Low-altitude aircraft cluster dynamic scheduling method and system based on multi-source data fusion

Abstract

The application provides a kind of low-altitude aircraft cluster dynamic scheduling method and system based on multi-source data fusion, it is related to information technology field, belongs to aircraft cooperative control technical field.In the method, each aircraft constructs local space-time resource tension field slice based on motion state extrapolation and environment perception information in initialization round, generates candidate predicted trajectory and broadcasts;Second round uses neighbor trajectory to replace extrapolation to update slice, obtains main trajectory and alternative trajectory library after Nash equilibrium is converged to by game iteration;Flight monitoring deviation between actual perception field and prediction field, when consistency deviation, adjust prediction model parameters, when sudden deviation, switch alternative trajectory;Periodically trigger group consensus process, modify game payoff function weight and / or tension field evolution equation parameters and carry out distributed diffusion when detecting common deviation mode.The application realizes distributed autonomous cooperative scheduling of aircraft cluster, significantly improves prediction accuracy, response speed and system robustness.

Description

Technical Field

[0001] This invention relates to the field of aircraft cooperative control, and more specifically, to a method and system for dynamic scheduling of low-altitude aircraft clusters based on multi-source data fusion. Background Technology

[0002] Low-altitude aircraft swarms are increasingly widely used in logistics, emergency rescue, and urban air traffic control. However, with the expansion of swarm size and the increasing complexity of airspace environments, achieving efficient collaborative scheduling of multiple aircraft within limited low-altitude airspace has become a key bottleneck restricting the large-scale application of this technology. Existing scheduling methods rely on centralized decision-making in some schemes, exhibiting excessive dependence on communication bandwidth and central nodes, resulting in insufficient system robustness. Other distributed schemes are largely based on static topology or single-dimensional environmental perception information, making it difficult to adapt to dynamically changing environments and the real-time interaction requirements between aircraft. Furthermore, existing air traffic management systems are primarily designed for mid-to-high-altitude manned aircraft and are not directly applicable to the operational scenarios of large-scale low-altitude unmanned aerial vehicle (UAV) swarms. Therefore, there is an urgent need for a technical solution that can integrate multi-source perception information within a distributed architecture to achieve autonomous collaborative scheduling of aircraft swarms. Summary of the Invention

[0003] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, the present invention provides a method for dynamic scheduling of low-altitude aircraft clusters based on multi-source data fusion, the method comprising: During the initialization round, each aircraft acquires its own local state information, local environmental perception information, and neighboring local state information broadcast by neighboring aircraft within its communication range within the low-altitude aircraft cluster. Based on its own local state information and the neighboring local state information, it extrapolates its motion state and, in conjunction with the local environmental perception information, constructs an initial local spatiotemporal resource tension field slice locally. The local spatiotemporal resource tension field slice is used to characterize the resource occupancy tension of the space surrounding the aircraft at different future times. Each aircraft generates at least two candidate predicted trajectories covering a future preset time domain based on its corresponding local spatiotemporal resource tension field slice, and assigns a prior confidence level to each candidate predicted trajectory. Each aircraft broadcasts its generated candidate predicted trajectory within its communication range and receives neighbor candidate predicted trajectories broadcast by neighboring aircraft. From the second round onwards, when each aircraft constructs the local spatiotemporal resource tension field slice, it uses the neighbor candidate predicted trajectory received in the previous round to replace the motion state extrapolation, updates the density tension field components, and forms an updated local spatiotemporal resource tension field slice. Each aircraft performs iterative trajectory adjustment based on its own candidate predicted trajectory, the neighbor's candidate predicted trajectory, and a preset distributed game payoff function, until the trajectory adjustment result of each aircraft converges to a Nash equilibrium state, thereby obtaining the main predicted trajectory and the candidate predicted trajectory library corresponding to each aircraft. Each aircraft flies using the main predicted trajectory as its current execution trajectory, and continuously monitors the prediction deviation between the actual perceived field and the predicted field during flight; the actual perceived field is determined by the local environmental perception information, and the predicted field is determined by the local spatiotemporal resource tension field slice; Each aircraft executes a feedback iterative update process based on the type of the prediction deviation: when the prediction deviation is a consistency deviation, the prediction model parameters inside the aircraft are adjusted; when the prediction deviation is a sudden deviation, the target candidate prediction trajectory is switched from the candidate prediction trajectory library as the new current execution trajectory. The group consensus process is periodically triggered. Prediction deviation patterns of multiple aircraft are collected within the communication range and clustered. When a common deviation pattern is detected, the weight parameters of the distributed game payoff function and / or the evolution equation parameters of the local spatiotemporal resource tension field slice are corrected, and the corrected parameters are distributed and diffused in the cluster.

[0004] Furthermore, the present invention also provides a dynamic scheduling system for low-altitude aircraft clusters based on multi-source data fusion, which is used to store and execute the above-mentioned methods.

[0005] Based on the above aspects, the method and system provided by the present invention have the following technical effects compared with the prior art.

[0006] First, by constructing local spatiotemporal resource tension field slices on each aircraft and switching from open-loop motion extrapolation to closed-loop trajectory information exchange with communication rounds as boundaries, the root mean square error of density tension field prediction is reduced from approximately 4.2 meters under uniform extrapolation conditions to approximately 1.8 meters, an improvement of over 57%. The aircraft can obtain the planning intentions of neighboring aircraft without relying on a ground-based central station, fundamentally overcoming the high dependence of centralized decision-making architectures on communication links and central nodes, and significantly improving the system's robustness under conditions of communication degradation or partial node failure.

[0007] Second, by employing spatiotemporal dual-dimensional statistical testing, prediction deviations are automatically classified into consistency deviations and sudden deviations. Two different feedback mechanisms are triggered: adjustment of prediction model parameters and switching of alternative trajectories. This achieves refined, differentiated control over deviation processing. This mechanism reduces the false alarm rate from approximately 8% in linear schemes to approximately 2%, while maintaining a true precursor detection rate above 96%. Trajectory switching for sudden deviations is completed within 100 milliseconds, significantly shortening the aircraft's response time to sudden obstacles or weather changes.

[0008] Third, by detecting the acceleration of changes in conflict intensity using second-order differentials, rather than focusing solely on the conflict intensity itself, early prediction of congestion precursors is achieved. Combined with monitoring the rate of decline in algebraic connectivity and dynamic optimization of the risk balance cost function, the weight coefficients change continuously and smoothly when the aircraft switches between two typical mission phases: dense formation and wide-area search. The weight change between adjacent cycles is controlled within 0.05, effectively avoiding control oscillations and trajectory jumps caused by phase switching.

[0009] Fourth, through a consensus confidence variable-driven disagreement arbitration mechanism and a hierarchical consensus process, the cluster can achieve orderly coordination when local perceptions and global consensus are inconsistent, overcoming the problem of each aircraft acting independently and making it difficult to form consistent decisions in existing solutions. The reflexive reconstruction operation enables the local collaborative topology graph to evolve autonomously with environmental changes, further enhancing the cluster's adaptability in complex dynamic environments.

[0010] In summary, this invention achieves deep fusion of multi-source sensing information and autonomous dynamic scheduling of aircraft clusters under a distributed architecture, demonstrating good engineering applicability and promising prospects for widespread application. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the process of local collaborative topology graph construction and early warning reconstruction in this invention; Figure 3 This is a schematic diagram of the system provided in this embodiment; Figure 4 A schematic diagram comparing the accuracy of density tension field prediction between the method provided by this invention and existing technologies; Figure 5 A schematic diagram comparing the key performance indicators of the method provided by this invention with those of existing technology systems. Detailed Implementation

[0012] like Figures 1-3 As shown.

[0013] This technical solution provides a method and system for dynamic scheduling of low-altitude aircraft clusters based on multi-source data fusion. The technical details are explained in detail below using a specific application scenario.

[0014] This embodiment uses an urban low-altitude logistics delivery scenario as an example. Assume that within a new urban area, 10 quadcopter logistics drones form a swarm, performing express delivery tasks from a central distribution station to multiple residential drop-off points. The airspace in this area is 50 to 150 meters above ground level, including three no-fly zones around high-rise buildings and two temporarily designated drone control zones. The weather conditions on that day are southeasterly winds of force 3 to 4, with gusts reaching force 5. Each drone's onboard computing platform is an ARM Cortex-A72 quad-core processor paired with a neural network acceleration unit, with a main frequency of 1.5 GHz. It is equipped with an onboard binocular vision sensor, millimeter-wave radar, inertial measurement unit, global navigation satellite system receiver, and air-to-air communication module. The communication module operates in the 2.4 GHz band, with a communication coverage radius of 500 meters and a communication cycle of 100 milliseconds. Each drone has been loaded with a high-precision digital map of the area and a mission waypoint sequence before takeoff.

[0015] The following detailed description, in conjunction with the flowchart shown in the attached figures, provides a specific implementation of the dynamic scheduling method for low-altitude aircraft clusters based on multi-source data fusion.

[0016] Round 1: Initialization and local spatiotemporal resource tension field slice construction; In the first round, or initialization round, each aircraft in the cluster first acquires its own local state information and local environmental awareness information.

[0017] Taking the UAV-01 as an example, its local status information includes at least the following data fields: its own three-dimensional spatial position coordinates in the northeast-northeast coordinate system, in meters; its own three-dimensional velocity vector, in meters per second; its own attitude angles, namely roll, pitch, and yaw, in degrees; and the current remaining battery power, expressed as a percentage. Local environmental perception information includes at least the following data fields: the coordinate sequence of the no-fly zone boundary detected by the fusion of airborne millimeter-wave radar and binocular vision sensors; the coordinates of the polygon vertices of the temporary control area; and wind vector field data obtained by airborne meteorological sensors and receiving ground meteorological broadcasts, i.e., wind speed and direction at various locations in three-dimensional space. Simultaneously, UAV-01 receives neighboring local status information broadcast by all neighboring aircraft within its communication range via its air-to-air communication module. The local status information of each neighboring aircraft includes at least its three-dimensional spatial position coordinates and three-dimensional velocity vector.

[0018] After completing the above information collection, UAV-01 constructs a local spatiotemporal resource tension field slice centered on itself on the local airborne computing platform.

[0019] This local spatiotemporal resource tension field slice is a scalar field function F(x,y,z,t) defined in four-dimensional spacetime, where x, y, and z are spatial coordinates, and t is a time variable. The spatial extent of this slice is a spherical region with a radius of 500 meters centered on the current location of UAV-01, and the time extent is the next 30 seconds from the current moment, with a time discretization step of 0.5 seconds. A higher value of the field function F indicates a higher degree of occupancy of that spatiotemporal location by aircraft, meaning the location is less usable or the cost of traversing it is greater. This field function is composed of four components superimposed point-by-point according to preset weights, and its formula is: F(p,t) = w_d · D_norm(p,t) + w_c · C_norm(p) + w_w · W_norm(p,t) +w_e · E_norm(p); Where p=(x,y,z) represents the spatial coordinates, and t represents time. D_norm is the normalized density tension field component, C_norm is the normalized constraint tension field component, W_norm is the normalized wind field perturbation tensor component, and E_norm is the normalized energy potential field component. The initial values ​​of the four weighting coefficients are set to w_d=0.35, w_c=0.30, w_w=0.20, and w_e=0.15, respectively, and the sum of all weighting coefficients is 1. Experiments show that when the weighting coefficients of the density tension field are in the range of 0.30 to 0.45, the swarm has the lowest conflict rate in medium-density scenarios; when the coefficient is below 0.25, the aircraft's perception of neighboring aircraft is insufficient, and close-range conflicts are prone to occur; when the coefficient is above 0.50, excessive avoidance by the aircraft leads to increased path detours, and the average mission completion time is extended by about 12%.

[0020] The density tension field component D(p,t) is calculated as follows. In the first round, UAV-01 extrapolates its own spatial position coordinates and velocity vector, as well as the received spatial position coordinates and velocity vectors of each neighboring aircraft, to perform uniform linear motion extrapolation for itself and each neighboring aircraft. The formula for uniform linear motion extrapolation is: p_i(t_k) = p_i(t_0) + v_i(t_0) · (t_k - t_0); Where p_i(t_k) represents the predicted spatial position of the i-th spacecraft at a future time t_k, p_i(t_0) is the position of the spacecraft at the current time t_0, v_i(t_0) is the velocity vector of the spacecraft at the current time, and t_k - t_0 is the prediction time increment. For each predicted spatial position p_i(t_k), its density contribution to any spatial position p within the slice at time t is calculated using a Gaussian kernel function: g(p, p_i(t_k)) = (1 / (σ · √(2π))) · exp(-||p - p_i(t_k)||² / (2σ²)); Where ||p - p_i(t_k)|| is the Euclidean distance between two spatial locations, and σ is the kernel bandwidth parameter, which is taken as 20 meters in this embodiment. This Gaussian kernel function has a smooth cutoff characteristic; when the distance exceeds 3 times σ, i.e., 60 meters, the density contribution value decays to approximately 1% of the peak value. The density contribution values ​​from all aircraft at the same spatial location at the same time are accumulated to obtain the predicted cluster density value for that spatiotemporal point. The predicted cluster density values ​​of all spatial locations at all future times are arranged in spatial and temporal order to form the density tension field D(p,t). Subsequently, D(p,t) is normalized to the interval [0,1] to obtain D_norm(p,t).

[0021] The calculation method for the constraint tension field component C(p) is as follows. UAV-01 calculates the hard constraint barrier value at each spatial location within the sliced ​​space based on the coordinates of the no-fly zone boundary and the temporary control area from the local environmental perception information. For any spatial location p, the formula for calculating its constraint barrier value c(p) is: c(p) = min_j (1 / (d_j(p) + ε)); Where d_j(p) is the shortest distance from spatial location p to the boundary of the j-th no-fly zone or controlled area, in meters, and ε is a very small positive number to prevent division by zero; in this embodiment, it is taken as 0.1 meters. When p is located inside the no-fly zone or controlled area, d_j(p) = 0, and c(p) reaches its maximum value. Normalizing c(p) to the interval [0,1] yields C_norm(p). This constraint tension field component does not change with time and remains unchanged in the time dimension of the slice.

