An unmanned aerial vehicle flight control method and system based on aerodynamic residual error learning, a terminal and a storage medium

Abstract

This invention belongs to the field of unmanned aerial vehicles (UAVs) and discloses a UAV flight control method, system, terminal, and storage medium based on aerodynamic residual learning. The method includes: acquiring real-time flight status data and desired reference signals of the UAV, calculating errors to obtain nominal feedback acceleration and error results, and performing low-pass filtering on the error results to obtain physical disturbance estimates; acquiring historical control thrust rate of change command sequences and historical disturbance rate of change baseline sequences, performing orthogonal reconstruction processing to obtain feature vectors; performing disturbance residual prediction processing on the feature vectors through a residual neural network to obtain residual prediction values; fusing the physical disturbance estimates and residual prediction values ​​to obtain total disturbance prediction values, and generating final control commands based on the total disturbance prediction values ​​and nominal feedback acceleration. This invention significantly enhances the anti-interference flight capability of rotary-wing UAVs in strong winds and complex airflow environments.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, and in particular to a UAV flight control method, system, terminal, and computer-readable storage medium based on aerodynamic residual learning. Background Technology

[0002] In real-world, complex operational scenarios, for rotary-wing UAVs to achieve highly reliable safe flight and autonomous obstacle avoidance, they must overcome the strong coupling effects of complex terrain constraints, dynamic disturbance uncertainties, and nonlinear mechanical models. Specifically: First, terrain constraints, such as dense urban buildings, forests, and narrow alleyways, severely limit the safe flight envelope of UAVs. Second, dynamic disturbances, such as variable weather conditions (e.g., sudden airflows and rainfall) and the interaction between wind fields and complex terrain, generate extremely strong and unpredictable environmental disturbances. Third, the inherent mechanical nonlinearity of the system, such as the nonlinear friction inherent in the UAV itself and parameter drift caused by dynamic load changes, further exacerbates the difficulty of low-level control.

[0003] To address the aforementioned issues, existing technologies generally employ three types of disturbance compensation methods: dynamic models based on linear or quadratic drag, predictive models based on computational fluid dynamics, and incremental nonlinear dynamic inverse control methods. However, none of these control schemes can accurately capture and compensate for the nonlinear disturbances caused by actuator lag and high-frequency wind field transients while ensuring extremely low system latency. This results in trajectory tracking divergence and even complete loss of control for UAVs when encountering strong winds or complex airflow environments.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide a flight control method, system, terminal, and computer-readable storage medium for unmanned aerial vehicles (UAVs) based on aerodynamic residual learning. This invention aims to solve the problem in the prior art that it is impossible to accurately capture and compensate for nonlinear disturbances caused by actuator lag and high-frequency wind field transients while ensuring extremely low system latency. This results in UAVs experiencing trajectory tracking divergence or even complete loss of control when encountering strong winds or complex airflow environments.

[0006] To achieve the above objectives, the present invention provides a UAV flight control method based on aerodynamic residual learning, the UAV flight control method based on aerodynamic residual learning comprising the following steps: The real-time flight status data and desired reference signal of the UAV are acquired. Error calculation is performed on the real-time flight status data and the desired reference signal to obtain the nominal feedback acceleration and error results. The error results are then subjected to low-pass filtering to obtain the physical disturbance estimate. Obtain the historical control thrust change rate command sequence and the historical disturbance change rate baseline sequence, and perform orthogonal reconstruction processing based on the historical control thrust change rate command sequence and the historical disturbance change rate baseline sequence to obtain the feature vector; A preset residual neural network is determined, and the feature vector is subjected to perturbation residual prediction processing through the residual neural network to obtain the residual prediction value; The physical disturbance estimate and the residual prediction are fused to obtain the total disturbance prediction. The final control command is then generated based on the total disturbance prediction and the nominal feedback acceleration, so that the UAV can complete the flight mission according to the final control command.

[0007] Optionally, the UAV flight control method based on aerodynamic residual learning, wherein acquiring real-time flight state data and desired reference signals of the UAV, calculating errors in the real-time flight state data and desired reference signals to obtain nominal feedback acceleration and error results, and performing low-pass filtering on the error results to obtain physical disturbance estimates, specifically includes: Acquire real-time flight status data of the UAV, wherein the real-time flight status data includes real-time position, real-time speed, and real-time acceleration; The real-time flight state data is predicted using Newton's kinematics formula to obtain a desired reference signal, wherein the desired reference signal includes desired position, desired velocity, and desired acceleration. Error calculations are performed on the real-time flight status data and the desired reference signal to obtain the nominal feedback acceleration and error results; The error results are processed by a low-pass filter to obtain the physical disturbance estimate.

[0008] Optionally, the UAV flight control method based on aerodynamic residual learning, wherein acquiring the historical control thrust rate of change command sequence and the historical disturbance rate of change baseline sequence, and performing orthogonal reconstruction processing on the historical control thrust rate of change command sequence and the historical disturbance rate of change baseline sequence to obtain a feature vector, specifically includes: The thrust change and sliding window size of the UAV are obtained, and the historical control thrust change rate command sequence is calculated based on the thrust change and sliding window size using the incremental nonlinear dynamic inverse control method. Obtain the historical perturbation rate of change baseline sequence, and perform feature dimension splicing and channel permutation processing based on the historical control thrust rate of change command sequence and the historical perturbation rate of change baseline sequence to obtain the feature vector.

