Adaptive scheduling method for communication unmanned aerial vehicle group

Abstract

The invention discloses a self-adaptive scheduling method for a communication unmanned aerial vehicle group, and relates to the field of unmanned aerial vehicle control, and the method comprises the steps: collecting water surface dynamic data in real time through a multi-mode sensing module, and constructing a water surface dynamic digital twinborn model; based on output data of the water surface dynamic digital twinning model, generating an electromagnetic environment simulation map by using a time-space map convolutional network; identifying the edge of a strong interference area in the electromagnetic environment simulation map, deploying an expandable intelligent reflector array, and dynamically optimizing the phase configuration of an intelligent reflector unit through a deep reinforcement learning algorithm; dividing the unmanned aerial vehicle group into a communication subgroup and a monitoring subgroup according to the spatial distribution of the strong interference area; the monitoring subgroup scans the strong interference area, obtains local fluctuation parameters, and feeds back data to the digital twinborn model in real time; and performing formation optimization on the communication subgroups by adopting a distributed model predictive control framework, solving a multi-target optimization problem, and generating an unmanned aerial vehicle position adjustment instruction.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle (UAV) control, and in particular to an adaptive scheduling method for a communication UAV swarm. Background Art

[0002] Current communication drone swarm scheduling technology primarily relies on radio communication networks (such as self-organizing networks and software-defined networks) to enable collaborative operations between drones, combined with swarm intelligence algorithms (such as ant colony algorithms and particle swarm optimization) for path planning and task allocation. Regarding environmental perception, traditional methods often use single sensors (such as radar or lidar) to collect data and construct static three-dimensional models using digital twin technology to support decision-making. For example, the Gaoyou Irrigation District uses drone oblique photography to construct digital twin models of water conservancy projects. Anti-interference technologies primarily utilize spectrum switching or low-power communication strategies. For example, the "firefly"-like optical communication drone proposed by Northwestern Polytechnical University utilizes the electromagnetic interference resistance of optical signals to enable short-range information transmission. Furthermore, existing drone formation optimization methods are mostly based on genetic algorithms or centralized control, and energy management utilizes a fixed energy consumption allocation model.

[0003] In this case, the existing communication drone swarm scheduling technology has at least the following problems: (1) Insufficient communication robustness. Radio communication is susceptible to electromagnetic interference, resulting in a high risk of link interruption. Although optical communication is anti-interference, it is limited by transmission distance and ambient light, making it difficult to cover a large area. The digital twin model has poor dynamics. Traditional modeling methods rely on offline data updates and cannot synchronize physical environment changes (such as water surface fluctuations and turbulent disturbances) in real time, resulting in delayed simulation results and an inability to support dynamic scheduling decisions. (3) The drone swarm has fixed functional divisions (such as dedicated communication or monitoring drones), and the sensor network density cannot be dynamically adjusted according to mission requirements, resulting in insufficient data collection accuracy in local areas or waste of resources. (4) Global optimization methods such as genetic algorithms take a long time to calculate and are difficult to meet the real-time adjustment needs of drone formations in dynamic environments. Summary of the Invention

[0004] The purpose of the present invention is to provide an adaptive scheduling method for a communication drone swarm, which aims to overcome the above-mentioned problems existing in the prior art.

