An unmanned aerial vehicle adaptive anti-wind measuring system fusing wind disturbance compensation

Abstract

The application relates to the technical field of unmanned aerial vehicle control and discloses an unmanned aerial vehicle adaptive wind-resistant wind measurement system fusing wind disturbance compensation. The system comprises an airborne sensing and measuring unit, an airborne flight control and computing unit and a ground station monitoring and calibration unit; the airborne end constructs a wind disturbance estimation and wind-resistant control closed loop: volume Kalman filtering is used to separate unmanned aerial vehicle movement and wind field state, light-weight neural network compensation is combined to compensate flow field distortion, high-precision wind measurement is realized; a feedforward compensation is generated based on a nonlinear disturbance observer, and wind disturbance is actively inhibited; the ground station performs parameter adaptive setting and neural network multi-granularity automatic pruning, and realizes model light-weight deployment. The application solves the coupling problem of measurement uncertainty and control instability, and significantly improves the wind measurement precision, flight stability and computing efficiency in a complex meteorological environment.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) control technology, and in particular to an adaptive wind-resistant wind measurement system for UAVs that integrates wind disturbance compensation. Background Technology

[0002] With the rapid development of the low-altitude economy, using drones equipped with meteorological sensors for detailed atmospheric boundary layer wind field detection has become an important technical means in fields such as wind energy resource assessment, pollution diffusion prediction, and aviation meteorological support. Compared with traditional methods such as wind measurement towers and weather balloons, drone wind measurement has significant advantages such as maneuverability, high spatial resolution, and reusability. However, existing drone wind measurement systems face severe challenges in practical applications: First, airborne anemometers measure relative wind speeds that include the wind speed itself, and traditional vector subtraction ignores dynamic differences in sensors and rotor downwash interference, resulting in limited compensation accuracy; second, random disturbances in the atmospheric wind field affect wind measurement accuracy and threaten flight safety, and traditional flight control systems lack the ability to actively sense and compensate for wind disturbances; third, most systems use fixed parameters, making it difficult to adapt to changes in complex meteorological environments.

[0003] Therefore, it is necessary to design an adaptive wind-resistant wind measurement system for UAVs that integrates wind disturbance compensation. By constructing a coupled model of the UAV's motion state and the wind field state, it is possible to accurately estimate and actively suppress wind disturbance, and complete high-precision and high-reliability atmospheric wind field measurement while ensuring flight stability. Summary of the Invention

[0004] This invention proposes an adaptive wind-resistant wind measurement system for UAVs that integrates wind disturbance compensation. To address the problems of existing UAV wind measurement systems being greatly affected by their own maneuvering, having weak resistance to sudden gusts, and having low model deployment efficiency, this invention constructs an end-to-end architecture that coordinates wind disturbance estimation, wind resistance control, and model compression, achieving multiple improvements in flight stability, wind measurement accuracy, and computational efficiency under strong disturbance environments.

[0005] This invention provides an adaptive wind-resistant wind measurement system for unmanned aerial vehicles (UAVs) that integrates wind disturbance compensation. The system includes: an airborne sensing and measurement unit, an airborne flight control and computing unit, and a ground station monitoring and calibration unit.

[0006] The airborne sensing and measurement unit is fixed to the fuselage of the UAV and collects inertial data, positioning and ground speed data, raw relative wind and atmospheric state data, and gust disturbance precursor signals as multi-source data.

[0007] The airborne flight control and computing unit, electrically connected to the airborne sensing and measurement unit, includes: a lightweight neural network inference engine, deployed with a dense sub-network generated by the ground station monitoring and calibration unit, outputting flow field distortion compensation; a wind field inversion and compensation module, which establishes a UAV kinematic model and a wind field coupling model, and uses a fluid-structure interaction state estimator based on capacitive Kalman filtering to separate and estimate the UAV's motion state and three-dimensional atmospheric wind field vector in real time, wherein the fluid-structure interaction state estimator contains preset model parameters; this module also calls the flow field distortion compensation network to output compensated real atmospheric wind field data and current gust intensity factor; and an adaptive wind-resistant flight control module, which takes real atmospheric wind field data, current gust intensity factor, and preset flight mission route as input, introduces a feedforward compensation term based on a nonlinear disturbance observer, and generates aircraft control signals.

[0008] The health management and data fusion module receives inertial data, positioning and ground speed data, raw relative wind measurement data, and gust disturbance precursor signals as multi-source data. It performs time synchronization and quality identification processing on the multi-source data to generate fused wind field data, UAV status data, and alarm information, and encapsulates them into standardized downlink data packets.

[0009] The ground station monitoring and calibration unit, deployed at the ground control terminal, communicates bidirectionally with the airborne flight control and computing unit, including:

[0010] A high-precision reference weather station interface is provided, which introduces measured data from standard weather stations as a reference benchmark; an adaptive system parameter tuning module performs offline iterative optimization of the parameters of the wind field inversion and compensation module and the adaptive wind-resistant flight control module; a multi-granularity automatic pruning framework is used to prune and train the initial flow field distortion compensation network, generate a dense sub-network and solidify it into a lightweight neural network inference engine; and a visual mission planning module displays the UAV status and provides route editing and mission command issuance.

