A wind field perception and anti-wind control device and method for cross-domain unmanned aerial vehicles

Abstract

The application relates to a wind field perception and wind resistance control device and method for cross-domain unmanned planes, belongs to the field of air-ground cooperation and land-air unmanned plane technology, and solves the problem of poor wind resistance flight capability of unmanned planes in the prior art under the condition that the wind environment is complex and the wind speed changes drastically, and comprises the following steps: S1, collecting original data through a multi-modal sensor group, and processing the original data to obtain a standardized data set; S2, constructing a PINN model, processing the standardized data set as input to obtain a wind field prediction result; S3, training the PINN model to obtain a trained PINN model; S4, inputting the standardized data set obtained by real-time collection and processing into the trained PINN model to obtain a real-time output wind field prediction result; S5, establishing a self-disturbance rejection controller, outputting a final control quantity based on the real-time output wind field prediction result; and S6, processing the final control quantity to generate motor distribution instructions and providing the motor distribution instructions to motors of the unmanned plane.

Description

Technical Field

[0001] The present invention relates to the field of land-to-air UAV technology in the field of air-ground collaboration, and in particular to a wind field sensing and wind resistance control device and method for cross-domain UAVs. Background Art

[0002] At present, with the rapid development and widespread application of drone technology, extensive and in-depth research has been carried out at home and abroad on the stable flight and anti-interference capabilities of drones under extreme weather conditions. Especially in the case of complex wind environment and drastic wind speed changes, improving the wind resistance of drones has become an important issue to ensure their reliable operation in key application fields such as military reconnaissance, urban rescue, environmental monitoring and logistics transportation. The research background is not only due to the ever-increasing operating environment requirements, but also accompanied by the rapid development of high-performance control technology and intelligent algorithms. Therefore, by introducing physical information neural network technology to realize real-time reconstruction and prediction of wind fields, it is possible to accurately capture wind speed, wind direction, turbulence intensity and its temporal and spatial distribution changes, providing dynamic and sufficient data support for wind resistance control strategies, thereby greatly improving the flight safety and operation accuracy of drones in variable wind fields. At the same time, it also provides support for multi-task collaborative operations in complex environments in the future.

[0003] The rapid development of unmanned ground and air systems requires drones to be more adaptable to their environments. However, in actual operation, these amphibious unmanned systems often face complex and variable wind conditions, such as gusts and turbulence. These extreme winds can cause the drone and unmanned vehicle platforms to lose control or even crash. Therefore, developing land and air drones with high wind resistance is crucial.

[0004] Currently, the following technical issues exist regarding the wind resistance of land and air drones: First, traditional methods for sensing wind speed estimate wind speed based on navigation data. The drone compares the ground speed measured by GPS with the airspeed (i.e., speed relative to the air) calculated by the flight control system, and the difference between the two is used to estimate wind speed. During low-speed flight or hovering, this method's estimation accuracy is limited by the accuracy of the IMU and control model. This method cannot provide more accurate wind speed information during takeoff and landing, thus affecting the drone's wind resistance. Second, the wind resistance controller has a response lag. Traditional PID or linear controllers struggle to adapt to the nonlinear disturbances of sudden wind changes, resulting in attitude adjustment delays and trajectory deviations. Third, due to its strong model dependency, offline wind field modeling based on CFD cannot adapt to environmental changes in real time and consumes a lot of computing resources.

[0005] A Chinese invention patent application, publication number CN112486204A, titled "A UAV Wind Resistance Control Method, Device, Equipment, and UAV," discloses using a linear controller to address UAV flight range errors. However, this method struggles to accurately and real-timely sense and adjust the attitude and trajectory of a UAV in wind fields subject to nonlinear interference.

[0006] Therefore, there is a need in this field for an improved land-air drone that can improve the wind resistance of the drone in complex wind environments through intelligent control and high-precision wind field prediction. Summary of the Invention

[0007] In view of the above problems, the present invention provides a wind field perception and wind resistance control device and method for cross-domain UAVs, which solves the following problems in the prior art: high-precision perception and prediction of the flight flow field of land-to-air UAVs is performed through a physical information neural network method, more accurate flow field information is constructed, and more accurate feedforward information is provided to the controller, thereby improving the wind resistance performance of land-to-air UAVs under the joint action of an anti-disturbance controller.

[0008] According to one embodiment of the present invention, a wind field sensing and wind resistance control method for a cross-domain UAV is provided, which is used to provide wind field sensing and wind resistance control for land-air UAVs including UAVs and unmanned vehicles, including the following steps:

[0009] Step S1, collecting the spatiotemporal coordinates, wind speed components, air pressure gradient, and turbulence intensity of the UAV as raw data through a multimodal sensor group, and processing them to obtain a standardized data set;

[0010] Step S2, constructing a PINN model for processing the standardized data set as input to obtain wind farm prediction results;

[0011] Step S3, training the PINN model to obtain a trained PINN model;

[0012] In step S4, the standardized data set collected and processed in real time is input into the trained PINN model to obtain real-time wind field prediction results, including the three-dimensional velocity field vector of the wind field, the equivalent wind disturbance distance, the dynamic torque coefficient of the UAV, and the equivalent drag center offset;

[0013] Step S5: Establish an active disturbance rejection controller and construct a complete wind resistance control loop to receive the real-time wind field prediction results and combine them with the real-time collected UAV attitude angle error and attitude angular velocity error to obtain the final control value;

[0014] In step S6, the final control amount is processed to generate a motor allocation instruction and provided to the motor of the UAV, thereby achieving wind resistance control of the UAV or the UAV and the unmanned vehicle as a whole.

