A UAV Control Method Combining Sequence Modeling and Predictive Control

By combining sequence modeling and predictive control methods, and utilizing the Transformer model and least squares estimation method, the problems of flight stability and control response lag of UAVs under complex wind fields were solved. Real-time compensation for unknown disturbances and accurate flight trajectory generation were achieved, thereby improving the control performance of UAVs in complex environments.

CN120686888BActive Publication Date: 2026-11-10SOUTHWEST JIAOTONG UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202510850349.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-06-24
Publication Date
2026-11-10
Estimated Expiration
2045-06-24

AI Technical Summary

Technical Problem

Existing UAV control methods are difficult to accurately model under complex wind field conditions, resulting in poor flight stability, lagging control response, and weak adaptability to unknown disturbances. Traditional methods also have large prediction residual errors in dynamic wind fields, which cannot meet real-time requirements.

Method used

A method combining sequence modeling and predictive control is adopted. The Transformer model is used to capture the long-term dependent variation characteristics of wind speed disturbance and flight state. The least squares estimation method is used for dynamic adjustment. The optimal flight trajectory is generated by combining MPC optimization decision. The residual aerodynamic force modeling mechanism and online estimation algorithm are introduced to realize real-time compensation for external disturbance forces.

Benefits of technology

It improves the robustness and generalization ability of UAVs in complex wind fields, reduces control response lag, enhances the accuracy and safety of flight paths, and meets the requirements of real-time control.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120686888B_ABST
    Figure CN120686888B_ABST
Patent Text Reader

Abstract

This application discloses a UAV control method combining sequence modeling and predictive control, comprising: acquiring UAV state information, including velocity information, attitude information, position information, and PWM motor control signals; obtaining state observations through a Transformer model based on the acquired state information, and predicting future residual aerodynamic forces based on the state observations and the current residual aerodynamic forces to obtain predicted residual aerodynamic forces; dynamically adjusting the predicted residual aerodynamic forces using a least squares estimation method; and generating an optimal flight trajectory based on the adjusted residual aerodynamic forces and the target trajectory to complete UAV control. This invention integrates a sequence modeling mechanism based on a Transformer structure with a model predictive control strategy, significantly improving the control stability, accuracy, and adaptability of UAVs under complex wind fields.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of human-machine intelligent control and autonomous decision-making technology, and more specifically, to a method for controlling unmanned aerial vehicles that combines sequence modeling and predictive control. Background Technology

[0002] Under complex wind conditions, UAVs face significant aerodynamic disturbances, such as sudden wind speed changes (5-15 m / s), turbulence intensity variations (>20%), and unsteady airflow separation. These disturbances are nonlinear, time-varying, and stochastic, making it difficult for traditional control methods based on linearization assumptions (such as PID and LQR) to accurately model them. For example, PID control is prone to overshoot or oscillations due to fixed gain parameters during sudden wind speed changes, while LQR's trajectory tracking error can increase to three times the baseline value when wind speed steps exceed 8 m / s. In recent years, data-driven methods (such as Neural-Fly) have learned the wind speed-aerodynamic mapping relationship through offline training, but their static neural networks cannot capture temporal correlations, increasing prediction residual errors by 40% in dynamic wind fields, and their reliance on pre-trained data limits generalization ability. Furthermore, existing methods have inference latency exceeding 20 ms on embedded platforms, which is insufficient to meet the real-time requirements (<10 ms) of high-speed maneuvering missions. In practical deployments, control response lags of 50-200 ms occur under dynamic wind fields, attitude angle tracking errors exceed 15° in complex trajectory tasks, and trajectory offsets exceed 0.5 m. Therefore, a technical solution integrating temporal dynamic modeling and real-time optimization control is urgently needed to address these issues. Summary of the Invention

[0003] To address the aforementioned issues, this application provides a UAV control method that combines sequence modeling and predictive control, aiming to solve the problems of poor flight stability, delayed control response, and weak adaptability to unknown disturbances in existing UAVs under dynamic environments such as complex wind fields and strong wind disturbances.

