Neural network enhancement-based model prediction contour control system and method

Abstract

The invention belongs to the technical field of unmanned aerial vehicle control, and discloses a model prediction contour control system and method based on neural network enhancement, the system processes multi-source sensing data of an unmanned aerial vehicle through an S4-Conv adaptive network architecture, and the system comprises an odometer, a thrust output and a PVDF flexible crash sensor. And accurately predicting the position and attitude of the unmanned aerial vehicle at the next moment. Based on the prediction model, a neural network enhanced model prediction contour control framework is provided for optimizing the thrust output of the unmanned aerial vehicle and performing real-time attitude adjustment. According to the method, external collision and dynamic disturbance in a complex environment can be quickly responded, the trajectory tracking precision of the unmanned aerial vehicle and the stability of task execution are remarkably improved, and the method is suitable for unmanned aerial vehicle flight control tasks in complex environments such as urban space and disaster area search and rescue.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned aerial vehicle control, and in particular relates to a model prediction profile control system and method based on neural network enhancement. Background Art

[0002] When performing complex tasks, drones are often faced with external disturbances and potential collision risks, especially in dynamic or narrow environments such as urban spaces and disaster area search and rescue, where external collisions are almost inevitable. Existing control methods are mainly based on fixed model design, ignoring the real-time response capability of external sudden collisions and the adaptability to dynamic environments. Although the traditional model prediction contour control framework can perform trajectory optimization, it is often difficult to make quick adjustments when faced with sudden collisions and complex disturbances, causing the drone to deviate from the predetermined trajectory or even lose control. In addition, the existing control algorithms lack sufficient adaptive capabilities and it is difficult to adjust control parameters based on real-time sensor data, resulting in poor performance of drones in dynamically changing environments. Therefore, a more adaptive drone control method is needed that can handle trajectory tracking problems in complex environments and quickly respond to external collisions and dynamic disturbances. Summary of the invention

[0003] The object of the present invention is to provide a model prediction contour control system and method based on neural network enhancement to solve the above-mentioned technical problems.

[0004] In order to solve the above technical problems, the specific technical solutions of a model prediction contour control system and method based on neural network enhancement of the present invention are as follows:

[0005] A neural network-enhanced model-predictive contour control system, comprising: an S4Conv adaptive network architecture for receiving and processing multi-source sensor data to predict the future position and attitude of a UAV;

[0006] PVDF flexible collision sensor, used to sense the information of the collision between the drone and the external object; the PVDF flexible collision sensor senses the collision information through the voltage signal and inputs it into the S4-Conv adaptive network architecture for processing;

[0007] A neural network-enhanced model-predictive contour control framework is used to optimize the thrust output of the UAV and adjust the attitude in real time based on the prediction results.

[0008] The present invention also discloses a control method of a model prediction profile control system based on neural network enhancement, comprising the following steps:

[0009] Step 1: Synchronization and fusion of multi-source sensor data;

[0010] Step 2: Data preprocessing and enhancement;

[0011] Step 3: Trajectory prediction of Sana prediction model;

[0012] Step 4: Trajectory optimization for model-predictive contour control.

[0013] Furthermore, step 1 includes data source and data synchronization. The data source is the data received by the Sana prediction model from multiple source sensors, including information from odometers, thrust sensors, and PVDF flexible collision sensors. These data fusions are used to perceive and predict the flight status of the drone. The data synchronization adopts a timestamp-based alignment and synchronization method:

[0014] Assume that different sensors (S1, S2, ..., S n ) are sampled at (t1, t2, …, t n ), through timestamp alignment, the data synchronization formula is:

[0015]

[0016] In this way, all data are aligned to a uniform time step t. aligned , ensuring the consistency of multi-dimensional sensing data.

[0017] Furthermore, the step 2 comprises:

[0018] Preprocessing: For sensor data x t , using the exponentially weighted moving average (EWMA) filtering formula:

[0019]

[0020] Where α is the smoothing coefficient, which is used to control the weight of historical data;

[0021] The sensor data is standardized so that data of different dimensions have consistent mean and variance:

[0022]

[0023] Among them, μ is the mean of the data, σ is the standard deviation, and after standardization, all sensor data have the same mean of 0 and variance of 1;

[0024] Data augmentation: includes simulation of voltage noise, thrust perturbations, random rotations and displacements.

