An ai-driven software-defined unmanned cluster management and control system and method

Abstract

The application discloses an AI-driven software-defined unmanned cluster management and control system and method, and belongs to the technical field of artificial intelligence. The system adopts a three-layer modular architecture, which is sequentially from bottom to top: a software-defined heterogeneous access module completes the bottom layer hardware resource pooling of various heterogeneous unmanned nodes through a hardware abstraction layer (HAL); the AI-driven collaborative scheduling module includes a large model command brain layer and a small model cluster layer, which constitute a layered intelligent architecture, perform layered decomposition and intelligent analysis on a global target task, generate unmanned cluster collaborative scheduling instructions, and distribute the instructions to various heterogeneous unmanned nodes; and a network switching and connectivity management module is integrated with a multi-mode link monitoring and switching submodule, an AI prediction and topology adjustment submodule, and a multi-unit collaborative scheduling implementation submodule. The application realizes technical breakthroughs through four dimensions of access, scheduling, network, and man-machine collaboration, and significantly improves the overall performance and adaptability of the unmanned cluster.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and specifically relates to an AI-driven software-defined unmanned cluster management system and method. Background Technology

[0002] With the widespread application of unmanned equipment such as drones, unmanned vehicles, unmanned ships, and unmanned dogs in various fields, unmanned systems are evolving from single platforms to large-scale, heterogeneous unmanned swarms. Unmanned swarms can accomplish complex tasks that are difficult for a single unit to handle through the collaborative cooperation of multiple unmanned nodes. However, existing unmanned swarm management systems still face the following significant technical challenges in practical applications: 1. Rigid architecture, poor compatibility with heterogeneous equipment: Traditional unmanned swarm management systems often employ a siloed architecture with highly coupled hardware and software. Custom-developed management software and hardware are frequently required for different types (e.g., aerial, ground, underwater) or different manufacturers' unmanned equipment. This monolithic architecture lacks unified hardware interface definitions and software module divisions, resulting in extremely poor system scalability and making it difficult to achieve rapid access and unified management of heterogeneous unmanned swarms.

[0003] 2. Rigid resource allocation and lack of task adaptability: Faced with complex and ever-changing external environments and unexpected tasks, traditional unmanned swarms often rely on preset static rules or heuristic algorithms for scheduling. This approach cannot dynamically allocate computing, storage, and communication resources based on real-time task load and the status of each unmanned node, resulting in slow response and low resource utilization when unmanned swarms perform complex collaborative tasks.

[0004] 3. The communication network is fragile, and connectivity is difficult to guarantee: In highly dynamic and complex mission environments, the topology of unmanned swarms changes frequently, and unmanned nodes are prone to losing contact due to interference or damage. Existing control systems lack dynamic communication network switching mechanisms and topology connectivity maintenance strategies. Once a critical unmanned node or link fails, it can easily lead to the paralysis of the entire unmanned swarm control system.

[0005] In summary, there is an urgent need for an unmanned cluster management system and method that can break down software and hardware barriers, has a high degree of environmental adaptability, and is robust to networks. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide an AI-driven software-defined unmanned swarm management system and method, which solves the technical problems faced by existing unmanned swarm management systems in practical applications, such as rigid architecture, poor adaptability to heterogeneous equipment, rigid resource allocation, lack of task self-adaptation capability, and fragile communication network with difficulty in ensuring connectivity.

[0007] To address the aforementioned technical problems, this invention provides an AI-driven software-defined unmanned cluster management system, the system comprising: The software-defined heterogeneous access module is configured to pool the underlying hardware resources of various heterogeneous unmanned nodes based on the hardware abstraction layer (HAL), and to formulate a globally unified standardized device access interface and a general data encapsulation protocol. The AI-driven collaborative scheduling module interacts bidirectionally with the software-defined heterogeneous access module; the AI-driven collaborative scheduling module is internally layered with a large model command brain layer, a small model cluster layer, and a human-machine collaborative switching unit. The network switching and connectivity management module communicates with the AI-driven collaborative scheduling module and all heterogeneous unmanned nodes, and has built-in multi-mode link monitoring sub-module, AI trajectory prediction sub-module and multi-unit collaborative scheduling sub-module.

[0008] In a further embodiment, the Hardware Abstraction Layer (HAL) is used to encapsulate the underlying drivers of various heterogeneous unmanned nodes, maintain a global device description table, assign a unique logical identifier to each connected unmanned node, and has single-point hardware fault isolation capability. The large model command brain layer integrates multimodal visual processing components, the small model cluster layer is configured with observation model, judgment model, decision model and control model; the human-machine collaboration switching unit is equipped with Mamdani fuzzy inference logic; The multi-unit collaborative scheduling submodule integrates an adaptive particle swarm region coverage algorithm, a convex hull combined quadrilateral Steiner tree topology reconstruction algorithm, and virtual potential field node motion control logic.

[0009] The management method for an AI-driven, software-defined, unmanned cluster management system, as described above, includes the following steps: S1. Call the Hardware Abstraction Layer (HAL) of the software-defined heterogeneous access module to complete the pooling of underlying hardware resources for various heterogeneous unmanned nodes; based on standardized device access interfaces and general data encapsulation protocols, receive the raw data periodically transmitted back by various heterogeneous unmanned nodes, and generate real-time situational information for the entire domain after unified formatting. S2. The AI-driven collaborative scheduling module takes real-time situational information from the entire domain as input and uses the large model command brain layer and the small model cluster layer to complete the intelligent analysis of situation-task coupling and output the initial cluster scheduling plan; combined with the human-machine collaborative switching unit to match and adapt the human-machine autonomous control level, and after correcting the scheduling constraints, the final unmanned cluster collaborative scheduling instruction is generated. S3. Call the network switching and connectivity management module, collect the signal-to-noise ratio of multi-mode communication links in real time and complete the link quality judgment through the built-in multi-mode link monitoring sub-module, use the AI ​​trajectory prediction sub-module to predict the network splitting risk caused by node movement, and execute the regional coverage optimization algorithm and topology reconstruction operation through the multi-unit collaborative scheduling sub-module, and output the node execution command to all heterogeneous unmanned nodes. S4: Receive the running status data returned by all heterogeneous unmanned nodes after executing instructions, and feed it back to the software-defined heterogeneous access module to update the real-time situation information of the entire domain; cyclically execute S1~S3 to achieve continuous closed-loop management and control of the unmanned cluster.

