] Furthermore, in S2, the roll subsystem is the main research object, and the dynamic model of the roll angle subsystem is constructed as follows: (2); in, Defined as roll angle, The roll rate is angular velocity. The parameters are known, and the input is the control input. This represents the lumped disturbance in the roll subsystem.

[0009] Furthermore, in S3, the function As the lumped disturbance experienced by the system, RBFNN is used to approximate the lumped disturbance. Estimated lumped disturbance for: (3); in, Ideal weights; The Gaussian function vector; The number of Gaussian function vectors; For the first Each element is defined as: (4); in For the set of real numbers, Represents all by A set of ordered tuples consisting of real numbers , Centered on, For the first The width of a neuron; definition Filtering parameters for: (5); in, yes Through a low-pass filter The change was obtained; It is the filter bandwidth.

[0010] Furthermore, in S3, the integrated disturbance after filtering... for: (6); Filtering operation is performed on both sides of the state dynamic equation (2): (7); in, It is a control input The filtered signal.

[0011] Furthermore, in S3, by combining and substituting equations (6) and (11), we obtain a complete dynamic description of the system's filtering state: (8); Define the intermediate matrix sum vector : (9); Among them, parameters ; Substituting (8) into (9) yields: (10); The solution to differential equation (10) is: (11); Finally, auxiliary variables From (11) and Calculated as: (12); in, For weight The estimated value; vector for The estimated value is obtained by applying the measured system dynamics. The low-pass filter is actually obtained; Rewriting instructions: Lemma 1: The low-pass filter The system dynamics applied to measurement, the vector defined by equation (12) It can be broken down into Among them, unknown parameters Its estimated value Error between Disturbance term It is generated by the unmodeled dynamics of the system and external disturbances, and satisfies the uniform boundedness, that is, there exists a normal constant. , making ; variable Includes The information can be used to update the weights of the RBFNN by designing adaptive laws. Therefore, it can be As a new loss term, it is incorporated into the adaptive law; therefore, the adaptive law is designed as follows: (13); in A diagonal matrix for learning gain; Let be the constant of the loss term; in (13), it can be appropriately set as . To ensure This avoids the possible singular values ​​in formula (26); Then, based on the convergence of the new adaptive rate, we can conclude that: Theorem 1: Consider system (2) with adaptive rate (13) and controller (25), if the augmented regression vector If the excitation conditions are met, the system is semi-globally stable, and the error... and It converges to a compact set near zero.

[0012] Furthermore, in S4, to study the performance of the roll angle tracking error under both transient and steady-state conditions, it is necessary to define the angle tracking error: (14); in, This represents the desired roll angle.

[0013] Furthermore, in S4, a decreasing function is selected in order to design the PPF. : (15); in, and The transient performance is guaranteed as follows: (16); in and These represent the lower and upper bounds of the constraint performance function, respectively. Indicates the convergence speed of the performance function; Indicates the final steady-state limit; Therefore, the error never exceeds the preset range throughout the entire control process, thereby improving the transient response performance and steady-state accuracy of the system; yes The strictly decreasing function of the transformation error. It has the following two properties: (1) ; (2) ; Therefore, (16) is rewritten as: (17); when Meet the conditions and , Therefore, equation (16) proves that the condition holds. Based on this, by introducing the error transformation function... The constrained original tracking error system (2) can be transformed into an equivalent unconstrained system (17). This transformation process is stable, and the error dynamics of the transformed system have a clear stability guarantee. Therefore, as long as the transformed error is bounded, the original tracking error can be guaranteed. The preset performance constraints (16) will always be met, thereby realizing the active control and guarantee of the transient and steady-state performance of the system; Using a uniform error function : (18); According to (16), the error is obtained. as follows: (19); Among them ; Therefore, it will be subject to constraints. Transform into a new unconstrained error variable Applied to controller design methods, i.e., when When bounded, the original error It will automatically meet the preset performance constraints.