[0022] The calculation method for the wind field disturbance tensor components is as follows. UAV-01 calculates the wind field disturbance values ​​at different times for each spatial location within the sliced ​​spatial range based on the wind vector field data from the local environmental perception information. It should be noted that the wind field disturbance tensor in this embodiment is actually a three-dimensional wind velocity vector field, meaning that each spatial location has three directional components W_x, W_y, and W_z at each time point, in meters per second. To facilitate superposition with other field components, this three-dimensional vector field is projected onto the aircraft's motion direction to obtain a scalar disturbance value w(p,t), calculated using the following formula: w(p,t) = W(p,t) · v_hat(p,t); Where W(p,t) is the wind speed vector, and v_hat(p,t) is the unit velocity vector of the aircraft at spatial position p along the planned path. The dot product of the two represents the positive or negative perturbation of the wind on the aircraft along the path. Normalizing w(p,t) to the interval [0,1] yields W_norm(p,t).

[0023] The energy potential component E(p) is calculated as follows. UAV-01 calculates the energy margin value for each spatial location based on the current remaining charge and the estimated energy consumption required to reach each spatial location within the slice. For spatial location p, the formula for calculating the energy margin value e(p) is: e(p) = (E_remaining - E_consume(p)) / E_remaining; Where E_remaining is the current remaining battery power, and E_consume(p) is the estimated energy consumption required to fly from the current position to point p, both in watt-hours. When E_consume(p) exceeds E_remaining, e(p) takes a negative value. E(p) is normalized to obtain E_norm(p). The normalization method is to map the energy consumption in the range [0, E_remaining] to [1, 0], and the excess part is mapped to the negative value range.

[0024] After completing the calculation of the four components, UAV-01 superimposes the four normalized components point by point according to the aforementioned weighted summation formula to generate a slice of the local spatiotemporal resource tension field centered on itself. This slice is stored in the dynamic random access memory of the onboard computing platform in the form of a four-dimensional array. The array dimension is the number of spatial grids multiplied by the number of time steps. The spatial grid resolution is 5 meters × 5 meters × 5 meters, and the number of time steps is 61.

[0025] Candidate predicted trajectory generation and broadcasting Based on the constructed local spatiotemporal resource tension field slices, UAV-01 generates five candidate predicted trajectories covering the next 30 seconds of the time domain. Each candidate predicted trajectory is parameterized using a B-spline curve, defined by a set of eight control points, with a 4-second time interval between adjacent control points. The specific method for trajectory generation is as follows: using the current position and velocity as the initial constraint and the target waypoint position as the terminal constraint, multiple candidate paths are generated in spatiotemporal regions with low tension field values ​​through random sampling and smooth connection. Each candidate predicted trajectory contains the three-dimensional spatial position coordinates and velocity vectors at 0.5-second intervals within the next 30 seconds.

[0026] Each candidate predicted trajectory is assigned a prior confidence score. The prior confidence score is a scalar value ranging from 0 to 1, calculated by normalizing the trajectory's total cost value. The total cost value of a trajectory consists of a weighted sum of three factors: the cumulative sum of tension field values ​​at each time step along the trajectory, the total trajectory length, and the deviation from the target waypoint's terminal position. A lower total cost value corresponds to a higher prior confidence score. In this embodiment, the trajectory with the highest prior confidence score is marked as the initial primary predicted trajectory, and the remaining four are stored in the candidate predicted trajectory library.

[0027] The UAV-01 broadcasts its five generated candidate predicted trajectories within its 500-meter communication range via its air-to-air communication module. The broadcast data packet includes the trajectory number, the sequence of B-spline control points for the trajectory, and the corresponding prior confidence level. Simultaneously, the UAV-01 receives neighboring candidate predicted trajectories broadcast by neighboring aircraft within its communication range and stores this information in its local cache.

[0028] Subsequent rounds: Iterative trajectory adjustment and Nash equilibrium Starting from the second round, when constructing local spatiotemporal resource tension field slices, UAV-01 no longer uses the uniform linear motion extrapolation method to predict the future positions of neighboring aircraft. Instead, it directly uses the candidate predicted trajectories of neighbors received in the previous round as the source of predicted spatial position coordinates for each neighboring aircraft in the future time domain. Specifically, when calculating the density tension field component D(p,t), UAV-01 uses its own candidate predicted trajectory and the candidate predicted trajectories of each neighboring aircraft as the predicted spatial position coordinate sequence for each aircraft, replacing the uniform motion extrapolation results from the first round, and calculates the predicted cluster density value for each spatiotemporal point accordingly. This substitution improves the prediction accuracy of the density tension field from open-loop extrapolation to prediction based on closed-loop information exchange, enabling it to more accurately reflect the actual planning intentions of each aircraft in the cluster.

[0029] In each round, UAV-01 iteratively adjusts its trajectory based on its own candidate predicted trajectory, the received candidate predicted trajectories from its neighbors, and a preset distributed game payoff function. The distributed game payoff function is a local cost function related to UAV-01's trajectory selection, and its mathematical form is: J_i(τ_i, τ_{-i}) = α · C_conflict(τ_i, τ_{-i}) + β · C_energy(τ_i)+ γ · C_time(τ_i); Where τ_i represents UAV-01's own candidate predicted trajectory, and τ_{-i} represents the set of candidate predicted trajectories for all neighboring aircraft. C_conflict is the conflict cost term, calculated based on the spatiotemporal overlap between its own trajectory and those of its neighbors; C_energy is the energy consumption cost term, calculated based on the total trajectory length and wind disturbance; and C_time is the time cost term, calculated based on the deviation between the actual arrival time and the planned arrival time at the target waypoint. The initial values ​​of the three weighting coefficients α, β, and γ are set to 0.50, 0.30, and 0.20, respectively. In each iteration, UAV-01 selects the trajectory that minimizes J_i as the optimal response for that round, while keeping the neighboring trajectories unchanged. Each aircraft simultaneously performs the above optimal response calculation and broadcasts the updated trajectory. This process is repeated until the trajectory adjustment amplitude of all aircraft is less than the position change threshold of 0.5 meters in three consecutive iterations, at which point the ε-Nash equilibrium state is determined. The trajectory at this point is used as the main predicted trajectory, while a set of suboptimal trajectories generated during the iteration process is retained to form a candidate predicted trajectory library.

[0030] Trajectory Execution and Prediction Deviation Handling The UAV-01 flies using the converged master predicted trajectory as its current execution trajectory. The flight control module generates flight control commands every 0.5 seconds based on the master predicted trajectory. The control commands include the desired thrust, desired attitude angle, and desired angular velocity, which drive the four rotor motors through the onboard flight controller.

[0031] During flight, UAV-01 continuously monitors the prediction deviation between the actual perceived field and the predicted field. The actual perceived field is determined by the local environmental perception information at the current moment, i.e., the instantaneous environmental field constructed based on the latest sensor data; the predicted field is determined by a slice of the local spatiotemporal resource tension field at the current moment. The prediction deviation is defined as the deviation field Δ(p,t) = F_actual(p,t) - F_predicted(p,t), where F_actual is the actual perceived field function value and F_predicted is the predicted field function value. UAV-01 maintains a sliding time window of 3 seconds, recording the deviation value of each spatial location at each moment within the window.

[0032] UAV-01 performs a spatiotemporal two-dimensional statistical test on prediction deviations to automatically distinguish the type of deviation. When the duration of the deviation within the sliding window exceeds 2 seconds, and its spatial influence range extends beyond the UAV-01's own 500-meter radius neighborhood and covers the corresponding perception areas of at least 3 neighboring aircraft, the prediction deviation is determined to be a consistency deviation; otherwise, it is determined to be a sudden deviation.

[0033] After identifying a consistency deviation, UAV-01 further collects confirmation signals from neighboring aircraft within its communication range affected by the deviation, indicating the same deviation pattern. When the number of collected confirmation signals reaches more than 50% of the number of valid neighbors within the communication range, UAV-01 confirms the execution of prediction model parameter adjustment operations. Specifically, prediction model parameter adjustment refers to correcting the bandwidth parameter σ of the Gaussian kernel in the density tension field calculation, or adjusting the time extrapolation coefficients of the wind field disturbance tensor components, to make the prediction model more closely reflect the current actual environmental conditions.

[0034] Upon determining a sudden deviation, UAV-01 immediately switches from the candidate predicted trajectory library to the target candidate predicted trajectory. The selection criterion for the target candidate predicted trajectory is: the trajectory with the smallest deviation from the current actual sensed field is selected as the new current execution trajectory. Simultaneously, UAV-01 broadcasts a sudden deviation warning to neighboring aircraft within its communication range. This warning includes three fields: the estimated location of the sudden deviation, the radius of influence, and the estimated duration.

[0035] Periodic group consensus and parameter adjustment Each aircraft in the cluster continuously monitors the topological critical indicators of its local cooperative topology graph, including the cluster local mean of conflict intensity and the cluster local variance of algebraic connectivity. UAV-01 checks these indicators every 500 milliseconds. When the cluster local mean of conflict intensity exceeds a preset mean threshold of 0.6 or the cluster local variance of algebraic connectivity exceeds a preset variance threshold of 0.15, UAV-01 initiates a group consensus request within its communication range.

[0036] The group consensus process employs a hierarchical consensus mechanism. UAV-01 first forms local clusters with its neighboring UAVs within its one-hop communication range. Within each local cluster, UAVs exchange their prediction bias patterns and perform local K-means clustering, with the number of clusters set to 2 to 4, forming local pre-consensus results. If the acceptance rate of the local pre-consensus results within the local cluster reaches 60% or more, the local cluster elects the UAV with the most direct communication links with its neighbors as the backbone UAV. The backbone UAV then disseminates the local pre-consensus results to adjacent clusters. The backbone UAVs in each adjacent cluster perform secondary alignment and clustering on the received multiple local pre-consensus results to generate a global consensus result, which is then distributed and disseminated to the entire cluster via broadcast.

[0037] After the global consensus result is confirmed, in addition to correcting the weight coefficients of the conflict cost, energy cost, and time cost terms in the distributed game payoff function, a topology reconstruction mask is directly output. This topology reconstruction mask specifies the set of edges that each aircraft should remove and the set of potential cooperative edges that should be added in subsequent reflexive reconstruction operations. Each aircraft synchronously performs reflexive reconstruction operations according to the received topology reconstruction mask, that is, removing edges corresponding to unreliable directions marked in the common deviation pattern, and adding new edges with specific neighboring aircraft based on the potential cooperative dependencies revealed by the common deviation pattern. The reconstructed local cooperative topology graph is used for trajectory adjustment games in the next cycle.

[0038] Construction and updating of density tension field Among the four components of a local spatiotemporal resource tension field slice, the density tension field component is the core component of the slice, and its construction accuracy directly affects the quality of subsequent trajectory planning. The specific construction process of the density tension field is explained in detail below.

[0039] Taking UAV-01 as an example, during the initialization round, UAV-01 acquires its spatial position coordinates p_self=(x_self, y_self, z_self) and velocity vector v_self=(vx_self, vy_self, vz_self) in the northeast-central coordinate system. Simultaneously, it receives local state information broadcast by all neighboring aircraft within its communication range via its air-to-air communication module. Each neighbor's state information includes at least its spatial position coordinates p_j and velocity vector v_j. UAV-01 then extrapolates its own and each neighbor's uniform linear motion using the following formula: p_i(t_k) = p_i(t_0) + v_i(t_0) · (t_k - t_0) The subscript i iterates through UAV-01 itself and all neighboring aircraft, t_0 is the current time, and t_k is the k-th discrete time in the future, with k ranging from 1 to 61, corresponding to a time point every 0.5 seconds within the next 30 seconds. The uniform linear motion extrapolation assumes that each aircraft maintains its current velocity vector unchanged in the short future time domain. This assumption is the only feasible prediction method when no planning intention information from neighboring aircraft is obtained during the initialization round.

[0040] After obtaining the predicted spatial positions of each aircraft at future time points, UAV-01 calculates the density contribution of each predicted spatial position to every grid node within the sliced ​​spatial area at each time point. The sliced ​​spatial area is a spherical region with a radius of 500 meters, centered on the current position of UAV-01, covering both the horizontal and vertical directions. The spatial grid is divided into 5m × 5m × 5m cubic grid cells, with 200 grid cells in the horizontal direction and 200 grid cells in the vertical direction, for a total of approximately 8 million spatial grid nodes. Considering such a large-scale computational load, UAV-01's onboard neural network acceleration unit performs hardware acceleration of this calculation process, mapping the calculation task of density contribution values ​​in parallel across thousands of computing cores of the acceleration unit.

[0041] The density contribution of each predicted spatial location to the target spatial location is calculated using a Gaussian kernel function, and its mathematical form is: g(p, p_i(t_k)) = (1 / (σ · √(2π))) · exp(-||p - p_i(t_k)||² / (2σ²)); Where p represents the target spatial coordinates, p_i(t_k) is the predicted spatial position of the i-th spacecraft at time t_k, and ||p - p_i(t_k)|| is the Euclidean distance between them. σ is the bandwidth parameter of the Gaussian kernel, which is 20 meters in this embodiment. This value is determined based on the average spacing and safety interval requirements of the spacecraft in the cluster. When the predicted spacing between two spacecraft is less than 10 meters, the density contribution value reaches a higher region of the peak value, indicating a high collision risk at this spatiotemporal location. When the predicted spacing is greater than 60 meters, the density contribution value decays to about 1% of the peak value, and its impact on the tension field is negligible. The density contribution values ​​from all spacecraft at the same spatial grid node at the same time are accumulated to obtain the cluster density prediction value of that spatiotemporal point. The cluster density prediction values ​​of all spatial grid nodes at all future times are arranged in spatial and temporal order to form the density tension field D(p,t). Then, D(p,t) is linearly mapped to the interval [0,1] to obtain the normalized density tension field D_norm(p,t). The mapping method is to take the maximum and minimum density values ​​D_max and D_min of all spatiotemporal points within the slice range, and calculate them according to the following formula: D_norm(p,t) = (D(p,t) - D_min) / (D_max - D_min); Starting from the second round, UAV-01 no longer uses uniform linear motion extrapolation when calculating the density tension field components. Instead, it directly utilizes the candidate predicted trajectories generated by itself in the previous round and adjusted through game theory, as well as the candidate predicted trajectories received from neighboring aircraft. The predicted spatial coordinate sequences from these trajectories serve as the source of each aircraft's position in the future time domain, replacing the uniform extrapolation results from the initialization round. This substitution upgrades the prediction basis of the density tension field from open-loop motion extrapolation to closed-loop planning information exchange. The actual motion intentions of each aircraft are known to its neighbors through trajectory broadcasting, thus significantly improving the accuracy of density field prediction. Experiments show that after replacing uniform extrapolation with candidate predicted trajectories, the root mean square error of density tension field prediction decreases from approximately 4.2 meters under uniform extrapolation conditions to approximately 1.8 meters, an improvement of over 57%.

