Unmanned aerial vehicle wind-resistant adaptive control method and system based on simulation driving and feature constraint

Abstract

This invention discloses a wind-resistant adaptive control method and system for unmanned aerial vehicles (UAVs) based on simulation data-driven and feature-constrained meta-learning. It primarily addresses the problems of existing adaptive control methods over-reliance on real flight data, leading to high data acquisition costs, difficulty in spatiotemporal alignment of high-dynamic wind field data, and limited generalization. The implementation scheme involves: utilizing a high-fidelity simulation environment to collect physically aligned flight data under multiple wind fields; constructing a neural network feature extraction model with spectral normalization; training the model offline using triplet loss to extract a general dynamic representation of wind invariance, which is then deployed to the airborne controller; during flight, inputting the real-time state into the model, and dynamically updating the wind field coefficients online using a composite adaptive law to generate a high-frequency feedforward compensation signal to counteract aerodynamic disturbances. This invention achieves robust decoupling between inherent characteristics and external disturbances, enabling the UAV to maintain high-precision trajectory tracking and stable response in unknown dynamic high-speed wind fields, effectively resisting complex wind disturbances.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) flight control technology, and in particular relates to a wind-resistant adaptive control method and system for UAVs. It can be used for logistics transportation, disaster relief and infrastructure inspection, resisting nonlinear aerodynamic disturbances suffered by UAVs in complex, dynamic, time-varying high-speed wind fields, and ensuring the safety of maneuvering flight and high-precision trajectory tracking. Background Technology

[0002] With the rapid development of drone technology, flight stability under dynamic and time-varying wind fields has become an increasingly significant technical bottleneck for the industry. Although existing research focuses on improving the wind resistance of drones by refining control algorithms, achieving low-cost, high-precision, and universal deployment still faces severe challenges.

[0003] Document "Accurate Tracking of Aggressive Quadrotor Trajectories Using Incremental Nonlinear Dynamic The paper "Inversion and Differential Flatness, IEEE Transactions on Control Systems Technology, 2021" proposes a UAV control framework based on incremental nonlinear dynamic inverse INDI and differential flatness techniques. This framework uses high-frequency estimation of angular acceleration to counteract external nonlinear aerodynamic disturbances. However, this method remains heavily reliant on extremely high control frequencies, such as 500Hz, and motor speed feedback provided by specific optical encoders. Therefore, in practical applications, the extremely demanding dedicated hardware and onboard computing power requirements limit it to specific high-end platforms. Even with high tracking accuracy, it still faces significant limitations in large-scale, low-cost, and general-purpose deployment on standard commercial UAV platforms.

[0004] The paper "Neural-fly enables rapid learning for agile flight in strong winds, Science Robotics, 2022" discloses an adaptive control method based on meta-learning. It utilizes Domain Adversarial Invariant Meta-Learning (DAIML) to decompose aerodynamic forces into wind-invariant basis functions and wind-dependent linear coefficients, achieving agile flight in strong winds. However, this method has two main drawbacks: First, its offline representation learning heavily relies on real-world wind tunnel systems and real UAV flight log data, resulting in extremely high economic costs and experimental risks. For example, collecting extreme maneuver data under extreme wind conditions can easily lead to UAV loss of control and crashes. Second, this method only collects data in real-world physical environments. The wind fields in real-world environments are highly random and non-repeatable, making it impossible to perfectly reproduce rigorously controlled samples of different wind conditions at the same time and under the same conditions in the physical world. This spatiotemporal misalignment of data fundamentally limits the effective decoupling and generalization capabilities of deep networks for aerodynamic features and environmental disturbances.

[0005] Patent document CN117539284A discloses a wind-resistant attitude control method for unmanned aerial vehicles (UAVs) based on an improved L1 adaptive algorithm. The main idea is to equivalently divide the complex nonlinear real-world wind field into multiple simple linear subdomain models, constructing a response surface model to replace the traditional controlled object module in the L1 adaptive algorithm, thereby achieving wind-resistant control. However, this method has significant limitations: First, the static strategy of dividing the subdomain based on wind speed and direction highly relies on prior assumptions about the environment, making it difficult to exhaustively represent the complex and ever-changing strongly coupled, highly dynamic airflow disturbances in the real world. Second, this method lacks the ability to obtain spatiotemporally strictly aligned comparison samples using a high-fidelity simulation environment, and it does not actively extract wind field environmental information, resulting in it only being able to adapt local parameters within the already divided wind condition subdomains, making it difficult to truly achieve zero-risk migration to unknown high-speed dynamic wind field environments.

