Unmanned aerial vehicle cluster intelligent cooperative control method

Abstract

The invention discloses an unmanned aerial vehicle cluster intelligent cooperative control method, and the method comprises the steps: optimizing a network topology through heterogeneous unmanned aerial vehicle cluster dynamic networking and a dynamic clustering algorithm, and guaranteeing the reliability of a communication link; a layered hybrid decision architecture is designed to improve the task allocation rationality and the dynamic adaptability of the unmanned aerial vehicle cluster; a distributed control strategy network is trained by using a multi-agent near-end strategy optimization MA-PPO algorithm, and unmanned aerial vehicle cluster behavior collaboration is ensured in combination with space-time consistency constraint; an asynchronous incremental consensus protocol AICP is provided, the data transmission amount is reduced, and topology reconstruction is accelerated; real-time three-dimensional environment reconstruction and dynamic threat prediction are realized based on a neural radiation field NeRF technology; a lightweight anti-interference communication middleware is developed, and the instruction transmission stability is enhanced by adopting a space-time coding diversity technology. The method solves the problems of high delay of centralized control of the unmanned aerial vehicle cluster, poor convergence of a distributed algorithm and the like, and is suitable for high-dynamic task scenes such as urban street battle and complex terrain search.

Description

Technical Field

[0001] The present invention relates to the fields of unmanned aerial vehicle (UAV) cluster control and computer vision, and in particular to an intelligent collaborative control method for UAV clusters. Background Art

[0002] Drone swarm collaborative control technology, with its strong mission flexibility and high survivability, has become a key enabler for modern unmanned combat and emergency response. In the military, swarm systems must complete complex missions such as coordinated reconnaissance, saturation strikes, and electronic countermeasures in highly dynamic environments. These include coordinated multi-target capture in urban warfare and distributed jamming suppression in complex electromagnetic environments. These missions place extreme demands on the swarm's autonomous decision-making capabilities, real-time response speed, and survivability. Traditional centralized control architectures pose a single point of failure risk, and the communication overhead caused by global information exchange leads to significant system latency. In highly adversarial environments, the average probability of mission interruption is as high as 42%.

[0003] The current mainstream distributed collaborative control methods face three major technical bottlenecks:

[0004] First, they lack adaptability to dynamic environments. Existing algorithms (such as those based on consistency theory and virtual navigator control strategies) rely on pre-set environmental models. They require manual re-tuning when encountering unexpected obstacles or member damage, and topology reconstruction takes over five seconds, making it difficult to meet the millisecond-level response requirements of highly dynamic battlefields.

[0005] Second, there's a significant conflict between collaborative efficiency and safety. While traditional reinforcement learning algorithms (such as MADDPG) can improve autonomy, they require frequent exchange of gradient information, increasing communication overhead by 300% and causing critical instruction packet loss rates exceeding 25% in electronic interference environments.

[0006] Third, heterogeneous swarm collaboration capabilities are lacking. Existing control frameworks are mostly designed for homogeneous drones. When a swarm includes heterogeneous platforms such as reconnaissance, attack, and relay platforms, the lack of a unified value function modeling method can cause collaborative effectiveness to drop by over 60%.

[0007] To address the above issues, hybrid collaborative architectures proposed in recent years (such as federated reinforcement learning and layered consensus protocols) have partially alleviated the computation-communication contradiction, but they still have two major flaws:

[0008] First, the disconnect between the environmental perception and decision-making loop: The 3D reconstruction output by the situational awareness module requires coordinate conversion before it can be input into the control algorithm, introducing a process delay of up to 200ms.

[0009] Second, the anti-interference strategy is static: the existing communication fault tolerance mechanism (such as retransmission and frequency hopping) adopts a fixed threshold strategy, and the spectrum switching success rate is less than 50% in the intelligent electromagnetic interference environment.

[0010] Therefore, it is meaningful to develop an intelligent control system that integrates real-time environmental understanding, dynamic anti-interference optimization and heterogeneous platform collaboration, which can not only break through the adaptability limits of existing algorithms in complex scenarios, but also build the next generation of drone cluster systems with strong battlefield survivability through cross-domain collaborative innovation at the algorithm level. Summary of the Invention

[0011] Aiming at the problems of poor dynamic adaptability, high communication overhead, and difficulty in heterogeneous collaboration in existing UAV swarm collaborative control algorithms, the present invention proposes an intelligent collaborative control method for UAV swarms to solve the problem of cluster state synchronization lag in highly dynamic environments.