[0011] Furthermore, the wind field inversion and compensation module outputs compensated real atmospheric wind field data and the current gust intensity factor, specifically including the following steps:

[0012] Step A1: Construct the state vector of the UAV-wind field coupled system;

[0013] Step A2: Establish the system state transition equation, using the previous state vector and the body inertial data as inputs, to obtain the predicted value of the current state vector;

[0014] Step A3: Establish the system observation equation, which maps the state vector to the observation space. Define the observation vector as consisting of positioning and ground speed data and raw relative wind measurement data.

[0015] Step A4: Introduce a fluid-structure interaction state estimator based on capacitive Kalman filtering, perform capacitive Kalman filtering time update, and obtain the one-step prediction value of the state vector and its error covariance;

[0016] Step A5: Perform a capacitive Kalman filter measurement update, substitute the current state vector prediction into the observation equation and correct the prediction to obtain the optimal state estimate that includes the current actual motion state of the UAV.

[0017] Step A6: Extract the three-dimensional atmospheric wind field vector from the optimal state estimate, call the flow field distortion compensation network in the lightweight neural network inference engine to obtain the flow field distortion compensation amount, superimpose the flow field distortion compensation amount onto the three-dimensional atmospheric wind field vector, and output the compensated real atmospheric wind field data; calculate the current gust intensity factor based on the rate of change and spatial gradient of the extracted three-dimensional atmospheric wind field vector and output it.

[0018] Furthermore, the specific process by which the adaptive wind-resistant flight control module outputs aircraft control signals is as follows:

[0019] Step B1: Construct a six-degree-of-freedom nonlinear dynamic model of the UAV that includes the effects of wind disturbance, and transform the real atmospheric wind field data into wind disturbance forces and wind disturbance moments acting on the UAV body;

[0020] Step B2: Design a nonlinear disturbance observer, construct the dynamic equation of the nonlinear disturbance observer to estimate the total disturbance, and obtain the total disturbance estimate;

[0021] Step B3: Design a robust controller based on backstepping, using the tracking error between the preset flight mission route and the current actual motion state of the UAV as input, and the total disturbance estimate output by the nonlinear disturbance observer as the feedforward compensation term to calculate the aircraft control signal.

[0022] Step B4: Based on the current gust intensity factor, dynamically adjust the observer gain of the nonlinear disturbance observer and the control parameters of the robust controller.

[0023] Furthermore, the process of generating dense subnetworks from the multi-granularity automatic pruning framework is as follows:

[0024] Step D1: Construct an initial flow field distortion compensation network, which contains multiple macroscopic structural blocks, each containing several microscopic structural units;

[0025] Step D2: Assign a macroscopic gating parameter to each macroscopic structural block and a microscopic gating parameter to each microscopic structural unit; transform the macroscopic and microscopic gating parameters into continuous gating values ​​using the Gumbel-Sigmoid function, which includes an adjustable temperature parameter to control the relaxation of the gating values; the gating values ​​correspond to the retention probabilities of the respective macroscopic structural blocks or microscopic structural units.

[0026] Step D3: Construct a joint loss function, which includes a task loss term, a computational complexity penalty term, and a structural feasibility penalty term. The computational complexity penalty term is designed based on the target computing power constraints of the embedded platform.

[0027] Step D4: During the training process, a temperature annealing strategy is used to gradually reduce the temperature parameter of the Gumbel-Sigmoid function, so that the gate value converges to 0 or 1, and the converged gate value is obtained to realize the sparsity selection of the network structure, and finally obtain a lightweight network structure that meets the computing power constraints.

[0028] Step D5: After training, extract the dense subnetwork based on the converged gating value and embed it into the lightweight neural network inference engine.

[0029] Furthermore, the workflow of the system parameter adaptive tuning module is as follows: receiving real atmospheric wind field data, and combining it with reference benchmarks and UAV state data, using a Bayesian optimization-based offline parameter tuning algorithm to iteratively optimize the model parameters included in the fluid-structure interaction state estimator and the observer gain of the nonlinear disturbance observer, and then sending the optimized parameters back to the airborne flight control and computing unit.

[0030] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0031] This invention overcomes the limitations of traditional UAV wind measurement systems that separate flight control from wind field measurement by constructing a closed-loop architecture of "wind disturbance estimation-wind resistance control." The wind field inversion and compensation module introduces a fluid-structure interaction state estimator based on capacitive Kalman filtering, which can separate the UAV's motion state from the three-dimensional atmospheric wind field vector in real time and accurately. It also utilizes a lightweight neural network to dynamically compensate for flow field distortion, significantly improving the accuracy of wind measurement data. Simultaneously, the adaptive wind resistance flight control module, based on a nonlinear disturbance observer, transforms the estimated wind field information into a feedforward compensation term and integrates it into the control law, achieving active suppression of sudden gusts and stable strong winds, ensuring the attitude stability and trajectory tracking accuracy of the UAV under complex weather conditions. The synergistic effect of these two modules solves the technical challenge of the coupling between "inaccurate measurement" and "unstable control."