[0004] The first aspect of this invention provides a method for controlling an unmanned aerial vehicle (UAV) that combines sequence modeling and predictive control, comprising: S1. Collect UAV status information, including speed information, attitude information, position information, and PWM motor control signals; S2. Based on the collected state information, obtain the state observation value through the Transformer model, and predict the future residual aerodynamic force based on the state observation value and the current residual aerodynamic force to obtain the predicted residual aerodynamic force. S3. The predicted residual aerodynamic forces are dynamically adjusted using the least squares estimation method. S4. Based on the adjusted residual aerodynamic forces and target trajectory, generate the optimal flight trajectory to complete the UAV control.

[0005] In one optional implementation, step S2 specifically includes the following steps: State observations are obtained using the Transformer model: (1) Where t is the current time and i is the time offset. These are state observations from the past T frames. For velocity vectors, The four elements of posture. This is a PWM motor control signal; Construct a residual aerodynamic force model, and obtain the current residual aerodynamic force based on the residual aerodynamic force model: (2) in, The aerodynamic residuals under the current flight conditions. The actual total aerodynamic force estimated by the sensor. The ideal aerodynamic force is calculated based on the dynamic model; The predicted residual aerodynamic forces were calculated as follows: (3) in, The predicted residual disturbance term is H, where H is the prediction window size of the MPC controller. It is a self-attention-based sequence encoder.

[0006] In one optional implementation, step S3 specifically includes the following steps: Dynamic adjustment is performed using the least squares estimation method: (4) in, The aerodynamic residual characteristics output by the Transformer structure. This represents the weight vector estimated online. Calculate the linear coefficients according to formula (4): (5) in, The coefficients are linear. This is a regularization term.

[0007] In one optional implementation, the optimal flight trajectory generated in S4 is: (6) The constraints are as follows: (7) (8) (9); Where H represents the prediction time domain, The control input sequence is from the current time t to the future time t+H. Let be the predicted state vector at time t+i. Let be the reference state vector at time t+i. Let be the control input vector at time t+i, Q be the symmetric positive definite weight matrix, and R be the symmetric positive definite weight matrix. For the predicted residual disturbance term, To control the minimum value of the physical limit of the input, To control the maximum value of the physical limit of the input, This represents the allowable range of change in the control quantity at adjacent time points.

[0008] In one alternative implementation, the residual aerodynamic model is trained using supervised learning.

[0009] A second aspect of this application provides a control system for an ATM used in an automated oil drilling rig, the system comprising: State perception unit: Collects UAV state information, including speed information, attitude information, position information, and PWM motor control signals; Sequence modeling unit: Based on the collected state information, it obtains state observations through the Transformer model, and predicts future residual aerodynamic forces based on the state observations and the current residual aerodynamic forces to obtain the predicted residual aerodynamic forces. Dynamic adjustment unit: The predicted residual aerodynamic forces are dynamically adjusted using the least squares estimation method; Control optimization unit: Based on the adjusted residual aerodynamic forces and target trajectory, it generates the optimal flight trajectory and completes the control of the UAV.

[0010] In one alternative implementation, the control system is deployed on a resource-constrained embedded platform.

[0011] In this embodiment, the present invention innovatively introduces a Transformer structure for sequence modeling, accurately capturing the long-term dependency characteristics between wind speed disturbances and flight state. Simultaneously, MPC is used to optimize the predicted trajectory, ensuring the accuracy and safety of the flight path. Furthermore, the introduced residual aerodynamic modeling mechanism and online least squares estimation algorithm can estimate and compensate for unmodeled external disturbances in real time, further enhancing the robustness and generalization ability of the control system under complex wind field disturbances. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart of an embodiment of an unmanned aerial vehicle (UAV) control method combining sequence modeling and predictive control, as proposed in this application. Figure 2 This is a structural diagram of an unmanned aerial vehicle (UAV) control system combining sequence modeling and predictive control, as proposed in an embodiment of this application. Detailed Implementation