[0009] Optionally, the UAV flight control method based on aerodynamic residual learning, wherein determining a preset residual neural network and performing perturbation residual prediction processing on the feature vector through the residual neural network to obtain residual prediction values, specifically includes: A preset residual neural network is determined, wherein the preset residual neural network includes a one-dimensional convolutional feature extraction layer and a fully connected output layer; The feature vector is input into the one-dimensional convolutional feature extraction layer in the preset residual neural network to obtain convolutional features. The one-dimensional convolutional feature extraction layer includes a one-dimensional convolutional layer, a batch normalization layer, a ReLU nonlinear activation layer, and a max pooling operation layer. The convolutional features are flattened and input into the fully connected output layer of the preset residual neural network to obtain the initial output result; Determine globally learnable scaling parameters, and perform channel-by-channel adaptive adjustment processing on the initial output results based on the scaling parameters to obtain residual prediction values.

[0010] Optionally, the UAV flight control method based on aerodynamic residual learning, wherein the step of determining a preset residual neural network, performing perturbation residual prediction processing on the feature vector through the residual neural network to obtain residual prediction values, further includes: The preset residual neural network is optimized using a self-supervised loss function to obtain the target residual prediction model.

[0011] Optionally, the UAV flight control method based on aerodynamic residual learning, wherein fusing the physical disturbance estimate and the residual prediction to obtain a total disturbance prediction, and generating a final control command based on the total disturbance prediction and the nominal feedback acceleration, so that the UAV completes the flight mission according to the final control command, specifically includes: The estimated physical disturbance and the predicted residual are summed to obtain the total predicted disturbance. The total disturbance prediction value is removed from the nominal feedback acceleration using feedforward compensation to obtain the target compensation value, and the final control command is generated based on the target compensation value. The final control command is sent to the UAV so that the UAV can complete the flight mission according to the final control command.

[0012] Optionally, the UAV flight control method based on aerodynamic residual learning, wherein fusing the physical disturbance estimate and the residual prediction to obtain a total disturbance prediction, and generating a final control command based on the total disturbance prediction and the nominal feedback acceleration, so that the UAV completes the flight mission according to the final control command, further includes: The final control command and the physical disturbance estimate are written to the end of the buffer, and the buffer is updated on a rolling basis according to the final control command and the physical disturbance estimate. If the UAV fails to complete the flight mission according to the final control command, the control cycle will be processed repeatedly until the UAV completes the flight mission.

[0013] Furthermore, to achieve the above objectives, the present invention also provides a UAV flight control system based on aerodynamic residual learning, wherein the UAV flight control system based on aerodynamic residual learning includes: The physical disturbance estimation calculation module is used to acquire real-time flight status data and desired reference signal of the UAV, perform error calculation on the real-time flight status data and desired reference signal to obtain nominal feedback acceleration and error results, and perform low-pass filtering on the error results to obtain physical disturbance estimation values. The orthogonalization reconstruction processing module is used to obtain the historical control thrust change rate command sequence and the historical disturbance change rate baseline sequence, and to perform orthogonalization reconstruction processing on the historical control thrust change rate command sequence and the historical disturbance change rate baseline sequence to obtain feature vectors; The residual prediction value calculation module is used to determine a preset residual neural network, and to perform perturbation residual prediction processing on the feature vector through the residual neural network to obtain the residual prediction value; The final control command generation module is used to fuse the physical disturbance estimate and the residual prediction value to obtain the total disturbance prediction value, and generate the final control command based on the total disturbance prediction value and the nominal feedback acceleration, so that the UAV can complete the flight mission according to the final control command.

[0014] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a UAV flight control program based on aerodynamic residual learning stored in the memory and executable on the processor, wherein when the UAV flight control program based on aerodynamic residual learning is executed by the processor, it implements the steps of the UAV flight control method based on aerodynamic residual learning as described above.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a UAV flight control program based on aerodynamic residual learning, and when the UAV flight control program based on aerodynamic residual learning is executed by a processor, it implements the steps of the UAV flight control method based on aerodynamic residual learning as described above.

[0016] In this invention, real-time flight status data and desired reference signals of a UAV are acquired. Error calculations are performed on the real-time flight status data and desired reference signals to obtain nominal feedback acceleration and error results. The error results are then low-pass filtered to obtain a physical disturbance estimate. Historical control thrust rate of change command sequences and historical disturbance rate of change baseline sequences are acquired. Orthogonal reconstruction processing is performed on the historical control thrust rate of change command sequences and historical disturbance rate of change baseline sequences to obtain feature vectors. A preset residual neural network is determined, and disturbance residual prediction processing is performed on the feature vectors through the residual neural network to obtain residual prediction values. The physical disturbance estimate and the residual prediction values ​​are fused to obtain a total disturbance prediction value. A final control command is generated based on the total disturbance prediction value and the nominal feedback acceleration, enabling the UAV to complete its flight mission according to the final control command. This invention decouples the disturbances experienced by the UAV into physical disturbance estimates and residual predictions, and then fuses the physical disturbance estimates and residual predictions to generate final control commands, which significantly enhances the anti-interference flight capability of rotary-wing UAVs in strong wind and complex airflow environments. Attached Figure Description

[0017] Figure 1 This is a flowchart of a preferred embodiment of the UAV flight control method based on aerodynamic residual learning of the present invention; Figure 2 This is a schematic diagram of the overall process of a preferred embodiment of the UAV flight control method based on aerodynamic residual learning of the present invention; Figure 3 This is a structural diagram of a preferred embodiment of the UAV flight control system based on aerodynamic residual learning of the present invention; Figure 4 This is a structural diagram of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] In real-world, complex operational scenarios, for rotary-wing UAVs to achieve highly reliable safe flight and autonomous obstacle avoidance, they must overcome the strong coupling effects of complex terrain constraints, dynamic disturbance uncertainties, and nonlinear mechanical models. Specifically: First, terrain constraints, such as dense urban buildings, forests, and narrow alleyways, severely limit the safe flight envelope of UAVs. Second, dynamic disturbances, such as variable weather conditions (e.g., sudden airflows and rainfall) and the interaction between wind fields and complex terrain, generate extremely strong and unpredictable environmental disturbances. Third, the inherent mechanical nonlinearity of the system, such as the nonlinear friction inherent in the UAV itself and parameter drift caused by dynamic load changes, further exacerbates the difficulty of low-level control.