[0025] Furthermore, the step 3 comprises:

[0026] S4-Conv adaptive network architecture: The Sana prediction model is based on the S4D state space sequence modeling method. It generates high-precision future trajectory predictions by fusing multi-source sensor data and provides the future position, speed and attitude of the drone.

[0027] Assume X t Representing the multi-source sensor data at time t, predicting the trajectory of the next k time steps, the output trajectory of the Sana prediction model is:

[0028]

[0029] Among them, f θ is a neural network function based on the S4D state space model, and θ represents the parameters of the network. Further, step 4 includes:

[0030] Trajectory generation and optimization: Based on the results of the drone path search and path optimization, the Mina control framework first calculates the ideal trajectory of the drone to ensure that the drone can follow the target trajectory in the optimal path. The current path is p(t) = {p1, p2, …, p m}, the target path is p target (t)={p1,p2,…,p m}, the position error is defined as:

[0031] e(t)=p target (t)-p(t)

[0032] In the calculation process, PID control is used:

[0033]

[0034] Among them, K p , K i , K d They are the proportional, integral and differential gains of PID control respectively.

[0035] Further, the step 5 comprises:

[0036] Collision Detection: When the PVDF flexible collision sensor detects that the drone collides with an external object, the Sana prediction model immediately updates the post-collision state, and the Mina control framework adjusts based on this real-time feedback information;

[0037] Adaptive adjustment: The Mina control framework uses the MPCC mechanism to adjust the thrust output and attitude control after a collision to restore the drone to normal flight status.

[0038] Further, the step 6 comprises:

[0039] Dynamic environmental adaptation: The MPCC framework monitors external disturbances in real time through the Sana prediction model and adjusts the trajectory based on this data. Through online optimization, the control system responds quickly to external disturbances.

[0040] Further, the step 7 comprises:

[0041] In the model predictive contour control framework MPCC, the optimization objectives of the control input include the following four items: minimization of position error, minimization of velocity error, smoothness of thrust change, and minimization of thrust power consumption. Assumptions: p(t) represents the current position of the UAV, v(t) represents the current velocity of the UAV, T(t) represents the current thrust of the UAV, T prev (t-1) is the thrust at the previous moment, P(t) is the current thrust power consumption, and multiple optimization objective functions are defined and combined into an overall optimization problem:

[0042] (1) Position error minimization: The goal is to minimize the difference between the current position of the drone and the ideal target position p target The error between (t):

[0043] J1=‖p target (t)-p(t)‖ 2

[0044] (2) Speed ​​error minimization: The goal is to minimize the difference between the current speed of the drone and the ideal speed v target The error between (t):

[0045] J2=‖v target (t)-v(t)‖ 2

[0046] (3) Smoothness of thrust change: The goal is to minimize the drasticness of thrust change and maintain smoothness of thrust adjustment:

[0047] J3=‖T(t)-T prev (t-1)‖ 2

[0048] (4) Minimization of thrust power consumption: The goal is to reduce energy consumption and optimize thrust power:

[0049]

[0050] Where P(t) is the power consumption of thrust T(t), which is usually related to the thrust and the load of the UAV, and the calculation formula is:

[0051] P(t)=T(t)·v(t)

[0052] (5) Combine these objective functions to construct a weighted comprehensive optimization objective function:

[0053] J=λ1J1+λ2J2+λ3J3+λ4J4

[0054] Among them, λ1,λ2,λ3,λ4 are the weights of each optimization target, which are used to balance the relative importance of each optimization target.

[0055] Finally, by optimizing the above objective function, considering trajectory error, external disturbance, and energy optimization factors, the thrust T(t) and attitude quaternion q(t) of the UAV are adjusted and output to the controller.

[0056] Further, the step 8 comprises:

[0057] Offline training: Based on historical flight data D = {X1, X2, ..., X N}, the loss function of the neural network model is:

[0058]

[0059] By minimizing the loss function L(θ), the offline training model learns typical flight patterns and solutions to emergencies;

[0060] Online optimization: During flight, the MPCC framework continuously updates model parameters based on sensor data.

[0061] Improve the adaptability and control accuracy in flight;

[0062] θ t+1 =θ t -η▽ θ L t

[0063] Where η is the learning rate, L t is the loss function at the current time step.

[0064] The model prediction profile control system and method based on neural network enhancement of the present invention has the following advantages:

[0065] 1. Strong real-time response capability: Through the S4-Conv adaptive network architecture in the Sana prediction model, it can quickly process the multi-source sensor data of the drone, accurately predict the position and attitude, and adjust the control strategy in real time according to external collisions or disturbances.