[0015] Optionally, step S2 specifically includes:

[0016] Step S2.1: Establish an 8-layer fully connected structure network to process the standardized data set and output the three-dimensional velocity field vector of the wind field;

[0017] Step S2.2, based on the three-dimensional velocity field vector of the wind field, the wind disturbance equivalent distance, dynamic moment coefficient, turbulence intensity and drag center offset are calculated;

[0018] Step S2.3: construct the compressible Navier-Stokes equations and embed them into an 8-layer fully connected structure network as hard constraints to couple the compressible Navier-Stokes equations to the PINN model as physical constraints.

[0019] Step S2.4: Establish data loss term, boundary condition loss term, and PDE residual term, and construct a total loss function to optimize the PINN model based on the data collected in real time by the multimodal sensor group;

[0020] In step S2.5, the compressible Navier-Stokes equation residual is calculated by automatic differentiation, a mean square error loss function is constructed, and the PINN model is optimized in combination with the data collected in real time by the multimodal sensor group.

[0021] Optionally, in step S2.3: constructing the compressible Navier-Stokes equations, including establishing the continuity equation, momentum equation and energy equation.

[0022] Optionally, the mean square error loss function constructed in step S2.5 is:

[0023]

[0024] in, is the mean square error between the predicted value and the measured value of the PINN model, is the weight coefficient, Indicates the output of the PINN model The predicted wind speed vector of the sampling points is Indicates the The true wind speed vector of the sampling points, Indicates the The coordinates of the drone’s position at each sampling point, Indicates the The timestamp of each sampling point, represents the residual function of the Navier-Stokes equations, is the total number of sampling points.

[0025] Optionally, step S5 specifically includes:

[0026] Step S5.1: Establish an ADRC controller. Based on the wind field prediction results output by the received PINN model as prior knowledge and the real-time measured attitude angle error and angular velocity error of the UAV, construct the state space equation and estimate the total disturbance of the UAV in real time.

[0027] Step S5.2: Construct a third-order linear extended state observer for the ADRC. Based on the obtained total disturbance of the UAV, the attitude angle and attitude angular velocity of the UAV, as well as the feedback control variable of the motor, are measured in real time to generate an updated control variable as the final control variable.

[0028] Step S5.3, using the hybrid particle swarm grey wolf optimization algorithm to tune the parameters of the ADRC, including the bandwidth of the ADRC, the bandwidth of the third-order linear extended state observer, and the control gain;

[0029] Step S5.4, setting the objective function of the ADRC and establishing a hybrid update strategy based on the hybrid particle swarm grey wolf optimization algorithm to optimize the ADRC.

[0030] Optionally, the final control variable includes a basic control variable, an angular velocity proportional control variable and an angular velocity error differential control variable.

[0031] According to another embodiment of the present invention, a wind field sensing and wind resistance control device for cross-domain UAVs is provided, which is used to provide wind field sensing and wind resistance control for land-air UAVs including UAVs and unmanned vehicles, and is characterized by including: a power system, a multimodal sensor group, a command and control module, and a communication module;

[0032] The multimodal sensor group includes a GPS, magnetic compass, wind speed sensor, IMU, and barometer installed on the drone. It is used to measure and provide wind speed, wind direction, air pressure gradient, and turbulence intensity as raw data, as well as real-time feedback of the drone's position, attitude angular velocity, attitude angular acceleration, and attitude angle error.

[0033] The power system includes propellers, motors, electronic speed regulators, and batteries mounted to the drone;

[0034] The command and control module includes a remote controller, a receiver, an autopilot, a data transmission radio, and an onboard computer installed on the drone. The onboard computer is equipped with a PINN model, and the autopilot is equipped with an active disturbance rejection controller.

[0035] The communication module is used to provide wind field perception of cross-domain UAVs and communication between various parts of the wind resistance control equipment.

[0036] Optionally, the PINN model in the command and control module receives the raw data collected by the multimodal sensor and generates a wind field prediction result, including the three-dimensional velocity field vector of the wind field, the equivalent distance of wind disturbance, the dynamic torque coefficient of the UAV and the equivalent drag center offset; the self-disturbance rejection controller in the command and control module receives the wind field prediction result, as well as the real-time feedback of the UAV's position, attitude angular velocity, attitude angular acceleration and attitude angular error from the multimodal sensor group, generates the final control quantity, and outputs it to the motor of the power system.

[0037] Compared with the prior art, the wind field perception and wind resistance control device and method for cross-domain UAVs provided in accordance with the embodiments of the present invention have at least the following beneficial effects.

