UAV trajectory control method and system

Abstract

The present invention relates to the field of unmanned aerial vehicle (UAV) trajectory control technology, and more particularly to a UAV trajectory control method and system. The method comprises the following steps: first, obtaining the UAV's flight braking parameters and delivery mission data, and extracting characteristic data of low-altitude delivery meteorological conditions; then, simulating the flight state based on the flight braking parameters and mission data, and quantifying flight attitude instability in combination with meteorological conditions, thereby learning trajectory avoidance behavior control; finally, applying a policy gradient algorithm to construct a trajectory control model, which is then sent to the UAV control center for trajectory control. This present invention further improves UAV trajectory control technology by optimizing it.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) trajectory control, and in particular to a UAV trajectory control method and system. Background Art

[0002] In low-altitude flight environments, meteorological factors (such as wind speed, temperature, humidity, and vortices) and the influence of bird flocks can significantly impact the flight attitude and trajectory of drones, leading to decreased flight stability and even the risk of flight loss of control. Therefore, effectively controlling the flight trajectory of drones to ensure their safety and stability in complex environments has become a critical issue in drone research and applications. Previous drone trajectory control methods have relied on preset flight paths and simple feedback control systems, making them difficult to adapt to dynamically changing environmental factors. Low-altitude flight missions, especially those requiring precise navigation like drone delivery, are limited by multiple uncertainties such as meteorological conditions and the influence of bird flocks. However, a traditional drone trajectory control method suffers from a lack of ability to effectively detect and respond to unexpected conditions (i.e., meteorological conditions and bird flocks) during simulated delivery flights, resulting in large errors in drone trajectory control. Summary of the Invention

[0003] Based on this, it is necessary to provide a UAV trajectory control method and system to solve at least one of the above technical problems.

[0004] To achieve the above object, a UAV trajectory control method is provided, the method comprising the following steps:

[0005] Step S1: Obtaining the UAV flight braking parameters and the UAV delivery mission data; extracting the low-altitude delivery meteorological conditions from the UAV delivery mission data to obtain the low-altitude delivery meteorological condition characteristic data;

[0006] Step S2: Based on the UAV flight braking parameters and the UAV delivery mission data, a UAV delivery flight state simulation is performed to obtain UAV delivery flight state simulation data; based on the low-altitude delivery meteorological condition characteristic data, a flight attitude instability simulation and quantification is performed on the UAV delivery flight state simulation data to obtain flight attitude instability quantification data; based on the flight attitude instability quantification data, a UAV flight trajectory avoidance behavior control learning is performed to obtain flight trajectory avoidance behavior control data;

[0007] Step S3: Based on the policy gradient algorithm, a trajectory control model is constructed for the flight trajectory avoidance behavior control data to obtain the UAV flight trajectory control model; the UAV flight trajectory control model is sent to the UAV control center to execute the UAV trajectory control.

[0008] Preferably, step S1 includes the following steps:

[0009] Step S11: Acquire UAV flight braking parameters and UAV delivery mission data;

[0010] Step S12: Cleaning the drone delivery task data to obtain drone delivery task cleaned data;

[0011] Step S13: extracting low-altitude delivery meteorological conditions from the UAV delivery task cleaning data to obtain low-altitude delivery meteorological condition data;

[0012] Step S14: Perform feature analysis on the low-altitude delivery meteorological condition data to obtain low-altitude delivery meteorological condition feature data.

[0013] Preferably, step S2 includes the following steps:

[0014] Step S21: simulating the UAV delivery flight state based on the UAV flight braking parameters and the UAV delivery mission data to obtain UAV delivery flight state simulation data;

[0015] Step S22: Obtain historical bird flock flight interference data for drone delivery; extract delivery load from the drone delivery mission data to obtain drone delivery load data;

[0016] Step S23: performing flight attitude instability simulation and quantification on the UAV delivery flight state simulation data based on the low-altitude delivery meteorological condition characteristic data and the UAV delivery payload data to obtain flight attitude instability quantification data;

[0017] Step S24: Based on the historical bird flock flight interference data and the flight attitude instability quantified data, the UAV flight trajectory avoidance behavior control learning is performed to obtain the flight trajectory avoidance behavior control data.

[0018] Preferably, step S23 includes the following steps:

[0019] Step S231: performing low-altitude turbulence structure identification on the low-altitude delivery meteorological condition characteristic data to obtain low-altitude turbulence structure data; performing wind direction shear phase plane state analysis on the low-altitude turbulence structure data to obtain wind direction shear phase plane state;

[0020] Step S232: calculating the average difference of asymmetric aerodynamic pressure of the UAV on the UAV delivery flight state simulation data based on the wind direction shear phase plane state to obtain the average difference of asymmetric aerodynamic pressure of the UAV;

[0021] Step S233: Calculating the rolling moment fluctuation of the UAV delivery flight state simulation data based on the UAV asymmetric aerodynamic pressure difference and the wind shear phase plane state to obtain the UAV flight rolling moment fluctuation data;

[0022] Step S234: performing a flight vibration multi-axis intensity regression analysis on the UAV delivery flight state simulation data based on the UAV flight rolling moment fluctuation data, the UAV asymmetric aerodynamic pressure average difference, and the UAV delivery payload data to obtain flight vibration multi-axis intensity regression data;

[0023] Step S235: performing flight attitude instability simulation and quantification based on the flight tremor multi-axis intensity regression data to obtain flight attitude instability quantification data.

[0024] Preferably, step S234 includes the following steps:

[0025] Perform multi-axis rotational inertia incremental integration on the UAV's flight rolling moment fluctuation data to obtain multi-axis rotational inertia incremental integral data;

[0026] The UAV angular velocity progressive growth ratio is calculated based on the UAV flight rolling moment fluctuation data to obtain the angular velocity progressive growth ratio;

[0027] The axial yaw moment momentum deviation is quantified based on the average difference of the asymmetric aerodynamic pressure of the UAV to obtain the axial yaw moment momentum deviation data;

[0028] The UAV aerodynamic pressure imbalance is quantified based on the multi-axis rotational inertia incremental integral data, the UAV delivery load data and the axial yaw moment momentum deviation data to obtain the aerodynamic pressure imbalance quantification data;

[0029] Based on the quantified data of aerodynamic pressure imbalance, the ratio of angular velocity progressive increase and the multi-axis moment of inertia incremental integral data, the aerodynamic structure flutter inter-axis strength coupling simulation analysis is carried out to obtain the aerodynamic structure flutter inter-axis strength coupling data;

[0030] Based on the inter-axis intensity coupling data of aerodynamic structure flutter, a multi-axis intensity regression analysis of flight flutter is performed on the simulation data of UAV delivery flight state to obtain the multi-axis intensity regression data of flight flutter.

[0031] Preferably, step S24 includes the following steps:

[0032] Step S241: Mark the multi-target bird flock movement trajectory of the historical bird flock flight interference data to obtain the multi-target bird flock movement trajectory;

[0033] Step S242: performing approach relative orientation clustering learning on the historical bird flock flight interference data based on the multi-target historical movement trajectories of the bird flock to obtain bird flock approach relative orientation clustering data;

[0034] Step S243: performing a quantitative assessment of the airflow disturbance caused by the approach of the bird flock on the historical bird flock flight interference data according to the historical movement trajectories of the bird flock multi-target, thereby obtaining quantitative data of the airflow disturbance caused by the approach of the bird flock;

[0035] Step S244: identifying the critical state of the UAV's flight attitude instability based on the quantitative data of the airflow disturbance caused by the approaching flock of birds and the clustered data of the relative orientation of the approaching flock of birds, and obtaining the critical state of the flight attitude instability;

[0036] Step S245: Based on the relative orientation clustering data of the approaching flock of birds and the critical state of flight attitude instability, the UAV flight trajectory avoidance behavior control learning is performed to obtain flight trajectory avoidance behavior control data.

[0037] Preferably, step S244 includes the following steps:

[0038] The multi-directional flapping wing rotating vortex intensity data were obtained by analyzing the multi-directional flapping wing rotating vortex intensity data based on the quantitative data of airflow disturbance and the relative orientation clustering data of the approaching birds.

[0039] The flapping rhythm vortex transient mutation data are identified based on the multi-directional flapping wing rotation vortex intensity data to obtain the flapping rhythm vortex transient mutation data;

[0040] The flight control surface deflection stall index of the UAV is evaluated based on the transient mutation data of the flapping rhythm vortex and the flight attitude instability quantitative data, and the flight control surface deflection stall index is obtained.

[0041] The critical state of UAV flight attitude instability is identified based on the flight attitude instability quantitative data according to the flight control surface deflection stall index, and the critical state of flight attitude instability is obtained.

[0042] Preferably, step S3 includes the following steps:

[0043] Step S31: performing convolution processing on the flight trajectory avoidance behavior control data to obtain flight trajectory avoidance behavior convolution data;

[0044] Step S32: performing iterative incremental learning on the flight trajectory avoidance behavior convolution data to obtain flight trajectory avoidance behavior iterative data;

[0045] Step S33: constructing a trajectory control model based on the flight trajectory avoidance behavior iterative data based on the policy gradient algorithm to obtain a UAV flight trajectory control model;

[0046] Step S34: Send the UAV flight trajectory control model to the UAV control center to perform UAV trajectory control.

[0047] Preferably, step S33 includes the following steps:

[0048] Step S331: dividing the flight trajectory avoidance behavior iteration data into a training set and a test set to obtain a flight trajectory avoidance behavior training set and a flight trajectory avoidance behavior test set respectively;

[0049] Step S332: performing clustering processing on the flight trajectory avoidance behavior training set to obtain a flight trajectory avoidance behavior cluster training set;

[0050] Step S333: performing feature sampling on the flight trajectory avoidance behavior cluster training set to obtain a trajectory avoidance behavior cluster sampling training set;

[0051] Step S334: constructing an initial trajectory control model for the trajectory avoidance behavior cluster sampling training set based on the policy gradient algorithm to obtain an initial UAV flight trajectory control model;

[0052] Step S335: Input the trajectory avoidance behavior test set into the initial UAV flight trajectory control model for model testing to obtain the UAV flight trajectory control model.

[0053] Preferably, the present invention further provides a UAV trajectory control system for executing the UAV trajectory control method described above, the UAV trajectory control system comprising:

[0054] The low-altitude delivery meteorological condition extraction module is used to obtain the UAV flight braking parameters and UAV delivery mission data; the low-altitude delivery meteorological condition extraction module is used to extract the low-altitude delivery meteorological condition characteristic data from the UAV delivery mission data;

[0055] The flight trajectory avoidance behavior learning module is used to simulate the UAV delivery flight state based on the UAV flight braking parameters and the UAV delivery mission data to obtain the UAV delivery flight state simulation data; simulate and quantify the flight attitude instability of the UAV delivery flight state simulation data based on the low-altitude delivery meteorological condition characteristic data to obtain the flight attitude instability quantification data; and perform UAV flight trajectory avoidance behavior control learning based on the flight attitude instability quantification data to obtain the flight trajectory avoidance behavior control data;

[0056] The trajectory control model construction module is used to construct a trajectory control model for the flight trajectory avoidance behavior control data based on the policy gradient algorithm to obtain the UAV flight trajectory control model; the UAV flight trajectory control model is sent to the UAV control center to execute UAV trajectory control.