[0066] 2. High adaptability and robustness: The Sana prediction model integrates multi-dimensional information from thrust, odometer and flexible collision sensors, and combines data enhancement technology to improve the drone's adaptability in complex and dynamic environments, making it suitable for efficient mission execution in cities, forests, disaster areas and other scenarios.

[0067] 3. Adaptive adjustment after collision: The Mina control framework integrates neural network enhancement, which can sense and make adaptive control adjustments in real time after a UAV collision, significantly improving the trajectory tracking accuracy and stability of the UAV in complex environments.

[0068] 4. High precision and energy saving: By combining model predictive contour control with high-precision prediction of neural networks, this method optimizes the thrust output of the UAV, ensuring flight efficiency and safety while effectively saving energy consumption.

[0069] 5. Improve the success rate of mission execution: The present invention can maintain flight stability when there is interference in the external environment, reduce the possibility of mission failure due to trajectory deviation or loss of control, and is suitable for various complex flight missions. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a front structural schematic diagram of a quad-rotor drone with a spherical shell of the present invention;

[0071] Figure 2 It is a schematic diagram of the top view of the structure of a quad-rotor drone with a spherical shell of the present invention;

[0072] Figure 3 It is a framework diagram of the model prediction contour control of the present invention combined with the collision neural network enhancement;

[0073] Figure 4 It is a network architecture diagram of the Sana prediction model of the present invention. DETAILED DESCRIPTION

[0074] In order to better understand the purpose, structure and function of the present invention, the following is a further detailed description of a model prediction contour control system and method based on neural network enhancement in conjunction with the accompanying drawings.

[0075] The present invention uses the S4-Conv adaptive network architecture to integrate the data of the odometer, thrust and PVDF flexible collision sensor, and uses the learning and generalization capabilities of the neural network to accurately predict the future position and attitude of the drone. Figure 1 Figure 2 The structure shown. By combining the neural network enhanced model prediction contour control framework, the present invention can quickly respond to sudden collision events and adjust the attitude and thrust output of the UAV in real time, improving the trajectory tracking accuracy and stability of the UAV in complex environments. It consists of two core parts: 1) prediction model; 2) control framework.

[0076] The first core part of this invention proposes the Sana prediction model (S4-Conv Adaptive NetArchitecture, which aims to use the powerful learning and generalization capabilities of neural networks to achieve high-precision prediction of the position and speed of the drone by fusing multi-source information such as flexible collision sensors PVDF, thrust and odometers, so as to adapt to dynamic changes in complex environments. Unlike traditional fixed models, the Sana prediction model can adapt to external disturbances in real time and improve the accuracy of trajectory prediction.

[0077] The second core part of the present invention proposes the Mina control framework (MPCC Integrated with NeuralAugmentation, neural network enhanced model prediction contour control), which introduces neural network enhancement on the basis of traditional model prediction contour control, so that it can cope with emergencies, such as collisions between drones and external objects. By integrating the results of the Sana prediction model, the Mina control framework realizes real-time perception and control adjustment of the post-collision state, solving the problems of slow response speed and inaccurate adjustment of traditional control methods in complex environments, thereby greatly improving the trajectory tracking capability and autonomous adaptability of drones in complex tasks.

[0078] Specifically, a model prediction contour control system based on neural network enhancement of the present invention includes:

[0079] The S4-Conv adaptive network architecture is used to receive and process multi-source sensor data to predict the future position and attitude of the drone. The S4-Conv adaptive network architecture integrates odometer data, thrust output data and collision sensor data, and improves the robustness of data processing through standardization and data enhancement techniques.

[0080] The PVDF flexible collision sensor is used to sense the information of the collision between the UAV and the external object; the PVDF flexible collision sensor senses the collision information through the voltage signal and inputs it into the S4-Conv adaptive network architecture for processing to realize the real-time state adjustment of the UAV after the collision.

[0081] The model prediction contour control framework based on neural network enhancement is used to optimize the thrust output of the drone and adjust the attitude in real time according to the prediction results. The model prediction contour control framework can quickly respond to external environmental changes and collision situations by combining neural network enhancement, and improve the trajectory tracking accuracy and stability of the drone.