[0042] Construction of constrained tension field Constraint tension field components are used to characterize the obstructive effects of various hard-constraint obstacles and restricted areas on aircraft passage in the airspace. UAV-01 constructs constraint tension field components based on the coordinates of no-fly zone boundaries, temporary control areas, and severe weather areas from local environmental perception information.

[0043] The coordinates of the no-fly zone boundaries are derived from a pre-installed high-precision digital map. In the urban logistics scenario of this embodiment, the three high-rise building no-fly zones are modeled as vertically extending prisms with polygonal horizontal sections and a vertical range from the ground to an altitude of 200 meters. The coordinates of the temporary control areas are derived from dynamic information transmitted in real time by the ground control center via data link. In this embodiment, the two temporary control areas are modeled as spheres or ellipsoids. The coordinates of the severe weather area are derived from real-time detection of precipitation areas ahead by airborne millimeter-wave radar. In this embodiment, a shower area with a diameter of approximately 300 meters is modeled as a horizontal disk-shaped area.

[0044] For any spatial location p within the slice space, the constraint barrier value c(p) is calculated using the nearest distance reciprocal method, as shown in the formula: c(p) = min_j (1 / (d_j(p) + ε)); Where d_j(p) is the shortest Euclidean distance from spatial location p to the boundary of the j-th no-fly zone, controlled area, or severe weather zone, in meters, and ε is the zero constant for prevention, taken as 0.1 meters. When p is located inside any restricted area, d_j(p) = 0, and c(p) reaches its maximum value, indicating that the location is completely impassable. When p is more than 100 meters from the boundaries of all restricted areas, c(p) approaches 0, indicating that the location is not affected by the constraint field. The constraint tension field components do not change with time and remain constant throughout the entire future time domain of the slice, reflecting the static characteristics of hard-constrained obstacles. Linearly mapping c(p) to the interval [0,1] yields C_norm(p).

[0045] Construction of wind field perturbation tensor field and energy potential field The construction of the wind field perturbation tensor field relies on wind vector field data from local environmental perception information. The UAV-01's onboard meteorological sensors measure wind speed and direction at its current location in real time, while simultaneously receiving regional wind field forecast data from ground-based meteorological broadcasts. To improve the spatial resolution of the wind field data, the UAV-01 exchanges its measured local wind field information with other neighboring aircraft within its communication range, and forms the wind field distribution at various spatial locations within the sliced ​​spatial range through data fusion.

[0046] In this embodiment, the wind field disturbance tensor is actually represented by a three-dimensional wind velocity vector field, meaning that each spatial location has three orthogonal wind velocity components W_x, W_y, and W_z at each time point, in meters per second. To convert the three-dimensional vector field into a form that can be superimposed with the other scalar field components, the wind velocity vector is projected onto the tangent direction of the planned path of the aircraft at that location, resulting in a scalar disturbance value w(p,t): w(p,t) = W(p,t) · v_hat(p,t); Where W(p,t) is the wind speed vector, and v_hat(p,t) is the unit tangent vector along the planned path when the aircraft passes through spatial position p. A positive dot product indicates a tailwind, which helps reduce flight energy consumption; a negative dot product indicates a headwind, which increases flight energy consumption and causes the aircraft to deviate from the predetermined path. Linearly mapping w(p,t) of all spatiotemporal points within the slice range to the interval [0,1] yields W_norm(p,t).

[0047] The energy potential field components are used to reflect the feasibility of the aircraft reaching various spatial locations under the current remaining power constraint. The energy potential field of UAV-01 is calculated as follows. For any spatial location p within the slice range, the estimated energy consumption E_consume(p) of the aircraft flying from the current location to point p is calculated by the following formula: E_consume(p) = ∫_{path} (P_hover + k_w · ||W(s)||²) · (ds / v_ground(s)); The integral is performed along the shortest path from the current position to point p. P_hover is the hovering power, taken as 120 watts, k_w is the drag coefficient, taken as 0.02, ||W(s)|| is the wind speed at point s, v_ground(s) is the aircraft's ground velocity, and ds is the path element. This formula comprehensively considers flight distance, wind drag, and hovering power consumption. The energy margin value e(p) is calculated using the following formula: e(p) = (E_remaining - E_consume(p)) / E_remaining; Where E_remaining represents the current remaining battery power. In this embodiment, UAV-01 has 100% battery power at takeoff, and after flying for a period of time, the current remaining battery power is 72%, which is approximately 432 watt-hours. When E_consume(p) is less than E_remaining, e(p) is positive, indicating that the aircraft can reach point p; when E_consume(p) is greater than E_remaining, e(p) is negative, indicating that the battery power is insufficient to support reaching that point. E_norm(p) is obtained by linearly mapping e(p) of all spatial locations within the slice range to the interval [0,1]. The mapping method is to map the maximum positive margin value to 1, the zero margin value to 0.5, and the negative margin value to the interval [0,0.5).

[0048] The superposition of four fields forms a slice of local spatiotemporal resource tension field. After calculating and normalizing the four field components, UAV-01 superimposes the four components point by point according to preset weights. The superposition formula is: F(p,t) = w_d · D_norm(p,t) + w_c · C_norm(p) + w_w · W_norm(p,t) +w_e · E_norm(p); The four weighting coefficients are set to w_d=0.35, w_c=0.30, w_w=0.20, and w_e=0.15, respectively. It should be noted that the constraint tension field component C_norm(p) is constant in the time dimension, therefore it takes the same value for each time t during superposition. The weighting coefficients are assigned based on the degree of influence of each component on flight safety: the density tension field directly reflects the collision risk between aircraft and is given the highest weight; the constraint tension field reflects hard obstacle limitations and is given the second highest weight; the wind field disturbance tensor field and the energy potential field reflect environmental interference and electrical constraints, respectively, and are given relatively lower weights. This weighting allocation was determined through multiple simulation experiments. Under the above weighting configuration, the cluster achieved a flight conflict rate of less than 2% and a mission completion rate of more than 95% in complex urban low-altitude airspace containing three no-fly zones and two controlled areas.

[0049] The final generated local spatiotemporal resource tension field slice F(p,t) is stored as a four-dimensional floating-point array in the dynamic random access memory of the UAV-01 airborne computing platform. The array dimension is 200×200×200×61, which is approximately 488 million floating-point numbers, occupying approximately 1.95 gigabytes of storage space. The airborne computing platform manages this array in blocks, loading data blocks of relevant spatiotemporal regions only as needed during trajectory planning calculations to reduce memory access latency.

[0050] For the dynamic management of the candidate prediction trajectory library, this embodiment pre-sets the library capacity limit K=10 during the initialization phase. After generating 5 candidate prediction trajectories, deterministic sampling is performed in the neighborhood space of each candidate trajectory, guided by the gradient direction of the local spatiotemporal resource tension field slice F(p,t). That is, a new sequence of control points is generated by offsetting a preset distance along the negative direction of the tension field gradient near each control point of the candidate trajectory. These control points are then smoothly connected into variant trajectories through cubic spline interpolation. After merging the variant trajectories with the original candidate trajectories, they are sorted from high to low prior confidence, and the top 10 trajectories are selected to form the initial candidate prediction trajectory library.

[0051] During subsequent iterative trajectory adjustments, UAV-01 continuously monitors the potential conflict intensity of each candidate trajectory in the candidate prediction trajectory library. The potential conflict intensity is calculated by performing a spatiotemporal proximity analysis between the candidate trajectory and the latest received neighboring candidate prediction trajectories, counting and weighting the pairs of trajectory points with a spatial distance of less than 30 meters at each time step. When the potential conflict intensity of a candidate trajectory exceeds a preset conflict tolerance threshold for five consecutive iterations, the candidate trajectory is marked as invalid and removed from the library. Simultaneously, UAV-01 resamples along the negative gradient direction (low tension direction) of the current local spatiotemporal resource tension field slice, using the current master prediction trajectory as a benchmark. Starting from the end point of the current master prediction trajectory, it searches for spatial locations with tension values ​​below 60% of the global mean in the negative gradient direction as new control points, generating supplementary trajectories and injecting them into the candidate prediction trajectory library to maintain the library capacity around K=10. This dynamic management mechanism ensures that the candidate trajectory library always maintains a diverse set of trajectories with low conflict risk to cope with the need for rapid switching when sudden deviations occur.

[0052] Construction of local collaborative topology graph During the iterative trajectory adjustment process, each aircraft constructs a local collaborative topology graph with itself as the root node, based on the distribution of neighboring aircraft within its communication range and high-tension regions in the local spatiotemporal resource tension field slice where the tension value exceeds a preset threshold. This graph is an undirected weighted graph, represented in graph theory as G=(V,E,W), where V is the set of nodes, E is the set of edges, and W is the edge weight matrix.

[0053] Taking UAV-01 as an example, the process of constructing a local cooperative topology graph is as follows. First, the node set V is determined. V contains three types of nodes: the first type is UAV-01 itself, serving as the root node; the second type is all neighboring aircraft within the communication range. In this embodiment, there are 6 neighboring aircraft within the 500-meter communication range of UAV-01, numbered UAV-02 to UAV-07 respectively; the third type is high-tension regions in the local spatiotemporal resource tension field slice where the tension value exceeds a preset threshold. In this embodiment, the preset threshold for high-tension regions is set to a tension field value greater than 0.70. This value was determined through simulation optimization. When the threshold is lower than 0.60, too many low-conflict-risk regions will be introduced into the graph, resulting in an overly dense graph and increasing the computational burden; when the threshold is higher than 0.80, some medium-conflict-risk regions will be missed, affecting the timeliness of early warning. Within the slice area of ​​UAV-01, three high-tension regions were identified, labeled HZ-01, HZ-02, and HZ-03, respectively. Two of them correspond to the areas near the no-fly zone, and one corresponds to the intersection of the predicted trajectories of multiple aircraft.

[0054] After determining the nodes, UAV-01 further determines the edge set E. Whether an edge exists between any two nodes in the graph depends on whether there is a trajectory conflict or resource competition relationship between them. For edges between aircraft nodes, the criterion is the spatiotemporal proximity between the candidate predicted trajectories of the two aircraft. Spatiotemporal proximity is defined as the normalized value of the sum of the reciprocals of the spatial distance between the two aircraft at each moment in the entire future 30-second time domain, and its formula is: S_ij = (1 / T) · ∫_0^T (1 / (||p_i(t) - p_j(t)|| + δ)) dt; Where p_i(t) and p_j(t) are the predicted positions of aircraft i and aircraft j at time t, respectively, T is the prediction time domain length of 30 seconds, and δ is a zero-prevention constant set to 0.5 meters. When S_ij exceeds the preset proximity threshold of 0.15, a conflict is determined between the two nodes, and an edge is added to the graph. For edges between aircraft nodes and high-tension region nodes, the determination criterion is whether the candidate predicted trajectory of the aircraft enters the spatial range of the high-tension region; if so, an edge is added between them. By default, no edges are added between high-tension region nodes.

[0055] After determining the nodes and edges, UAV-01 assigns an initial weight to each edge. The weight of an edge is determined by the superposition of the spatiotemporal proximity between two nodes and the local spatiotemporal resource tension field slice at the node positions at both ends of the edge. The specific calculation formula is as follows: W_e = λ · S_ab + (1-λ) · (F(p_a) + F(p_b)) / 2; Where S_ab represents the spatiotemporal proximity between node a and node b, F(p_a) and F(p_b) are the tension field values ​​at the spatial locations of node a and node b, respectively, and λ is the harmonic coefficient with a value of 0.50. This weight value is between 0 and 1, and the larger the value, the stronger the conflict or competition between the two nodes.

[0056] Conflict intensity monitoring and detection of precursors to sudden congestion changes; During the iterative trajectory adjustment process, UAV-01 calculates the conflict intensity of each edge in the local cooperative topology graph in real time. Conflict intensity is defined as the instantaneous value of the edge weight at the current moment, reflecting the real-time severity of the conflict between two nodes. Crucially, UAV-01 performs second-order differential calculations on the conflict intensity of the corresponding edges of each neighboring aircraft.

[0057] Since the data in the system is updated in discrete time steps, the second-order differential is implemented using the discrete second-order difference method. For the edge between the system and its neighboring UAV-02, let the conflict intensity in the nth iteration be c(n). Then the first-order difference is Δc(n) = c(n) - c(n-1), and the second-order difference is Δ²c(n) = Δc(n) - Δc(n-1) = c(n) - 2c(n-1) + c(n-2). This second-order difference value reflects the acceleration of the rate of change of conflict intensity, that is, the trend of the conflict intensity increasing or decreasing at an accelerated pace.

[0058] When UAV-01 detects that the second derivative of the conflict intensity of an edge with any neighboring aircraft exceeds a preset critical warning threshold, it determines that there is a precursor to a congestion abrupt change. The preset critical warning threshold is 0.05, a value determined based on extensive simulation experiments: when the second derivative exceeds 0.05, it indicates that the conflict intensity is rapidly increasing at an acceleration of more than 0.05 per round, predicting a serious conflict that may occur within the next 3 to 5 iteration cycles. The determination of a precursor to a congestion abrupt change means that the conflict intensity of that edge is not only currently high but is also rapidly deteriorating, requiring immediate intervention rather than waiting for the conventional game to converge.