[0057] The present invention achieves this by first acquiring UAV flight braking parameters and UAV delivery mission data. This data includes the UAV's flight performance, braking system characteristics, flight plan, and specific delivery mission requirements, such as the flight path and mission start and end points. Next, by extracting low-altitude delivery meteorological conditions, meteorological characteristic data relevant to flight safety, such as wind speed, air pressure, temperature, humidity, and airflow variations, is identified. This low-altitude meteorological data provides the basis for subsequent flight state simulation, enabling the prediction and optimization of flight performance under different meteorological conditions. The flight and meteorological data acquired in step S1 are then used to simulate the UAV's flight state during an actual delivery mission. This flight state simulation considers multiple variables, including altitude, speed, direction, and attitude, to further analyze the risks and challenges associated with different flight states. Next, based on the low-altitude delivery meteorological condition characteristic data, flight attitude instability is simulated to quantify the risk of instability. For example, wind speed fluctuations or unstable airflow can cause UAV flight attitude instability, posing a threat to mission completion. Therefore, by simulating and quantifying instability, flight trajectory optimization and avoidance strategies can be developed, ensuring flight stability and safety. Based on the quantitative data of flight attitude instability from the first two steps, flight trajectory avoidance behavior control learning is performed. This process primarily analyzes unstable flight trajectories to construct a control strategy that can effectively avoid dangerous flight paths. This process utilizes a policy gradient algorithm to optimize the UAV's flight trajectory under complex weather conditions and avoid flight attitude instability. Finally, the trained and optimized trajectory control model is transmitted to the UAV control center, where it is used for real-time flight trajectory control, ensuring that the UAV can adaptively adjust its flight trajectory during mission execution, avoid unsafe areas, and successfully complete the mission. Therefore, this invention optimizes a traditional UAV trajectory control method, addressing the problem of traditional UAV trajectory control methods' weak ability to effectively judge and respond to unexpected conditions (i.e., weather conditions, bird flocks) during simulated delivery flights, which results in large UAV trajectory control errors. This method improves the ability to effectively judge and respond to unexpected conditions (i.e., weather conditions, bird flocks), and reduces UAV trajectory control errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A schematic diagram of the steps of a UAV trajectory control method;

[0059] Figure 2 for Figure 1 Detailed implementation steps of step S2 in FIG.

[0060] Figure 3 for Figure 1 Detailed implementation steps of step S3 in FIG. DETAILED DESCRIPTION

[0061] See also Figures 1 to 3 , a UAV trajectory control method, the method comprising the following steps:

[0062] Step S1: Obtaining the UAV flight braking parameters and the UAV delivery mission data; extracting the low-altitude delivery meteorological conditions from the UAV delivery mission data to obtain the low-altitude delivery meteorological condition characteristic data;

[0063] Step S2: Based on the UAV flight braking parameters and the UAV delivery mission data, a UAV delivery flight state simulation is performed to obtain UAV delivery flight state simulation data; based on the low-altitude delivery meteorological condition characteristic data, a flight attitude instability simulation and quantification is performed on the UAV delivery flight state simulation data to obtain flight attitude instability quantification data; based on the flight attitude instability quantification data, a UAV flight trajectory avoidance behavior control learning is performed to obtain flight trajectory avoidance behavior control data;

[0064] Step S3: Based on the policy gradient algorithm, a trajectory control model is constructed for the flight trajectory avoidance behavior control data to obtain the UAV flight trajectory control model; the UAV flight trajectory control model is sent to the UAV control center to execute the UAV trajectory control.

[0065] In the embodiment of the present invention, reference Figure 1 The above is a schematic flow chart of the steps of a drone trajectory control method of the present invention. In this example, the drone trajectory control method includes the following steps:

[0066] Step S1: Obtaining the UAV flight braking parameters and the UAV delivery mission data; extracting the low-altitude delivery meteorological conditions from the UAV delivery mission data to obtain the low-altitude delivery meteorological condition characteristic data;

[0067] In this embodiment of the present invention, when acquiring UAV flight braking parameters and delivery mission data, a ground control station synchronization scheduling system receives measured flight control parameters from the UAV's onboard sensors. These parameters include the maximum pitch angle setting (±35°), maximum roll angle (±45°), maximum vertical velocity (5 m / s), maximum horizontal velocity (18 m / s), propeller reaction time delay (0.2 s), and flight control feedback delay (150 ms). These parameters are used as flight braking parameters. During the mission data acquisition process, the backend logistics scheduling management system exports delivery mission information including the mission origin and destination latitude and longitude, takeoff and landing times, delivery weight, mission number, meteorological conditions within the flight path, and flight altitude level (30–120 m below sea level). Simultaneously, low-altitude meteorological observation data within the corresponding mission time window is retrieved. This data is sourced from the accompanying LIDAR low-altitude scanning equipment. Data is extracted, including low-altitude wind speed (instantaneous wind speed, 10-second average wind speed), wind direction, humidity, temperature, rainfall, air pressure, and thunderstorm probability, with a resolution of one group per 1 km. The original mission data was then inspected field by field to remove missing, duplicate, or incorrectly formatted records. Data cleaning was achieved through regular expression matching-based field format verification. After cleaning, the retained fields were uniformly stored in UTF-8 encoding. K-means clustering was then used to perform a preliminary classification and identification of the meteorological data. Weather conditions were categorized into stable, moderately disturbed, and highly disturbed types based on wind speed, direction, and humidity. Principal component analysis (PCA) was then applied to reduce the dimensionality of the multidimensional meteorological variables and extract the main influencing factors. The first five principal components were used as characteristic data for low-altitude delivery meteorological conditions and fed into the next step.

[0068] Step S2: Based on the UAV flight braking parameters and the UAV delivery mission data, a UAV delivery flight state simulation is performed to obtain UAV delivery flight state simulation data; based on the low-altitude delivery meteorological condition characteristic data, a flight attitude instability simulation and quantification is performed on the UAV delivery flight state simulation data to obtain flight attitude instability quantification data; based on the flight attitude instability quantification data, a UAV flight trajectory avoidance behavior control learning is performed to obtain flight trajectory avoidance behavior control data;

[0069] In this embodiment of the present invention, a three-dimensional trajectory simulation engine is used to simulate the flight process based on flight braking parameters and delivery mission data. This simulator utilizes a six-degree-of-freedom dynamic model based on physical modeling, taking into account factors such as air resistance, propulsion force, gravity, and attitude adjustment torque. It inputs mission path points, a wind field disturbance model, and flight control delay parameters to generate spatial trajectory data, attitude angle data (pitch, roll, and yaw angles), acceleration vectors, and attitude stability records for the simulated flight process. This simulated flight state data is then fused with the meteorological characteristic data extracted in step S1. Using a wind field disturbance model constructed based on CFD calculation results, the flight attitude deviation trend is identified by analyzing lift fluctuations within localized airflow instability regions. Numerical integration methods are then used to quantify attitude stability at each time point. Quantitative indicators include attitude deviation rate, return-to-center response time, and attitude oscillation frequency, ultimately yielding quantitative data on flight attitude instability. Combined with the quantitative results of attitude instability, a flight avoidance behavior control learning task was constructed. Historical bird disturbance events from the Birdstrike dataset were used as training samples. A graph neural network (GNN) was used to map the location, timing, and intensity of obstacle sources, mapping risky sections of the flight path. Simulations compared the attitude recovery outcomes of different behavioral decisions (such as early detours, reduced altitude, and accelerated passage) under corresponding disturbance conditions. The Q-learning algorithm, a reinforcement learning algorithm, was used for policy training, selecting the behavior with the highest value in each state. During training, the state space was defined as attitude angular rate, obstacle approach velocity, and disturbance intensity, while the action space consisted of flight acceleration and deceleration, steering angle correction, and altitude adjustment. The reward function was a weighted combination of minimizing attitude deviation and minimizing flight energy consumption. Ultimately, the flight trajectory avoidance behavior control data was output.

[0070] Step S3: Based on the policy gradient algorithm, a trajectory control model is constructed for the flight trajectory avoidance behavior control data to obtain the UAV flight trajectory control model; the UAV flight trajectory control model is sent to the UAV control center to execute the UAV trajectory control.

[0071] In an embodiment of the present invention, a trajectory control model is constructed for the flight trajectory avoidance behavior control data generated in the previous stage. First, three-dimensional convolution processing is performed on the control data, and 3D-CNN is used to extract the spatial motion trend features in the time series. Each set of data has a time series length of 32 frames, and each frame contains 12-dimensional features including position, attitude angle, and motion control amount. The convolution kernel size is set to 3×3×3, the step size is 1, and the extraction result is input into the GRU structure for time series modeling. After the feature sequence is output, data normalization is performed. The data set is then divided, and the ratio of the training set to the test set is 8:2. The K-means++ algorithm is used to cluster the trajectory control behavior, and trajectory control data samples under different types of avoidance modes are extracted. Each cluster center represents an avoidance strategy, such as large side flight, small altitude adjustment, rapid descent, etc. A policy gradient algorithm (such as Proximal Policy Optimization (PPO)) with Bayesian optimization parameters was then applied to each clustered sample to train the trajectory control policy. A loss function focused on minimizing trajectory deviation was used, with 300 training iterations and evaluation of the validation set trajectory deviations during each round. Finally, the final policy weight parameters were output to construct the UAV flight trajectory control model. This model was exported in ONNX format and sent to the UAV ground control center via a 5G communication module. It was then embedded in the UAV's mission scheduling system for real-time trajectory control. This control model supports a control frame rate of 20 Hz and is capable of real-time dynamic trajectory adjustment to ensure flight stability and safety.

[0072] Step S1 includes the following steps:

[0073] Step S11: Acquire UAV flight braking parameters and UAV delivery mission data;

[0074] Step S12: Cleaning the drone delivery task data to obtain drone delivery task cleaned data;

[0075] Step S13: extracting low-altitude delivery meteorological conditions from the UAV delivery task cleaning data to obtain low-altitude delivery meteorological condition data;

[0076] Step S14: Perform feature analysis on the low-altitude delivery meteorological condition data to obtain low-altitude delivery meteorological condition feature data.

