A method and system for controlling a cluster of unmanned aerial vehicles

Abstract

This invention relates to the field of unmanned aerial vehicle (UAV) swarm control technology. To address the problem in existing UAV control methods where the inability to synchronize states between UAVs leads to a lack of inter-UAV collaboration, this invention provides a UAV swarm control method and system. The method executes the following steps: Step S1, collecting the local state of each UAV to generate standardized local state information; Step S2, performing distributed full-state synchronization processing on the standardized local state information to generate globally consistent state information; Step S3, performing local intelligent decision-making processing on the globally consistent state information to generate local control commands; Step S4, performing execution control processing on the local control commands to generate UAV actions; Step S5, performing effect feedback processing on the UAV actions to generate updated local state information, and using the updated local state information as the new local state information, returning to step S2.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) swarm control technology, and in particular to a UAV swarm control method and system. Background Technology

[0002] With the expansion of drone application scenarios and the liberalization of policies, the number of drone registrations has surged, highlighting the importance of drone swarm control. Traditional "one-to-one" and "one-controller-one-drone" drone control methods are no longer suitable for scenarios where multiple drones collaborate to complete tasks, making drone swarm control technology urgently needed.

[0003] Traditional drone control methods prevent drones from synchronizing their states, resulting in a lack of inter-drone collaboration. When multiple drones perform collaborative tasks, issues such as slow task planning, lack of communication between nodes, inability to collaborate, and lack of survivability arise. Summary of the Invention

[0004] This invention addresses the problems in existing drone control methods where drones cannot synchronize their states, leading to a lack of inter-drone collaboration, slow task planning, node incompatibility and inability to collaborate when multiple drones are performing collaborative tasks. It provides a drone swarm control method and system.

[0005] The technical solution adopted in this invention is: The UAV swarm control method disclosed in this invention performs the following steps: Step S1: Collect the local status of each drone and generate standardized local status information; Among them, the collected local status is the raw measurement data; Step S2: Perform distributed full-state synchronization processing on the standardized local state information to generate globally consistent state information; Step S3: Perform local intelligent decision-making processing on the globally consistent state information to generate local control commands; Step S4: Execute control processing on local control commands to generate drone actions; Step S5: Perform effect feedback processing on the drone's actions, generate updated local state information, and use the updated local state information as the new local state information, then return to step S2.

[0006] Furthermore, step S1 specifically includes the following sub-steps: Step S11: Read the raw measurement data collected by the sensors of each UAV; The sensing devices include an inertial measurement unit, a global positioning system receiver, a barometer, a visual sensor, and a lidar. The sensing device is used to collect position, velocity, attitude, and angular velocity data; the position, velocity, attitude, and angular velocity data collected by the sensing device are raw state data; Step S12: Perform error correction and filtering on the original state data; Among them, error correction is based on parameter compensation of the preset sensor error model, and filtering is achieved by constructing an extended Kalman filter with the UAV motion model as the state equation and the sensor observation model as the observation equation. The Kalman filter is used to fuse and smooth the multi-source heterogeneous original state data to generate stable and reliable effective state data. Step S13: Format and timestamp the valid state data, encapsulate it into state information tuples with a predefined data structure, and attach a unique and monotonically increasing timestamp to each state information tuple based on a high-precision clock source to generate standardized local state information.

[0007] Furthermore, step S2 specifically includes the following sub-steps: Step S21: Perform information encapsulation processing on the standardized local status information, add a unique local identifier, current information version number, timestamp, and hash value for integrity verification, and generate a status information packet in a standard format. Step S22: Broadcast and receive standard format status information packets from neighboring nodes. Through the UAV ad hoc network communication protocol, broadcast standard format status information packets to neighboring UAVs within the communication range periodically, and simultaneously receive standard format status information packets from neighboring UAVs, generating a set of status information copies containing the local device and all received neighboring UAVs. Step S23: Consensus verification and state merging are performed on the set of state information replicas. By comparing the version number, timestamp, and hash value, the temporal correctness and integrity of each information packet are verified. The state information with differences is coordinated according to the preset conflict resolution rules to generate a temporary global state view containing the latest state estimate of all members of the cluster. Step S24: Process the temporary global state view using a distributed consensus algorithm. Utilize the practical Byzantine fault tolerance algorithm to enable all drone nodes in the cluster to reach a consensus on the validity of the temporary global state view, and finally generate a globally consistent state information that is jointly confirmed and locked by all honest nodes.

[0008] Furthermore, step S3 specifically includes the following sub-steps: Step S31: Perform situation understanding and target parsing processing on the globally consistent state information. Based on the preset cluster task objectives, extract situation elements related to the local task from the globally consistent state information to generate a situation assessment result that includes quantifiable task indicators, environmental constraints, and the local role in the cluster. Step S32: Perform control decision calculation processing on the situation assessment results. Based on the preset decision function, input the situation assessment results, calculate the expected motion trajectory or direct control quantity of the machine required to achieve the mission objective, and generate preliminary control decisions. Step S33: Perform decision verification and instruction generation processing on the preliminary control decision. Based on the dynamic constraints, obstacle avoidance rules and swarm coordination rules of the UAV, verify the feasibility and safety of the preliminary control decision. Convert the verified control decision into a standardized instruction format that can be recognized by the UAV motion control module and generate local control instructions.