[0014] Furthermore, in step S4, in order to achieve the desired tracking error performance... We derive the error after transformation. Combining (13), we obtain the stability error: (20); in, And satisfy ; Therefore, it is deduced that : (twenty one); When constraints Down, Under this derivative: (twenty two); To simplify the proof, auxiliary variables are designed. : (twenty three); Among them, when When bounded, guarantee Bounded; right Taking the derivative in the time domain, we get: (twenty four); Therefore, the controller is designed as follows: (25); in, For feedback gain; Feedback items Includes the redefined error variable Furthermore, the tracking error is controlled within a certain range by our proposed algorithm, thus keeping the system's tracking error accuracy within the specified range.

[0015] Furthermore, in S5, it is proven that we choose the Lyapunov function as follows: (26); right Differentiating, we get: (27); Substitute (26) and (27) into: (28); The transformation yields: (29); Using Yang's inequality: (30); (31); Therefore, when hour, It is a positive parameter, therefore we can conclude that: (32); A positive parameter, therefore: (33); Based on Lyapunov's theorem, the conclusion is as follows: and It is uniformly and eventually bounded; tracking error and control input It is also completely bounded; Note 1: Unlike the classical adaptive law, the adaptive law (25) It is a new excitation-adjustable gain used to maintain convergence. Includes estimation error information Therefore, use the disclosure item This will result in a quadratic term in the Lyapunov function (31). Therefore, control error and estimation error Simultaneous convergence, which is consistent with widely used... Correction or The adaptive law of the modified (1) is significantly different; Note 2: As shown in the above proof, the error convergence rate and the final error bound depend on the feedback gain. Incentive levels Learning parameters and Generally speaking, a larger control gain Learning gain Adjusting the filter parameters can achieve faster convergence and smaller errors, while excessively large learning gain may induce oscillations. This can be addressed by using an appropriate gain in the adaptive law (25). This problem can be partially avoided.

[0016] In summary, this application includes at least one of the following beneficial technical effects: 1. This scheme unifies model uncertainty and external disturbances into lumped disturbances and uses RBFNN for online estimation and approximation, which significantly improves the system's ability to suppress composite disturbances. Combined with auxiliary variable design based on filtering technology, it can achieve real-time compensation for unknown dynamics without directly measuring disturbances or relying on accurate models. This enables the controller to maintain high-precision attitude tracking even in complex environments such as strong wind disturbances and parameter perturbations, overcoming the shortcomings of insufficient robustness of traditional PID controllers. 2. This scheme introduces a preset performance function (PPF) and applies explicit time-domain constraints to the tracking error, which can strictly guarantee the transient performance (overshoot and convergence speed) and steady-state accuracy of the system simultaneously. By using error transformation technology, the constrained original system is transformed into an equivalent unconstrained system, so that the controller design no longer depends on repeated trial and error and the error response envelope can be predetermined. Compared with the chattering problem of sliding mode control and the lack of transient performance guarantee in traditional adaptive control, this invention significantly improves the designability and predictability of control quality.

[0017] This invention estimates lumped disturbances online using RBFNN, combines a preset performance function to constrain tracking errors, and designs a novel adaptive law. The system is proven to be uniformly eventually bounded by the Lyapunov method, ensuring both transient and steady-state performance in complex environments. It overcomes problems such as insufficient robustness of PID, sliding mode chattering, and lack of transient protection in traditional adaptive systems. Furthermore, the parameters are easy to adjust, balancing performance and engineering practicality. Attached Figure Description

[0018] Figure 1 This is a flowchart of the anti-interference attitude control method for unmanned aerial vehicles based on adaptive weighted neural networks proposed in this invention; Figure 2 This is a schematic diagram illustrating the principle of the anti-interference attitude control method for unmanned aerial vehicles based on adaptive weighted neural networks proposed in this invention. Figure 3 This is a simulation image of angle tracking under sinusoidal signal expectation for the UAV anti-interference attitude control method based on adaptive weighted neural network proposed in this invention. Figure 4 This is a simulation image of the tracking error under the sinusoidal signal expectation of the UAV anti-interference attitude control method based on adaptive weighted neural network proposed in this invention; Figure 5 This is a simulation image of the perturbation estimation under sinusoidal signal expectation for the UAV anti-interference attitude control method based on adaptive weighted neural network proposed in this invention. Detailed Implementation

[0019] 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.