[0006] In summary, existing UAV anti-disturbance control technologies generally suffer from three major drawbacks: excessive reliance on dedicated hardware and high computing power, resulting in high deployment costs; high cost and risk of real data collection, and difficulty in cross-referencing samples; and reliance on prior environmental information and linear modeling, leading to weak generalization and transfer capabilities for unknown and complex dynamic disturbances. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of the existing technologies by proposing a simulation-driven and feature-constrained wind-resistant adaptive control method and system for unmanned aerial vehicles (UAVs). This avoids reliance on dedicated hardware, high computing power, and costly measured data, enhances the generalization and migration deployment capabilities of UAVs under unknown, highly dynamic, and nonlinear wind disturbances, and achieves universal, low-cost, and high-precision wind-resistant adaptive flight control.

[0008] To achieve the above objectives, the technical solution of the present invention includes:

[0009] 1. A wind-resistant adaptive control method for unmanned aerial vehicles (UAVs) based on simulation-driven and feature-constrained approaches, characterized by comprising:

[0010] (1) Using a high-fidelity simulation environment, flight state data of UAVs executing standardized maneuver trajectories under static atmosphere and multi-dimensional dynamic wind field interference are collected. Different wind fields and maneuver trajectory data are grouped and grouped according to the mission intention to obtain a triplet simulation feature dataset including anchor points, positive samples and negative samples.

[0011] (2) The unknown aerodynamic disturbance is decoupled into a general dynamic representation and a linear coefficient matrix using a deep learning model, and a fully connected neural network containing a spectral normalization layer in the model is selected as the basic feature extractor.

[0012] (3) Input the triplet simulation feature dataset into the basic feature extractor, perform offline feature loss meta-learning training on it to obtain an inherent feature extractor containing general dynamic representations independent of wind conditions, and fix its network weights.

[0013] (4) Deploy the aforementioned feature extractor with fixed network weights into the online controller of the UAV onboard computing platform;

[0014] (5) During the real-time flight of the UAV, the system state features including the current linear velocity, attitude quaternion and motor pulse width modulation (PWM) are acquired. After normalization preprocessing to eliminate data distribution offset, the data is input into the inherent feature extractor for forward propagation and the reduced basis function feature vector is output.

[0015] (6) Real-time acquisition of the current UAV expected trajectory and actual response to calculate the composite tracking error, and the error is introduced into the composite adaptive law with forgetting factor as the core feedback signal. The linear coefficient matrix is ​​updated online in real time, and the matrix and the eigenvector of the reduced basis function are solved and converted into wind-resistant feedforward compensation acceleration.

[0016] (7) The wind-resistant feedforward compensation acceleration is superimposed on the feedforward control quantity based on the desired trajectory reference acceleration to generate a comprehensive control command output to the UAV's underlying controller for execution.

[0017] Furthermore, the deep learning model in (2) includes an input layer, two hidden layers, and an output layer connected in sequence, wherein:

[0018] The input layer contains 11 neurons: 3 for linear velocity, 4 for attitude quaternions, and 4 for PWM motor outputs. These neurons are used to receive the UAV's state features. ;

[0019] Two hidden layers, each containing 64 neurons;

[0020] An eigenvector for outputting basis functions The output layer contains 8 neurons;

[0021] The linear transformation matrix of the hidden layer and the output layer Both introduce the concept of matrix maximum singular value. The scaling spectral normalization operation uses Update network weights and use a linear rectified function. As an activation function, it restricts the local Lipschitz continuity constant of the neural network, ensuring the stability of online control;

[0022] The input layer, two hidden layers, and output layer are sequentially connected to form a feedforward network topology. The inter-layer connections are as follows: the outputs of the 11 neurons in the input layer are fully connected to the inputs of the 64 neurons in the first hidden layer; the outputs of the neurons in the first hidden layer are fully connected to the inputs of the 64 neurons in the second hidden layer; and the outputs of the neurons in the second hidden layer are fully connected to the inputs of the 8 neurons in the output layer. This enables the layer-by-layer forward propagation and dimensionality reduction mapping of the UAV state feature signals to the basis function feature vectors.

[0023] Furthermore, the offline feature loss function training performed on the basic feature extractor in step (3) includes:

[0024] 3a) Calculate the optimal linear coefficients on the fitness set using the least squares method, and calculate the mean square error loss of the physical prediction reconstruction error based on the optimal linear coefficients;

[0025] 3b) Construct a triplet feature loss function to minimize the feature distance between the anchor state in a windless environment and the positive sample state in a windy environment under the same maneuver, and maximize the feature distance between the anchor state and the negative sample state under different maneuvers.

[0026] 3c) The gradient descent algorithm is used to jointly optimize the mean squared error loss and the triplet feature loss, update the network weights of the basic feature extractor, and force the neural network to strip away the wind field environmental features and extract the kinematic essential features with wind condition invariance.