[0009] Furthermore, step S4 specifically includes the following sub-steps: Step S41: Perform instruction parsing and allocation processing on the local control instructions. Based on the configuration of the UAV and the actuator configuration, calculate and allocate the unified control instructions to each rotor motor and control surface servo mechanism, and generate sub-instructions for each independent actuator; wherein, the actuators include each rotor motor and control surface servo mechanism. Step S42: Perform dynamic model feedforward and feedback processing on the sub-commands of each actuator. Calculate the feedforward control quantity based on the nonlinear dynamic model of the UAV, and combine it with feedback from the real-time sensing device. Calculate the feedback control quantity through the proportional-integral-derivative control law. Fuse the feedforward control quantity and the feedback control quantity to generate the final control quantity of each actuator that can accurately achieve the desired motion and resist interference. Step S43: The final control quantities of each actuator are converted into drive signals and amplified by power, and converted into pulse width modulation signals and analog voltages. The motors are then driven by electronic speed controllers and servo drivers to generate physical actions, and finally the drone actions are synthesized. Among them, drone actions include drones flying, hovering, and maneuvering in space.

[0010] Furthermore, step S5 specifically includes the following sub-steps: Step S51: Perceive and process the environmental changes and changes in the drone's own state caused by the drone's actions. Use onboard sensors to perceive new changes in the drone's own pose and speed caused by the execution of local control commands, as well as changes in environmental factors, in real time, and generate new raw state data. Step S52: Perform error correction and filtering on the new raw state data. Use the same data processing flow and model as in step S12 to clean, correct and filter the perceived new data in real time to prepare new effective state data for decision-making in the next control cycle. Step S53: Format and timestamp the new valid state data. Using the same standardized process as step S13, add a new timestamp to the new valid state data and encapsulate it to generate updated local state information that serves as input for the next control cycle and is used to evaluate the effect of action execution.

[0011] Furthermore, the core architecture of the practical Byzantine fault-tolerant algorithm in step S24 is as follows: A three-phase consensus protocol is adopted, including a pre-preparation phase, a preparation phase, and a submission phase; During the pre-preparation phase, the master node of the current view assigns a sequence number to the received client request and broadcasts a pre-preparation message. During the preparation phase, each backup node, after verifying the validity of the pre-preparation message, broadcasts the preparation message to the entire network and collects preparation messages from other nodes. During the commit phase, once a node has collected valid preparation messages from more than two-thirds of the nodes for the same view and sequence number, it broadcasts a commit message. After receiving more than two-thirds of the valid commit messages, it commits the temporary global state view as globally consistent state information stored locally and executes the corresponding state machine application.

[0012] Furthermore, the parameters of the Byzantine fault-tolerant algorithm used in step S24 include the view number v, the current primary node identifier primary_id, the latest stable checkpoint sequence number stable_checkpoint_n, the current consensus sequence number n, and the total number of nodes in the cluster N and the maximum number of fault-tolerant nodes f; Where f satisfies N≥3f+1; during runtime, the algorithm also maintains an acknowledgment set containing 2f+1 nodes to record the message collection status at each stage. When a valid acknowledgment for the same message is received from more than two-thirds (i.e., at least 2f+1) of the nodes in the set, it is considered that the message has reached consensus in this stage.

[0013] Furthermore, the decision function in step S31 is expressed as: u_i(t)=π(s_global(t),g,τ_i,C_{dynamic},C_{static}) in: u_i(t) represents the local control command generated by UAV i at time t; π(·) represents a pre-defined decision algorithm model, which is a mapping function based on optimization theory, rule reasoning, or deep neural networks; s_global(t) represents the globally consistent state information reached at time t, which includes the state of all members of the cluster; g represents the preset cluster-level task objective, including formation parameters, target point coordinates, and coverage area; τ_i represents the role of drone i in the swarm, including leader, follower, and scout; C_{dynamic} represents the set of dynamic constraints, including real-time obstacle avoidance rules, minimum safe distance from neighboring drones, and communication range limitations; C_{static} represents the set of static constraints, including the UAV's own dynamic limits, energy constraints, and mission area boundaries.

[0014] As attached Figure 2 As shown, based on the same inventive concept, the present invention also discloses a drone swarm control system, which executes the aforementioned drone swarm control method. The drone swarm control system includes: The state awareness module 1 is used to access the raw measurement data of each UAV, collect and standardize the raw measurement data, and output the standardized local state information of each UAV. State synchronization module 2 is used to receive standardized local state information, perform distributed full state synchronization processing on the standardized local state information, and output globally consistent state information. Intelligent decision module 3 is used to receive globally consistent status information, perform local intelligent decision processing on the globally consistent status information, and output local control commands; Control execution module 4 is used to receive local control commands, perform execution control processing on the local control commands, and output UAV action commands; Feedback evaluation module 5 is used to process the feedback based on the actual effect after the drone's action commands are executed, generate updated local status information, and send the updated local status information to status synchronization module 2.

[0015] Based on the same inventive concept, the present invention also discloses a drone swarm control system, which executes the aforementioned drone swarm control method. The drone swarm control system includes: The state awareness module is used to access the raw measurement data of each UAV, collect and standardize the raw measurement data, and output standardized local state information of each UAV. The state synchronization module is used to receive standardized local state information, perform distributed full-state synchronization processing on the standardized local state information, and output globally consistent state information. The intelligent decision-making module is used to receive globally consistent status information, perform local intelligent decision-making processing on the globally consistent status information, and output local control commands. The control execution module is used to receive local control commands, perform execution control processing on the local control commands, and output UAV action commands; The feedback evaluation module is used to process the feedback based on the actual effect of the drone's action commands, generate updated local status information, and send the updated local status information to the status synchronization module.