[0012] This application achieves the above effects through the following technical solutions: a method for intelligent collaborative control of drone clusters, the method comprising the following steps:

[0013] S1 builds a heterogeneous drone cluster consisting of reconnaissance nodes, attack nodes, and relay nodes. The reconnaissance node is equipped with a FLIR Boson 640 infrared thermal imager and a Hesai PandarXT lidar for environmental perception; the attack node is equipped with a DJI M300 RTK platform and an NVIDIA Jetson AGX Orin computing unit for executing attack missions; the relay node is equipped with an Ettus USRP B210 software radio that supports dual-band communication to ensure the stability of the communication link. A dynamic clustering algorithm is used to optimize the network topology, and cluster heads are dynamically elected based on indicators such as signal strength, remaining power, computing load, and node spacing to ensure that the communication link has redundant paths. A hierarchical hybrid decision-making architecture is designed, including a global task planning layer and a local autonomous control layer, to improve the rationality of task allocation and the dynamic adaptability of the drone cluster. Spectrum sensing technology optimized by meta-learning is used to establish an anti-interference communication link, dynamically select the optimal frequency band, and provide a reliable communication foundation for subsequent collaborative control;

[0014] S2 is based on the heterogeneous UAV cluster constructed in S1 and its dynamic networking topology. It uses the Multi-Agent Proximal Policy Optimization (MA-PPO) algorithm to train the distributed control strategy network of the local autonomous control layer. It also proposes an asynchronous incremental consensus protocol (AICP) that only transmits incremental data of key state variables. In addition, it builds a collaborative situational awareness model based on Neural Radiance Field (NeRF) technology. Multiple UAVs jointly input multi-view RGB images and LiDAR point clouds to achieve real-time 3D environment reconstruction and dynamic threat prediction, providing accurate environmental information for the cluster's collaborative decision-making.

[0015] S3 develops lightweight communication middleware based on the distributed collaborative control strategy and dynamic environment perception results generated in S2, and uses space-time coding diversity technology (time domain + frequency domain + spatial domain redundancy) to improve the reliability of command transmission. A dynamic spectrum perception module is designed to monitor changes in the electromagnetic environment in real time. When the interference intensity exceeds the set threshold, the adaptive frequency hopping mechanism is triggered. At the same time, a flexible task migration mechanism is established. When a node fails or communication is interrupted, the health assessment model is used to reallocate tasks based on the computing power and location of the remaining drones to ensure an overall task completion rate of ≥85%. The Hungarian algorithm is used to solve the optimal allocation for the position reallocation task to ensure the task continuity of the cluster in complex environments.

[0016] Building on S3, S4 establishes an unsupervised federated incremental learning system, updating the global control strategy through a momentum-weighted aggregation algorithm to reduce model drift. In the momentum-weighted aggregation expression, the momentum vector is used to store historical gradient information, and the momentum coefficient controls the weight of historical gradient information in the current update. A differentiated parameter transmission strategy is adopted, synchronizing only key network layer parameters, reducing communication bandwidth usage by 50%. Multi-level collaborative performance evaluation is performed, covering metrics such as mission completion rate, latency, and energy consumption, and dynamically adjusting control algorithm parameter weights based on different combat scenarios. Cluster collaborative performance indicators are monitored and optimized in real time, outputting optimal cluster behavior instructions, further enhancing the system's intelligence and adaptability.

[0017] Furthermore, the S1 is specifically:

[0018] S11 builds a heterogeneous drone cluster, where reconnaissance nodes are equipped with infrared thermal imagers and lidar for real-time perception of complex environments; attack nodes are equipped with high-performance computing units and flight platforms, possessing powerful attack capabilities; and relay nodes support multi-band communication through software-defined radio equipment, ensuring the stability and redundancy of communication links. A dynamic clustering algorithm is used to dynamically elect cluster heads based on indicators such as signal strength, remaining battery power, computing load, and node spacing. Each cluster head maintains at least two backup paths.

[0019] S12, a hierarchical hybrid decision-making architecture is designed. The global task planning layer uses a federated reinforcement learning algorithm to comprehensively consider task urgency, resource requirements, and environmental factors to generate a task priority sequence, ensuring the rationality and efficiency of task allocation. The local autonomous control layer uses dynamic topological potential field theory to adjust the drone's motion strategy in real time to avoid collisions and resource conflicts, thereby improving the dynamic adaptability of the cluster.

[0020] S13 uses spectrum sensing technology optimized through meta-learning to monitor the electromagnetic environment in real time and dynamically select the optimal frequency band to avoid interference and ensure the stability and reliability of the communication link. This technology uses historical signal-to-noise ratio and frequency band utilization to optimize the frequency band selection strategy, improving the cluster's communication capabilities in complex electromagnetic environments.