[0027] 2. A wind-resistant adaptive control system for unmanned aerial vehicles (UAVs) based on simulation data-driven and feature-constrained meta-learning, characterized in that it comprises:

[0028] The simulation data collection and preprocessing module is used to collect multi-wind field simulation flight data, and after combining them into the same group or groups, construct and output a triplet simulation feature dataset.

[0029] The model offline training module is used to receive the triplet simulation feature dataset, perform offline training using a deep neural network, and output the inherent feature extractor weights after separating the wind field features.

[0030] The real-time status processing module is used to acquire the status characteristics of the UAV system in real time and perform scaling and normalization preprocessing on them.

[0031] The online feature inference module is used to receive the preprocessed system state features, call the onboard fixed weight feature extractor to perform forward propagation, and output the dimensionality-reduced basis function feature vector in real time.

[0032] The composite adaptive control module is used to receive the desired trajectory, the actual response and the basis function eigenvector, and after adaptively updating the linear coefficient matrix, calculate and output the wind-resistant feedforward compensation acceleration.

[0033] The integrated command generation module is used to receive the wind-resistant feedforward compensation acceleration, process it by amplitude limiting, generate integrated control commands, and send them to the UAV's underlying external control interface.

[0034] Compared with the prior art, the present invention has the following advantages:

[0035] Firstly, this invention introduces a triplet loss mechanism during the offline training phase, which can actively and explicitly impose geometric constraints on the feature space and forcibly strip away wind field environmental information based on spatiotemporally aligned simulation feature data, thereby eliminating feature confusion and error accumulation caused by unmodeled aerodynamic disturbances.

[0036] Secondly, since the present invention uses a composite adaptive law driven by both physical reconstruction error and composite trajectory tracking error, it ensures the accuracy and smoothness of high-frequency updates of the wind field linear coefficient matrix. Therefore, it will not cause transient instability or continuous trajectory deviation loss when facing unknown sudden gusts or strong turbulence.

[0037] Third, this invention extends the theoretical stability guarantee of neural networks to the precise constraint of spectral normalization of the full-layer linear transformation matrix, effectively constraining the local Lipschitz continuity constant of the network. This allows the controller to have strict closed-loop stability when facing unknown extreme wind conditions, which not only avoids system divergence and crash due to feature mutations, but also enables normal and stable high-precision maneuvering flight.

[0038] Fourth, since the present invention obtains the benchmark ground value of physical alignment under dynamic wind field through a high-fidelity simulation environment, it can carry out large-scale high dynamic limit feature training of deep learning models with zero risk, low cost and more targeted approach.

[0039] Fifth, by deploying the basis function feature extraction model that extracts and solidifies wind condition invariance into the UAV airborne controller, this invention enables the UAV to autonomously predict and feedforward aerodynamic disturbances in real time after encountering complex and unknown wind fields, ensuring the efficiency and robustness of the system when deployed on a standard commercial platform. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating the implementation of the UAV wind-resistant adaptive control method based on simulation-driven and feature-constrained principles of this invention.

[0041] Figure 2 This is a schematic diagram of the feature-constrained neural network constructed in the method of this invention;

[0042] Figure 3 This is a block diagram of a wind-resistant adaptive control system for unmanned aerial vehicles (UAVs) based on simulation-driven and feature-constrained architecture. Detailed Implementation

[0043] The embodiments of this method are described in further detail below with reference to the accompanying drawings.

[0044] Example 1: A wind-resistant adaptive control method for unmanned aerial vehicles based on simulation-driven and feature-constrained approaches.

[0045] This method is applicable to commercial or customized multi-rotor UAVs equipped with a basic flight control unit and inertial measurement sensors. During normal flight or in scenarios encountering complex external aerodynamic disturbances such as strong gusts and turbulence, the sensors acquire environmental data of the UAV. The UAV acquires this sensor data and performs internal calculations based on the desired trajectory and actual response to obtain comprehensive control commands including wind-resistant feedforward compensation, which are then sent to the actuators for execution.

[0046] Reference Figure 1 The implementation steps of this example include the following:

[0047] Step 1: Configure the Gazebo simulation wind field plugin and determine the standardized maneuver trajectory to construct a triplet dataset.

[0048] 1.1) During the flight of the UAV, by configuring the Gazebo simulation engine and custom wind field plugin, various deterministic multidimensional dynamic wind field environments are set, including constant wind, sinusoidal gusts and random turbulence;

[0049] 1.2) Determine the standardized maneuvering trajectories of the UAV, including hovering, figure-eight flight, and spiral ascent and descent;

[0050] 1.3) Control the UAV to execute the standardized maneuver trajectory under different wind field environments, and collect environmental data and internal status of the UAV during flight at high frequency;

[0051] 1.4) Extract state features and dynamic response data from flight log data, including three-dimensional linear velocity, attitude quaternion, measured values ​​of pulse width modulation signals of the four motors, and three-dimensional true acceleration.