[0005] To achieve the purpose, the present invention provides the following technical solutions: An adaptive scheduling method for a communication drone swarm comprises the following steps: Step (1) collects water surface dynamic data in real time through the multimodal perception module of the drone swarm to build a water surface dynamic digital twin model; Step (2): Based on the output data of the water surface dynamic digital twin model, a time-space graph convolutional network is used to extract the spatial correlation characteristics and time evolution laws of wave propagation, generate an electromagnetic environment simulation map including signal reflection path, interference intensity distribution and multipath fading characteristics, and mark the signal interference level at each spatial position; Step (3), identifying areas in the electromagnetic environment simulation map where the signal interference intensity exceeds a first threshold, and marking them as strong interference areas; deploying a deployable smart reflector array at the edge of the strong interference area, and dynamically optimizing the phase configuration of the smart reflector unit through a deep reinforcement learning algorithm to form a beam null for suppressing reflection interference; Step (4) dynamically divides the drone swarm into a communication subgroup and a monitoring subgroup based on the spatial distribution of the strong interference area; the monitoring subgroup uses a topology reconstruction algorithm to form a high-density sensor array, performs adaptive focus scanning on the strong interference area, obtains local fluctuation parameters with sub-meter accuracy, and feeds data back to the digital twin model in real time for incremental learning; Step (5) uses a distributed model predictive control framework to optimize the communication subgroup formation, solve the multi-objective optimization problem, and generate UAV position adjustment instructions; wherein the optimization objectives include communication quality indicators, energy consumption balance and topology stability; Step (6) dynamically configures the communication parameters of the UAV based on the spectrum situation knowledge graph. When it is detected that the residual interference exceeds the second threshold, the federated learning mechanism is triggered to collaboratively decide the optimal frequency band switching solution and synchronously update the phase distribution of the intelligent reflective surface.

[0006] Furthermore, in step (1), the multimodal perception module includes a millimeter-wave radar array, a distributed laser ranging unit, a surface acoustic wave sensor network, and a polarization sensor; wherein the millimeter-wave radar array measures the slope change rate and three-dimensional motion vector of the water surface wave crest, the laser ranging unit detects the water surface spectrum characteristics through coherent interference, the surface acoustic wave sensor network captures the air turbulence velocity field at a height of 10 cm near the water surface, and the polarization sensor inverts the dynamic change of the complex refractive index of the water surface medium; and the above-mentioned multi-source heterogeneous data are integrated to construct a dynamic digital twin model of the water surface.

[0007] Furthermore, the dynamic digital twin model of the water surface includes a dynamic mapping layer, an evolution prediction layer and a physical-signal coupling layer; among them, the dynamic mapping layer adopts a spatial topology matching algorithm to align the above-mentioned multi-source heterogeneous data with the three-dimensional water surface grid model in real time; the evolution prediction layer simulates the dynamic evolution of waves based on the finite element method and computational fluid dynamics, combines Kalman filtering for data fusion and anomaly correction, and predicts the water surface fluctuation trend; the physical-signal coupling layer embeds an electromagnetic wave propagation model to quantify the interference of changes in water surface inclination on the signal reflection path, and generate a dynamic electromagnetic environment map.

[0008] Furthermore, the step (3) specifically includes: Step (301): using a deployable smart reflective surface mounted on a drone, a ring array is formed around the strong interference area, the spacing between adjacent smart reflective surface units satisfies d<λ / 2, where λ is the signal wavelength, and the phase resolution is not less than 4 bits; Step (302): Establish a mapping relationship model between the phase distribution of the intelligent reflector unit and the interference source orientation:

[0009] in, is the phase of the nth smart reflector unit, is the direct channel response, is the reflection channel response of the smart reflection surface; Step (303) is based on a double-delayed deep deterministic policy gradient algorithm, and the phase distribution is updated according to the real-time electromagnetic environment feedback, and the spacing between adjacent smart reflective surface units satisfies d<λ / 2, and the phase resolution is not less than 4 bits; where λ is the signal wavelength.

[0010] Going further, the direct channel response is calculated as: ;in, is the direct path attenuation coefficient (including path loss and shadow fading), is the direct path delay, is the specific frequency value of the kth frequency sampling point; The reflection channel response is calculated as: ;in, is the reflection coefficient of the mth smart reflector unit, is the initial phase offset of the mth smart reflector unit, is the propagation delay from the transmitter to the mth smart reflector unit, represents the equivalent delay adjustment introduced by the phase configuration θ.