[0010] In a further embodiment, the steps for generating the real-time global situational information are as follows: S101. After the heterogeneous unmanned node is powered on, it periodically sends raw heartbeat data packets through its own communication link. S102. The Hardware Abstraction Layer (HAL) receives heterogeneous format raw heartbeat data and sensor data, and converts them into node status information with a unified format according to a preset encapsulation protocol. S103. Kalman filtering is used to smooth and reduce noise in sensor data and track target position, eliminating random noise in perception and outputting smoothed target position time series data. S104. The Hardware Abstraction Layer (HAL) maintains the global device description table in real time, synchronously updates the device status of newly added access nodes and offline disconnected nodes, and assigns a unique logical addressing identifier to all online nodes. S105. Collect node status information of all online nodes and time-series data of target locations to complete full-domain data fusion and construct a... A global real-time situational awareness matrix is ​​generated, and this matrix is ​​used to push global real-time situational information to an AI-driven collaborative scheduling module; among which, For the number of nodes, For feature dimensions; S106 The Hardware Abstraction Layer (HAL) has a built-in fault isolation mechanism. If a single node hardware failure or communication interruption is detected, the data transmission path of that node will be immediately cut off, without interfering with the data acquisition and global situation generation process of other nodes.

[0011] In a further embodiment, the output process of the initial cluster scheduling scheme is as follows: Build a large model to command the brain, and construct a question-and-answer dataset for fine-tuning instructions; Supervised fine-tuning is performed using LoRA low-rank adaptation, freezing the original parameters of the large model's "brain" and utilizing the loss function. Train a low-rank incremental matrix to obtain a domain-wide model adapted to unmanned cluster management scenarios; The large-scale model of the domain is compressed, multimodal features are fused through a cross-attention mechanism, and the overall target task is interpreted in conjunction with real-time situational information of the entire domain, and high-level scheduling and guidance instructions are output. A large domain model calls smaller model clusters, with each model deployed at the edge forming a hierarchical collaborative architecture: running and observing models utilizes the detection loss function. Conduct environmental situational awareness, output target category, location, and confidence level, and push them to the judgment model; The judgment model receives the perception results and uses the cross-entropy loss function. Conduct threat level assessments, output standardized risk scores, and push them to the decision-making model; The decision-making model receives risk scores and high-level scheduling guidance instructions, and employs a Kido agent deep deterministic policy gradient algorithm and reward function. Solve for the optimal cluster task allocation strategy and output the preliminary cluster scheduling scheme.

[0012] In a further embodiment, the generation process of the final unmanned cluster collaborative scheduling instruction is as follows: The real-time situational awareness matrix is ​​input into the graph neural network threat assessment model to calculate the risk index of each node. Cluster connectivity risk ,like If so, the connectivity control module will be triggered, and the global topology risk assessment results will be output. Integrating the initial cluster scheduling scheme, high-level scheduling guidance instructions, real-time global situational information, and task objectives, the population is iterated through a multi-objective evolutionary algorithm. After selection, crossover, mutation, and Pareto non-dominated sorting, the results were ranked according to task completion. Total energy consumption of the cluster Network resilience Using fitness metrics, we screen for optimal Pareto front scheduling plans; Utilizing the fuzzy inference logic built into the human-machine collaborative switching unit, the task complexity is reduced. Environmental uncertainty and operator trust As input variables, the corresponding main control level is adapted through fuzzy rule reasoning and centroid method defuzzification, and the optimal scheduling plan is constrained and modified by combining preset safety thresholds and manual intervention permissions. The revised scheduling plan is converted into standardized instructions and issued to each unmanned node. The movement instructions are smoothed to generate continuous path points, and the communication switching instructions are executed by the network management module. The unmanned nodes transmit back the running status update of the global real-time situation matrix, continuously iterate in the OODA closed loop, and output the final unmanned cluster collaborative scheduling command after convergence.

[0013] In a further embodiment, the AI ​​trajectory prediction submodule predicts the network splitting risk caused by node movement in the following way: Collect the historical motion trajectory of each unmanned node, input it into the time series prediction model, and output the predicted coordinates of each node at future time. A global real-time situational awareness matrix is ​​constructed based on predicted coordinates. This matrix is ​​then input into a graph neural network threat assessment model to calculate cluster connectivity risk. ;like If node movement is determined to cause network splitting risk, the topology adjustment process is initiated in advance.

[0014] In a further embodiment, the execution steps of the region coverage optimization algorithm are as follows: Discretize the target region into Uniform grid, defining a single unmanned node The detection radius is Set the grid coordinate identifier as If grid With any single unmanned node plane coordinates The Euclidean distance between them is less than or equal to the detection radius. If so, the corresponding raster is recorded as a valid covering raster; Calculate the overall regional coverage rate based on all valid coverage grids. Construct an adaptive particle swarm optimization model, where a single particle... A corresponding set of unmanned node spatial location configurations; Population iterative updates are performed using particle velocity and particle position iteration formulas, with adaptive inertia weights introduced during the iteration process. Dynamically adjust the search step size; Based on the particle's current fitness Average fitness of the population and minimum fitness of the population Update inertia weights For iterations that exceed the region boundary The particle coordinates are subjected to boundary truncation correction. The maximum boundary dimension parameter of a single side of the target monitoring area; Continue iterative calculations until full coverage is achieved. Once the preset threshold is reached or the number of iterations reaches the upper limit, the optimal deployment coordinates of each unmanned node are output, and a regional coverage movement control command is issued.

[0015] In a further embodiment, the topology reconstruction operation is performed as follows: S301. Based on the optimized output of the area coverage of each unmanned node, the node's heartbeat information is used to divide multiple independent communication partitions using the connected component algorithm. S302. For each independent communication partition, determine the corresponding representative node; aggregate the representative nodes of all partitions to form a global representative node set. ; S303, Based on the global representative node set Each representative node within the region is used to construct a convex quadrilateral by combining any four representative nodes, and the planar coordinates of the Steiner relay nodes used to connect different partitions are solved. S304. Traverse all idle unmanned nodes in the cluster except for the representative node, and select the idle node with the smallest cost value through the scheduling cost function as the cross-partition communication relay carrier; use the virtual artificial potential field method to control the autonomous motion of the relay node, and rely on the resultant force formed by the superposition of the target gravity vector and the obstacle repulsion vector to drive the node to move to the target relay coordinate. S305. After the relay node arrives at the target coordinates and all isolated communication partitions restore full connectivity, the cluster topology connectivity reconstruction is completed, and the set of representative nodes for the entire domain is iteratively updated. This enables continuous maintenance of dynamic topology.