[0052] 1.5) The extracted data is timestamped and synchronized, and the actual aerodynamic residual force is calculated by inverse solving of the rigid body dynamics equations;

[0053] 1.6) Select the flight state features of the UAV at a certain moment and under a certain wind field environment as anchor samples, select the features at the same time but under different wind field environments as positive samples, and select the features at different times or under different actions as negative samples.

[0054] 1.7) Combine the above anchor samples, positive samples and negative samples to form a triplet simulation feature dataset for network training.

[0055] Step 2: Build a feature-constrained neural network model.

[0056] This step aims to construct a fully connected neural network to reduce the dimensionality of the UAV's physical flight state to a basis function feature vector. To prevent the model from overfitting to dynamic wind field disturbances and causing control oscillations, a feature constraint mechanism is introduced. This mechanism limits the Lipschitz continuity constant by introducing a spectral normalization operation into the network's linear transformation matrix, and uses the ReLU activation function to filter negative disturbance signals, thereby limiting the network's amplification factor for input disturbances and ensuring that the final extracted dynamic representation features are smooth and bounded.

[0057] Reference Figure 2 This step involves constructing a fully connected neural network consisting of an input layer, two hidden layers, and an output layer connected in series, where:

[0058] The input layer includes 11 neurons, used to receive the UAV's three-dimensional linear velocity, four-dimensional attitude quaternion, and four-dimensional motor normalized pulse width modulation signal.

[0059] The two hidden layers have the same structure, each containing 64 neurons. The linear transformation matrix introduces a spectral normalization operation based on the scaling of the maximum singular value of the matrix, and uses the ReLU activation function to perform a nonlinear transformation on the initial calculation results of the neurons, so that the results less than 0 are mapped to 0, while the results greater than 0 remain unchanged.

[0060] The output layer consists of 8 neurons, and its linear transformation matrix also incorporates spectral normalization operations. It does not use the ReLU activation function and is used to output the dimensionality-reduced basis function feature vector.

[0061] The linear transformation matrix of the hidden layer and the output layer Both introduce the concept of matrix maximum singular value. The scaling spectral normalization operation uses Update network weights and use a linear rectified function. As an activation function, it restricts the local Lipschitz continuity constant of the neural network, ensuring the stability of online control;

[0062] The input layer, the two hidden layers, and the output layer are sequentially connected using a fully connected method to form a feedforward network topology. The inter-layer connections are as follows: the output terminals of the 11 neurons in the input layer are fully connected to the input terminals of the 64 neurons in the first hidden layer; the output terminals of the neurons in the first hidden layer are fully connected to the input terminals of the 64 neurons in the second hidden layer; and the output terminals of the neurons in the second hidden layer are fully connected to the input terminals of the 8 neurons in the output layer. This forms a feature-constrained neural network model to achieve layer-by-layer forward propagation and dimensionality reduction mapping from the UAV state feature signals to the basis function feature vectors.

[0063] Step 3: Train the feature-constrained neural network model.

[0064] 3.1) Divide the triplet simulation feature dataset into training and adaptation sets in batches, and input them into the neural network batch by batch for forward propagation;

[0065] 3.2) On the fitness set, with the aerodynamic residual force as the target value, the optimal linear coefficient matrix corresponding to the current characteristic state is calculated by the least squares method;

[0066] 3.3) On the training set, the mean squared error loss function is used to calculate the regression error loss between the aerodynamic reconstructed force predicted based on the optimal linear coefficient and the actual aerodynamic residual force;

[0067] 3.4) Construct the triplet feature loss function;

[0068] 3.4.1) Obtain the basis function feature vectors output by the feature-constrained neural network during forward propagation of anchor samples, positive samples, and negative samples, respectively;

[0069] 3.4.2) Calculate the Euclidean distance between the anchor sample feature vector and the positive sample feature vector as the first feature distance, and calculate the Euclidean distance between the anchor sample feature vector and the negative sample feature vector as the second feature distance;

[0070] 3.4.3) Introduce a preset boundary margin hyperparameter Margin, add the boundary margin hyperparameter to the first feature distance, and subtract it from the second feature distance. Take the maximum value of the difference and zero as the triplet feature loss function.

[0071] 3.5) Use the triplet feature loss function constructed above to minimize the distance between the anchor sample and the positive sample, and maximize the distance between the anchor sample and the negative sample;

[0072] 3.6) Use the gradient descent algorithm to jointly optimize the mean squared error loss and the triplet feature loss, calculate the gradient of the above loss with respect to each parameter in the network, and backpropagate to update the parameters of each layer in the neural network.

