Bionic swarm intelligence low-altitude logistics unmanned aerial vehicle cluster anti-wind interference cooperation method

Abstract

The invention discloses a bionic group intelligent low-altitude logistics unmanned aerial vehicle cluster anti-wind interference cooperation method, and the method comprises the steps: collecting the historical flight data and three-dimensional wind field data of an unmanned aerial vehicle cluster, and generating a bionic formation feature set with a wind field label; inputting the bionic formation feature set into a swarm intelligence model fused with fluid mechanics, and generating a dynamic formation topology instruction; according to the dynamic formation topology instruction, adjusting the relative position and attitude angle of each unmanned aerial vehicle through a distributed cooperative control algorithm, and generating an anti-wind disturbance cooperative flight state; and continuously monitoring the deviation between the three-dimensional wind field change and the cooperative flight state, dynamically correcting the weight of the formation density-anti-wind disturbance intensity mapping relation through a reinforcement learning algorithm, updating a dynamic formation topology instruction, and realizing adaptive control of bionic group anti-wind disturbance cooperation. According to the embodiment of the invention, high-disturbance-rejection cooperative flight of the unmanned aerial vehicle cluster in the dynamic wind field can be realized, the formation energy consumption is reduced, and the obstacle avoidance capability under the sudden wind condition is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned aerial vehicles (UAVs), and in particular to a bionic swarm intelligence low-altitude logistics UAV cluster anti-wind disturbance collaborative method. Background Art

[0002] With the rapid development of low-altitude logistics drone swarms, their ability to resist wind disturbances under complex meteorological conditions has become a key challenge. Traditional drone formation control mostly adopts fixed formations or simple following strategies, which are difficult to cope with dynamic wind field changes, especially in sudden gusts or turbulent environments, which can easily lead to formation instability, surge in energy consumption and even collision risks. In existing technologies, single optimization methods based on fluid mechanics or swarm intelligence have limitations: the fluid mechanics model is computationally complex and difficult to respond to wind field changes in real time; and the swarm intelligence algorithm lacks accurate modeling of aerodynamic effects, resulting in insufficient formation coordination efficiency. In addition, the correlation analysis between historical flight data and wind field characteristics has not been fully explored, and the application of bionics in anti-wind disturbance coordination is mostly limited to simple behavioral imitation of birds or insects, without combining flow field interaction and dynamic topology optimization. Summary of the Invention

[0003] The purpose of the present invention is to provide a bionic swarm intelligence low-altitude logistics drone cluster anti-wind disturbance coordination method to address the deficiencies in the existing technology, enable high-interference-resistant coordinated flight of drone clusters in dynamic wind fields, reduce formation energy consumption and improve obstacle avoidance capabilities in sudden wind conditions.

[0004] One embodiment of the present application provides a bionic swarm intelligence low-altitude logistics UAV cluster anti-wind disturbance collaboration method, the method comprising: Collect historical flight data and three-dimensional wind field data of the UAV swarm, extract the spatial topological features and wind field adaptation features of the formation through a bionic feature extraction algorithm, and generate a bionic formation feature set with wind field labels; The bionic formation feature set is input into a swarm intelligence model integrated with fluid mechanics to construct a formation density-wind disturbance resistance mapping relationship, and a multi-objective optimization is used to solve the optimal formation structure and generate dynamic formation topology instructions; According to the dynamic formation topology instructions, the relative positions and attitude angles of each UAV are adjusted through a distributed cooperative control algorithm, and the updraft of the wake of the leading UAV is used to reduce the energy consumption of the trailing UAV. At the same time, group obstacle avoidance in the event of sudden gusts is achieved through local information interaction, generating a wind-resistant cooperative flight state. Continuously monitor the deviation between the three-dimensional wind field changes and the coordinated flight status, dynamically correct the weight of the formation density-anti-wind disturbance intensity mapping relationship through the reinforcement learning algorithm, update the dynamic formation topology instructions, and realize the adaptive control of bionic group anti-wind disturbance coordination.

[0005] Optionally, the historical flight data and three-dimensional wind field data of the UAV cluster are collected, and the spatial topological features and wind field adaptation features of the formation are extracted by a bionic feature extraction algorithm to generate a bionic formation feature set with wind field labels, including: Synchronously collect historical flight data and 3D wind field data from drone swarms, eliminate data collection delays through timestamp alignment, use Kalman filtering to remove measurement noise in flight data, and output pre-processed flight and wind field datasets that are synchronized in time and space. Based on the preprocessed flight data set, the Voronoi diagram partitioning algorithm is used to calculate the relative distance and azimuth between each UAV. Combining the diamond and wedge topological characteristics of the bird formation, the spatial adjacency matrix and density distribution entropy of the formation are extracted, and the spatial topological feature vector of the formation is output. The wind field dataset is associated with the pre-processed flight data. The dynamic time warping algorithm is used to match the UAV's attitude adjustment amplitude, lift coefficient correction value, and energy consumption fluctuation coefficient under different wind field intensities. Principal component analysis is used to extract the core indicators that best reflect the wind field adaptability and output the wind field adaptation feature matrix. The formation space topology feature vector and the wind field adaptation feature matrix are bound in time series, and the corresponding wind field intensity and wind direction labels are marked. The feature fusion network is used to generate a bionic formation feature set with wind field labels.

[0006] Optionally, the bionic formation feature set is input into a swarm intelligence model integrated with fluid mechanics, a formation density-wind disturbance resistance mapping relationship is constructed, a multi-objective optimization is used to solve the optimal formation structure, and a dynamic formation topology instruction is generated, including: The feature importance of the bionic formation feature set is evaluated to select features that are strongly correlated with wind disturbance resistance, eliminate redundant features, and output a streamlined feature subset. The simplified feature subset is input into the swarm intelligence model, integrated with the wake interference model in computational fluid dynamics, and the airflow disturbance coefficients under different formation structures are obtained through flow field simulation, and the formation-flow field interaction parameter matrix is ​​output; Based on the formation-flow field interaction parameter matrix, with formation density as input variable and anti-wind disturbance strength as output variable, Gaussian process regression is used to construct the nonlinear mapping relationship between the two and output the formation density-anti-wind disturbance strength mapping function. Set multi-objective optimization goals, use mapping functions as constraints, use genetic algorithms to search for Pareto optimal solutions, and output multiple sets of candidate formation structure parameters; A wind field simulation test is conducted on the candidate formation structure parameters to screen out structures that meet the wind disturbance resistance threshold and have energy consumption lower than the preset value. The structures are converted into relative position instructions and communication topology relationships for each UAV to generate dynamic formation topology instructions.

[0007] Optionally, the method of adjusting the relative position and attitude angle of each UAV through a distributed collaborative control algorithm based on the dynamic formation topology instruction, utilizing the updraft of the wake of the leading UAV to reduce the energy consumption of the trailing UAV, and achieving group obstacle avoidance in the event of sudden gusts through local information interaction to generate an anti-wind disturbance collaborative flight state includes: Parse the dynamic formation topology instructions, extract the target relative position and communication neighbor list of each drone, calculate the position deviation based on the current position, and output the target position-deviation comparison table; Based on the target position-deviation comparison table, a distributed model predictive control algorithm is adopted. Each UAV only exchanges position information with its communication neighbors, autonomously calculates the speed adjustment, and realizes position closed-loop regulation through the PID controller, outputting a real-time position correction sequence. Based on the real-time position correction sequence and the wind direction angle of the three-dimensional wind field, the pitch and yaw angles of each UAV are optimized through the aerodynamic model, the attitude angle adjustment amount is calculated, and the attitude collaborative control instructions are output; Based on attitude coordinated control instructions, the system uses onboard meteorological sensors to identify the updraft area of ​​the leading aircraft's wake, guides the following aircraft into this area and adjusts its altitude. It then calculates the lift gain and energy consumption reduction of each following aircraft and outputs the wake utilization parameters. Each UAV exchanges gust monitoring data in real time through short-distance communication, adopts a local obstacle avoidance algorithm based on the potential field method, adjusts the velocity vector to achieve group obstacle avoidance, integrates position correction, attitude collaborative control, and wake utilization parameters to generate a wind-resistant collaborative flight state.

[0008] Optionally, the method of continuously monitoring the deviation between the three-dimensional wind field changes and the coordinated flight state, dynamically correcting the weight of the formation density-anti-wind disturbance intensity mapping relationship through a reinforcement learning algorithm, updating the dynamic formation topology instructions, and realizing adaptive control of bionic group anti-wind disturbance coordination includes: The distributed sensors of the UAV cluster collect the current three-dimensional wind field data and the flight status data of each aircraft in real time, calculate the deviation between the actual flight status and the anti-wind disturbance coordinated flight status, and output the state deviation matrix; The state deviation matrix and the current wind field characteristics are used as state inputs to construct a Markov decision process for reinforcement learning. The deep deterministic policy gradient algorithm is used to train the weight correction model and output the weight adjustment gradient of the mapping relationship. Adjust the gradient based on the weights, dynamically correct the key weights in the formation density-wind disturbance resistance mapping relationship, test the accuracy of the corrected mapping relationship through the validation set, and output the optimized mapping relationship; Based on the optimized mapping relationship, the multi-objective optimization solution process is re-executed to generate updated dynamic formation topology instructions, which are then sent to each UAV to adjust the coordination status, forming a closed loop of monitoring-correction-optimization-execution, and realizing adaptive control of bionic group anti-wind disturbance coordination.

[0009] Another embodiment of the present application provides a bionic swarm intelligence low-altitude logistics drone cluster anti-wind disturbance collaborative system, the system comprising: The extraction module is used to collect historical flight data and three-dimensional wind field data of the UAV cluster, extract the spatial topological features and wind field adaptation features of the formation through the bionic feature extraction algorithm, and generate a bionic formation feature set with wind field labels; A construction module is used to input the bionic formation feature set into a swarm intelligence model integrated with fluid mechanics, construct a formation density-anti-wind disturbance strength mapping relationship, use multi-objective optimization to solve the optimal formation structure, and generate dynamic formation topology instructions; An adjustment module is used to adjust the relative positions and attitude angles of each UAV through a distributed cooperative control algorithm according to the dynamic formation topology instructions, utilize the updraft of the wake of the leading UAV to reduce the energy consumption of the trailing UAV, and achieve group obstacle avoidance in the event of sudden gusts through local information interaction, thereby generating a wind-resistant cooperative flight state; The correction module is used to continuously monitor the deviation between the three-dimensional wind field changes and the coordinated flight status, dynamically correct the weight of the formation density-anti-wind disturbance intensity mapping relationship through the reinforcement learning algorithm, update the dynamic formation topology instructions, and realize the adaptive control of bionic group anti-wind disturbance coordination.

[0010] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when run.

[0011] Yet another embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the above methods.

[0012] Compared with the existing technology, the present invention provides a bionic swarm intelligent low-altitude logistics drone cluster anti-wind disturbance collaboration method, which collects historical flight data and three-dimensional wind field data of the drone cluster to generate a bionic formation feature set with wind field labels; the bionic formation feature set is input into a swarm intelligence model integrated with fluid mechanics to generate dynamic formation topology instructions; according to the dynamic formation topology instructions, the relative position and attitude angle of each drone are adjusted through a distributed collaborative control algorithm to generate an anti-wind disturbance collaborative flight state; the deviation between the three-dimensional wind field changes and the collaborative flight state is continuously monitored, the weight of the formation density-anti-wind disturbance intensity mapping relationship is dynamically corrected through a reinforcement learning algorithm, the dynamic formation topology instructions are updated, and adaptive control of bionic swarm anti-wind disturbance collaboration is realized, thereby realizing high anti-interference collaborative flight of the drone cluster in a dynamic wind field, reducing formation energy consumption and improving obstacle avoidance capability under sudden wind conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A hardware block diagram of a computer terminal for a bionic swarm intelligence low-altitude logistics drone cluster anti-wind disturbance collaborative method provided by an embodiment of the present invention; Figure 2 A schematic diagram of a flow chart of a bionic swarm intelligence low-altitude logistics UAV cluster anti-wind disturbance collaborative method provided by an embodiment of the present invention; Figure 3 A schematic structural diagram of a bionic swarm intelligent low-altitude logistics drone cluster anti-wind disturbance collaborative system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0014] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0015] The embodiment of the present invention first provides a bionic swarm intelligence low-altitude logistics drone cluster anti-wind disturbance collaboration method, which can be applied to electronic devices such as computer terminals, specifically ordinary computers.

[0016] The following describes it in detail by taking running on a computer terminal as an example. Figure 1 The hardware structure block diagram of the computer terminal of a bionic swarm intelligence low-altitude logistics UAV cluster anti-wind disturbance collaborative method provided by the embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0017] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions that, when executed, enable the processor to execute any one of the bionic swarm intelligence low-altitude logistics drone cluster anti-wind interference collaborative methods.

[0018] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0019] The internal memory provides an environment for the operation of computer programs in non-volatile storage media. When the computer program is executed by the processor, the processor can execute any bionic swarm intelligence low-altitude logistics drone cluster anti-wind interference collaborative method.