[0032] This invention introduces a ground station monitoring and calibration unit, enabling the system to continuously learn and self-optimize. On one hand, the system parameter adaptive tuning module utilizes the data transmitted after each flight, combined with a high-precision reference meteorological station benchmark, to perform offline iterative tuning of the wind field inversion model and wind resistance control parameters using Bayesian optimization, allowing the system performance to continuously evolve with the increase in flight frequency. On the other hand, the multi-granularity automatic pruning framework adopts an innovative hierarchical gating mechanism to perform end-to-end multi-granularity pruning training on the initial flow field distortion compensation network. This can automatically generate a compact, dense sub-network that does not require sparse library support while ensuring model accuracy, and then deploy it in the airborne lightweight neural network inference engine. This solves the bottleneck of limited airborne computing resources and ensures low latency and high energy efficiency for model inference.

[0033] This invention utilizes a health management and data fusion module to synchronize and quality-identify heterogeneous data from multiple sources, including inertial measurement, satellite positioning, atmospheric data, and gust detection, and monitors sensor health and flight status in real time. When data anomalies or sensor malfunctions are detected, the system automatically triggers redundancy switching, significantly improving its fault tolerance and survivability in complex environments. The final output downlink data packet integrates rigorously verified and fused wind field data, UAV status information, and alarm information, providing users with integrated, high-confidence data support. It can be widely applied in high-value-added scenarios such as wind energy resource assessment, pollution diffusion prediction, and aviation meteorological support. Attached Figure Description

[0034] Figure 1 This is a system structure block diagram of an adaptive wind-resistant wind measurement system for unmanned aerial vehicles that integrates wind disturbance compensation, as proposed in this invention.

[0035] Figure 2 This is a flowchart illustrating the training and deployment process of the multi-granularity automatic pruning framework in Embodiment 3 of the present invention. Detailed Implementation

[0036] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0037] Example 1:

[0038] according to Figure 1This invention provides an adaptive wind-resistant wind measurement system for UAVs that integrates wind disturbance compensation. It is applicable to various platforms such as fixed-wing UAVs, multi-rotor UAVs, and compound-wing UAVs. The system includes an airborne sensing and measurement unit, an airborne flight control and computing unit, and a ground station monitoring and calibration unit.

[0039] The airborne sensing and measurement unit is fixed to the fuselage of the UAV and collects multi-source data. This unit specifically includes:

[0040] Inertial Measurement Module: Employs high-precision fiber optic gyroscopes and quartz flexural accelerometers to collect real-time inertial data of the UAV, including three-axis angular rate and three-axis acceleration, with an update frequency of 200 Hz;

[0041] Satellite positioning module: It adopts a multi-mode multi-frequency RTK-GPS / INS combined navigation board to collect the UAV's positioning and ground speed data in real time, including geographical location, ground speed and heading angle, with an update frequency of 50 Hz;

[0042] Atmospheric data module: includes a multi-point ultrasonic anemometer and temperature, humidity and pressure sensor fixed to the upper part of the UAV arm, which measures raw relative wind measurement data and atmospheric state data, including relative wind speed, relative wind direction, atmospheric temperature, humidity and air pressure, with an update frequency of 20 Hz;

[0043] Gust detection module: Employs forward-facing millimeter-wave radar or an airborne multi-hole probe to detect sudden airflow disturbances within a certain distance in front of the UAV and collect precursor signals of gust disturbances;

[0044] The airborne flight control and computing unit is electrically connected to the airborne sensing and measurement unit. It employs a high-performance embedded computer running a real-time operating system and integrates the following modules:

[0045] Lightweight neural network inference engine: Deployed with a dense subnetwork generated by the multi-granularity automatic pruning framework in the ground station monitoring and calibration unit; This engine outputs the flow field distortion compensation in real time in forward inference mode, and its inference latency is strictly controlled at an extremely low level to ensure that it does not affect the overall control cycle.

[0046] Wind field inversion and compensation module: This module establishes a kinematic model of the UAV and a coupled wind field model, and uses a fluid-structure interaction (FSI) state estimator based on capillary Kalman filtering to separate and estimate the UAV's motion state and the three-dimensional atmospheric wind field vector in real time. The FSI state estimator includes preset model parameters. This module takes inertial data, positioning and ground speed data, and raw relative wind measurement data as input. By introducing the FSI state estimator based on capillary Kalman filtering, the module separates and estimates the UAV's motion state (three-axis position, three-axis velocity, quaternion attitude, and three-axis angular rate bias) and the three-dimensional atmospheric wind field vector in real time. Simultaneously, it calls a flow field distortion compensation network deployed in a lightweight neural network inference engine to calculate the flow field distortion compensation amount, and finally outputs the compensated real atmospheric wind field data and the current gust intensity factor.

[0047] The fluid-structure interaction state estimator includes the following preset model parameters:

[0048] Process noise covariance matrix Q: describes the statistical characteristics of process noise of each state component (including three-axis position, three-axis velocity, quaternion attitude, three-axis angular rate bias and three-dimensional atmospheric wind field vector) in the system state transition model. Its diagonal element values ​​are preset according to the sensor noise and wind field fluctuation.