[0020] To address the aforementioned issues, existing disturbance compensation methods can be broadly categorized into three types, but all of them exhibit significant limitations when dealing with high-frequency dynamic disturbances: 1. Dynamic models based on linear or quadratic drag: These models are mostly constructed based on steady-state airflow assumptions, which are too idealistic. Therefore, these methods are difficult to accurately characterize and adapt to the complex wind conditions of high frequency, violent and dynamic changes in real-world environments, and are prone to failure under strong disturbances.

[0021] 2. Predictive Models Based on Computational Fluid Dynamics: This method aims to accurately predict flow field states by numerically solving complex physical equations such as the Navier-Stokes equations. However, its drawback lies in its extremely high computational complexity, making it completely unsuitable for real-time operation on the limited computing power of UAV-borne microcomputers. Furthermore, current engineering applications are mainly limited to offline steady-state condition analysis.

[0022] 3. Incremental Nonlinear Dynamic Inversion (INDI): INDI directly estimates the aerodynamic disturbance force vector based on model inversion and instantaneous linearization. However, the actual motor response has inertial delay (actuator lag), and the low-pass filter (LPF) inside the INDI observer will inevitably filter out high-frequency gust signals in order to filter out noise, resulting in serious estimation lag and hardware dependency problems under strong wind disturbances.

[0023] Overall, existing control schemes are unable to accurately capture and compensate for nonlinear disturbances caused by actuator lag and high-frequency wind transients while ensuring extremely low system latency. This technical bottleneck can easily lead to trajectory tracking divergence or even complete loss of control for drones when encountering strong winds or complex airflow environments.

[0024] To address the technical problems of estimation lag and difficulty in modeling high-frequency nonlinear disturbances in traditional INDI control technologies, this invention proposes a flight control method based on aerodynamic residual learning. The basic idea of ​​this invention is to decouple the disturbance into a low-frequency macroscopic baseline (i.e., the physical disturbance estimate in this invention) and a high-frequency nonlinear residual (i.e., the residual prediction in this invention). A traditional INDI observer is used to process the low-frequency physical baseline, while a lightweight multi-layer perceptron (MLP) network is constructed. By introducing historical thrust sequences for causal alignment and feature orthogonalization, this network specifically learns the high-frequency physical components (such as gusts and actuator delays) missed by the low-pass filter.

[0025] This invention achieves zero-shot online feedforward compensation by embedding a forward physics dynamics formula into the loss function and introducing a temporal smoothing penalty. This invention enables self-supervised training without pseudo-labels, significantly improving the real-time performance of disturbance estimation while ensuring the smoothness of control commands, thus significantly enhancing the anti-interference flight capability of rotary-wing UAVs in strong winds and complex airflow environments.

[0026] The preferred embodiment of the UAV flight control method based on aerodynamic residual learning described in this invention, such as... Figure 1 and Figure 2 As shown, the UAV flight control method based on aerodynamic residual learning includes the following steps: Step S10: Obtain the real-time flight status data and expected reference signal of the UAV, perform error calculation on the real-time flight status data and the expected reference signal to obtain the nominal feedback acceleration and error result, and perform low-pass filtering on the error result to obtain the physical disturbance estimate.

[0027] This invention proposes a flight control framework based on aerodynamic residual learning. This framework can estimate disturbance information in complex environments in real time and accurately compensate for nonlinear disturbances caused by actuator hysteresis and high-frequency wind field transients, thereby significantly improving the anti-interference flight capability of rotary-wing UAVs in strong wind environments.

[0028] like Figure 1 As shown, the present invention can be generally divided into the following three main steps: 1. Decoupled Architecture Design Based on Newtonian Physical Information Residual Learning: This invention creatively decouples the total perturbation into two parts: a "low-frequency macroscopic baseline" and a "high-frequency nonlinear residual." Specifically, a traditional incremental nonlinear dynamic inverse (INDI) observer is used to extract the low-frequency physical baseline (i.e., the physical perturbation estimate), while a neural network is constructed to specifically predict the high-frequency nonlinear terms (i.e., the residual predictions) that are difficult to represent by physical formulas. The two are superimposed and fused to finally generate a high-precision perturbation estimate (i.e., the total perturbation prediction) with extremely low latency.

[0029] 2. Orthogonal Reconstruction of Multi-Source Heterogeneous Temporal Features: To ensure that the neural network can accurately learn the causal relationship of perturbations, this invention reconstructs the input feature space. In the offline dataset generated from real flight tests, by introducing the historical control thrust rate of change command sequence, the multiple mapping ambiguities of single pure acceleration data under different flight attitudes are broken, and orthogonal separation of the system's endogenous response and external environmental disturbances is achieved.