[0016] In a further embodiment, the matching process for the three levels of human-led, machine-led, and hybrid human-machine autonomous control is as follows: Task complexity Environmental uncertainty and operator trust Quantize to intervals respectively ; Pre-determine the task complexity Environmental uncertainty and operator trust Set corresponding low-fuzzy sets, high-fuzzy sets, and membership functions respectively, and obtain the corresponding membership values ​​based on the membership functions: Task complexity membership value. Membership value of environmental uncertainty Operator trust membership value ; If the task complexity and membership value Adapting to its own low-fuzzy set and environmental uncertainty membership value Adapt to its own low fuzzy set and operator trust membership value If the set of high fuzziness adapts to itself, it switches to the machine-dominated control level; If the task complexity and membership value Adapt to its own highly fuzzy set or the membership value of environmental uncertainty Adapt to its own highly fuzzy set, or operator trust membership value If the low-fuzzy set that adapts to itself is used, then switch to the hybrid master control level; If the overall topology risk assessment results are output, then switch to the personnel primary control level.

[0017] The beneficial effects of this invention are as follows: 1. By defining standardized access interfaces and data encapsulation protocols through the Hardware Abstraction Layer (HAL), plug-and-play functionality and unified status reporting of each unmanned node are achieved, which greatly reduces the complexity of system integration and significantly improves the scalability and device compatibility of the unmanned cluster.

[0018] 2. Based on the multi-agent deep deterministic policy gradient algorithm, the algorithm takes the real-time state matrix of the unmanned swarm as input and the minimization of global task completion and resource consumption as the reward function, and dynamically outputs the optimal scheduling strategy. This breaks through the limitations of preset rules and can flexibly allocate computing resources and adaptively distribute heterogeneous tasks according to environmental changes and task requirements, effectively improving the task completion efficiency and resource utilization of the swarm.

[0019] 3. Monitor the communication quality of multi-mode links to achieve seamless switching between primary and backup links. Combine AI prediction models to predict the movement trajectory of unmanned nodes and schedule relay points or adjust the topology in advance to ensure network connectivity of the unmanned cluster. This greatly enhances the resilience and communication reliability of the unmanned cluster in strong interference and highly dynamic environments. Attached Figure Description

[0020] Figure 1 This is an architecture diagram of the AI-driven software-defined unmanned cluster management system in Example 1.

[0021] Figure 2 This is the architecture diagram of the software-defined heterogeneous access module in Example 1.

[0022] Figure 3 This is a flowchart of the multi-agent reinforcement learning algorithm in the AI-driven collaborative scheduling module of Example 2.

[0023] Figure 4 This is the logic flowchart of the network switching and connectivity management module in Example 2.

[0024] Figure 5 This is a flowchart of the intelligent switching process for the human-machine collaboration mode in Example 2.

[0025] Figure 6 This is a schematic diagram of the multi-unit collaborative scheduling implementation in Example 2. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] Example 1 like Figure 1 As shown in the figure, this embodiment discloses an AI-driven software-defined unmanned cluster management and control system, the system comprising: The software-defined heterogeneous access module is configured to pool the underlying hardware resources of various heterogeneous unmanned nodes based on the Hardware Abstraction Layer (HAL), and to define a globally unified standardized device access interface and a general data encapsulation protocol. The various heterogeneous unmanned nodes described in this embodiment include at least aerial, ground, and underwater unmanned devices. Preferably, aerial unmanned devices can be drones, ground unmanned devices can be unmanned vehicles, and underwater unmanned devices can be unmanned boats. The HAL incorporates a single-point hardware fault isolation mechanism, ensuring that hardware failure or communication anomalies in a single node will not interfere with the normal access and data transmission of other nodes, effectively improving the overall robustness and horizontal scalability of the unmanned cluster.

[0028] The AI-driven collaborative scheduling module interacts bidirectionally with the software-defined heterogeneous access module; the AI-driven collaborative scheduling module is internally layered with a large model command brain layer, a small model cluster layer, and a human-machine collaborative switching unit. The network switching and connectivity management module communicates with the AI-driven collaborative scheduling module and all heterogeneous unmanned nodes, and has built-in multi-mode link monitoring sub-module, AI trajectory prediction sub-module and multi-unit collaborative scheduling sub-module.

[0029] In a further embodiment, the Hardware Abstraction Layer (HAL) is used to encapsulate the underlying drivers of various heterogeneous unmanned nodes, maintain a global Device Descriptor Table, assign a unique logical identifier to each connected unmanned node, and possess single-point hardware fault isolation capabilities (such as sensor type, communication interface, and battery life parameters). Specifically, in combination... Figure 2 The architecture features a hardware abstraction layer that connects to multiple types of heterogeneous unmanned nodes. This layer integrates four core mechanisms: hardware abstraction and encapsulation, resource pooling, interface standardization, and data encapsulation protocol. It supports plug-and-play functionality for heterogeneous devices and ultimately outputs standardized status data that has undergone unified formatting.

[0030] The large model command brain layer integrates multimodal visual processing components, the small model cluster layer is configured with observation model, judgment model, decision model and control model; the human-machine collaboration switching unit is equipped with Mamdani fuzzy inference logic; The multi-unit collaborative scheduling submodule integrates an adaptive particle swarm region coverage algorithm, a convex hull combined quadrilateral Steiner tree topology reconstruction algorithm, and virtual potential field node motion control logic.

[0031] By using a software-defined heterogeneous access module, and based on a Hardware Abstraction Layer (HAL), the underlying hardware resources of various unmanned nodes, including those in the air, on the ground, and underwater, are pooled. Standardized access interfaces and data encapsulation protocols are defined, enabling plug-and-play functionality and unified status reporting for heterogeneous nodes. Compared to traditional siloed architectures, this invention significantly reduces system integration complexity and substantially improves cluster scalability and device compatibility.