[0077] In this embodiment of the present invention, the UAV ground control system first accesses real-time data transmitted from the UAV's onboard sensors to extract flight braking-related physical control parameters, including the pitch rate limit (set to 10° / s), roll rate limit (set to 15° / s), yaw response delay (fixed to 0.25s), maximum climb rate (set to 4.5 m / s), maximum descent rate (set to 4.0 m / s), and maximum horizontal cruising speed (set to 16 m / s). Simultaneously, the task scheduling platform accesses the UAV's mission history database to obtain complete UAV delivery mission data, including mission number, origin and destination latitude and longitude coordinates, takeoff and estimated landing timestamps, mission route number, flight segment level (uniformly 50–120 m low-altitude), meteorological conditions within the flight path, payload weight (accurate to 0.1 kg), mission creation time, and delivery category (e.g., medical, documents, or emergency supplies). All parameters are exported in a unified CSV file format and saved in UTF-8 encoding for ease of subsequent program processing. First, the integrity of the aforementioned CSV-formatted drone delivery mission data file was verified field by field. The Pandas library in Python was used to read the data frame and perform column-by-column validation for null values ​​and unusual formats. For example, regular expression matching was used to ensure that the latitude and longitude fields were between [-180, 180]. The timestamp field was validated by matching the "YYYY-MM-DD hh:mm:ss" format. The weight field was checked to be a non-negative floating-point number and not exceed the maximum payload of the drone (5.0 kg). Data rows that failed these checks were recorded in the error log and deleted from the dataset. Next, the Z-score method was used to identify outliers in the weight field, with a Z-score threshold of ±3 to flag and remove extreme outliers. The mission timestamp field was uniformly formatted for the Eastern Time Zone (GMT+8). Time differences were calculated for multi-day delivery missions to identify scheduling logic errors. For example, records where the landing time was earlier than the takeoff time were considered logical errors and deleted. After cleaning, the data is once again verified for uniqueness to ensure that the mission numbers are not repeated. The fields retained include the mission number, starting and ending latitude and longitude, timestamp, delivery weight, and mission category, which are input into the next processing flow as the cleaned data of the drone delivery mission. Based on the cleaned mission time and latitude and longitude information, meteorological radar and ground observation network data from public meteorological websites are called to extract meteorological data for the corresponding time window within the delivery route coverage area. The extracted fields include ground wind speed (in m / s), ground wind direction (in degrees), average wind speed at 10m altitude, vertical wind shear gradient (in m / s / m), relative humidity (percentage), temperature (in degrees Celsius), air pressure (in hPa), rainfall intensity (in mm / h), and probability of thunderstorm occurrence (percentage).To ensure the spatial and temporal resolution of the extracted data matches the mission path, interpolation is used to insert sampling points at key locations along the path, with temporal frames calculated every 5 minutes and spatially interpolated at meteorological sampling points every 250 meters along the path, creating a composite mission-level meteorological data set. After data merging, consistency checks are performed to ensure that all missions have complete meteorological records for their flight paths and flight periods. All data is stored in a structured JSON format, with key-value fields clearly identifying the corresponding spatial location and timestamp. This data is then input into the next step as low-altitude delivery meteorological condition data. The multidimensional meteorological data matrix generated in step S13 is normalized using Min-Max normalization to map values ​​such as wind speed, humidity, and temperature to the [0, 1] range to unify the dimensions and facilitate subsequent feature processing. Subsequently, principal component analysis (PCA) is used to perform dimensionality reduction on this normalized multidimensional data, extracting the top n principal components (typically four) with cumulative contributions exceeding 95%. These components correspond to the principal dimensional factors of wind intensity, humidity-precipitation characteristics, temperature-pressure combination characteristics, and thunderstorm risk level. At the same time, a variable importance analysis method based on the entropy weight method was introduced to recalculate the relative weights of various meteorological factors on flight stability to verify the physical rationality of the PCA. On this basis, a composite feature, including the wind speed gradient change rate (Δv / Δt), wind direction deflection angle change rate, vertical wind shear amplitude, and thunderstorm forecast intensity index, was constructed as a refined representation of the high-dimensional raw meteorological data. Ultimately, these features were stored in a multidimensional array, with each array sample corresponding to a delivery mission. The array dimensions were [number of missions × number of features]. These features served as characteristic data for low-altitude delivery meteorological conditions, which were used for subsequent flight state simulation and attitude instability analysis.

[0078] Step S2 includes the following steps:

[0079] Step S21: simulating the UAV delivery flight state based on the UAV flight braking parameters and the UAV delivery mission data to obtain UAV delivery flight state simulation data;

[0080] Step S22: Obtain historical bird flock flight interference data for drone delivery; extract delivery load from the drone delivery mission data to obtain drone delivery load data;

[0081] Step S23: performing flight attitude instability simulation and quantification on the UAV delivery flight state simulation data based on the low-altitude delivery meteorological condition characteristic data and the UAV delivery payload data to obtain flight attitude instability quantification data;

[0082] Step S24: Based on the historical bird flock flight interference data and the flight attitude instability quantified data, the UAV flight trajectory avoidance behavior control learning is performed to obtain the flight trajectory avoidance behavior control data.

[0083] As an example of the present invention, refer to Figure 2 As shown, in this example, step S2 includes:

[0084] Step S21: simulating the UAV delivery flight state based on the UAV flight braking parameters and the UAV delivery mission data to obtain UAV delivery flight state simulation data;

[0085] In this embodiment of the present invention, based on the UAV flight braking parameters and delivery mission data acquired in step S1, a dynamics-based three-degree-of-freedom flight state simulation framework is established to simulate the UAV's flight state. This simulation framework uses the flight speed, pitch angle, roll angle, yaw angle, acceleration, and control steering control value specified in the flight delivery path as core state variables, and represents the UAV's spatial attitude using Euler angles. During the simulation, state iterations are performed with a time step of 0.5 seconds, and the change in the UAV's position in the three-dimensional spatial coordinate system is numerically integrated using the Runge-Kutta fourth-order integration method. Control inputs are constrained based on the UAV's maximum pitch change rate (10° / s) and maximum roll change rate (15° / s) to prevent state divergence. The simulation path is piecewise linearly interpolated based on the mission's starting and target longitudes and latitudes to generate a sequence of spatial path points. To account for the impact of payload on flight state, the total mass of each UAV is set based on its delivery payload, and the thrust distribution is adjusted by substituting it into Newton's equations of motion to ensure physical consistency of the simulation state. During the simulation process, the velocity vector, acceleration, attitude angle, geographic coordinates and propulsion force are recorded at each moment, and the final output is the UAV delivery flight status simulation data.

[0086] In another embodiment, when simulating a drone delivery flight based on drone flight braking parameters and drone delivery mission data, the braking torque in the drone flight braking parameters is first limited to a range of 0.2 N·m to 5.8 N·m, and the braking response time is controlled to a range of 0.1s to 0.5s. A three-dimensional virtual environment is established using a digital twin simulation platform. The starting and target delivery coordinates from the drone delivery mission data are set to a flight altitude range of 50m to 150m, and the delivery payload weight is limited to a range of 0.5kg to 10kg. The simulation system uses a numerical integration method with a discrete time step of 0.01s to gradually calculate the drone's position, velocity, and attitude angle changes in three-dimensional space. Using a six-degree-of-freedom motion equation solver, the system calculates the numerical changes in the drone's pitch, yaw, and roll angles in real time, and quantifies and records the lift, drag, and side force coefficients during flight. During the simulation, the propeller speed of the UAV is monitored and controlled within the range of 3000r / min to 8000r / min, and the battery voltage variation is limited to within 5%. The flight status parameters are sampled and stored at a frequency of 100 times per second through the real-time data acquisition module, and finally the UAV delivery flight status simulation data including time series, position coordinates, attitude angles, and power parameters are generated.

[0087] Step S22: Obtain historical bird flock flight interference data for drone delivery; extract delivery load from the drone delivery mission data to obtain drone delivery load data;

[0088] In this embodiment of the present invention, records of bird interference encountered during drone delivery operations are collected. This data is derived from multiple historical flight records, capturing bird flight patterns using sensors (such as radar, infrared, and acoustic sensors). This data typically includes information such as the bird's flight speed, altitude, trajectory, and density. To extract this interference data, data cleaning techniques are used to remove noise, and clustering algorithms (such as K-means clustering) are used to group the bird's flight paths into clusters, helping to identify different types of flight interference. Delivery payloads are then extracted from the drone delivery mission data. Delivery payload data includes the actual payloads carried by the drone during different delivery missions and is typically determined by the weight, volume, and packaging method of the delivered items. Through data analysis, a weighted average method is used to extract a reliable payload dataset from the historical mission data. The payload data extraction process takes into account both the maximum and actual payloads of each delivery mission to ensure that the impact of different payloads on flight performance is accurately reflected.

[0089] Step S23: performing flight attitude instability simulation and quantification on the UAV delivery flight state simulation data based on the low-altitude delivery meteorological condition characteristic data and the UAV delivery payload data to obtain flight attitude instability quantification data;

[0090] In this embodiment of the present invention, the low-altitude delivery meteorological condition characteristic data obtained in step S14 and the delivery payload data obtained in step S22 are jointly input into the flight attitude stability assessment module. This module, based on a multivariate perturbation analysis method, utilizes state sensitivity analysis to quantitatively calculate the stability of the UAV's simulated flight state under complex meteorological and payload conditions. During the analysis, perturbations are applied to characteristics such as the wind speed gradient change rate, wind direction mutation amplitude, and thunderstorm probability, introducing random fluctuations into the simulated state data (with the standard deviation set to ±5% of the original value). Response indicators such as the attitude angle change rate, heave rate change, and thrust offset angle are calculated at each time step. The cumulative attitude deviation integral and the second-order derivative of the attitude angle are used to calculate the probability of the UAV instability within a specific time period. A quantitative attitude instability vector is constructed by counting the frequency of violations within each time period. Finally, the attitude instability characteristics for each mission are output as a tensor structure of [time × attitude dimension × indicator item], including fields such as pitch angle violation frequency, roll angle fluctuation energy, instability duration, and maximum attitude deviation angle. This quantitative flight attitude instability data is used by the subsequent control learning module.

[0091] Step S24: Based on the historical bird flock flight interference data and the flight attitude instability quantified data, the UAV flight trajectory avoidance behavior control learning is performed to obtain the flight trajectory avoidance behavior control data.

[0092] In this embodiment of the present invention, a trajectory avoidance behavior learning method based on a heuristic strategy is used to deduce a UAV's obstacle avoidance strategy based on the historical bird flock flight interference matrix constructed in step S22 and the quantitative flight attitude instability data generated in step S23. The behavior control learning module uses the bird flock interference probability, instability risk level, and spatial path nodes as inputs. Control actions include avoidance angle adjustment, flight altitude correction, and speed descent rate adjustment. By setting the state space (bird flock density, attitude disturbance index, spatial position) and the action space (adjustment of path node direction angles, altitude layer switching amplitude, and speed adjustment amplitude), a greedy strategy-based path backtracking algorithm is used to perform multiple rounds of trajectory evaluation, with the objective function of minimizing the cumulative instability time and interference overlap rate. After each simulation round, the avoidance response action, operation amplitude, and corresponding reward value (including the instability risk reduction ratio and path deviation degree) of each trajectory point are recorded to generate a trajectory behavior dataset. The resulting flight trajectory avoidance behavior control data is indexed by the task number and contains fields such as the path adjustment operation sequence, operation corresponding timestamp, operation type identifier, and action intensity value, which are used in the trajectory control model construction step.

[0093] Step S23 includes the following steps:

[0094] Step S231: performing low-altitude turbulence structure identification on the low-altitude delivery meteorological condition characteristic data to obtain low-altitude turbulence structure data; performing wind direction shear phase plane state analysis on the low-altitude turbulence structure data to obtain wind direction shear phase plane state;

[0095] Step S232: calculating the average difference of asymmetric aerodynamic pressure of the UAV on the UAV delivery flight state simulation data based on the wind direction shear phase plane state to obtain the average difference of asymmetric aerodynamic pressure of the UAV;

[0096] Step S233: Calculating the rolling moment fluctuation of the UAV delivery flight state simulation data based on the UAV asymmetric aerodynamic pressure difference and the wind shear phase plane state to obtain the UAV flight rolling moment fluctuation data;

[0097] Step S234: performing a flight vibration multi-axis intensity regression analysis on the UAV delivery flight state simulation data based on the UAV flight rolling moment fluctuation data, the UAV asymmetric aerodynamic pressure average difference, and the UAV delivery payload data to obtain flight vibration multi-axis intensity regression data;

[0098] Step S235: performing flight attitude instability simulation and quantification based on the flight tremor multi-axis intensity regression data to obtain flight attitude instability quantification data.