[0021] Furthermore, the S2 is specifically:

[0022] S21 uses the Multi-Agent Proximal Policy Optimization (MA-PPO) algorithm, which effectively improves the autonomous decision-making capabilities of drone swarms through a multi-layer perceptron (MLP) network structure, combined with discount factors, GAE parameters, and large-batch training. The introduction of spatiotemporal consistency constraints ensures the coordination of drones' behaviors in time and space, avoiding conflicts and uncoordinated behaviors, and improving the overall performance of the swarm.

[0023] S22 proposes the Asynchronous Incremental Consensus Protocol (AICP), which significantly reduces communication overhead by transmitting only incremental data of key state variables rather than the complete state information. This protocol combines bitwise exclusive OR (XOR) operations and entropy coding technology to achieve efficient state synchronization, significantly improving the cluster's responsiveness in highly dynamic environments.

[0024] S23, based on Neural Radiance Field (NeRF) technology, builds a collaborative situational awareness model, and realizes real-time three-dimensional environment reconstruction and dynamic threat prediction through multi-perspective RGB images and lidar point cloud data jointly input by multiple drones.

[0025] Furthermore, the S3 is specifically:

[0026] S31, developed lightweight communication middleware, uses space-time coding diversity technology, through time-domain Hamming code, space-domain MIMO technology and frequency-domain pseudo-random code diversity, significantly improving the reliability of command transmission and ensuring the stability of cluster communication in complex environments;

[0027] S32, a dynamic spectrum sensing module, monitors electromagnetic environment changes in real time and automatically triggers an adaptive frequency hopping mechanism when interference exceeds a set threshold. This mechanism uses a decision tree with a depth of 5 layers and a branching factor of 3 to quickly select the optimal frequency band, keeping spectrum switching delays within 100 milliseconds, effectively avoiding electromagnetic interference.

[0028] S33, when a node fails or communication is interrupted, activates the elastic task migration mechanism. Through the health assessment model, it dynamically evaluates the communication capabilities, remaining power and computing load of the remaining drones, and uses the Hungarian algorithm to reallocate tasks, ensuring that the cluster can still maintain a task completion rate of more than 85% when some nodes fail, significantly improving the robustness and task continuity of the cluster.

[0029] Furthermore, the S4 is specifically:

[0030] S41, establishes a federated incremental learning system, which uses a momentum-weighted aggregation algorithm to combine historical gradient information and current updates to dynamically adjust the global control strategy, reduce model drift, and improve system stability and convergence speed;

[0031] S42 adopts a differentiated parameter transmission strategy and only synchronizes key network layer parameters, significantly reducing communication bandwidth usage, lowering communication overhead, and improving system communication efficiency;

[0032] S43 performs multi-level collaborative effectiveness evaluations, monitoring the cluster's collaborative performance indicators in real time, including task completion rate, latency, and energy consumption. It dynamically adjusts control algorithm parameter weights based on different combat scenarios to ensure optimal system performance under different mission requirements, further enhancing the cluster's intelligence and adaptability.

[0033] The beneficial effects of the present invention are:

[0034] 1. By building a heterogeneous drone swarm and dynamic networking, optimizing the network topology using a dynamic clustering algorithm, and combining it with meta-learning-optimized spectrum sensing technology, this approach addresses the strong single-point dependency and poor anti-interference capabilities inherent in traditional drone swarm communications. The dynamic clustering algorithm dynamically selects cluster heads based on signal strength, remaining battery life, computational load, and node spacing. Each cluster head maintains at least two backup paths, significantly improving communication reliability and resilience. Meta-learning-optimized spectrum sensing technology monitors the electromagnetic environment in real time and dynamically selects the optimal frequency band, achieving an initial communication link success rate exceeding 95%. The application of these technologies has reduced network topology reconstruction time from 5 seconds to 500 milliseconds, improving communication link reliability by 40%.

[0035] 2. Based on a distributed collaborative control strategy, the Multi-Agent Proximal Policy Optimization (MA-PPO) algorithm and spatiotemporal consistency constraints ensure coordinated swarm behavior. Combined with the real-time 3D environment reconstruction capabilities of Neural Radiance Field (NeRF) technology, this approach addresses the limited adaptability of existing methods in dynamic and complex environments. The MA-PPO algorithm utilizes a multi-layer perceptron (MLP) network structure, combined with discount factors, GAE parameters, and large-batch training, to effectively enhance the autonomous decision-making capabilities of drone swarms. Spatiotemporal consistency constraints ensure coordinated behavior across time and space, avoiding conflicts and uncoordinated behavior. Neural Radiance Field (NeRF) technology utilizes multi-view RGB images and LiDAR point cloud data from multiple drones to achieve real-time 3D environment reconstruction and dynamic threat prediction, with reconstruction accuracy within 0.5 meters. The application of these technologies has increased the swarm's mission completion rate by 40% in densely obstructed environments and reduced collaborative response latency to 150 milliseconds.