[0020] Example 1 Reference Figures 1-5 An anti-interference attitude control method for unmanned aerial vehicles based on adaptive weighted neural networks includes the following steps: S1: Establish an attitude control model for the quadcopter UAV: (1); in, , , These represent roll angle, pitch angle, and yaw angle, respectively. , , These represent the angular velocities of roll, pitch, and yaw, respectively. , , These represent the moments for roll, pitch, and yaw, respectively. , , It is a quadcopter drone axis, axis, Moment of inertia of the shaft; The moment of inertia of a single motor rotor and propeller assembly about its axis of rotation; , , This is the drag coefficient; This refers to the total residual yaw moment caused by the speed imbalance of the quadcopter. S2: The attitude control of the quadcopter UAV is decoupled into a roll subsystem, a pitch subsystem, and a yaw subsystem. The roll subsystem is the main research object, and the dynamic model of the roll angle subsystem is constructed as follows: (2); in, Defined as roll angle, The roll rate is angular velocity. The parameters are known, and the input is the control input. This represents the lumped disturbance in the roll subsystem.

[0021] Furthermore, in S3, the function As the lumped disturbance experienced by the system, RBFNN is used to approximate the lumped disturbance. Estimated lumped disturbance for: (3); in, Ideal weights; The Gaussian function vector; The number of Gaussian function vectors; For the first Each element is defined as: (4); in For the set of real numbers, Represents all by A set of ordered tuples consisting of real numbers , Centered on, For the first The width of a neuron; definition Filtering parameters for: (5); in, yes Through a low-pass filter The change was obtained; It is the filter bandwidth; S3: RBFNN estimator design, with filtered integrated perturbation for: (6); Filtering operation is performed on both sides of the state dynamic equation (2): (7); in, It is a control input Filtered signal; By combining and substituting equations (6) and (11), we obtain a complete dynamic description of the system's filtering state: (8); Define the intermediate matrix sum vector : (9); Among them, parameters ; Substituting (8) into (9) yields: (10); The solution to differential equation (10) is: (11); Finally, auxiliary variables From (11) and Calculated as: (12); in, For weight The estimated value; vector for The estimated value is obtained by applying the measured system dynamics. The low-pass filter is actually obtained; Rewriting instructions: Lemma 1: The low-pass filter The system dynamics applied to measurement, the vector defined by equation (12) It can be broken down into Among them, unknown parameters Its estimated value Error between Disturbance term It is generated by the unmodeled dynamics of the system and external disturbances, and satisfies the uniform boundedness, that is, there exists a normal constant. , making ; variable Includes The information can be used to update the weights of the RBFNN by designing adaptive laws. Therefore, it can be As a new loss term, it is incorporated into the adaptive law; therefore, the adaptive law is designed as follows: (13); in A diagonal matrix for learning gain; Let be the constant of the loss term; in (13), it can be appropriately set as . To ensure This avoids the possible singular values ​​in formula (26); Then, based on the convergence of the new adaptive rate, we can conclude that: Theorem 1: Consider system (2) with adaptive rate (13) and controller (25), if the augmented regression vector If the excitation conditions are met, the system is semi-globally stable, and the error... and Converging to a compact set near zero; S4: Based on the error transformation of the preset performance function (PPF), the performance of roll angle tracking error under transient and steady-state conditions needs to be studied. The angle tracking error needs to be defined: (14); in, This represents the desired roll angle. A decreasing function was chosen for designing the PPF. : (15); in, and The transient performance is guaranteed as follows: (16); in and These represent the lower and upper bounds of the constraint performance function, respectively. Indicates the convergence speed of the performance function; Indicates the final steady-state limit; Therefore, the error never exceeds the preset range throughout the entire control process, thereby improving the transient response performance and steady-state accuracy of the system; yes The strictly decreasing function of the transformation error. It has the following two properties: (1) ; (2) ; Therefore, (16) is rewritten as: (17); when Meet the conditions and , Therefore, equation (16) proves that the condition holds. Based on this, by introducing the error transformation function... The constrained original tracking error system (2) can be transformed into an equivalent unconstrained system (17). This transformation process is stable, and the error dynamics of the transformed system have a clear stability guarantee. Therefore, as long as the transformed error is bounded, the original tracking error can be guaranteed. The preset performance constraints (16) will always be met, thereby realizing the active control and guarantee of the transient and steady-state performance of the system; Using a uniform error function : (18); According to (16), the error is obtained. as follows: (19); Among them ; Therefore, it will be subject to constraints. Transform into a new unconstrained error variable Applied to controller design methods, i.e., when When bounded, the original error It will automatically meet the preset performance constraints; To achieve the desired tracking error performance We derive the error after transformation. Combining (13), we obtain the stability error: (20); in, And satisfy ; Therefore, it is deduced that : (twenty one); When constraints Down, Under this derivative: (twenty two); To simplify the proof, auxiliary variables are designed. : (twenty three); Among them, when When bounded, guarantee Bounded; right Taking the derivative in the time domain, we get: (twenty four); Therefore, the controller is designed as follows: (25); in, For feedback gain; Feedback items Includes the redefined error variable Furthermore, the tracking error accuracy of the system is controlled within a certain range by our proposed algorithm, thus keeping the tracking error accuracy within the specified range. S5: Prove the system stability. Proof: We choose the Lyapunov function as follows: (26); right Differentiating, we get: (27); Substitute (26) and (27) into: (28); The transformation yields: (29); Using Yang's inequality: (30); (31); Therefore, when hour, It is a positive parameter, therefore we can conclude that: (32); A positive parameter, therefore: (33); Based on Lyapunov's theorem, the conclusion is as follows: and It is uniformly and eventually bounded; tracking error and control input It is also completely bounded; Note 1: Unlike the classical adaptive law, the adaptive law (25) It is a new excitation-adjustable gain used to maintain convergence. Includes estimation error information Therefore, use the disclosure item This will result in a quadratic term in the Lyapunov function (31). Therefore, control error and estimation error Simultaneous convergence, which is consistent with widely used... Correction or The adaptive law of the modified (1) is significantly different; Note 2: As shown in the above proof, the error convergence rate and the final error bound depend on the feedback gain. Incentive levels Learning parameters and Generally speaking, a larger control gain Learning gain Adjusting the filter parameters can achieve faster convergence and smaller errors, while excessively large learning gain may induce oscillations. This can be addressed by using an appropriate gain in the adaptive law (25). This problem can be partially avoided.