[0049] Measurement noise covariance matrix R: describes the statistical characteristics of measurement noise in the positioning and ground speed data and raw relative wind measurement data in the observation equation, and is preset according to the nominal accuracy of the airborne satellite positioning module and ultrasonic anemometer;

[0050] Wind field related time constant τ: In the wind field evolution model (first-order Markov process), it controls the memory decay rate between the wind field at the current moment and the wind field at the previous moment. Its value is preset according to the typical atmospheric boundary layer wind field change law, and is usually set to 80 seconds.

[0051] Initial state error covariance matrix P0: characterizes the uncertainty of the initial estimate of the state vector, and its value is set according to the prior statistical information of the initial alignment accuracy and atmospheric wind field before the UAV takes off.

[0052] Number of volume points and weight coefficients: Based on the standard algorithm of volumetric Kalman filtering, the third-order spherical radial volume rule is adopted to determine 2n volume points (n is the dimension of the state vector) and their corresponding weights. The coefficients are fixed preset values.

[0053] The aforementioned preset model parameters can be iteratively optimized offline in the system parameter adaptive tuning module of the ground station monitoring and calibration unit;

[0054] Adaptive wind-resistant flight control module: Taking real atmospheric wind field data, gust intensity factor and preset flight mission route as input, this module is based on inner loop attitude control and outer loop trajectory control, and introduces a feedforward compensation term based on nonlinear disturbance observer to generate aircraft control signal containing wind-resistant correction in real time, so as to suppress wind disturbance on the attitude and trajectory of UAV, with a control cycle of 5 ms.

[0055] The health management and data fusion module receives and processes the multi-source data, generating fused wind field data, UAV status data, and alarm information, and encapsulates them into standardized downlink data packets. This module performs time synchronization and quality identification processing on the multi-source data based on a precise time protocol, monitors the UAV flight status and sensor health status in real time, and can trigger redundancy switching when data anomalies or sensor malfunctions are detected. Finally, this module generates verified and fused wind field data, UAV status data, and alarm information, encapsulates them into standardized downlink data packets, and sends them to the ground station monitoring and calibration unit via a data transmission link.

[0056] The ground station monitoring and calibration unit is deployed at the ground control terminal and communicates bidirectionally with the airborne flight control and computing unit via a data transmission link. This unit integrates the following components:

[0057] High-precision reference weather station interface: Accesses measured data from nearby standard weather stations via 4G / 5G network, serving as a reference benchmark for system calibration and post-calibration;

[0058] System parameter adaptive tuning module: Receives downlink real atmospheric wind field data and UAV status data from the airborne flight control and computing unit, and combines it with the reference benchmark obtained from the high-precision reference meteorological station interface. It uses an offline parameter tuning algorithm based on Bayesian optimization to iteratively optimize the model parameters included in the fluid-structure interaction state estimator and the observer gain of the nonlinear disturbance observer, and transmits the optimized parameters back to the airborne flight control and computing unit through the data transmission link.

[0059] Multi-granularity automatic pruning framework: Using historical flight data and reference benchmarks from the downlink of the airborne flight control and computing unit as training datasets, an initial flow field distortion compensation network is constructed and pruned and trained to generate a dense sub-network, which is then embedded into a lightweight neural network inference engine. This framework minimizes the computational load of the model while ensuring the accuracy of wind field estimation, and finally outputs a compact, directly deployable dense sub-network, which is then embedded into the lightweight neural network inference engine of the airborne flight control and computing unit.

[0060] Visualized mission planning module: Developed based on a map engine, it displays the drone's real-time location, flight trajectory, 3D wind field distribution map, and system health status to users, and provides graphical editing of flight routes and the function of issuing mission commands.

[0061] Example 2:

[0062] This embodiment, based on Embodiment 1, focuses on illustrating the collaborative workflow between the wind field inversion and compensation module and the adaptive wind-resistant flight control module. In this embodiment, the modules in the airborne flight control and computing unit work together through the following process to achieve accurate wind measurement and stable flight under wind-resistant conditions:

[0063] First, the wind field inversion and compensation module, based on a capacitive Kalman filter-based fluid-structure interaction state estimator, outputs the optimal state estimate for the current moment in the following specific process:

[0064] Step A1: Construct the system state vector of UAV-wind field coupling, which includes three-axis position, three-axis velocity, quaternion attitude, three-axis angular rate bias and three-dimensional atmospheric wind field vector;

[0065] Step A2: Establish the system state transition equation. Using the body's inertial data as input, the state transition function describes the rigid body kinematics and wind field evolution model. Specifically, the current state vector is calculated from the previous state vector, the current angular velocity, and the acceleration through the state transition function, and process noise is added. The wind field is modeled as a first-order Markov process, meaning that the wind field at the current moment is linearly decreasing compared to the wind field at the previous moment, and process noise is added. The decreasing coefficient is determined by the reciprocal of the wind field-related time constant and the time step.