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0015] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a UAV control method combining sequence modeling and predictive control, as proposed in an embodiment of this application. Figure 1 As shown, the UAV control method combining sequence modeling and predictive control includes: S1. Collect UAV status information, including speed information, attitude information, position information, and PWM motor control signals; S2. Based on the collected state information, obtain the state observation value through the Transformer model, and predict the future residual aerodynamic force based on the state observation value and the current residual aerodynamic force to obtain the predicted residual aerodynamic force. S3. The predicted residual aerodynamic forces are dynamically adjusted using the least squares estimation method. S4. Based on the adjusted residual aerodynamic forces and target trajectory, generate the optimal flight trajectory to complete the UAV control.

[0016] In this embodiment, various sensors carried by the UAV, such as IMU (Inertial Measurement Unit) and GPS (Global Positioning System), are used to collect the UAV's speed, attitude, position and PWM motor control signals in real time and with high accuracy.

[0017] We utilize the self-attention mechanism and multi-layer structure of the Transformer model to analyze time-series data. By mining the complex long-range dependencies between state information such as wind, attitude, and velocity, we capture the dynamic patterns of wind field disturbances and predict future residual aerodynamic forces based on the current residual aerodynamic forces.

[0018] Due to the high complexity and uncertainty of the actual wind field environment, the least squares estimation method is used to dynamically calibrate the predicted future residual aerodynamic forces.

[0019] Based on the principle of Model Predictive Control (MPC), an optimization problem is constructed. MPC takes into account the physical dynamics equations of the UAV and control constraints (such as maximum thrust, attitude angle limits, attitude rate limits, etc.), and generates the optimal flight trajectory by solving the optimization problem.

[0020] Furthermore, S2 specifically includes the following steps: State observations are obtained using the Transformer model: (1) Where t is the current time and i is the time offset. These are state observations from the past T frames. For velocity vectors, The four elements of posture. This is a PWM motor control signal; Construct a residual aerodynamic force model, and obtain the current residual aerodynamic force based on the residual aerodynamic force model: (2) in, The aerodynamic residuals under the current flight conditions. The actual total aerodynamic force estimated by the sensor. The ideal aerodynamic force is calculated based on the dynamic model; The predicted residual aerodynamic forces were calculated as follows: (3) in, The predicted residual disturbance term is H, where H is the prediction window size of the MPC controller. It is a self-attention-based sequence encoder.

[0021] In this embodiment, considering the significant time-dependent characteristics of residual aerodynamic forces, such as gusts and periodic disturbances, this invention employs an improved lightweight Transformer model to replace the traditional RNN or LSTM as the encoder for the aircraft state sequence. This invention uses a distilled version of the Transformer (such as TinyTransformer) structure, with each inference step taking less than 5ms. The Transformer model takes the state observations of the past T frames as input, and through a multi-head attention mechanism and a positional encoding structure, the Transformer model can mine the long-range dependency structure between wind speed disturbances and aircraft response, generating residual aerodynamic forces with temporal memory.

[0022] Residual aerodynamic force is defined as the aerodynamic residual under the current flight condition; it is the difference between the actual total aerodynamic force and the ideal aerodynamic force. The actual total aerodynamic force is estimated by sensors, which are affected by various factors such as wind field and atmospheric environment. The ideal aerodynamic force, on the other hand, is calculated based on a dynamic model. Due to the complexity and variability of the actual flight environment, the ideal aerodynamic model cannot completely and accurately describe the real situation. Therefore, the existence of residual aerodynamic force reflects the combined effects of modeling errors and actual wind field disturbances.

[0023] Based on the preceding time series analysis and residual aerodynamic force model, the Transformer model calculates future residual aerodynamic forces using a self-attention-based sequence encoder. The prediction window size H determines the model's prediction range for the future. Within this window, the model uses a self-attention mechanism to weight historical state data, focusing on information related to future residual aerodynamic forces, thereby outputting the residual force prediction result for the next H steps. This prediction result serves as the predicted residual disturbance term, providing crucial information for subsequent control decisions.