[0022] Example 2 The difference between this embodiment and Embodiment 1 is that in S2, before the controller is implemented, the nominal value of model parameter b1 and its possible range of variation are obtained through ground testing or offline identification. Simultaneously, the expected upper limit of the external disturbance amplitude is determined based on flight experience. This information is used as input for selecting controller parameters, guiding the initial setting of key parameters such as feedback gain and learning gain. This step provides a quantitative basis for subsequent controller parameter configuration, enhancing the relevance and rationality of parameter selection.

[0023] Example 3 The difference between this embodiment and Embodiment 1 is as follows: In S3, in actual implementation, the input Z of the RBFNN is taken as the system state quantity. The center distribution of neurons is obtained by clustering offline collected normal flight data using the K-means clustering method. The width is uniformly set to 1.2 to 1.5 times the average distance between the cluster centers to ensure that the Gaussian function forms an effective coverage in the state space. The initial weights are set to zero vectors, indicating that there is no prior knowledge of the disturbance at the initial time and it relies entirely on online learning; the learning gain is selected according to the state magnitude. Through the above standardized initialization process, the dependence of parameter tuning on manual experience is effectively reduced.

[0024] Example 4 The difference between this embodiment and Embodiment 1 is that in S4, the boundary parameters, convergence speed, and steady-state limits in the preset performance function can be flexibly configured according to specific task requirements. Taking roll angle tracking as an example, the boundary parameter is usually selected as 1, the convergence speed is set according to the actuator bandwidth, and the steady-state limit is determined according to the expected tracking accuracy. This parameter configuration step makes the requirements for transient performance and steady-state accuracy clear and quantifiable in the design stage, and the controller design goal is clear and predictable, avoiding the passive situation of indirectly meeting performance indicators through repeated simulation and trial and error in the traditional method.