[0082] like Figure 3 As shown, a model prediction contour control method based on neural network enhancement of the present invention comprises the following steps:

[0083] Step 1: Synchronization and fusion of multi-source sensor data

[0084] like Figure 4 As shown, data source: The Sana prediction model receives data from multiple source sensors, including information from odometers, thrust sensors, and flexible collision sensors (PVDF). These data fusions are used to accurately perceive and predict the flight status of the drone. Unlike traditional fixed models, the Sana prediction model can adapt to external disturbances in real time and improve the accuracy of trajectory prediction.

[0085] Data synchronization: Since different sensors have different sampling frequencies and delays, a timestamp-based alignment and synchronization method is used to ensure the consistency of multidimensional data at each time step, so that the prediction model can be processed sequentially.

[0086] Assume that different sensors (S1, S2, ..., S n ) are sampled at (t1, t2, ..., t n ), through timestamp alignment, the data synchronization formula is:

[0087]

[0088] In this way, all data are aligned to a uniform time step t. aligned , ensuring the consistency of multi-dimensional sensing data.

[0089] Step 2: Data preprocessing and enhancement

[0090] Preprocessing: For sensor data x t , using the exponentially weighted moving average (EWMA) filtering formula:

[0091]

[0092] Where α is the smoothing coefficient, which is used to control the weight of historical data.

[0093] The sensor data is further standardized so that data of different dimensions have consistent means and variances, thereby avoiding negative impacts on model training.

[0094]

[0095] Among them, μ is the mean of the data, σ is the standard deviation, and after standardization, all sensor data have the same mean 0 and variance 1.

[0096] Data enhancement: Data enhancement technology is introduced to simulate complex situations that may be encountered in different flight scenarios and enhance the robustness of the model. This includes voltage noise simulation, thrust disturbance, random rotation and displacement to enhance the model's adaptability to sensor instability and external disturbances.

[0097] Step 3: Trajectory prediction with Sana prediction model

[0098] S4-Conv adaptive network architecture: The Sana prediction model is based on the S4D state space sequence modeling method. It generates high-precision future trajectory predictions by fusing multi-source sensor data and provides the future position, speed and attitude of the drone.

[0099] Assume X t Representing the multi-source sensor data at time t, predicting the trajectory of the next k time steps, the output trajectory of the Sana prediction model is:

[0100]

[0101] Among them, f θ It is a neural network function based on the S4D state space model, and θ represents the parameters of the network.

[0102] Step 4: Trajectory Optimization for Model Predictive Contour Control

[0103] Trajectory generation and optimization: Based on the results of the drone path search and path optimization, the Mina control framework first calculates the ideal trajectory of the drone to ensure that the drone can follow the target trajectory in the optimal path. The current path is p(t) = {p1, p2, ..., p m}, the target path is p target (t)={p1,p2,…,p m}. The position error is defined as

[0104] e(t)=p target (t)-p(t)

[0105] In the calculation process, using the traditional PID control

[0106]

[0107] Adapt to changes in the external environment in real time, generate initial values ​​of thrust and attitude, and provide a numerical basis for subsequent optimization. p , K i , K d They are the proportional, integral and differential gains of PID control respectively.

[0108] Step 5: Collision Sensing and Adaptive Control

[0109] Collision Detection: When the PVDF flexible collision sensor detects that the drone collides with an external object, the Sana prediction model immediately updates the post-collision state, and the Mina control framework adjusts based on this real-time feedback information.

[0110] Adaptive adjustment: The Mina control framework uses the MPCC mechanism to quickly adjust thrust output and attitude control after a collision, restore the drone to normal flight status, and prevent it from deviating from the trajectory or losing control.

[0111] Step 6: Trajectory tracking in a disturbed environment

[0112] Dynamic environmental adaptation: The MPCC framework monitors external disturbances in real time (such as wind speed changes or sudden obstacles) through the Sana prediction model and adjusts the trajectory based on this data. Through online optimization, the control system can quickly respond to external disturbances and ensure the flight stability of the drone in complex environments.

[0113] Step 7: Real-time Optimization of Control Inputs

[0114] In the framework of model predictive contour control (MPCC), the optimization objectives of the control input include the following four items: minimizing the position error, minimizing the speed error, smoothing the thrust change, and minimizing the thrust power consumption. Assumptions: p(t) represents the current position of the drone, v(t) represents the current speed of the drone, T(t) represents the current thrust of the drone, T prev (t-1) is the thrust at the previous moment, and P(t) is the current thrust power consumption.