[0020] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 1The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0021] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0022] See also Figure 2 The embodiment of the present invention provides a bionic swarm intelligence low-altitude logistics UAV cluster anti-wind disturbance collaboration method, which may include the following steps: S201, collecting historical flight data and three-dimensional wind field data of the UAV cluster, extracting the spatial topological features and wind field adaptation features of the formation through a bionic feature extraction algorithm, and generating a bionic formation feature set with wind field labels; Specifically, it can synchronously collect historical flight data and three-dimensional wind field data of the drone cluster, eliminate data collection delays through timestamp alignment, use Kalman filtering to remove measurement noise in the flight data, and output spatiotemporally synchronized pre-processed flight data sets and wind field data sets; Multi-source data synchronous acquisition and timestamp alignment: The historical flight data of the drone cluster is collected in real time through the onboard sensors of each drone, including: GPS positioning data (longitude, latitude, altitude, with centimeter-level accuracy); IMU inertial measurement data (three-axis accelerometer, gyroscope, sampling frequency of 100 Hz); airspeed meter data (flight speed, in meters per second); attitude angle data (pitch angle, roll angle, yaw angle, in degrees).

[0023] Meanwhile, three-dimensional wind data is collected by Doppler lidar and meteorological sondes deployed in the flight airspace. These data include the three-dimensional distribution of wind speed (in meters per second), wind direction (in degrees), and vertical turbulence intensity (a dimensionless parameter). Due to differences in data collection frequency and transmission delays between different devices (e.g., GPS update frequency of 10 Hz and radar update frequency of 5 Hz), the system uses the Precision Time Protocol (PTP) to timestamp all data with nanosecond-level accuracy. Using a sliding time window matching algorithm (with a window size of 500 milliseconds), the drone's position at the same moment is aligned with the wind data at the corresponding spatial coordinates, eliminating data misalignment caused by communication delays. For example, if the wind speed data for a drone's coordinate point (X, Y, Z) at timestamp T is missing, bilinear interpolation is performed using wind data from the adjacent times T-100 milliseconds and T+100 milliseconds.

[0024] Kalman filter noise reduction processing: The measurement noise in flight data mainly comes from sensor errors and environmental interference (such as electromagnetic interference). Take the altitude data of a drone as an example: State equation: predicts the height at the next moment based on the physical motion model (such as a uniform ascent / descent model); Observation equation: Fusion of GPS altitude, barometer altitude, and ultrasonic ranging data.

[0025] The Kalman filter sets the process noise covariance matrix Q (representing model prediction error) and the observation noise covariance matrix R (representing sensor error) and outputs the optimal estimate through recursive calculation (prediction-update loop). For example, when the GPS signal is obscured by clouds, the filter automatically increases the weight of GPS noise in the R matrix and relies on barometer data instead, ensuring smooth and continuous altitude data. The standard deviation (SD) of the processed flight data is reduced from 1.5 meters in the original data to less than 0.2 meters.

[0026] Generation of spatiotemporal synchronized datasets: The aligned and filtered data are organized into structured datasets in time series: Preprocessed flight data set: Each row contains fields such as timestamp, drone ID, 3D coordinates, attitude angle, speed, etc. Wind field dataset: Each row contains a timestamp, three-dimensional grid coordinates (grid resolution 10m × 10m × 5m), wind speed vector, and turbulence intensity value.

[0027] The final output data undergoes spatiotemporal consistency verification (such as verifying whether the drone position is within the wind field grid at the same time) to ensure the reliability of subsequent analysis.

[0028] Based on the preprocessed flight data set, the Voronoi diagram partitioning algorithm is used to calculate the relative distance and azimuth between each UAV. Combining the diamond and wedge topological characteristics of the bird formation, the spatial adjacency matrix and density distribution entropy of the formation are extracted, and the spatial topological feature vector of the formation is output. Voronoi diagram construction and neighborhood relationship analysis: For each drone cluster location at a given moment, each drone is used as a generator point and its Voronoi cells (i.e., the convex polygons formed by the points closest to it in space) are calculated. This is achieved quickly using the Fortune sweep algorithm (computational complexity O(n log n)). For example, in a formation of 10 drones, the algorithm outputs 10 non-overlapping Voronoi polygons. Based on the polygon adjacency relationships, two drones are considered spatial neighbors if their Voronoi cells share an edge. Based on this, the adjacency matrix A (A) is generated: if drone i is adjacent to j, then A[i, j] = 1; otherwise, it is 0. The relative distance matrix D (element D[i, j] represents the Euclidean distance from drone i to j) and the azimuth matrix Θ (the azimuth from i to j is calculated, with true north as 0 degrees).

[0029] Bird formation topology feature mapping: Compare the actual formation with typical structures in biomimetic science: Diamond topology: The main drone is located at the front, and the drones on both sides are arranged in a symmetrical diagonal line, with an azimuth angle difference of a fixed angle (such as 60 degrees); Wedge topology: Similar to a goose formation, all drones point in the same direction, and the azimuth angles are symmetrically distributed along the center line.

[0030] The topological similarity was quantified by calculating the Azimuth Consistency Index (ACI): if the standard deviation of the formation's azimuth distribution was less than 15 degrees and the formation was centrally symmetric, it was considered a wedge topology (ACI>0.8); if there were multiple groups of 60-degree angles and the adjacency matrix was connected in a ring shape, it was considered a diamond topology (ACI>0.7).

[0031] Density distribution entropy calculation: Measuring the spatial density of the formation: Step 1: Divide the flight airspace into a 1-cubic-meter voxel grid. Step 2: Count the number of drones within each grid and generate a density distribution histogram. Step 3: Calculate the Shannon Entropy (SE). High entropy values ​​(e.g., SE > 2.5) indicate a uniform formation distribution, while low entropy values ​​(SE < 1.0) indicate a dense or sparse formation. The final output is a topological feature vector: [adjacency matrix compression value, average relative distance, azimuth consistency index, density entropy].

[0032] The wind field dataset is associated with the pre-processed flight data. The dynamic time warping algorithm is used to match the UAV's attitude adjustment amplitude, lift coefficient correction value, and energy consumption fluctuation coefficient under different wind field intensities. Principal component analysis is used to extract the core indicators that best reflect the wind field adaptability and output the wind field adaptation feature matrix. Wind field-flight behavior dynamic correlation: Taking a single UAV as an example, extract its response parameters in the wind field: Attitude Adjustment Amplitude (AAA): The change in roll angle from 0 to 8 degrees in a force 5 crosswind (10 m / s). Lift Coefficient Correction (LCC): The ratio of actual lift to theoretical lift, reflecting aerodynamic efficiency (e.g., LCC drops from 1.0 to 0.85 in a headwind). Energy Consumption Fluctuation (ECF): The instantaneous fluctuation rate of battery current (e.g., the current surges from 20 amps to 35 amps in a gust of wind).

[0033] Because there is a time lag between wind field changes and drone responses (for example, attitude adjustment occurs 200 milliseconds after a sudden change in wind speed), the Dynamic Time Warping (DTW) algorithm is used to align time series: the wind speed time series W(t) is nonlinearly stretched / compressed to match the attitude angle series P(t); the time difference offset under the optimal path (for example, a maximum lag of 300 milliseconds) and the matching similarity (DTW distance < 0.1 indicates strong correlation) are output.

[0034] Multi-parameter principal component dimensionality reduction: The initial parameter set includes 20 indicators (such as AAA, LCC, ECF, rudder angle, etc.), which are compressed through principal component analysis (PCA): Step 1: Calculate the covariance matrix and find the eigenvalue and eigenvector. Step 2: Select principal components with a Cumulative Contribution Rate (CCR) greater than 85%. For example, the first principal component PC1 (contribution rate 60%) is strongly correlated with energy consumption fluctuations; the second principal component PC2 (25%) reflects the sensitivity of posture adjustment. Step 3: Generate a core indicator set: Wind field adaptability index = 0.6 × PC1 + 0.25 × PC2. This index comprehensively represents the ability to resist wind disturbance (the larger the value, the stronger the adaptability).

[0035] Feature matrix construction: Organize the time series into a matrix, for example: Timestamp Wind speed level Wind field adaptation index Attitude adjustment amplitude Energy consumption fluctuation coefficient T1 4th level 0.72 5.3 degrees 1.15 T2 6th level 0.58 8.7 degrees 1.42 The number of matrix rows is the number of samples, and the number of columns is the feature dimension.

[0036] The formation space topology feature vector and the wind field adaptation feature matrix are bound in time series, and the corresponding wind field intensity and wind direction labels are marked. The feature fusion network is used to generate a bionic formation feature set with wind field labels.

[0037] Spatiotemporal feature binding and labeling: Based on the same time: Input 1: Formation space topology feature vector (e.g., [adjacency matrix: 110011, average distance: 15.2 meters, ACI: 0.75, density entropy: 1.8]); Input 2: Wind field adaptation feature matrix (e.g., [wind field adaptation index: 0.68, attitude adjustment range: 6.4 degrees]); Labels: wind field intensity (classified by Beaufort wind scale, e.g., level 5), wind direction (e.g., northeast wind 45 degrees). After binding, a complete sample is formed. Example: { Timestamp: T, Topological features: [0.75, 15.2, 1.8], Adaptive features: [0.68, 6.4], Tags: [wind speed:8.5m / s, wind direction:45°]}. Feature fusion network processing: A three-layer fully connected neural network is used to achieve feature fusion: Input layer: 3D topological features + 2D adaptation features = 5D input; hidden layer: 8 neurons, activation function is ReLU (Rectified Linear Unit); output layer: 4D fused feature vector.

[0038] During training, wind field labels are used as supervisory signals: Loss function: Mean Squared Error (MSE); Optimizer: Adaptive Moment Estimation (Adam, learning rate 0.001). For example, when the input topological feature ACI = 0.8 (strong wedge structure) and the adaptation index = 0.7 (high wind resistance), the output fusion feature [0.92, 0.15, 0.33, 0.07] represents the optimal state of the formation in northeasterly winds.

[0039] Bionic formation feature set generation: Perform fusion operations on data from all historical moments to generate a structured data set, including: sample ID, fused feature vector, wind speed label, and wind direction label.

[0040] S202, inputting the bionic formation feature set into a swarm intelligence model integrated with fluid mechanics, constructing a formation density-wind disturbance resistance mapping relationship, solving the optimal formation structure using multi-objective optimization, and generating dynamic formation topology instructions; Specifically, the feature importance of the bionic formation feature set can be evaluated to screen out features that are strongly correlated with wind resistance performance, eliminate redundant features, and output a streamlined feature subset. The system first receives the Bionic Formation Feature Set (BFFS) generated in the previous step. This dataset contains multidimensional features (such as the spatial adjacency matrix, density distribution entropy, attitude adjustment amplitude, lift coefficient correction value, etc.) and their corresponding wind field intensity labels. The core of feature importance assessment is to identify which features have a decisive impact on wind resistance. A random forest algorithm is used for preliminary screening: a random forest model consisting of 1,000 decision trees is constructed, each trained on a random subset of the BFFS and a random subset of features. During training, the mean decrease in impurity (MDI) of each feature across all trees is recorded—the sum of the information gain when the feature is used as a split node. For example, an MDI value of 0.15 (ranging from 0 to 1) for the feature "density distribution entropy" indicates a significant contribution to predicting wind resistance performance; whereas an MDI value of only 0.02 for the feature "azimuth variance" may indicate a weak correlation. At the same time, the permutation feature importance (PFI) is calculated: after randomly perturbing the value of a certain feature (such as the lift coefficient correction value), the decrease in model accuracy (such as an 8% decrease in accuracy) is re-evaluated to further verify its importance.

[0041] To avoid bias from single-method approaches, a recursive feature elimination (RFE) strategy was supplemented. Using a support vector machine (SVM) as the base model, a set containing all features was initialized. The following steps were performed iteratively: the SVM model was trained and feature weights were obtained; features with the smallest absolute weights (e.g., weights less than 0.01) were removed. After 10 rounds of iteration, the top 30% of features by cumulative weight were retained. During this process, it was found that the "energy consumption fluctuation coefficient" was eliminated in early iterations, confirming its redundancy. Finally, a correlation redundancy analysis was performed: the Pearson Correlation Coefficient (PCC) between features was calculated. If the absolute value of the PCC between two features exceeded 0.85 (e.g., the PCC for "horizontal spacing" and "diamond topology index" was 0.92), the one with the lower mean dilation index (MDI) was removed. Through a triple filtering mechanism (MDI + PFI + RFE + correlation analysis), the original 128-dimensional features were compressed to 35 dimensions, forming a refined feature subset (RFS).

[0042] To verify the effectiveness of the feature reduction, feature sensitivity testing was conducted in a wind tunnel simulation environment, maintaining a fixed wind intensity of level 6 (12 m / s). Prediction models were trained using both the original feature set and the RFS. The results showed that using RFS increased model inference speed by 3.2 times, while the prediction error for wind resistance performance increased by only 0.8% (absolute error 1.2% to 1.3%), demonstrating the effectiveness of redundant feature removal. The RFS ultimately included key features such as the Density Distribution Entropy (DDE) (reflecting formation compactness), the Wake Interference Index (WII) (quantifying the influence of the lead aircraft's wake), and the Crosswind Offset Response Rate (CORR) (indicating crosswind resistance), laying the foundation for subsequent flow modeling.