[0038] 1) In terms of flow field reconstruction, the physical information neural network technology is used to integrate the traditional fluid mechanics control equations with the self-learning ability of neural networks to establish a PINN model. The real-time sensor data and environmental parameters are input into the PINN model. The wind field is reconstructed and predicted online through boundary conditions, initial conditions, and physical constraints such as fluid continuity and momentum conservation. The output results not only include spatial wind speed distribution, wind direction changes, and turbulence characteristics, but are also presented in the form of high-dimensional data vectors and tensors, providing accurate and real-time updated data support for subsequent wind resistance control.

[0039] 2) In terms of controller design, after obtaining real-time wind field information from the PINN model output, the command and control module quickly and accurately adjusts the posture of each degree of freedom of the UAV by integrating advanced control methods such as active disturbance rejection control. In a complex environment with multiple wind speeds and uncertain disturbances, by adjusting the control amount of the motor in real time, it provides thrust distribution, rudder deflection angle and multi-rotor pitch to eliminate the dynamic influence caused by the wind field, thereby achieving stable flight and precise positioning of the UAV. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. By referring to the drawings, the features and advantages of the present invention can be more clearly understood. The drawings are schematic and should not be understood as limiting the present invention in any way. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0041] Figure 1 2 is a schematic diagram of a wind field sensing and wind resistance control device for a cross-domain UAV provided according to a first embodiment of the present invention.

[0042] Figure 2 This is a side view of a wind field sensing and wind resistance control device for a cross-domain UAV provided according to a first embodiment of the present invention.

[0043] Figure 3 This is a flow chart of a wind field perception and wind resistance control method for a cross-domain UAV provided according to the second embodiment of the present invention.

[0044] Explanation of the reference numerals: 1-UAV, 2-Unmanned vehicle, 3-IMU, 4-Wind speed sensor. DETAILED DESCRIPTION

[0045] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present invention and the features therein can be combined with each other without conflict.

[0046] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0047] The following describes in detail a wind field perception and wind resistance control device and method for a cross-domain UAV provided in accordance with an embodiment of the present invention with reference to the accompanying drawings.

[0048] like Figure 1 and Figure 2 As shown, according to the first embodiment of the present invention, a wind field sensing and wind resistance control device for a cross-domain UAV is provided, which is used to provide wind field sensing and wind resistance control for land-air UAVs. The overall structure of the land-air UAV adopts an integrated design, integrating the multi-modal functions of the UAV in air flight and on land.

[0049] The first embodiment provides a wind field sensing and wind resistance control device for a cross-domain UAV, which is installed in a land-air UAV and includes key subsystems such as a power system, a multimodal sensor group, a command and control module, and a communication module.

[0050] The land-air drone comprises a detachably assembled drone 1 and an unmanned vehicle 2. Drone 1 comprises a frame (fuselage and landing gear). The power system, command and control module, and multimodal sensor array can be installed in the fuselage of drone 1. The power system and command and control module can be mounted on the frame using various connection methods, including threaded connections and adhesive fixation.

[0051] Drone 1 and vehicle 2 are connected by a square-shaped landing gear on the bottom of drone 1, which is combined with a locking device on the top of vehicle 2 to secure drone 1 and vehicle 2 together. Drone 1's body structure is made of lightweight, high-strength carbon fiber material, which not only meets weight requirements but also ensures sufficient mechanical rigidity and impact resistance.

[0052] In terms of wind field perception and wind resistance control, the land-air unmanned vehicle of the embodiment is different from the conventional unmanned vehicle in that: when the unmanned vehicle and the unmanned vehicle are separated, the unmanned vehicle can fly autonomously; after perceiving the wind field information, the unmanned vehicle can be landed on the unmanned vehicle, and the unmanned vehicle can be carried by the unmanned vehicle to pass through the adverse wind field, avoid the interference of the non-steady wind field on the unmanned vehicle, and save energy; according to the need, when encountering a land obstacle, the unmanned vehicle can be locked to the unmanned vehicle to be carried by the unmanned vehicle to fly over the obstacle.

[0053] The multi-modal sensor group of the embodiment can include a GPS, a magnetic compass, a wind speed sensor, a barometer, an IMU, etc. Specifically, the multi-modal sensor group can include a wind speed sensor (anemometer) 4 installed on the top of the unmanned vehicle 1 and a barometer and an IMU (a nine-axis inertial measurement unit) 3 installed inside the unmanned vehicle 1, for measuring and providing raw data such as wind speed, wind direction, air pressure gradient, turbulence intensity, etc., fusing the data, and feeding back information such as pose, attitude angular velocity, attitude angular acceleration, and attitude angular error of the unmanned vehicle 1 in real time.

[0054] The power system can include propellers, motors, electronic speed controllers, batteries, etc. installed to the unmanned vehicle 1. Specifically, it can include a lithium battery module, a flight controller, an electronic speed controller, a data transmission module, a laser radar, etc. installed in the middle of the fuselage of the unmanned vehicle 1. Optionally, a protection frame structure can be provided below the propeller disc of the unmanned vehicle 1.

[0055] The power system can also include a transmission system of the unmanned vehicle 2, which adopts a tracked drive system including wheels, tracks, and a motor drive system. The top of the unmanned vehicle 2 is provided with a device for docking and separating with the square frame landing gear of the unmanned vehicle 1, including a docking locking device and a slope servo mechanism. The docking locking device is a wheel disc transmission structure driven by a reduction brush motor, and the tooth disc slot is driven to lock the end horizontally and vertically.