[0016] The beneficial effects of this invention are: This invention discloses a UAV swarm control method and system that achieves efficient autonomous collaboration among UAV swarms through a technical solution involving standardized state acquisition, distributed full-state synchronization, local intelligent decision-making, and a closed-loop command execution and effect feedback mechanism. Distributed full-state synchronization enables the swarm to obtain a globally consistent state view without a central node, ensuring unified decision-making and coordinated actions among multiple UAVs, significantly improving system resilience and scalability. Standardized state processing provides reliable data support for swarm collaboration, while local intelligent decision-making considers both the overall situation and individual UAV characteristics, and closed-loop control enables continuous optimization. Overall, this significantly improves swarm autonomy, collaboration accuracy, and task execution efficiency, meeting the requirements for large-scale, intelligent, and highly reliable multi-UAV operations in complex scenarios. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 Flowchart of the unmanned aerial vehicle (UAV) swarm control method; Figure 2 This is a block diagram of an unmanned aerial vehicle (UAV) swarm control system. Detailed Implementation

[0019] The following disclosure provides many different embodiments or examples for implementing different structures of the present invention. To simplify the disclosure of the present invention, the components and arrangements of specific examples are described below. Of course, these are merely examples and are not intended to limit the present invention.

[0020] The embodiments of the invention will now be described in detail with reference to the accompanying drawings.

[0021] As attached Figure 1 As shown, the UAV swarm control method disclosed in this embodiment performs the following steps: As attached Figure 1 As shown, the UAV swarm control method disclosed in this embodiment performs the following steps: Step S1: Collect the local status of each drone and generate standardized local status information; Among them, the collected local status is the raw measurement data; Step S2: Perform distributed full-state synchronization processing on the standardized local state information to generate globally consistent state information; Step S3: Perform local intelligent decision-making processing on the globally consistent state information to generate local control commands; Step S4: Execute control processing on local control commands to generate drone actions; Step S5: Process the feedback of the drone's actions to generate updated local state information, and use this updated local state information as the new local state information, returning to step S2. This process is continuously iterated to form a continuous autonomous collaborative control closed loop, which will only terminate when the mission ends or the system actively shuts down.

[0022] The beneficial effects of steps S1-S5 are as follows: By constructing a complete closed-loop control process of "state acquisition - distributed synchronization - intelligent decision-making - command execution - effect feedback," the method achieves autonomous collaborative operation of the UAV swarm. Specifically, it first standardizes the raw, heterogeneous local state data, providing a consistent and reliable data foundation for subsequent collaborative decision-making. Next, through distributed full-state synchronization processing, it ensures that all UAVs in the swarm make decisions based on the same global state view without a central node, guaranteeing the consistency of collaborative behavior. Then, each UAV makes localized intelligent decisions based on this global information, generating control commands adapted to its own role and task, achieving a combination of distributed autonomous decision-making and centralized global situational awareness. These commands, after execution control processing, are ultimately transformed into precise UAV flight maneuvers. After the maneuver is executed, the local state is updated in real time through effect feedback processing and used as input for the next control cycle, forming an autonomous collaborative closed loop capable of continuous perception, decision-making, execution, learning, and optimization. This process comprehensively improves the robustness, scalability, autonomy, and task completion efficiency of the swarm system.

[0023] Furthermore, step S1 specifically includes the following sub-steps: Step S11: Read the raw measurement data collected by the sensors of each UAV; The sensing devices include an inertial measurement unit, a global positioning system receiver, a barometer, a visual sensor, and a lidar. The sensing device is used to collect position, velocity, attitude, and angular velocity data; the position, velocity, attitude, and angular velocity data collected by the sensing device are raw state data; Step S12: Perform error correction and filtering on the original state data; The error correction is based on a preset sensor error model for parameter compensation (specifically, based on the factory calibration parameters and field calibration data of sensors such as inertial measurement units, satellite positioning receivers, vision and radar, an error model including zero bias error, proportional error, installation error and temperature drift is established; when the original state data is input, the measured values ​​are corrected in real time for offset, proportional coefficient correction and axis alignment compensation according to the error model, eliminating the inherent system error of the sensor and providing accurate original input for subsequent filtering processing). The filtering processing is achieved by constructing an extended Kalman filter with the UAV motion model as the state equation and the sensor observation model as the observation equation. The Kalman filter is used to fuse and smooth the multi-source heterogeneous original state data to generate stable and reliable effective state data. Step S13: Format and timestamp the valid state data, encapsulate it into state information tuples with a predefined data structure, and attach a unique and monotonically increasing timestamp to each state information tuple based on a high-precision clock source to generate standardized local state information.

[0024] The beneficial effects of steps S11-S13 are as follows: By standardizing the raw measurement data through a "read-correction filtering-formatting and marking" process, this step significantly improves the accuracy, reliability, and usability of the UAV's own state perception. Specifically, by fusing data from multiple heterogeneous sensors such as inertial, satellite, and vision sensors, the limitations of a single sensor are overcome, and a more comprehensive raw state is obtained. Using compensation and extended Kalman filtering techniques based on a preset error model, sensor system errors are effectively corrected, random noise is smoothed, and multi-source data is optimally fused under the constraints of the UAV's motion model, thereby generating stable, continuous, and highly accurate effective state data. Finally, through formatting and timestamp marking, the effective data is encapsulated into standardized information tuples with a unified structure and strict temporal relationships. This not only ensures the consistency of state information processing within the UAV but, more importantly, provides a precise time reference and data format foundation for state synchronization and comparison between multiple UAVs in subsequent steps, which is a prerequisite for achieving high-precision collaboration.