[0025] Example 5 The difference between this embodiment and Embodiment 1 is that in S5, based on the stability analysis results, the feedback gain, learning gain, and filter bandwidth jointly determine the system's convergence speed and steady-state accuracy, while also affecting the control input amplitude and noise sensitivity. During actual debugging, the following tuning sequence should be followed: First, select the filter parameters based on the actuator bandwidth limit; second, increase the feedback gain as much as possible within the allowable control range to improve the response speed; finally, gradually increase the learning gain until the disturbance suppression effect meets the requirements and does not induce oscillations. This tuning process provides clear operational guidance for engineering implementation.

[0026] Comparative Example Comparative Example 1 The difference between this comparative example and Example 1 is that it does not introduce an RBFNN estimator for online estimation and compensation of lumped disturbances. Instead, it directly uses a traditional PID controller for roll attitude control. The PID parameters are initially obtained through the Ziegler-Nichols tuning method and then fine-tuned through repeated trials to achieve the best control effect. This scheme relies solely on error feedback for control and lacks the ability to actively estimate and compensate for model uncertainties and external disturbances, thus having inherent limitations in high-precision tracking and strong disturbance suppression. Comparative Example 2 The difference between this comparative example and Example 1 is that: RBFNN is used to estimate the lumped disturbance online, but a preset performance function (PPF) is not introduced to impose time-domain constraints on the tracking error. Instead, the estimated disturbance is directly superimposed on the PID control law to form a composite control through feedforward compensation. Although this scheme has a certain disturbance suppression capability, it lacks pre-design and active guarantee means for the transient performance of the system (such as overshoot and convergence speed). The dynamic performance still depends on the repeated tuning of the controller parameters and cannot be quantitatively guaranteed in the design stage. Comparative Example 3 The difference between this comparative example and Example 1 is that the adaptive law design does not introduce an excitation adjustment term containing estimation error information. Instead, it uses the classic σ-correction adaptive law to update the neural network weights. The rest of the controller structure is consistent with Example 1. Although this scheme can also achieve online updating of neural network weights, it lacks a coupling convergence mechanism between estimation error and tracking error, resulting in a slow weight convergence speed. Under continuous disturbance, the estimation error may not be effectively attenuated, affecting further improvement of control accuracy.

[0027] Experimental Example Experimental Objective To fully verify the effectiveness and superiority of the method proposed in this invention, the online estimation accuracy and compensation effect of the RBFNN estimator for lumped disturbances were tested by simultaneously applying composite disturbances such as periodic wind disturbance, random noise, and parameter perturbation to the system. The ability of the preset performance function to actively guarantee the transient and steady-state indices of the tracking error was also examined. Furthermore, traditional PID control, RBFNN composite control without preset performance constraints, and neural network control using σ-correction adaptive laws were compared to comprehensively evaluate the performance advantages of this invention in terms of tracking accuracy, convergence speed, and control energy consumption, providing experimental evidence for the engineering application of the method.

[0028] II. Experimental Conditions To verify the effectiveness and superiority of the method of this invention, a quadcopter UAV attitude control simulation environment was built on the Matlab / Simulink simulation platform. The simulation mainly focused on verifying the roll channel. The quadcopter model parameters were configured according to a typical small UAV: ​​the moment of inertia was set on each of the three axes, and the rotor inertia and drag coefficient were selected based on actual physical characteristics. The controller parameters were uniformly configured according to the method described in Example 1: the performance function boundary parameters were taken as symmetrical values, the convergence speed was set to a reasonable value to ensure the response rate, the steady-state limit was determined according to the desired accuracy, and the feedback gain was selected to ensure fast response while avoiding the excitation of oscillations.

[0029] III. Simulation Working Conditions In the simulation, the desired tracking signal was set as a sinusoidal signal with moderate amplitude and reasonable frequency to simulate the typical desired trajectory of the UAV during attitude adjustment. To verify the robustness of the method under complex disturbance environments, a composite disturbance was added to the system, specifically including: a sinusoidal wind disturbance with a large amplitude to simulate periodic external aerodynamic interference; superimposed random noise components to simulate sensor noise and turbulence effects; and a certain proportion of perturbation was applied to the model parameters to simulate the uncertainty of rotational inertia and drag coefficient in actual flight. The total simulation time was set to 20 seconds, and the methods described in Example 1 and Comparative Examples 1 to 3 were used for comparative verification.