[0011] Furthermore, the objective function of the multi-objective optimization problem in step (5) is

[0012] Where X represents the coefficient of the drone group position matrix; α, β, γ are multi-objective weights, and α + β + γ = 1; QoS represents the quality of communication service, Energy represents the energy consumption balance of the group, and Topo_Stab represents the formation deformation index; Constraints include: Drone spacing ;in, , Represents the three-dimensional coordinates of the i / j-th UAV; Indicates the minimum safe distance; Ground user receiving signal strength ;in, represents the received signal strength of the kth ground user, Indicates the communication assurance threshold; Maximum battery charge difference ;in, Indicates the maximum battery power difference between drones in a group.

[0013] Furthermore, the communication service quality is calculated as follows: ; Where SINR is the signal-to-interference-plus-noise ratio, Throughput is the effective throughput, ω1, ω2 are weight coefficients, and ω1+ω2=1; Calculation method of group energy consumption balance: ;in, is the remaining battery percentage of the i-th drone, The average remaining power of the group; The calculation method of formation deformation index is: ;in, is the matrix Frobenius norm, The time interval between adjacent optimization cycles.

[0014] Furthermore, in step (6), the spectrum situation knowledge graph is constructed through a graph attention network.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This solution leverages core technologies such as multi-physics field perception, intelligent reflective surface collaboration, and dynamic evolution of digital twins to address the communication vulnerabilities, model lags, and resource rigidity inherent in traditional drone swarm scheduling. Field tests have demonstrated 98% communication availability in force 5 wind and wave conditions, significantly improving formation optimization efficiency. This provides a highly robust and adaptive scheduling solution for scenarios such as environmental reconnaissance and disaster monitoring.

[0016] First, this invention integrates data from multiple sources, including millimeter-wave radar, laser ranging, and surface acoustic wave sensors, to construct a high-precision dynamic digital twin model of the water surface. This model updates three-dimensional deformation characteristics and electromagnetic interference predictions in real time, integrating multimodal perception with the dynamic digital twin. Compared to traditional static modeling, this method reduces data update latency to seconds, improves model accuracy to sub-meter levels, and supports the dynamic generation of electromagnetic environment simulation maps.

[0017] Secondly, the present invention combines the active beamforming of the intelligent reflecting surface (RIS) with optical communication technology to build a dual-mode anti-interference communication architecture, forming an "active elimination + passive avoidance" mechanism, which can effectively suppress the reflection interference of the water surface.

[0018] Third, in this invention, a drone swarm is divided into communication and monitoring subgroups as needed. The monitoring subgroup forms a high-density sensor array through topological reconfiguration, improving local data acquisition accuracy. By employing a distributed model predictive control (DMPC) framework and a parallel particle swarm algorithm, formation optimization response time is reduced to milliseconds, while simultaneously achieving multi-objective optimization for communication quality, energy consumption balance, and topological stability. This demonstrates that this invention enables dynamic resource reconfiguration and distributed optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Flowchart of the present invention. DETAILED DESCRIPTION

[0020] The specific embodiments of the present invention are described below with reference to the accompanying drawings. In order to fully understand the present invention, many details are described below, but for those skilled in the art, the present invention can be implemented without these details.

[0021] like Figure 1 As shown, an adaptive scheduling method for a communication drone group includes the following steps: Step (1) collects water surface dynamic data in real time through the multimodal perception module of the drone swarm to build a water surface dynamic digital twin model.

[0022] Specifically, the multimodal perception module includes a millimeter-wave radar array, a distributed laser ranging unit, a surface acoustic wave sensor network and a polarization light sensor.