[0066] The specific calculation formula for the above wind field evolution model is as follows:

[0067] ;

[0068] in, This represents the current three-dimensional wind field vector. For time step; The wind field-related time constant; This is process noise; This represents the wind field vector from the previous moment; The coefficient determines the decay rate of wind field memory;

[0069] Step A3: Establish the system observation equations. The observation vector consists of positioning and ground speed data, as well as raw relative wind measurement data. The observation function maps the state vector to the observation space. Its key part is the observation equation describing the raw relative wind measurement data: the raw relative wind measurement data equals the UAV's true velocity relative to the atmosphere minus the flow field distortion error term caused by rotor downwash and fuselage obstruction, plus measurement noise. The UAV's true velocity relative to the atmosphere is obtained by subtracting the UAV's velocity from the three-dimensional atmospheric wind field vector in the state vector. The flow field distortion error term transforms the distorted velocity in the body coordinate system to the navigation coordinate system using a rotation matrix. This distorted velocity is modeled as a nonlinear function related to the current flight state, specifically:

[0070] ;

[0071] in, The distortion rate; This is the rotation matrix from the body coordinate system to the navigation coordinate system; For nonlinear functions of flow field distortion; relative speed of the drone ; The angular velocity of the organism;

[0072] Step A4: Introduce a fluid-structure interaction state estimator based on capacitive Kalman filtering and perform time updates using capacitive Kalman filtering. The time update utilizes the state transition equation to predict the state vector and its error covariance in one step, specifically as follows:

[0073] Volume point generation: A third-order spherical radial volume rule is adopted for... 3D state vector (in this embodiment) Given the total dimension of the state vector (17 dimensions), generate Volume point and their weights The standard volume point set is taken as ,in express A unit vector in the 1st dimension is... Elements in each direction (i.e., the front) Take the positive axis first, then (taking the negative axis);

[0074] Propagation of volume point: The volume point from the previous time step is propagated to the current time step through the state transition function, i.e.

[0075] ;

[0076] in, This is an estimate of the optimal state at the previous moment; For the first The first moment One volume point; The data represents the body's inertial data (including triaxial angular velocity and triaxial acceleration) from the previous moment. This is the state transition function;

[0077] ;

[0078] in, For the first The optimal state estimate at time (i.e., the filtered result from the previous time step); For the first The lower triangular matrix of the time-time covariance matrix;

[0079] Calculate the predicted state value: ;

[0080] Calculate the predicted covariance: ;

[0081] in, Here is the state prediction error covariance matrix. (This is the process noise covariance matrix, describing the uncertainty of the system state transition model);

[0082] Step A5: Perform measurement updates using capacitive Kalman filtering. Calculate the Kalman gain using the observation equation and the predicted state value, and then fuse the actual multi-source measurement data at the current time to correct the predicted state vector, obtaining the optimal state estimate for the current time. The specific process is as follows:

[0083] According to the observation function Calculate the observed predicted value for each volume point: ;

[0084] Calculate the observed and predicted mean: ;

[0085] Calculate the new information covariance: Where R is the measurement noise covariance matrix;

[0086] Calculate cross-covariance: ;

[0087] Calculate the Kalman gain: ;

[0088] Correcting the state estimate: using the actual observation vector at the current time. (Including positioning and ground speed data, and raw relative wind measurement data) Correct the state prediction values:

[0089] ;

[0090] Update state error covariance: ;

[0091] Finally, the optimal state estimate at the current moment is obtained. From this, the motion state of the UAV and the three-dimensional atmospheric wind field vector can be extracted;

[0092] Step A6: Extract the three-dimensional atmospheric wind field vector from the optimal state estimate at the current moment. Based on the current flight state, an input feature vector is constructed, and the flow field distortion compensation network in the lightweight neural network inference engine is invoked for forward inference to obtain the flow field distortion compensation amount. The compensation amount is superimposed onto the three-dimensional atmospheric wind field vector, and the final output is the compensated real atmospheric wind field data; at the same time, the current gust intensity factor is calculated and output based on the rate of change and spatial gradient of the real atmospheric wind field data.

[0093] The flow field distortion compensation network is a pre-deployed miniature Transformer structure (4-layer encoder, each layer has 4 attention heads with a head dimension of 32), and its input feature vector is a combination of the current flight state and the preliminary wind field estimate. ;in For relative velocity, For the body's angular velocity, For attitude quaternions;

[0094] The compensated true atmospheric wind field data are as follows:

[0095] ;

[0096] Simultaneously, the current gust intensity factor is calculated based on the extracted three-dimensional atmospheric wind field vector. This factor comprehensively reflects the temporal rate of change and spatial gradient of the wind field, and the specific calculation formula is as follows:

[0097] ;

[0098] in, The wind field's temporal rate of change (approximately by the difference between two consecutive frames); The wind field spatial gradient (spatial difference along the UAV flight path); , Preset weighting coefficients ( The sigmoid function maps the weighted sum to the interval (0, 1) and outputs the current gust intensity factor; the final output is the compensated real atmospheric wind field data and the current gust intensity factor.

[0099] Furthermore, the adaptive wind-resistant flight control module, based on a feedforward compensation controller of a nonlinear disturbance observer, outputs the aircraft control signal as follows:

[0100] Step B1: Construct a six-degree-of-freedom nonlinear dynamic model of the UAV that includes wind disturbance effects, transforming real atmospheric wind field data into wind disturbance forces and moments acting on the UAV body; the specific dynamic model is as follows:

[0101] ;

[0102] in, The flight state vector (including position, velocity, attitude angle, and angular rate) For control input (motor speed or rudder deflection angle), and Given a known nonlinear function, This represents the total disturbance term caused by wind disturbance.