[0059] Emergency Refactoring Requests and Reflexive Refactoring Operations Upon detecting a congestion abrupt change precursor, UAV-01 immediately activates a rerouting trajectory from its pool of candidate predicted trajectories, cooperating with the neighboring aircraft to replace the current primary predicted trajectory. The selection criteria for this rerouting trajectory are: choosing the trajectory from the pool with the lowest spatiotemporal proximity to the corresponding neighboring aircraft's primary predicted trajectory and the lowest accumulated tension field cost, thus achieving proactive avoidance. Simultaneously, UAV-01 broadcasts an emergency reconfiguration request carrying the intention to change the topology to neighboring aircraft within its communication range. This request includes the target neighboring aircraft number involved in the congestion abrupt change precursor, the estimated coordinates of the conflict zone, and the suggested avoidance direction.

[0060] In the periodically triggered group consensus process, when a common deviation pattern is detected, each aircraft performs a reflexive reconstruction operation on its maintained local collaborative topology graph based on the consensus result. Reflexive reconstruction refers to each aircraft autonomously performing a closed-loop operation of adding and deleting edges on its local topology graph based on the received consensus result. The reconstructed graph structure, in turn, affects the aircraft's subsequent trajectory adjustment game, forming a closed loop of "reconstruction-decision-reconstruction".

[0061] The specific rules for reflexive reconstruction operations include two types of operations: edge removal and edge addition. Edge removal involves removing edges corresponding to directions marked as unreliable in the common deviation pattern. For example, when consensus results reveal a widespread systematic bias in the tension field prediction for a certain region, each aircraft removes edges pointing to that region from its topology graph to avoid making incorrect decisions based on unreliable information. Edge addition involves adding new edges with specific neighboring aircraft based on the potential cooperative dependencies revealed by the common deviation pattern. For example, when consensus results find a long-term coupling relationship between the trajectories of two aircraft, a new cooperative edge is established between them to more closely coordinate their trajectories. The reconstructed local cooperative topology graph is used for trajectory adjustment games in the next cycle.

[0062] Algebraic connectivity monitoring and critical collaborative edge protection To further enhance the robustness of the cluster topology, UAV-01 calculates the algebraic connectivity of its local cooperative topology graph upon detecting precursory congestion changes. Algebraic connectivity is the second smallest eigenvalue λ² of the graph's Laplacian matrix, used to measure the strength of the graph's connectivity. A larger λ² value indicates better graph connectivity and higher information propagation efficiency within the cluster; a λ² value close to 0 indicates that the graph is nearing a fragmented state.

[0063] UAV-01 calculates algebraic connectivity using the following steps. First, it constructs a Laplacian matrix L based on the local cooperative topology graph G=(V,E,W), with diagonal elements L_ii = Σ_j W_ij and off-diagonal elements L_ij = -W_ij, where W_ij is the weight of edge (i,j). Then, it calculates the eigenvalues ​​of L using a power iteration method, taking the second smallest eigenvalue as the algebraic connectivity λ2. UAV-01 simultaneously monitors the rate of change of λ2 within a preset sliding time window, which has a length of 5 iteration cycles.

[0064] If the rate of decrease in algebraic connectivity exceeds a preset split warning threshold, a precursor to a sudden split in the cluster network is identified. The preset split warning threshold is a decrease of 0.02 per iteration cycle, corresponding to the critical point where the cluster's connectivity in a graph theory sense deteriorates rapidly within a short period. When the rate of decrease of λ² is below this threshold, it indicates that the structural changes in the topology graph are gradual, and cluster communication remains unimpeded. When the rate of decrease exceeds this threshold, it indicates that the cluster has a tendency to split into multiple disconnected subgroups, which will seriously affect subsequent information exchange and cooperative game performance.

[0065] After identifying the precursor to a splitting mutation, UAV-01 attaches an algebraic connectivity protection instruction to its broadcast emergency reconfiguration request. This instruction instructs neighboring aircraft receiving the instruction to refrain from removing critical cooperative edges with the sending aircraft during subsequent reflexive reconfiguration operations. The critical cooperative edge identification method employs the Fiedler vector method from graph theory: the eigenvector corresponding to the algebraic connectivity λ² is calculated, i.e., the Fiedler vector f. For each edge (i,j) in the graph, (f_i - f_j)² is calculated; the edge with the largest value is the critical cooperative edge that contributes the most to maintaining algebraic connectivity, because the edge with the largest difference between its two ends in the Fiedler vector is the bottleneck restricting overall connectivity. UAV-01 encodes the calculated critical cooperative edge information in the protection instruction and broadcasts it along with the request.

[0066] Dynamic adjustment of edge density In the reflexive reconfiguration operation, UAV-01 dynamically adjusts the edge density of its local cooperative topology graph based on the available bandwidth of the current communication channel and the mission coordination accuracy requirements. The edge density is defined as the ratio of the actual number of edges in the graph to the maximum possible number of edges.

[0067] The available bandwidth of the communication channel is estimated in real time by monitoring the number of data packets successfully received per unit time. When the available bandwidth is lower than the preset bandwidth threshold of 1 megabits per second, it indicates that communication resources are strained. UAV-01 prioritizes removing edges with a collision intensity lower than the preset collision threshold of 0.30 to reduce communication overhead. When the mission coordination accuracy requirement is higher than the preset accuracy threshold, such as requiring a position deviation of less than 0.5 meters during the close formation flight phase, UAV-01 prioritizes adding edges with neighboring aircraft with high predictable trajectory proximity to enhance coordinated control. The mission coordination accuracy requirement is automatically determined by the aircraft's current mission phase, with high accuracy requirements during the assembly and close formation phases and low accuracy requirements during the wide-area search phase.

[0068] Construction of the Risk Balance Cost Function and Dynamic Decision Making During actual flight, the cluster may simultaneously face the combined risks of impending congestion and network fragmentation. Taking UAV-01 as an example, when it detects that the second-order derivative of the conflict intensity of the edge with UAV-02 reaches 0.065, exceeding the critical warning threshold of 0.05, its local cooperative topology graph's algebraic connectivity decrease rate also reaches 0.028 per iteration cycle, exceeding the fragmentation warning threshold of 0.02. In this situation, UAV-01 needs to address both types of risks simultaneously, rather than handling one in isolation.

[0069] To address the aforementioned concurrent risk scenarios, UAV-01 constructs a risk balancing cost function. This cost function consists of a weighted sum of conflict risk terms and separation risk terms, and its formula is as follows: R_total = μ · R_conflict + ν · R_disconnect; Where R_conflict is the conflict risk term and R_disconnect is the disconnect risk term, μ and ν are the weight coefficients of the two terms, satisfying μ+ν=1. The value of the conflict risk term is determined based on the extent to which the second derivative of the conflict intensity exceeds a preset critical warning threshold. The extent to which it exceeds the threshold is defined as the difference between the measured value of the second derivative and the threshold, i.e., Δ_conf = 0.065 - 0.05 = 0.015, which is normalized and used as the value of R_conflict. The value of the disconnect risk term is determined based on the extent to which the rate of decrease in algebraic connectivity exceeds a preset disconnect warning threshold, i.e., Δ_conn = 0.028 - 0.02 = 0.008, which is also normalized and used as the value of R_disconnect. The two normalized baseline values ​​are the maximum permissible value for conflict exceeding the limit of 0.03 and the maximum permissible value for disconnect exceeding the limit of 0.015. Therefore, the initial risk of UAV-01 is R_conflict=0.015 / 0.03=0.50 and R_disconnect=0.008 / 0.015≈0.53, and the total risk is at a relatively high level.

[0070] The weighting coefficients μ and ν are dynamically adjusted based on the current mission phase and mission priority of UAV-01. The mission phase is autonomously identified by the aircraft, based on the average spatial distance between itself and all neighboring aircraft within its communication range. When the average distance is less than 30 meters, it is determined to be in a gathering or dense formation phase; when the average distance is greater than 80 meters, it is determined to be in a wide-area search or sparse distribution phase; and when it is between 30 and 80 meters, it is a transition phase. In this embodiment, UAV-01 is in a dense formation phase with an average neighbor distance of 22 meters. At this time, the spatial coupling between aircraft is tight, and the collision risk is much higher than the network connectivity risk. Therefore, the conflict risk weight μ is set to 0.75, and the separation risk weight ν is set to 0.25. Experiments show that when a conflict weight in the range of 0.70 to 0.80 is used in the dense formation phase, the formation maintenance error is less than 0.8 meters and there are no collision records; when the conflict weight is lower than 0.60, the probability of close-range collisions within the formation increases to over 5%. During the wide-area search phase, the weight configuration is reversed to μ=0.30 and ν=0.70 to prioritize the connectivity of the cluster network and the information coverage.

[0071] UAV-01 dynamically determines the values ​​of three key decision variables by minimizing the risk balance cost function: the edge removal threshold in the reflexive reconstruction operation, the protection priority of key cooperative edges, and the transmission rate of broadcasting emergency reconstruction requests to neighboring aircraft within communication range. The edge removal threshold is continuously adjustable between 0.20 and 0.60, and is reduced to remove more potentially dangerous edges when the total risk R_total is high. The protection priority of key cooperative edges is sorted from high to low according to the contribution of each edge to algebraic connectivity, and the higher the total risk, the more edges are in the protection priority sequence. The transmission rate of emergency reconstruction requests is adjusted between 1 and 5 times per iteration cycle, and the transmission rate is increased when the risk is high to ensure timely dissemination of early warning information.

[0072] Mapping relationship between task priority and weight coefficient Mission priorities are assigned by the ground mission planning system when a mission is issued, and are divided into four levels: emergency obstacle avoidance, formation maintenance, path tracking, and energy-saving cruise, corresponding to priority values ​​of 4, 3, 2, and 1, respectively. UAV-01 is currently performing formation maintenance, with a priority value of 3. The priority value directly affects the baseline value of the risk balance weighting coefficient: higher priority missions are more sensitive to risk, meaning that a more aggressive adjustment strategy is adopted under the same risk level. Specifically, for every 1 increase in priority value, the adjustment amplification factor of the weighting coefficient increases by 0.15.

[0073] Initial construction of the alternative predicted trajectory library During the trajectory planning phase, UAV-01 needs to maintain a limited but sufficiently diverse pool of alternative predicted trajectories to cope with various unforeseen circumstances that may occur during flight. The generation and maintenance of the alternative predicted trajectory pool is divided into two phases: initial construction and online evolution.

[0074] The initial construction phase is executed immediately after UAV-01 generates five candidate predicted trajectories. Guided by the gradient direction of the local spatiotemporal resource tension field slice, UAV-01 performs deterministic sampling in the neighborhood of each candidate predicted trajectory. It should be noted that although this sampling is named random sampling, to ensure the repeatability of the results and the determinism of the patented technical solution, a deterministic low-difference sequence is actually used for sampling. The sampling seed is determined by the fixed device number of UAV-01 and the current round number, ensuring that each run generates completely consistent sampling results under the same conditions. Specifically, six sampling points are generated around each control point of each candidate trajectory, within a ring-shaped area with a radius of 5 to 15 meters centered on the control point, in the vertical plane along the tension field gradient direction. The sampling point and control point sequences are then concatenated in chronological order and smoothly connected using cubic spline interpolation to generate a complete variant trajectory. Cubic spline interpolation ensures that the position, velocity, and acceleration of the trajectory are continuous at the connection point, satisfying the C² continuity requirement, making the generated variant trajectory dynamically executable.

[0075] After merging the variant trajectories generated in the above manner with the original five candidate predicted trajectories, UAV-01 sorts each trajectory from highest to lowest prior confidence and selects the top K trajectories to form the initial candidate predicted trajectory library. K is a preset upper limit for the library capacity; in this embodiment, K is set to 10. This value is determined based on the storage capacity of the airborne computing platform and the real-time requirements of the trajectory planning algorithm: when K is less than 6, the diversity of candidate trajectories is insufficient, and effective replacement trajectories cannot be provided when sudden deviations occur; when K is greater than 15, the computational overhead of trajectory library maintenance leads to an extension of approximately 30 milliseconds in each iteration cycle, affecting the system's real-time response capability. Experiments show that the overall performance is optimal when K is between 10 and 12.

[0076] Online evolution and dynamic maintenance of the alternative predicted trajectory library During the iterative trajectory adjustment process, UAV-01 continuously monitors the potential conflict intensity of each candidate trajectory in its candidate predicted trajectory library. The potential conflict intensity is calculated as follows: a time-by-time spatiotemporal proximity analysis is performed between the candidate trajectory and the latest candidate predicted trajectories broadcast by neighboring aircraft. For each time point, if the Euclidean distance between the spatial position on the candidate trajectory at that time and the spatial position on any neighboring trajectory at the same time is less than 30 meters, then that time point is marked as a conflict time. The potential conflict intensity is defined as the proportion of conflict times to the total number of times in the entire prediction time domain, ranging from 0 to 1.

[0077] When the potential conflict intensity of a candidate trajectory exceeds a preset conflict tolerance threshold for N consecutive iterations, the candidate trajectory is marked as invalid and removed from the database. In this embodiment, N is set to 5, and the conflict tolerance threshold is set to 0.25. That is, if there are conflicts with a proximity of less than 30 meters at more than 25% of the times within 5 consecutive iterations, the candidate trajectory is determined to be no longer safe and usable. The threshold is set based on the following: when the potential conflict intensity is below 0.25, the trajectory only has sporadic moments of proximity with neighboring trajectories, which can be resolved through conventional game theory adjustments; when it exceeds 0.25, it indicates that the trajectory has systematic conflicts with its neighbors over a long period of time, and its retention value is low.