[0036] 3. The system utilizes space-time coding diversity technology and a flexible task migration mechanism. This triple guarantee, achieved through time-domain Hamming coding, spatial MIMO optimization, and frequency-domain dynamic spread spectrum, complemented by an intelligent spectrum management module, addresses communication stability issues in high-interference environments. The lightweight communication middleware significantly improves command transmission reliability through the use of time-domain Hamming coding, spatial MIMO, and frequency-domain pseudo-random code diversity. Even with a 30% packet loss rate, command delivery rates remain above 90%. The dynamic spectrum sensing module monitors electromagnetic environment changes in real time. When interference exceeds a set threshold, it automatically triggers adaptive frequency hopping, keeping spectrum switching latency to under 100 milliseconds. These technologies reduce spectrum switching latency to under 50 milliseconds, significantly improving cluster communication stability in complex electromagnetic environments.

[0037] 4. The federated incremental learning system, combined with a momentum-weighted aggregation algorithm and a differentiated parameter transmission strategy, addresses the model drift problem in traditional distributed learning. Using a momentum-weighted aggregation algorithm, the federated incremental learning system combines historical gradient information with current updates to dynamically adjust the global control strategy, reducing model drift and improving system stability and convergence speed. The differentiated parameter transmission strategy significantly reduces communication bandwidth usage and overhead by synchronizing only key network layer parameters, ultimately reducing bandwidth usage by 50%. Furthermore, a multi-dimensional performance evaluation system enables autonomous optimization of control parameters, further enhancing the system's stability during long-term missions.

[0038] 5. The system innovatively combines a hierarchical decision-making architecture with flexible task migration. The global layer updates task priorities every 5 seconds, and the local layer adjusts movement strategies every 100 milliseconds. Combined with a node health assessment model, this system maintains a task completion rate exceeding 85% even when some nodes fail. This hierarchical decision-making architecture ensures rational and efficient task allocation through the coordinated work of the global task planning layer and the local autonomous control layer. The flexible task migration mechanism uses a health assessment model to dynamically assess the communication capabilities, remaining battery power, and computational load of the remaining drones. It then uses the Hungarian algorithm to reallocate tasks, significantly improving the swarm's mission continuity and survivability in the event of partial node failure. The application of these technologies has significantly enhanced the swarm's survivability and task completion rate in battlefield environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The figure is a flow chart of a method for intelligent collaborative control of UAV swarms;

[0040] Figure 2 A flowchart of the method for building a dynamic networking and hierarchical decision-making architecture for heterogeneous clusters;

[0041] Figure 3A flow chart of the distributed collaborative control strategy generation and dynamic environment perception method;

[0042] Figure 4 A flowchart of an anti-interference communication and flexible task scheduling method;

[0043] Figure 5 Schematic diagram of the process for online learning optimization and collaborative effectiveness evaluation. DETAILED DESCRIPTION

[0044] In order to facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and drawings. The contents mentioned in the embodiments are not intended to limit the present invention.

[0045] This application proposes a method for intelligent collaborative control of drone swarms, which realizes autonomous collaboration of swarms through a hierarchical hybrid decision-making architecture; adopts a multi-agent proximal policy optimization (MA-PPO) algorithm under spatiotemporal consistency constraints, combined with a dynamic topological potential field to generate local optimal control instructions; proposes an asynchronous incremental consensus protocol (AICP) to solve the problem of cluster state synchronization lag in highly dynamic environments; designs a collaborative situational awareness model based on neural radiant fields (NeRF) to achieve real-time three-dimensional environment reconstruction and threat prediction; develops lightweight anti-interference communication middleware, and improves the reliability of control instruction transmission through spatiotemporal coding diversity technology. This method breaks through the bottlenecks of high latency of traditional centralized control, poor convergence of distributed algorithms, and weak adaptability to dynamic environments, and provides a highly robust swarm control solution for tasks such as urban street fighting and complex terrain search.