[0025] Furthermore, step S2 specifically includes the following sub-steps: Step S21: Perform information encapsulation processing on the standardized local status information, add a unique local identifier, current information version number, timestamp, and hash value for integrity verification, and generate a status information packet in a standard format. Step S22: Broadcast and receive standard format status information packets from neighboring nodes. Through the UAV ad hoc network communication protocol, broadcast standard format status information packets to neighboring UAVs within the communication range periodically, and simultaneously receive standard format status information packets from neighboring UAVs, generating a set of status information copies containing the local device and all received neighboring UAVs. Step S23: Consensus verification and state merging are performed on the set of state information replicas. By comparing the version number, timestamp, and hash value, the temporal correctness and integrity of each information packet are verified. The state information with differences is coordinated according to the preset conflict resolution rules to generate a temporary global state view containing the latest state estimate of all members of the cluster. Step S24: Process the temporary global state view using a distributed consensus algorithm. Utilize the practical Byzantine fault tolerance algorithm to enable all drone nodes in the cluster to reach a consensus on the validity of the temporary global state view, and finally generate a globally consistent state information that is jointly confirmed and locked by all honest nodes.

[0026] The beneficial effects of steps S21-S24 are as follows: Through a complete distributed processing flow of "encapsulation-broadcast-verification-consensus," highly reliable and strongly consistent synchronization of UAV cluster state information is achieved. First, the information encapsulation process adds identity, time sequence, and integrity identifiers to the state data, laying the foundation for subsequent communication, verification, and conflict coordination. Second, periodic broadcasting and receiving via a self-organizing network protocol enables efficient and decentralized diffusion of state information in dynamically changing network topologies. Next, consensus verification and state merging processes, based on version, timestamp, and hash value, verify and resolve conflicts of the received multi-copy information, initially forming a temporary global view containing the latest estimates from all nodes, effectively addressing information inconsistencies caused by network latency, out-of-order delivery, and packet loss. Finally, distributed consensus processing using a practical Byzantine fault-tolerant algorithm ensures that all honest nodes in the cluster can reach a final consensus on this temporary view. Even with a few node failures or malicious behavior (Byzantine faults), a unique, deterministic, and globally consistent state information locked by all honest nodes is generated, fundamentally guaranteeing the reliability and consistency of the cluster's collaborative decision-making foundation.

[0027] Furthermore, step S3 specifically includes the following sub-steps: Step S31: Perform situation understanding and target parsing processing on the globally consistent state information. Based on the preset cluster task objectives, extract situation elements related to the local task from the globally consistent state information to generate a situation assessment result that includes quantifiable task indicators, environmental constraints, and the local role in the cluster. Step S32: Perform control decision calculation processing on the situation assessment results. Based on the preset decision function, input the situation assessment results, calculate the expected motion trajectory or direct control quantity of the machine required to achieve the mission objective, and generate preliminary control decisions. Step S33: Perform decision verification and instruction generation processing on the preliminary control decision. Based on the dynamic constraints, obstacle avoidance rules and swarm coordination rules of the UAV, verify the feasibility and safety of the preliminary control decision. Convert the verified control decision into a standardized instruction format that can be recognized by the UAV motion control module and generate local control instructions.

[0028] The beneficial effects of steps S31-S33 are as follows: Through the progressive processing of "situational understanding - decision calculation - verification and generation," autonomous intelligent decision-making based on globally consistent information, while possessing both mission adaptability and flight safety, is achieved. First, the situational understanding and target resolution processing extracts key elements highly relevant to the current mission from massive amounts of global information and clarifies the drone's role and constraints, transforming abstract mission objectives into concrete, quantifiable decision inputs, thus improving the targeting and efficiency of decision-making. Second, the control decision calculation processing utilizes a pre-set decision model and, based on the quantified situational assessment results, quickly calculates the preliminary control strategy to achieve the mission objective, demonstrating the intelligence and autonomy of decision-making. Finally, the decision verification and command generation processing is crucial. It rigorously verifies the dynamic feasibility, real-time obstacle avoidance, and compliance with collaborative rules of the preliminary decisions, filtering out decisions that may lead to danger or conflict, and converting verified decisions into standard control commands. This process ensures that the generated local control commands not only serve high-level mission objectives but also strictly adhere to the physical limitations of the UAV itself and swarm flight safety rules, achieving a unity of intelligence, autonomy, safety, and reliability.

[0029] Furthermore, step S4 specifically includes the following sub-steps: Step S41: Perform instruction parsing and allocation processing on the local control instructions. Based on the configuration of the UAV and the actuator configuration, calculate and allocate the unified control instructions to each rotor motor and control surface servo mechanism, and generate sub-instructions for each independent actuator; wherein, the actuators include each rotor motor and control surface servo mechanism. Step S42: Perform dynamic model feedforward and feedback processing on the sub-commands of each actuator. Calculate the feedforward control quantity based on the nonlinear dynamic model of the UAV, and combine it with feedback from the real-time sensing device. Calculate the feedback control quantity through the proportional-integral-derivative control law. Fuse the feedforward control quantity and the feedback control quantity to generate the final control quantity of each actuator that can accurately achieve the desired motion and resist interference. Step S43: The final control quantities of each actuator are converted into drive signals and amplified by power, and converted into pulse width modulation signals and analog voltages. The motors are then driven by electronic speed controllers and servo drivers to generate physical actions, and finally the drone actions are synthesized. Among them, drone actions include drones flying, hovering, and maneuvering in space.