[0078] While removing failed trajectories, UAV-01 uses the current master predicted trajectory as a benchmark, resamples along the low-tension direction of the current local spatiotemporal resource tension field slice, and generates supplementary trajectories, which are then injected into the candidate predicted trajectory library to maintain the library capacity around K. The low-tension direction is defined as the negative gradient direction of the tension field function, i.e., the direction in which the tension value decreases the fastest. UAV-01 applies an offset along the negative gradient direction to the terminal control point of the current master predicted trajectory, with an offset step size of 10 meters, generating a new control point sequence, and then generates a complete supplementary trajectory through cubic spline interpolation. The cumulative tension field cost of this supplementary trajectory is usually lower than that of the removed failed trajectories, thus having a higher prior confidence and effectively improving the overall quality of the candidate trajectory library. Through the above cyclical operation of removal and replenishment, the candidate predicted trajectory library always maintains K diverse trajectories that can be switched at any time.

[0079] During flight, UAV-01 uses the main predicted trajectory as its current execution trajectory for flight control, while continuously monitoring the prediction deviation between the actual perceived field and the predicted field. The actual perceived field is constructed in real-time based on the latest environmental perception information acquired by the onboard sensors at the current moment. Its spatial range and temporal resolution are consistent with the predicted field, both being spherical regions with a radius of 500 meters and a temporal resolution of 0.5 seconds. The predicted field is a slice of the local spatiotemporal resource tension field F(p,t) at the current moment, stored in the memory of the onboard computing platform.

[0080] Prediction bias is quantified in the form of a bias field, defined as the difference between the field values ​​of the actual sensed field and the predicted field at the same spatial location and time, expressed by the formula ΔF(p,t) = F_actual(p,t) - F_predicted(p,t). UAV-01 calculates the bias field values ​​at all spatial grid nodes within each communication cycle, resulting in a bias field array with the same dimensions as the tension field slice. This bias field reflects the degree of deviation between the actual environmental state and the prediction model, serving as the basis for subsequent bias classification and feedback adjustments.

[0081] UAV-01 maintains a sliding time window to record historical data on prediction bias. This sliding time window is 3 seconds long, meaning it saves all bias field data for every 0.5 seconds within the past 3 seconds, totaling 7 snapshots, starting from the current moment. The 3-second window length is chosen because it covers the shortest time required for UAV-01 to complete a full trajectory switch or parameter adjustment from detecting a bias. A window that is too short cannot distinguish between short-term fluctuations and persistent biases, while a window that is too long will introduce outdated environmental information, affecting classification accuracy.

[0082] UAV-01 performs a spatiotemporal dual-dimensional statistical test on the prediction deviation data within the sliding window to automatically distinguish between consistent deviations and sudden deviations. The specific implementation of this test method is as follows: In the spatial dimension, the number of spatial grid nodes in the UAV-01 statistical deviation field whose deviation values ​​exceed a preset deviation threshold of 0.15 is counted, and the spatial range covered by these threshold-exceeding nodes is calculated. When the radius of the threshold-exceeding region exceeds the radius of UAV-01's own neighborhood (500 meters), and this threshold-exceeding region contains the current positions of at least three neighboring aircraft, the deviation is judged to have a wide-area diffusion characteristic in the spatial dimension. In the temporal dimension, the duration for which the UAV-01 statistical deviation value exceeds the preset deviation threshold of 0.15 is counted. When the duration exceeds a preset duration threshold of 2 seconds, the deviation is judged to have a persistent characteristic in the temporal dimension. When the deviation meets the above conditions in both the spatial and temporal dimensions, it is classified as a consistent deviation; otherwise, it is classified as a sudden deviation.

[0083] It should be noted that all preset thresholds involved in the above testing methods were calibrated through offline simulation. The deviation threshold of 0.15 is calibrated based on the fact that this value corresponds to the smallest resolvable environmental change under the normalized dimensions of the tension field. Deviations smaller than this value usually originate from sensor noise and do not require triggering feedback operations. The duration threshold of 2 seconds corresponds to the sum of the response delay of the UAV-01 airborne flight control system and the time required for a complete trajectory replanning. The neighbor number threshold of 3 is set based on the engineering experience that each aircraft in the cluster needs to maintain coordination with at least 3 neighbors to maintain formation stability.

[0084] Once UAV-01 determines that the current prediction deviation is a consistency deviation, it enters the confirmation phase. UAV-01 collects confirmation signals from neighboring aircraft affected by the deviation within its communication range via its air-to-air communication module, alleging the same deviation pattern. The confirmation signal is generated as follows: each neighboring aircraft independently performs the same spatiotemporal two-dimensional statistical test. If the neighboring aircraft also detects a consistency deviation, it sends a confirmation data packet back to UAV-01. UAV-01 continuously collects confirmation signals. When the number of confirmation signals reaches a preset consensus threshold, it confirms and executes the prediction model parameter adjustment operation. In this embodiment, the preset consensus threshold is set to more than 50% of the number of effective neighbors within the communication range. Since UAV-01 currently has 6 effective neighbors within its communication range, confirmation signals from at least 4 neighboring aircraft are required. This threshold is set based on the majority confirmation principle to avoid a single aircraft mistakenly triggering global parameter adjustments due to sensor anomalies.

[0085] After receiving sufficient confirmation signals, UAV-01 performs prediction model parameter adjustment operations. Specifically, this adjustment involves correcting the bandwidth parameter σ of the Gaussian kernel in the density tension field calculation, and adjusting the time extrapolation coefficients of the wind field disturbance tensor components. The bandwidth parameter σ is adjusted as follows: when the consistency deviation indicates a systematic underestimation of the density field prediction value, σ is increased from its current value of 20 meters to 25 meters, expanding the spatial influence range of the density contribution value and thus improving the sensitivity to nearby aircraft; when the consistency deviation indicates a systematic overestimation of the density field prediction value, σ is decreased to 16 meters, narrowing the influence range to reduce the false alarm rate. The time extrapolation coefficients of the wind field disturbance tensor components are adjusted by increasing or decreasing the weight decay rate of the wind field forecast data in the future time domain, making the predicted wind field more closely match the changing trend of the measured wind field. The adjusted parameters take effect immediately and are used for the next round of density tension field calculations.

[0086] Once UAV-01 determines that the current prediction deviation is a sudden deviation, it immediately switches from the candidate prediction trajectory library to the target candidate prediction trajectory. The selection criteria for the target candidate prediction trajectory are as follows: traverse all K candidate trajectories in the candidate prediction trajectory library, calculate the deviation field norm between the prediction field and the actual sensed field for each candidate trajectory, and select the candidate trajectory with the smallest norm, i.e., the one that best matches the current actual sensed field, as the new current execution trajectory. This switching operation is completed within the same communication cycle, with a switching latency of less than 100 milliseconds, ensuring that the aircraft can respond quickly to sudden situations.

[0087] Simultaneously with the trajectory switch, UAV-01 broadcasts a sudden deviation warning to neighboring aircraft within its communication range. This warning data packet contains three key fields. The first field is the estimated location of the sudden deviation, calculated by taking the coordinates of the spatial grid node with the largest absolute deviation value in the deviation field as the estimated location of the deviation source. The second field is the radius of influence, calculated by taking the estimated location as the center and calculating the maximum distance between the center and the spatial grid node whose absolute deviation value exceeds a preset deviation threshold of 0.15. The third field is the estimated duration, calculated by linearly extrapolating the deviation field's trend over the last three moments to estimate the time required for the deviation value to decay below the preset deviation threshold, in seconds.

[0088] Each aircraft in the cluster continuously monitors the topological criticality indicators of its local collaborative topology graph to determine whether a group consensus process needs to be initiated. The topological criticality indicators are of two types. The first type is the cluster's local mean of conflict intensity, calculated as the arithmetic mean of the conflict intensity of each edge connected to all neighboring aircraft in the aircraft's local collaborative topology graph. The second type is the cluster's local variance of algebraic connectivity, calculated as the variance of the algebraic connectivity of the aircraft's local collaborative topology graph over the most recent five iterations.

[0089] When any topological critical index meets the preset adaptive triggering condition, the aircraft initiates a group consensus request within its communication range. The specific formula for the preset adaptive triggering condition is: the cluster local mean of conflict intensity is greater than a preset mean threshold of 0.60, or the cluster local variance of algebraic connectivity is greater than a preset variance threshold of 0.15. These two thresholds are set based on the following: a mean threshold of 0.60 corresponds to a high overall conflict level in the cluster, requiring coordination and adjustment; a variance threshold of 0.15 corresponds to drastic fluctuations in algebraic connectivity, indicating an unstable cluster topology. These thresholds are fixed, but UAV-01 performs online calibration every 50 iterations based on the statistical distribution of historical data to maintain them at a suitable level sensitive to changes in the current environment; hence, this is called an adaptive triggering condition.

[0090] Once UAV-01 meets the triggering conditions and initiates a group consensus request, the cluster enters a layered consensus process. This process adopts a two-layer consensus architecture, and the specific execution flow is as follows.

[0091] The first layer is the local cluster pre-consensus phase. UAV-01 and all its neighboring aircraft within its one-hop communication range form a local cluster. The cluster size is the total number of aircraft within the current communication range; in this embodiment, it is 7. Each aircraft within the cluster exchanges its prediction bias patterns. The data structure of the prediction bias pattern is a feature vector containing three dimensions: the spatial gradient direction of the bias field, the rate of change over time, and the mean bias value. After each aircraft aggregates its bias pattern feature vectors, a K-means clustering algorithm is executed locally on UAV-01. The number of clusters is preset to 3, resulting in a local clustering result. If the proportion of members in a certain cluster exceeds the preset cluster consensus threshold of 60%, the cluster center of that cluster is output as the local pre-consensus result. If the proportion of members in all clusters is less than 60%, the local consensus fails and waits for the next cycle to trigger again. The local pre-consensus result includes the feature vector of the consensus bias pattern and the list of members who have reached a consensus.

[0092] The second layer is the backbone aircraft diffusion and alignment phase. Within a local cluster, each aircraft elects a backbone aircraft based on the criterion of having direct communication links with the most neighboring aircraft. This criterion ensures the backbone aircraft has optimal information propagation capabilities. In this embodiment, UAV-03 has direct links with 5 aircraft, more than UAV-01's 4, therefore UAV-03 is elected as the backbone aircraft. UAV-03 encodes the local pre-consensus results and broadcasts them to neighboring clusters. Upon receiving multiple local pre-consensus results, the backbone aircraft in neighboring clusters perform secondary alignment and clustering locally, merging the pre-consensus results from different clusters based on feature vector similarity to generate a global consensus result. This global consensus result, after confirmation, is distributed and diffused throughout the entire cluster via broadcast. The diffusion uses a multi-hop relay method, with each aircraft receiving the global consensus result forwarding it to its neighbors until all reachable aircraft within the cluster are covered.

[0093] After the global consensus result is confirmed, two types of information are output. The first type is the adjusted weight parameters of the distributed game payoff function and the adjusted parameters of the local spatiotemporal resource tension field slice evolution equation. Each spacecraft synchronously updates its local weight coefficients and evolution equation coefficients based on the received adjusted values. The second type is a topology reconstruction mask, which specifies the set of edges to be removed and the set of potential cooperative edges to be added by each spacecraft in the reflexive reconstruction operation. The data structure of the topology reconstruction mask is a list of edge operations. Each record in the list contains the operation type (removal or addition), source node number, target node number, and the round number of the operation execution. After receiving the topology reconstruction mask, each spacecraft synchronously executes the edge operations at the specified round boundaries, ensuring that the topology graph structure switch of the entire cluster remains time-consistent and avoiding inconsistencies in the topology state caused by some spacecraft performing reconstruction prematurely.

[0094] To further improve the accuracy and stability of decision-making in concurrency risk scenarios, this embodiment has refined the construction method of the cost function and the minimization solution method.

[0095] When UAV-01 simultaneously detects both congestion precursors and cluster network discontinuity precursors, the conflict risk term and discontinuity risk term in its constructed risk balance cost function are valued as follows: The conflict risk term no longer uses the linearly normalized value of the excess amplitude, but instead takes the square of the excess amplitude of the second derivative of the conflict intensity exceeding the preset critical warning threshold. Let the measured value of the second derivative of the conflict intensity be d_conf, and the preset critical warning threshold be T_conf = 0.05, then the excess amplitude Δ_conf = d_conf - T_conf, and the conflict risk term R_conflict = Δ_conf². Similarly, let the measured value of the algebraic connectivity decrease rate be r_conn, and the preset discontinuity warning threshold be T_conn = 0.02, then the excess amplitude Δ_conn = r_conn - T_conn, and the discontinuity risk term R_disconnect = Δ_conn².

[0096] Replacing linear values ​​with squared values ​​has clear physical significance and technical benefits. When the over-limit amplitude is small, squaring further compresses it, preventing minor over-limits caused by noise from triggering overly aggressive responses. When the over-limit amplitude is large, squaring dramatically amplifies it, ensuring that deviations approaching the critical state receive high attention from the decision-making system. Taking actual data from UAV-01 as an example, when Δ_conf=0.015, the linear value is 0.015, which becomes 0.000225 after squaring, a compression of approximately 98.5%. However, if Δ_conf reaches 0.03, the linear value is 0.03, which becomes 0.0009 after squaring, amplifying it to 30 times the linear value. This nonlinear mapping allows the aircraft to maintain a moderate response in the early stages of risk accumulation and rapidly increase decision priority when the risk approaches the critical state, effectively distinguishing between ordinary disturbances and true precursor signals. Experiments show that after adopting squared nonlinearity, the false alarm rate decreases from approximately 8% in the linear scheme to approximately 2%, while the detection rate of true precursors remains above 96%.

[0097] The adjustment method for the weighting coefficients has also been improved. The conflict risk weight μ and the separation risk weight ν no longer change discretely in a phase-switching manner, but are continuously and smoothly adjusted according to the aircraft's current mission phase and mission priority. UAV-01 defines two key variables to drive the weighting adjustment. The first variable is the mission progress percentage, which is calculated as the proportion of the number of waypoints completed by the aircraft to the total number of waypoints, with a value ranging from 0 to 100%. The second variable is the current local formation density, which is calculated as the average of the reciprocals of the distances between the aircraft and all neighboring aircraft within the communication range, in meters. A local formation density greater than 0.03 per meter indicates a dense formation state, while a density less than 0.012 per meter indicates a sparse formation state.