[0073] 3.7) Repeat steps 3.1) to 3.6) until the objective function of the feature-constrained neural network converges, and you will get the trained feature-constrained neural network model.

[0074] Step 4, Airborne Deployment of the Model.

[0075] 4.1) Freeze all network weights of the feature-constrained neural network model after training in step 3;

[0076] 4.2) Convert the model with frozen weights into a lightweight format suitable for airborne computing and deploy it to the online controller of the UAV's onboard companion computer as the base model for online forward inference.

[0077] Step 5: Real-time state assessment and online composite adaptive control of the UAV during flight.

[0078] 5.1) During the actual operation of the UAV, the sensor measurement values ​​and internal state variables of the UAV are acquired and recorded in real time, and the state characteristics are input into the feature constraint neural network model deployed on the airborne, and the basis function feature vector is output.

[0079] 5.2) Determine whether the drone is currently in flight mode, i.e., whether it has taken off and entered external control mode;

[0080] If the drone is in flight, then perform step 6 to perform online composite adaptive update and feedforward compensation output;

[0081] Otherwise, determine whether the drone has already landed;

[0082] If the drone has already landed, the parameter update will be terminated and the system operation will end.

[0083] If the drone has not landed, and is in the ground unlocking preparation stage, return to step 5.1) to continue monitoring of online composite adaptive control.

[0084] Step 6: Output wind field coefficient update and compensation commands during flight.

[0085] 6.1) Obtain the reference position and reference velocity of the UAV's desired trajectory at the current moment, and calculate the composite tracking error by subtracting them from the actual position and actual velocity; at the same time, obtain the actual acceleration and thrust, and calculate the observed aerodynamic residual force.

[0086] 6.2) Online composite adaptive update of wind field linear coefficients: Using the composite tracking error and residual prediction error obtained in step 6.1), the wind field linear coefficient matrix is ​​updated in real time and at high frequency through the covariance evolution equation with forgetting factor;

[0087] 6.3) Multiply the basis function eigenvectors output in step 5.1) with the real-time updated wind field linear coefficient matrix to calculate the predicted aerodynamic disturbance force, and combine it with the mass of the UAV to convert it into wind-resistant feedforward compensation acceleration.

[0088] 6.4) The wind-resistant feedforward compensation acceleration is superimposed on the reference acceleration based on the desired trajectory to generate a comprehensive feedforward acceleration and limit it with a maximum safety threshold. Then, it is sent to the UAV's underlying position controller for execution to counteract dynamic wind field disturbances.

[0089] 6.5) After completing the control command output for the current cycle, return to step 5.1) to enter the next control cycle.

[0090] The flowchart representations or method representations of the above embodiments can be understood as representing code, fragments, or portions of executable instructions comprising one or more steps configured to implement a specific logical function or process. This invention is not limited to the disclosed preferred embodiments, and its implementation may not follow the order shown or discussed. That is, the step numbers in the specification and claims are only for clear description and understanding of the embodiments of the invention, and their order is not limited.

[0091] Example 2: A wind-resistant adaptive control system for unmanned aerial vehicles based on simulation-driven and feature-constrained design.

[0092] Reference Figure 3This example includes a simulation data collection and preprocessing module 1, an offline model training module 2, a real-time state processing module 3, an online feature inference module 4, a composite adaptive control module 5, a comprehensive command generation module 6, and an unmanned aerial vehicle (UAV) airframe 7. The composite adaptive control module 5 includes a composite tracking error submodule 51, a feedback update construction submodule 52, an adaptive law update submodule 53, and a compensation command generation submodule 54.

[0093] The working principle of the entire system is as follows:

[0094] The simulation data collection and preprocessing module 1 is used to collect multi-wind field simulation flight data in a high-fidelity simulation environment, construct a triplet simulation feature dataset after grouping or grouping, and output the dataset to the model offline training module 2.

[0095] The offline training module 2 is used to receive the triplet simulation feature dataset from the simulation data collection and preprocessing module 1, perform offline meta-learning training using a deep neural network, and output and solidify the pre-trained inherent feature extractor weights after stripping the wind field features, so as to provide a model basis for airborne online inference.

[0096] The real-time state processing module 3 is used to acquire the state features of the UAV system during the real-time flight of the UAV, and after scaling and normalization preprocessing operations, output the state features at a uniform scale and transmit them to the online feature reasoning module 4.

[0097] The online feature reasoning module 4 receives the preprocessed system state features from the real-time state processing module 3, calls the onboard fixed weight feature extractor to perform forward propagation, outputs the dimensionality-reduced basis function feature vector in real time, and transmits the feature vector to the composite adaptive control module 5.