[0030] IV. Evaluation Indicators To comprehensively evaluate the control performance of each method, the following four types of quantitative indicators were selected: Maximum absolute error: The maximum value among the absolute values ​​of the tracking error throughout the entire process, reflecting the maximum deviation that the system may exhibit during the transient process, and is a direct indicator for measuring overshoot; Root mean square error: The square root of the mean square of the tracking error throughout the entire operation, reflecting the overall tracking accuracy of the system over the entire operating time, and is a comprehensive indicator for measuring steady-state performance; Settling time: The time required for the error to move from its initial value and remain within the steady-state limits, reflecting the convergence speed of the system; Maximum control torque: The maximum absolute value of the control input throughout the entire process, reflecting the level of control energy consumption.

[0031] V. Data Comparison

[0032] VI. Test Report Regarding tracking accuracy: Example 1 shows that both the maximum absolute error and root mean square error are significantly better than the three comparative examples. Compared with pure PID control, the maximum absolute error is reduced by more than 80%, and the root mean square error is reduced by more than 85%. Compared with Comparative Example 2, which also uses RBFNN but has no preset performance constraints, both error indicators are reduced by more than 50%. Compared with Comparative Example 3, which uses σ-correction adaptive law, the maximum absolute error is reduced by about 47%, and the root mean square error is reduced by about 56%. The results show that the combined effect of preset performance constraints and novel excitation adjustment terms significantly improves tracking accuracy. Regarding convergence speed: Example 1 has the shortest settling time. It is reduced by about 75% compared to pure PID control; by about 52% compared to Comparative Example 2 without preset performance constraints; and by about 43% compared to the σ-correction adaptive scheme. The results show that the preset performance function has a direct effect on improving the transient response speed and can actively plan the convergence rate during the design phase. In terms of control efficiency: the maximum control torque of Example 1 is the lowest among all schemes, indicating that the present invention achieves a good balance between performance and energy consumption without increasing the control quantity while ensuring high-precision tracking and fast convergence. Regarding perturbation suppression: In Example 1, the RBFNN's estimate of the lumped perturbation can quickly track the actual perturbation changes, and the estimation error stably converges to a small neighborhood near zero. However, the perturbation estimation performance of each comparative example is inferior to that of this invention, verifying the effectiveness of the designed adaptive law.

[0033] 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 method for anti-interference attitude control of unmanned aerial vehicles based on adaptive weighted neural networks, characterized in that: Includes the following steps: S1: Establish an attitude control model for a quadcopter UAV; S2: Decouple the attitude control of the quadcopter UAV into a roll subsystem, a pitch subsystem, and a yaw subsystem; S3: RBFNN estimator design; S4: Error transformation based on a preset performance function; S5: Prove the stability of the system.

2. The UAV anti-interference attitude control method based on adaptive weighted neural network according to claim 1, characterized in that: In step S1, an attitude control model for a quadcopter UAV is established: （1）； in, , , These represent roll angle, pitch angle, and yaw angle, respectively. , , These represent the angular velocities of roll, pitch, and yaw, respectively. , , These represent the moments for roll, pitch, and yaw, respectively. , , It is a quadcopter drone axis, axis, Moment of inertia of the shaft; The moment of inertia of a single motor rotor and propeller assembly about its axis of rotation; , , This is the drag coefficient; This refers to the total residual yaw moment caused by the speed imbalance of the quadcopter.

3. The UAV anti-interference attitude control method based on adaptive weighted neural network according to claim 2, characterized in that: In S2, the roll angle subsystem is the main research object, and the system dynamics model of the roll angle subsystem is constructed as follows: （2）； in, Defined as roll angle, The roll rate is angular velocity. The parameters are known, and the input is the control input. This represents the lumped disturbance in the roll subsystem.