[0030] The beneficial effects of steps S41-S43 are as follows: Through the refined control chain of "analysis and allocation - feedforward and feedback fusion - drive execution," the high-level decision-making commands are accurately, reliably, and robustly transformed into actual UAV flight maneuvers. First, command parsing and allocation decomposes the unified control commands into sub-commands that match the specific actuators of the UAV (such as motors and servos), achieving accurate mapping from control commands to the physical execution layer. Second, dynamic model feedforward and feedback processing is the core of precise control: feedforward control based on nonlinear models can compensate for the dynamic characteristics of the system in advance, achieving rapid response; PID control combined with real-time sensor feedback can effectively suppress model errors and external disturbances. The fusion of these two aspects ensures that the generated control quantity can accurately achieve the desired motion while possessing strong anti-interference capabilities. Finally, drive signal conversion and power amplification complete the final conversion from digital control quantities to physical actions, driving the actuators through mature power electronics technology, ultimately synthesizing complex and precise UAV flight maneuvers. The entire process achieves high-fidelity, low-latency conversion from control commands to physical actions.

[0031] Furthermore, step S5 specifically includes the following sub-steps: Step S51: The changes in environment and the UAV's own state caused by its actions are processed through perception. Onboard sensors are used to perceive in real time the new changes in its own pose and velocity resulting from the execution of local control commands, as well as changes in environmental factors, generating new raw state data. (The perception processing process is specifically as follows: While the UAV performs physical actions driven by the actuators, its onboard sensing system works synchronously. Inertial measurement units, GPS receivers, visual odometry, and other state perception sensors continuously measure the motion parameters of the UAV itself. Through recursive calculation or direct observation, the UAV's own pose caused by the execution of control commands is calculated in real time.) The system tracks new changes in position, attitude, and velocity (linear and angular velocities). Simultaneously, environmental sensors such as airspeed indicators, visual sensors, and lidar continuously monitor the external environment. By detecting the forces exerted by airflow on the aircraft or scanning point cloud data of surrounding obstacles and comparing it with information from the previous cycle, the system detects changes in environmental factors (such as wind disturbance forces / torques and the relative positions of obstacles). Finally, the flight control system uses a unified time base to package the multi-source heterogeneous measurement values ​​collected by all the aforementioned sensors, reflecting the new state after the "action," into a new, unprocessed set of raw state data with timestamps, which serves as input for subsequent data processing. Step S52: Perform error correction and filtering on the new raw state data. Use the same data processing flow and model as in step S12 to clean, correct and filter the perceived new data in real time to prepare new effective state data for decision-making in the next control cycle. Step S53: Format and timestamp the new valid state data. Using the same standardized process as step S13, add a new timestamp to the new valid state data and encapsulate it to generate updated local state information that serves as input for the next control cycle and is used to evaluate the effect of action execution.

[0032] The beneficial effects of steps S51-S53 are as follows: Through the feedback processing flow of "perceiving new states - processing new data - standardizing new information," real-time evaluation of the control action's execution effect is achieved, and the key connection of the control closed loop is completed. First, by real-time perception of state changes (such as new poses and velocities) and environmental changes (such as new obstacle positions) caused by its own actions, raw data reflecting the action execution results and the latest environmental conditions are generated. This forms the perceptual basis for effect evaluation and subsequent decision-making. Second, the new data is processed using the same error correction and filtering process as in step S12, ensuring the consistency of the system's state data processing standards and quality, eliminating the influence of sensor noise and errors on feedback information, and generating reliable and effective state data. Finally, using the same formatting and timestamping process as in step S13, the new effective data is encapsulated into standardized state information tuples, generating "updated local state information." This information serves two purposes: firstly, it assesses whether the control actions of the previous cycle have achieved the expected results; secondly, it serves as the input for step S2 in the next control cycle, thereby re-injecting the results of "execution-perception" into the control loop. This enables the entire cluster system to make continuous collaborative decisions and adaptations based on the latest and most accurate environment and its own state, achieving closed-loop feedback and dynamic optimization of the system.

[0033] Furthermore, the core architecture of the practical Byzantine fault-tolerant algorithm in step S24 is as follows: It adopts the classic three-phase consensus protocol, including the pre-preparation phase, the preparation phase, and the commit phase; During the pre-preparation phase, the master node of the current view assigns a sequence number to the received client request (in this case, a temporary global state view) and broadcasts a pre-preparation message. During the preparation phase, each backup node, after verifying the validity of the pre-preparation message, broadcasts the preparation message to the entire network and collects preparation messages from other nodes. During the commit phase, once a node has collected valid preparation messages from more than two-thirds of the nodes for the same view and sequence number, it broadcasts a commit message. After receiving more than two-thirds of the valid commit messages, it commits the temporary global state view as globally consistent state information stored locally and executes the corresponding state machine application.

[0034] The specific steps for training the practical Byzantine fault-tolerant algorithm in further step S24 are as follows: In the offline training phase before drone swarm deployment, an adversarial training environment is constructed using a simulator, incorporating various network latency, packet loss, node failures, and Byzantine nodes (sending contradictory or erroneous information). Nodes using the Byzantine fault-tolerant algorithm (including master and backup nodes) are placed as agents in this environment. The consensus achievement speed, communication overhead, and robustness in outputting correct and consistent states in the face of failures and malicious behavior are used as joint optimization objectives. Deep reinforcement learning algorithms are used to train the message processing strategies, timeout retransmission mechanisms, and view switching strategies of nodes at each stage of consensus (pre-preparation, preparation, and commit), enabling the algorithm parameters to be adaptively optimized to cope with the complex and ever-changing communication conditions and node behaviors in real physical networks.