[0030] 3. Self-Supervised Physical Model Training for Actuator Constraints: Addressing the engineering constraints of the underlying motor actuator's extreme sensitivity to high-frequency control commands, this invention directly embeds Newton's forward dynamics formula into the loss function of the neural network. By directly fitting the predicted acceleration with the actual IMU acceleration collected at the next moment, self-supervised kinematic consistency training is achieved, completely eliminating pseudo-label noise.

[0031] Specifically, real-time flight status data of the UAV is acquired, including real-time position, real-time velocity, and real-time acceleration; Newton's kinematics formula is used to predict the state of the real-time flight status data to obtain a desired reference signal, including desired position, desired velocity, and desired acceleration; error calculation is performed on the real-time flight status data and the desired reference signal to obtain nominal feedback acceleration and error results; a low-pass filter is used to perform low-pass filtering on the error results to obtain a physical disturbance estimate.

[0032] First, the decoupled architecture design based on Newtonian physical information residual learning in this invention is as follows: In one specific embodiment, the translational rigid body dynamics equations of a rotary-wing UAV can be described by Newton's second law: ; in, Indicates the quality of the drone. This represents the position vector of the drone in three-dimensional space. This represents the actual velocity vector of the drone in three-dimensional space. This represents the actual acceleration vector of the drone in three-dimensional space. This represents the total thrust generated by the four propellers of the drone. This represents the set of total disturbance terms (including unmodeled aerodynamic drag, external wind field disturbances, and motor actuator hysteresis).

[0033] The traditional INDI method for estimating perturbations can be described as follows: ; in, This represents the instantaneous disturbance calculated by the drone at time k. This represents the acceleration vector of the drone at time k. This represents the thrust vector of the drone at the previous moment.

[0034] However, because the raw measurement signal contains a large amount of high-frequency sensor noise, direct feedback can cause severe vibration in the UAV rotor motor. Therefore, traditional INDI must use a low-pass filter (LPF) in series in the disturbance feedback loop for signal smoothing. Its typical first-order discretization update law is expressed as: ; in, Represents the filter coefficients. This represents the disturbance representation of the UAV at time k after low-pass filtering (i.e., the physical disturbance estimate in this invention). This represents the perturbation representation of the UAV at time k-1 after low-pass filtering.

[0035] This invention performs model prediction based on Newton's kinematics formulas to obtain the predicted acceleration for the next moment (i.e., the expected acceleration in the expected reference signal). This invention directly performs L2 norm fitting between the predicted acceleration and the original true acceleration collected by the IMU (Inertial Measurement Unit) at the next moment (i.e., the real-time acceleration in the real-time flight status data). This design completely avoids using inaccurate external observers to generate "pseudo-labels," ensuring absolute kinematic consistency. Due to the target variance decay, the network convergence speed is exponentially improved. Furthermore, by limiting the global Lipschitz constant of the network weight gradient, the residual estimation network is mathematically guaranteed to smoothly suppress small input jitters, ensuring the smoothness of the final output control command.

[0036] Step S20: Obtain the historical control thrust change rate command sequence and the historical disturbance change rate baseline sequence, and perform orthogonal reconstruction processing based on the historical control thrust change rate command sequence and the historical disturbance change rate baseline sequence to obtain the feature vector.

[0037] Specifically, the thrust change and sliding window size of the UAV are obtained, and the historical control thrust change rate command sequence is calculated based on the thrust change and sliding window size using the incremental nonlinear dynamic inverse control method; the historical disturbance change rate baseline sequence is obtained, and feature dimension splicing and channel permutation processing are performed on the historical control thrust change rate command sequence and the historical disturbance change rate baseline sequence to obtain the feature vector.

[0038] This invention employs orthogonal reconstruction of multi-source heterogeneous temporal features, the specific implementation process of which is as follows: To ensure that the neural network can correctly deduce high-frequency wind disturbances from causal relationships, this invention modifies the input state space of the residual neural network. Rigorous feature engineering reconstruction was performed, specifically including: ; ; ; in, This indicates the sequence of historical control thrust change rate commands. This represents the baseline sequence of historical perturbation change rates calculated using traditional INDI. Indicates control of thrust, Indicates the size of the sliding window. Indicates the change in a variable. This indicates the change in thrust.

[0039] The core purpose of introducing this historical command sequence is to break the ambiguity of multiple mappings of pure acceleration data under different flight attitudes and establish a strict temporal causal relationship between "thrust issuance and acceleration response". This invention not only extracts local first-order differences to eliminate the influence of absolute numerical drift, but also eliminates the nominal thrust response at the analytical level, breaking the serious collinearity problem between thrust command and environmental response characteristics, and realizing orthogonal regularization of the input feature space.

[0040] Step S30: Determine the preset residual neural network, and perform perturbation residual prediction processing on the feature vector through the residual neural network to obtain the residual prediction value.

[0041] Specifically, a preset residual neural network is determined, wherein the preset residual neural network includes a one-dimensional convolutional feature extraction layer and a fully connected output layer; the feature vector is input into the one-dimensional convolutional feature extraction layer in the preset residual neural network to obtain convolutional features, wherein the one-dimensional convolutional feature extraction layer includes a one-dimensional convolutional layer, a batch normalization layer, a ReLU nonlinear activation layer, and a max pooling operation layer; the convolutional features are flattened and input into the fully connected output layer in the preset residual neural network to obtain an initial output result; a globally learnable scaling parameter is determined, and the initial output result is adaptively adjusted channel by channel according to the scaling parameter to obtain a residual prediction value.