[0103] Step B2: Design a nonlinear disturbance observer. First, define an auxiliary variable to construct the intermediate state of the disturbance observer. Then, construct the dynamic equation of the nonlinear disturbance observer to drive the disturbance estimate to converge to the true total disturbance and output the total disturbance estimate.

[0104] The specific process for designing a nonlinear perturbation observer is as follows:

[0105] First, define the auxiliary variable:

[0106] ;

[0107] in, As an auxiliary variable; This is the estimated total disturbance. For the design of the system state The nonlinear function; the auxiliary variable Used to construct intermediate states of the perturbation observer to avoid direct measurement of the state derivative;

[0108] Then, the dynamic equations of the nonlinear perturbation observer are constructed:

[0109] ;

[0110] in, Total disturbance estimate The first derivative with respect to time, i.e., the rate of change of the disturbance estimate; The observer gain matrix; This refers to the nonlinear function related to disturbances separated from the UAV dynamics model; To control the input; this dynamic equation ensures that, in the absence of model error, the disturbance estimate is... It converges exponentially to the true total disturbance. This allows for the output of an accurate estimate of the total disturbance.

[0111] Step B3: Design a robust controller based on backstepping, using the tracking error between the desired trajectory and the actual state as input, and the total disturbance estimate as the feedforward compensation term, to progressively back-calculate the virtual control quantity and the final actual control quantity that can offset the wind disturbance and achieve trajectory tracking.

[0112] The backstepping method is a recursive design approach suitable for nonlinear systems with strict feedback. It takes the eye-tracking error between the desired trajectory and the actual state as input, starting from the outermost layer of the system (position error), and progressively designs the virtual control quantity in reverse order until the final actual control quantity is obtained. Specifically, the tracking error is defined... ; This is the flight state vector (containing position, velocity, attitude angles, and angular rate). To determine the desired trajectory, a Lyapunov function is designed, and a total perturbation estimate is introduced. As a feedforward compensation term, the control law is obtained:

[0113] ;

[0114] in, This is a feedback control term based on the backstepping method; for The generalized inverse matrix; this control law can counteract the effects of wind disturbance and achieve trajectory tracking;

[0115] Step B4: Based on the current gust intensity factor output in step A6 Dynamically adjust the observer gain of the nonlinear disturbance observer and the damping coefficient in the backstepping robust controller This enables adaptive adjustment of the control system to the meteorological environment; the adjustment rules are as follows:

[0116] when When the wind is strong (i.e., under strong gust conditions), increase the observer gain to improve the response speed of disturbance estimation, and at the same time increase the damping coefficient to enhance the system's ability to suppress wind disturbances.

[0117] when When the conditions are relatively mild (i.e., weak winds or stable weather conditions), reduce the observer gain to decrease sensitivity to measurement noise, while reducing the damping coefficient to maintain control compliance.

[0118] Specifically, the adaptive law adopts a linear weighted form:

[0119] ;

[0120] ;

[0121] in, The baseline value for the observer gain (corresponding to) (Minimum gain at time) This is the adaptive coefficient for the observer gain, controlling the magnitude of the gain increase with gust intensity; The reference value for the damping coefficient (corresponding to) (Minimum damping at time) The adaptive coefficients for the controller damping coefficient; the two formulas above respectively give the observer gain and controller damping coefficient as a function of the current gust intensity factor. The specific calculation method of linear adjustment is used to quickly estimate disturbances and enhance robustness under strong wind conditions, and to avoid excessive gain and noise under weak wind conditions.

[0122] The above coefficients and The value of is predetermined in the following way: under typical wind field disturbance conditions (gust intensity factor) Using parameters of 0.2, 0.5, and 0.8 respectively, parameter scanning experiments were conducted using a UAV six-degree-of-freedom simulation platform. With trajectory tracking error and root mean square of attitude fluctuation as optimization objectives, the coefficient combination that optimizes the overall system performance was selected. In this embodiment, , These are the optimal values ​​verified by simulation.

[0123] Example 3:

[0124] This embodiment, based on Embodiment 1, focuses on illustrating the workflow of the multi-granularity automatic pruning framework. Figure 2 The framework achieves automatic pruning and deployment of airborne flow field distortion compensation networks through the following steps:

[0125] Step D1: Construct the initial flow field distortion compensation network. In this embodiment, the network adopts a micro Transformer structure, which includes 4 encoder layers, each layer contains 8 attention heads (macro blocks), and each attention head has a dimension of 32 (micro unit). The network input is the flight state features (10-dimensional), and the output is the flow field distortion compensation amount (3-dimensional).

[0126] Step D2: Assign gating parameters. Assign a macro-gating parameter γi (i=1…32) to each attention head (macro-block), and a micro-gating parameter β{i,j} (j=1…32) to each dimension (micro-unit) within each attention head. Transform the macro-gating parameters and micro-gating parameters into continuous gating values ​​using the Gumbel-Sigmoid function. The Gumbel-Sigmoid function includes an adjustable temperature parameter to control the relaxation of the gating values, making the pruning decision differentiable. The output of the Gumbel-Sigmoid function is between 0 and 1, representing the retention probability.