[0099] In this embodiment of the present invention, the wind speed component, wind direction change rate, and pressure disturbance data from the low-altitude delivery meteorological condition characteristic data are first reconstructed into a spatial grid. The flight area is divided into a three-dimensional grid with a spatial resolution of 100m × 100m and an altitude interval of 50m. Three-dimensional wind speed vector data is extracted at each grid point. Combined with pressure gradient and temperature profile data, a structure function-based turbulence identification algorithm is used to extract the low-altitude turbulence structure. This algorithm calculates the first- and second-order structure functions of the wind speed time series to identify the nonlinear growth intervals of local wind speed variations and, based on these, define the turbulence intensity level. The identified turbulent regions are classified according to the eddy scale, which is set within a range of 5–30m. The peak wind speed gradient, turbulence intensity index, and duration within each region are recorded to form low-altitude turbulence structure data. This turbulence structure data is then subjected to wind direction shear phase plane state analysis. A two-dimensional phase plane is constructed based on the wind direction change rate and altitude gradient at each grid point within the turbulent region. In the phase plane, the horizontal axis is set to altitude difference, and the vertical axis is wind direction difference. Wind shear types are classified based on the phase plane trajectory change pattern: single peak, oscillatory, and sudden change. The maximum wind direction gradient and phase duration interval are used as key parameters to generate wind shear phase plane state data for subsequent aerodynamic analysis. Based on the wind shear phase plane state, asymmetric aerodynamic pressure difference calculations are performed on the generated UAV delivery flight simulation data. First, the attitude angle, velocity vector, and flight direction at each moment in the flight state data are mapped to a turbulent structure grid coordinate system, matching the wind shear regions traversed in the flight path. At the instant when the flight attitude and wind shear angle form an angle greater than 15°, aerodynamic pressure estimation is performed based on the difference in angle of attack between the windward and leeward surfaces of the wing. The pressure estimation process is based on the Bernoulli equation and, combined with the direction of the disturbance vector in the wind velocity field, calculates the dynamic pressure difference per unit area between the left and right wing surfaces. Using the wing area as a multiplication factor, the pressure difference between the left and right wings at each time step is integrated to calculate the asymmetric aerodynamic pressure difference. Finally, the asymmetric pressure sequence generated in each flight mission is normalized, and the maximum mean difference, mean, and standard deviation are recorded. The output is the asymmetric aerodynamic pressure mean difference data of the UAV, which serves as the input parameter for the roll moment calculation. Based on the asymmetric aerodynamic pressure mean difference and the wind shear phase plane state data obtained in step S232, combined with the speed, pitch angle, and roll angle in the flight state simulation, the roll moment fluctuation calculation is performed. The calculation method is based on the basic formula of aerodynamic torque, and combines the flight speed and the rate of change of the angle of attack to perform an instantaneous calculation of the wing surface torque difference caused by the asymmetric pressure. In the specific implementation, the pressure difference is mapped to the geometric center of the wing, and the resultant torque generated by the left and right wing resultant arms is calculated with the rotation axis as the reference. The change curve of the roll moment is recorded for each flight time step.To assess the intensity of the fluctuations, a time-frequency analysis of the torque variation sequence was further performed. Fast Fourier transform (FFT) was used to calculate its frequency spectrum and identify the dominant frequency and energy density of the roll torque fluctuations. The roll torque fluctuation results for each mission were organized into a four-dimensional matrix containing timestamps, torque values, frequency components, and fluctuation intensity factors. The final output of the UAV's flight roll torque fluctuation data includes the maximum torque peak, torque fluctuation frequency bandwidth, fluctuation duration, and average torque rate of change, which serves as a basis for modeling subsequent flight trajectory control behavior.

[0100] The UAV's flight roll moment fluctuation data, asymmetric aerodynamic pressure difference data, and UAV delivery payload data were aligned to a unified time series, and a 1 Hz time step was used to construct a multi-axis state analysis sample. The dominant frequency, fluctuation amplitude, and average torque change rate of the roll moment fluctuation data were used as one set of input variables. The time series mean, peak, and root mean square value of the asymmetric aerodynamic pressure difference data served as a second set of input variables. The UAV delivery payload data was encoded based on payload mass, load distribution location, and center of gravity offset. After standardization preprocessing these variables, a typical multivariate linear regression method was used to construct a multi-axis flight tremor intensity regression analysis model. In practice, a three-dimensional output vector was set to represent the tremor intensity index in the roll, pitch, and yaw directions. The input variable weight coefficients were determined using the least squares error criterion. A total of 100 flight records were selected as the training sample for modeling. Each record contained a complete aerodynamic pressure distribution, payload condition, and torque fluctuation response. After the regression model was trained, it was applied to the test data to predict tremor intensity for each flight mission. The output is represented by the three-axis tremor intensity factors at each moment, forming the flight tremor multi-axis intensity regression data. The data format includes a timestamp, axis identifier, tremor intensity value, and corresponding input variable identifier, which is used for subsequent instability simulation and quantification. Flight attitude instability is simulated and quantified based on the flight tremor multi-axis intensity regression data. First, the tremor intensity factor is mapped to the attitude angle change rate input in the UAV attitude dynamic model, and the disturbance injection conditions are set based on the three-axis tremor intensity. In the specific implementation, the UAV attitude evolution process is described by a system of ordinary differential equations. The input variables include pitch angle, roll angle, yaw angle, and their first-order derivatives. A disturbance term with the tremor intensity as the amplitude is added to the attitude angle at each control time step. The perturbation frequency is consistent with the dominant frequency in the tremor regression data, and the perturbation duration is set according to the torque fluctuation duration. The dynamic evolution of the attitude angle during flight is simulated through numerical integration. The fourth-order Runge-Kutta method is used to solve the integral and calculate the attitude angle change curve. During each simulation, the attitude angle offset, fluctuation amplitude, and rate of change of the attitude angle derivative per unit time are recorded to determine attitude stability. Statistical calculations are then performed on the simulation curves to extract key indicators such as maximum offset angle, average fluctuation amplitude, and frequency of instability, forming quantitative data on flight attitude instability. This data is stored in a three-dimensional structure, including the attitude angle value at each time point, instability determination flag, instability amplitude, recovery time, and attitude angle change rate. This data serves as direct input variables for subsequent avoidance behavior learning. The entire instability quantification process is simulated and verified in a real-world experimental environment by replaying flight data to ensure data consistency and engineering usability.

[0101] In another embodiment, when low-altitude distribution meteorological condition characteristic data is used to identify low-altitude turbulent structures, wind speed and direction data within a vertical range of 50m to 150m are obtained, the observation point spacing is set to 20m, and the data acquisition frequency is 10Hz. During the turbulent structure identification process, a spectral analysis method is used to segment the continuous wind speed data according to a time window length of 60s. The spectral characteristics within each time period are calculated using a fast Fourier transform algorithm to identify turbulent vortex structures with a frequency range of 0.01Hz to 10Hz. The turbulence intensity quantitative index is set as the ratio of the wind speed standard deviation to the average wind speed, and the ratio range is controlled between 0.05 and 0.35. When the ratio exceeds 0.2, it is determined to be moderate turbulence, and when it exceeds 0.3, it is determined to be strong turbulence. When performing wind direction shear phase plane state analysis on low-altitude turbulent structure data, a two-dimensional phase plane diagram is established with the horizontal wind speed component as the horizontal coordinate and the vertical wind speed component as the vertical coordinate, and the phase plane grid resolution is set to 0.1m / s. The trajectory of the wind vector within the phase plane was analyzed using a trajectory tracking algorithm, identifying three typical phase plane state patterns: spiral, focal, and saddle point. Wind shear intensity was quantified by calculating the wind direction difference between adjacent altitude layers. The altitude layer spacing was set at 10m, and the wind direction shear threshold was set to a significant shear when the wind direction change exceeded 15° for every 10m altitude difference. Wind direction shear phase plane state data was generated, including turbulent vortex scale, turbulence intensity level, and shear type identification. When calculating the asymmetric aerodynamic pressure difference of the drone based on the simulated drone delivery flight data based on the wind direction shear phase plane state, the drone body was first divided into six aerodynamic analysis regions: left and right wings, front and rear fuselage, and upper and lower surfaces. The number of pressure sampling points in each region was set to 100. The pressure distribution on the surface of each region was calculated using a computational fluid dynamics numerical solver, with a pressure calculation accuracy of 0.1Pa. During the asymmetric aerodynamic pressure analysis, the pressure difference at corresponding locations on the left and right wings was calculated, based on the longitudinal symmetry plane of the UAV. The pressure difference range was controlled between -50 Pa and +50 Pa. Combined with shear intensity data from the wind shear phase plane, the discrete pressure measurement points were expanded into a continuous pressure field distribution using a pressure interpolation algorithm. The interpolation method used was cubic spline interpolation, and the interpolation grid size was set to 0.05 m × 0.05 m. During the mean difference calculation, the mean pressure of the corresponding regions on the left and right wings and the front and rear fuselage was calculated. The total pressure in each region was calculated using a numerical integration method with an integration step of 0.01 m. The asymmetric aerodynamic pressure mean difference value was normalized by subtracting the right total pressure from the left total pressure and dividing it by the sum of the pressures on both sides. The normalized result was limited to the range of -0.2 to +0.2. A significant asymmetric pressure distribution was identified when the absolute value of the mean difference exceeded 0.1. The UAV asymmetric aerodynamic pressure mean difference data was generated, including pressure distribution maps for each region, asymmetry indicators, and pressure gradient change rates.To calculate the roll moment fluctuations in simulated UAV delivery flight data based on the asymmetric aerodynamic pressure difference and wind shear phase plane, a coordinate system was established with the center of mass of the UAV as the origin, the nose direction as the forward axis, the right wing direction as the lateral axis, and the vertical axis as the longitudinal axis. The roll moment calculation was performed by dividing the UAV into 20 equally spaced sections, each with a spacing of 0.1 m. The roll moment contribution of each section to the center of mass was calculated using the cross-sectional integration method. Combined with the asymmetric aerodynamic pressure difference data of the UAV, the pressure distribution of each section was converted into a roll moment component using the moment superposition principle. The maximum roll moment of a single section was limited to 0.5 N·m. During the analysis of the impact of wind shear, the shear intensity was quantified into three levels: weak, moderate, and strong, with the corresponding roll moment amplification factors set to 1.2, 1.5, and 1.8, respectively. The fluctuation calculation utilizes a time-domain analysis method. Roll moment data is sampled at 0.1-second intervals and the instantaneous rate of change of the roll moment is calculated using a sliding window algorithm with a window length of 1 second. The roll moment fluctuation amplitude is quantified by calculating the peak-to-peak value of the time series data, and the fluctuation frequency is statistically analyzed using a zero-crossing counting method with a statistical window length of 10 seconds. During the calculation process, the percentage of time when the roll moment exceeds the steady-state value within a range of ±20% is monitored. Significant fluctuation is determined when the percentage exceeds 30%. This generates UAV flight roll moment fluctuation data, including roll moment time series, fluctuation amplitude statistics, and frequency distribution characteristics. A multi-axis intensity regression analysis of flight vibration intensity is performed on UAV delivery flight simulation data based on the UAV flight roll moment fluctuation data, the UAV asymmetric aerodynamic pressure difference, and the UAV delivery payload data. A three-dimensional vibration analysis coordinate system is established, including the three rotational degrees of freedom: roll, pitch, and yaw. During the quantification of tremor intensity, roll moment fluctuation data were classified into five levels based on amplitude: Level 1 corresponds to fluctuation amplitudes below 0.1 N·m, Level 2 corresponds to 0.1-0.3 N·m, Level 3 corresponds to 0.3-0.6 N·m, Level 4 corresponds to 0.6-1.0 N·m, and Level 5 corresponds to 1.0 N·m or greater. Combined with the UAV's asymmetric aerodynamic pressure mean difference data, the tremor excitation intensity for each axis was calculated using a weighted average algorithm, with weight coefficients set to 0.5 for roll, 0.3 for pitch, and 0.2 for yaw. During the load impact analysis, the center of gravity offset in the UAV delivery payload data was used as a correction factor. Each 0.01m increase in center of gravity offset corresponds to a 5% increase in tremor intensity. The least squares method was used for parameter estimation in the multi-axis intensity regression analysis. The regression variables included the aerodynamic pressure mean difference, the payload distribution ratio, and the center of gravity offset, and the dependent variable was the tremor intensity for each axis. During the regression process, the confidence level was set to 95%, the correlation coefficient threshold was set to 0.8, and the corresponding data points were eliminated when the correlation coefficient was lower than the threshold.The analysis results establish a vibration intensity prediction relationship using a polynomial fitting method. The polynomial order is limited to third order, and the fitting accuracy requires a coefficient of determination greater than 0.9. This generates multi-axis vibration intensity regression data, including the vibration intensity values ​​for each axis, a regression coefficient matrix, and a prediction accuracy assessment. When quantifying flight attitude instability based on this multi-axis vibration regression data, an attitude stability evaluation index system is established, comprising three aspects: static stability, dynamic response, and vibration suppression capability. During the static stability quantification process, the vibration intensity for each axis in the multi-axis vibration regression data is used as a disturbance input. A linearization analysis method is used to calculate the steady-state deviation of the UAV's attitude angle. The steady-state deviation is limited to ±3° for pitch, ±5° for roll, and ±2° for yaw. The dynamic response is quantified using a step response analysis method. Given a unit step disturbance input, the three characteristic parameters of the attitude angle response, namely, rise time, settling time, and overshoot, are calculated. The rise time is controlled to be within 0.5, the settling time is limited to 2 seconds, and the overshoot does not exceed 10%. During the vibration suppression capability assessment process, the vibration intensity of each axis is decomposed according to the frequency domain. The low-frequency component corresponds to below 0.1Hz, the medium-frequency component corresponds to 0.1 to 1Hz, and the high-frequency component corresponds to above 1Hz. The transfer function analysis method is used to calculate the suppression ratio of the drone control system to vibrations of different frequencies. The suppression ratio is quantified by the ratio of output amplitude to input amplitude. When the ratio is less than 0.5, it is determined to be effectively suppressed. The weighted summation method is used for the quantitative comprehensive score of instability. The static stability weight is 0.4, the dynamic response weight is 0.4, and the vibration suppression capability weight is 0.2. The comprehensive score range is set from 0 to 100 points. When the score is less than 60 points, it is determined to be in an attitude instability state. The flight attitude instability quantitative data including stability level classification, instability risk assessment, and control difficulty coefficient are generated.