[0032] The network switching and connectivity management module monitors the quality of multi-mode communication links in real time. When the signal-to-noise ratio of the main link falls below a threshold, it automatically triggers protocol stack reconstruction to achieve seamless switching. Simultaneously, it uses an AI prediction model to anticipate node movement trajectories and pre-schedule relay nodes or adjust the topology to ensure that the cluster graph connectivity never falls below a preset safety threshold. This mechanism significantly enhances the cluster's resilience and communication reliability in environments with strong interference and high dynamics.

[0033] The system introduces a large language model as the command brain, which collaborates with a cluster of small models to form a rapid closed loop of observation-judgment-decision-action. It also supports intelligent switching between three human-machine collaboration modes and can dynamically adjust the level of autonomy according to the complexity of the task and the operator's intentions, thereby improving the accuracy of decision-making and the efficiency of human-machine collaboration in complex scenarios.

[0034] Example 2 Based on the software-defined unmanned cluster management system disclosed in Embodiment 1, this embodiment discloses an AI-driven management method for a software-defined unmanned cluster management system. This embodiment selects the open-source large language model (Qwen2.5-32B) as the base model and includes the following steps: S1. Call the Hardware Abstraction Layer (HAL) of the software-defined heterogeneous access module to complete the pooling of underlying hardware resources for various heterogeneous unmanned nodes; based on standardized device access interfaces and general data encapsulation protocols, receive the raw data periodically transmitted back by various heterogeneous unmanned nodes, and generate real-time situational information for the entire domain after unified formatting. S2. The AI-driven collaborative scheduling module takes real-time situational information from the entire domain as input and uses the large model command brain layer and the small model cluster layer to complete the intelligent analysis of situation-task coupling and output the initial cluster scheduling plan; combined with the human-machine collaborative switching unit to match and adapt the human-machine autonomous control level, and after correcting the scheduling constraints, the final unmanned cluster collaborative scheduling instruction is generated. S3. Invoke the network switching and connectivity management module, and use the built-in multi-mode link monitoring submodule to collect the signal-to-noise ratio of multi-mode communication links in real time and complete the link quality determination; such as Figure 4As shown, the module monitors the signal-to-noise ratio (SNR) of multi-mode links (RF, microwave, satellite, etc.) between nodes in real time. When the SNR of the primary link falls below a set threshold, it automatically triggers software reconfiguration of the network protocol stack, achieving seamless switching to the backup link. The switching process is transparent to upper-layer applications, ensuring the continuity of data transmission.

[0035] The AI ​​trajectory prediction submodule is used to predict the network splitting risk caused by node movement. The multi-unit collaborative scheduling submodule executes the regional coverage optimization algorithm and topology reconstruction operation, and outputs the node execution command to all heterogeneous unmanned nodes. S4: Receive the running status data returned by all heterogeneous unmanned nodes after executing instructions, and feed it back to the software-defined heterogeneous access module to update the real-time situation information of the entire domain; cyclically execute S1~S3 to achieve continuous closed-loop management and control of the unmanned cluster.

[0036] The steps for generating the real-time situational information across the entire domain are as follows: S101. After the heterogeneous unmanned node is powered on, it periodically sends raw heartbeat data packets through its own communication link. Specifically, it periodically pushes heartbeat data packets to the cluster management system through its own onboard communication unit (Wi-Fi, 4G / 5G wireless data transmission, dedicated data transmission radio, underwater acoustic communication module, etc.). The heartbeat packet adopts a lightweight binary encoding format and selects the Protocol Buffers standard format or a custom TLV tag-length-value encapsulation format. The data packet carries multi-dimensional status information such as the node's unique identifier, device type, air / surface / underwater positioning coordinates, remaining power, current task execution progress, and instantaneous snapshots of sensor data (temperature, humidity, image feature values, etc.).

[0037] S102. The Hardware Abstraction Layer (HAL) receives raw heartbeat data and sensor data in heterogeneous formats, and converts them into node status information with a unified format according to a preset encapsulation protocol. The standardized status messages after conversion are transmitted through any high-performance message middleware such as DDS, MQTT, or Kafka, which can shield the differences in data formats between different manufacturers and different types of devices. S103. Kalman filtering is used to smooth and reduce noise in sensor data and track target position, eliminating random noise in perception and outputting smoothed target position time series data. S104. The Hardware Abstraction Layer (HAL) maintains the global device description table in real time, synchronously updates the device status of newly added access nodes and offline disconnected nodes, and assigns a unique logical addressing identifier to all online nodes. S105. Collect node status information of all online nodes and time-series data of target locations to complete full-domain data fusion and construct a... A global real-time situational awareness matrix is ​​generated, and this matrix is ​​used to push global real-time situational information to an AI-driven collaborative scheduling module; among which, For the number of nodes, For feature dimensions; S106 The Hardware Abstraction Layer (HAL) has a built-in fault isolation mechanism. If a single node hardware failure or communication interruption is detected, the data transmission path of that node will be immediately cut off, without interfering with the data acquisition and global situation generation process of other nodes.

[0038] The above data conversion and distribution process is completely transparent to the upper-layer business modules, thereby enabling the plug-and-play feature of heterogeneous unmanned nodes: adding new nodes to the network requires no manual configuration, the system automatically identifies the devices and includes them in the global hardware resource pool; when a node goes offline, the hardware abstraction layer automatically updates the device description table and pushes an offline notification to the upper layer, supporting dynamic reconfiguration and scheduling of unmanned clusters.

[0039] The Kalman filtering described in this embodiment follows a four-layer progressive calculation relationship iterative operation: state prediction, covariance prediction, Kalman gain calculation, and state correction. Specifically: Based on the optimal state estimate from the previous time step, the prior state of the target at the current time step is predicted. The calculation relationship is as follows: ; Based on the state transition matrix and the process noise matrix, the current prior covariance is predicted, and the calculation relationship is as follows: ; The Kalman gain is calculated by combining the prior covariance, observation mapping matrix, and observation noise matrix, and the relationship is as follows: ; By correcting the prior state using Kalman gain and actual sensor observations, the optimal target position estimate for the current moment is obtained. The calculation relationship is as follows: .

[0040] In the formula, For a moment Prior state estimation of the target at any given time. For a moment Posterior optimal state Here is the state transition matrix. To control the input matrix, To control the input amount; The prior covariance matrix, For a moment The prior covariance matrix, The process noise covariance matrix; For Kalman gain, For the observation mapping matrix, To observe the noise covariance matrix; For a moment Sensor observations Let T be the optimal target position state after filtering, and T be the matrix transpose.