[0024] In one optional implementation, step S3 specifically includes the following steps: Dynamic adjustment is performed using the least squares estimation method: (4) in, The aerodynamic residual characteristics output by the Transformer structure. This represents the weight vector estimated online. Calculate the linear coefficients according to formula (4): (5) in, The coefficients are linear. This is a regularization term.

[0025] In this embodiment, the least squares estimation method includes recursive least squares (RLS) or least squares update with a forgetting factor, used for online real-time adjustment of coefficients. To adapt to current wind speed disturbances.

[0026] Using the least squares estimation method to recursively estimate linear coefficients has the following advantages: High computational efficiency: The time complexity of updating each frame is only O(d). 2 In scenarios with high real-time requirements, it can process data quickly without causing excessive computational burden; Strong dynamic adaptability: It can effectively track changes in disturbance patterns and respond promptly to dynamic changes in external disturbances such as wind fields, ensuring the effectiveness of system control; Excellent noise resistance: It has the optimal solution characteristics under Gaussian noise conditions, and can accurately estimate parameters in environments with noise interference, thereby improving the reliability of the estimation results.

[0027] In one optional implementation, the optimal flight trajectory generated in S4 is: (6) The constraints are as follows: (7) (8) (9); Where H represents the prediction time domain, The control input sequence is from the current time t to the future time t+H. Let be the predicted state vector at time t+i. Let be the reference state vector at time t+i. Let be the control input vector at time t+i, Q be the symmetric positive definite weight matrix, and R be the symmetric positive definite weight matrix. For the predicted residual disturbance term, To control the minimum value of the physical limit of the input, To control the maximum value of the physical limit of the input, This represents the allowable range of change in the control quantity at adjacent time points.

[0028] In this embodiment, the cost function minimized by the MPC controller during the optimization of control input includes: control input smoothing term: the magnitude of control input change between adjacent time steps; residual disturbance compensation loss term: the impact of the compensation term on the overall control effect; and control constraint penalty term: used to apply the physical constraint boundary of the aircraft.

[0029] The control constraints and penalties include: control input restrictions: the motor PWM signal must not exceed the upper and lower limits; attitude angle restrictions: pitch angle and roll angle must be controlled within the safe and stable range of the aircraft (optional); acceleration and speed change restrictions: control continuity and flight safety (optional).

[0030] The MPC optimization problem is solved through rapid iteration using a quadratic programming (QP) solver, and convergence is accelerated by a warm-start mechanism. The output control variable... Used to drive the actual flight control system to execute.

[0031] In one alternative implementation, the residual aerodynamic model is trained using supervised learning.

[0032] In this embodiment, the dataset collected in the flight test wind tunnel is used as the training data for the residual aerodynamic model, and the training objective is to minimize the prediction residual.

[0033] Please refer to Figure 2 , Figure 2 This is a structural diagram of an unmanned aerial vehicle (UAV) control system combining sequence modeling and predictive control, as proposed in an embodiment of this application. Figure 2 As shown, the system includes: State perception unit 201: Collects UAV state information, including speed information, attitude information, position information and PWM motor control signals; Sequence modeling unit 202: Based on the collected state information, it obtains state observation values ​​through the Transformer model, and predicts future residual aerodynamic forces based on the state observation values ​​and the current residual aerodynamic forces to obtain the predicted residual aerodynamic forces. Dynamic adjustment unit 203: The least squares estimation method is used to dynamically adjust the predicted residual aerodynamic forces; Control optimization unit 204: Generates the optimal flight trajectory based on the adjusted residual aerodynamic forces and target trajectory, and completes the control of the UAV.

[0034] Furthermore, the control system is deployed on a resource-constrained embedded platform.

[0035] In this embodiment, the predictive control optimizer runs on a Raspberry Pi 4 platform with a total inference latency of less than 15ms. The Transformer model is lightweighted through pruning, quantization, and other means to ensure that its inference latency does not exceed 10ms.