[0046] Specifically, the present invention provides a method for intelligent collaborative control of drone clusters, such as Figure 1 As shown, the following steps are included:

[0047] S1 builds a heterogeneous drone cluster consisting of reconnaissance nodes, attack nodes, and relay nodes. The reconnaissance node is equipped with a FLIR Boson 640 infrared thermal imager and a Hesai PandarXT lidar for environmental perception; the attack node is equipped with a DJI M300 RTK platform and an NVIDIA Jetson AGX Orin computing unit for executing attack missions; the relay node is equipped with an Ettus USRP B210 software radio that supports dual-band communication to ensure the stability of the communication link. A dynamic clustering algorithm is used to optimize the network topology, and cluster heads are dynamically elected based on indicators such as signal strength, remaining power, computing load, and node spacing to ensure that the communication link has redundant paths. A hierarchical hybrid decision-making architecture is designed, including a global task planning layer and a local autonomous control layer, to improve the rationality of task allocation and the dynamic adaptability of the drone cluster. Anti-interference communication links are established through meta-learning optimized spectrum sensing technology, and the optimal frequency band is dynamically selected to provide a reliable communication foundation for subsequent collaborative control, such as Figure 2 As shown;

[0048] S11: Build a heterogeneous drone swarm consisting of reconnaissance nodes, attack nodes, and relay nodes. The reconnaissance nodes are equipped with a FLIR Boson 640 thermal imager (640×512 resolution, 30Hz frame rate) and a Hesai PandarXT lidar (10Hz, 300m ranging). The attack nodes are equipped with a DJI M300 RTK platform and an NVIDIA Jetson AGX Orin compute unit (32TOPS). The relay nodes are equipped with an Ettus USRP B210 software-defined radio (SDR) supporting 2.4GHz / 5.8GHz dual-band communication. A dynamic clustering algorithm is used to optimize the network topology and ensure communication link redundancy. The clustering indicator weights are distributed as follows:

[0049] W cluster =0.4RSSI+0.3E batt +0.2C load +0.1D inter

[0050] Where RSSI is the signal strength, E batt is the remaining power, C load To calculate the load, D inter is the node spacing.

[0051] The cluster head election cycle is 60 seconds, and the link redundancy requirement is: each cluster head maintains ≥ 2 backup paths;

[0052] S12: A hierarchical hybrid decision-making architecture was designed. The global task planning layer uses federated reinforcement learning (FRL) to generate task priorities, and the local autonomous control layer adjusts motion strategies in real time based on a dynamic topological potential field. In the global task planning layer, the FRL update cycle is 5 seconds, and the task priority matrix dimension is N×M (N number of nodes, M number of tasks). In the local control layer, the dynamic topological potential field update frequency is 10Hz, and the potential field parameters are set as follows: repulsion coefficient α = 0.8, attraction coefficient β = 0.2.

[0053] S13, establishes an interference-resistant communication link through spectrum sensing technology optimized by meta-learning, and dynamically selects the optimal frequency band. The input feature dimension is 128 dimensions (including historical signal-to-noise ratio, frequency band utilization, etc.), the meta-training cycle is 1000 iterations, and the frequency band switching response time is ≤50ms.

[0054] S2 is based on the heterogeneous UAV cluster constructed in S1 and its dynamic networking topology. It uses the multi-agent proximal policy optimization (MA-PPO) algorithm to train the distributed control strategy network of the local autonomous control layer. At the same time, it proposes an asynchronous incremental consensus protocol (AICP) to transmit only incremental data of key state variables. In addition, it builds a collaborative situational awareness model based on the neural radiance field (NeRF) technology. Multiple UAVs jointly input multi-view RGB images and lidar point clouds to achieve real-time three-dimensional environment reconstruction and dynamic threat prediction, providing accurate environmental information for the collaborative decision-making of the cluster, such as Figure 3 As shown;

[0055] Incorporating spatiotemporal consistency constraints: Introducing spatiotemporal consistency constraints into the MA-PPO algorithm ensures the coordinated behavior of the drone swarm. This integration is uncommon in multi-agent reinforcement learning and is an innovation tailored to the specific needs of drone swarm collaborative control.

[0056] Application to the local autonomous control layer: The MA-PPO algorithm is applied to the local autonomous control layer, combined with dynamic topological potential field theory to generate real-time motion control instructions. This combination is innovative in drone swarm control and can better adapt to dynamic environments.

[0057] S21 uses the multi-agent proximal policy optimization (MA-PPO) algorithm to train the distributed control strategy, introducing spatiotemporal consistency constraints to ensure the coordination of cluster behavior. The network structure of the MA-PPO algorithm includes a 3-layer MLP (256-128-64) policy network and a 3-layer MLP (256-128-64) value network. The training parameters include: discount factor γ = 0.95, GAE parameter λ = 0.9 and batch size 1024.