[0056] The command and control module can include a remote controller and a receiver, an autopilot, a data transmission radio, an onboard computer, etc. To realize real-time detection and prediction of complex wind fields, the embodiment sets a PINN (Physical Information Neural Network) model in the onboard computer, combines the traditional fluid mechanics equations (including continuity equation, momentum equation, and energy conservation basic physical constraints) with the multi-source sensor data collected by the wind speed sensor 4 on the top of the unmanned vehicle 1 in real time, constructs a deep neural network model combining data driving and physical constraints, which can reconstruct the flow field in real time under the preset boundary conditions and initial conditions, accurately capture the wind speed, wind direction, turbulence intensity, and their spatiotemporal distribution changes, and output the real-time updated wind field prediction results in the form of high-dimensional data vectors and tensors, ensuring a comprehensive description of complex atmospheric disturbances.

[0057] The command and control module can also include an active disturbance rejection controller (ARDC) set in the autopilot. In terms of controller design, the ARDC utilizes advanced active disturbance rejection control and incremental feedback linearization methods, takes the real-time wind field prediction results (characteristic parameters) from the PINN model as the core input, and coordinates and controls the UAV's various degrees of freedom through a multi-level closed-loop feedback control algorithm. Furthermore, by adjusting parameters such as the thrust distribution and motor speed of each motor in real time, it achieves rapid compensation for dynamic wind field interference and precise attitude control, thereby significantly improving the UAV's anti-disturbance stability and control accuracy in complex and variable wind environments.

[0058] The working process of the wind field perception and wind resistance control device for cross-domain UAVs provided in this first embodiment is as follows: a multimodal sensor group (for example, a wind speed sensor) measures the local speed and wind speed and other data of the UAV, and transmits the data to the PINN model of the command and control module (for example, an onboard computer) via the communication module; the wind field prediction result is obtained after processing by the PINN model and provided to the active disturbance rejection controller; the active disturbance rejection controller processes the wind field prediction result to obtain the final control quantity (i.e., motor speed data), and transmits it to the power system through the communication module; in the power system, the electronic speed regulator then provides the motor speed to the motor, changes the motor speed, and thus changes the thrust of each propeller of the UAV, thereby realizing wind field perception and wind resistance control of the UAV.

[0059] See also Figure 3 According to a second embodiment of the present invention, a wind field perception and wind resistance control method for a cross-domain UAV is provided, which includes the following steps.

[0060] In step S1, the spatiotemporal coordinates, wind speed components, air pressure gradient, and turbulence intensity of the UAV are collected as raw data by a multimodal sensor group, and processed to obtain a standardized data set.

[0061] Specifically, the time and space coordinates of the drone 1 are collected by GPS and a multimodal sensor group (eg, wind speed sensor 4, barometer, etc.) provided on the drone 1. , wind speed component , air pressure gradient and turbulence intensity as raw data, and process them to obtain a standardized data set. The sampling frequency of the wind speed sensor 4 to collect the raw data of the flow field can be set to 100Hz. The time and space coordinates of the drone 1 For drone 1 The coordinates of the three-dimensional axis at a given moment. Wind speed components The IMU3 in the multimodal sensor group also measures the UAV’s position, attitude angular velocity, attitude angular acceleration, and attitude angular error in real time, which are used as inputs for the ADRC.

[0062] The original data is processed by Kalman filter noise reduction and time-space alignment to form the position coordinates of the drone ,time , wind speed component ,pressure and temperature The standardized dataset:

[0063]

[0064]

[0065] in, represents the sampling point index and , represents the total number of sampling points, Indicates the The standardized data of the sampling points, Indicates the The coordinates of the drone’s position at each sampling point, Indicates the The timestamp of each sampling point, Indicates the The wind speed component at each sampling point, Indicates the The pressure at each sampling point, Indicates the The temperature of the sampling point.

[0066] Step S2 constructs a PINN model. Using the standardized dataset as input, the model performs real-time reconstruction and prediction of the flow field around UAV 1, generating wind field prediction results. These wind field prediction results include the predicted three-dimensional velocity field vector (i.e., predicted flow field variables) and predicted flow field characteristic parameters. These parameters include the wind field's equivalent wind disturbance distance, turbulence intensity, the UAV's dynamic torque coefficient, and the equivalent drag center offset.

[0067] This implementation uses a PINN (Physical Information Neural Network) model to achieve real-time reconstruction and prediction of the flow field around the drone 1. By establishing a deep learning model that integrates physical laws, the accuracy and real-time performance of wind field prediction are significantly improved. This step S2 specifically includes the following steps.

[0068] Step S2.1, establish an 8-layer fully connected structure network, process the standardized data set and output the predicted flow field variables. Each layer of the fully connected structure network has 128 neurons, and the activation function is selected ,in is a learnable parameter and can be set to 1.0.