[0043] The simplified feature subset is input into the swarm intelligence model, integrated with the wake interference model in computational fluid dynamics, and the airflow disturbance coefficients under different formation structures are obtained through flow field simulation, and the formation-flow field interaction parameter matrix is ​​output; The swarm intelligence model (SIM) utilizes an improved particle swarm optimization-fluid dynamics (PSO)-CFD framework. First, the formation structure parameters (such as horizontal spacing, vertical height difference, and yaw angle) in the RFS are mapped into a set of three-dimensional coordinate points. An unstructured mesh (with a resolution of 0.1 meters) is generated from this point set and used as input for the computational fluid dynamics (CFD) simulation. The core CFD model is the wake interference model (WIM), which simulates airflow by solving the Reynolds-Averaged Navier-Stokes (RANS) equations. The k-omega shear stress transport model (SST) is used to balance computational accuracy and efficiency. Boundary conditions are set as follows: an inlet wind speed of 8 to 15 m / s (covering the range of common wind disturbances), a turbulence intensity of 5%, and enhanced wall treatment (EWT) for the wall function.

[0044] Flow field simulation performs a three-step process: Steady-state solution: Under a fixed formation structure, iterative calculation is performed until the flow field residual (Residual) drops below 10^-4, and the velocity field and pressure field distribution are output; Transient sampling: Introduce gust disturbance (Gust Disturbance), lasting 5 seconds, with a time step of 0.01 seconds, and record flow field data at 500 time points; Disturbance Quantification: Extract the airflow disturbance coefficient (ADC) at each drone's position using the following formula: ADC = (σ_v / V_mean) × 100%, where σ_v is the standard deviation of wind speed and V_mean is the mean wind speed. For example, the ADC value of the rear aircraft in a diamond formation is 18.7%, while in a loose formation it can reach 32.5%.

[0045] The swarm intelligence model uses an adaptive particle swarm (APS) to drive the simulation: 50 particles represent different combinations of formation structure parameters (such as particle position encoding horizontal spacing), and each particle triggers a CFD simulation. Particle velocity updates incorporate fluid gradient information, accelerating convergence to low-disturbance regions.

[0046] Finally, the formation-flow interaction parameter matrix (FFIPM) is generated, with a dimension of [N×M]: Row dimension N: the number of simulated formation structures (e.g., 50 particles in APS correspond to 50 structures); Column dimension M: contains three types of parameters: (1) Structural parameters: Mean Horizontal Spacing (MHS), Height Variance (HV); (2) Flow field parameters: mean ADC (MAD), peak ADC (PAD); (3) Performance parameter: Lift-to-Drag Ratio Variation (LDRV).

[0047] This matrix establishes an explicit association between the geometric characteristics of the formation and the flow field response. For example, records show that when MHS = 3.2 meters and HV = 0.5 meters, the MAD drops to 15.3%, which is better than the baseline value of 25.1%.

[0048] Based on the formation-flow field interaction parameter matrix, with formation density as input variable and anti-wind disturbance strength as output variable, Gaussian process regression is used to construct the nonlinear mapping relationship between the two and output the formation density-anti-wind disturbance strength mapping function. Formation Density (FD) is the core input variable, defined as a composite of the number of drones per cubic meter (DCM) and the space filling ratio (SFR): FD = α × DCM + β × SFR (for example, α = 0.6, β = 0.4). DCM = number of drones in the cluster / volume of the formation bounding box, and SFR is calculated using the α-shape algorithm to calculate the convex hull volume fraction. Wind Resistance Intensity (WRI) is the output variable, a weighted composite of three metrics: WRI = γ1 × (1-ADC_norm) + γ2 × LDRV_norm + γ3 × Stability_Index (γ1 = 0.5, γ2 = 0.3, γ3 = 0.2). ADC_norm is the normalized airflow disturbance coefficient, and Stability_Index is calculated using the Lyapunov exponent.

[0049] Gaussian Process Regression (GPR) was chosen as the modeling tool because it excels at handling small sample nonlinear problems. The implementation consists of four steps: Kernel selection: The Matérn kernel (parameter ν = 2.5) is used, whose expression includes the distance scale parameter l = 1.2 and the signal variance σ_f = 0.8. This combination can capture the local mutation characteristics of the FD-WRI relationship; Training data construction: 80% of the data (40 groups) were extracted from the FFIPM matrix as the training set, with the input being FD (range 0.35-0.82) and the output being WRI (range 0-1); Hyperparameter optimization: Automatically adjust kernel parameters by maximizing marginal likelihood, with convergence condition of likelihood function change rate < 0.01%; Mapping function generation: After training, the GPR outputs the formation density-wind resistance intensity mapping function (FD-WRIMapping Function). Its mathematical essence is a probability distribution, but it appears to the user as a black box function: input the FD value and output the predicted mean and 95% confidence interval of the WRI.

[0050] The remaining 20% ​​of the data (10 groups) were used for model validation: the FD values ​​were input into the mapping function to obtain the predicted WRI values, which were then compared with the actual WRI values. The results showed that the mean absolute error (MAE) was 0.036 and the coefficient of determination R 2 =0.928. Key findings include: When FD = 0.48, WRI reaches a peak of 0.86 (confidence interval [0.83, 0.89]), corresponding to the optimal density; When FD < 0.35, the WRI decreases sharply (the sparse formation has weak wind resistance); When FD>0.75, WRI decreases (over-dense wake superposition deteriorates the flow field).

[0051] This function clearly reveals the non-monotonic relationship between density and wind resistance, providing a constraint basis for multi-objective optimization.

[0052] Set multi-objective optimization goals, use mapping functions as constraints, use genetic algorithms to search for Pareto optimal solutions, and output multiple sets of candidate formation structure parameters; The multi-objective optimization problem is defined as: Objective 1: Maximize wind resistance intensity (Max WRI); Objective 2: Minimize formation total energy consumption (Min Total Energy Consumption, TEC); Constraint: The FD-WRI mapping function requires WRI ≥ 0.75 (wind resistance threshold).

[0053] Among them, TEC is calculated by integrating the motor power: TEC = Σ(motor voltage × current × flight time), and is fitted into the formation density FD and wind speed v through historical data: TEC = k1 × FD 2 + k2×v 3 (k1=8.3, k2=0.17).

[0054] Genetic Algorithm (GA) parameter configuration: Coding scheme: real number coding, the chromosome contains 5 genes: horizontal spacing mean, height variance, yaw angle, pitch angle, roll angle; Population size: 200 individuals, initialized using Latin Hypercube Sampling (LHS) to cover the parameter space; Fitness evaluation: Calculate the WRI (by calling the mapping function through FD) and TEC (by the energy consumption model) of each individual; Selection operation: an elite retention strategy based on non-dominated sorting (NDS) and crowding distance (CD), retaining the top 50% of individuals; Genetic operations: simulated binary crossover (SBX, distribution index η_c = 15), polynomial mutation (PM, η_m = 20), and mutation probability 0.08.

[0055] The algorithm converges after 100 iterations and outputs the Pareto Front, which contains 45 non-dominated solutions. Each solution corresponds to a set of formation structure parameters, for example: Solution A (focusing on wind resistance): WRI=0.89, TEC=1200W, parameters=[horizontal spacing 2.8m, height variance 0.3m...]; Solution B (focusing on energy saving): WRI=0.76, TEC=850W, parameters=[horizontal spacing 4.2m, height variance 0.7m...].

[0056] These solutions are constructed into candidate formation structure parameter sets (CFSPs), covering the optimal trade-off solutions under different wind conditions.

[0057] A wind field simulation test is conducted on the candidate formation structure parameters to screen out structures that meet the wind disturbance resistance threshold and have energy consumption lower than the preset value. The structures are converted into relative position instructions and communication topology relationships for each UAV to generate dynamic formation topology instructions.

[0058] Wind farm simulation tests are performed on the digital twin platform: Scenario construction: set wind speed gradient (8 / 10 / 12 / 14m / s), wind direction change (±30° pulsation); Model loading: Import the formation structure in CFSP and test 10 random wind field sequences for each structure; Performance monitoring: Record key indicators: actual WRI (simulated by flow field sensors), TEC (calculated by motor model), position offset (PO); Screening rules: (a) Average WRI ≥ 0.75; (b) TEC < 1000W (preset energy consumption threshold); (c) PO < 1.0m (stability requirement) must all be met.

[0059] After testing, only 28 structures met the standards. Further robustness ranking was performed: under a wind speed disturbance of 12±2m / s, the performance standard deviation (σ_WRI) of each structure was calculated. The top 5 structures with σ_WRI < 0.05 were selected as the final solutions, for example: Structure γ: Under wind conditions of 12 m / s, WRI=0.83±0.02, TEC=920±15 W, PO=0.6±0.1 m; Structure δ: Under wind conditions of 14 m / s, WRI=0.78±0.03, TEC=880±20 W, PO=0.7±0.2 m.

[0060] Convert the preferred structure into executable instructions: Relative position command: Calculate the relative coordinates (Δx, Δy, Δz) of the member drones with the pilot drone as the origin. For example, the coordinates of drone No. 3 in structure γ are: [-2.8m, -1.6m, +0.3m]; Communication topology: Determine neighbor nodes based on the Voronoi diagram and generate an adjacency matrix. For example, machine 4 needs to maintain communication with machines {1, 3, 5}. Dynamic adaptability parameters: including density adjustment rate (FD_change_rate=0.05 / s), gust response threshold (Gust_threshold=3m / s 2 ).

[0061] The data is finally packaged into a Dynamic Formation Topology Command (DFTC) and sent to the cluster for execution via the data transmission radio.

[0062] S203, adjusting the relative positions and attitude angles of the UAVs based on the dynamic formation topology command using a distributed collaborative control algorithm, utilizing the updraft from the wake of the leading UAV to reduce energy consumption of the trailing UAVs, and simultaneously achieving group obstacle avoidance during sudden gusts through local information exchange, thereby generating a wind-resistant collaborative flight state; Specifically, it can parse the dynamic formation topology instructions, extract the target relative position and communication neighbor list of each UAV, calculate the position deviation based on the current position, and output the target position-deviation comparison table; Dynamic formation topology commands are the core operational basis for UAV swarms to conduct wind-resistant coordinated flight. Their data structure typically includes fields such as a command header, a timestamp, a formation topology type identifier, each UAV's unique device number, target relative coordinates, a communication range threshold, and a checksum. For example, the command header for a dynamic formation topology command might contain the fixed byte "0xFAFA" to identify the command's start. The timestamp, accurate to the millisecond level, such as "2025-07-25 10:30:15.123," ensures the command's temporal validity. The formation topology type identifier is represented by an enumerated value, such as "01" for a diamond formation and "02" for a wedge formation, enabling each UAV to quickly identify the formation pattern. Upon receiving the command, each UAV first performs an integrity check on the command through the command parsing module. This check can be performed using the CRC32 algorithm, which performs a cyclic redundancy check on all bytes in the command except the checksum. If the result matches the checksum, the command is considered valid. Otherwise, it is considered invalid and a retransmission request is made.

[0063] When extracting the relative position of targets, a local coordinate system is established with the lead drone in the formation as the origin, with the X-axis pointing directly in front of the formation, the Y-axis pointing to the right of the formation, and the Z-axis pointing upward, perpendicular to the XY plane. The target relative position of each follower drone is represented by three-dimensional coordinates (Δx, Δy, Δz). For example, the target relative position of a follower drone is (5.2m, 3.1m, 0.8m), indicating that the drone should be 5.2 meters in front of the lead drone, 3.1 meters to the right, and 0.8 meters above it. The communication neighbor list is extracted based on a preset communication distance threshold. For example, if the communication distance threshold is set to 10 meters, the Euclidean distance between the relative target positions of each drone is calculated, and the drone numbers with a distance of less than or equal to 10 meters are added to the corresponding drone's communication neighbor list. For example, the target relative position of the drone numbered UAV-001 is (0,0,0), and the target relative position of the drone numbered UAV-002 is (5.2m, 3.1m, 0.8m). The distance between the two is calculated to be approximately 6.1 meters, which is less than 10 meters. Therefore, UAV-002 will be included in the communication neighbor list of UAV-001, and vice versa.

[0064] Current location is determined by the drone's onboard multi-sensor fusion positioning system, which integrates a GPS module, an inertial measurement unit (IMU), and a visual odometry system. The GPS module provides global location information, such as longitude, latitude, and altitude, but signal drift can occur in complex environments. For example, in urban canyons or densely vegetated areas, positioning errors can reach 3-5 meters. The IMU achieves short-term, high-precision positioning by measuring acceleration and angular velocity, but errors accumulate over time. For example, a position error of 0.5 meters can occur within 10 seconds. The visual odometry calculates relative motion by capturing ground features, providing auxiliary positioning when GPS signals fail. For example, in indoor environments or tunnels, positioning accuracy can reach 0.1 meters. By fusing the data from these three sensors using a Kalman filter algorithm, the current position is output with centimeter-level accuracy. For example, the current fused position of a drone is (E116.3845°, N39.9088°, 120.5m).

[0065] Position deviation is calculated by converting each drone's current position into a local coordinate system centered on the pilot drone and then performing a difference calculation with the target's relative position. For example, if a drone's current local coordinates are (5.0m, 3.3m, 0.7m) and the target's relative position is (5.2m, 3.1m, 0.8m), its position deviation is (-0.2m, 0.2m, -0.1m). A target position-deviation comparison table is presented in tabular format, containing information such as the drone number, the target's relative position (Δx, Δy, Δz), the current local position (x, y, z), and the position deviation (dx, dy, dz), providing a clear basis for subsequent position adjustments. This comparison table is updated in real time, matching the sensor data acquisition frequency—typically 10Hz, or every 0.1 seconds—to ensure the timeliness of position deviation information.