[0041] In a further embodiment, the output process of the initial cluster scheduling scheme is as follows: A large-scale command brain was built based on the Qwen2.5-32B platform, and a command fine-tuning question-and-answer dataset was constructed based on the combat mission manual, equipment condition data, and historical cluster scheduling cases. Supervised fine-tuning is performed using LoRA low-rank adaptation, freezing the original parameters of the large model's "brain" and utilizing the loss function. A low-rank increment matrix is ​​trained to obtain a domain-wide model adapted to unmanned cluster management scenarios; the loss function described in this embodiment... The expression is as follows: ; In the formula, To monitor and fine-tune the loss value, The total number of samples; For input and model parameters Lower output The probability of; These are pre-trained parameters; These are incremental parameters.

[0042] The large-scale model of the domain is compressed, multimodal features are fused through a cross-attention mechanism, and the overall target task is interpreted in conjunction with real-time situational information of the entire domain, and high-level scheduling and guidance instructions are output.

[0043] Furthermore, 4-bit quantization is employed to compress the model, using either GPTQ or AWQ quantization algorithms. This is based on the vLLM high-performance inference framework to achieve high concurrency and low latency. Simultaneously, a semantic caching mechanism is established to identify duplicate queries by comparing semantic similarity. When a semantically duplicated query is detected, the cached historical inference results are directly returned. 4-bit quantization is a post-training quantization technique that compresses the model's weights from FP16 / BF16 (16 bits) to only 4 integers (INT4). The GPTQ algorithm is based on the second-order Hessian matrix approximation, quantizing layer by layer and weight by weight, and using error compensation to back-allocate the quantization error to the unquantized weights in the same layer, minimizing the overall output bias. The AWQ algorithm identifies key channels based on activation value amplitude, applies scaling protection to these channels before quantization (amplifying weights and correspondingly shrinking activations), and then unifies low-bit quantization.

[0044] Image and text cross-modal feature alignment is achieved through joint image and text pre-training. Based on the aligned cross-modal features, real-time images transmitted back by unmanned equipment are analyzed to identify target categories and assess scene situation.

[0045] The formula for fusing multimodal features through a cross-attention mechanism is as follows: ,in, For multimodal fusion embedding vectors, For query purposes, the query is from a text modality; The key comes from the image modality; The value is derived from the image modality.

[0046] The formal formula for high-level instruction reasoning is as follows: ,in, For a large language model, this is treated as a function; The output is the instruction / action / decision; For historical dialogue and task knowledge, i.e., contextual information; The output for the small model is the observation information; For the current target task, the generated instructions are sent to the decision model via API to adjust its behavioral parameters, such as reward function weights and exploration rate. This is a vector concatenation and fusion operation.

[0047] The large domain model calls the smaller model cluster layer and the various models deployed at the edge to form a layered collaborative architecture: Run the observation model, based on YOLOv8 and utilizing the detection loss function Environmental situation awareness is conducted, and the target category, location, and confidence level are output and pushed to the judgment model; the detection loss function described in this embodiment... The expression is as follows: , For cross-entropy loss, The bounding box regression loss is (CIoU). For target confidence loss, , and These are the weighting coefficients for each item, and their values ​​range from [value range missing]. The specific value can be adjusted adaptively according to the task focus. For example, it can be appropriately increased when the target positioning accuracy requirement is high. The value of The judgment model receives the perception results and uses the cross-entropy loss function. A threat level assessment is conducted, a standardized risk score is output, and this score is fed into the decision-making model; furthermore, the cross-entropy loss function described in this embodiment... The calculation formula is as follows: ; in, This represents the total number of categories at a fine-grained level. For the current fine-grained category, For the sample in the first The actual label of the class is usually one-hot encoded: the actual class is 1, and the rest are 0; The sample predicted by the model belongs to the current fine-grained level. The probability of this is between 0 and 1. Combining environmental data and historical development trends, a long short-term memory network is used to predict the rate of deterioration of defective areas and generate a risk assessment score.

[0048] The decision-making model receives risk scores and high-level scheduling guidance instructions, and is based on a multi-agent deep deterministic policy gradient algorithm and reward function. Solve for the optimal cluster task allocation strategy; where the reward function is... The calculation formula is as follows: ,in, and These are dynamically adjusted weighting coefficients, generally... and The value of is between 0 and 1, and .

[0049] The task completion rate includes the percentage of targets successfully detected or covered. This represents the total resource consumption of the cluster, including flight energy consumption, communication energy consumption, and onboard computing energy consumption.

[0050] like Figure 3 As shown, the multi-agent deep deterministic policy gradient algorithm adopts a centralized training and distributed execution architecture, and the specific process is as follows: Each agent All are equipped with actor networks and the network of critics Among them, actor network It is used to output deterministic actions based on its own local observations, while the critic network evaluates the value of global joint actions based on the joint observations and joint actions of all agents; For intelligent agents Its own local observation information, For actors network Trainable parameters, For all intelligent agents to jointly observe the world, For all intelligent agents Combined action set, For the Critics Network Trainable parameters.

[0051] A globally unified shared experience replay pool is set up. Used for storing all interaction samples generated by the interaction between intelligent agents and the environment. ,in, For global coordinated action, For single-step, real-time global rewards, determined by the reward function. Calculations show that This is the environment termination flag, a Boolean variable. A value of 1 indicates the end of this round of interactive tasks, while a value of 0 indicates the task continues.

[0052] During the training phase, the following critic loss function is minimized. Update the commentator network for each agent. Trainable parameters: ,in, Let be the mathematical expectation operator, representing the shared experience replay pool. Random sampling interactive samples And take the average. For intelligent agents The target value, For the Critics Network Input, global observation Global coordinated actions Output the current joint action value estimate.

[0053] Among them, intelligent agents target value The calculation formula is as follows: ;in, For intelligent agents Single-step instant rewards, For future reward discount factors, . For intelligent agents The target critic network, For the next moment of global joint observation, For all intelligent agents in the next moment Combined action set, For intelligent agents The next output action is synchronized with the main actor's network with parameter delay. For intelligent agents Target actor network For intelligent agents The next moment of its own local observation. As a constraint term, it indicates that the actions of each agent in the next moment are all generated by the corresponding target actor network based on its own observations in the next moment, and uniformly input into the target critic network to calculate future value.