[0069] The input layer of the 8-layer fully connected structure network receives the spatiotemporal coordinates of the standardized dataset , the output layer outputs the predicted flow field variables The 8-layer fully connected network function is expressed as:

[0070]

[0071] in, represents the network parameters of the fully connected structure network, represents the neural network mapping function, Represents the three-dimensional velocity field vector of the wind field output by the fully connected structure network, that is, the wind field velocity field output by the PINN model. The three-dimensional velocity field vector of the wind field output by the PINN model , which may include predicting flow field variables , which can include predicted wind speed components, pressure and temperature.

[0072] Step S2.2 calculates the flow field characteristic parameters of the PINN model. This involves obtaining the wind disturbance equivalent distance, dynamic moment coefficient, turbulence intensity, and drag center offset based on the three-dimensional velocity field vector of the wind field. The specific process is as follows.

[0073] Establish the wind disturbance equivalent distance:

[0074]

[0075] in, is the integration domain and the unit is m 3 .

[0076] Establish the dynamic torque coefficient of the UAV body:

[0077]

[0078] in, is the surface area of ​​the drone, in m 2 ; is the center of mass position of the UAV, is the unit normal vector, Indicates the points measured in real time The air pressure at the

[0079] Establish turbulence intensity:

[0080]

[0081] in, is the turbulent kinetic energy of the flow field (unit: m 2 / s 2 ), The average wind speed measured in real time (unit: m / s).

[0082] Establish the drag center offset of the flow field:

[0083]

[0084] in, is the center of drag offset and is measured in meters.

[0085] In step S2.3, the compressible Navier-Stokes equations, including the continuity equation, momentum equation, and energy equation, are constructed and embedded as hard constraints in an 8-layer fully connected structure network, thereby coupling the compressible Navier-Stokes equations to the PINN model as physical constraints.

[0086] Formulating the compressible Navier-Stokes equations involves:

[0087] Establish the continuity equation: ;

[0088] Establish the momentum equation: ;

[0089] Establish the energy equation: ;

[0090] in, is the air flow field density, is the dynamic viscosity of the air flow field, is the specific heat capacity of the air flow field, is the thermal conductivity of the air flow field, is the viscous dissipation term of the air flow field, is the acceleration due to gravity, To find the divergence of a vector, is the fluid pressure.

[0091] Step S2.4: Establish the data loss term, boundary condition loss term, and PDE residual term, and construct the total loss function. The specific execution process of this step is as follows.

[0092] First, construct the data loss term:

[0093]

[0094] in, is the number of sampling data points, Indicates the first The three-dimensional velocity field vector of the wind field of the sampling data point is the predicted wind speed component, Indicates the The wind speed components of the sampled data points are from a standardized dataset.

[0095] Next, construct the PDE (partial differential equation solution) residual term:

[0096]

[0097] in, is the number of configuration points, represents the collocation point index and , Indicates the first The three-dimensional velocity field vector of the wind field at each configuration point, represents the continuity equation residual, represents the residual of the momentum equation, represents the residual of the energy equation.

[0098] Boundary condition loss term:

[0099]

[0100] in, is the number of boundary sampling points, represents the boundary sampling point index and , Indicates the first The three-dimensional velocity field vector of the wind field at each boundary sampling point is the predicted wind speed component. Indicates the The wind speed components at the boundary sampling points are from the standardized dataset.

[0101] Establish the total loss function and get the total loss:

[0102]

[0103] in, 、 and is the weight coefficient. In this embodiment, the weight coefficient can be =1.0, =0.5, =0.8.

[0104] In step S2.5, for the PINN model with coupled compressible Navier-Stokes equations as physical constraints, the residuals of the equations are calculated by automatic differentiation and the mean square error loss function is constructed:

[0105]

[0106] in, is the mean square error between the PINN model prediction value and the measured value, and the right side of the equation is the residual constraint of the Navier-Stokes equation. is the weight coefficient, Indicates the output of the PINN model The predicted wind speed vector of each sampling point, in meters per second (m / s), comes from the three-dimensional velocity field vector of the wind field predicted by the PINN model. Indicates the The true wind speed vector of the sampling point is the wind speed component measured by the sensor, in meters per second (m / s). Indicates the The coordinates of the drone’s position at each sampling point, Indicates the The timestamp of each sampling point, Represents the residual function of the Navier-Stokes equation, which represents the residual of the physical equation corresponding to the predicted wind field velocity field at the sampled UAV position and time. The actual wind speed vector of the sampling point The first The wind speed components at each sampling point . No. The coordinates of the drone's position at each sampling point , and The timestamp of the sampling point From the standardized dataset.

[0107] The PINN model is optimized based on the total loss function and the mean square error loss function, and the loss is minimized by combining the data collected in real time by the multimodal sensor group (a standardized dataset can be used).

[0108] Thus, a PINN model is constructed. In this implementation, the physical conservation laws (continuity, momentum, and energy equations) are embedded as hard constraints within the PINN model's neural network through the aforementioned steps. Automatic differentiation is then used to accurately calculate the residuals of the partial differential equations (PDEs), resulting in a computational efficiency improvement of over 100 times compared to traditional CFD methods while maintaining physical plausibility. Experiments have shown that, for example, in wind conditions of force 8, the predicted wind speed error is less than 0.5 m / s, and the relative error of the pressure field is less than 3%, meeting the requirements for real-time wind control for drones.