[0035] Simulator Construction in Practical Byzantine Fault Tolerance Algorithm This adversarial simulator for training practical Byzantine Fault Tolerance (PBFT) algorithms employs a four-layer integrated architecture: UAV dynamics simulation, wireless ad hoc network communication simulation, node fault and malicious behavior simulation, and training evaluation closed loop. Its core simulation kernel is built upon simulation engines such as ROS-Gazebo, AirSim, and NS-3. It simultaneously integrates a UAV nonlinear motion model and an Ad-Hoc ad hoc network communication protocol module, and independently encapsulates network interference plugins and node behavior scripting control modules. It also interfaces with deep reinforcement learning agents and algorithm performance evaluation modules. Using a purely digital offline simulation mode, it supports configurable cluster node numbers and can synchronously simulate UAV cluster spatial motion, dynamic network topology changes, various communication anomalies, node faults, and malicious behaviors. This constructs an adversarial training environment highly consistent with real UAV cluster deployment scenarios, providing complete simulation support for PBFT algorithm strategy optimization.

[0036] Environmental parameter settings in practical Byzantine fault-tolerant algorithms The simulator employs a gradient-based approach to setting adversarial parameters such as network latency, packet loss, node failures, and Byzantine nodes. Network latency is categorized into three types: fixed latency, random jitter latency, and step-change latency, with values ​​adjusted gradients across basic, advanced, and extreme training stages within the 10ms~500ms range. Packet loss is configured with three modes: random, burst, and targeted, with the packet loss rate gradually increasing from 1% to 40%. Regular node failures strictly adhere to the PBFT algorithm constraint N≥3f+1, ensuring the number of faulty nodes does not exceed the maximum fault-tolerant node number f. Simulations of crashes, offline events, and restarts are provided, along with corresponding triggering times and durations. The number of malicious Byzantine nodes also does not exceed f. Targeted configurations are available for the trigger probability, action stage, and specific parameters of malicious behaviors such as data tampering, sending contradictory messages, malicious silence, and identity forgery. All parameters are set in conjunction with the actual control cycle of the UAV and engineering constraints, covering the full gradient training requirements from stable scenarios to extreme adversarial scenarios.

[0037] Furthermore, the parameters of the Byzantine fault-tolerant algorithm used in step S24 include the view number v, the current primary node identifier primary_id, the latest stable checkpoint sequence number stable_checkpoint_n, the current consensus sequence number n, and the total number of nodes in the cluster N and the maximum number of fault-tolerant nodes f; Where f satisfies N≥3f+1; during runtime, the algorithm also maintains an acknowledgment set containing 2f+1 nodes to record the message collection status at each stage. When a valid acknowledgment for the same message is received from more than two-thirds (i.e., at least 2f+1) of the nodes in the set, it is considered that the message has reached consensus in this stage.

[0038] Furthermore, the decision function in step S31 is expressed as: u_i(t)=π(s_global(t),g,τ_i,C_{dynamic},C_{static}) in: u_i(t) represents the local control command generated by UAV i at time t; π(·) represents a pre-defined decision algorithm model, which is a mapping function based on optimization theory, rule reasoning, or deep neural networks; s_global(t) represents the globally consistent state information reached at time t, which includes the state of all members of the cluster; g represents the preset cluster-level task objective, including formation parameters, target point coordinates, and coverage area; τ_i represents the specific role that drone i plays in the swarm, including leader, follower, and scout; C_{dynamic} represents the set of dynamic constraints, including real-time obstacle avoidance rules, minimum safe distance from neighboring drones, and communication range limitations; C_{static} represents the set of static constraints, including the UAV's own dynamic limits, energy constraints, and mission area boundaries.

[0039] π(・) employs a distributed model predictive control method based on optimization theory. This model takes the globally consistent state information achieved at the current moment, the preset cluster task objectives, and the role of the UAV in the cluster as its core basis. At the same time, it takes dynamic constraints such as real-time obstacle avoidance rules, safe distance limits with neighboring UAVs, and communication range constraints, as well as static conditions such as the UAV's maximum flight speed, acceleration, dynamic limits, and energy constraints as hard constraints. Relying on the nonlinear dynamic model of the UAV, it makes rolling predictions of the flight state in the future. In each control cycle, it constructs an optimization solution logic with the goals of accurately following the task trajectory, maintaining the cluster's collaborative stability, and reducing unnecessary energy consumption. Through online rolling calculations, it selects the optimal control scheme that meets all constraints and outputs the optimal control quantity corresponding to the current moment as the final local control command. This ensures that the control of a single UAV conforms to its own physical characteristics and safety specifications, and that the decision of a single UAV is highly consistent with the global situation of the cluster and the overall task objectives.

[0040] The significance of this decision function formula lies in its clear definition of how a single UAV combines global information from the cluster with its own specific conditions to generate safe, efficient control commands that meet the overall mission objectives.

[0041] Specifically, the significance of this formula is reflected in the following aspects: This embodies the collaborative paradigm of "global perception and local decision-making": the core input of the function, s_global(t), is the globally consistent state information achieved through step S2. This means that when each drone makes a decision, it does not rely solely on its own limited local perspective, but rather on a unified and reliable panoramic view of the cluster. This ensures that the decisions of all individuals are based on a shared situational awareness, which is fundamental to achieving precise collaboration.

[0042] The mapping process from mission objectives to control commands is quantified: the function π(·) represents a specific decision-making algorithm (such as an optimization controller, rule engine, or neural network). It maps the abstract high-level mission objective g (e.g., "maintaining a diamond formation to a point") and the UAV's own role τ_i (e.g., "left wing member of the formation") to specific, executable flight control commands u_i(t) through computation. This quantitative description makes the autonomous decision-making process designable, optimizable, and verifiable.