4. The UAV anti-interference attitude control method based on adaptive weighted neural network according to claim 3, characterized in that: In S3, the function As the lumped disturbance experienced by the system, RBFNN is used to approximate the lumped disturbance. Estimated lumped disturbance for: （3）； in, Ideal weights; The Gaussian function vector; The number of Gaussian function vectors; For the first Each element is defined as: （4）； in For the set of real numbers, Represents all by A set of ordered tuples consisting of real numbers , Centered on, For the first The width of a neuron; definition Filtering parameters for: （5）； in, yes Through a low-pass filter The change was obtained; It is the filter bandwidth.

5. The UAV anti-interference attitude control method based on adaptive weighted neural network according to claim 4, characterized in that: In S3, the integrated disturbance after filtering... for: （6）； Filtering operation is performed on both sides of the state dynamic equation (2): （7）； in, It is a control input The filtered signal.

6. The UAV anti-interference attitude control method based on adaptive weighted neural network according to claim 5, characterized in that: In S3, by combining and substituting equations (6) and (11), a complete dynamic description of the system's filtering state is obtained: （8）； Define the intermediate matrix sum vector : （9）； Among them, parameters ; Substituting (8) into (9) yields: （10）； The solution to differential equation (10) is: （11）； Finally, auxiliary variables From (11) and Calculated as: （12）； in, For weight The estimated value; vector for The estimated value is obtained by applying a measurement to the system dynamics. The low-pass filter was actually obtained.

7. The UAV anti-interference attitude control method based on adaptive weighted neural network according to claim 6, characterized in that: In S4, to study the performance of the roll angle tracking error under both transient and steady-state conditions, it is necessary to define the angle tracking error: （14）； in, This represents the desired roll angle.

8. The UAV anti-interference attitude control method based on adaptive weighted neural network according to claim 7, characterized in that: In S4, a decreasing function is chosen in order to design the PPF. : （15）； in, and The transient performance is guaranteed as follows: （16）； in and These represent the lower and upper bounds of the constraint performance function, respectively. Indicates the convergence speed of the performance function; Indicates the final steady-state limit; Therefore, the error never exceeds the preset range throughout the entire control process, thereby improving the transient response performance and steady-state accuracy of the system; yes The strictly decreasing function of the transformation error. It has the following two properties: （1） ； （2） ； Therefore, (16) is rewritten as: （17）； when Meet the conditions and , Therefore, equation (16) proves that the condition holds. Based on this, by introducing the error transformation function... The constrained original tracking error system (2) can be transformed into an equivalent unconstrained system (17). This transformation process is stable, and the error dynamics of the transformed system have a clear stability guarantee. Therefore, as long as the transformed error is bounded, the original tracking error can be guaranteed. The preset performance constraints (16) will always be met, thereby realizing the active control and guarantee of the transient and steady-state performance of the system; Using a uniform error function : （18）； According to (16), the error is obtained. as follows: （19）； Among them ; Therefore, it will be subject to constraints. Transform into a new unconstrained error variable Applied to controller design methods, i.e., when When bounded, the original error It will automatically meet the preset performance constraints.

9. The UAV anti-interference attitude control method based on adaptive weighted neural network according to claim 8, characterized in that: In step S4, in order to achieve the desired tracking error performance Derive the error after transformation Combining (13), we obtain the stability error: （20）； in, And satisfy ; Therefore, it is deduced that : （21）； When constraints Down, Under this derivative: （22）； To simplify the proof, auxiliary variables are designed. : （23）； Among them, when When bounded, guarantee Bounded; right Taking the derivative in the time domain, we get: （24）； Therefore, the controller is designed as follows: （25）； in, For feedback gain; Feedback items Includes the redefined error variable Furthermore, the proposed algorithm controls the tracking error within a certain range, thus keeping the system's tracking error accuracy within the specified range.

10. The UAV anti-interference attitude control method based on adaptive weighted neural network according to claim 9, characterized in that: In S5, it is proven that the Lyapunov function is selected as follows: （26）； right Differentiating, we get: （27）； Substitute (26) and (27) into: （28）； The transformation yields: （29）； Using Yang's inequality: （30）； （31）； Therefore, when hour, It is a positive parameter, therefore we can conclude that: （32）； A positive parameter, therefore: （33）； Based on Lyapunov's theorem, the conclusion is as follows: and It is uniformly and eventually bounded; tracking error and control input It is also completely bounded.

Images