[0098] The composite adaptive control module 5 is used to receive the desired trajectory, actual response, and the basis function eigenvector. After adaptively updating the linear coefficient matrix, it calculates the wind-resistant feedforward compensation acceleration. Specifically, the composite tracking error submodule 51 receives the UAV's desired trajectory and actual response data, calculates the composite tracking error through comparison, and outputs it to the feedback update construction submodule 52. This feedback update construction submodule 52 receives the composite tracking error, constructs a first feedback update term based on the physical prediction error and a second feedback update term based on the trajectory yaw, and outputs these two update terms to the self-adaptive control module 52. The adaptive law update submodule 53 is used to receive the feedback update term from the feedback update construction submodule 52, dynamically adjust the gain through the matrix evolution equation, update the linear coefficient matrix in real time, and output the updated linear coefficient matrix to the compensation instruction generation submodule 54. The compensation instruction generation submodule 54 is used to receive the updated linear coefficient matrix from the adaptive law update submodule 53 and the basis function eigenvector from the online feature inference module 4, combine the two and solve them to generate the wind-resistant feedforward compensation acceleration, and output it to the comprehensive instruction generation module 6.

[0099] The integrated command generation module 6 superimposes the wind-resistant feedforward compensation acceleration from the composite adaptive control module 5 with the currently acquired UAV feedforward control terms and feedback control terms, generates the final integrated control command after amplitude limiting processing, and sends it to the UAV's underlying external control interface.

[0100] It should be noted that the above components can be implemented, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, they can be implemented, in whole or in part, as a program instruction product. A program instruction product includes one or a set of program instructions. When the program instructions are loaded and executed on a computer, the described process or function is generated, in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The program instructions can be stored in a computer-readable and writable storage medium or transferred from one computer's readable and writable storage medium to another.

[0101] In this embodiment, the direct coupling or communication connection between the modules can be achieved through indirect coupling or communication connection via interfaces, devices, or parts. The components in this embodiment can dynamically reside within a single processing unit, exist as separate physical entities, or consist of two or more dynamically integrated components within a single processing unit. When these dynamic components are implemented as software functions and sold or used as independent products, they can also be stored in a computer-readable and writable storage medium. This storage medium can be a memory, disk, or optical disc, etc.

[0102] The above description is merely a specific example of the present invention and does not constitute any limitation on the present invention. Obviously, those skilled in the art, after understanding the content and principles of the present invention, may make various modifications and changes in form and detail without departing from the principles and structure of the present invention. For example, when building a feature-constrained neural network and selecting a deep learning framework (such as using TensorFlow instead of PyTorch), the construction function called may be a function in another library or a custom function. However, these modifications and changes based on the ideas of the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. A wind-resistant adaptive control method for unmanned aerial vehicles (UAVs) based on simulation-driven and feature-constrained approaches, characterized in that, include: (1) Using a high-fidelity simulation environment, flight state data of UAVs executing standardized maneuver trajectories under static atmosphere and multi-dimensional dynamic wind field interference are collected. Different wind fields and maneuver trajectory data are grouped and grouped according to the mission intention to obtain a triplet simulation feature dataset including anchor points, positive samples and negative samples. (2) The unknown aerodynamic disturbance is decoupled into a general dynamic representation and a linear coefficient matrix using a deep learning model, and a fully connected neural network containing a spectral normalization layer in the model is selected as the basic feature extractor. (3) Input the triplet simulation feature dataset into the basic feature extractor, perform offline feature loss meta-learning training on it to obtain an inherent feature extractor containing general dynamic representations independent of wind conditions, and fix its network weights. (4) Deploy the aforementioned feature extractor with fixed network weights into the online controller of the UAV onboard computing platform; (5) During the real-time flight of the UAV, acquire data including the current linear velocity, attitude quaternion, and motor pulse width modulation. The system state characteristics, after being normalized and preprocessed to eliminate data distribution bias, are input into the inherent feature extractor for forward propagation, and the reduced basis function feature vector is output. (6) Real-time acquisition of the current UAV expected trajectory and actual response to calculate the composite tracking error, and the error is introduced into the composite adaptive law with forgetting factor as the core feedback signal. The linear coefficient matrix is ​​updated online in real time, and the matrix and the eigenvector of the reduced basis function are solved and converted into wind-resistant feedforward compensation acceleration. (7) The wind-resistant feedforward compensation acceleration is superimposed on the feedforward control quantity based on the desired trajectory reference acceleration to generate a comprehensive control command output to the UAV's underlying controller for execution.