[0115] We can define multiple optimization objective functions and combine them into an overall optimization problem:

[0116] (1) Position error minimization: The goal is to minimize the difference between the current position of the drone and the ideal target position p target The error between (t)

[0117] J1=‖p target (t)-p(t)‖ 2

[0118] (2) Speed ​​error minimization: The goal is to minimize the difference between the current speed of the drone and the ideal speed v target The error between (t)

[0119] J2=‖v target (t)-v(t)‖ 2

[0120] (3) Smoothness of thrust change: The goal is to minimize the drasticness of thrust change and maintain smoothness of thrust adjustment.

[0121] J3=‖T(t)-T prev (t-1)‖ 2

[0122] (4) Minimization of thrust power consumption: The goal is to reduce energy consumption and optimize thrust power.

[0123]

[0124] Where P(t) is the power consumption of thrust T(t), which is usually related to the thrust and the load of the UAV, and the calculation formula is:

[0125] P(t)=T(t)·v(t)

[0126] (5) Combine these objective functions to construct a weighted comprehensive optimization objective function:

[0127] J=λ1J1+λ2J2+λ3J3+λ4J4

[0128] Among them, λ1, λ2, λ3, and λ4 are the weights of each optimization objective, which are used to balance the relative importance of each optimization objective.

[0129] Finally, by optimizing the above objective function and comprehensively considering factors such as trajectory error, external disturbance, and energy optimization, the thrust T(t) and attitude quaternion q(t) of the drone are adjusted and output to the controller.

[0130] Step 8: Offline and Online Training

[0131] Offline training: Through offline training, the neural network learns typical flight patterns and emergency response plans so that it can respond quickly in actual flight. Based on historical flight data D = {X1, X2, ..., X N}, the loss function of the neural network model is:

[0132]

[0133] By minimizing the loss function L(θ), the offline training model learns typical flight patterns and solutions to emergencies.

[0134] Online optimization: During the flight, the MPCC framework continuously updates the model parameters based on the sensor data to improve the adaptive capability and control accuracy during flight.

[0135] θ t+1 =θ t -η▽ θ L t

[0136] Where η is the learning rate, L t is the loss function at the current time step.

[0137] Step 9: Final Control Output

[0138] Control output: After completing trajectory optimization, the MPCC framework outputs the final control instructions to the UAV's flight control system to ensure that the UAV continues to perform high-precision trajectory tracking tasks in complex environments. Throughout the process, the Mina control framework combines the feedback from the Sana prediction model to continuously perform real-time optimization and adjustments.

[0139] Through the above steps, the present invention combines the Sana prediction model with the Mina control framework to form a model prediction contour control system based on neural network enhancement, which can perform high-precision trajectory tracking and adaptive control adjustment in real time in complex and dynamic environments, effectively improving the execution efficiency and stability of drones in complex tasks.

[0140] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. A model prediction contour control system based on neural network enhancement, characterized in that: include: S4Conv adaptive network architecture for receiving and processing multi-source sensor data to predict the future position and attitude of the drone; PVDF flexible collision sensor, used to sense the information of the collision between the drone and the external object; the PVDF flexible collision sensor senses the collision information through the voltage signal and inputs it into the S4-Conv adaptive network architecture for processing; A neural network-enhanced model-predictive contour control framework is used to optimize the thrust output of the UAV and adjust the attitude in real time based on the prediction results.

2. A control method for a model prediction profile control system based on neural network enhancement as claimed in claim 1, characterized in that: The steps include: Step 1: Synchronization and fusion of multi-source sensor data; Step 2: Data preprocessing and enhancement; Step 3: Trajectory prediction of Sana prediction model; Step 4: Trajectory optimization for model-predictive contour control.

3. The control method according to claim 2, characterized in that: The step 1 includes data source and data synchronization. The data source is the data received by the Sana prediction model from multiple source sensors, including information from odometers, thrust sensors, and PVDF flexible collision sensors. These data fusions are used to perceive and predict the flight status of the drone. The data synchronization adopts a timestamp-based alignment and synchronization method: Assume that different sensors (S1, S2, ..., S n ) are sampled at (t1, t2, …, t n ), through timestamp alignment, the data synchronization formula is: In this way, all data are aligned to a uniform time step t. aligned , ensuring the consistency of multi-dimensional sensing data.