[0066] In practical applications, fault-tolerant processing of the target position-deviation comparison table is also required to address the effects of data transmission delays and sensor noise. For example, if a drone does not receive a new dynamic formation topology instruction within three consecutive sampling cycles, it automatically activates the locally cached instruction from the previous moment and predicts the target position based on its own motion state to avoid formation confusion caused by missing instructions. At the same time, the calculated position deviation is checked for rationality. When the deviation exceeds a preset threshold (e.g., 10 meters), it is determined to be a sensor failure or data anomaly, triggering a redundant sensor switching mechanism to ensure the accuracy of the position deviation information.

[0067] Based on the target position-deviation comparison table, a distributed model predictive control algorithm is adopted. Each UAV only exchanges position information with its communication neighbors, autonomously calculates the speed adjustment, and realizes position closed-loop regulation through the PID controller, outputting a real-time position correction sequence. The core of the distributed model predictive control algorithm lies in each drone constructing an optimization problem locally and achieving coordinated control through rolling optimization, eliminating the need for centralized decision-making by a central controller. In its implementation, the prediction and control domains are first defined for each drone. For example, the prediction domain Np is set to 10 (predicting the position state for the next 10 sampling periods) and the control domain Nc is set to 5 (optimizing the control input for the next 5 sampling periods). Based on the position deviation in the target position-deviation comparison table, a kinematic model of the drone is established. This model uses the current position (x, y, z) and velocity (vx, vy, vz) as state variables, and acceleration (ax, ay, az) as control input. The future position state is predicted using the Euler integral formula. For example, if the position of a drone at the current time t is (x(t), y(t), z(t)) and the velocity is (vx(t), vy(t), vz(t)), then the predicted position at time t+1 is x(t+1) = x(t) + vx(t)・Δt + 0.5・ax(t)・Δt 2 , where Δt is the sampling period, which is set to 0.1 seconds. Similarly, y (t+1) and z (t+1) can be calculated.

[0068] In the design of the optimization objective function, three aspects are comprehensively considered: position tracking accuracy, control energy consumption, and neighbor coordination. The position tracking accuracy term uses the square difference between the predicted position and the target position. For example, for the kth moment in the prediction time domain, the cost is (x_pred (k) - x_ref (k)) 2 + (y_pred (k) - y_ref (k)) 2 + (z_pred(k)-z_ref(k)) 2 , where x_ref (k) is the target position at the kth moment. The control energy consumption term is the sum of the squares of the control input (acceleration), for example, Σ(ax (k) 2 + ay (k) 2 + az (k) 2 ), to avoid excessive acceleration leading to mechanical loss of the drone. The neighbor cooperation term is reflected by the predicted position deviation from the communication neighbor. For example, for neighbor drone j, the cost is (x_pred_i (k)-x_pred_j (k)-Δx_ij) 2 , where Δx_ij is the component of the target relative distance between drones i and j on the x-axis, ensuring the stability of the relative position of the formation.

[0069] Each drone uses its local communication module to exchange predicted position information for the next three sampling periods with drones in its communication neighbor list. The communication data format includes the drone number, predicted time, predicted position (x, y, z), and a data checksum. The communication frequency is 10Hz, which matches the sampling period. After receiving the neighbor's predicted position, it incorporates it into its own optimization objective function to implement group coordination constraints. For example, when calculating the optimization problem, drone i considers the predicted position of neighbor drone j and uses constraints to ensure that the relative position deviation between the two does not exceed ±0.5 meters of the target value, that is, |x_pred_i (k) - x_pred_j (k) - Δx_ij| ≤ 0.5m.

[0070] The optimization problem is solved using the interior point method, which iteratively finds the control input sequence (acceleration) that minimizes the objective function. In each iteration, each drone uses only local information and the predicted positions of its neighbors, without requiring global information, significantly reducing the communication burden and computational complexity. For example, when solving a problem, a drone is initialized to zero acceleration and then adjusted using the gradient descent method until the objective function converges to the minimum value. The first Nc control inputs obtained at this time are the optimal acceleration sequence at the current moment. The speed adjustment is calculated by integrating the acceleration. For example, the speed adjustment corresponding to the acceleration ax (1) of the first control input is Δvx = ax (1)・Δt. Similarly, Δvy and Δvz can be obtained.

[0071] The PID controller, as the actuator for closed-loop position control, converts the speed adjustment output by the distributed model predictive control algorithm into a specific motor control signal. The PID controller's parameters need to be adjusted based on the drone's dynamic characteristics. For example, the proportional coefficient Kp, integral coefficient Ki, and differential coefficient Kd are set to 5.0, 0.1, and 0.5, respectively. The proportional term is used to quickly respond to position deviations. For example, when the position deviation dx = 0.5 meters, the proportional term output is Kp・dx = 2.5 m / s. The integral term is used to eliminate steady-state errors. When there are persistent small deviations (such as 0.1 meters), the integral term accumulates the error and outputs the control variable until the deviation reaches zero. The differential term is used to suppress overshoot by predicting the deviation trend and adjusting the control variable in advance. For example, when the deviation increases over time, the differential term outputs a negative value, slowing down the control action.

[0072] The closed-loop position adjustment process is a continuous feedback control process. The drone collects its current position in real time, calculates the deviation from the target position, and inputs this deviation into the PID controller. The controller calculates a speed command based on the proportional, integral, and differential components of the deviation. This command is then combined with the speed adjustment output by the distributed model predictive control algorithm to produce the final speed control command. For example, if the PID controller outputs a speed command of vx_cmd = 3.0 m / s and the model predictive control outputs a speed adjustment of Δvx = 0.5 m / s, the final speed command is vx = vx_cmd + Δvx = 3.5 m / s. The motor drive module adjusts the propeller speed based on the speed command, changing the drone's flight speed and, in turn, its position, forming a closed loop of "position measurement - deviation calculation - control output - speed adjustment - position change."

[0073] The real-time position correction sequence is the output of closed-loop position control. It contains position corrections (dx, dy, dz) for the next five sampling periods. Each correction corresponds to a specific moment in time. For example, the first correction is the position adjustment value for the current moment + 0.1 seconds, the second is the position adjustment value for the current moment + 0.2 seconds, and so on. Position corrections are calculated based on the current velocity adjustment and the drone's kinematic model. For example, at a velocity of vx = 3.5 m / s, the position correction in 0.1 seconds is dx = vx 0.1 = 0.35 meters. This sequence is stored locally and shared with neighbors for optimization calculations in the next round of distributed model predictive control, ensuring the continuity and stability of formation adjustments.

[0074] In actual operation, in order to improve the robustness of the system, the PID controller needs to be anti-saturation processed. When the speed command exceeds the maximum speed limit of the drone (such as 10m / s), it is automatically truncated to the maximum value, and the saturation error of the integral term is recorded and compensated in subsequent control to avoid system instability caused by integral saturation. At the same time, the position deviation signal is preprocessed through sliding window filtering, and the window size is set to 5 sampling periods to reduce the impact of noise on the control effect. For example, the original position deviation at a certain moment is 0.52 meters, and the deviations at the first four moments are 0.48, 0.50, 0.51, and 0.49 meters respectively. The deviation after filtering is (0.48+0.50+0.51+0.49+0.52) / 5=0.50 meters, making the control output smoother.

[0075] Based on the real-time position correction sequence and the wind direction angle of the three-dimensional wind field, the pitch and yaw angles of each UAV are optimized through the aerodynamic model, the attitude angle adjustment amount is calculated, and the attitude collaborative control instructions are output; The real-time position correction sequence provides the UAV's position adjustment plan for the next period of time, while the wind direction angle in the three-dimensional wind field directly influences the UAV's attitude adjustment strategy. The wind direction angle in the three-dimensional wind field typically includes horizontal and vertical wind angles. The horizontal wind angle refers to the angle between the wind's direction on the horizontal plane and due north, ranging from 0° to 360°. For example, a north wind has a horizontal wind direction angle of 0° and an east wind has a horizontal wind direction angle of 90°. The vertical wind direction angle refers to the angle between the vertical component of the wind and the horizontal plane, with upward winds being positive and ranging from -90° to 90°. For example, an updraft has a vertical wind direction angle of 30° and a downdraft has a vertical wind direction angle of -20°. These wind direction angle data are collected in real time by the UAV's onboard ultrasonic anemometer and wind vane, with a sampling frequency of 5Hz, meaning it is updated every 0.2 seconds. High-frequency noise is removed through a low-pass filter, such as a Butterworth filter with a cutoff frequency of 1Hz, to smooth the wind direction angle data.

[0076] The aerodynamic model is the core tool for optimizing attitude angles. Based on the lift formula, drag formula, and torque balance equation, the model calculates the aerodynamic force and torque of the drone at different attitude angles. The calculation formula for lift L is L=0.5・ρ・v 2 ・S・Cl, where ρ is the air density (e.g. 1.225 kg / m 3 ), v is the speed of the drone relative to the airflow (including flight speed and wind speed), S is the wing area (such as 0.5m 2 ), Cl is the lift coefficient, and its value is related to the pitch angle α. For example, when α=5°, Cl=0.8; when α=10°, Cl=1.2. The calculation formula of drag D is D=0.5・ρ・v 2 ・S・Cd, where Cd is the drag coefficient, which is related to the yaw angle β. For example, when β=0°, Cd=0.05; when β=3°, Cd=0.07.

[0077] Based on a real-time position correction sequence, the drone can predict its future flight speed and calculate the magnitude and direction of the relative airflow velocity based on the wind direction angle in the three-dimensional wind field. For example, if the drone's flight speed is (10 m / s, 0, 0), the horizontal wind direction angle is 90° (easterly wind), the wind speed is 3 m / s, and the vertical wind direction angle is 10°, the wind speed is 1 m / s. The x-axis component of the relative airflow velocity in the drone's coordinate system is 10-3 sin (90°) = 7 m / s, the y-axis component is -3 cos (90°) = 0 m / s, and the z-axis component is -1 sin (10°) ≈ -0.17 m / s. Based on the relative airflow velocity, an aerodynamic model calculates the lift, drag, and torque at the current attitude angle. This is then compared with the desired lift (used to offset gravity and achieve position correction) to determine the required attitude angle adjustment.

[0078] The optimization of pitch angle is mainly to balance lift and drag to ensure the vertical position correction of the UAV. For example, when the real-time position correction sequence requires the UAV to rise 0.3 meters in 0.5 seconds, it is necessary to increase lift. In this case, the lift coefficient Cl is increased by increasing the pitch angle α. Assuming that α = 5°, Cl = 0.8, the lift L = 0.5・1.225・7 2 0.5-0.8 ≈ 11.8 N, which is less than the drone's gravity of 15 N and shows a downward trend. Through iterative calculations, when α = 8°, Cl = 1.0, and the lift force L ≈ 14.7 N, close to gravity. At this point, α is fine-tuned to 8.5°, bringing the lift slightly above gravity and meeting the required ascent. At the same time, the effects of drag must be considered to avoid excessively large pitch angles that can cause a significant increase in drag. For example, when α exceeds 15°, stalling may occur, and Cl drops sharply. Therefore, the pitch angle adjustment range is limited to between -5° and 15°.

[0079] Yaw angle optimization is used to reduce lateral drag and ensure horizontal position correction accuracy. For example, when a 3D wind field contains a crosswind (horizontal wind angle of 45°, wind speed of 2 m / s), the relative airflow against the drone has a component in the y-axis direction, resulting in increased lateral drag. By adjusting the yaw angle β to align the longitudinal axis of the drone with the relative airflow, that is, β equals the lateral angle of the relative airflow, Cd is minimized, and lateral drag is minimized. Assuming the lateral angle of the relative airflow is 3°, adjusting the yaw angle to 3° will align the drone directly with the airflow, reducing energy consumption. The yaw angle adjustment range is typically limited to -10° to 10° to prevent excessive yaw from disrupting the relative position of the formation.

[0080] The attitude angle adjustment is calculated using a gradient descent method, with the lift and drag deviations as the objective function. Pitch and yaw angles are iteratively optimized. For example, if the initial attitude angles are (α=5°, β=0°), the calculated lift deviation is ΔL=3.2N (needing an increase of 3.2N) and the drag deviation is ΔD=0.5N (needing a decrease of 0.5N). By calculating the partial derivatives of the objective function with respect to α and β, the adjustment direction is determined: α should be increased, and β should be adjusted appropriately. Each iteration adjusts the angle by 0.5°, recalculating lift and drag until the deviations fall below a preset threshold (e.g., ΔL<0.1N, ΔD<0.05N). The resulting attitude angle is then considered the optimized angle, and the adjustment amount is the difference between the optimized and current angles.

[0081] The coordinated attitude control command contains each UAV's target pitch angle, target yaw angle, and adjustment rate, for example, "UAV-001, α = 8.5°, β = 3°, adjustment rate 1° / s." This command is sent to the UAV's attitude control system via an internal bus. The attitude control system uses PID control to achieve closed-loop adjustment of the attitude angles. The proportional, integral, and differential coefficients are tuned for pitch and yaw, respectively. For example, the pitch PID parameters are Kp = 2.0, Ki = 0.05, Kd = 0.1, and the yaw PID parameters are Kp = 1.5, Ki = 0.03, Kd = 0.08. During the adjustment process, each UAV exchanges attitude angle information through communication with its neighbors to ensure coordinated attitude changes and avoid increased airflow interference caused by sudden attitude changes. For example, when UAV i needs to increase its pitch angle, it sends a warning message to its neighbor UAV j, prompting j to adjust its attitude in advance to adapt to the airflow change.