2. The method according to claim 1, characterized in that, In (1), a high-fidelity simulation environment is used to collect flight state data of the UAV when it executes a standardized maneuver trajectory under static atmosphere and multi-dimensional dynamic wind field interference. The data of different wind fields and maneuver trajectories are grouped and combined according to the mission intention. 1a) Collect flight status data of the UAV in a stationary atmosphere: 1a1) For windless environments Extracting UAV state features from simulated UAV flight data. and dynamic response data, the state features Including three-dimensional linear velocity Posture Quaternions Pulse width modulation signals of four motors The dynamic response data includes three-dimensional real acceleration. ; 1a2) The extracted features are timestamped and synchronized with the response data; 1a3) Using the rigid body dynamics equations of UAVs Combined with the drone's gravity vector Thrust vector corresponding to the body attitude Calculate the baseline aerodynamic residual force This step is based on the actual acceleration. Inverse kinematics calculation yields the baseline aerodynamic residual force under windless conditions. ; 1a4) State features after aligning timestamps With the reference aerodynamic residual force Merge them into Anchor sample datasets; 1b) Collect flight status data of UAVs under multi-dimensional dynamic wind fields: 1b1) For different dynamic wind field configurations The data includes simulated flight data of UAVs under constant wind, sinusoidal gusts, and random turbulence, and extracts the state characteristics of the UAVs under the same baseline conditions. , And dynamic response data; 1b2) The extracted features are timestamped and synchronized with the response data; 1b3) Using the above rigid body dynamics equations of the UAV, the actual aerodynamic residual force under the condition of wind field interference is calculated by inverse solution; 1b4) State characteristics of the same maneuver performed at the same time but under different wind field disturbances. Actual aerodynamic residual force Merge them into a positive sample dataset (Positive). 1b5) State characteristics at different times or when performing different maneuvers Actual aerodynamic residual force Merge them into the negative sample dataset; 1b6) will contain feature triples and its corresponding aerodynamic residual force The triplet simulation feature datasets are combined in batches to form a triplet simulation feature dataset for offline feature loss learning training.

3. The method according to claim 1, characterized in that, The deep learning model in (2) includes an input layer, two hidden layers, and an output layer connected in sequence, wherein: The input layer contains 11 neurons: 3 for linear velocity, 4 for attitude quaternions, and 4 for PWM motor outputs. These neurons are used to receive the UAV's state features. ; Two hidden layers, each containing 64 neurons; An eigenvector for outputting basis functions The output layer contains 8 neurons; The linear transformation matrix of the hidden layer and the output layer Both introduce the concept of matrix maximum singular value. The scaling spectral normalization operation uses Update network weights and use a linear rectified function. As an activation function, it restricts the local Lipschitz continuity constant of the neural network, ensuring the stability of online control.

4. The method according to claim 1, characterized in that, In step (2), a deep learning model is used to decouple the unknown aerodynamic disturbance into a general dynamic representation and a linear coefficient matrix, including: 2a) The unknown aerodynamic disturbance force experienced by the UAV in a complex wind field is structurally decomposed into a combination of the linear product of a general aerodynamic basis function vector and an environment-specific linear coefficient matrix; 2b) A deep neural network is used as the basic feature extractor. The system state features of the UAV are used as input. The general aerodynamic basis function vector is output through forward nonlinear mapping. This vector only represents the inherent aerodynamic response law of the UAV body that is invariant across wind fields. 2c) The environment-specific linear coefficient matrix is ​​used as a dedicated parameter to reflect the dynamic changes of the current external wind field. It is multiplied with the general aerodynamic basis function vector to reconstruct and approximate the actual aerodynamic disturbance under the current specific wind field. 2d) Restrict the deep neural network to learn only the underlying motion essence that is independent of wind conditions, and forcibly remove the disturbance component that changes drastically with the wind field from the weights of the deep network and transfer it to the linear coefficient matrix that is independent of the neural network, thereby completing the structural decoupling of nonlinear UAV dynamics and linear wind condition variables.

5. The method according to claim 1, characterized in that, The offline feature loss function training performed on the basic feature extractor in (3) includes: 3a) Calculate the optimal linear coefficients on the fitness set using the least squares method, and calculate the mean square error loss of the physical prediction reconstruction error based on the optimal linear coefficients; 3b) Construct a triplet feature loss function to minimize the feature distance between the anchor state in a windless environment and the positive sample state in a windy environment under the same maneuver, and maximize the feature distance between the anchor state and the negative sample state under different maneuvers. 3c) The gradient descent algorithm is used to jointly optimize the mean squared error loss and the triplet feature loss, update the network weights of the basic feature extractor, and force the neural network to strip away the wind field environmental features and extract the kinematic essential features with wind condition invariance.