[0042] This invention sets up a self-supervised physical model training method oriented towards actuator constraints, and the specific implementation process is as follows: In the training of a self-supervised physical model oriented towards actuator constraints, this invention constructs a one-dimensional residual convolutional neural network with Spectral Normalization (SN) constraints to accurately fit high-frequency nonlinear perturbations. This network calculates the current residual by deeply extracting the temporal memory features of historical thrust and baseline perturbations. Its specific mathematical operation flow and network structure definition are as follows: Set the sliding window size to The historical control thrust change rate command sequence and the historical disturbance change rate baseline sequence are concatenated along the feature dimension and channel permutation is performed to form a multi-channel time-series feature tensor (i.e., the feature vector in this invention) as input. eigenvectors First, the data is fed into the convolutional module. The convolutional module includes one-dimensional convolution (Conv1d), batch normalization (BN), ReLU non-linear activation, and max pooling (MaxPool1d) with a kernel size of 2, to obtain the convolutional features. Its mathematical expression is: ; in, Indicates the convolution kernel weights. This indicates the convolution bias.

[0043] Convolutional features Flattened into a one-dimensional vector Then, the input is fed into a cascaded fully connected layer network. To enhance generalization ability, a dropout mechanism is introduced in the hidden layers, expressed as: ; ; in, The output of the intermediate activation function layer of the network. For the output of a fully connected layer network, Indicates the weights of the fully connected layer. This indicates a fully connected bias.

[0044] Finally, to accommodate the differences in perturbation amplitudes across different dimensions of multi-axis UAV dynamics, a globally learnable scaling parameter was set at the network output. The output is adaptively adjusted channel by channel to generate the final three-dimensional high-frequency nonlinear perturbation residual components. The expression is: ; in, This represents the Hadamard product (element-by-element multiplication).

[0045] Furthermore, a self-supervised loss function is used to optimize the preset residual neural network to obtain the target residual prediction model.

[0046] In the model training phase, this invention addresses the engineering challenge of the UAV's underlying motor actuators being extremely sensitive to high-frequency control commands (high-frequency chatter can easily lead to motor overheating and runaway). To address this, this invention designs a self-supervised loss function. The expression is as follows: ; ; ; in, Represents the parameters of the neural network. Indicates the total sequence length. This represents a weight parameter. Indicates the predicted acceleration at the next moment. This represents the raw, true acceleration collected by the IMU. Indicates spectral normalization smoothing penalty. For neural network weights, For neural networks to calculate the gradient of the target, This represents the global Lipschitz constant.

[0047] Step S40: The physical disturbance estimate and the residual prediction are fused to obtain the total disturbance prediction value. The final control command is generated based on the total disturbance prediction value and the nominal feedback acceleration so that the UAV can complete the flight mission according to the final control command.

[0048] Specifically, the estimated physical disturbance value and the predicted residual value are summed to obtain the total disturbance prediction value; the total disturbance prediction value is removed from the nominal feedback acceleration using feedforward compensation to obtain the target compensation value, and a final control command is generated based on the target compensation value; the final control command is sent to the UAV so that the UAV can complete the flight mission according to the final control command.

[0049] Existing traditional architectures exhibit severe physical limitations under strong dynamic wind conditions: on the one hand, real rotor actuators have a non-negligible inertial delay (i.e., actuator lag), causing the current acceleration feedback to be unable to accurately correspond to the command at the previous moment; on the other hand, while filtering out noise, low-pass filters (LPFs) inevitably cause irreversible phase lag and filter out the real high-frequency gust components that cannot be explained by physical formulas, ultimately leading to the failure of high-frequency anti-interference compensation.

[0050] To address the aforementioned theoretical shortcomings, this invention decouples the disturbance into a low-frequency macroscopic baseline and a high-frequency nonlinear residual. This allows the traditional INDI filter output to extract only the macroscopic low-frequency physical baseline. Simultaneously, a neural network is designed to specifically learn the high-frequency nonlinear terms (such as actuator delay and instantaneous high-frequency gusts) that are mistakenly missed by the low-pass filter. Finally, the total disturbance estimate (i.e., the total disturbance prediction) is reconstructed as follows: ; ; in, This indicates a preset residual neural network. This represents the state representation (i.e., feature vector) of the UAV at time k. This represents the high-frequency nonlinear term (i.e., the residual prediction value) of the UAV after low-pass filtering at time k. This represents the estimated value of the physical disturbance.

[0051] The design of this invention significantly expands the parameter space of the neural network, eliminating the need to fit macroscopic rigid body kinematics and concentrating all its representational capabilities on highly nonlinear transient aerodynamic spikes.

[0052] After acquiring the low-frequency physical baseline and the high-frequency network residual respectively, this invention implements a joint compensation mechanism for physical estimation and network residual. Specifically, the disturbance estimate output by the physical prior channel (i.e., physical disturbance estimate) is added to the neural network residual prediction output by the data-driven channel (i.e., residual prediction) to obtain the total disturbance prediction value for the current control cycle (i.e., total disturbance prediction value). Subsequently, the total disturbance prediction value is subtracted from the feedback (i.e., nominal feedback acceleration) calculated from the position tracking error using a feedforward compensation method. Finally, after a set amplitude limiting process, the final safety control command (i.e., final control command) is formed and directly sent to the underlying flight controller (flight controller) for execution.

[0053] Furthermore, the final control command and the physical disturbance estimate are written to the end of the buffer, and the buffer is updated on a rolling basis according to the final control command and the physical disturbance estimate; when the UAV fails to complete the flight mission according to the final control command, the control cycle is cyclically processed until the UAV completes the flight mission.