[0058] S22 proposes an asynchronous incremental consensus protocol (AICP), which only transmits critical state incremental data, reducing the topology reconstruction time to less than 500ms. The state incremental encoding of AICP is shown in the following formula:

[0059] △S t =Huffman(S t ⊕S t-1 )

[0060] Where S t is the state vector, which represents the complete state (position / power, etc.) at the current moment, S t-1 is the complete state at the previous moment, ⊕ is the bitwise exclusive OR (XOR) operator, and Huffman represents entropy coding based on frequency statistics.

[0061] The compression rate is ≥70%. In the synchronization mechanism, the key states are position and power (100ms synchronization), and the non-critical state is sensor reading (1s synchronization).

[0062] S23, based on the Neural Radiance Field (NeRF) to build a collaborative situational awareness model, multiple UAVs jointly complete real-time 3D environment reconstruction and dynamic threat prediction, including the input multi-view RGB image (20 million pixels) and LiDAR point cloud (10Hz). The model's network structure includes 64-dimensional position encoding and 8 layers of 256-dimensional MLP hidden layers, with a reconstruction accuracy of ≤0.5m.

[0063] S3 develops lightweight communication middleware based on the distributed collaborative control strategy and dynamic environment perception results generated in S2, and uses space-time coding diversity technology (time domain + frequency domain + spatial domain redundancy) to improve the reliability of command transmission. Design a dynamic spectrum perception module to monitor changes in the electromagnetic environment in real time. When the interference intensity exceeds the set threshold, trigger the adaptive frequency hopping mechanism. At the same time, establish a flexible task migration mechanism. When a node fails or communication is interrupted, based on the computing power and position of the remaining drones, use the health assessment model to reallocate tasks to ensure that the overall task completion rate is ≥85%. The Hungarian algorithm is used to solve the optimal allocation for the position reallocation task to ensure the continuity of the cluster's tasks in complex environments, such as Figure 4 As shown;

[0064] S31 develops lightweight communication middleware, using space-time coding diversity technology (time domain + frequency domain + spatial domain redundancy) to improve command transmission reliability. The time domain coding uses (7, 4) Hamming code with 3-bit error correction capability, the spatial domain coding uses a MIMO 2×2 configuration with a maximum ratio combining algorithm, and the pseudo-random code period in the frequency domain coding is 255 chips.

[0065] S32, design a dynamic spectrum sensing module to monitor the electromagnetic interference intensity in real time and trigger the adaptive frequency hopping mechanism, where the interference detection threshold is P int =-85dBm, the depth of the frequency hopping decision tree is 5 layers, and the branching factor is 3;

[0066] S33, when a node fails or communication is interrupted, the elastic task migration mechanism is activated to redistribute tasks based on the computing power and location of the remaining drones to ensure that the overall task completion rate is ≥85%. The elastic task migration mechanism uses a health assessment model as follows:

[0067] H=0.6Q comm +0.3E batt +0.1C load

[0068] Where Q comm For the communication capability of the UAV, E batt is the remaining power, Cload To calculate the load.

[0069] The Hungarian algorithm is used to solve the optimal allocation for the location reallocation task.

[0070] Based on S3, S4 establishes an unsupervised federated incremental learning system, updates the global control strategy through the momentum weighted aggregation algorithm, and reduces model drift. In the momentum weighted aggregation expression, the momentum vector is used to store historical gradient information, and the momentum coefficient controls the weight of the historical gradient information in the current update. A differentiated parameter transmission strategy is adopted to synchronize only key network layer parameters, reducing communication bandwidth occupancy by 50%. Perform multi-level collaborative performance evaluation, covering indicators such as task completion rate, latency, and energy consumption, and dynamically adjust the control algorithm parameter weights according to different combat scenarios. Real-time monitoring and optimization of cluster collaborative performance indicators, output of optimal cluster behavior instructions, and further improve the intelligence level and adaptability of the system, such as Figure 5 shown.

[0071] S41, establish a federated incremental learning system and update the global control strategy through momentum weighted aggregation to reduce model drift. The momentum weighted aggregation expression is as follows:

[0072] θ t =βθ t-1 +(1-β)△θ

[0073] Where θ t-1 is the momentum vector at the t-1th iteration, which is used to store historical gradient information. △θ is the global model momentum, and β is the momentum coefficient, which is used to control the weight of historical gradient information in the current update, where β = 0.9;

[0074] S42, adopts a differentiated parameter transmission strategy and only synchronizes key network layer parameters;

[0075] S43, perform multi-level collaborative effectiveness evaluation (including task completion rate Latency Energy consumption E sum =∑E i and other indicators), dynamically adjust the control algorithm parameters (the weight distribution is 0.6:0.2:0.2 in combat scenarios and 0.3:0.4:0.3 in reconnaissance scenarios), and output the optimal cluster behavior instructions.