[0102] Step S234 includes the following steps:

[0103] Perform multi-axis rotational inertia incremental integration on the UAV's flight rolling moment fluctuation data to obtain multi-axis rotational inertia incremental integral data;

[0104] The UAV angular velocity progressive growth ratio is calculated based on the UAV flight rolling moment fluctuation data to obtain the angular velocity progressive growth ratio;

[0105] The axial yaw moment momentum deviation is quantified based on the average difference of the asymmetric aerodynamic pressure of the UAV to obtain the axial yaw moment momentum deviation data;

[0106] The UAV aerodynamic pressure imbalance is quantified based on the multi-axis rotational inertia incremental integral data, the UAV delivery load data and the axial yaw moment momentum deviation data to obtain the aerodynamic pressure imbalance quantification data;

[0107] Based on the quantified data of aerodynamic pressure imbalance, the ratio of angular velocity progressive increase and the multi-axis moment of inertia incremental integral data, the aerodynamic structure flutter inter-axis strength coupling simulation analysis is carried out to obtain the aerodynamic structure flutter inter-axis strength coupling data;

[0108] Based on the inter-axis intensity coupling data of aerodynamic structure flutter, a multi-axis intensity regression analysis of flight flutter is performed on the simulation data of UAV delivery flight state to obtain the multi-axis intensity regression data of flight flutter.

[0109] In the embodiment of the present invention, the first thing to be performed is the multi-axis moment of inertia increment integration operation of the UAV flight roll torque fluctuation data. This operation is based on the instantaneous torque fluctuation values ​​in the roll direction, pitch direction and yaw direction recorded during the flight. Based on the relationship between torque, moment of inertia and angular acceleration in Newton's second law, combined with the angular acceleration data under discrete time steps, the instantaneous moment of inertia increment in each axis is calculated by numerical integration. The trapezoidal integration method is adopted, and the time step is set to 0.05s. The actual torque fluctuation value and the instantaneous angular acceleration value at each moment are combined to calculate the moment of inertia changes of the roll axis, pitch axis and yaw axis respectively. In actual operation, taking a certain type of six-rotor UAV as an example, the basic moment of inertia parameters are the roll axis and the pitch axis, respectively. 0.0024kg·m² , pitch axis 0.0022 kg·m² , yaw axis 0.0041 kg·m² When encountering sudden wind shear during flight, the peak value of torque fluctuation reaches 0.08 N·m , the corresponding angular acceleration reaches 35 rad / s² After trapezoidal integration, the increment of the moment of inertia of the roll axis within a 0.5s period is The incremental moment of inertia results for each axis are integrated into a three-dimensional time series array, forming multi-axis incremental moment of inertia integral data for subsequent tremor intensity analysis. The roll moment fluctuation data is then used to calculate the angular velocity progressive growth ratio. This calculation step uses the angular velocity data in the time series as input, extracts the starting value and peak value of the angular velocity increase segment in each time window, and calculates the growth ratio based on this. A sliding window mechanism is used, with a window length of 1s and a step size of 0.1s. The first-order derivative of the angular velocity curve within each window is calculated to determine the angular velocity growth segment. The velocity values ​​before and after the growth are recorded, and the degree of growth is expressed as a ratio. In the specific experiment, a roll disturbance event from a typical flight mission was selected. In this event, the angular velocity increased from 12 rad / s to 48 rad / s in 0.3s. The calculated growth ratio for this segment was 4.0. To ensure the stability of the growth ratio calculation, the calculation results of each time window are smoothed at three points, and the maximum growth ratio value and its corresponding time point for each window are output. The final result is output as a two-dimensional table, containing fields such as time index, initial angular velocity value, peak value, and ratio, which constitute the angular velocity progressive acceleration ratio data. The axial yaw moment momentum deviation is quantified based on the average difference in asymmetric aerodynamic pressure of the UAV. First, the total aerodynamic torque of the left and right wings is calculated based on the measured aerodynamic pressure data of the UAV in the direction of the left and right symmetric axes, and the wing surface pressure data is torqued using the integral area method. The aerodynamic pressure difference on the left and right sides is applied to the yaw axis in the positive and negative directions, forming an asymmetric torque distribution. Then, combined with the current yaw angular velocity, the corresponding yaw momentum deviation is derived. The momentum deviation is expressed as the change relative to the steady-state yaw angular momentum. The steady-state momentum calculation uses the product of the yaw angular velocity and inertia under calm wind conditions as the benchmark. In experimental verification, under conditions of a wind speed disturbance of 6 m / s and a sudden change in angle of attack of 2°, the pressure difference between the left and right wing surfaces of the test aircraft reached 15 N, corresponding to a yaw moment of 0.12 N·m, resulting in a momentum deviation of 0.007 kg·m²·rad / s. This momentum deviation data is expressed in a single-axis quantized format, with timestamp, yaw moment difference, and angular momentum increment as the primary fields. This constitutes the axial yaw moment momentum deviation data, providing quantitative input for subsequent attitude stability analysis and control behavior learning. In the specific implementation of the UAV aerodynamic pressure deviation quantification based on multi-axis moment of inertia incremental integral data, UAV delivery payload data, and axial yaw moment momentum deviation data, the first step is to extract the rate of change of moment of inertia per unit time for the roll, pitch, and yaw axes based on the multi-axis moment of inertia incremental integral data to determine the contribution of each inertia change to the dynamic response of the overall aircraft structure. Combined with the actual load data of the UAV, for example, in a multi-rotor delivery mission, a delivery package with a mass of 2.8 kg is carried. The load value causes the asymmetric increment of the pitch axis and yaw axis inertia to reach and , this asymmetric inertia will change the aerodynamic response timing and load transfer mechanism of each axis during flight. Subsequently, combined with the momentum deviation peak under different yaw angular rate conditions in the axial yaw moment momentum deviation data, the torque applied to each wing surface is instantaneously decoupled based on the known pressure distribution coefficient, and the time series of the aerodynamic imbalance force is calculated. It is then integrated along the axial direction to obtain the final pressure imbalance quantification index. This index is measured in Newtons and outputs the net imbalance force components in the roll, pitch, and yaw directions, forming a complete aerodynamic pressure imbalance quantification data, which provides input for subsequent vibration intensity analysis. When performing the inter-axis strength coupling simulation analysis of aerodynamic structure flutter, the aerodynamic pressure imbalance quantification data is first used as the load boundary condition, and the velocity increment perturbation is applied to each axis in combination with the angular velocity progressive increase ratio data, thereby constructing the initial conditions for transient stress driving. The inertia change rate from the multi-axis moment of inertia incremental integral data is then introduced as a control factor for inertial coupling. Using an inter-axis disturbance coupling mapping mechanism, the inertia-torque interaction between the roll, pitch, and yaw axes is constructed as a tensor coupling mapping matrix. In an experiment, using a 0.6 m wingspan UAV subjected to a wind speed disturbance of 7.5 m / s, cross-correlation analysis of the roll and pitch cross-transient pressure response sequences measured by sensors revealed an inter-axis coupling coefficient of 0.67, indicating strong coupling between the force-inertia responses of the different axes. This process uses a multidimensional time series differentiation method to extract characteristic frequencies, calculate the dominant frequency components and intermodulation frequencies of each axis' response, and identify the main flutter modes and the influence range of inter-axis coupling. Ultimately, the inter-axis coupling strength data for aerodynamic structure flutter, primarily roll-pitch, pitch-yaw, and roll-yaw, is output in N / m²·rad. A multi-axis flutter intensity regression analysis of UAV delivery flight simulation data based on inter-axis flutter intensity coupling data from aerodynamic structures was performed. First, the core dynamic feature vectors of the flight state simulation data, such as flight attitude, velocity, and angular velocity, were matched with the flutter coupling data through a stepwise regression method. A multivariate linear regression function incorporating the principal and mutual axis coupling factors was constructed using the stepwise regression method. In practice, data was sampled at 20 Hz during the experimental flight process, and the dynamic flight attitude variation trend within each cycle was extracted. Using angular acceleration as the dependent variable and inter-axis coupling stress terms as the independent variable, the flutter response in the roll direction caused by pitch axis coupling was calculated to have a 1.9-fold increase. A first-order autocorrelation test of the regression residuals confirmed the convergence and accuracy stability of the regression process. The relationship between the flutter response intensity among the three axes was quantified within a 95% confidence interval, ultimately generating multi-axis flight flutter intensity regression data. This data is used to characterize the intensity variation trend and load impact of each axis' flutter behavior during dynamic flight missions, directly reflecting the dynamic adjustment of the flight control trajectory control system input.

[0110] Step S24 includes the following steps:

[0111] Step S241: Mark the multi-target bird flock movement trajectory of the historical bird flock flight interference data to obtain the multi-target bird flock movement trajectory;

[0112] Step S242: performing approach relative orientation clustering learning on the historical bird flock flight interference data based on the multi-target historical movement trajectories of the bird flock to obtain bird flock approach relative orientation clustering data;

[0113] Step S243: performing a quantitative assessment of the airflow disturbance caused by the approach of the bird flock on the historical bird flock flight interference data according to the historical movement trajectories of the bird flock multi-target, thereby obtaining quantitative data of the airflow disturbance caused by the approach of the bird flock;

[0114] Step S244: identifying the critical state of the UAV's flight attitude instability based on the quantitative data of the airflow disturbance caused by the approaching flock of birds and the clustered data of the relative orientation of the approaching flock of birds, and obtaining the critical state of the flight attitude instability;

[0115] Step S245: Based on the relative orientation clustering data of the approaching flock of birds and the critical state of flight attitude instability, the UAV flight trajectory avoidance behavior control learning is performed to obtain flight trajectory avoidance behavior control data.