[0054] The buffer rolling update feature in this invention ensures continuous operation of the control cycle and real-time performance of characteristic states. After the control command for the current cycle is issued, the system performs a buffer rolling update operation. Specifically, the control command issued in the current cycle (the smoothed final control command) and the estimated physical disturbance value for the current cycle are synchronously written to the end of the sliding window buffer, while the oldest data frame at the front of the buffer is removed, thus completing the rolling update of the historical state time series. This operation provides a continuous and accurate data foundation for constructing the sliding window characteristics for the next control cycle, forming a complete closed-loop control data flow.

[0055] The key innovation of this invention is: 1. Newton's Physics Information Residual Learning Architecture: This is the first architecture to propose splitting perturbation compensation into a dual-channel processing of "physical shorting wire (low frequency)" and "neural network residual (high frequency)". This not only relieves the neural network of the burden of fitting macroscopic rigid body kinematics, but also allows all the dense parameter capacity to be used for transient aerodynamic fitting.

[0056] 2. Smooth Self-Supervised Learning Integrating Dynamics and Spectral Normalization Constraints: By directly embedding the physical dynamics model into the forward prediction loss, an unsupervised training paradigm that requires no real wind field data is proposed. Self-supervised learning without pseudo-labels is achieved by embedding the underlying Newton's formula into the loss function and directly comparing it with the next frame of raw IMU data. A creative introduction of a global constant constraint for spectral normalization achieves the optimal physical balance between ultra-low latency and control smoothness.

[0057] 3. Input space reconstruction based on causal alignment and feature orthogonalization: By binding the "historical thrust differential command sequence" and the "physical observer perturbation differential sequence" in the feature window, the ambiguity of acceleration mapping is eliminated, and the orthogonal separation of the system's endogenous response and external environmental disturbance is realized. This solves the problem that traditional perturbation observers based on black-box neural networks are prone to overfitting and significantly improves the robustness of wind resistance estimation.

[0058] This invention provides a low-latency residual disturbance compensation control system for a rotary-wing unmanned aerial vehicle (UAV), comprising: 1. a physical observer (such as INDI) for processing low-frequency macroscopic physical baselines and a neural network for processing high-frequency residuals; 2. an orthogonal feature extraction mechanism using historical thrust and baseline disturbance difference sequences; and 3. a self-supervised network training method that includes physical forward prediction and spectral normalization smoothing penalty, and an execution logic for instantaneous feedforward compensation implemented in the flight controller through algebraic inversion.

[0059] Possible alternatives or modifications to the present invention are as follows: 1. Replacement of Neural Network Structure and Feature Extractor: This invention currently uses MLP for residual learning. On platforms with sufficient computing power, it can be replaced with Temporal Convolutional Network (TCN) or a lightweight Transformer architecture to further capture complex wind field memory features over long time periods.

[0060] 2. Alternative Designs for Low-Frequency Physical Baseline Disturbance Observers: Current designs are based on traditional INDI (Independent Dynamic Interference Device) and its internal low-pass filter. Competitors may propose disturbance observers based on adaptive control, utilizing real-time acquired flight data for self-learning to adapt to different environmental changes. Another possible variation is to use a disturbance observation and compensation strategy based on Model Predictive Control (MPC), which allows for real-time optimization of control inputs during path planning, further enhancing the system's adaptability.

[0061] 3. Alternatives to the Smoothing Penalty Term and Self-Supervised Loss Function: To ensure the smoothness of control commands, the current solution of this invention introduces spectral normalization (SN) constraints into the loss function. As an alternative, a penalty can be added to the time series of the network output in the loss function, or a differentiable discrete low-pass filter layer can be added outside the network output to participate in joint training. Both can achieve the engineering effect of suppressing high-frequency chatter of the actuator.

[0062] Possible design solutions for avoiding the present invention: 1. Based on other disturbance estimation or prediction techniques: Competitors may propose using sensor fusion techniques to obtain disturbance information, rather than relying solely on dynamic models. This can estimate disturbances by fusing data from different sensors (e.g., IMU, GPS, LiDAR, etc.), reducing reliance on system models.

[0063] 2. End-to-end reinforcement learning (RL) control: Competitors may completely abandon the explicit Newtonian physics baseline and observer separation architecture, and use reinforcement learning algorithms to directly map the IMU and historical state to motor PWM signals.

[0064] 3. Adding an external sensor array: Competitors may abandon the pure IMU and dynamic estimation scheme and instead add a miniature multi-directional Pitot tube array and ultrasonic anemometer to the fuselage. They can directly measure the real wind field vector in three-dimensional space through physical sensors and feed it forward to the controller, thereby completely avoiding the dependence on "residual learning network" and "feature orthogonalization reconstruction".

[0065] In addition, the present invention may also include the following supplementary contents: 1. Cross-platform integration and semantic collaboration of complex robot systems: The low-latency residual wind disturbance resistance algorithm of this invention is not only applicable to standard rotary-wing UAVs, but can also be seamlessly integrated into the underlying control system of autonomous navigation of Land-Air Bimodal Robots (LABRs) in the future. Especially in agricultural inspection or complex unstructured scenarios, when the system accesses a Vision-Language Model (VLM) or Large Language Model (LLM) for advanced semantic reasoning and target recognition, the upper-level large model inevitably introduces high inference latency. The underlying zero-latency transient wind disturbance compensation algorithm of this invention can greatly enhance the robot's hovering stability, providing a valuable stable time window for upper-level VLM inference.