[0054] Actor Network Parameter updates are performed using the policy gradient, which is calculated using the following formula: ;in, For the objective function Actor network parameters The policy gradient is used for backpropagation to update the actor network weights. To share the experience replay pool globally Internally sample global observation samples and obtain the mathematical expectation; For actors, the network outputs trainable parameters of their movements. The gradient; To provide commentators with insights into the value of joint actions for intelligent agents action The gradient; As a constraint, actions are only output to the actor's network. We calculate the critic gradient to improve our strategy.

[0055] After completing the iterative training of the critic network and the actor network, the actor policy network of each agent is lightweighted and compressed through knowledge distillation. The lightweight policy networks are then deployed to the corresponding edge devices of the drones, and a preliminary cluster scheduling scheme is output.

[0056] In a further embodiment, the final unmanned cluster collaborative scheduling instruction generation process is as follows: The real-time situational awareness matrix is ​​input into the graph neural network threat assessment model to calculate the risk index of each node. Cluster connectivity risk ,like If this occurs, the connectivity management module will be triggered, and a global topology risk assessment result will be output; the cluster connectivity risk described in this embodiment... The calculation formula is as follows: ,in, For the pre-set security threshold, This represents the connectivity of the current graph. As an optional implementation, The specific value can be determined using a graph theory-based connected component algorithm, that is, by counting the number of nodes contained in the largest connected component in the current unmanned swarm communication topology; if the topology is already fully connected, then... It equals the total number of nodes in the cluster, N. It can be preset according to the reliability requirements of the task for cluster connectivity. For example, it can be set to ⌈0.8N⌉ in a typical monitoring task scenario.

[0057] The node risk index The method for quantifying the degree of external threat faced by a single unmanned node in the current situation is as follows: Where m is the node index, ω1, ω2, and ω3 are the risk weight coefficients for each item, all ranging from [0.1, 1.0], and satisfying ω1 + ω2 + ω3 = 1; For nodes The intensity of enemy firepower threat or interference can be normalized and quantified based on the strength of enemy radar detection signals, missile warning information or electronic interference power in the area, with a value range of [0,1]. For nodes Natural risks in the environment, including but not limited to wind speed / sea state level, terrain / obstacle density, electromagnetic environment complexity, etc., are normalized to [0,1] after fusion of environmental sensor data; For nodes The risk of its own health status is comprehensively assessed based on the remaining power, communication link error rate, and actuator fault detection results, and normalized to [0,1]. The larger the value, the worse the node status and the higher the probability of failure.

[0058] Alternatively, a graph neural network threat assessment model can be used to directly output the risk score. This involves taking a real-time global situational awareness matrix as input, aggregating the node's own features and those of its neighbors through a multi-layer graph convolutional network, and then outputting the risk score for each node. ∈[0,1] represents the model inference result, and no manual weighting coefficients are required.

[0059] Integrating the initial cluster scheduling scheme, high-level scheduling guidance instructions, real-time global situational information, and task objectives, the population is iterated through a multi-objective evolutionary algorithm. After selection, crossover, mutation, and Pareto non-dominated sorting, the results were ranked according to task completion. Total energy consumption of the cluster Network resilience Using fitness metrics, optimal scheduling plans with the Pareto front are selected; among them, Based on the task execution progress data returned by each node, the data is directly obtained and normalized to [0,1]. Based on the remaining power of each node and the power consumption model of the propulsion system, real-time estimation is performed. The calculation is based on the number of connected components and link redundancy in the current communication topology.

[0060] Utilizing the fuzzy inference logic built into the human-machine collaborative switching unit, the task complexity is reduced. Environmental uncertainty and operator trust As input variables, the corresponding main control level is adapted through fuzzy rule reasoning and centroid method defuzzification. The optimal scheduling plan is then constrained and corrected by combining preset safety thresholds and manual intervention permissions. Among these, task complexity... The number of subtasks to be executed and their coupling dependencies are directly quantified and determined, taking into account environmental uncertainty. Based on the environmental sensor data (including wind speed, visibility, electromagnetic interference intensity, etc.) returned by each node, the operator trust level is calculated in real time after weighted fusion. Determined by reverse mapping based on the historical intervention frequency of operators to the system's automatic output commands, i.e. All input variables are normalized to the interval [0,10]. The revised scheduling plan is converted into standardized instructions and issued to each unmanned node. The movement instructions are smoothed to generate continuous path points, and the communication switching instructions are executed by the network management module. The unmanned nodes transmit back the running status update of the global real-time situation matrix, continuously iterate in the OODA closed loop, and output the final unmanned cluster collaborative scheduling command after convergence.

[0061] It is worth noting that, for reference Figure 5 The matching process for the three levels of human-led, machine-led, and hybrid human-machine autonomous control is as follows: Task complexity Environmental uncertainty and operator trust Quantize to intervals respectively The task complexity Quantified based on the number of subtasks and their dependencies, and considering environmental uncertainty. Based on interference intensity and obstacle density quantification, operator trust level Quantified by reverse mapping of historical intervention frequency; Pre-determine the task complexity Environmental uncertainty and operator trust Set corresponding low-fuzzy sets, high-fuzzy sets, and membership functions respectively, and obtain the corresponding membership values ​​based on the membership functions: Task complexity membership value. Membership value of environmental uncertainty Operator trust membership value ; If the task complexity and membership value Adapting to its own low-fuzzy set and environmental uncertainty membership value Adapt to its own low fuzzy set and operator trust membership value If the set of high fuzziness adapts to itself, it switches to the machine-dominated control level; If the task complexity and membership value Adapt to its own highly fuzzy set or the membership value of environmental uncertainty Adapt to its own highly fuzzy set, or operator trust membership value If the low-fuzzy set that adapts to itself is used, then switch to the hybrid master control level; If the overall topology risk assessment results are output, then switch to the personnel primary control level.

[0062] Correspondingly, the AI ​​trajectory prediction submodule predicts the risk of network splitting caused by node movement in the following way: Collect the historical motion trajectory of each unmanned node, input it into the time series prediction model, and output the predicted coordinates of each node at future time. A global real-time situational awareness matrix is ​​constructed based on predicted coordinates. This matrix is ​​then input into a graph neural network threat assessment model to calculate cluster connectivity risk. ;like If node movement is determined to cause network splitting risk, the topology adjustment process is initiated in advance.