[0023] A millimeter-wave radar array measures the slope change rate and three-dimensional motion vector of surface wave crests to quantify the three-dimensional deformation characteristics of the waves. A laser ranging unit detects the spectral characteristics of the water surface through coherent interferometry, identifying the frequency distribution and energy density of the waves. A surface acoustic wave sensor network captures the turbulent air velocity field at a height of 10 cm near the water surface and analyzes the impact of turbulent disturbances on signal propagation. A polarization sensor inverts the dynamic changes in the complex refractive index of the water surface medium, reflecting the water quality and medium characteristics. Finally, the above multi-source heterogeneous data is integrated to construct a dynamic digital twin model of the water surface.

[0024] The water surface dynamic digital twin model includes a dynamic mapping layer, an evolution prediction layer, and a physical-signal coupling layer.

[0025] Dynamic mapping layer: Using a spatial topology matching algorithm, the above multi-source heterogeneous data are aligned with the three-dimensional water surface grid model in real time, achieving millimeter-level precision mapping from the physical water surface to the digital space.

[0026] Evolution prediction layer: Based on the finite element method (FEM) and computational fluid dynamics (CFD) to simulate the dynamic evolution of waves, and combined with the Kalman filter for data fusion and anomaly correction, it predicts the water surface fluctuation trend in the next few minutes (such as five minutes).

[0027] Physics-signal coupling layer: Embeds electromagnetic wave propagation models (such as ray tracing) to quantify the interference of water surface inclination changes on signal reflection paths and generate dynamic electromagnetic environment maps.

[0028] Step (2): Based on the output data of the water surface dynamic digital twin model, the space-time graph convolutional network (ST-GCN) is used to extract the spatial correlation characteristics and temporal evolution laws of wave propagation, generate an electromagnetic environment simulation map including signal reflection path, interference intensity distribution and multipath fading characteristics, and mark the signal interference level at each spatial location.

[0029] The spatiotemporal graph convolutional network (ST-GCN) generates an electromagnetic environment simulation map through the following steps: Step 1: Based on the fluctuation data of the water surface grid nodes, spatial correlation features are extracted through graph convolution (GCN).

[0030] Specifically, taking the three-dimensional grid node data (such as wave crest phase and turbulent velocity) output by the water surface dynamic digital twin model as input, the graph convolutional layer (GCN) models the water surface grid as a graph structure, where the nodes represent spatial positions and the edge weights reflect the correlation between fluctuations in adjacent regions, and outputs a spatial correlation matrix (such as the inter-regional fluctuation synchronization index).

[0031] Step 2: Based on historical time series data, extract the time series rules of wave evolution through the long short-term memory network (LSTM).

[0032] Specifically, taking historical time series data (such as the evolution of the wave spectrum in the past 10 seconds) as input, the long short-term memory network (LSTM) extracts the periodicity and mutation in the historical time series data and outputs a time evolution feature vector (such as the fluctuation forecast for the next 5 seconds).

[0033] Step 3: Integrate spatial correlation features, temporal patterns, and electromagnetic parameters, calculate the signal reflection path, interference intensity distribution, and multipath fading characteristics through the fully connected layer, and generate an electromagnetic environment simulation map.

[0034] Step (3) identifies areas in the electromagnetic environment simulation map where the signal interference intensity exceeds a first threshold (e.g., -90 dBm) and marks them as strong interference areas; deploys a deployable smart reflector (RIS) array at the edge of the strong interference area, and dynamically optimizes the phase configuration of the smart reflector units through a deep reinforcement learning algorithm to form a beam null for suppressing reflection interference.

[0035] Specifically include: Step (301): A ring array is formed around the strong interference area by using deployable smart reflective surface units mounted on the drone, forming a smart reflective surface array. The spacing between adjacent smart reflective surface units satisfies d < λ / 2, where λ is the signal wavelength, and the phase resolution is not less than 4 bits.

[0036] Step (302): Establish a mapping relationship model between the phase distribution of the intelligent reflector unit and the interference source orientation: .