[0043] Integrating multi-dimensional constraints ensures the feasibility and safety of decision-making: the formula explicitly uses dynamic constraints C_dynamic (such as real-time obstacle avoidance and collision prevention) and static constraints C_static (such as the physical limits of the organism) as key inputs to the decision. This means that when the decision algorithm π solves for the optimal or feasible control instructions, it must treat these constraints as hard conditions or optimization objectives, thereby ensuring that the generated instructions are not only "effective" but also "safe" and "executable," preventing loss of control or danger due to improper decision-making.

[0044] This lays the foundation for system adaptability and intelligence: the function defines a clear input-output relationship. In practical systems, the decision model π can be trained and optimized using machine learning methods to adapt to complex and ever-changing environments. Simultaneously, the roles τ_i and task objectives g can be dynamically assigned, allowing the entire cluster to flexibly reorganize by adjusting the decision function parameters of each drone to cope with different task scenarios.

[0045] In summary, this formula is not merely a mathematical expression; it fundamentally describes how this invention achieves an autonomous, secure, and task-oriented decision-making mechanism supported by unified global information. It serves as a crucial intelligent bridge connecting "cluster collaborative perception" (step S2) and "individual precise execution" (step S4), and is the core of the entire method's efficient, robust, and autonomous collaboration.

[0046] As attached Figure 2 As shown, based on the same inventive concept, this embodiment also discloses a drone swarm control system, which executes the aforementioned drone swarm control method. The drone swarm control system includes: The state awareness module 1 is used to access the raw measurement data of each UAV, collect and standardize the raw measurement data, and output the standardized local state information of each UAV. State synchronization module 2 is used to receive standardized local state information, perform distributed full state synchronization processing on the standardized local state information, and output globally consistent state information. Intelligent decision module 3 is used to receive globally consistent status information, perform local intelligent decision processing on the globally consistent status information, and output local control commands; Control execution module 4 is used to receive local control commands, perform execution control processing on the local control commands, and output UAV action commands; Feedback evaluation module 5 is used to process the feedback based on the actual effect after the drone's action commands are executed, generate updated local status information, and send the updated local status information to status synchronization module 2.

Claims

1. A method for controlling a swarm of unmanned aerial vehicles (UAVs), characterized in that, Perform the following steps: Step S1: Collect the local status of each drone and generate standardized local status information; Among them, the collected local status is the raw measurement data; Step S2: Perform distributed full-state synchronization processing on the standardized local state information to generate globally consistent state information; Step S3: Perform local intelligent decision-making processing on the globally consistent state information to generate local control commands; Step S4: Execute control processing on local control commands to generate drone actions; Step S5: Perform effect feedback processing on the drone's actions, generate updated local state information, and use the updated local state information as the new local state information, then return to step S2.

2. The UAV swarm control method according to claim 1, characterized in that, Step S1 specifically includes the following sub-steps: Step S11: Read the raw measurement data collected by the sensors of each UAV; The sensing devices include an inertial measurement unit, a global positioning system receiver, a barometer, a visual sensor, and a lidar. The sensing device is used to collect position, velocity, attitude, and angular velocity data; the position, velocity, attitude, and angular velocity data collected by the sensing device are raw state data; Step S12: Perform error correction and filtering on the original state data; Among them, error correction is based on the preset sensor error model for parameter compensation, and filtering is achieved by constructing an extended Kalman filter with the UAV motion model as the state equation and the sensor observation model as the observation equation. The Kalman filter is used to fuse and smooth the multi-source heterogeneous original state data to generate stable and reliable effective state data. Step S13: Format and timestamp the valid state data, encapsulate it into state information tuples with a predefined data structure, and attach a unique and monotonically increasing timestamp to each state information tuple based on a high-precision clock source to generate standardized local state information.

3. The UAV swarm control method according to claim 1, characterized in that, Step S2 specifically includes the following sub-steps: Step S21: Perform information encapsulation processing on the standardized local status information, add a unique local identifier, current information version number, timestamp, and hash value for integrity verification, and generate a status information packet in a standard format. Step S22: Broadcast and receive standard format status information packets from neighboring nodes. Through the UAV ad hoc network communication protocol, broadcast standard format status information packets to neighboring UAVs within the communication range periodically, and simultaneously receive standard format status information packets from neighboring UAVs, generating a set of status information copies containing the local device and all received neighboring UAVs. Step S23: Consensus verification and state merging are performed on the set of state information replicas. By comparing the version number, timestamp, and hash value, the temporal correctness and integrity of each information packet are verified. The state information with differences is coordinated according to the preset conflict resolution rules to generate a temporary global state view containing the latest state estimate of all members of the cluster. Step S24: Process the temporary global state view using a distributed consensus algorithm. Utilize the practical Byzantine fault tolerance algorithm to enable all drone nodes in the cluster to reach a consensus on the validity of the temporary global state view, and finally generate a globally consistent state information that is jointly confirmed and locked by all honest nodes.