[0063] In a further embodiment, reference is made to... Figure 6 The execution steps of the region coverage optimization algorithm are as follows: Discretize the target region into Uniform grid, defining a single unmanned node The detection radius is Set the grid coordinate identifier as If grid With any single unmanned node plane coordinates The Euclidean distance between them is less than or equal to the detection radius. If so, the corresponding raster is recorded as a valid covering raster; Calculate the overall regional coverage rate based on all valid coverage grids. Construct an adaptive particle swarm optimization model, where a single particle... This corresponds to a set of unmanned node spatial location configurations; for ease of understanding, the full-area coverage rate described in this embodiment... The calculation formula is: ; Population iterative updates are performed using particle velocity and particle position iteration formulas, with adaptive inertia weights introduced during the iteration process. Dynamically adjust the search step size.

[0064] Furthermore, the particle velocity iteration formula described in this embodiment takes the following form: ;in, For particles No. Vi in the The speed of round iteration, For particles No. Vi in the The speed after each iteration update and These are individual learning factors and global learning factors, respectively. and For interval Uniformly distributed random coefficients; For particles Its own historical best Dimensional position, The globally optimal first of all particles in the population Dimensional position, For particles No. Vi in the Round iteration position.

[0065] Correspondingly, the particle position iteration formula is expressed in the following form: Dimensional explanation For particles No. Vi in the The position after each iteration update. Note the meaning of "velocity" in the formula. It is not a physical unit of velocity (m / s), but rather a dimensionless abstract quantity used in particle swarm optimization algorithms to characterize the search step size; its value is related to its position. The dimensions of both are consistent, both being spatial coordinates (m). Therefore, their direct addition maintains dimensional consistency and does not involve conversion of physical kinematic units. Similarly, in this invention, all "velocity" and "force" based on particle swarm optimization, adaptive inertial weight update, and virtual potential field method are abstract numerical values ​​or normalized vectors at the algorithm level. After boundary constraints and step size scaling, they are directly used for coordinate iterative updates without requiring additional physical dimension conversion.

[0066] Based on the particle's current fitness Average fitness of the population and minimum fitness of the population Update inertia weights In a further embodiment, the inertial weight described in this embodiment The update formula is as follows: ;in, This represents the upper limit of the inertia weight value. This represents the lower limit of the inertia weight value.

[0067] For iterations that exceed the region boundary The particle coordinates are subjected to boundary truncation correction. The maximum boundary dimension parameter of a single side of the target monitoring area; the corresponding boundary truncation correction formula is: .

[0068] Continue iterative calculations until full coverage is achieved. Once the preset threshold is reached or the number of iterations reaches the upper limit, the optimal deployment coordinates of each unmanned node are output, and a regional coverage movement control command is issued.

[0069] Correspondingly, the operation process of the topology reconstruction operation is as follows: S301. Based on the deployment location of each unmanned node optimized by the regional coverage, all unmanned nodes periodically exchange heartbeat data packets to obtain the communication status information of neighboring nodes. The connected component algorithm is used to group all unmanned nodes in the cluster, resulting in multiple groups of isolated independent communication partitions that cannot communicate directly with each other. S302. For each independent communication partition, the Graham convex hull scan algorithm is used to extract the outer boundary nodes of the partition. Then, the Prim algorithm is combined to construct the minimum spanning tree of the partition nodes. The intersection of the convex hull node set and the minimum spanning tree node set is taken as the representative node of the partition. The representative nodes selected from all independent communication partitions are summarized and integrated to form the global representative node set. ; S303, Based on the global representative node set For each representative node within the region, any four representative nodes are combined to construct a convex quadrilateral, and the following formula is used to solve for the planar coordinates of the Steiner relay nodes used to connect different partitions. : ;in, , Vertex of a convex quadrilateral plane coordinates, As vertex ,vertex The straight-line distance between them For line segments The angle between the horizontal axis of the plane; It should be noted that if all interior angles of the constructed convex quadrilateral are less than 120°, the corresponding relay point is calculated; when there are obstacles in the node connection path, the tangent method is used to correct the relay coordinates; if the straight connection path intersects with the obstacle, two tangents combined with an arc are used to replace the straight link, and the coordinates of the tangent point are obtained through geometric relationships.

[0070] S304. Traverse all idle unattended nodes in the cluster except for the representative node, and select the idle node with the smallest cost value as the cross-partition communication relay carrier through the scheduling cost function; furthermore, the scheduling cost function described in this embodiment is expressed in the following form: ,in, Idle unattended nodes The overall cost of relay scheduling, , These are the distance weight and communication load weight coefficients, respectively. Idle unattended nodes Real-time planar coordinates To obtain the target relay coordinates, Idle unattended nodes The number of currently established communication links, This represents the maximum number of communication connections allowed for a single node.

[0071] The virtual artificial potential field method is used to control the autonomous motion of the relay node. The node moves to the target relay coordinates by relying on the resultant force formed by the superposition of the target gravity vector and the obstacle repulsion vector. Specifically, the formula for calculating gravity is as follows: ,in, The gravitational vector pointing towards the relay target point. This is the gravity adjustment coefficient. The real-time planar coordinates of the relay node. The coordinates of the relay target.

[0072] Correspondingly, the formula for calculating repulsive force segments is expressed as follows: ; in, The repulsive force vector is the force moving away from the obstacle. This is the repulsion adjustment coefficient. Let the plane coordinates of the obstacle be... This is the effective distance threshold for the repulsive force.

[0073] S305. After the relay node arrives at the target coordinates and all isolated communication partitions restore full connectivity, the cluster topology connectivity reconstruction is completed, and the set of representative nodes for the entire domain is iteratively updated. This enables continuous maintenance of dynamic topology.

[0074] In summary, the AI-driven collaborative scheduling module embeds a deep multi-agent reinforcement learning (MARL) algorithm engine into the control plane. Taking the cluster's real-time state matrix as input and minimizing global task completion and resource consumption as the reward function, it dynamically outputs the optimal scheduling strategy. This mechanism overcomes the limitations of preset rules, enabling flexible allocation of computing resources and adaptive distribution of heterogeneous tasks based on environmental changes and task requirements, effectively improving the cluster's task completion efficiency and resource utilization.