[0082] To cope with dynamic wind field changes, attitude angle optimization must be performed continuously. The optimization cycle is consistent with the wind direction angle sampling cycle, which is 0.2 seconds. During each optimization cycle, the attitude angle adjustment is recalculated based on the latest three-dimensional wind field data and the real-time position correction sequence, and the attitude coordinated control instructions are updated. Simultaneously, the optimized attitude angle is checked for feasibility. For example, if the calculated pitch angle exceeds 15°, it is automatically limited to 15°, and the flight speed is increased to compensate for the lack of lift, ensuring the safety of the UAV. In addition, a historical database of attitude angle adjustments is established, recording the optimal attitude angles under different wind field conditions. Through data mining, patterns in attitude adjustment are discovered, which are used to guide subsequent optimization processes and improve optimization efficiency.

[0083] Based on attitude collaborative control instructions, the airborne meteorological sensor is used to identify the rising airflow area of ​​the wake of the leading aircraft, guide the following aircraft into the area and adjust the altitude, calculate the lift gain and energy consumption reduction value of each following aircraft, and output the wake utilization parameters; each UAV uses short-range communication to exchange gust monitoring data in real time, adopts a local obstacle avoidance algorithm based on the potential field method, adjusts the velocity vector to achieve group obstacle avoidance, integrates position correction, attitude collaborative control, and wake utilization parameters, and generates an anti-wind disturbance collaborative flight state.

[0084] Airborne meteorological sensors are key to identifying areas of updrafts in the wake of preceding aircraft. They primarily include LiDAR (Light Detection and Ranging) and micro-weather stations. LiDAR scans the airspace ahead with a laser beam, measuring changes in the atmospheric refractive index to detect air velocity and turbulence intensity. Its detection range is 50 meters ahead, with a horizontal and vertical resolution of 1 meter and a sampling frequency of 2 Hz. For example, during the LiDAR scan, it generates a three-dimensional air velocity matrix measuring 50 × 50 × 50 cubic meters, where positive values ​​indicate updrafts and negative values ​​indicate downdrafts. The micro-weather station measures meteorological parameters such as temperature, air pressure, and humidity, and combines them with aerodynamic models to calculate the density and lift potential of the airflow. For example, if the air pressure in a certain area is detected to be 0.5 hPa lower than the surrounding air pressure, an updraft may be present.

[0085] The updraft areas in the wake of the leading aircraft exhibit specific spatial distribution characteristics, similar to the "updraft vortex" formed by the wake of the leading bird in a flying formation. For example, in a diamond formation, the wake of the leading aircraft (the pilot) creates two symmetrical updraft areas on either side of it, located 10-20 meters from the leading aircraft, 1-3 meters lower in height, and approximately 5 meters wide. These characteristic areas are identified by clustering lidar data using the DBSCAN algorithm (density-based spatial clustering of noise applications), with the eps parameter set to 2 meters (cluster radius) and the min_samples parameter set to 5 (minimum number of points). For example, if more than five adjacent lidar sampling points within a region show updraft velocities greater than 0.5 m / s, the region is identified as an updraft area, and its center coordinates (relative to the leading aircraft) and extent are output.

[0086] Guiding a trailing aircraft into an updraft requires precise path planning. Based on the leading aircraft's position and the updraft coordinates, the trailing aircraft's target entry point and flight path are calculated. For example, if the leading aircraft's current position is (x0, y0, z0), and the updraft center coordinates are (x0+15m, y0+3m, z0-2m), the trailing aircraft's target entry point is set to these center coordinates. The flight path is a straight line from the trailing aircraft's current position to the target entry point. The path length is calculated using Euclidean distance, for example, 20 meters. The flight time is set to 5 seconds based on the trailing aircraft's current speed. During the guidance process, the trailing aircraft's pitch and yaw angles are adjusted in conjunction with attitude coordinated control commands. For example, when approaching an updraft, the pitch angle can be reduced by 0.5° to slow the descent rate and ensure a smooth entry into the updraft.

[0087] Altitude adjustments after the trailing aircraft enters an updraft are based on real-time air velocity measurements. A micro-weather station monitors the updraft velocity at the current location. When the velocity exceeds 0.8 m / s, the aircraft lowers its altitude by 0.5 m to position itself within the strongest updraft. When the velocity drops below 0.3 m / s, the aircraft raises its altitude by 0.5 m to seek out a stronger updraft. Altitude adjustments are limited to a ±2-meter range to avoid significant deviation from the formation's overall altitude. For example, if the trailing aircraft initially enters at an altitude of z0-2 m and measures an air velocity of 1.0 m / s, it is then lowered to z0-2.5 m. At this point, the air velocity increases to 1.2 m / s, achieving optimal lift.

[0088] Lift gain is calculated based on the difference in lift before and after entering the updraft. The formula is ΔL = L_after - L_before, where L_after is the lift after entering the updraft and L_before is the lift before entering. Both are calculated using an aerodynamic model. For example, if the relative air velocity before entering is 7 m / s, the lift L_before = 14.7 N. After entering the updraft, the relative air velocity increases to 8.2 m / s due to the updraft, and the lift L_after = 0.5・1.225・8.2 2 0.5 / 1.0 ≈ 16.5 N, resulting in a lift gain of ΔL = 1.8 N. Energy reduction is directly related to lift gain, as increased lift reduces motor output power. Energy reduction ΔE = (P_before - P_after) t, where P_before and P_after are the motor power before and after entry, respectively, and t is the time spent in the updraft. Assuming P_before = 500 W, P_after = 450 W, and a dwell time of t = 60 seconds, ΔE = 50 W × 60 seconds = 3000 J.

[0089] Wake utilization parameters include information such as the center coordinates of the updraft region, its extent, average ascent velocity, the target entry point for the following aircraft, lift gain, and energy consumption reduction. For example, "Updraft region: (x0+15m, y0+3m, z0-2m), extent 5×5×2m, average ascent velocity 1.2m / s; target entry point for the following aircraft, UAV-002: same as the region center, lift gain 1.8N, energy consumption reduction 3000J." These parameters are transmitted to the relevant following aircraft via a communication network and updated in real time at a frequency consistent with the lidar sampling rate of 2Hz. A wake utilization effectiveness evaluation mechanism is also established. If the lift gain of a following aircraft is less than 0.5N for three consecutive cycles, the wake region is deemed ineffective and the following aircraft is directed to leave and seek out the wake region of another preceding aircraft.

[0090] Real-time exchange of gust monitoring data is achieved through a short-range communication module using the ZigBee protocol. The communication range is 30 meters, and the data transmission rate is 250kbps, ensuring data exchange within 100ms. Gust monitoring data includes gust intensity (sudden increase in wind speed), gust direction, and duration. For example, "UAV-003 detected a gust of 5 m / s, direction 135°, and duration of approximately 2 seconds." After receiving gust data from its neighbors, each drone combines its own monitoring results with a data fusion algorithm (such as weighted averaging) to determine the overall impact and intensity of the gust. For example, if three adjacent drones all detect gusts of 4-6 m / s and direction 130°-140°, the fusion results will be determined as a regional gust of 5 m / s and direction 135°.

[0091] The local obstacle avoidance algorithm based on the potential field method treats the drone as a point mass moving in a virtual force field. Obstacles (including other drones and areas affected by sudden gusts of wind) generate repulsive forces, while the target position generates attractive forces. The direction of the drone's movement is determined by the direction of the resultant force. The magnitude of the repulsive force is inversely proportional to the distance, and the formula is F_rep = k_rep / d 2 , where k_rep is the repulsion coefficient (set to 10) and d is the distance to the obstacle. The attractive force is proportional to the distance, as shown in the formula F_attr = k_attr·d, where k_attr is the attraction coefficient (set to 0.5). For example, if a drone is 3 meters away from the edge of the gust area, the repulsive force it experiences is F_rep = 10 / 3 2 ≈1.11N; 5 meters away from the target position, the attractive force F_attr=0.5×5=2.5N, the direction of the resultant force is toward the target position, but slightly away from the gust area.

[0092] The velocity vector is adjusted based on the direction of the resultant force, which is converted into a velocity direction. The velocity is adjusted based on the distance from the target location, decreasing with distance. For example, at a distance of 5 meters, the velocity is 3 m / s, while at a distance of 1 meter, it is 1 m / s. For example, if the resultant force is at a 15° angle to the original flight direction, the velocity vector is adjusted by 15°, maintaining a constant velocity of 3 m / s. This allows the drone to continue moving toward the target while avoiding gusts of wind. During the adjustment process, each drone synchronizes velocity vector information through communication to ensure that adjacent drones' adjustment directions do not conflict. For example, if drone i adjusts to the left, drone j on its right will be notified to temporarily halt its rightward adjustment to avoid a collision.

[0093] Collaborative group obstacle avoidance is achieved through a "follow-and-avoid" mechanism. When a drone detects a gust and begins adjusting its velocity vector, the drones behind it monitor its trajectory and make preemptive avoidance preparations. For example, if drone i adjusts 5° to the left due to a gust, drone j, upon receiving this information, will also adjust 3° to the left, maintaining their relative positions. The obstacle avoidance process continues until the gust monitoring data indicates that the gust intensity has dropped below a safe threshold (e.g., 2 m / s). At this point, each drone gradually restores its original velocity vector and returns to its normal formation position.

[0094] The generation of anti-wind disturbance cooperative flight state is the integration of parameters such as position correction, attitude coordinated control and wake utilization, forming a comprehensive data set including the position (x, y, z), speed (vx, vy, vz), attitude angle (α, β, γ), lift, energy consumption, wake utilization state and obstacle avoidance state of each UAV. For example, the cooperative flight status of UAV-001 is "position (100.5, 50.3, 120.8), speed (10.2, 0.5, 0.3), attitude (8.5°, 3°, 0°), lift 16.2N, energy consumption 420W, wake utilization: none, obstacle avoidance: none"; the status of UAV-002 is "position (115.3, 53.2, 118.3), speed (10.1, 0.4, 0.2), attitude (7.8°, 2.5°, 0°), lift 17.5N (gain 2.8N), energy consumption 380W (reduction 40W), wake utilization: UAV-001 wake zone, obstacle avoidance: none".

[0095] This dataset is stored in each drone's local database using distributed storage technology and is periodically transmitted (every second) to the ground control center for status monitoring. A collaborative status evaluation indicator system was established, including formation relative position error (less than 1 meter), average energy consumption reduction rate (greater than 10%), and obstacle avoidance success rate (100%). These indicators are calculated in real time to evaluate the effectiveness of wind interference coordination. When these indicators fall short (e.g., relative position error exceeds 1.5 meters), re-optimization of collaborative control parameters is triggered to continuously improve wind interference performance. Furthermore, collaborative flight status data is used to train a machine learning model to predict the optimal collaborative strategy under different wind conditions, enhancing the system's adaptability.

[0096] S204 continuously monitors the deviation between the three-dimensional wind field changes and the coordinated flight status, dynamically corrects the weight of the formation density-anti-wind disturbance intensity mapping relationship through the reinforcement learning algorithm, updates the dynamic formation topology instructions, and realizes the adaptive control of the bionic group's anti-wind disturbance coordination.

[0097] Specifically, the distributed sensors of the UAV cluster can collect the current three-dimensional wind field data and the flight status data of each aircraft in real time, calculate the deviation between the actual flight status and the anti-wind disturbance collaborative flight status, and output the state deviation matrix; The drone swarm's distributed sensor system is the foundation for real-time data collection. This system comprises multiple sensor types, with each drone equipped with a complete sensor suite to ensure comprehensive and independent data collection. Three-dimensional wind field data is primarily collected using ultrasonic anemometers and five-hole probes. Ultrasonic anemometers measure three-dimensional wind speed components using the principle that the speed of sound waves propagating through air is affected by wind speed. Their sampling frequency is 10 Hz, meaning they output data every 0.1 seconds. The measurement range is 0-20 m / s, with an accuracy of ±0.1 m / s. For example, at a given moment, the three-dimensional wind speed data collected by the ultrasonic anemometer is (vx = 3.2 m / s, vy = -1.5 m / s, vz = 0.8 m / s), representing the wind speeds along the x-axis (flight direction), y-axis (lateral), and z-axis (vertical), respectively. The five-hole probe calculates the airflow speed and direction by measuring the pressure difference in different directions. It serves as a redundant backup for the ultrasonic anemometer and automatically switches when the ultrasonic anemometer fails. Its measurement accuracy is slightly lower, at ±0.2m / s, but its response speed is faster, and it can provide more timely data when gusts occur.

[0098] The flight status data collection of each aircraft is completed by the global navigation satellite system (GNSS), inertial measurement unit (IMU) and barometric altimeter. GNSS includes GPS and Beidou dual-mode positioning, providing centimeter-level location information (latitude, longitude and altitude), with a sampling frequency of 5Hz. For example, the GNSS data of a certain drone is (longitude = 116.3845°, latitude = 39.9088°, altitude = 120.5m). The IMU consists of a three-axis accelerometer, a three-axis gyroscope and a three-axis magnetometer, which measures the acceleration, angular velocity and heading angle of the drone. The sampling frequency is 100Hz and is used to capture high-frequency attitude changes. For example, the IMU data at a certain moment is (acceleration ax = 0.2m / s 2 , ay=-0.1m / s 2 , az=-9.8m / s 2 Angular velocities ωx = 0.05 rad / s, ωy = -0.03 rad / s, ωz = 0.01 rad / s; heading angle ψ = 30.5°). The barometric altimeter calculates relative altitude by measuring atmospheric pressure, assisting GNSS altitude measurement. The sampling frequency is 20 Hz, and the accuracy is ±0.1 m.