[0066] 2. Evolutionary Potential of Online Incremental Learning: The current network weights in this invention are trained offline based on a discrete orthogonal dataset. A lightweight online incremental learning mechanism can be introduced in the future. If the drone encounters extreme weather conditions not covered in the training set during actual flight (such as high-altitude freezing airflow or vortex washing), the algorithm can utilize onboard computing power to fine-tune the network weights in real time in the background based on a pseudo-label-free, physically self-supervised loss, enabling the drone to possess a self-evolving anti-interference capability that becomes increasingly stable over time.

[0067] 3. Highly Agile Applications in Specialized Counter-Drone Operations and Anti-UAV Fields: The extremely strong anti-disturbance capabilities and low-latency response characteristics imparted by this invention have significant value for specialized applications. In anti-UAV combat scenarios or close-range engagements, friendly UAVs often face strong airflow caused by enemy interception networks, jet interference, or high-speed maneuvering. This invention can instantly compensate for these high-frequency nonlinear physical disturbances, ensuring the maneuverability and survivability of aircraft in extreme combat environments, and providing underlying technical support for the development of specialized UAV modules with high escape capabilities and high payload stability.

[0068] Furthermore, such as Figure 3 As shown, based on the above-mentioned UAV flight control method based on aerodynamic residual learning, the present invention also provides a UAV flight control system based on aerodynamic residual learning, wherein the UAV flight control system based on aerodynamic residual learning includes: The physical disturbance estimation calculation module 51 is used to acquire real-time flight status data and expected reference signal of the UAV, perform error calculation on the real-time flight status data and expected reference signal to obtain nominal feedback acceleration and error results, and perform low-pass filtering on the error results to obtain physical disturbance estimation value. The orthogonalization reconstruction processing module 52 is used to obtain the historical control thrust change rate command sequence and the historical disturbance change rate baseline sequence, and to perform orthogonalization reconstruction processing based on the historical control thrust change rate command sequence and the historical disturbance change rate baseline sequence to obtain feature vectors; The residual prediction value calculation module 53 is used to determine a preset residual neural network, and to perform perturbation residual prediction processing on the feature vector through the residual neural network to obtain the residual prediction value; The final control command generation module 54 is used to fuse the physical disturbance estimate and the residual prediction value to obtain the total disturbance prediction value, and generate the final control command based on the total disturbance prediction value and the nominal feedback acceleration, so that the UAV can complete the flight mission according to the final control command.

[0069] Furthermore, such as Figure 4 As shown, based on the above-mentioned UAV flight control method and system based on aerodynamic residual learning, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 4 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0070] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, SmartMediaCard (SMC), SecureDigital (SD) card, or FlashCard. Furthermore, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a UAV flight control program 40 based on aerodynamic residual learning, which can be executed by the processor 10 to implement the UAV flight control method based on aerodynamic residual learning described in this application.

[0071] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the UAV flight control method based on aerodynamic residual learning.

[0072] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. The display 30 is used to display information on the terminal and to display a visual user interface.

[0073] In one embodiment, when the processor 10 executes the UAV flight control program 40 based on aerodynamic residual learning in the memory 20, it implements the steps of the UAV flight control method based on aerodynamic residual learning as described above.

[0074] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a UAV flight control program based on aerodynamic residual learning, and the UAV flight control program based on aerodynamic residual learning, when executed by a processor, implements the steps of the UAV flight control method based on aerodynamic residual learning as described above.

[0075] In summary, this invention provides a UAV flight control method, system, terminal, and storage medium based on aerodynamic residual learning. The method includes: acquiring real-time flight state data and desired reference signals of the UAV; calculating errors in the real-time flight state data and desired reference signals to obtain nominal feedback acceleration and error results; performing low-pass filtering on the error results to obtain physical disturbance estimates; acquiring historical control thrust rate of change command sequences and historical disturbance rate of change baseline sequences; performing orthogonal reconstruction processing on the historical control thrust rate of change command sequences and historical disturbance rate of change baseline sequences to obtain feature vectors; determining a preset residual neural network; performing disturbance residual prediction processing on the feature vectors through the residual neural network to obtain residual prediction values; fusing the physical disturbance estimates and residual prediction values ​​to obtain total disturbance prediction values; and generating final control commands based on the total disturbance prediction values ​​and the nominal feedback acceleration, so that the UAV completes its flight mission according to the final control commands. This invention decouples the disturbances experienced by the UAV into physical disturbance estimates and residual predictions, and then fuses the physical disturbance estimates and residual predictions to generate final control commands, which significantly enhances the anti-interference flight capability of rotary-wing UAVs in strong wind and complex airflow environments.

[0076] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0077] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0078] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for unmanned aerial vehicle flight control based on aerodynamic residual learning, characterized in that, The UAV flight control method based on aerodynamic residual learning includes: The real-time flight status data and desired reference signal of the UAV are acquired. Error calculation is performed on the real-time flight status data and the desired reference signal to obtain the nominal feedback acceleration and error results. The error results are then subjected to low-pass filtering to obtain the physical disturbance estimate. Obtain the historical control thrust change rate command sequence and the historical disturbance change rate baseline sequence, and perform orthogonal reconstruction processing based on the historical control thrust change rate command sequence and the historical disturbance change rate baseline sequence to obtain the feature vector; A preset residual neural network is determined, and the feature vector is subjected to perturbation residual prediction processing through the residual neural network to obtain the residual prediction value; The physical disturbance estimate and the residual prediction are fused to obtain the total disturbance prediction. The final control command is then generated based on the total disturbance prediction and the nominal feedback acceleration, so that the UAV can complete the flight mission according to the final control command.