[0037] in, is the phase of the nth smart reflector unit; is the direct channel response, calculated as: ;in, is the direct path attenuation coefficient (including path loss and shadow fading), is the direct path delay, is the specific frequency value of the kth frequency sampling point; is the reflection channel response of the smart reflector, calculated as: ;in, is the reflection coefficient of the mth smart reflector unit, is the initial phase offset of the mth smart reflector unit, is the propagation delay from the transmitter to the mth smart reflector unit, represents the equivalent delay adjustment introduced by the phase configuration θ.

[0038] Step (303): Based on a double-delayed deep deterministic policy gradient algorithm (TD3), the phase distribution is updated according to real-time electromagnetic environment feedback, and the spacing between adjacent smart reflector units satisfies d < λ / 2, and the phase resolution is not less than 4 bits; where λ is the signal wavelength. The double-delayed deep deterministic policy gradient (TD3) algorithm is used to update the phase distribution of the smart reflector units in real time.

[0039] Step (4) dynamically divides the drone swarm into a communication subgroup and a monitoring subgroup based on the spatial distribution of the strong interference area. The monitoring subgroup uses a topology reconstruction algorithm to form a high-density sensor array, performs adaptive focus scanning on the strong interference area, obtains local fluctuation parameters with sub-meter accuracy, and feeds the data back to the digital twin model in real time for incremental learning.

[0040] Step (5) uses a distributed model predictive control framework to optimize the communication subgroup formation, solve the multi-objective optimization problem, and generate UAV position adjustment instructions. The optimization objectives include communication quality indicators, energy consumption balance, and topological stability.

[0041] Specifically, the objective function of the multi-objective optimization problem in step (5) is:

[0042] Among them, X represents the coefficient of the drone group position matrix, X∈R N×3 are the three-dimensional coordinates of N drones.

[0043] α, β, and γ are multi-objective weights, where α + β + γ = 1. By default, α = 0.5, β = 0.3, and γ = 0.2. Of course, α, β, and γ can be adjusted dynamically based on task requirements.

[0044] QoS stands for Quality of Service (QoS) and is calculated as follows: ; Among them, SINR is the signal-to-interference-plus-noise ratio, Throughput is the effective throughput, ω1 and ω2 are weight coefficients, and ω1+ω2=1. The default values are ω1=0.6 and ω2=0.4.

[0045] Energy represents the energy consumption balance of the group, which is calculated as follows: ;in, is the remaining battery percentage of the i-th drone, The average remaining power of the group.

[0046] Topo_Stab represents the formation deformation index, calculated as follows: ;in, is the matrix Frobenius norm, The time interval between adjacent optimization cycles.

[0047] The constraints of the multi-objective optimization problem include the following three.

[0048] Constraint 1 is used to prevent collisions between drones: ;in, , , represents the three-dimensional coordinates of the i / j-th UAV; : Indicates the minimum safety distance.

[0049] Constraint 2 is used to ensure user service quality: ;in, represents the received signal strength of the kth ground user, Indicates the communication guarantee threshold, such as -90dBm, Constraint three is used to prevent individual drones from running out of power prematurely: ;in, Indicates the maximum battery power difference between drones in a group.

[0050] Step (6) dynamically configures the communication parameters of the UAV based on the spectrum situation knowledge graph. When the residual interference is detected to exceed the second threshold (such as -80dBm), the federated learning mechanism is triggered to collaboratively decide the optimal frequency band switching solution and synchronously update the phase distribution of the intelligent reflector.

[0051] Specifically, the spectrum situation knowledge graph is constructed through the Graph Attention Network (GAT), which includes: Historical interference signature database, used to store <frequency, polarization mode, interference pattern> triples; Dynamic association rule base, used to establish the mapping relationship between interference characteristics and <available frequency band, modulation and coding scheme>; The federated learning framework aggregates the local spectrum decision models of multiple drone swarms to generate the globally optimal frequency band switching strategy.