[0099] To ensure data consistency and timeliness, the distributed sensor system uses a time synchronization protocol (such as the IEEE1588 Precision Time Protocol) for clock calibration, keeping the timestamp error of each sensor within 1ms. The collected raw data first undergoes preprocessing, including outlier detection and filtering. Outlier detection uses the 3σ criterion: when a data point deviates from the mean by more than three standard deviations, it is identified as an outlier and replaced using linear interpolation. For example, in 3D wind field data, vx = 15.3 m / s at a certain moment. The mean for that period is 3.5 m / s, the standard deviation is 1.2 m / s, and three standard deviations equal 3.6 m / s. Since 15.3 - 3.5 = 11.8, which is greater than 3.6, it is considered an outlier and replaced with the average of the preceding and following moments. The filtering process uses Kalman filtering, which establishes the system state equation and observation equation, fuses multi-sensor data and removes noise. For example, it fuses the GNSS position data and the IMU acceleration data to obtain smoother and more accurate position and velocity information.

[0100] The actual flight state refers to preprocessed real-time UAV flight data, including parameters such as position (x, y, z), velocity (vx, vy, vz), attitude angles (pitch angle θ, roll angle φ, heading angle ψ), acceleration (ax, ay, az), and angular velocity (ωx, ωy, ωz). The wind-resistance coordinated flight state is the ideal flight state expected based on the dynamic formation topology instructions and wind-resistance strategy. This state is precalculated by the system based on the current wind conditions and formation structure. For example, the wind-resistance coordinated flight state for a given UAV is (x = 100.0 m, y = 50.0 m, z = 120.0 m; vx = 10.0 m / s, vy = 0 m / s, vz = 0 m / s; θ = 5.0°, φ = 0°, ψ = 30.0°).

[0101] Deviations are calculated by subtracting each parameter from the actual flight state from the corresponding parameter for wind-resistance coordinated flight. For example, if the actual position x = 100.5 m and the expected x = 100.0 m, the position deviation dx = 0.5 m; the actual velocity vx = 9.8 m / s and the expected vx = 10.0 m / s, the velocity deviation dvx = -0.2 m / s; the actual pitch angle θ = 5.3° and the expected θ = 5.0°, the attitude angle deviation dθ = 0.3°. For vector parameters (such as position and velocity), the deviation of each component is calculated separately; for scalar parameters (such as energy consumption), the numerical difference is directly calculated.

[0102] The state deviation matrix is ​​a matrix formed by arranging the deviation values ​​of all drones in a certain order. The number of rows in the matrix is ​​the number of drones, and the number of columns is the number of deviation parameters. For example, for a swarm of 3 drones, considering the position (3 components), velocity (3 components), and attitude angle (3 components), a total of 9 deviation parameters, the state deviation matrix is ​​a 3×9 matrix: [dx1, dy1, dz1, dvx1, dvy1, dvz1, dθ1, dφ1, dψ1; dx2, dy2, dz2, dvx2, dvy2, dvz2, dθ2, dφ2, dψ2; dx3, dy3, dz3, dvx3, dvy3, dvz3, dθ3, dφ3, dψ3] Here, dx1 represents the position deviation of the first UAV in the x-direction, and so on. The element values ​​of the state deviation matrix can be positive or negative. A positive value indicates that the actual parameter is greater than the expected parameter, and a negative value indicates that the actual parameter is less than the expected parameter.

[0103] To facilitate subsequent reinforcement learning and control decision-making, the state deviation matrix needs to be normalized, mapping each deviation value to the range [-1, 1]. The normalization formula is: Deviation value normalization = Deviation value / Maximum deviation threshold. The maximum deviation threshold is set based on the physical meaning of the parameter and system requirements. For example, the maximum threshold for position deviation is 10m, the maximum threshold for velocity deviation is 5m / s, and the maximum threshold for attitude angle deviation is 10°. For example, if dx = 0.5m and the maximum threshold is 10m, the normalized dx_norm = 0.5 / 10 = 0.05.

[0104] The state deviation matrix output frequency is consistent with the data acquisition frequency, 10Hz, ensuring real-time reflection of any deviations between the drone's flight status and the expected state. The system also performs continuity checks on the state deviation matrix. If a drone's deviation value exceeds a preset warning threshold (e.g., position deviation exceeding 2 meters, speed deviation exceeding 1 meter / s) for five consecutive sampling periods, an alarm is triggered, indicating that the drone may be experiencing an anomaly and requires focused monitoring and adjustment.

[0105] In addition, the state deviation matrix is ​​associated with the three-dimensional wind field data, recording the wind field conditions corresponding to each deviation value, such as wind speed and wind direction, to facilitate subsequent analysis of the wind field's impact on flight state deviations. For example, by adding the corresponding wind field parameters (vx_wind, vy_wind, vz_wind) to the end of each row in the state deviation matrix, an expanded state deviation matrix is ​​formed, providing richer input information for subsequent reinforcement learning.

[0106] The state deviation matrix and the current wind field characteristics are used as state inputs to construct a Markov decision process for reinforcement learning. The deep deterministic policy gradient algorithm is used to train the weight correction model and output the weight adjustment gradient of the mapping relationship. The state deviation matrix and the current wind field characteristics together constitute the state input for reinforcement learning. These inputs require feature engineering to extract information useful for decision-making. The current wind field characteristics include parameters such as three-dimensional wind speed (vx_wind, vy_wind, vz_wind), the rate of change of wind speed (dvx_wind / dt, dvy_wind / dt, dvz_wind / dt), the turbulence intensity of the wind field (calculated by the variance of wind speed, I = σ_v / v_mean, where σ_v is the standard deviation of wind speed and v_mean is the mean wind speed), and wind direction (horizontal wind direction angle α and vertical wind direction angle β). For example, the wind field characteristics at a certain moment are (vx_wind = 3.0 m / s, vy_wind = -1.0 m / s, vz_wind = 0.5 m / s; dvx_wind / dt = 0.2 m / s). 2 , dvy_wind / dt=-0.1m / s 2 , dvz_wind / dt=0.05m / s 2 ; I=0.15; α=135°, β=10°).

[0107] The state deviation matrix is ​​fused with the current wind field characteristics to form a high-dimensional state vector. This fusion method uses feature concatenation, which expands the state deviation matrix row by row into a one-dimensional vector and then concatenates it with the one-dimensional wind field feature vector. For example, the 9-dimensional state deviation matrix of three drones is expanded into a 27-dimensional vector, which is concatenated with the 9-dimensional wind field feature vector to form a 36-dimensional state input vector. To improve the model's generalization, these features can also be normalized so that each feature has a mean of 0 and a standard deviation of 1. For example, if the original value of a feature is 5.0, the mean is 3.0, and the standard deviation is 1.0, the normalized value is (5.0 - 3.0) / 1.0 = 2.0.

[0108] The Markov decision process (MDP) is the basic framework of reinforcement learning, consisting of state (S), action (A), reward (R), state transition probability (P), and discount factor (γ). In this method, the state S is the processed state input vector (the fusion vector of the state deviation matrix and the wind field characteristics); the action A is the adjustment of each weight in the formation density-anti-wind disturbance strength mapping relationship. These weights determine the importance of different features in the mapping relationship. For example, if a mapping relationship has 5 weight parameters, then the action is a 5-dimensional vector, and each element represents the adjustment value of the corresponding weight (positive or negative, ranging from [-0.1, 0.1]); the reward R is an indicator to measure the effect of the action. It is designed to encourage the reduction of state deviation and energy consumption. The calculation formula is R = -λ1×|deviation energy|-λ2×energy consumption increase+λ3×anti-wind disturbance strength improvement, where λ1, λ2, λ3 are weight coefficients (e.g., λ1=0.5, λ2=0.3, and λ3=0.2). The deviation energy is the sum of the squares of all elements in the state deviation matrix. The energy consumption increment is the difference between the cluster's energy consumption before and after the action is executed. The wind resistance improvement is the difference between the wind resistance after and before the action is executed. For example, after an action is executed, the deviation energy is 10, the energy consumption increment is 5, and the wind resistance improvement is 3. Then the reward R = -0.5×10 -0.3×5 +0.2×3 = -5 -1.5 +0.6 = -5.9.

[0109] The state transition probability P describes the probability of transitioning to the next state after executing an action in the current state. Due to the dynamic characteristics of drone swarms and environmental uncertainty, state transition probabilities are difficult to accurately model. Therefore, in practical applications, state transition patterns are estimated based on empirical data. The discount factor γ is used to balance immediate rewards with future rewards. Its value range is [0, 1]. Here, it is set to 0.9, indicating that it prioritizes immediate rewards while also considering future long-term rewards.

[0110] The Deep Deterministic Policy Gradient (DDPG) algorithm is a reinforcement learning algorithm for continuous action spaces. It combines deep neural networks with deterministic policy gradients and consists of an actor network and a critic network. The actor network outputs a deterministic action based on the current state. Its architecture uses a three-layer fully connected neural network. Its input is a 36-dimensional state vector. The hidden layers have 128 and 64 neurons, respectively, using Reluctant Unit (ReLU) as the activation function. The output layer has five neurons (corresponding to five weight adjustments). The activation function uses the tanh function, which constrains the output to the range [-0.1, 0.1]. For example, after inputting a state vector, the actor network outputs the action vector (0.02, -0.01, 0.03, 0.00, -0.02), indicating adjustments to the five weights by +0.02, -0.01, +0.03, 0, and -0.02, respectively.

[0111] The Critic network evaluates the value of the actions output by the Actor network. Given the current state and action, it calculates the Q-value (state-action value function). Its network structure is also a three-layer fully connected neural network. Its input is the concatenation of a 36-dimensional state vector and a 5-dimensional action vector (a total of 41 dimensions). The hidden layers have 128 and 64 neurons, respectively, with a Reluctant Linear Unit (ReLU) activation function. The output layer consists of one neuron, representing the Q-value. For example, for state S and action A, the Critic network outputs Q(S,A)=-3.5, indicating that the expected cumulative reward for executing action A in state S is -3.5.

[0112] The training process of the weight correction model consists of three main steps: exploration and exploitation, experience replay, and parameter updates. Exploration and exploitation aims to strike a balance between acquiring more environmental information (exploration) and leveraging known information to maximize reward (exploitation). This is achieved by adding Gaussian noise (mean 0, standard deviation 0.01) to the actions output by the actor network. For example, the actor network output actions (0.02, -0.01, 0.03, 0.00, -0.02) become (0.025, -0.008, 0.032, 0.005, -0.018) after adding noise. Experience replay stores the experience (S, A, R, S') of each interaction in an experience pool. When the number of samples in the experience pool reaches a preset threshold (e.g., 10,000), a batch of samples (e.g., 64) is randomly sampled for training to break down correlation between samples and improve training stability.

[0113] Parameters are updated using gradient descent. The goal of the Actor network is to maximize the Q value, and its loss function is -Q(S,A). The gradient of the Actor network parameters is calculated and updated using the chain rule. The goal of the Critic network is to make the Q value close to the target Q value (calculated by the target network), and its loss function is (Q(S,A)-(R + γ×Q'(S',A'))). 2 , where Q' is the output of the target critic network, and A' is the output of the target actor network. The parameters of the target network are copied from the main network via soft updates (update rate τ = 0.001), i.e., target parameters = τ × main network parameters + (1 - τ) × target parameters. During training, the learning rate is set to 0.001, and the number of iterations is determined by the size of the experience pool and the quality of the samples. For example, training is performed for 10,000 epochs, with 64 samples per epoch.

[0114] During training, the model's performance needs to be regularly verified, and the effectiveness of the weight correction model is evaluated using a test set. Training is stopped when the model's average reward on the test set no longer increases and stabilizes. The trained weight correction model can output a weight adjustment gradient for the mapping relationship based on the input state deviation matrix and the current wind field characteristics. This gradient indicates the direction and magnitude of the adjustment for each weight. For example, a gradient vector of (0.05, -0.03, 0.02, 0, -0.01) indicates that the first weight should be increased by 0.05, the second weight should be decreased by 0.03, and so on.

[0115] The output of the weight adjustment gradient is subject to a rationality check. When the absolute value of a weight adjustment gradient exceeds the preset maximum adjustment range (e.g., 0.1), it is truncated to the maximum range to prevent excessive adjustments from causing drastic changes in the mapping relationship and affecting system stability. For example, if the calculated gradient is 0.12, it is truncated to 0.1. At the same time, the system records each output weight adjustment gradient and its corresponding state input and reward value for subsequent model optimization and improvement.

[0116] Adjust the gradient based on the weights, dynamically correct the key weights in the formation density-wind disturbance resistance mapping relationship, test the accuracy of the corrected mapping relationship through the validation set, and output the optimized mapping relationship; The formation density-wind disturbance resistance mapping relationship is a mathematical model that describes the nonlinear relationship between formation density and wind disturbance resistance. It typically takes the form of a weighted summation or neural network. Key weights are parameters that determine the influence of various input features (such as the formation's spatial topology and wind field adaptability) on the output (wind disturbance resistance). For example, a mapping relationship might be: wind disturbance resistance = w1 × formation density + w2 × average wind speed + w3 × turbulence intensity + b, where w1, w2, and w3 are key weights, and b is a bias term. The size of these weights directly influences the output of the mapping relationship.