[0116] In this embodiment of the present invention, during the specific implementation of labeling the multi-target bird flock movement trajectory from historical bird flock flight interference data to obtain the multi-target historical bird flock movement trajectory, optical image data and radar target echo data from a real low-altitude flight area were first selected. The time range was set to March to May 2023, with a daily sampling period of 6:00 to 18:00, and a sampling frequency of 5 frames / s. The image resolution was set to 1920×1080 pixels, the radar echo data sampling radius was 3 km, and the spatial resolution was 2 m. Target state prediction was performed using a multi-target Kalman filtering algorithm based on Euclidean geometric centers. Static background interference signals were eliminated using a Gaussian mixture background modeling method. Identifiable bird bodies within each frame were assigned motion trajectories, and the inter-frame difference method was used to track flight trajectory points. In the presence of occlusion or sudden changes in flight speed, nearest neighbor interpolation was used to compensate for lost frame trajectory points to ensure the integrity and stability of continuous target labeling. The final output is a collection of two- or three-dimensional trajectory points indexed by timestamps, recording the spatial position sequence, velocity vector, and flight direction angle of each bird, forming a historical multi-target bird flock movement trajectory dataset. In the specific implementation of relative azimuth clustering learning based on the historical multi-target bird flock flight interference data, the historical trajectory data is first partitioned according to spatial azimuth. Using the drone's own GPS heading angle at each moment as the reference zero point, the relative azimuth angle, relative distance, and flight velocity component of each bird in this heading coordinate system are calculated. Using the average relative azimuth angle change rate (in degrees per second) and the relative velocity distribution density as similarity indicators, the trajectory point cloud is unsupervisedly classified using the Density Peaks Clustering algorithm to extract typical bird flock interference azimuth patterns. Taking 2,456 bird trajectories recorded during a UAV flight experiment as an example, by clustering the bird's motion characteristics in the directions of 0–60° in front of the UAV, 60–120° to the side, and 120–180° behind the UAV, seven typical approaches were identified, including frontal collisions, right-front oblique parallel flights, and left-rearward maneuvers. The outputs include key characteristic values ​​for each cluster center, such as the average azimuth angle, bird density, average relative velocity, and frequency of occurrence. A cluster label matrix was also generated for subsequent airflow disturbance modeling. In the implementation of quantitatively assessing the airflow disturbance caused by bird approach based on historical multi-target bird flight disturbance data, the clustered relative azimuth trajectories were first used as input. For each trajectory type, the velocity gradient change and aerodynamic disturbance coefficient generated by the disturbance source bird in the airflow field around the UAV were calculated. Using the empirical fluid disturbance function, the typical wingspan of the bird is set to 0.8 m and the average flight speed is set to 14.2 m / s, and the disturbance propagation area is constructed in various typical directions.Using a spatially discrete grid, a 3D grid was created within a range of 0–5 m from the drone, with a cell volume of 0.125 m³. The local airflow velocity variations caused by multiple birds flying within each time period were spatially superimposed to extract a vortex intensity distribution map. For example, a group of birds flying in parallel from the left front of the drone in a specific experimental scenario exhibited a maximum velocity perturbation difference of 3.7 m / s at a distance of 2.3 m from the drone's fuselage, 30° to the left front, resulting in a local lift fluctuation exceeding 9.5 N. After completing the perturbation statistics at all evaluation points, the three metrics of perturbation peak intensity, duration, and frequency were integrated to generate quantitative data on the airflow perturbation caused by the approaching birds. This data serves as a key parameter for predicting drone flight stability and adjusting its avoidance trajectory.

[0117] In the specific implementation of critical state identification for UAV flight attitude instability based on quantitative data of flight attitude instability and clustered data of relative orientation of bird flocks approaching, the quantitative data of flight attitude instability is first segmented by timestamp, with attitude angle changes, angular velocity fluctuations, and lift vector deviations within every 5 seconds considered as an analysis period. Within each analysis period, periods with pitch angle fluctuations exceeding 6.2°, roll angle change rates exceeding 38.4° / s, and yaw angle fluctuations exceeding 7.1° are selected as candidate instability signals. Using the disturbance intensity (in m / s²) and disturbance duration (in seconds) in different directions from the quantitative data of bird flock approaching airflow disturbances as reference variables, a state threshold map based on three-dimensional vector correlation is constructed to correlate high-intensity disturbances with sudden attitude changes. Based on the typical interference patterns represented by cluster labels in the relative orientation clustering data of approaching bird flocks, the projected disturbance magnitude of the corresponding interference pattern on the principal force direction of the UAV's airframe structure is calculated. When this disturbance magnitude exceeds the critical disturbance threshold (for example, the rolling moment caused by lifting surface deflection exceeds 1.7 N·m) multiple times within two seconds, a critical flight attitude instability state is identified. In the experimental sample, for the left front-facing converging bird flock interference pattern, at a low-altitude flight speed of 15 m / s, the UAV's roll angle experienced three consecutive sudden changes exceeding 12.3° within 1.2 seconds. The corresponding disturbance quantification data in this direction exhibited a disturbance intensity of 4.6 m / s² and a duration of 1.8 seconds, meeting the criticality criteria. Therefore, this state is considered a critical flight attitude instability state. The final output is a critical state trigger record indexed by time point, containing parameters such as the trigger location coordinates, disturbance source direction, attitude change rate, and interference pattern label. In the specific implementation of learning UAV flight trajectory avoidance behavior control based on relative position clustering data of approaching bird flocks and critical flight attitude instability states, the identified critical instability states are first used as input labels. A sample set of control behavior learning is constructed by retrospectively analyzing the flight control commands, external disturbance parameters, and relative position change sequences of the bird flock within three seconds before the state is triggered. Using state-action pairs as the basic unit, the attitude change response of the UAV under specific relative position disturbance patterns is combined with the flight control system output commands (including throttle percentage and pitch / roll / yaw control amounts) for analysis. The incremental time difference learning algorithm (TD-Lambda) is used to learn the trajectory adjustment strategy for each typical disturbance pattern in the sample set, optimizing the response delay and control smoothness of the adjustment path in the initial disturbance phase. In the actual data set, a typical case was a flock of birds approaching from the right rear at a flight altitude of approximately 90 meters. The original flight trajectory and final obstacle avoidance trajectory of the drone in this case were extracted. The trajectory correction angle was calculated to be 26.8°, and the avoidance command was issued 1.4 seconds before the disturbance was triggered, successfully achieving the avoidance operation.The final learning results are output in tabular form. Each cluster direction corresponds to a standard trajectory avoidance strategy, including the trajectory deflection angle (in degrees), the initial avoidance time advance (in seconds), the manipulator output correction value (expressed as the increase or decrease of the PWM signal), and the corresponding disturbance scenario parameters. These are used for subsequent online trajectory avoidance module calls in the flight control system.

[0118] Step S244 includes the following steps:

[0119] The multi-directional flapping wing rotating vortex intensity data were obtained by analyzing the multi-directional flapping wing rotating vortex intensity data based on the quantitative data of airflow disturbance and the relative orientation clustering data of the approaching birds.

[0120] The flapping rhythm vortex transient mutation data are identified based on the multi-directional flapping wing rotation vortex intensity data to obtain the flapping rhythm vortex transient mutation data;

[0121] The flight control surface deflection stall index of the UAV is evaluated based on the transient mutation data of the flapping rhythm vortex and the flight attitude instability quantitative data, and the flight control surface deflection stall index is obtained.

[0122] The critical state of UAV flight attitude instability is identified based on the flight attitude instability quantitative data according to the flight control surface deflection stall index, and the critical state of flight attitude instability is obtained.

[0123] In an embodiment of the present invention, when performing a multi-directional analysis of the intensity of rotating vortices during flapping flight based on the quantitative data of airflow disturbances caused by the approach of a flock of birds and the clustering data of the relative orientations of the flock of birds, a circular monitoring area with a radius of 500 meters centered on the UAV is established, and the monitoring area is divided into 36 sectors according to the azimuth angle, with each sector corresponding to a 10° azimuth angle range. A Doppler radar array is used in the flapping flight vortex detection process, with the radar operating frequency set to 9.4 GHz and the detection accuracy reaching a speed resolution of 0.2 m / s. Rotating vortices are identified by analyzing the air disturbance pattern generated by the flapping wings of the flock of birds, with the flapping frequency range set to 3 times / s to 8 times / s, and the vortex diameter generated by a single flapping wing ranging from 0.5 m to 2 m. The vortex intensity is quantified using a vortex calculation method, by measuring the velocity distribution on a circular path around the vortex center, with the circular path radius set to half the vortex diameter, and the number of velocity sampling points being 24 evenly distributed points. During the intensity analysis, vortices are classified as either clockwise or counterclockwise, with the angular velocity range controlled between 0.5 rad / s and 3.0 rad / s. The multi-directional vortex superposition effect is calculated using a vector synthesis method. When multiple vortices exist within the same azimuth range, an intensity-weighted average algorithm is used to calculate the composite vortex intensity, with the weight coefficient being inversely proportional to the vortex's distance from the drone. The vortex duration statistics range is set to 0.2s to 2.5s, with vortices lasting longer than 1s considered stable. Multi-directional flapping vortex intensity data is generated, including vortex intensity distribution at each azimuth, rotation direction identification, and duration statistics.