[0052] Step (7): When a UAV in the monitoring subgroup fails, execute the following: (a) redivide the sensor array responsibility area based on the Voronoi diagram; (b) start the neighboring UAV sensor redundancy mode, and increase the sampling frequency to 1.5 times the baseline value; (c) limit the formation optimization search space radius to 60% of the initial value to ensure real-time response speed.

[0053] Experimental test verification 1. Communication Availability Comparison Data

[0054] 2. Comparative Data on Formation Optimization Efficiency

[0055] 3. Experimental Design and Statistical Verification 1. Variable control (1) Fixed environmental parameters: wind speed 14 m / s (level 5 wind and wave), water temperature 20 ± 2 °C, and humidity 60 ± 5%.

[0056] (2) The control group used the same hardware platform (DJI Matrice 300 RTK drone, H20T sensor payload).

[0057] (3) Eliminate interference from traffic fluctuations: Only analyze the communication data of the same user group (reconnaissance task group).

[0058] 2. Significance test (1) Using two-sample T test (α=0.05): Communication availability difference p = 1.2 × 10⁻ 9 (much less than 0.05) Optimized response time difference p=3.7×10⁻¹² (2) Power analysis (Power = 0.8): The minimum sample size required was 512 groups, and the actual sample size of 864 groups met the requirement.

[0059] 3. Robustness Verification (1) Extreme condition test: In the scenario of sudden gust of wind (instantaneous wind speed 18m / s), the communication availability still remains at 96.2%.

[0060] (2) Fault injection test: After randomly removing 20% of the drone nodes, the success rate of formation optimization only dropped by 4.7% (the traditional solution dropped by 31%). 4. Comparison Benchmarks with Industry Standards

[0061] The above data show that the present invention has achieved a breakthrough improvement in communication reliability and formation efficiency in complex water environments through innovative designs such as multimodal perception fusion and intelligent reflective surface collaboration.

[0062] The above is only a specific implementation of the present invention, but the design concept of the present invention is not limited to this. Any non-substantial changes to the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. An adaptive scheduling method for a communication drone swarm, characterized by: The following steps are involved: Step (1) collects water surface dynamic data in real time through the multimodal perception module of the drone swarm to build a water surface dynamic digital twin model; Step (2): Based on the output data of the water surface dynamic digital twin model, a time-space graph convolutional network is used to extract the spatial correlation characteristics and time evolution laws of wave propagation, generate an electromagnetic environment simulation map including signal reflection path, interference intensity distribution and multipath fading characteristics, and mark the signal interference level at each spatial position; Step (3), identifying areas in the electromagnetic environment simulation map where the signal interference intensity exceeds a first threshold, and marking them as strong interference areas; deploying a deployable smart reflector array at the edge of the strong interference area, and dynamically optimizing the phase configuration of the smart reflector unit through a deep reinforcement learning algorithm to form a beam null for suppressing reflection interference; Step (4) dynamically divides the drone swarm into a communication subgroup and a monitoring subgroup based on the spatial distribution of the strong interference area; the monitoring subgroup uses a topology reconstruction algorithm to form a high-density sensor array, performs adaptive focus scanning on the strong interference area, obtains local fluctuation parameters with sub-meter accuracy, and feeds data back to the digital twin model in real time for incremental learning; Step (5) uses a distributed model predictive control framework to optimize the communication subgroup formation, solve the multi-objective optimization problem, and generate UAV position adjustment instructions; wherein the optimization objectives include communication quality indicators, energy consumption balance and topology stability; Step (6) dynamically configures the communication parameters of the UAV based on the spectrum situation knowledge graph. When it is detected that the residual interference exceeds the second threshold, the federated learning mechanism is triggered to collaboratively decide the optimal frequency band switching solution and synchronously update the phase distribution of the intelligent reflective surface.