[0117] The process of dynamically correcting key weights based on the weight adjustment gradient essentially updates the weights in the direction of the ascending (or descending) gradient, so that the mapping more accurately reflects the actual relationship between formation density and wind resistance. Specifically, for each key weight wi, the update formula is wi_new = wi_old + η×gi, where wi_old is the weight before the update, gi is the weight adjustment gradient, and η is the learning rate (step size), which controls the magnitude of the update. The learning rate needs to take into account both convergence speed and stability and is typically set to 0.1. For example, if a weight wi_old = 0.5, the corresponding gradient gi = 0.05, and the learning rate η = 0.1, then wi_new = 0.5 + 0.1×0.05 = 0.505.

[0118] During the correction process, the weight range needs to be restricted to prevent distortion of the mapping relationship caused by excessively large or small weights. For example, the weights w1, w2, and w3 are restricted to the range [0, 1]. When the calculated wi_new exceeds 1, it is set to 1; when wi_new is less than 0, it is set to 0. For the bias term b, a wider range, such as [-1, 1], can be set based on actual conditions. Furthermore, the correction process is gradual, with only small updates based on the latest weight adjustment gradient, rather than a large all-at-once adjustment, to ensure a smooth transition of the mapping relationship.

[0119] To ensure the effectiveness of the correction, it is necessary to conduct a causal analysis of the correction of key weights, that is, to verify the correlation between the weight adjustment and the change in wind resistance intensity. By comparing the output changes of the mapping relationship under the same input before and after the correction, it is analyzed whether the weight adjustment is in the direction of reducing the state deviation. For example, before the correction, the predicted value of wind resistance intensity corresponding to a certain input is 80, the actual value is 85, and the deviation is 5; after the correction, the predicted value is 83, and the deviation is 2, indicating that the weight adjustment is effective. If the deviation increases after the correction, it is necessary to check whether the calculation of the weight adjustment gradient is correct, or whether the learning rate is appropriate, and if necessary, backtrack to the weight value of the previous state.

[0120] The validation set is an independent dataset used to evaluate the accuracy of the corrected mapping relationship. This data is derived from historical flight data from the drone swarm, but is not used in the model training process to ensure the objectivity of the evaluation results. The validation set is typically 20% of the training set and includes samples under different wind conditions (such as light winds, gusts, and turbulence) and different formation structures. For example, it contains 1,000 samples, each of which includes information such as formation density, wind characteristics, and actual wind disturbance resistance.

[0121] The test indicators of mapping relationship accuracy mainly include mean square error (MSE), mean absolute error (MAE) and determination coefficient (R 2 ). The mean square error is the average of the squares of the differences between the predicted value and the actual value, and the formula is MSE=Σ(yi_pred-yi_actual) 2 / n, where yi_pred is the predicted wind resistance strength, yi_actual is the actual wind resistance strength, and n is the number of samples. The smaller the MSE, the higher the accuracy. For example, the MSE of a validation set is 10. The mean absolute error is the average of the absolute values ​​of the difference between the predicted value and the actual value. The formula is MAE=Σ|yi_pred-yi_actual| / n. The smaller the value, the better, for example, MAE=2.5. The coefficient of determination R 2 It is used to measure the degree of fit of the mapping relationship to the data. The formula is R 2 =1-Σ(yi_actual-yi_pred) 2 / Σ(yi_actual-y_mean) 2 , where y_mean is the mean of the actual values, R 2 The closer it is to 1, the better the fitting effect is. For example, R 2 =0.92.

[0122] The test process is to input the formation density and wind field characteristics in the validation set into the modified mapping relationship to obtain the predicted wind disturbance strength, and then calculate the above-mentioned accuracy indicators. If the accuracy indicators meet the preset requirements (such as MSE < 15, MAE < 3, R 2 >0.9), the corrected mapping relationship is determined to be valid, and the optimized mapping relationship is output; if it does not meet the requirements, it is necessary to readjust the learning rate of the weight adjustment gradient or retrain the weight correction model, and repeat the weight correction and accuracy test process until the mapping relationship accuracy meets the requirements.

[0123] During testing, we also need to analyze the accuracy of the mapping relationship under different wind conditions, identify specific scenarios with low accuracy (such as strong gusts), and optimize the weights in these scenarios. For example, if we find that the MSE is large (such as 20) in strong wind conditions with wind speeds greater than 10m / s, we will collect more training samples under strong wind conditions and retrain the weight correction model, focusing on correcting the weights in this scenario to improve the accuracy of the mapping relationship under strong wind conditions.

[0124] The optimized mapping is a precision-verified model that more accurately reflects the relationship between formation density and wind resistance. While the structure of the original mapping is retained, key weights have been dynamically modified. For example, the optimized mapping might be: wind resistance = 0.52 × formation density + 0.31 × average wind speed + 0.18 × turbulence intensity + 0.05. Compared to the original model, these weights have undergone subtle but effective adjustments.

[0125] The optimized mapping relationship needs to undergo actual flight testing to verify its performance in a real low-altitude logistics environment. Its wind disturbance resistance in dynamic wind conditions will be evaluated using actual flight data from drone swarms. For example, if the optimized mapping relationship can reduce the position deviation of the drone swarm by 30% and reduce energy consumption by 15% in gusty wind conditions in actual testing, it means that the mapping relationship has achieved the expected optimization effect.

[0126] Furthermore, the optimized mapping relationships are stored in the system's model library and associated with the corresponding wind field characteristics and timestamps for easy query and analysis. When the system accumulates sufficient new flight data, it periodically re-executes the weight correction and optimization process, allowing the mapping relationships to continuously adapt to new environmental and flight conditions, maintaining long-term high accuracy.

[0127] Based on the optimized mapping relationship, the multi-objective optimization solution process is re-executed to generate updated dynamic formation topology instructions, which are then sent to each UAV to adjust the coordination status, forming a closed loop of monitoring-correction-optimization-execution, and realizing adaptive control of bionic group anti-wind disturbance coordination.

[0128] Re-executing the multi-objective optimization process based on the optimized mapping relationship aims to find the optimal formation structure under current wind conditions based on a more accurate relationship between formation density and wind resistance. Multi-objective optimization objectives include maximizing wind resistance, minimizing total cluster energy consumption, minimizing formation relative position error, and maximizing mission completion efficiency. These objectives may conflict with each other. For example, improving wind resistance may require increasing formation density, which in turn increases energy consumption. Therefore, a balance must be struck between these objectives.

[0129] Each goal needs to be quantified into a specific objective function. For example, the objective function for wind disturbance resistance is f1 = wind disturbance resistance (calculated by the optimized mapping relationship), and the goal is to maximize f1; the objective function for total energy consumption is f2 = Σ energy consumption i (the sum of the energy consumption of each drone), and the goal is to minimize f2; the objective function for relative position error is f3 = Σ√(dx_i 2 +dy_i 2 +dz_i 2 ) / n (average position error), with the goal of minimizing f3. The task completion efficiency objective function, f4, is the inverse of the task completion time, with the goal of maximizing f4. Furthermore, weight coefficients need to be set for each objective function (e.g., w1 = 0.4, w2 = 0.3, w3 = 0.2, w4 = 0.1), transforming the multi-objective optimization problem into a single-objective optimization problem. Specifically, the overall objective function is F = w1 × f1 - w2 × f2 - w3 × f3 + w4 × f4, with the goal of maximizing F.

[0130] Optimization constraints include the physical performance limitations of the drones (such as maximum speed, maximum acceleration, and maximum flight time), communication range limitations (the communication distance between drones must not exceed 50 meters), obstacle avoidance safety distance (the minimum distance between drones must be no less than 2 meters), and wind field adaptability limitations (the formation structure must be able to withstand the current wind field intensity). For example, if a drone has a maximum speed limit of 15m / s, its speed parameter must not exceed this value during the optimization process.

[0131] The multi-objective optimization solution still uses a genetic algorithm, which simulates the biological evolution process and searches for Pareto optimal solutions through selection, crossover, and mutation. The genetic algorithm's parameters need to be adjusted according to the complexity of the problem. For example, a population size of 100 simultaneously considers 100 different formation structure options; a maximum number of iterations of 50 ensures sufficient time for the algorithm to converge; a crossover probability of 0.8, meaning 80% of individuals participate in the crossover operation; and a mutation probability of 0.05, meaning 5% of genes mutate, increasing population diversity.

[0132] During the genetic algorithm, each individual represents a formation structure solution, including information such as the relative positions of each drone, communication topology, and flight parameters. First, the wind resistance of each individual is calculated based on the optimized mapping relationship, and its fitness (the value of the overall objective function F) is calculated in combination with other objective functions. Next, a roulette wheel selection method is used to select individuals with high fitness as parents, with individuals with higher fitness being more likely to be selected. Next, a crossover operation is performed on the selected parents, for example, combining the formation structure information of two parents in a certain ratio to generate offspring individuals. Finally, a mutation operation is performed on the offspring, randomly changing some parameters, such as fine-tuning the relative positions of individual drones. This process is repeated until the maximum number of iterations is reached. At this point, the individual with the highest fitness in the population is the optimal solution in the Pareto optimal solution.

[0133] Generating an updated dynamic formation topology instruction involves converting the optimal formation structure solution obtained through multi-objective optimization into an executable instruction format for the drones. This instruction contains information such as the instruction identifier, timestamp, each drone's device number, target relative position (3D coordinates with the pilot aircraft as the origin), target speed, target attitude angle, communication neighbor list, and execution validity period. For example, in the updated dynamic formation topology instruction, the target relative position of drone UAV001 is (0,0,0) (pilot aircraft), the target relative position of UAV002 is (5,3,0), and the target relative position of UAV003 is (5,-3,0). The communication neighbor list indicates that UAV001 is neighbors with UAV002 and UAV003, and UAV002 is neighbors with UAV003. The execution validity period is 10 seconds.

[0134] Dynamic formation topology commands are issued using reliable communication protocols (such as TCP / IP) to ensure accurate and timely transmission to each drone. During the delivery process, the system encrypts the commands (e.g., using the AES encryption algorithm) to prevent tampering or theft. Furthermore, a combination of broadcast and unicast is used: commands are first broadcast to all drones in the cluster, followed by a unicast confirmation to each drone to ensure successful reception. For drones that do not confirm receipt, the command is resent up to three times. If the command fails, it is marked as an abnormal node and appropriate fault tolerance measures are implemented.

[0135] After receiving the updated dynamic formation topology command, each drone first parses and verifies the command, checking its integrity, legitimacy, and timeliness. For example, it verifies that the command's signature is correct and the timestamp is within its validity period. Once verified, it compares the target parameters in the command (such as target relative position and target speed) with its current state, calculates an adjustment, and uses its control system (such as distributed model predictive control and PID control) to adjust its flight state to achieve the coordinated state required by the command. For example, if UAV002's current relative position is (4.8, 2.9, 0.1) and the target relative position is (5, 3, 0), the calculated position adjustment is (0.2, 0.1, -0.1), gradually approaching the target position by adjusting its speed and attitude.

[0136] During the coordinated state adjustment process, each drone shares its progress and status in real time through local information exchange. For example, it sends its current position and speed information to its communication neighbors every 0.1 second. This ensures coordinated adjustments across the entire cluster and avoids collisions or formation disruptions. For example, while adjusting its position, UAV002 discovers that its distance to UAV003 is approaching a safety threshold. It promptly sends a slowdown prompt to UAV003, which then appropriately reduces its speed and waits for UAV002 to adjust.

[0137] The closed loop of monitoring, correction, optimization, and execution involves the system continuously repeating the following process: Distributed sensors monitor three-dimensional wind field changes and the actual flight status of the drones, calculating deviations from the coordinated anti-wind disturbance flight state; based on these deviations and the current wind field characteristics, reinforcement learning is used to modify the weights of the formation density-anti-wind disturbance intensity mapping relationship; based on this optimized mapping relationship, multi-objective optimization is re-performed to generate updated dynamic formation topology instructions; these instructions are then distributed to each drone for execution, adjusting the coordinated state. The period of this closed loop is dynamically adjusted based on the severity of wind field fluctuations. When the wind field is stable, the period can be set to 5 seconds; when the wind field fluctuates drastically (such as with frequent gusts), the period is shortened to 1 second to rapidly respond to wind field changes.

[0138] During closed-loop operation, the system monitors the operational status of each link in real time. If an anomaly occurs in a link (such as sensor failure, non-convergence of the optimization algorithm, or command failure), the corresponding fault-tolerance mechanism is triggered. For example, if a drone's sensor fails and wind data cannot be collected, the system automatically interpolates wind data from adjacent drones. If the optimization algorithm fails to find a feasible solution after multiple iterations, a pre-set emergency formation plan is activated to ensure the safety of the drone cluster.

[0139] Adaptive control of the bionic swarm's wind-resistance coordination is achieved through the aforementioned closed-loop system. The system automatically adjusts its control strategy (formation structure, flight parameters, etc.) based on environmental (wind) changes and deviations from its own state, enabling the drone swarm to maintain a well-coordinated flight state under various wind conditions, achieving an optimal balance between wind resistance, energy consumption, and mission efficiency. For example, in stable wind conditions, the system adopts a sparse formation to reduce energy consumption; in gusty conditions, it automatically adjusts to a dense formation to improve wind resistance; and in the event of sudden changes in the wind, it quickly responds and re-optimizes the formation, ensuring the swarm can safely and efficiently complete low-altitude logistics missions.