[0076] The method provided in this application constructs a heterogeneous drone cluster including reconnaissance, attack, and relay nodes, adopts a hierarchical hybrid decision-making architecture (global task planning + local autonomous control), and optimizes the cluster motion strategy in combination with a dynamic topological potential field; realizes distributed collaborative decision-making through the multi-agent proximal policy optimization (MA-PPO) algorithm, and introduces spatiotemporal consistency constraints to ensure the coordination of cluster behavior; proposes an asynchronous incremental consensus protocol (AICP) to solve the problem of cluster state synchronization lag in highly dynamic environments, shortening the topology reconstruction time to less than 500ms; designs a collaborative situational awareness model based on neural radiant fields (NeRF) to achieve real-time three-dimensional environment reconstruction and dynamic threat prediction, improves adaptability to complex scenarios, increases the task completion rate by 40% in dense obstacle scenarios, and reduces the collaborative response delay to the 150ms level; develops lightweight anti-interference communication middleware, adopts spatiotemporal coding diversity technology to enhance the reliability of control command transmission, and can still maintain a command delivery rate of more than 90% in an environment with a 30% packet loss rate; establishes a flexible task migration mechanism, and automatically adjusts the task allocation strategy when some nodes fail to ensure that the overall task completion rate of the cluster is not less than 85%.

[0077] Compared with traditional methods, this solution increases the cluster task completion rate by 40% in dense obstacle environments, reduces the collaborative response delay to 150ms, reduces the communication load by 50%, and supports efficient collaboration of heterogeneous drones (reconnaissance, attack, and relay types). It is suitable for highly dynamic and highly confrontational scenarios such as urban street fighting, complex terrain search, and electronic countermeasures, significantly improving the battlefield survivability and combat effectiveness of drone clusters.

[0078] The present invention has many specific application paths. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements can be made without departing from the principles of the present invention. These improvements should also be considered as the scope of protection of the present invention.

Claims

1. A method for intelligent collaborative control of drone swarms, characterized in that: The method comprises the following steps: S1 first builds a heterogeneous drone cluster consisting of reconnaissance nodes, attack nodes, and relay nodes; A dynamic clustering algorithm is used to optimize the network topology. Cluster heads are dynamically selected based on signal strength, remaining power, computing load, and node spacing to ensure redundant communication paths. Design a hierarchical hybrid decision-making architecture, including a global task planning layer and a local autonomous control layer, to improve the rationality of task allocation and the dynamic adaptability of drone swarms; Spectrum sensing technology optimized through meta-learning establishes an anti-interference communication link and dynamically selects the optimal frequency band, providing a reliable communication foundation for subsequent collaborative control. S2 trains the distributed control strategy network of the local autonomous control layer based on the heterogeneous drone cluster and its dynamic networking topology constructed in S1. At the same time, it adopts the asynchronous incremental consensus protocol AICP to transmit only the incremental data of key state variables. In addition, a collaborative situational awareness model is built based on Neural Radiance Field (NeRF) technology. Multiple drones jointly input multi-view RGB images and LiDAR point clouds to achieve real-time 3D environment reconstruction and dynamic threat prediction, providing accurate environmental information for collaborative decision-making in the swarm. S3 develops lightweight communication middleware based on the distributed collaborative control strategy and dynamic environment perception results generated in S2, and uses space-time coding diversity technology for command transmission; A dynamic spectrum sensing module is designed to monitor electromagnetic environment changes in real time. When the interference intensity exceeds a set threshold, an adaptive frequency hopping mechanism is triggered. Furthermore, a flexible task migration mechanism is established. When a node fails or communication is interrupted, tasks are reallocated using a health assessment model based on the computing power and location of the remaining drones, ensuring an overall task completion rate of ≥85%. The Hungarian algorithm is used to find the optimal allocation for position reallocation tasks, ensuring task continuity in complex environments. S4 establishes an unsupervised federated incremental learning system and updates the global control strategy through a momentum-weighted aggregation algorithm.