[0109] Step S3: training the established PINN model to obtain a trained PINN model.

[0110] In one example, an adaptive learning rate Adam optimizer can be used, with an initial learning rate of 5e-4 and a 10% decay every 1000 iterations. The training process uses a curriculum learning strategy, first optimizing the data loss and then gradually increasing the PDE constraint weights. The batch size is 1024, and the training is repeated for 50,000 iterations.

[0111] Specifically, the training of the PINN model includes using the Monte Carlo Dropout method to evaluate the prediction confidence, maintaining a 20% dropout rate during testing, and For example, 50 forward propagations are used to calculate the mean of the predicted results. and variance :

[0112]

[0113]

[0114] Among them, the number of sampling times Can be set to 50 times, Indicates the The prediction result obtained by sampling is The prediction results obtained by subsampling the input to the PINN model.

[0115] This step maintains the Dropout activation (e.g., 20%) during the test phase and simulates model parameter perturbations through multiple forward propagations to estimate the predictive distribution.

[0116] If the average uncertainty (i.e., mean and variance) is observed to be stable over several consecutive rounds of training, the variation (i.e., mean and variance) is less than the set threshold of 1%, and the coverage of the true value (i.e., mean and variance) by the 95% prediction confidence interval meets the expected requirement (i.e., exceeds 90%), the model is considered to have converged and training is complete, resulting in a trained PINN model. Otherwise, it indicates that the PINN model still has high uncertainty areas or insufficient prediction credibility, and further training is required to improve model robustness and prediction confidence.

[0117] Step S4: Input the wind field data sensed in real time by the multi-source sensor group carried by the UAV into the trained PINN model, and output the predicted three-dimensional velocity field vector of the wind field in real time, including the spatiotemporal evolution sequence of the wind field velocity field, pressure field, vorticity field and gradient field within a future time period (for example, 1 second); and generate the predicted flow field characteristic parameters based on the predicted three-dimensional velocity field vector, including the wind disturbance equivalent distance , turbulence intensity , dynamic moment coefficient and equivalent resistance center offset , as the real-time output of wind field prediction results.

[0118] In step S5, an active disturbance rejection controller (ADRC) is established. This controller receives the wind field prediction results generated by the PINN model and combines them with the real-time collected UAV attitude angle error and attitude angular velocity error to process the final control variable, namely the motor speed data. The ADRC can be installed in the autopilot.

[0119] Specifically, the wind field prediction result is used as the feedforward control quantity and input into the active disturbance rejection controller, and the active disturbance rejection controller is based on the wind disturbance equivalent distance in the wind field prediction result provided by the PINN model. The state space equation is constructed with the UAV attitude angle error measured in real time by the IMU, and the total disturbance of the UAV is estimated in real time.

[0120] The wind farm prediction results output by the PINN model in real time are fed into the active disturbance rejection controller as feedforward control variables to build a complete wind control loop. The active disturbance rejection controller can include a third-order linear extended state observer.

[0121] Step S5.1, establish an active disturbance rejection controller for receiving wind field prediction results output by the PINN model, including wind disturbance equivalent distance (Unit: m), dynamic moment coefficient (Unit: N·m) and the center of resistance offset (Unit: m) as prior knowledge; and the attitude angle error obtained by receiving IMU real-time measurement (unit: rad) and angular velocity error (Unit: rad / s), where Indicates the true value of the attitude angle, represents the expected value of the attitude angle, represents the true value of angular velocity, Express the expected value of angular velocity, construct the state space equation, and estimate the total disturbance of the UAV in real time.

[0122] Step S5.2: Build a dynamic model based on the output and sensor feedback of the PINN model and the estimated total disturbance, and construct a third-order linear extended state observer (LESO):

[0123]

[0124]

[0125]

[0126] in, is the attitude angle of the UAV, is the attitude angular velocity of the UAV (unit: rad, rad / s), which is measured in real time by the multimodal sensor group; is the total disturbance estimate (unit: rad / s²), obtained from step S5.1 above; , , is the gain parameter of the third-order linear extended state observer (units are: s , s , s ); is the control gain (dimensionless); Indicates the controlled output; represents the derivative of the attitude angle of the UAV, Indicates the attitude angular velocity of the drone, represents the derivative of the total disturbance, Represents the feedback control quantity for the motor. The feedback control quantity for the motor is the feedback input of the final control quantity generated by the active disturbance rejection controller at the previous moment.

[0127] The final control variable generated by the ADRC is composed of three parts: basic control variable, angular velocity proportional control variable and angular velocity error differential control variable.

[0128] The final control quantity can be expressed as:

[0129]

[0130] in, is the basic control quantity (unit: rad / s²); is the disturbance compensation term (unit: rad / s²); is the fuzzy thrust correction (unit: N), is the proportional coefficient adjustment (dimensionless); is the differential coefficient adjustment (dimensionless). is the angular velocity proportional control quantity; is the angular velocity error differential control quantity.