[0124] To identify transient changes in flapping rhythm vortex intensity from multi-directional flapping wing data, a time-frequency analysis system was established. A short-time Fourier transform algorithm was used to perform spectral analysis on the vortex intensity time series, with a time window length of 0.5 seconds and a window overlap of 50%. During the flapping rhythm feature extraction process, the flapping frequency of the bird flocks was classified into slow rhythms of 3 to 4 times / s, medium rhythms of 4 to 6 times / s, and fast rhythms of 6 to 8 times / s. The corresponding vortex intensity change periods for each rhythm were 0.25 to 0.33 seconds, 0.17 to 0.25 seconds, and 0.12 to 0.17 seconds, respectively. Transient changes were detected using an edge detection algorithm, calculating the first- and second-order differences of the vortex intensity data. The first-order difference threshold was set at 0.5 units of intensity change / s, and the second-order difference threshold was set at 0.2 units of acceleration / s². During the mutation identification process, the amplitude of the vortex intensity change within 0.1 seconds is monitored. A transient mutation event is identified when the amplitude exceeds 30% of the average intensity. Rhythmic vortex pattern analysis is implemented using an autocorrelation function algorithm. The calculation time delay range is set to 0 to 2 seconds, and the correlation coefficient threshold is set to 0.7. A rhythmic pattern is confirmed when the delay time corresponds to an integer multiple of the flapping period and the correlation coefficient exceeds the threshold. Mutation types are classified into three basic types: intensity jump, intensity drop, and frequency mutation. The jump amplitude threshold is 50%, the drop amplitude threshold is 40%, and the frequency mutation threshold is 1 flapping frequency change / s. Flapping rhythmic vortex transient mutation data is generated, including the mutation event timestamp, mutation type identifier, and mutation amplitude quantification. Based on the flapping rhythmic vortex transient mutation data, a control surface response analysis system is established to evaluate the UAV flight control surface stall index using quantitative flight attitude instability data. The system includes three main control surfaces: elevator, rudder, and aileron. The deflection angle range of each control surface is limited to -30° to +30°. The stall index calculation process first extracts the peak intensity of the transient mutation data from the flapping rhythm vortex. The mutation intensity is graded as 0.1 to 0.3 units for weak mutations, 0.3 to 0.6 units for medium mutations, and 0.6 to 1 unit for strong mutations. The rudder deflection response is analyzed using a step response test method. The amplitude of a standard step signal is set at a 10° deflection angle. The response time from the initial position to the target position is measured. The normal response time range is 0.1s to 0.3s. The rudder deflection demand is calculated using a numerical differentiation algorithm, combining the attitude angle change rate from the quantitative flight attitude instability data. The deflection demand calculation uses a proportional-integral-derivative control algorithm with a proportional coefficient of 2.0, an integral coefficient of 1.5, and a differential coefficient of 0.5. Stall detection is achieved by monitoring the rudder deflection angular velocity and angular acceleration. The angular velocity threshold is set at 90° / s² and the angular acceleration threshold is set at 300° / s². When the actual deflection parameters exceed the threshold range, the rudder response is considered saturated.Stall index quantification utilizes a dimensionless method, dividing the actual deflection demand by the maximum deflection capacity. The ratio is controlled between 0 and 1.5. A ratio exceeding 1 is considered a mild stall, exceeding 1.2 is considered a moderate stall, and exceeding 1.4 is considered a severe stall. This generates flight control surface deflection stall index data, which includes the stall level, response saturation, and control margin for each control surface. To identify critical states of UAV flight attitude instability using flight control surface deflection stall index data, a multidimensional state space analysis framework is established. The state variables include pitch angle, roll angle, yaw angle, and their corresponding angular velocities, and the state space is set to six dimensions. The critical state identification process first determines the safe flight envelope boundaries: the pitch angle safety range is set to -20° to +20°, the roll angle safety range is limited to -30° to +30°, and the yaw angle change rate is controlled within the range of -10° / s to +10° / s. Combined with flight control surface deflection stall index data, a state boundary approximation algorithm is used to identify critical state points. When the stall index of any control surface exceeds 0.8 and the corresponding attitude angle approaches 80% of the safety margin, the system is deemed to have entered the critical state warning zone. Instability criticality is determined using a multi-index comprehensive assessment method, including a weighted scoring system with a weight of 0.4 for attitude angle deviation, 0.3 for angular velocity excess, and 0.3 for control surface stall severity. During state identification, a sliding time window analysis is used with a window length of 3 seconds. A critical state is confirmed when the combined score exceeds 0.7 in two consecutive time windows. Critical states are categorized into three levels: mild critical (with a score of 0.7 to 0.8), moderate critical (with a score of 0.8 to 0.9), and severe critical (with a score of 0.9 to 1.0). During the identification process, state trends are monitored, and linear regression analysis is used to predict the state trajectory over the next 5 seconds. The regression fit is required to reach a minimum of 0.85. When the predicted trajectory shows that the attitude angle will exceed the safety boundary within 10 seconds, the critical state alarm mechanism is activated to generate flight attitude instability critical state data including critical state level, remaining safety time, and state development trend prediction.

[0125] Step S3 includes the following steps:

[0126] Step S31: performing convolution processing on the flight trajectory avoidance behavior control data to obtain flight trajectory avoidance behavior convolution data;

[0127] Step S32: performing iterative incremental learning on the flight trajectory avoidance behavior convolution data to obtain flight trajectory avoidance behavior iterative data;

[0128] Step S33: constructing a trajectory control model based on the flight trajectory avoidance behavior iterative data based on the policy gradient algorithm to obtain a UAV flight trajectory control model;

[0129] Step S34: Send the UAV flight trajectory control model to the UAV control center to perform UAV trajectory control.

[0130] As an example of the present invention, refer to Figure 3 As shown, in this example, step S3 includes:

[0131] Step S31: performing convolution processing on the flight trajectory avoidance behavior control data to obtain flight trajectory avoidance behavior convolution data;

[0132] In an embodiment of the present invention, a three-dimensional tensor structure is first constructed based on the acquired flight trajectory avoidance behavior control data, with the time series as the first dimension, the lateral heading offset angle as the second dimension, and the longitudinal pitch height change as the third dimension. The time interval between each frame of data is 50ms, and the control command value is quantized into the interval [-1, 1]. A one-dimensional convolution kernel is used to extract features from the time series. The convolution kernel length is set to 5 and the step size is 1. The ReLU activation function is used to handle negative value truncation and extract the trajectory offset change trend within the sliding window. Then, a two-dimensional convolution kernel is used to perform a spatial joint analysis of the coordinated changes in lateral and longitudinal offsets. The convolution kernel size is set to 3×3, and the padding method is edge replication. After convolution, a maximum pooling layer is used to downsample the feature map, and the pooling window size is 2×2. The output of the convolution layer is flattened to form a feature vector used to represent the flight trajectory avoidance behavior convolution data. The input trajectory data in the experiment is 148 seconds of data collected by a 2.5 kg quadrotor drone crossing a path between urban buildings. The total sample size is 2960 frames. After convolution processing, an output dataset with 64-dimensional feature dimensions is formed.

[0133] Step S32: performing iterative incremental learning on the flight trajectory avoidance behavior convolution data to obtain flight trajectory avoidance behavior iterative data;

[0134] In the embodiment of the present invention, in the specific implementation process of iterative incremental learning of flight trajectory avoidance behavior convolution data, an online iterative incremental training method is adopted. Each time a set of trajectory samples after convolution processing is received, the gradient is updated based on the last learning weight. The optimizer used is the Adam optimizer, the initial learning rate is set to 0.001, and the gradient decay coefficient is set to 0.001. and They are set to 0.9 and 0.999 respectively, and the L2 norm loss function with a regularization term is used to control the error convergence of the trajectory control output and the target trajectory offset difference. The samples are labeled for consistency before being added to the incremental training. If the current trajectory offset value deviates from the historical learned control behavior by more than 0.15 unit lengths, it is marked as a "pseudo-abnormal" sample and removed from the training to improve the robustness of the model. Each iteration cycle is set to 200 gradient steps. After the cycle, all learned sample sets are re-trained with gradient replay to prevent catastrophic forgetting. The experimental training data comes from flight control data after approaching a flock of birds during low-altitude flight in the city. The training is stopped when the trajectory error drops from the initial 2.7m to an average of 0.83m. The final output is the iterative data of the flight trajectory avoidance behavior.

[0135] Step S33: constructing a trajectory control model based on the flight trajectory avoidance behavior iterative data based on the policy gradient algorithm to obtain a UAV flight trajectory control model;

[0136] In an embodiment of the present invention, during the specific implementation of a trajectory control model based on iterative data of flight trajectory avoidance behavior using a policy gradient algorithm, the policy gradient method from reinforcement learning was used to parameterize the flight path control strategy. A state-action space was constructed, where the state vector contained 12 features: the current flight attitude angle, inertial velocity, radar range scan results, and the surrounding airflow disturbance intensity. The action outputs included thrust control commands for up / down, left / right, and forward / backward offsets. The policy function used was a Gaussian distribution parameterized policy, employing the REINFORCE policy gradient method. A trajectory deviation reward function obtained after each flight round was used for reward-weighted updates. The reward function targets minimal track deviation, with smaller average trajectory deviation values ​​resulting in higher rewards. An experience replay mechanism was used to reinforce and evaluate the action sequences and state transition trajectories over nearly 500 iterations, with cumulative gradient updates. During the experiment, iterative control behavior data generated by five drones during a 120-second urban building crossing mission was selected as the training data source. After 1,320 policy iterations, the final drone flight trajectory control model was constructed. The model output had a dimension of 3, corresponding to the three-axis thrust vector offset.

[0137] Step S34: Send the UAV flight trajectory control model to the UAV control center to perform UAV trajectory control.

[0138] In this embodiment of the present invention, the specific implementation process of sending a UAV flight trajectory control model to a UAV control center for trajectory control is as follows: the constructed flight trajectory control model is first exported in tensor format to a standard binary file, using a Protobuf structure to define the model parameter sequence. The model, with a total size of 18.6 MB, is then transmitted to the UAV control center via an Ethernet connection. A SHA-256 hash integrity check is performed before transmission to ensure accurate reception by the control center. The control center embeds a control distribution module based on the PX4 flight control system. After the model is imported, it is mapped to the flight control output decision logic through the control interface. The current state is periodically read every 10ms and the model is called to generate the next control thrust vector. During actual deployment verification, no model call delays or communication delays were observed during the control center's control command cycle. The control module successfully implemented dynamic trajectory control behavior for three UAVs in an urban high-altitude test platform through the control link.

[0139] Step S33 includes the following steps:

[0140] Step S331: dividing the flight trajectory avoidance behavior iteration data into a training set and a test set to obtain a flight trajectory avoidance behavior training set and a flight trajectory avoidance behavior test set respectively;

[0141] Step S332: performing clustering processing on the flight trajectory avoidance behavior training set to obtain a flight trajectory avoidance behavior cluster training set;

[0142] Step S333: performing feature sampling on the flight trajectory avoidance behavior cluster training set to obtain a trajectory avoidance behavior cluster sampling training set;

[0143] Step S334: constructing an initial trajectory control model for the trajectory avoidance behavior cluster sampling training set based on the policy gradient algorithm to obtain an initial UAV flight trajectory control model;

[0144] Step S335: Input the trajectory avoidance behavior test set into the initial UAV flight trajectory control model for model testing to obtain the UAV flight trajectory control model.

[0145] In this embodiment of the present invention, during the process of partitioning the iterative flight trajectory avoidance behavior data into training and test sets, the collected iterative flight trajectory avoidance behavior data was first numbered and integrated in chronological order. The total data set consisted of 35,240 data sets, each containing three dimensions: lateral offset thrust command, longitudinal attitude adjustment angle, and current velocity vector change. A proportional partitioning method was used to divide the data into training and test sets in an 8:2 ratio, with the training set containing 28,192 data sets and the test set containing 7,048 data sets. To avoid training bias caused by uneven data distribution, the data was stratified according to flight state stability labels before partitioning to ensure a consistent proportion of each state in the two data sets. After partitioning, each data set was subjected to minimum-maximum normalization, and all control variables were uniformly scaled to the range [0, 1] to facilitate stable convergence of the subsequent algorithm. During the clustering process of the flight trajectory avoidance behavior training set, an unsupervised cluster analysis of the training set data was performed using the density-based DBSCAN algorithm. The algorithm's initial parameters were set to a minimum sample number, MinPts, of 30, and a neighborhood radius, ε, of 0.12. Euclidean distances were first calculated for all sample points to construct a sample adjacency matrix. Density neighborhoods were then determined for each point, and density-reachable samples were grouped into the same cluster. Five cluster centers were identified, each representing a pattern of flight avoidance behavior under varying interference intensities. After clustering, each sample was labeled with a corresponding cluster label, forming a flight trajectory avoidance behavior cluster training set. Clusters were numbered C0 to C4, corresponding to 6432, 5217, 6146, 5821, and 6576 sample groups, respectively. During feature sampling for the flight trajectory avoidance behavior cluster training set, based on the principle of balanced sampling, an equal proportion of samples were selected from each cluster to construct a highly representative training set. 1200 feature sequences were randomly selected from each cluster to construct the trajectory avoidance behavior cluster training set, totaling 6000 sequences. After sampling, each sample group is rearranged in time sequence and timestamps are recalibrated according to the control moment to ensure continuity and temporal logic. Each sample group retains the original three-dimensional control parameters and is assigned a corresponding cluster label, forming an input sample set with cluster category identification. Sampling is performed using a random walk algorithm, with a sampling step size of 15 frames to avoid high duplication of adjacent sampled data. During the implementation of the policy gradient algorithm to construct the initial trajectory control model for the trajectory avoidance behavior cluster sampling training set, a policy gradient-based reinforcement learning framework was used to train the model on the sampled data. A state-action-reward triple is constructed, where the state vector is a 12-dimensional feature consisting of the current trajectory position, the control behavior within the past three frames, the relative disturbance direction, and the flight speed. The action space is the adjustment of the thrust vector, and the reward function is the negative squared difference of the trajectory deviation.The policy function used was a parameterized Gaussian distribution policy, and the REINFORCE algorithm was used for parameter updates. During training, 64 samples were used per batch, with a total of 1500 training rounds. The initial learning rate was 0.0008, and a dynamic learning rate decay was performed every 250 rounds with a decay coefficient of 0.85. During training, the action probability distribution was compared with the actual trajectory control results, and the policy function parameters were updated based on the expected reward. The resulting output control policy became the initial UAV flight trajectory control model, which has 4.2 million parameters and consists of a three-layer policy network structure and a one-layer value function network structure. To test the initial UAV flight trajectory control model by inputting the flight trajectory avoidance behavior test set into the model, the constructed initial UAV flight trajectory control model was first loaded into the simulation platform, using MATLAB Simulink as the base environment, with a simulation frequency set to 100 Hz. The test data was input into the model in chronological order. Each set of test data went through three stages: state preprocessing, model inference, and control action generation, ultimately outputting a thrust control vector. The output actions were compared with the actual control commands in the test data labels to determine trajectory deviation. Evaluation metrics such as average trajectory error, maximum attitude deviation, and control stability time were calculated. The test data included eight flight disturbance conditions, including crosswinds, approaching birds, and thermal disturbances. Across all 7,048 test data sets, the average Euclidean distance between the model's output control actions and the labeled control commands was 0.73, with a standard deviation of 0.28, demonstrating the model's stable control strategy performance. Finally, the validated model parameters were solidified into a UAV flight trajectory control model for deployment on the flight control platform.