[0036] This system has been validated in the following example scenarios: Example 1: Wind tunnel environment test. The wind tunnel environment can simulate specific wind speed conditions. At a wind speed of 43.6 km / h, the average trajectory error of the system is less than 8cm.

[0037] Example 2: Outdoor wind field test. The outdoor wind field environment is complex and contains various uncontrollable factors, such as unstable airflow and sudden changes in wind direction. In an outdoor wind field with a wind speed of 17 km / h, the systematic error is maintained at 7.5 cm.

[0038] Example 3: Complex trajectory tasks. Complex trajectories mean that the drone's attitude and speed need to be adjusted frequently, placing higher demands on the control system. When faced with complex trajectory tasks such as elliptical paths, the algorithm can quickly adapt to changes in wind speed and maintain stable flight.

[0039] Example 4: When using drones with different architectures or changing flight control platforms, there is no need to retrain the Transformer; only the basis function weights need to be re-estimated.

[0040] In summary, traditional control methods in the field of UAV intelligent control suffer from numerous drawbacks in complex wind field environments. Traditional methods, such as Neural-Fly, which formulate control strategies based on feature encoding and linear weights, possess some adaptive capabilities. However, when faced with sudden wind changes and frequent trajectory shifts, they lack sufficient understanding of temporal disturbance patterns, leading to increased errors and delayed response. Furthermore, their use of static regression models for residual aerodynamic modeling fails to accurately characterize dynamic disturbance structures, hindering the full utilization of predictive compensation and resulting in poor UAV flight stability, delayed control response, and weak adaptability to unknown disturbances.

[0041] This invention innovatively proposes a UAV control method combining sequence modeling and predictive control, integrating the advantages of data-driven and physics-driven approaches to construct a closed-loop control framework of "perception-prediction-compensation-decision". Compared with traditional static neural networks, the Transformer structure has stronger long-term dependency modeling capabilities, accurately capturing time-varying disturbance trends under dynamic wind fields, realizing the transformation of wind disturbance modeling from "outcome modeling" to "dynamic process modeling", significantly improving disturbance identification accuracy and foresight. Based on physical dynamic equations, MPC integrates the residual aerodynamic force prediction results output by the Transformer and introduces control constraints such as maximum thrust, attitude angle limits, and attitude rate limits. Within the finite prediction time domain, MPC generates the optimal control sequence in real time by solving an optimization problem, ensuring that the control strategy is executed within the physically feasible space, improving the physical interpretability and executability of the control strategy, and effectively enhancing flight safety and system stability. Furthermore, this invention introduces an online least squares estimation algorithm to continuously update the linear structure parameters of the disturbance. This enables adaptive adjustment of specific wind speed parameters, enhancing the system's real-time response capability. Meanwhile, the model decoupling structure separates sequence modeling and control strategies into independent functional modules, which facilitates deployment and optimization.

[0042] The control method of this invention, through a lightweight Transformer model and Model Predictive Control (MPC) algorithm, can be efficiently deployed on edge computing platforms such as Raspberry Pi. Addressing the 20ms latency issue of existing methods, this invention achieves a breakthrough of <10ms single inference latency, fully meeting the real-time requirements of UAV flight control systems. Its technical advantages are mainly reflected in the following application scenarios: Complex wind field environment: As verified in Example Scenario 1, under a wind speed of 43.6 km / h (approximately 12 m / s), the average trajectory error of the system is less than 8 cm, which is significantly better than the error level of the traditional method described in the background technology; High-maneuverability control: MPC optimization framework supports high-speed maneuvering missions under control constraints (including maximum thrust, attitude angle limits, etc.); In dense environment operations: Transformer's sequence modeling capabilities can effectively handle airflow disturbances in occluded environments such as building clusters, improving flight stability.

[0043] This technology has been verified through wind tunnel experiments and field tests, and is compatible with mainstream flight control hardware platforms. As described in the specific implementation, the system can adapt to different UAV architectures without retraining the Transformer model, and has significant application value in fields where traditional methods are insufficient, such as logistics delivery, power line inspection, and disaster relief.