6. The method according to claim 1, characterized in that, The intrinsic feature extractor with fixed network weights, as described in (4), is deployed in the online controller of the UAV onboard computing platform, including: 4a) Convert the format of the intrinsic feature extractor with fixed network weights to generate a deployable model file adapted to the UAV onboard computing platform; 4b) Load the deployable model file into the onboard storage space of the UAV for storage; 4c) Configure a feature inference thread in the control architecture of the airborne flight control system. During the operation of the UAV, the feature inference thread reads and initializes the inherent feature extractor from the airborne storage space, receives the preprocessed system state features in real time, and calls the inherent feature extractor to perform forward propagation calculations according to a preset operating cycle.

7. The method according to claim 1, characterized in that, The process in (5) that eliminates data distribution bias after normalization preprocessing includes: 5a) Based on the data format and corresponding effective value range of the motor pulse width modulation signal at the current moment, perform normalization preprocessing to eliminate data distribution offset: If the data format is duty cycle, it is scaled to a preset symmetrical continuous interval through a linear mapping relationship to match the data distribution characteristics during offline training and eliminate online control oscillations caused by input data distribution offset. If the data format is an absolute pulse width time value, then the upper and lower limits of the pulse width extreme values ​​allowed by the underlying control interface are obtained, and they are scaled proportionally to the symmetrical continuous interval through affine transformation, so as to unify the data scale of different motor control inputs in the feature space. 5b) Input the scaled system state features into the intrinsic feature extractor for forward propagation calculation and output the dimensionality-reduced basis function feature vector in real time.

8. The method according to claim 1, characterized in that, The process of solving the updated linear coefficient matrix and the dimensionality-reduced basis function eigenvectors in (6) and converting them into wind-resistant feedforward compensation acceleration includes: 6a) Construct a first feedback update term based on physical prediction error to compensate for the deviation between the real-time observed aerodynamic residual force and the model's predicted aerodynamic disturbance force; 6b) Construct a second feedback update term based on the composite tracking error to eliminate the deviation between the actual flight trajectory and the desired trajectory of the UAV; 6c) Using the covariance matrix evolution equation with a forgetting factor, the adaptive gain of the first feedback update term and the second feedback update term is dynamically adjusted in real time, thereby completing the online update of the linear coefficient matrix.

9. The method according to claim 1, wherein step (7) superimposes the wind-resistant feedforward compensation acceleration onto the feedforward control quantity based on the desired trajectory reference acceleration to generate a comprehensive control command, comprising: 7a) The feedforward control term of the UAV nominal dynamics model, the feedback control term based on the composite tracking error, and the wind-resistant feedforward compensation acceleration are superimposed and mapped into a comprehensive feedforward acceleration in the position controller; 7b) The amplitude of the integrated feedforward acceleration is limited to constrain the safe output range of the underlying control commands; 7c) The combined feedforward acceleration after amplitude limiting, along with the reference position and reference velocity of the desired trajectory, are output to the external control interface of the UAV's underlying position controller to perform flight control.

10. A wind-resistant adaptive control system for unmanned aerial vehicles (UAVs) based on simulation data-driven and feature-constrained meta-learning, characterized in that, include: The simulation data collection and preprocessing module is used to collect multi-wind field simulation flight data, and after combining them into the same group or groups, construct and output a triplet simulation feature dataset. The model offline training module is used to receive the triplet simulation feature dataset, perform offline training using a deep neural network, and output the inherent feature extractor weights after separating the wind field features. The real-time status processing module is used to acquire the status characteristics of the UAV system in real time and perform scaling and normalization preprocessing on them. The online feature inference module is used to receive the preprocessed system state features, call the onboard fixed weight feature extractor to perform forward propagation, and output the dimensionality-reduced basis function feature vector in real time. The composite adaptive control module is used to receive the desired trajectory, the actual response and the basis function eigenvector, and after adaptively updating the linear coefficient matrix, calculate and output the wind-resistant feedforward compensation acceleration. The integrated command generation module is used to receive the wind-resistant feedforward compensation acceleration, process it by amplitude limiting, generate integrated control commands, and send them to the UAV's underlying external control interface.

11. The system according to claim 9, characterized in that, The composite adaptive control module includes: The composite tracking error submodule is used to receive the expected trajectory and actual response data of the UAV, and output the composite tracking error after comparison and calculation. The feedback update construction submodule is used to receive the composite tracking error, and construct and output the first feedback update term based on the physical prediction error and the second feedback update term based on the trajectory yaw, respectively. The adaptive law update submodule is used to receive the feedback update term, dynamically adjust the gain through the matrix evolution equation, and update and output the linear coefficient matrix in real time; The compensation instruction generation submodule is used to receive the updated linear coefficient matrix and the basis function eigenvector, and after combining and solving them, generate and output the wind-resistant feedforward compensation acceleration.

Images