[0146] The present invention also provides a UAV trajectory control system for executing the UAV trajectory control method described above, the UAV trajectory control system comprising:

[0147] The low-altitude delivery meteorological condition extraction module is used to obtain the UAV flight braking parameters and UAV delivery mission data; the low-altitude delivery meteorological condition extraction module is used to extract the low-altitude delivery meteorological condition characteristic data from the UAV delivery mission data;

[0148] The flight trajectory avoidance behavior learning module is used to simulate the UAV delivery flight state based on the UAV flight braking parameters and the UAV delivery mission data to obtain the UAV delivery flight state simulation data; simulate and quantify the flight attitude instability of the UAV delivery flight state simulation data based on the low-altitude delivery meteorological condition characteristic data to obtain the flight attitude instability quantification data; and perform UAV flight trajectory avoidance behavior control learning based on the flight attitude instability quantification data to obtain the flight trajectory avoidance behavior control data;

[0149] The trajectory control model construction module is used to construct a trajectory control model for the flight trajectory avoidance behavior control data based on the policy gradient algorithm to obtain the UAV flight trajectory control model; the UAV flight trajectory control model is sent to the UAV control center to execute UAV trajectory control.

[0150] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A UAV trajectory control method, characterized in that: The following steps are involved: Step S1: Acquire the UAV flight braking parameters and the UAV delivery mission data; extract the low-altitude delivery meteorological conditions from the UAV delivery mission data to obtain the low-altitude delivery meteorological condition characteristic data; Step S2 is specifically as follows: Step S21: simulating the UAV delivery flight state based on the UAV flight braking parameters and the UAV delivery mission data to obtain UAV delivery flight state simulation data; Step S22: Acquire historical bird flock flight interference data for drone delivery; extract delivery load from drone delivery mission data to obtain drone delivery load data; Step S23: performing flight attitude instability simulation and quantification on the UAV delivery flight state simulation data based on the low-altitude delivery meteorological condition characteristic data and the UAV delivery payload data to obtain flight attitude instability quantification data; Step S24: performing flight trajectory avoidance behavior control learning for the UAV based on historical flock flight interference data and flight attitude instability quantified data to obtain flight trajectory avoidance behavior control data; wherein step S24 includes: Step S241: Marking the multi-target bird flock movement trajectory of the historical bird flock flight interference data to obtain the multi-target bird flock movement trajectory; Step S242: performing approach relative orientation clustering learning on the historical bird flock flight interference data based on the multi-target historical movement trajectories of the bird flock to obtain bird flock approach relative orientation clustering data; Step S243: performing a quantitative assessment of the airflow disturbance caused by the approach of the bird flock on the historical bird flock flight interference data according to the historical movement trajectories of the bird flock multi-target, thereby obtaining quantitative data of the airflow disturbance caused by the approach of the bird flock; Step S244: identifying the critical state of the UAV's flight attitude instability based on the quantitative data of the airflow disturbance caused by the approaching flock of birds and the clustered data of the relative orientation of the approaching flock of birds, and obtaining the critical state of the flight attitude instability; Step S245: performing flight trajectory avoidance behavior control learning for the UAV based on the relative orientation clustering data of the approaching flock of birds and the critical state of flight attitude instability to obtain flight trajectory avoidance behavior control data; Step S3: Based on the policy gradient algorithm, a trajectory control model is constructed for the flight trajectory avoidance behavior control data to obtain the UAV flight trajectory control model; the UAV flight trajectory control model is sent to the UAV control center to execute the UAV trajectory control.

2. The UAV trajectory control method according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: Acquire UAV flight braking parameters and UAV delivery mission data; Step S12: performing data cleaning on the drone delivery task data to obtain drone delivery task cleansing data; Step S13: extracting low-altitude delivery meteorological conditions from the UAV delivery task cleaning data to obtain low-altitude delivery meteorological condition data; Step S14: Perform feature analysis on the low-altitude delivery meteorological condition data to obtain low-altitude delivery meteorological condition feature data.

3. The UAV trajectory control method according to claim 1, characterized in that: Step S23 includes the following steps: Step S231: performing low-altitude turbulence structure identification on the low-altitude delivery meteorological condition characteristic data to obtain low-altitude turbulence structure data; performing wind direction shear phase plane state analysis on the low-altitude turbulence structure data to obtain wind direction shear phase plane state; Step S232: calculating the average difference of asymmetric aerodynamic pressure of the UAV on the UAV delivery flight state simulation data based on the wind direction shear phase plane state to obtain the average difference of asymmetric aerodynamic pressure of the UAV; Step S233: Calculating the rolling moment fluctuation of the UAV delivery flight state simulation data based on the UAV asymmetric aerodynamic pressure difference and the wind shear phase plane state to obtain the UAV flight rolling moment fluctuation data; Step S234: performing a flight vibration multi-axis intensity regression analysis on the UAV delivery flight state simulation data based on the UAV flight rolling moment fluctuation data, the UAV asymmetric aerodynamic pressure average difference, and the UAV delivery payload data to obtain flight vibration multi-axis intensity regression data; Step S235: performing flight attitude instability simulation and quantification based on the flight tremor multi-axis intensity regression data to obtain flight attitude instability quantification data.

4. The UAV trajectory control method according to claim 3, characterized in that: Step S234 includes the following steps: Perform multi-axis rotational inertia incremental integration on the UAV's flight rolling moment fluctuation data to obtain multi-axis rotational inertia incremental integral data; The UAV angular velocity progressive growth ratio is calculated based on the UAV flight rolling moment fluctuation data to obtain the angular velocity progressive growth ratio; The axial yaw moment momentum deviation is quantified based on the average difference of the asymmetric aerodynamic pressure of the UAV to obtain the axial yaw moment momentum deviation data; The UAV aerodynamic pressure imbalance is quantified based on the multi-axis rotational inertia incremental integral data, the UAV delivery load data and the axial yaw moment momentum deviation data to obtain the aerodynamic pressure imbalance quantification data; Based on the quantified data of aerodynamic pressure imbalance, the ratio of angular velocity progressive increase and the multi-axis moment of inertia incremental integral data, the aerodynamic structure flutter inter-axis strength coupling simulation analysis is carried out to obtain the aerodynamic structure flutter inter-axis strength coupling data; Based on the inter-axis intensity coupling data of aerodynamic structure flutter, a multi-axis intensity regression analysis of flight flutter is performed on the simulation data of UAV delivery flight state to obtain the multi-axis intensity regression data of flight flutter.

5. The UAV trajectory control method according to claim 1, characterized in that: Step S244 includes the following steps: The multi-directional flapping wing rotating vortex intensity data were obtained by analyzing the multi-directional flapping wing rotating vortex intensity data based on the quantitative data of airflow disturbance and the relative orientation clustering data of the approaching birds. The flapping rhythm vortex transient mutation data are identified based on the multi-directional flapping wing rotation vortex intensity data to obtain the flapping rhythm vortex transient mutation data; The flight control surface deflection stall index of the UAV is evaluated based on the transient mutation data of the flapping rhythm vortex and the flight attitude instability quantitative data, and the flight control surface deflection stall index is obtained. The critical state of UAV flight attitude instability is identified based on the flight attitude instability quantitative data according to the flight control surface deflection stall index, and the critical state of flight attitude instability is obtained.

6. The UAV trajectory control method according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: performing convolution processing on the flight trajectory avoidance behavior control data to obtain flight trajectory avoidance behavior convolution data; Step S32: performing iterative incremental learning on the flight trajectory avoidance behavior convolution data to obtain flight trajectory avoidance behavior iterative data; Step S33: constructing a trajectory control model based on the flight trajectory avoidance behavior iterative data based on the policy gradient algorithm to obtain a UAV flight trajectory control model; Step S34: Send the UAV flight trajectory control model to the UAV control center to perform UAV trajectory control.

7. The UAV trajectory control method according to claim 6, characterized in that: Step S33 includes the following steps: Step S331: dividing the flight trajectory avoidance behavior iteration data into a training set and a test set to obtain a flight trajectory avoidance behavior training set and a flight trajectory avoidance behavior test set respectively; Step S332: performing clustering processing on the flight trajectory avoidance behavior training set to obtain a flight trajectory avoidance behavior cluster training set; Step S333: performing feature sampling on the flight trajectory avoidance behavior cluster training set to obtain a trajectory avoidance behavior cluster sampling training set; Step S334: constructing an initial trajectory control model for the trajectory avoidance behavior cluster sampling training set based on the policy gradient algorithm to obtain an initial UAV flight trajectory control model; Step S335: Input the trajectory avoidance behavior test set into the initial UAV flight trajectory control model for model testing to obtain the UAV flight trajectory control model.

8. A UAV trajectory control system, characterized in that: Used to execute the UAV trajectory control method according to claim 1, the UAV trajectory control system comprises: The low-altitude delivery meteorological condition extraction module is used to obtain the UAV flight braking parameters and UAV delivery mission data; the low-altitude delivery meteorological condition extraction module is used to extract the low-altitude delivery meteorological condition characteristic data from the UAV delivery mission data; The flight trajectory avoidance behavior learning module is used to simulate the UAV delivery flight state based on the UAV flight braking parameters and the UAV delivery mission data to obtain the UAV delivery flight state simulation data; simulate and quantify the flight attitude instability of the UAV delivery flight state simulation data based on the low-altitude delivery meteorological condition characteristic data to obtain the flight attitude instability quantification data; and perform UAV flight trajectory avoidance behavior control learning based on the flight attitude instability quantification data to obtain the flight trajectory avoidance behavior control data; The trajectory control model construction module is used to construct a trajectory control model for the flight trajectory avoidance behavior control data based on the policy gradient algorithm to obtain the UAV flight trajectory control model; the UAV flight trajectory control model is sent to the UAV control center to execute UAV trajectory control.

Images