[0131] Based on the total disturbance of the drone, the multimodal sensor array measures the drone's attitude angle and angular velocity in real time, as well as the feedback control variables for the motors. This generates updated control variables as the final control variables. The final control variables include the basic control variable, the angular velocity proportional control variable, and the angular velocity error differential control variable.

[0132] In step S5.3, the hybrid particle swarm optimization algorithm (PSO-Grey Wolf Optimization) is used to tune the key parameters of the ADRC:

[0133]

[0134] in, represents the bandwidth of the ADRC (unit: rad / s); represents the bandwidth of the third-order linear extended state observer (unit: rad / s); represents the control gain (dimensionless).

[0135] Step S5.4, set the objective function of the ADRC:

[0136]

[0137] in, represents the loss function, and is the weight coefficient, Represents the feedback control quantity of the motor, Indicates the controlled output. As an example, you can set is 0.7, and set is 0.3.

[0138] The hybrid update strategy based on the hybrid particle swarm grey wolf optimization algorithm is established as follows:

[0139]

[0140] in, is the inertia weight; and is the learning factor; 、 and is the gray wolf optimization coefficient, Indicates the Particle No. The updated speed, represents the particle number, represents the iteration step number, and Represents a random number (0~1), used to enhance exploration randomness, Indicates the The historical optimal position of a particle, Indicates the Particle No. The position of the step, represents the global optimal position of the group, represents the current optimal solution, represents a suboptimal solution, Indicates the third best solution. Inertia weight Used to balance global and local search.

[0141] As an example, the inertia weight can be set, for example is 0.5, and sets the learning factor and Both are 1.5.

[0142] Through the above steps, the hybrid particle swarm-grey wolf optimization algorithm is used to dynamically adjust the controller parameters and control gains. During the optimization process, the particles are optimized according to the individual historical optimal, the group optimal and the gray wolf optimization. 、 、 The wolf's guidance direction updates its position, and the inertia weight decreases with each iteration, achieving a smooth transition from global exploration to local optimization. Ultimately, the optimized Active Disturbance Rejection Controller (ARDC) generates precise control variables, which offset wind field disturbances by adjusting motor thrust distribution in real time, ensuring stable flight in complex wind environments.

[0143] Optionally, a dual-input and three-output Mamdani fuzzy controller can be designed to adjust the PD controller parameters of the ADRC. The inputs include: the angular velocity error measured in real time by the IMU (unit: rad / s), domain ; Angular velocity error change rate (unit: rad / s²), domain . And the design output variable: proportional coefficient adjustment (dimensionless) ; Differential coefficient adjustment (dimensionless) ;Thrust distribution correction (Unit: N). The defuzzification method uses the centroid method to obtain the accurate output value.

[0144] In the above step S5, a hybrid particle swarm grey wolf optimization algorithm is used to tune the bandwidth parameters and gain of the active disturbance rejection controller. The optimization goal is to minimize the attitude angle error and energy consumption. The angular velocity error and the angular velocity error change rate are input, and a Mamdani fuzzy controller is used to adjust the PD controller parameters online to optimize the final control quantity.

[0145] In step S6, the ADRC processes the final control variable to generate motor allocation instructions, which are then provided to the UAV's motors to control the propeller speeds of the UAV and provide wind resistance control for the UAV or the UAV and UAV as a whole. Specifically, the ADRC processes the final control variable to generate motor allocation instructions, which are then provided to the power system. In the power system, an electronic tachometer provides the motor speeds to each motor, changing the motor speeds to change and control the thrust distribution of each motor of the UAV, the thrust of each propeller, the deflection angle of the rudder, the pitch of the multi-rotor blades, the propeller speed, and other factors. This achieves accurate wind field perception and precise wind resistance control for the UAV or the land-air UAV as a whole.

[0146] All of the above optional technical solutions can be combined in any way to form optional embodiments of the present application, and will not be described in detail here.

[0147] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0148] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A wind field sensing and wind resistance control method for cross-domain UAVs, which is used to provide wind field sensing and wind resistance control for land-air UAVs including UAVs and unmanned vehicles, characterized in that: The following steps are involved: Step S1, collecting the spatiotemporal coordinates, wind speed components, air pressure gradient, and turbulence intensity of the UAV as raw data through a multimodal sensor group, and processing them to obtain a standardized data set; Step S2, constructing a PINN model for processing the standardized data set as input to obtain wind farm prediction results; Step S3, training the PINN model to obtain a trained PINN model; In step S4, the standardized data set collected and processed in real time is input into the trained PINN model to obtain real-time wind field prediction results, including the three-dimensional velocity field vector of the wind field, the equivalent wind disturbance distance, the dynamic torque coefficient of the UAV, and the equivalent drag center offset; Step S5: Establish an active disturbance rejection controller and construct a complete wind resistance control loop to receive the real-time wind field prediction results and combine them with the real-time collected UAV attitude angle error and attitude angular velocity error to obtain the final control value; Step S6: Processing the final control amount to generate a motor allocation instruction and providing it to the motor of the UAV to achieve wind resistance control of the UAV or the UAV and the unmanned vehicle as a whole; Wherein, step S2 specifically includes: Step S2.1: Establish an 8-layer fully connected structure network to process the standardized data set and output the three-dimensional velocity field vector of the wind field; Step S2.2, based on the three-dimensional velocity field vector of the wind field, the wind disturbance equivalent distance, dynamic moment coefficient, turbulence intensity and drag center offset are calculated; Step S2.3: construct the compressible Navier-Stokes equations and embed them into an 8-layer fully connected structure network as hard constraints to couple the compressible Navier-Stokes equations to the PINN model as physical constraints. Step S2.4: Establish data loss term, boundary condition loss term, and PDE residual term, and construct a total loss function to optimize the PINN model based on the data collected in real time by the multimodal sensor group; In step S2.5, the compressible Navier-Stokes equation residual is calculated by automatic differentiation, a mean square error loss function is constructed, and the PINN model is optimized in combination with the data collected in real time by the multimodal sensor group.