[0140] To verify the effectiveness of adaptive control, the system regularly evaluates key swarm performance indicators, such as the wind disturbance resistance success rate (the ratio of the number of times the formation remained stable in gusty conditions to the total number of gusts), average energy consumption, and task completion on-time rate. These indicators are compared with preset targets. If a metric falls short, the system analyzes the cause and adjusts closed-loop parameters (such as the weight of the reward function in reinforcement learning and the target weights in multi-objective optimization) to continuously improve adaptive control performance. Through long-term operation and optimization, the system continuously accumulates experience, improves its adaptability to complex wind farm environments, and achieves true biomimetic swarm intelligent wind disturbance coordinated control.

[0141] It can be seen that the historical flight data and three-dimensional wind field data of the UAV cluster are collected to generate a bionic formation feature set with wind field labels; the bionic formation feature set is input into the swarm intelligence model integrated with fluid mechanics to generate dynamic formation topology instructions; according to the dynamic formation topology instructions, the relative position and attitude angle of each UAV are adjusted through the distributed collaborative control algorithm to generate an anti-wind disturbance collaborative flight state; the deviation between the three-dimensional wind field changes and the collaborative flight state is continuously monitored, and the weight of the formation density-anti-wind disturbance intensity mapping relationship is dynamically corrected through the reinforcement learning algorithm, and the dynamic formation topology instructions are updated to realize the adaptive control of the bionic group anti-wind disturbance collaboration, thereby realizing high anti-interference collaborative flight of the UAV cluster in the dynamic wind field, reducing the formation energy consumption and improving the obstacle avoidance capability in sudden wind conditions.

[0142] Another embodiment of the present invention provides a bionic swarm intelligence low-altitude logistics UAV cluster anti-wind disturbance collaborative system, see Figure 3 , the system may include: Extraction module 301 is used to collect historical flight data and three-dimensional wind field data of the UAV cluster, extract the spatial topological features and wind field adaptation features of the formation through a bionic feature extraction algorithm, and generate a bionic formation feature set with wind field labels; A construction module 302 is configured to input the bionic formation feature set into a swarm intelligence model integrated with fluid mechanics, construct a formation density-wind disturbance resistance mapping relationship, solve the optimal formation structure using multi-objective optimization, and generate dynamic formation topology instructions; Adjustment module 303 is used to adjust the relative position and attitude angle of each UAV according to the dynamic formation topology instruction through a distributed cooperative control algorithm, utilize the updraft of the wake of the leading UAV to reduce the energy consumption of the trailing UAV, and realize group obstacle avoidance in the event of sudden gusts through local information interaction, thereby generating a wind-resistant cooperative flight state; Correction module 304 is used to continuously monitor the deviation between the three-dimensional wind field changes and the coordinated flight state, dynamically correct the weight of the formation density-anti-wind disturbance intensity mapping relationship through the reinforcement learning algorithm, update the dynamic formation topology instructions, and realize the adaptive control of bionic group anti-wind disturbance coordination.

[0143] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps of any one of the above method embodiments when running.

[0144] An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any one of the above method embodiments.

[0145] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0146] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. A bionic swarm intelligence low-altitude logistics UAV cluster anti-wind disturbance collaborative method, characterized by: The method comprises: Collect historical flight data and three-dimensional wind field data of the UAV swarm, extract the spatial topological features and wind field adaptation features of the formation through a bionic feature extraction algorithm, and generate a bionic formation feature set with wind field labels; The bionic formation feature set is input into a swarm intelligence model integrated with fluid mechanics to construct a formation density-wind disturbance resistance mapping relationship, and a multi-objective optimization is used to solve the optimal formation structure and generate dynamic formation topology instructions; According to the dynamic formation topology instructions, the relative positions and attitude angles of each UAV are adjusted through a distributed cooperative control algorithm, and the updraft of the wake of the leading UAV is used to reduce the energy consumption of the trailing UAV. At the same time, group obstacle avoidance in the event of sudden gusts is achieved through local information interaction, generating a wind-resistant cooperative flight state. Continuously monitor the deviation between the three-dimensional wind field changes and the coordinated flight status, dynamically correct the weight of the formation density-anti-wind disturbance intensity mapping relationship through the reinforcement learning algorithm, update the dynamic formation topology instructions, and realize the adaptive control of bionic group anti-wind disturbance coordination.

2. The method according to claim 1, characterized in that The method collects historical flight data and three-dimensional wind field data of the drone cluster, extracts the spatial topological features and wind field adaptation features of the formation through a bionic feature extraction algorithm, and generates a bionic formation feature set with wind field labels, including: Synchronously collect historical flight data and 3D wind field data from drone swarms, eliminate data collection delays through timestamp alignment, use Kalman filtering to remove measurement noise in flight data, and output pre-processed flight and wind field datasets that are synchronized in time and space. Based on the preprocessed flight data set, the Voronoi diagram partitioning algorithm is used to calculate the relative distance and azimuth between each UAV. Combining the diamond and wedge topological characteristics of the bird formation, the spatial adjacency matrix and density distribution entropy of the formation are extracted, and the spatial topological feature vector of the formation is output. The wind field dataset is associated with the pre-processed flight data. The dynamic time warping algorithm is used to match the UAV's attitude adjustment amplitude, lift coefficient correction value, and energy consumption fluctuation coefficient under different wind field intensities. Principal component analysis is used to extract the core indicators that best reflect the wind field adaptability and output the wind field adaptation feature matrix. The formation space topology feature vector and the wind field adaptation feature matrix are bound in time series, and the corresponding wind field intensity and wind direction labels are marked. The feature fusion network is used to generate a bionic formation feature set with wind field labels.

3. The method according to claim 2, characterized in that The bionic formation feature set is input into a swarm intelligence model integrated with fluid mechanics, a formation density-wind disturbance resistance mapping relationship is constructed, a multi-objective optimization is used to solve the optimal formation structure, and a dynamic formation topology instruction is generated, including: The feature importance of the bionic formation feature set is evaluated to select features that are strongly correlated with wind resistance performance, eliminate redundant features, and output a streamlined feature subset. The simplified feature subset is input into the swarm intelligence model, integrated with the wake interference model in computational fluid dynamics, and the airflow disturbance coefficients under different formation structures are obtained through flow field simulation, and the formation-flow field interaction parameter matrix is ​​output; Based on the formation-flow field interaction parameter matrix, with formation density as input variable and anti-wind disturbance strength as output variable, Gaussian process regression is used to construct the nonlinear mapping relationship between the two and output the formation density-anti-wind disturbance strength mapping function. Set multi-objective optimization goals, use mapping functions as constraints, use genetic algorithms to search for Pareto optimal solutions, and output multiple sets of candidate formation structure parameters; A wind field simulation test is conducted on the candidate formation structure parameters to screen out structures that meet the wind disturbance resistance threshold and have energy consumption lower than the preset value. The structures are converted into relative position instructions and communication topology relationships for each UAV to generate dynamic formation topology instructions.

4. The method according to claim 3, characterized in that The method adjusts the relative positions and attitude angles of the UAVs according to the dynamic formation topology instructions through a distributed cooperative control algorithm, utilizes the updraft of the wake of the leading UAV to reduce the energy consumption of the trailing UAV, and achieves group obstacle avoidance in the event of sudden gusts through local information interaction, thereby generating a wind-resistant cooperative flight state, including: Parse the dynamic formation topology instructions, extract the target relative position and communication neighbor list of each drone, calculate the position deviation based on the current position, and output the target position-deviation comparison table; Based on the target position-deviation comparison table, a distributed model predictive control algorithm is adopted. Each UAV only exchanges position information with its communication neighbors, autonomously calculates the speed adjustment, and realizes position closed-loop regulation through the PID controller, outputting a real-time position correction sequence. Based on the real-time position correction sequence and the wind direction angle of the three-dimensional wind field, the pitch and yaw angles of each UAV are optimized through the aerodynamic model, the attitude angle adjustment amount is calculated, and the attitude collaborative control instructions are output; Based on attitude coordinated control instructions, the system uses onboard meteorological sensors to identify the updraft area of ​​the leading aircraft's wake, guides the following aircraft into this area and adjusts its altitude. It then calculates the lift gain and energy consumption reduction of each following aircraft and outputs the wake utilization parameters. Each UAV exchanges gust monitoring data in real time through short-distance communication, adopts a local obstacle avoidance algorithm based on the potential field method, adjusts the velocity vector to achieve group obstacle avoidance, integrates position correction, attitude collaborative control, and wake utilization parameters to generate a wind-resistant collaborative flight state.

5. The method according to claim 4, characterized in that The method continuously monitors the deviation between the three-dimensional wind field changes and the coordinated flight state, dynamically corrects the weight of the formation density-anti-wind disturbance intensity mapping relationship through a reinforcement learning algorithm, updates the dynamic formation topology instructions, and realizes adaptive control of bionic group anti-wind disturbance coordination, including: The distributed sensors of the UAV cluster collect the current three-dimensional wind field data and the flight status data of each aircraft in real time, calculate the deviation between the actual flight status and the anti-wind disturbance coordinated flight status, and output the state deviation matrix; The state deviation matrix and the current wind field characteristics are used as state inputs to construct a Markov decision process for reinforcement learning. The deep deterministic policy gradient algorithm is used to train the weight correction model and output the weight adjustment gradient of the mapping relationship. Adjust the gradient based on the weights, dynamically correct the key weights in the formation density-wind disturbance resistance mapping relationship, test the accuracy of the corrected mapping relationship through the validation set, and output the optimized mapping relationship; Based on the optimized mapping relationship, the multi-objective optimization solution process is re-executed to generate updated dynamic formation topology instructions, which are then sent to each UAV to adjust the coordination status, forming a closed loop of monitoring-correction-optimization-execution, and realizing adaptive control of bionic group anti-wind disturbance coordination.

6. A bionic swarm intelligence low-altitude logistics UAV cluster anti-wind disturbance collaborative system, characterized by: The system comprises: The extraction module is used to collect historical flight data and three-dimensional wind field data of the UAV cluster, extract the spatial topological features and wind field adaptation features of the formation through the bionic feature extraction algorithm, and generate a bionic formation feature set with wind field labels; A construction module is used to input the bionic formation feature set into a swarm intelligence model integrated with fluid mechanics, construct a formation density-anti-wind disturbance strength mapping relationship, use multi-objective optimization to solve the optimal formation structure, and generate dynamic formation topology instructions; An adjustment module is used to adjust the relative positions and attitude angles of each UAV through a distributed cooperative control algorithm according to the dynamic formation topology instructions, utilize the updraft of the wake of the leading UAV to reduce the energy consumption of the trailing UAV, and achieve group obstacle avoidance in the event of sudden gusts through local information interaction, thereby generating a wind-resistant cooperative flight state; The correction module is used to continuously monitor the deviation between the three-dimensional wind field changes and the coordinated flight status, dynamically correct the weight of the formation density-anti-wind disturbance intensity mapping relationship through the reinforcement learning algorithm, update the dynamic formation topology instructions, and realize the adaptive control of bionic group anti-wind disturbance coordination.

7. The system according to claim 6, characterized in that The extraction module is specifically used to: Synchronously collect historical flight data and 3D wind field data from drone swarms, eliminate data collection delays through timestamp alignment, use Kalman filtering to remove measurement noise in flight data, and output pre-processed flight and wind field datasets that are synchronized in time and space. Based on the preprocessed flight data set, the Voronoi diagram partitioning algorithm is used to calculate the relative distance and azimuth between each UAV. Combining the diamond and wedge topological characteristics of the bird formation, the spatial adjacency matrix and density distribution entropy of the formation are extracted, and the spatial topological feature vector of the formation is output. The wind field dataset is associated with the pre-processed flight data. The dynamic time warping algorithm is used to match the UAV's attitude adjustment amplitude, lift coefficient correction value, and energy consumption fluctuation coefficient under different wind field intensities. Principal component analysis is used to extract the core indicators that best reflect the wind field adaptability and output the wind field adaptation feature matrix. The formation space topology feature vector and the wind field adaptation feature matrix are bound in time series, and the corresponding wind field intensity and wind direction labels are marked. The feature fusion network is used to generate a bionic formation feature set with wind field labels.

8. The system according to claim 7, characterized in that The building blocks are specifically used for: The feature importance of the bionic formation feature set is evaluated to select features that are strongly correlated with wind resistance performance, eliminate redundant features, and output a streamlined feature subset. The simplified feature subset is input into the swarm intelligence model, integrated with the wake interference model in computational fluid dynamics, and the airflow disturbance coefficients under different formation structures are obtained through flow field simulation, and the formation-flow field interaction parameter matrix is ​​output; Based on the formation-flow field interaction parameter matrix, with formation density as input variable and anti-wind disturbance strength as output variable, Gaussian process regression is used to construct the nonlinear mapping relationship between the two and output the formation density-anti-wind disturbance strength mapping function. Set multi-objective optimization goals, use mapping functions as constraints, use genetic algorithms to search for Pareto optimal solutions, and output multiple sets of candidate formation structure parameters; A wind field simulation test is conducted on the candidate formation structure parameters to screen out structures that meet the wind disturbance resistance threshold and have energy consumption lower than the preset value. The structures are converted into relative position instructions and communication topology relationships for each UAV to generate dynamic formation topology instructions.

9. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 5 when executed.

10. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 5.

Images