[0044] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0045] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0046] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0047] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0048] The above provides a detailed description of a UAV control method combining sequence modeling and predictive control. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and its core ideas. At the same time, those skilled in the art will recognize that there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for controlling an unmanned aerial vehicle (UAV) combining sequence modeling and predictive control, characterized in that, include: S1. Collect UAV status information, including speed information, attitude information, position information, and PWM motor control signals; S2. Based on the collected state information, state observations are obtained through a Transformer model. Then, based on these state observations and the current residual aerodynamic forces, future residual aerodynamic forces are predicted to obtain the predicted residual aerodynamic forces. The state observations are obtained through a Transformer model. (1) Where t is the current time and i is the time offset. These are state observations from the past T frames. For velocity vectors, The four elements of posture. This is a PWM motor control signal; Construct a residual aerodynamic force model, and obtain the current residual aerodynamic force based on the residual aerodynamic force model. (2) in, The aerodynamic residuals under the current flight conditions. The actual total aerodynamic force estimated by the sensor. For ideal aerodynamic forces calculated based on dynamic models, The predicted residual aerodynamic forces were calculated as follows: (3) in, For the predicted residual aerodynamic forces, The predicted residual disturbance term is H, where H is the prediction window size of the MPC controller. It is a self-attention-based sequence encoder; S3. The predicted residual aerodynamic forces are dynamically adjusted using the least squares estimation method. (4) in, Let be the aerodynamic residual at time i. The aerodynamic residual characteristics output by the Transformer structure. The transpose of the dynamic aerodynamic residual characteristics. This represents the weight vector estimated online. Calculate the linear coefficients according to formula (4). (5) in, The coefficients are linear. Let be the weight vector to be solved. For regularization terms, The dynamic and aerodynamic residual characteristics at time i; S4. Based on the adjusted residual aerodynamic forces and target trajectory, generate the optimal flight trajectory to complete the UAV control.

2. The UAV control method combining sequence modeling and predictive control according to claim 1, characterized in that, The optimal flight trajectory generated in S4 is: (6) The constraints are as follows: (7) (8) (9); Where H represents the prediction time domain, The control input sequence is from the current time t to the future time t+H. Let be the predicted state vector at time t+i. Let be the reference state vector at time t+i. Let be the control input vector at time t+i, Q be the symmetric positive definite weight matrix, and R be the symmetric positive definite weight matrix. The physical dynamics equations for the unmanned aerial vehicle (UAV) are as follows: For the predicted residual disturbance term, To control the minimum value of the physical limit of the input, To control the maximum value of the physical limit of the input, This refers to the allowable range of change in the control quantity at adjacent time points. To control the physical limiting of the input. for Physical limit of the control input at any time.

3. The UAV control method combining sequence modeling and predictive control according to claim 2, characterized in that, The residual aerodynamic model is trained using supervised learning.

4. The UAV control system combining sequence modeling and predictive control according to any one of claims 1-3, characterized in that, The system includes: State perception unit: Collects UAV state information, including speed information, attitude information, position information, and PWM motor control signals; Sequence modeling unit: Based on the collected state information, it obtains state observations through the Transformer model, and predicts future residual aerodynamic forces based on the state observations and the current residual aerodynamic forces to obtain the predicted residual aerodynamic forces. Dynamic adjustment unit: The predicted residual aerodynamic forces are dynamically adjusted using the least squares estimation method; Control optimization unit: Based on the adjusted residual aerodynamic forces and target trajectory, it generates the optimal flight trajectory and completes the control of the UAV.

5. The UAV control system combining sequence modeling and predictive control according to claim 4, characterized in that, The control system is deployed on a resource-constrained embedded platform.

Citation Information

Patent Citations

  • Fixed-wing unmanned aerial vehicle incremental element flight pneumatic modeling method considering wind interference

    CN117875090A

  • Self-adaptive control unmanned aerial vehicle flight path correction method and system

    CN120143870A