2. The wind field perception and wind resistance control method for cross-domain UAV according to claim 1 is characterized in that: In step S2.3: Construct the compressible Navier-Stokes equations, including establishing the continuity equation, momentum equation, and energy equation.

3. The wind field perception and wind resistance control method for cross-domain UAV according to claim 1 is characterized in that: The mean square error loss function constructed in step S2.5 is: in, is the mean square error between the predicted value and the measured value of the PINN model, is the weight coefficient, Indicates the output of the PINN model The predicted wind speed vector of the sampling points is Indicates the The true wind speed vector of the sampling points, Indicates the The coordinates of the drone’s position at each sampling point, Indicates the The timestamp of each sampling point, represents the residual function of the Navier-Stokes equations, is the total number of sampling points.

4. The wind field perception and wind resistance control method for cross-domain UAV according to claim 1 is characterized in that: Step S5 specifically includes: Step S5.1: Establish an ADRC controller. Based on the wind field prediction results output by the received PINN model as prior knowledge and the real-time measured attitude angle error and angular velocity error of the UAV, construct the state space equation and estimate the total disturbance of the UAV in real time. Step S5.2: Construct a third-order linear extended state observer for the ADRC. Based on the obtained total disturbance of the UAV, the attitude angle and attitude angular velocity of the UAV, as well as the feedback control variable of the motor, are measured in real time to generate an updated control variable as the final control variable. Step S5.3, using the hybrid particle swarm grey wolf optimization algorithm to tune the parameters of the ADRC, including the bandwidth of the ADRC, the bandwidth of the third-order linear extended state observer, and the control gain; Step S5.4, setting the objective function of the ADRC and establishing a hybrid update strategy based on the hybrid particle swarm grey wolf optimization algorithm to optimize the ADRC.

5. The wind field perception and wind resistance control method for cross-domain UAV according to claim 4 is characterized in that: The final control variable includes the basic control variable, the angular velocity proportional control variable and the angular velocity error differential control variable.

6. A wind field sensing and wind resistance control device for cross-domain UAVs, used to provide wind field sensing and wind resistance control for land and air UAVs including UAVs and unmanned vehicles, characterized by: include: power systems, multimodal sensor suites, command and control modules, and communications modules; The multimodal sensor group includes a GPS, magnetic compass, wind speed sensor, IMU, and barometer installed on the drone. It is used to measure and provide wind speed, wind direction, air pressure gradient, and turbulence intensity as raw data, as well as real-time feedback of the drone's position, attitude angular velocity, attitude angular acceleration, and attitude angle error. The power system includes propellers, motors, electronic speed regulators, and batteries mounted to the drone; The command and control module includes a remote controller, a receiver, an autopilot, a data transmission radio, and an onboard computer installed on the drone. The onboard computer is equipped with a PINN model, and the autopilot is equipped with an active disturbance rejection controller. The communication module is used to provide wind field perception for cross-domain UAVs and communication between various parts of the wind resistance control equipment; Among them, setting up the PINN model in the onboard computer includes: An 8-layer fully connected network is established to process the standardized data set and output the three-dimensional velocity field vector of the wind field; Based on the three-dimensional velocity field vector of the wind field, the wind disturbance equivalent distance, dynamic moment coefficient, turbulence intensity and drag center offset are calculated; Construct the compressible Navier-Stokes equations and embed them into an 8-layer fully connected structure network as hard constraints to couple the compressible Navier-Stokes equations to the PINN model as physical constraints. Establish data loss term, boundary condition loss term and PDE residual term, and construct the total loss function, and optimize the PINN model based on the data collected in real time by the multimodal sensor group; The compressible Navier-Stokes equation residuals are calculated by automatic differentiation, and the mean square error loss function is constructed. The PINN model is optimized by combining the data collected in real time by the multimodal sensor group.

7. The wind field sensing and wind resistance control device for a cross-domain UAV according to claim 6 is characterized in that: The PINN model in the command and control module receives the raw data collected by the multimodal sensors and generates wind field prediction results, including the three-dimensional velocity field vector of the wind field, the equivalent distance of wind disturbance, the dynamic torque coefficient of the UAV, and the equivalent center of drag offset; The active disturbance rejection controller in the command and control module receives the wind field prediction results and the real-time feedback of the UAV's position, attitude angular velocity, attitude angular acceleration and attitude angular error from the multimodal sensor group, generates the final control quantity, and outputs it to the motor of the power system.

Images