[0098] The UAV-01 uses a preset fuzzy membership function to map the two variables mentioned above into weight adjustment factors. The fuzzy membership function adopts an S-curve form, and its formula is as follows: S(x; a, b) = 1 / (1 + exp(-(x - a) / b)); Where x is the input variable, a is the center point parameter of the S-curve, and b is the slope parameter of the curve. For the mission progress percentage, a = 50% and b = 15%. When the mission progress percentage is less than 50%, the aircraft is in the early stage of the mission, and the formation and assembly requirements are high. At this time, the S value is small, and the weight is tilted towards the conflict risk item. When the mission progress percentage is greater than 50%, the S value increases, and the weight is tilted towards the separation risk item. For the local formation density, a = 0.02 per meter and b = 0.005 per meter. When the local formation density is greater than 0.02 per meter, the S value is large, and the conflict risk weight increases; when the local formation density is less than 0.02 per meter, the S value is small, and the separation risk weight increases.

[0099] The overall weighting coefficient is calculated using the following formula: μ = S(progress) · (1 - α) + S(density) · α; ν = 1 – μ; Where "progress" represents the mission progress percentage, "density" represents the local formation density, and "α" is a balancing factor of 0.5, used to adjust the relative contribution of the two variables to the weights. This S-curve mapping method ensures that the weights change continuously during mission phase transitions, without any abrupt jumps, thus avoiding abrupt weight changes. Simulation results show that when the change in weight coefficients between two adjacent iteration cycles exceeds 0.15, the aircraft's trajectory adjustment exhibits significant oscillations, extending the average convergence time by approximately 40%. However, by adopting the continuous smooth adjustment method of this embodiment, the weight change between adjacent cycles is controlled within 0.05, resulting in a smooth and stable trajectory adjustment process.

[0100] The risk balance cost function is minimized using gradient descent with a momentum term. The cost function that UAV-01 needs to minimize is R_total = μ·R_conflict + ν·R_disconnect, with three decision variables: the edge removal threshold θ_remove, the protection priority ranking vector p_protect for critical collaborative edges, and the emergency reconfiguration request sending rate f_broadcast. Since p_protect is a discrete ranking variable, it is relaxed to a continuous protection priority score vector in actual optimization, with the score for each edge ranging from 0 to 1; a higher score indicates a higher protection priority.

[0101] The iterative update rule of the gradient descent method is as follows: x_{n+1} = x_n - η · ∇R_total(x_n) + ρ · (x_n - x_{n-1}); Where x is the decision variable vector, η is the learning rate, ∇R_total is the gradient of the cost function with respect to the decision variables, and ρ is the momentum term coefficient. The momentum term ρ·(x_n-x_{n-1}) is used to accumulate historical gradient directions. Its mechanism is as follows: when the gradient direction is consistent in multiple consecutive iterations, the momentum term accelerates the movement of the decision variables along that direction, making the optimization process converge to the extreme point faster; when the gradient direction oscillates, the momentum term plays a damping role, suppressing optimization oscillations caused by instantaneous fluctuations in local topology. In this embodiment, the learning rate η is taken as 0.15, and the momentum term coefficient ρ is taken as 0.85. This configuration achieves a balance between response speed and stability, enabling the optimization process to converge to a stable solution within 5 to 8 iterations. Experiments show that, without the momentum term, when conflict intensity and connectivity alternately dominate, the decision variable will oscillate back and forth with an amplitude of 0.2 each time, and the convergence time will exceed 20 iteration cycles. After adopting the momentum term, the oscillation amplitude is suppressed to within 0.04, and the convergence time is shortened to 7 iteration cycles.

[0102] Through the gradient descent optimization described above, UAV-01 dynamically determines the optimal edge removal threshold, protection priority ranking, and broadcast transmission rate in each round. For example, in a dense formation scenario with only 30% mission progress, the optimized edge removal threshold is 0.52, meaning only edges with a conflict intensity higher than 0.52 are removed, indicating that the aircraft tends to retain more topological connections to maintain formation stability. The three edges with the highest protection priority scores for key cooperative edges correspond to connections with UAV-02, UAV-04, and UAV-06, respectively. The broadcast transmission rate is 4 times per iteration cycle to ensure rapid dissemination of warning information. In a wide-area search scenario with 80% mission progress, the optimized edge removal threshold decreases to 0.28, and the broadcast transmission rate decreases to 2 times per iteration cycle, indicating that the aircraft tends to remove conflicting edges for more flexible maneuverability, and the demand for communication speed is correspondingly reduced.

[0103] In actual cluster operation, each aircraft independently performs spatiotemporal two-dimensional statistical tests on prediction deviations based on its own sensor data. The results of these tests may differ from the cluster consensus results due to sensor noise, communication delays, or local environmental differences. This embodiment addresses scenarios where there is a discrepancy between the collection of confirmation signals for consistency deviations and the local pre-consensus results in the group consensus process. It introduces a consensus confidence-driven arbitration mechanism to ensure that each aircraft can make reasonable decisions when its local understanding differs from the global consensus.

[0104] Taking UAV-01 as an example, suppose that in a certain iteration cycle, it determines that the current prediction deviation is a consistency deviation through a spatiotemporal two-dimensional statistical test and broadcasts a confirmation signal request to neighboring aircraft within its communication range. During the subsequent confirmation signal collection process, UAV-01 only receives confirmation signals from two neighboring aircraft, failing to reach the preset consensus threshold of four. Simultaneously, the local pre-consensus phase of the group consensus process initiated by UAV-01 yields the opposite conclusion—the local clustering results show that the current deviation pattern is a sudden deviation rather than a consistency deviation. At this point, UAV-01 faces a direct discrepancy between its own judgment and the cluster consensus, requiring an arbitration mechanism to guide subsequent actions.

[0105] UAV-01 maintains a consensus confidence variable C_conf for local prediction bias. This variable is a continuous value ranging from 0 to 1, initially set to 0.50, representing the neutral level of confidence in its own judgment. The consensus confidence variable is dynamically updated based on the time decay factor and the strength of the neighbor acknowledgment signal.

[0106] The time decay factor applies to the consensus confidence variable in an exponential decay manner. The specific update rule is as follows: at the beginning of each iteration cycle, UAV-01 first applies a time decay to the consensus confidence variable, with the decay formula being C_conf = C_conf · exp(-Δt / τ), where Δt is the duration of a single iteration cycle (100 milliseconds), and τ is a time constant, which in this embodiment is 2 seconds. The time constant is chosen because within approximately 20 iteration cycles (2 seconds), the consensus confidence without external confirmation will decay from the initial value of 0.50 to 0.50·exp(-2 / 2)≈0.184, a decrease of approximately 63%. When no confirmation signal is received within a preset silent time window, the consensus confidence variable decreases according to the aforementioned exponential decay factor. The length of the silent time window is set to 2 seconds, matching the time constant τ. When there are no confirmation signals continuously within the window, it indicates that the aircraft's confidence in its own judgment should decrease over time.

[0107] The effect of neighbor acknowledgment signal strength is the opposite of time decay. When UAV-01 receives acknowledgment signals from neighboring aircraft for the same deviation pattern, the consensus confidence variable increases by a preset increment step. The increment step for each acknowledgment signal is 0.08, meaning that C_conf increases by 0.08 for each acknowledgment received. Acknowledgment signals from different neighbors have the same increment step, and signal strength is no longer differentiated because the binary nature of the acknowledgment signal—whether it is acknowledged or not—determines its information value, eliminating the need for further strength quantification. When UAV-01 receives two acknowledgment signals in a single iteration cycle, C_conf increases by 0.16, partially offsetting the effect of time decay. This increment step setting ensures that the consensus confidence can quickly climb above the confidence threshold when receiving acknowledgments from a majority of neighbors, while gradually decaying to zero over time in the absence of external support.

[0108] The consensus confidence variable plays a crucial role in the local pre-consensus stage of the group consensus process. During local K-means clustering, UAV-01 uses the consensus confidence variable as a weighting coefficient for its own bias patterns in the clustering. Specifically, the UAV-01 bias pattern feature vector is multiplied by a weighting factor w_i = 0.50 + C_conf in the clustering distance calculation, where C_conf is the current consensus confidence value. When C_conf is high, the weighting factor is large, and the UAV-01 bias pattern has a higher cluster center attraction in the clustering. That is, the bias pattern feature vectors of other aircraft will move towards UAV-01 when calculating the distance, thereby accelerating the convergence of the local pre-consensus. In this embodiment, UAV-01's C_conf is 0.72, and the weighting factor is 1.22, higher than the default value of 1.0, which enhances the influence of the aircraft's bias pattern in the clustering. When the consensus confidence level is lower than the preset confidence threshold of 0.40, the weight factor is less than 0.90, and the influence of the aircraft's deviation pattern in clustering is weakened, thus preventing low-confidence judgments from interfering with consensus formation. The preset confidence threshold of 0.40 is set based on the following: when the consensus confidence level is lower than 0.40, it indicates that the aircraft has not obtained sufficient external confirmation in the most recent iterations, and the confidence level of its judgment results is lower than randomness, and it should not dominate the consensus direction.

[0109] If the deviation judgment result of UAV-01 is inconsistent with the local pre-consensus result, and the consensus confidence variable of UAV-01 is lower than the preset doubt threshold of 0.35, UAV-01 abandons its own judgment result, adopts the local pre-consensus result, and adjusts its internal prediction model parameters accordingly. The preset doubt threshold of 0.35 is set based on the fact that this value is lower than the confidence threshold of 0.40, and a hysteresis interval of 0.05 is retained to prevent frequent switching of judgment strategies near the threshold. Simulation verification shows that the introduction of the hysteresis interval reduces the frequency of strategy switching of UAV-01 in the consensus confidence critical region by about 70%, effectively avoiding the oscillation of prediction model parameters caused by repeated switching of judgment strategies.

[0110] If the consensus confidence variable is higher than or equal to the preset doubt threshold of 0.35, UAV-01 retains its own judgment result and reports its own judgment result with the consensus confidence variable as an independent opinion when the backbone aircraft performs secondary alignment clustering. The reported data packet contains three fields: the type of deviation judgment result of UAV-01, the current consensus confidence value C_conf, and the feature vector of the deviation pattern. After receiving independent opinions from multiple aircraft, the backbone aircraft performs weighted arbitration based on the consensus confidence of each independent opinion. The specific rules of weighted arbitration are as follows: in secondary alignment clustering, the cluster center attraction weight of each independent opinion is proportional to its reported consensus confidence value. Independent opinions with higher consensus confidence have a larger proportion in the calculation of cluster centers, and the final output global consensus result tends to trust the judgment of the aircraft with higher consensus confidence.

[0111] Taking UAV-01's actual experience as an example, during one flight, the consensus confidence level C_conf = 0.72, while the local pre-consensus result indicated a sudden deviation. Since C_conf was higher than the doubt threshold of 0.35, UAV-01 reported its own consensus deviation judgment as an independent opinion. UAV-03 and UAV-05 reported sudden deviation judgments, with C_conf values ​​of 0.58 and 0.63, respectively. The backbone aircraft arbitrated the results in a second-order aligned clustering process, weighted by confidence level. The weighted scores for the three types of opinions were 0.72 for UAV-01, 0.58 for UAV-03, and 0.63 for UAV-05. UAV-01 scored the highest, and the final global consensus result adopted UAV-01's consensus deviation judgment. This global consensus result was subsequently adopted by the entire cluster through distributed diffusion. Based on this, each aircraft corrected the bandwidth parameter of the Gaussian kernel in the density tension field, successfully eliminating the systematic prediction bias caused by local weather changes.

[0112] This embodiment provides a dynamic scheduling system for low-altitude aircraft clusters based on multi-source data fusion. The system consists of multiple aircraft nodes, each of which has complete sensing, computing, communication and control capabilities.

[0113] Taking UAV-01 as an example, the airborne computing platform of this aircraft node uses an ARM Cortex-A72 quad-core processor as the main control unit, with a main frequency of 1.5 GHz, paired with a neural network acceleration unit for parallel computing acceleration, and equipped with 4 gigabytes of dynamic random access memory and 32 gigabytes of flash memory for data storage. The airborne sensor group includes a binocular vision sensor, millimeter-wave radar, inertial measurement unit, global navigation satellite system receiver, and air-to-air communication module. The communication module operates in the 2.4 GHz frequency band, with a communication coverage radius of 500 meters and a communication cycle of 100 milliseconds. The airborne flight controller drives four rotor motors through pulse width modulation signals to achieve closed-loop control of flight attitude and thrust. All the above modules are connected to the airborne computing platform through the controller area network bus to realize bidirectional transmission of data and control commands.

[0114] Each aircraft node's onboard computing platform stores a computer program that, when executed, implements the functions of the following modules. Modules exchange data via shared memory, and the timing of calls between modules is uniformly scheduled by the real-time operating system, with a scheduling cycle consistent with the communication cycle of 100 milliseconds.

[0115] Perception and Field Establishment Module The perception and field construction module is responsible for collecting and processing the aircraft's state and environmental information, and constructing local spatiotemporal resource tension field slices.

[0116] This module acquires the aircraft's own three-dimensional spatial position coordinates, three-dimensional velocity vector, attitude angle, and current remaining battery power in the northeast-northeast coordinate system through an inertial measurement unit and a global navigation satellite system receiver, forming local state information. It acquires obstacle distribution data of the surrounding environment through a binocular vision sensor and millimeter-wave radar, combining this with wind vector field data obtained from an onboard meteorological sensor to form local environmental perception information. This module receives neighboring local state information broadcast by neighboring aircraft within its communication range via an air-to-air communication module; each neighbor's information includes its three-dimensional spatial position coordinates and three-dimensional velocity vector. These three types of information are sent to the input buffer of the onboard computing platform in the form of data frames, with a timestamp alignment accuracy of 1 millisecond.