2. The method of claim 1, wherein, The process of acquiring real-time flight status data and desired reference signals of the UAV, calculating errors in the real-time flight status data and desired reference signals to obtain nominal feedback acceleration and error results, and performing low-pass filtering on the error results to obtain physical disturbance estimates specifically includes: Acquire real-time flight status data of the UAV, wherein the real-time flight status data includes real-time position, real-time speed, and real-time acceleration; The real-time flight state data is predicted using Newton's kinematics formula to obtain a desired reference signal, wherein the desired reference signal includes desired position, desired velocity, and desired acceleration. Error calculations are performed on the real-time flight status data and the desired reference signal to obtain the nominal feedback acceleration and error results; The error results are processed by a low-pass filter to obtain the physical disturbance estimate.

3. The method of claim 1, wherein, The process of acquiring the historical control thrust change rate command sequence and the historical disturbance change rate baseline sequence, and then performing orthogonal reconstruction processing on the historical control thrust change rate command sequence and the historical disturbance change rate baseline sequence to obtain a feature vector, specifically includes: The thrust change and sliding window size of the UAV are obtained, and the historical control thrust change rate command sequence is calculated based on the thrust change and sliding window size using the incremental nonlinear dynamic inverse control method. Obtain the historical perturbation rate of change baseline sequence, and perform feature dimension splicing and channel permutation processing based on the historical control thrust rate of change command sequence and the historical perturbation rate of change baseline sequence to obtain the feature vector.

4. The method of claim 1, wherein, The step of determining a preset residual neural network and performing perturbation residual prediction processing on the feature vector through the residual neural network to obtain residual prediction values ​​specifically includes: A preset residual neural network is determined, wherein the preset residual neural network includes a one-dimensional convolutional feature extraction layer and a fully connected output layer; The feature vector is input into the one-dimensional convolutional feature extraction layer in the preset residual neural network to obtain convolutional features. The one-dimensional convolutional feature extraction layer includes a one-dimensional convolutional layer, a batch normalization layer, a ReLU nonlinear activation layer, and a max pooling operation layer. The convolutional features are flattened and input into the fully connected output layer of the preset residual neural network to obtain the initial output result; Determine globally learnable scaling parameters, and perform channel-by-channel adaptive adjustment processing on the initial output results based on the scaling parameters to obtain residual prediction values.

5. The method of claim 1, wherein, The process of determining a preset residual neural network, performing perturbation residual prediction processing on the feature vector using the residual neural network to obtain residual prediction values, further includes: The preset residual neural network is optimized using a self-supervised loss function to obtain the target residual prediction model.

6. The method of claim 1, wherein, The process of fusing the physical disturbance estimate and the residual prediction to obtain the total disturbance prediction, and generating a final control command based on the total disturbance prediction and the nominal feedback acceleration, so that the UAV can complete the flight mission according to the final control command, specifically includes: The estimated physical disturbance and the predicted residual are summed to obtain the total predicted disturbance. The total disturbance prediction value is removed from the nominal feedback acceleration using feedforward compensation to obtain the target compensation value, and the final control command is generated based on the target compensation value. The final control command is sent to the UAV so that the UAV can complete the flight mission according to the final control command.

7. The method of claim 1, wherein, The process of fusing the physical disturbance estimate and the residual prediction to obtain the total disturbance prediction, and generating a final control command based on the total disturbance prediction and the nominal feedback acceleration, so that the UAV can complete the flight mission according to the final control command, further includes: The final control command and the physical disturbance estimate are written to the end of the buffer, and the buffer is updated on a rolling basis according to the final control command and the physical disturbance estimate. If the UAV fails to complete the flight mission according to the final control command, the control cycle will be processed repeatedly until the UAV completes the flight mission.

8. A flight control system for unmanned aerial vehicles (UAVs) based on aerodynamic residual learning, characterized in that, The UAV flight control system based on aerodynamic residual learning includes: The physical disturbance estimation calculation module is used to acquire real-time flight status data and desired reference signal of the UAV, perform error calculation on the real-time flight status data and desired reference signal to obtain nominal feedback acceleration and error results, and perform low-pass filtering on the error results to obtain physical disturbance estimation values. The orthogonalization reconstruction processing module is used to obtain the historical control thrust change rate command sequence and the historical disturbance change rate baseline sequence, and to perform orthogonalization reconstruction processing on the historical control thrust change rate command sequence and the historical disturbance change rate baseline sequence to obtain feature vectors; The residual prediction value calculation module is used to determine a preset residual neural network, and to perform perturbation residual prediction processing on the feature vector through the residual neural network to obtain the residual prediction value; The final control command generation module is used to fuse the physical disturbance estimate and the residual prediction value to obtain the total disturbance prediction value, and generate the final control command based on the total disturbance prediction value and the nominal feedback acceleration, so that the UAV can complete the flight mission according to the final control command.

9. A terminal, characterized in that, The terminal includes: a memory, a processor, and a UAV flight control program based on aerodynamic residual learning stored in the memory and executable on the processor. When the UAV flight control program based on aerodynamic residual learning is executed by the processor, it implements the steps of the UAV flight control method based on aerodynamic residual learning as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a UAV flight control program based on aerodynamic residual learning, which, when executed by a processor, implements the steps of the UAV flight control method based on aerodynamic residual learning as described in any one of claims 1-7.

Images