2. The method for intelligent collaborative control of drone swarms according to claim 1, characterized in that: The S1 is specifically: S11 builds a heterogeneous drone cluster, where reconnaissance nodes are equipped with infrared thermal imagers and lidar for real-time perception of complex environments; attack nodes are equipped with high-performance computing units and flight platforms; and relay nodes support multi-band communication through software-defined radio equipment. A dynamic clustering algorithm is used to dynamically elect cluster heads based on indicators such as signal strength, remaining battery power, computing load, and node spacing. Each cluster head maintains at least two backup paths. S12, a hierarchical hybrid decision-making architecture is designed. The global task planning layer uses a federated reinforcement learning algorithm to comprehensively consider the urgency of the task, resource requirements, and environmental factors to generate a task priority sequence. The local autonomous control layer uses dynamic topological potential field theory to adjust the drone's motion strategy in real time to avoid collisions and resource conflicts, thereby improving the dynamic adaptability of the cluster. S13, through meta-learning optimized spectrum sensing technology, monitors the electromagnetic environment in real time, dynamically selects the optimal frequency band, avoids interference, and ensures the stability and reliability of the communication link.

3. The method for intelligent collaborative control of drone swarms according to claim 1, characterized in that: The S2 is specifically: S21 uses a multi-agent proximal strategy to optimize the MA-PPO algorithm, using a multi-layer perceptron (MLP) network structure, combined with discount factors, GAE parameters, and large-batch training; The spatiotemporal consistency constraint is introduced into the multi-agent proximal policy optimization MA-PPO algorithm to ensure the behavior coordination of UAVs in time and space; S22 uses the asynchronous incremental consensus protocol AICP to transmit only incremental data of key state variables; S23 builds a collaborative situational awareness model based on Neural Radiation Field (NeRF) technology, and realizes real-time three-dimensional environment reconstruction and dynamic threat prediction through multi-perspective RGB images and lidar point cloud data jointly input by multiple drones.

4. The method for intelligent collaborative control of drone swarms according to claim 1, characterized in that: The S3 is specifically: S31, developed lightweight communication middleware, uses space-time coding diversity technology, through time-domain Hamming code, space-domain MIMO technology and frequency-domain pseudo-random code diversity, significantly improving the reliability of command transmission and ensuring the stability of cluster communication in complex environments; S32, a dynamic spectrum sensing module, monitors electromagnetic environment changes in real time and automatically triggers an adaptive frequency hopping mechanism when interference exceeds a set threshold. This mechanism uses a decision tree with a depth of 5 layers and a branching factor of 3 to quickly select the optimal frequency band, keeping spectrum switching delays within 100 milliseconds, effectively avoiding electromagnetic interference. S33, when a node fails or communication is interrupted, the elastic task migration mechanism is activated. Through the health assessment model, the communication capabilities, remaining power and computing load of the remaining drones are dynamically evaluated, and the Hungarian algorithm is used to reallocate tasks to ensure that the cluster can still maintain a task completion rate of more than 85% when some nodes fail, significantly improving the robustness and task continuity of the cluster.

5. The method for intelligent collaborative control of a drone swarm according to claim 1, characterized in that: The S4 is specifically: S41, establishes a federated incremental learning system, which uses a momentum-weighted aggregation algorithm to combine historical gradient information and current updates to dynamically adjust the global control strategy, reduce model drift, and improve system stability and convergence speed; S42 adopts a differentiated parameter transmission strategy and only synchronizes key network layer parameters, significantly reducing communication bandwidth usage, lowering communication overhead, and improving system communication efficiency; S43 performs multi-level collaborative effectiveness evaluations, monitoring the cluster's collaborative performance indicators in real time, including task completion rate, latency, and energy consumption. It dynamically adjusts control algorithm parameter weights based on different combat scenarios to ensure optimal system performance under different mission requirements, further enhancing the cluster's intelligence and adaptability.

6. The method for intelligent collaborative control of drone swarms according to claim 1, characterized in that: The heterogeneous drone cluster includes: Reconnaissance nodes, equipped with a FLIR Boson 640 thermal imager and a Hesai PandarXT lidar for environmental awareness; An attack node, wherein the attack node is equipped with a DJI M300 RTK platform and an NVIDIA Jetson AGX Orin computing unit, and is used to perform attack missions; The relay node is configured with an Ettus USRP B210 software radio, supports dual-band communication, and is used to ensure the stability of the communication link.

7. The method for intelligent collaborative control of drone swarms according to claim 1, characterized in that: In said S4: In the momentum-weighted aggregation expression, the momentum vector is used to store historical gradient information, and the momentum coefficient controls the weight of historical gradient information in the current update; Adopting a differentiated parameter transmission strategy, only synchronizing key network layer parameters, and dynamically adjusting the control algorithm parameter weights according to different combat scenarios; Monitor and optimize cluster collaboration performance indicators in real time, and output optimal cluster behavior instructions.

Images