[0117] The perception and field establishment module operates three parallel processing sub-units. The first sub-unit is responsible for the extrapolation calculation of uniform linear motion, performing the aforementioned operation p_i(t_k)=p_i(t_0)+v_i(t_0)·(t_k-t_0). The second sub-unit is responsible for the parallel calculation of the Gaussian kernel function of the density contribution value. This calculation is mapped to thousands of computational cores of the neural network acceleration unit for execution, with a spatial grid resolution of 5m×5m×5m and a time step of 61 steps. The third sub-unit is responsible for the point-by-point calculation of the constraint tension field, wind field perturbation tensor field, and energy potential field. The outputs of the three sub-units are weighted and summed to generate a local spatiotemporal resource tension field slice centered on the aircraft, which is stored in dynamic random access memory as a four-dimensional floating-point array for subsequent modules to call. From the second round onwards, this module stops calling the uniform extrapolation sub-unit and instead obtains the neighbor candidate predicted trajectories from the previous round from the trajectory generation and communication module as the input source for density calculation.

[0118] Trajectory Generation and Communication Module The trajectory generation and communication module is responsible for generating candidate predicted trajectories, managing the communication and transmission of trajectories, and maintaining prior confidence levels.

[0119] This module reads local spatiotemporal resource tension field slice data from the perception and field construction module. Using the current position and velocity as initial constraints and the mission waypoint position as terminal constraints, it generates five candidate prediction trajectories covering the next 30 seconds of the time domain using a B-spline curve parameterization method. Each trajectory is defined by eight control points, with adjacent control points spaced 4 seconds apart. This module calculates the total cost for each candidate trajectory. The total cost is a weighted sum of three factors: the cumulative sum of the tension field values ​​along the trajectory at each moment, the total trajectory length, and the deviation from the target waypoint's terminal position. After normalizing the total cost, it is converted into a priori confidence score, with the confidence score ranging from 0 to 1.

[0120] This module broadcasts candidate predicted trajectories via an air-to-air communication module. The broadcast data packet format includes the aircraft number, trajectory number, B-spline control point sequence, and corresponding prior confidence level. Simultaneously, this module receives broadcast data packets from neighboring aircraft within its communication range, parses out the neighboring candidate predicted trajectories, and stores them in a local trajectory buffer. This buffer is indexed by aircraft number, with each number storing trajectory data from the most recent five iterations, for use by the game adjustment module and the perception and field construction module.

[0121] Game Theory Adjustment Module The game adjustment module is responsible for performing iterative trajectory adjustments until a Nash equilibrium is reached.

[0122] This module obtains its own candidate predicted trajectory and neighbor candidate predicted trajectories from the trajectory generation and communication module, and calls a preset distributed game payoff function J_i to calculate the optimal response. The payoff function includes a conflict cost term calculated based on the spatiotemporal overlap between its own trajectory and those of its neighbors; an energy consumption cost term calculated based on the total trajectory length and wind disturbance; and a time cost term calculated based on the deviation between the actual arrival time and the planned arrival time at the target waypoint. In each iteration, the module keeps the neighbor trajectories constant and selects the trajectory that minimizes J_i as the optimal response for that round by traversing its own candidate trajectory set. The iteration convergence criterion is that the trajectory adjustment amplitude of all aircraft is less than the position change threshold of 0.5 meters for three consecutive iterations. When this criterion is met, the main predicted trajectory is output, and a set of suboptimal trajectories generated during the iteration process is stored in the candidate predicted trajectory library.

[0123] Execution and Deviation Processing Module The execution and deviation processing module is responsible for converting the master predicted trajectory into flight control commands and monitoring and processing prediction deviations.

[0124] This module reads the main predicted trajectory from the game-theoretic adjustment module and generates a set of flight control commands every 0.5 seconds, including the desired thrust value, desired roll angle, desired pitch angle, and desired yaw rate. These commands are then sent to the onboard flight controller via the controller area network bus. The flight controller then drives the four rotor motors accordingly.

[0125] This module simultaneously reads actual sensed field and predicted field data from the sensing and field establishment modules. Within each communication cycle, it calculates the deviation field ΔF(p,t) = F_actual(p,t) - F_predicted(p,t), resulting in a deviation field array of the same dimension as the tension field slice. This array is then sent to the deviation classification module for further processing. When the deviation classification module returns a sudden deviation determination result, this module immediately selects the trajectory with the smallest deviation from the current actual sensed field from the candidate predicted trajectory library to replace the currently executed trajectory; the switch is completed within 100 milliseconds. When a consistency deviation determination result is returned, this module waits for parameter correction instructions output by the group consensus module and adjusts the prediction model parameters accordingly.

[0126] Deviation Classification Module The deviation classification module is responsible for performing spatiotemporal two-dimensional statistical tests on prediction deviations and automatically distinguishing between consistency deviations and sudden deviations.

[0127] This module receives deviation field data from the execution and deviation processing module and maintains a sliding time window with a length of 3 seconds, or 7 snapshots. Spatially, this module counts the radius of the area covered by spatial grid nodes whose deviation values ​​exceed a preset deviation threshold of 0.15, determining whether this radius exceeds the aircraft's own neighborhood radius of 500 meters, and whether the number of neighboring aircraft within the threshold area is 3 or more. Temporally, this module counts the duration of deviation values ​​exceeding the threshold, determining whether it exceeds 2 seconds. If both conditions are met simultaneously, a consistency deviation judgment is output; otherwise, a sudden deviation judgment is output.

[0128] Upon identifying a consistency deviation, the module initiates a confirmation signal collection process, broadcasting confirmation requests to neighboring aircraft via the communication module to collect independent judgment results from each neighbor. When the number of confirmation signals reaches a preset consensus threshold, the prediction model parameters are adjusted. Upon identifying a sudden deviation, the module immediately generates a deviation sudden warning, including three fields: estimated location, radius of influence, and estimated duration, and broadcasts it to neighbors via the communication module.

[0129] Topology Construction and Critical Early Warning Module The topology construction and critical warning module is responsible for constructing a local collaborative topology graph with the aircraft as the root node and monitoring for signs of impending congestion.

[0130] This module reads the locations of high-tension regions with tension values ​​exceeding 0.70 from the local spatiotemporal resource tension field slice from the perception and field construction module, and reads information about neighboring aircraft within the communication range from the trajectory generation and communication module. The node set includes the aircraft itself, each neighboring aircraft, and the high-tension regions. Edge determination is based on whether the spatiotemporal proximity between two nodes exceeds a preset threshold of 0.15. The edge weight is determined by the weighted sum of the spatiotemporal proximity and the tension field value, with a harmonic coefficient of 0.50. This module stores the graph in the form of an adjacency matrix in dynamic random access memory.

[0131] In each iteration, this module calculates the conflict intensity of each edge in real time and performs a discrete second-order difference Δ²c(n) = c(n) - 2c(n-1) + c(n-2) on each edge connected to a neighboring aircraft. When the second-order difference value exceeds the preset critical warning threshold of 0.05, the module determines that there is a precursor to a sudden congestion and immediately notifies the trajectory database management module to activate a detour trajectory. At the same time, it notifies the communication module to broadcast an emergency reconstruction request to the neighbors.

[0132] Connectivity maintenance and edge density adjustment module The connectivity maintenance and edge density adjustment module is responsible for calculating algebraic connectivity and monitoring the risk of cluster network fragmentation.

[0133] This module reads the Laplacian matrix L of the local collaborative topology graph from the topology construction and critical warning module, calculates the eigenvalues ​​using the power iteration method, and takes the second smallest eigenvalue as the algebraic connectivity λ2. This module maintains a sliding time window of 5 iteration cycles, records the trajectory of λ2 within the window, and calculates its rate of decrease. When the rate of decrease exceeds 0.02 per iteration cycle, it is determined that there is a precursor to a cluster network split or abrupt change.

[0134] After determining the presence of a precursor to a break, this module calculates the eigenvector f corresponding to λ2. For each edge (i,j) in the graph, it calculates (f_i-f_j)², marking the edge with the largest value as a critical cooperative edge. This module generates an algebraic connectivity protection command, which is broadcast along with the emergency reconfiguration request via the communication module, prohibiting neighboring aircraft from removing this edge during reflexive reconfiguration.

[0135] This module also dynamically adjusts the edge density based on the current available bandwidth of the communication channel and the task coordination accuracy requirements. When the available bandwidth is less than 1 megabit per second, edges with a collision intensity of less than 0.30 are removed first. The task coordination accuracy requirements are automatically determined by the current task stage. In the dense formation stage, the accuracy requirements are high, so edges with high proximity to neighbors with high predicted trajectories are added first.

[0136] Risk balancing module The risk balancing module is responsible for constructing and minimizing the risk balancing cost function in concurrency risk scenarios.

[0137] This module receives the measured second-order differential value of conflict intensity from the topology construction and critical warning module, and the measured value of algebraic connectivity decrease rate from the connectivity maintenance and edge density adjustment module. When both exceed the critical warning threshold of 0.05 and the split warning threshold of 0.02 respectively, this module calculates the square of the over-limit amplitude as the conflict risk term and the split risk term. The weighting coefficients are determined by a continuous smooth mapping of the task progress percentage and local formation density using an S-curve, with the mapping formula S(x;a,b)=1 / (1+exp(-(xa) / b)), where the center point of the task progress percentage is a=50% and the slope is b=15%, and the center point of the local formation density is a=0.02 per meter and the slope is b=0.005 per meter.

[0138] This module uses gradient descent with a momentum term to minimize the cost function: R_total=μ·R_conflict+ν·R_disconnect; The decision variables are the edge removal threshold, the protection priority score, and the broadcast transmission rate. The iterative update rule is: x_{n+1}=x_n-η·∇R_total(x_n)+ρ·(x_n-x_{n-1}); The learning rate η is set to 0.15, and the momentum term coefficient ρ is set to 0.85. The optimization results are output to the topology construction and critical warning module and the connectivity maintenance and edge density adjustment module to guide subsequent reconstruction and protection operations.

[0139] Track Library Management Module The trajectory library management module is responsible for the initial construction and online evolution and maintenance of the candidate prediction trajectory library.

[0140] This module receives five candidate predicted trajectories and their prior confidence scores from the trajectory generation and communication module. Centered on the control point of each candidate trajectory, six sampling points are generated within a ring-shaped region with a radius of 5 to 15 meters, perpendicular to the plane of the tension field gradient direction. These points are then smoothly connected using cubic spline interpolation to form variant trajectories. The variant trajectories are merged with the original candidate trajectories, sorted by prior confidence scores, and the top K=10 trajectories are selected to form the initial candidate predicted trajectory library.

[0141] During the online evolution phase, this module reads the potential conflict intensity of each candidate trajectory from the topology construction and critical warning module. When the potential conflict intensity of a trajectory exceeds the threshold of 0.25 within 5 consecutive iterations, it is marked as invalid and removed. At the same time, based on the control point at the end of the current master predicted trajectory, supplementary trajectories are generated by sampling along the negative gradient direction of the tension field with a step size of 10 meters, and injected into the library to maintain a capacity of around 10 trajectories.

[0142] Group consensus module and adaptive consensus reconstruction module The group consensus module is responsible for initiating periodic group consensus processes, while the adaptive consensus reconstruction module is responsible for executing layered consensus and distributing topology reconstruction masks.

[0143] The group consensus module continuously monitors critical topological indicators, including the cluster local mean of conflict intensity and the cluster local variance of algebraic connectivity. When the mean exceeds 0.60 or the variance exceeds 0.15, the module initiates a group consensus request.

[0144] The adaptive consensus reconstruction module takes over the execution of the consensus process. In the first layer, the local pre-consensus phase, this module collects the prediction bias pattern feature vectors of each aircraft within its local cluster, performs K-means clustering (preset to 3 clusters), and outputs the local pre-consensus result when a certain cluster's members account for more than 60%, and elects the aircraft with the highest connectivity as the backbone. In the second layer, the diffusion alignment phase, the backbone aircraft broadcasts the local pre-consensus result to neighboring clusters. After receiving multiple pre-consensus results, it performs secondary alignment clustering to generate the global consensus result.

[0145] After the global consensus result is confirmed, a topology reconstruction mask is output. This mask is a list of edge operations, and each record includes the operation type (removal or addition), source node number, target node number, and execution round number. Each spacecraft synchronously executes edge operations at the specified round boundaries to ensure the temporal consistency of the topology switch. The global consensus result also outputs parameter correction values, including the weight coefficients of the distributed game payoff function and the evolution equation parameters of the local spatiotemporal resource tension field slice. Each spacecraft receives these values ​​and synchronously updates its local configuration.

[0146] All the above modules operate collaboratively under the unified scheduling of the real-time operating system. Within each communication cycle, i.e. 100 milliseconds, they sequentially complete the following tasks: perception acquisition, tension field construction, trajectory generation and broadcasting, game adjustment, deviation monitoring, topological criticality detection, consensus triggering and parameter correction, forming a complete perception-decision-communication-adjustment closed loop, supporting the dynamic scheduling and collaborative flight of low-altitude aircraft clusters.

[0147] The above embodiments use a quadcopter drone swarm in an urban low-altitude logistics delivery scenario as an example to describe in detail the dynamic scheduling method and system for low-altitude aircraft swarms based on multi-source data fusion provided by this invention. This application scenario is merely illustrative and does not constitute any limitation on the scope of application of this invention. Those skilled in the art should understand that the technical solutions described in this invention are also applicable to other types of aircraft swarms, including but not limited to fixed-wing drone swarms, tiltrotor aircraft swarms, compound-wing drone swarms, and collaborative scenarios involving mixed manned and unmanned aircraft formations. In different types of aircraft swarms, the dynamic parameters, communication performance indicators, sensor configurations, and mission characteristic parameters of each aircraft can be adaptively adjusted according to the specific aircraft model and operating environment; such adjustments do not depart from the basic framework of the technical solutions of this invention.

[0148] The application scope of this invention is not limited to the field of low-altitude logistics and distribution. In application scenarios such as urban air traffic management, emergency rescue collaborative search, agricultural plant protection drone swarm operations, power line inspection cluster collaboration, and border patrol surveillance, when multiple aircraft are cooperating in the same airspace and there is a risk of dynamic conflict, the method described in this invention can be used to achieve distributed and adaptive trajectory scheduling. For different application scenarios, the constraint tension field components in the local spatiotemporal resource tension field slice can be customized according to the obstacle type in the specific scenario. For example, in the power line inspection scenario, transmission towers and high-voltage line corridors can be used as no-fly zones; in the emergency rescue scenario, temporarily designated casualty transfer channels can be treated as priority passage areas.