[0127] Step D3: Construct the joint loss function

[0128] in, The weighting coefficients for the computational complexity penalty term are used to balance model accuracy and computational cost. The weight coefficients for the structural feasibility penalty term ensure the rationality of the network structure after pruning; the methods for determining these two factors are as follows: (log sampling) and A grid search is performed within the (linear sampling) range. After pruning training for each set of coefficients, the model's compensation error and inference latency on the target embedded platform are evaluated on the validation set. Models that meet the computational constraints (inference latency not exceeding the control cycle) are retained, and the set with the smallest compensation error is selected as the optimal value. In this example, the optimal value is... , ;

[0129] in, and This is the balance coefficient; the composition and calculation method of each part of this function are as follows:

[0130] Task loss items The error between the compensation output of the flow field distortion compensation network and the true value provided by the high-precision reference weather station is measured using the mean square error method.

[0131] ;

[0132] in, For training batch size; It is the network to the first The 3D prediction compensation amount for each sample output; Three-dimensional real compensation quantity provided for high-precision reference weather stations; The square of the Euclidean norm of the vector is represented by the sum of squares of the components of the residual. The mean square error is obtained by averaging the squared errors of all samples. Supervised training is then used to make the network output approximate the true distortion value.

[0133] Computational complexity penalty term The target computing power constraint design based on the embedded platform is obtained by differentiable modeling of the multiplication-accumulation operands of all operations in the network; its calculation method is as follows:

[0134] ;

[0135] in, The total number of operations is accumulated for the network multiplication; For the first The gating value for each operation; This represents the basic computational cost of the operation; This represents the maximum computational load allowed by the embedded platform.

[0136] Structural feasibility penalty To prevent pruning from causing complete degradation of network layers, ensure that at least a certain number of micro-units are retained within each macroscopic structural block; the calculation method is as follows:

[0137] ;

[0138] in, For the first A macroscopic structural block; The minimum number of micro-units to be retained; It is an exponential function, when the gate value of the micro-unit... When the value is greater than 0.5, the value is 1, indicating that the cell is retained; otherwise, the value is 0, indicating that it is pruned. This represents the number of missing valid cells in the current macroblock (i.e., the gap from the minimum retention number). If the retention number has been met or exceeded... If the value is negative or zero, then the value is negative or zero. It is the summation of all macroscopic structural blocks; Ensure that only when the number of reserves is insufficient A positive penalty is generated at that time; this penalty forces the pruning process to maintain the rationality of the network structure;

[0139] Step D4: During the training process, a temperature annealing strategy is used to gradually reduce the temperature parameter of the Gumbel-Sigmoid function from the initial 5.0 to 0.1, so that the gate value converges to 0 or 1, in order to achieve the sparsity selection of the network structure and finally obtain a lightweight network structure that meets the computing power constraints.

[0140] Step D5: After training, extract the dense subnetwork based on the converged gating value (greater than 0.5 is considered retained); the possible result is: 2 encoder layers are retained (2 layers are removed by macro-pruning), 4 attention heads are retained in each layer, and 24 dimensions are retained in each head. This subnetwork can run directly and efficiently without sparse computing libraries.

[0141] Step D6: Solidify the weights and structure of the dense subnetwork and burn it into the lightweight neural network inference engine via the data transmission link to complete the automatic pruning and deployment of the model. In subsequent flights, the airborne flight control and computing unit will use the lightweight network to calculate the flow field distortion compensation in real time.

[0142] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention. The actual structure, parameters, and module division are not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. An adaptive wind-resistant wind measurement system for unmanned aerial vehicles (UAVs) integrating wind disturbance compensation, characterized in that, include: Airborne sensing and measurement unit, airborne flight control and computing unit, and ground station monitoring and calibration unit; The airborne sensing and measurement unit collects inertial data of the aircraft, positioning and ground speed data, raw relative wind and atmospheric state data, and gust disturbance precursor signals to form multi-source data; The airborne flight control and computing unit, electrically connected to the airborne sensing and measurement unit, includes: a lightweight neural network inference engine that outputs flow field distortion compensation; a wind field inversion and compensation module that establishes a UAV kinematic model and a wind field coupling model, and uses a fluid-structure interaction state estimator based on capacitive Kalman filtering to separately estimate the UAV's motion state and the three-dimensional atmospheric wind field vector, the estimator containing preset model parameters; this module calls the flow field distortion compensation network to output compensated real atmospheric wind field data and gust intensity factor; an adaptive wind-resistant flight control module that generates aircraft control signals; and a health management and data fusion module that receives and processes multi-source data, generates fused wind field data, UAV status data, and alarm information, and encapsulates them into standardized downlink data packets. The ground station monitoring and calibration unit communicates bidirectionally with the airborne flight control and computing unit, including: a high-precision reference meteorological station interface, which introduces measured data as a reference benchmark; a system parameter adaptive tuning module, which optimizes the parameters of the wind field inversion and compensation module and the adaptive wind-resistant flight control module; a multi-granularity automatic pruning framework, which constructs an initial flow field distortion compensation network, performs pruning training, generates a dense subnetwork, and embeds it into a lightweight neural network inference engine; and a visual mission planning module, which provides route editing and mission command issuance. The process of generating dense subnetworks using the multi-granularity automatic pruning framework is as follows: Step D1: Construct an initial flow field distortion compensation network, which contains multiple macroscopic structural blocks, each containing several microscopic structural units; Step D2: Assign a macroscopic gating parameter to each macroscopic structural block and a microscopic gating parameter to each microscopic structural unit; convert the macroscopic and microscopic gating parameters into continuous gating values ​​using the Gumbel-Sigmoid function, which includes an adjustable temperature parameter to control the relaxation of the gating values; Step D3: Construct a joint loss function, which includes a task loss term, a computational complexity penalty term, and a structural feasibility penalty term, wherein the computational complexity penalty term is designed based on the target computing power constraints of the embedded platform; Step D4: During training, a temperature annealing strategy is used to gradually reduce the temperature parameter, causing the gating values ​​to converge to 0 or 1, resulting in converged gating values; Step D5: After training, extract the dense subnetwork based on the converged gating values ​​and embed it into the lightweight neural network inference engine.