2. The method according to claim 1, wherein: In the step (1), the multimodal perception module includes a millimeter wave radar array, a distributed laser ranging unit, a surface acoustic wave sensor network and a polarization light sensor; wherein the millimeter wave radar array measures the slope change rate and the three-dimensional motion vector of the water surface wave crest, the laser ranging unit detects the water surface spectrum characteristics through coherent interference, the surface acoustic wave sensor network captures the air turbulence velocity field at a height of 10 cm near the water surface, and the polarization light sensor inverts the dynamic change of the complex refractive index of the water surface medium; and the above multi-source heterogeneous data are integrated to construct a water surface dynamic digital twin model.

3. The adaptive scheduling method for a communication drone swarm according to claim 1 or 2, characterized in that: The water surface dynamic digital twin model includes a dynamic mapping layer, an evolution prediction layer, and a physical-signal coupling layer; the dynamic mapping layer uses a spatial topology matching algorithm to align the above-mentioned multi-source heterogeneous data with the three-dimensional water surface grid model in real time; the evolution prediction layer simulates the dynamic evolution of waves based on the finite element method and computational fluid dynamics, combines Kalman filtering for data fusion and anomaly correction, and predicts the water surface fluctuation trend; the physical-signal coupling layer embeds an electromagnetic wave propagation model to quantify the interference of changes in water surface inclination on the signal reflection path, and generate a dynamic electromagnetic environment map.

4. The adaptive scheduling method for a communication drone swarm according to claim 1, characterized in that: The step (3) specifically includes: Step (301): using the deployable intelligent reflective surface mounted on the drone to form a ring array outside the strong interference area; Step (302): Establish a mapping relationship model between the phase distribution of the intelligent reflector unit and the interference source orientation: in, is the phase of the nth smart reflector unit, is the direct channel response, is the reflection channel response of the smart reflection surface; Step (303) is based on a double-delayed deep deterministic policy gradient algorithm, and the phase distribution is updated according to the real-time electromagnetic environment feedback, and the spacing between adjacent smart reflective surface units satisfies d<λ / 2, and the phase resolution is not less than 4 bits; where λ is the signal wavelength.

5. The adaptive scheduling method for a communication drone swarm according to claim 4, characterized in that: The direct channel response is calculated as: ;in, is the direct path attenuation coefficient (including path loss and shadow fading), is the direct path delay, is the specific frequency value of the kth frequency sampling point; The reflection channel response is calculated as: ;in, is the reflection coefficient of the mth smart reflector unit, is the initial phase offset of the mth smart reflector unit, is the propagation delay from the transmitter to the mth smart reflector unit, represents the equivalent delay adjustment introduced by the phase configuration θ.

6. The adaptive scheduling method for a communication drone swarm according to claim 1, characterized in that: The objective function of the multi-objective optimization problem in step (5) is Where X represents the coefficient of the drone group position matrix; α, β, γ are multi-objective weights, and α + β + γ = 1; QoS represents the quality of communication service, Energy represents the energy consumption balance of the group, and Topo_Stab represents the formation deformation index; Constraints include: Drone spacing ;in, , Represents the three-dimensional coordinates of the i / j-th UAV; Indicates the minimum safe distance; Ground user receiving signal strength ;in, represents the received signal strength of the kth ground user, Indicates the communication assurance threshold; Maximum battery charge difference ;in, Indicates the maximum battery power difference between drones in a group.

7. The adaptive scheduling method for a communication drone swarm according to claim 6, characterized in that: The communication service quality is calculated as follows: ; Where SINR is the signal-to-interference-plus-noise ratio, Throughput is the effective throughput, ω1, ω2 are weight coefficients, and ω1+ω2=1; Calculation method of group energy consumption balance: ;in, is the remaining battery percentage of the i-th drone, The average remaining power of the group; The calculation method of formation deformation index is: ;in, is the matrix Frobenius norm, The time interval between adjacent optimization cycles.

8. The adaptive scheduling method for a communication drone swarm according to claim 1, characterized in that: In step (6), the spectrum situation knowledge graph is constructed through the graph attention network.

Images