[0149] In the specific parameter configurations of the above embodiments, the given values ​​include, but are not limited to, a spatial radius of 500 meters, a prediction time domain of 30 seconds, a time step of 0.5 seconds, a Gaussian kernel bandwidth of 20 meters, weight coefficients w_d=0.35, w_c=0.30, w_w=0.20, w_e=0.15, a second-order differential threshold of 0.05, an algebraic connectivity decrease threshold of 0.02, a consensus threshold of 50%, a group consensus threshold of 60%, a candidate trajectory library capacity of K=10, a learning rate of η=0.15, and a momentum term coefficient of ρ=0.85, etc. These are all preferred configurations under specific implementation conditions, intended to help those skilled in the art better understand and implement the present invention. The actual selection of the above numerical parameters should not be construed as a limitation of the present invention. In specific engineering practice, those skilled in the art can determine the parameter configurations suitable for their respective application scenarios through a limited number of simulation experiments or on-site debugging, based on factors such as the dynamic characteristics of the aircraft, communication bandwidth and latency constraints, the processing power of the airborne computing platform, the spatial scale of the mission environment, and the cluster size, without requiring creative effort.

[0150] The functional modules described in this invention include a perception and field establishment module, a trajectory generation and communication module, a game theory adjustment module, an execution and deviation processing module, a deviation classification module, a topology construction and critical early warning module, a connectivity maintenance and edge density adjustment module, a risk balancing module, a trajectory library management module, a group consensus module, and an adaptive consensus and topology reconstruction module. In specific implementations, these modules can be deployed on the onboard computing platform of each aircraft to form a fully distributed architecture, or they can be deployed heterogeneously based on the differences in computing power among the aircraft in the cluster. This involves offloading the tasks of some computationally intensive modules to backbone aircraft or ground edge computing nodes with stronger computing power. Regardless of the deployment method, the data interaction logic and core algorithm flow between the modules remain consistent with the above embodiments. Those skilled in the art can flexibly choose the deployment scheme for the above modules based on the actual system's computing power distribution and communication topology characteristics.

[0151] Through the above description of specific embodiments, the technical solution of the low-altitude aircraft swarm dynamic scheduling method and system provided by the present invention has been fully and clearly explained. Those skilled in the art can implement all the technical solutions of the present invention based on the above description without excessive experimentation, and can make several adjustments and modifications without departing from the concept of the present invention. Therefore, the above embodiments should be understood as illustrative disclosures of the present invention, rather than limitations on the scope of protection of the present invention. The scope of protection of the present invention should be determined by the appended claims.

Claims

1. A dynamic scheduling method for low-altitude aircraft clusters based on multi-source data fusion, characterized in that, The method includes: During the initialization round, each aircraft acquires its own local state information, local environmental perception information, and the local state information of its neighbors; it extrapolates its motion state based on its own and its neighbors' local state information, and constructs an initial local spatiotemporal resource tension field slice by combining the local environmental perception information; the slice is used to characterize the resource occupancy tension of the space around the aircraft at different future times. Each aircraft generates at least two candidate predicted trajectories based on the slice, and assigns a prior confidence level to each candidate predicted trajectory; Each aircraft broadcasts its candidate predicted trajectory and receives candidate predicted trajectories from its neighbors; From the second round onwards, when each aircraft constructs the slice, it uses the neighbor candidate predicted trajectory received in the previous round to replace the motion state extrapolation, updates the density tension field components, and forms an updated slice. Each aircraft iteratively adjusts its trajectory based on its own and its neighbors' candidate predicted trajectories and a preset distributed game payoff function until it converges to a Nash equilibrium state, thus obtaining the main predicted trajectory and a library of alternative predicted trajectories. Each aircraft flies along the main predicted trajectory and continuously monitors the prediction deviation between the actual perceived field and the predicted field, wherein the actual perceived field is determined by local environmental perception information and the predicted field is determined by the slice. Each aircraft performs feedback iterations based on the type of prediction deviation: when it is a consistency deviation, the prediction model parameters are adjusted; when it is a sudden deviation, the target candidate trajectory is switched from the alternative prediction trajectory library. The group consensus process is periodically triggered to collect prediction deviation patterns of multiple aircraft for cluster analysis. When a common deviation pattern is detected, the weight parameters of the distributed game payoff function and / or the evolution equation parameters of the slice are corrected, and the corrected parameters are distributed and diffused in the cluster.

2. The method according to claim 1, characterized in that, The construction of local spatiotemporal resource tension field slices includes: Based on its own spatial coordinates and velocity vector, as well as the spatial coordinates and velocity vectors of its neighbors, the predicted clustering density of each spatial location within a preset spatial radius in a future preset time domain is calculated to obtain the density tension field. Based on the coordinates of the no-fly zone boundary, the coordinates of the temporary control area, and the coordinates of the severe weather area, the hard constraint barrier value of each spatial location within the preset spatial radius is calculated to obtain the constraint tension field; Based on the wind vector field data, the wind field disturbance tensor at each spatial location within the preset spatial radius at different times is calculated to obtain the wind field disturbance tensor field. Based on the current remaining power and the estimated energy consumption required to reach each spatial location, calculate the energy margin value of each spatial location within the preset spatial radius to obtain the energy potential field. The density tension field, the constraint tension field, the wind field disturbance tensor field, and the energy potential field are superimposed point by point according to preset weights to generate a local spatiotemporal resource tension field slice.

3. The method according to claim 2, characterized in that, The calculation of the density tension field includes: During the initialization round, the system extrapolates at a constant speed based on its own and its neighbors' spatial position coordinates and velocity vectors to obtain the predicted spatial position coordinates of each in the future preset time domain. From the second round onwards, the received neighbor candidate predicted trajectory is used instead of the uniform motion extrapolation as the source of predicted spatial location coordinates. Based on the predicted spatial position coordinates of each aircraft, the density contribution value of each predicted spatial position to each spatial position within the preset spatial radius is calculated, and the density contribution value is inversely proportional to the distance. The density contribution values ​​of all the same spatial locations at the same time are summed to generate the cluster density prediction value. The density tension field is generated by arranging the predicted cluster density values ​​of all spatial locations at all future times in spatial and temporal order.

4. The method according to claim 1, characterized in that, Also includes: Based on neighboring aircraft within the communication range and high-tension regions in the slice whose tension values ​​exceed a preset threshold, a local collaborative topology graph is constructed with the aircraft as the root node. The nodes in the topology graph include the aircraft, each neighboring aircraft, and each high-tension region. Edges represent trajectory conflicts or resource competition relationships, and edge weights are determined by spatiotemporal proximity and slice superposition values. During the iterative trajectory adjustment process, the conflict intensity of each edge in the topology graph is calculated in real time, and the second-order derivative of each neighboring edge is performed. When the second derivative value exceeds the preset critical warning threshold, it is determined that there is a sign of impending congestion. An alternative trajectory is activated from the alternative prediction trajectory library to replace the current main prediction trajectory, and an emergency reconstruction request is broadcast. When a common deviation pattern is detected during the group consensus process, each aircraft performs a reflexive reconstruction of its local cooperative topology graph according to the common deviation pattern: removing edges in unreliable directions, adding new edges with potential cooperative dependencies, and using the reconstructed topology graph for trajectory adjustment game in the next cycle.

5. The method according to claim 4, characterized in that, When the second derivative value exceeds the preset critical warning threshold, the following further applies: Calculate the algebraic connectivity of the local collaborative topology graph and monitor its rate of change. If the rate of decrease exceeds the preset split warning threshold, it is determined that there is a precursor to a split change in the cluster network. When broadcasting the emergency reconfiguration request, an algebraic connectivity protection instruction is attached, instructing neighboring aircraft to prohibit the removal of critical cooperative edges during reflexive reconfiguration. The critical cooperative edge is the edge that contributes the most to maintaining algebraic connectivity. In the reflexive reconstruction, the edge density is dynamically adjusted according to the available bandwidth of the current communication channel and the task coordination accuracy requirements: when the available bandwidth is lower than the preset bandwidth threshold, edges with conflict intensity lower than the preset conflict threshold are removed first; when the task coordination accuracy requirements are higher than the preset accuracy threshold, edges with high proximity neighbors are added first.

6. The method according to claim 5, characterized in that, When both the congestion precursor and the cluster network disruption precursor are detected simultaneously, the method further includes: Construct a risk balance cost function consisting of a weighted sum of conflict risk terms and separation risk terms. The conflict risk terms are determined based on the overlimit magnitude of the second-order differential value, and the separation risk terms are determined based on the overlimit magnitude of the algebraic connectivity decrease rate. By minimizing the risk balance cost function, the edge removal threshold, the priority of key collaborative edge protection, and the emergency reconstruction request sending rate are dynamically determined. The weights of the conflict risk item and the fragmentation risk item are dynamically adjusted according to the current task stage and task priority: the weight of the conflict risk item is increased during the assembly or dense formation stage, and the weight of the fragmentation risk item is increased during the wide-area search or sparse distribution stage.

7. The method according to claim 4, characterized in that, The generation and maintenance of the alternative predicted trajectory library includes: Initial construction phase: Guided by the gradient direction of the slice, sample and smoothly connect in the space near the candidate predicted trajectory to generate variant trajectories. After merging with the candidate predicted trajectories, select the top K trajectories in descending order of prior confidence to form a candidate predicted trajectory library. Online evolution phase: Monitor the potential conflict intensity of each candidate trajectory. When the conflict intensity exceeds the preset conflict tolerance threshold for N consecutive iterations, the trajectory is removed. At the same time, a supplementary trajectory is generated by resampling along the low-tension direction of the slice and injected into the library to maintain the library capacity.

8. The method according to claim 1, characterized in that, The step of performing a feedback iterative update process based on the prediction deviation type includes: To distinguish between consistency deviation and sudden deviation, a spatiotemporal two-dimensional statistical test is performed on the prediction deviation: a sliding time window is maintained, and the duration and spatial influence range of the prediction deviation within the window are recorded. When the duration exceeds a preset duration threshold and the spatial influence range exceeds its own neighborhood and covers at least M neighbors, it is determined to be a consistency deviation; otherwise, it is a sudden deviation. After determining that a consistency deviation has occurred, the confirmation signals of neighboring aircraft affected by the deviation for the same deviation pattern are collected. When the number of confirmation signals reaches a preset consensus threshold, the prediction model parameters are adjusted. Once a sudden deviation is identified, an alternative trajectory switch is immediately executed, and a sudden deviation warning is broadcast. The warning includes the estimated location, the radius of influence, and the estimated duration.

9. The method according to claim 4, characterized in that, The periodic triggering of the group consensus process includes: Each aircraft continuously monitors the topological critical indicators of the local collaborative topology graph, including the cluster local mean of conflict intensity and the cluster local variance of algebraic connectivity. When any aircraft detects that its topological critical indicators meet the preset adaptive triggering conditions, it initiates a group consensus request. A hierarchical consensus mechanism is adopted: the initiating aircraft and its single-hop neighbors form a local group, exchange prediction bias patterns within the group and perform local clustering to form a local pre-consensus result. After the consensus rate reaches the preset group consensus threshold, the backbone aircraft is elected to spread the local pre-consensus result to the adjacent groups. Each backbone aircraft performs secondary alignment and clustering on the multiple local pre-consensus results received to generate a global consensus result and spread it to the cluster in a distributed manner. After the global consensus result is confirmed, a topology reconstruction mask is output, specifying the set of edges that should be removed and added by each aircraft in the reflexive reconstruction. Each aircraft synchronously performs reflexive reconstruction according to the received topology reconstruction mask.

10. A dynamic scheduling system for low-altitude aircraft clusters based on multi-source data fusion, comprising multiple aircraft nodes, each of which is configured with: The perception and field construction module is used to acquire the status and environmental information of the aircraft, and to construct and update the local spatiotemporal resource tension field slice; The trajectory generation and communication module is used to generate candidate predicted trajectories and assign prior confidence levels, as well as broadcast and receive candidate predicted trajectories. The game adjustment module is used to iteratively adjust the trajectory based on its own and its neighbors' candidate predicted trajectories and the preset distributed game payoff function. After convergence, it outputs the main predicted trajectory and the library of alternative predicted trajectories. The execution and deviation processing module is used to execute flight with the main predicted trajectory, monitor prediction deviations, and adjust prediction model parameters or switch alternative trajectories according to the type of deviation. The group consensus module is used to periodically trigger the group consensus process, collect prediction deviation patterns for cluster analysis, and correct parameters and distribute them when common deviation patterns are detected. The topology construction and critical warning module is used to construct a local cooperative topology graph, calculate the conflict intensity and its second derivative, activate the detour trajectory and broadcast an emergency reconstruction request when the warning threshold is exceeded. The connectivity maintenance and edge density adjustment module is used to calculate algebraic connectivity and monitor its rate of change, attach protection instructions when determining the risk of disconnection, and dynamically adjust the edge density according to communication bandwidth and task accuracy requirements. The risk balancing module is used to construct and minimize the risk balancing cost function in concurrency risk scenarios, and dynamically determine the edge removal threshold, protection priority and broadcast sending rate. The trajectory library management module is used for the initial construction and online evolution of the candidate prediction trajectory library. It selects the top K trajectories according to confidence level and dynamically removes failed trajectories and adds new trajectories to maintain the library capacity. The deviation classification module is used to perform spatiotemporal two-dimensional statistical tests on prediction deviations to distinguish the types of deviations, and trigger corresponding parameter adjustments or trajectory switching and early warning broadcasts after classification. The adaptive consensus and topology reconstruction module is used to monitor critical topology indicators, initiate a hierarchical consensus process when the triggering conditions are met, and output a topology reconstruction mask based on the consensus results to drive each aircraft to synchronously perform reflexive reconstruction.

Images