4. The UAV swarm control method according to claim 1, characterized in that, Step S3 specifically includes the following sub-steps: Step S31: Perform situation understanding and target parsing processing on the globally consistent state information. Based on the preset cluster task objectives, extract situation elements related to the local task from the globally consistent state information to generate a situation assessment result that includes quantifiable task indicators, environmental constraints, and the local role in the cluster. Step S32: Perform control decision calculation processing on the situation assessment results. Based on the preset decision function, input the situation assessment results, calculate the expected motion trajectory or direct control quantity of the machine required to achieve the mission objective, and generate preliminary control decisions. Step S33: Perform decision verification and instruction generation processing on the preliminary control decision. Based on the dynamic constraints, obstacle avoidance rules and swarm coordination rules of the UAV, verify the feasibility and safety of the preliminary control decision. Convert the verified control decision into a standardized instruction format that can be recognized by the UAV motion control module and generate local control instructions.

5. The UAV swarm control method according to claim 1, characterized in that, Step S4 specifically includes the following sub-steps: Step S41: Perform instruction parsing and allocation processing on the local control instructions. Based on the configuration of the UAV and the actuator configuration, calculate and allocate the unified control instructions to each rotor motor and control surface servo mechanism, and generate sub-instructions for each independent actuator; wherein, the actuators include each rotor motor and control surface servo mechanism. Step S42: Perform dynamic model feedforward and feedback processing on the sub-commands of each actuator. Calculate the feedforward control quantity based on the nonlinear dynamic model of the UAV, and combine it with feedback from the real-time sensing device. Calculate the feedback control quantity through the proportional-integral-derivative control law. Fuse the feedforward control quantity and the feedback control quantity to generate the final control quantity of each actuator that can accurately achieve the desired motion and resist interference. Step S43: The final control quantities of each actuator are converted into drive signals and amplified by power, and converted into pulse width modulation signals and analog voltages. The motors are then driven by electronic speed controllers and servo drivers to generate physical actions, and finally the drone actions are synthesized. Among them, drone actions include drones flying, hovering, and maneuvering in space.

6. The UAV swarm control method according to claim 2 or 4, characterized in that, Step S5 specifically includes the following sub-steps: Step S51: Perceive and process the environmental changes and changes in the drone's own state caused by the drone's actions. Use the onboard sensing device to perceive the new changes in the drone's own pose and speed caused by the execution of local control commands, as well as the changes in environmental factors, and generate new raw state data in real time. Step S52: Perform error correction and filtering on the new raw state data. Use the same data processing flow and model as in step S12 to clean, correct and filter the perceived new data in real time to prepare new effective state data for decision-making in the next control cycle. Step S53: Format and timestamp the new valid state data. Using the same standardized process as step S13, add a new timestamp to the new valid state data and encapsulate it to generate updated local state information that serves as input for the next control cycle and is used to evaluate the effect of action execution.

7. The UAV swarm control method according to claim 3, characterized in that, The core architecture of the practical Byzantine fault-tolerant algorithm in step S24 is as follows: A three-phase consensus protocol is adopted, including a pre-preparation phase, a preparation phase, and a submission phase; During the pre-preparation phase, the master node of the current view assigns a sequence number to the received client request and broadcasts a pre-preparation message. During the preparation phase, each backup node, after verifying the validity of the pre-preparation message, broadcasts the preparation message to the entire network and collects preparation messages from other nodes. During the commit phase, once a node has collected valid preparation messages from more than two-thirds of the nodes for the same view and sequence number, it broadcasts a commit message. After receiving more than two-thirds of the valid commit messages, it commits the temporary global state view as globally consistent state information stored locally and executes the corresponding state machine application.

8. The UAV swarm control method according to claim 3, characterized in that, The parameters of the Byzantine fault-tolerant algorithm used in step S24 include the view number v, the current primary node identifier primary_id, the latest stable checkpoint sequence number stable_checkpoint_n, the current consensus sequence number n, and the total number of nodes in the cluster N and the maximum number of fault-tolerant nodes f; Where f satisfies N≥3f+1; during runtime, the algorithm also maintains an acknowledgment set containing 2f+1 nodes to record the message collection status at each stage. When a valid acknowledgment for the same message is received from more than two-thirds of the nodes in the set, it is considered that the message has reached consensus in this stage.

9. The UAV swarm control method according to claim 4, characterized in that, The decision function in step S32 is expressed as: u_i(t)=π(s_global(t),g,τ_i,C_{dynamic},C_{static}) in: u_i(t) represents the local control command generated by UAV i at time t; π(·) represents a pre-defined decision algorithm model, which is a mapping function based on optimization theory, rule reasoning, or deep neural networks; s_global(t) represents the globally consistent state information reached at time t, which includes the state of all members of the cluster; g represents the preset cluster-level task objective, including formation parameters, target point coordinates, and coverage area; τ_i represents the role of drone i in the swarm, including leader, follower, and scout; C_{dynamic} represents the set of dynamic constraints, including real-time obstacle avoidance rules, minimum safe distance from neighboring drones, and communication range limitations; C_{static} represents the set of static constraints, including the UAV's own dynamic limits, energy constraints, and mission area boundaries.

10. A drone swarm control system, characterized in that, The unmanned aerial vehicle (UAV) swarm control method as described in any one of claims 1-9, wherein the UAV swarm control system comprises: The state awareness module is used to access the raw measurement data of each UAV, collect and standardize the raw measurement data, and output standardized local state information of each UAV. The state synchronization module is used to receive standardized local state information, perform distributed full-state synchronization processing on the standardized local state information, and output globally consistent state information. The intelligent decision-making module is used to receive globally consistent status information, perform local intelligent decision-making processing on the globally consistent status information, and output local control commands. The control execution module is used to receive local control commands, perform execution control processing on the local control commands, and output UAV action commands; The feedback evaluation module is used to process the feedback based on the actual effect of the drone's action commands, generate updated local status information, and send the updated local status information to the status synchronization module.

Images