2. The system according to claim 1, characterized in that, The wind field inversion and compensation module outputs compensated real atmospheric wind field data and the current gust intensity factor, specifically including the following steps: Step A1: Construct the state vector of the UAV-wind field coupled system; Step A2: Establish the system state transition equation, using the previous state vector and the body inertial data as inputs, to obtain the predicted value of the current state vector; Step A3: Establish the system observation equation, which maps the state vector to the observation space. Define the observation vector as consisting of positioning and ground speed data and raw relative wind measurement data. Step A4: Introduce a fluid-structure interaction state estimator based on capacitive Kalman filtering, perform capacitive Kalman filtering time update, and obtain the one-step prediction value of the state vector and its error covariance; Step A5: Perform a capacitive Kalman filter measurement update, substitute the current state vector prediction into the observation equation and correct the prediction to obtain the optimal state estimate that includes the current actual motion state of the UAV. Step A6: Extract the three-dimensional atmospheric wind field vector from the optimal state estimate, call the flow field distortion compensation network in the lightweight neural network inference engine to obtain the flow field distortion compensation amount, superimpose the flow field distortion compensation amount onto the three-dimensional atmospheric wind field vector, and output the compensated real atmospheric wind field data; calculate the current gust intensity factor based on the rate of change and spatial gradient of the extracted three-dimensional atmospheric wind field vector and output it.

3. The system according to claim 1, characterized in that, The specific process by which the adaptive wind-resistant flight control module outputs aircraft control signals is as follows: Step B1: Construct a six-degree-of-freedom nonlinear dynamic model of the UAV that includes the effects of wind disturbance, and transform the real atmospheric wind field data into wind disturbance forces and wind disturbance moments acting on the UAV. Step B2: Design a nonlinear disturbance observer, construct the dynamic equation of the nonlinear disturbance observer to estimate the total disturbance, and obtain the total disturbance estimate; Step B3: Design a robust controller based on backstepping, using the tracking error between the preset flight mission route and the current actual motion state of the UAV as input, and the total disturbance estimate output by the nonlinear disturbance observer as the feedforward compensation term to calculate the aircraft control signal. Step B4: Based on the current gust intensity factor, dynamically adjust the observer gain of the nonlinear disturbance observer and the control parameters of the robust controller.

4. The system according to claim 3, characterized in that, The workflow of the system parameter adaptive tuning module is as follows: It receives real atmospheric wind field data, and combines it with reference benchmarks and UAV state data. It uses a Bayesian optimization-based offline parameter tuning algorithm to iteratively optimize the model parameters included in the fluid-structure interaction state estimator and the observer gain of the nonlinear disturbance observer, and then sends the optimized parameters back to the airborne flight control and computing unit.

5. An adaptive wind measurement method for unmanned aerial vehicles (UAVs) applied to the system described in any one of claims 1 to 4, characterized in that, The process includes the following: The airborne sensing and measurement unit collects in real time the aircraft's inertial data, positioning and ground speed data, raw relative wind and atmospheric state data, and gust disturbance precursor signals; The wind field inversion and compensation module obtains the compensated real atmospheric wind field data and the current gust intensity factor by combining the flow field distortion compensation amount output by the lightweight neural network inference engine with the fluid-structure interaction state estimator based on volumetric Kalman filtering. The adaptive wind-resistant flight control module estimates wind disturbance force and wind disturbance moment based on real atmospheric wind field data and current gust intensity factor through a nonlinear disturbance observer, and generates aircraft control signals to control the UAV's flight. The health management and data fusion module receives inertial data of the body, positioning and ground speed data, raw relative wind and atmospheric state data, and gust disturbance precursor signals. It synchronizes and fuses these data in time and generates downlink data packets to send to the ground station. The ground station monitoring and calibration unit receives downlink data packets, combines them with the reference benchmark obtained from the high-precision reference meteorological station interface, updates the neural network through a multi-granularity automatic pruning framework, and optimizes the model parameters contained in the fluid-structure interaction state estimator and the observer gain of the nonlinear disturbance observer through the system parameter adaptive tuning module. The optimized parameters are then sent back to the airborne flight control and computing unit.

Images