4. The control method according to claim 2, characterized in that: The step 2 comprises: Preprocessing: For sensor data x t , using the exponentially weighted moving average EWMA filtering formula: Where α is the smoothing coefficient, which is used to control the weight of historical data; The sensor data is standardized so that data of different dimensions have consistent mean and variance: Among them, μ is the mean of the data, σ is the standard deviation, and after standardization, all sensor data have the same mean of 0 and variance of 1; Data augmentation: includes simulation of voltage noise, thrust perturbations, random rotations and displacements.

5. The control method according to claim 2, characterized in that: The step 3 comprises: S4-Conv adaptive network architecture: The Sana prediction model is based on the S4D state space sequence modeling method. It generates high-precision future trajectory predictions by fusing multi-source sensor data and provides the future position, speed and attitude of the drone. Assume X t Representing the multi-source sensor data at time t, predicting the trajectory of the next k time steps, the output trajectory of the Sana prediction model is: Among them, f θ It is a neural network function based on the S4D state space model, and θ represents the parameters of the network.

6. The control method according to claim 2, characterized in that: The step 4 comprises: Trajectory generation and optimization: Based on the results of the drone path search and path optimization, the Mina control framework first calculates the ideal trajectory of the drone to ensure that the drone can follow the target trajectory in the optimal path. The current path is p(t) = {p1, p2, …, p m }, the target path is p target (t)={p1,p2,…,p m }, the position error is defined as: e(t)=p target (t)-p(t) In the calculation process, PID control is used: Among them, K p , K i , K d They are the proportional, integral and differential gains of PID control respectively.

7. The control method according to claim 2, characterized in that: The step 5 comprises: Collision Detection: When the PVDF flexible collision sensor detects that the drone collides with an external object, the Sana prediction model immediately updates the post-collision state, and the Mina control framework adjusts based on this real-time feedback information; Adaptive adjustment: The Mina control framework uses the MPCC mechanism to adjust the thrust output and attitude control after a collision to restore the drone to normal flight status.

8. The control method according to claim 2, characterized in that: The step 6 comprises: Dynamic environmental adaptation: The MPCC framework monitors external disturbances in real time through the Sana prediction model and adjusts the trajectory based on this data. Through online optimization, the control system responds quickly to external disturbances.

9. The control method according to claim 2, characterized in that: The step 7 comprises: In the model predictive contour control framework MPCC, the optimization objectives of the control input include the following four items: minimization of position error, minimization of velocity error, smoothness of thrust change, and minimization of thrust power consumption. Assumptions: p(t) represents the current position of the UAV, v(t) represents the current velocity of the UAV, T(t) represents the current thrust of the UAV, T prev (t-1) is the thrust at the previous moment, P(t) is the current thrust power consumption, Define multiple optimization objective functions and combine them into an overall optimization problem: (1) Position error minimization: The goal is to minimize the difference between the current position of the drone and the ideal target position p target The error between (t): J1=‖p target (t)-p(t)‖ 2 (2) Speed ​​error minimization: The goal is to minimize the difference between the current speed of the drone and the ideal speed v target The error between (t): J2=‖v target (t)-v(t)‖ 2 (3) Smoothness of thrust change: The goal is to minimize the drasticness of thrust change and maintain smoothness of thrust adjustment: J3=‖T(t)-T prev (t-1)‖ 2 (4) Minimization of thrust power consumption: The goal is to reduce energy consumption and optimize thrust power: Where P(t) is the power consumption of thrust T(t), which is usually related to the thrust and the load of the UAV, and the calculation formula is: P(t)=T(t)·v(t) (5) Combine these objective functions to construct a weighted comprehensive optimization objective function: J=λ1J1+λ2J2+λ3J3+λ4J4 Among them, λ1,λ2,λ3,λ4 are the weights of each optimization target, which are used to balance the relative importance of each optimization target. Finally, by optimizing the above objective function, considering trajectory error, external disturbance, and energy optimization factors, the thrust T(t) and attitude quaternion q(t) of the UAV are adjusted and output to the controller.

10. The control method according to claim 2, characterized in that: The step 8 comprises: Offline training: Based on historical flight data D = {X1, X2, ..., X N }, the loss function of the neural network model is: By minimizing the loss function L(θ), the offline training model learns typical flight patterns and solutions to emergencies; Online optimization: During flight, the MPCC framework continuously updates model parameters based on sensor data to improve in-flight adaptability and control accuracy; i t+1 =θ t -η▽ θ L t Where η is the learning rate, L t is the loss function at the current time step.

Images