Claims

1. An AI-driven, software-defined, unmanned cluster management and control system, characterized in that, The system includes: The software-defined heterogeneous access module is configured to pool the underlying hardware resources of various heterogeneous unmanned nodes based on the hardware abstraction layer (HAL), and to formulate a globally unified standardized device access interface and a general data encapsulation protocol. The AI-driven collaborative scheduling module interacts bidirectionally with the software-defined heterogeneous access module; the AI-driven collaborative scheduling module is internally layered with a large model command brain layer, a small model cluster layer, and a human-machine collaborative switching unit. The network switching and connectivity management module communicates with the AI-driven collaborative scheduling module and all heterogeneous unmanned nodes, and has built-in multi-mode link monitoring sub-module, AI trajectory prediction sub-module and multi-unit collaborative scheduling sub-module.

2. The AI-driven software-defined unmanned cluster management system according to claim 1, characterized in that, The Hardware Abstraction Layer (HAL) is used to encapsulate the underlying drivers of various heterogeneous unmanned nodes, maintain a global device description table, assign a unique logical identifier to each connected unmanned node, and has single-point hardware fault isolation capability. The large model command brain layer integrates multimodal visual processing components, the small model cluster layer is configured with observation model, judgment model, decision model and control model; the human-machine collaboration switching unit is equipped with Mamdani fuzzy reasoning logic; The multi-unit collaborative scheduling submodule integrates an adaptive particle swarm region coverage algorithm, a convex hull combined quadrilateral Steiner tree topology reconstruction algorithm, and virtual potential field node motion control logic.

3. A management method using an AI-driven software-defined unmanned cluster management system as described in claim 1, characterized in that, Includes the following steps: S1. Call the hardware abstraction layer (HAL) of the software-defined heterogeneous access module to complete the pooling of underlying hardware resources for various heterogeneous unmanned nodes. Based on standardized equipment access interfaces and universal data encapsulation protocols, it receives raw data periodically transmitted back from various heterogeneous unmanned nodes, and generates real-time situational information across the entire domain after unified formatting. S2. The AI-driven collaborative scheduling module takes real-time situational information from the entire domain as input and uses the large model command brain layer and the small model cluster layer to work together to complete the intelligent analysis of situation-task coupling and output the preliminary cluster scheduling plan. By combining the human-machine collaborative switching unit to match and adapt the human-machine autonomous control level, and after correcting the scheduling constraints, the final unmanned cluster collaborative scheduling instruction is generated. S3. Call the network switching and connectivity management module, collect the signal-to-noise ratio of multi-mode communication links in real time and complete the link quality judgment through the built-in multi-mode link monitoring sub-module, use the AI ​​trajectory prediction sub-module to predict the network splitting risk caused by node movement, and execute the regional coverage optimization algorithm and topology reconstruction operation through the multi-unit collaborative scheduling sub-module, and output the node execution command to all heterogeneous unmanned nodes. S4: Receive the running status data returned by all heterogeneous unmanned nodes after executing instructions, and feed it back to the software-defined heterogeneous access module to update the real-time situation information of the entire domain; cyclically execute S1~S3 to achieve continuous closed-loop management and control of the unmanned cluster.

4. The control method for an AI-driven software-defined unmanned cluster control system according to claim 3, characterized in that, The steps for generating the real-time situational information across the entire domain are as follows: S101. After the heterogeneous unmanned node is powered on, it periodically sends raw heartbeat data packets through its own communication link. S102. The Hardware Abstraction Layer (HAL) receives heterogeneous format raw heartbeat data and sensor data, and converts them into node status information with a unified format according to a preset encapsulation protocol. S103. Kalman filtering is used to smooth and reduce noise in sensor data and track target position, eliminating random noise in perception and outputting smoothed target position time series data. S104. The Hardware Abstraction Layer (HAL) maintains the global device description table in real time, synchronously updates the device status of newly added access nodes and offline disconnected nodes, and assigns a unique logical addressing identifier to all online nodes. S105. Collect node status information of all online nodes and time-series data of target locations to complete full-domain data fusion and construct a... A global real-time situational awareness matrix is ​​generated, and this matrix is ​​used to push global real-time situational information to an AI-driven collaborative scheduling module; among which, For the number of nodes, For feature dimensions; S106 The Hardware Abstraction Layer (HAL) has a built-in fault isolation mechanism. If a single node hardware failure or communication interruption is detected, the data transmission path of that node will be immediately cut off, without interfering with the data acquisition and global situation generation process of other nodes.

5. The control method for an AI-driven software-defined unmanned cluster control system according to claim 3, characterized in that, The output process of the initial cluster scheduling scheme is as follows: Build a large model to command the brain, and construct a question-and-answer dataset for fine-tuning instructions; Supervised fine-tuning is performed using LoRA low-rank adaptation, freezing the original parameters of the large model's "brain" and utilizing the loss function. Train a low-rank incremental matrix to obtain a domain-wide model adapted to unmanned cluster management scenarios; The large-scale model of the domain is compressed, multimodal features are fused through a cross-attention mechanism, and the overall target task is interpreted in conjunction with real-time situational information of the entire domain, and high-level scheduling and guidance instructions are output. A large domain model calls smaller model clusters, with each model deployed at the edge forming a hierarchical collaborative architecture: running and observing models utilizes the detection loss function. Conduct environmental situational awareness, output target category, location and confidence level and push them to the judgment model; The judgment model receives the perception results and uses the cross-entropy loss function. Conduct threat level assessments, output standardized risk scores, and push them to the decision-making model; The decision-making model receives risk scores and high-level scheduling guidance instructions, and employs a Kido agent deep deterministic policy gradient algorithm and reward function. Solve for the optimal cluster task allocation strategy and output the preliminary cluster scheduling scheme.

6. The control method for an AI-driven software-defined unmanned cluster control system according to claim 3, characterized in that, The process of generating the final unmanned cluster collaborative scheduling instruction is as follows: The real-time situational awareness matrix is ​​input into the graph neural network threat assessment model to calculate the risk index of each node. Cluster connectivity risk ,like If so, the connectivity control module will be triggered, and the global topology risk assessment results will be output. Integrating the initial cluster scheduling scheme, high-level scheduling guidance instructions, real-time global situational information, and task objectives, the population is iterated through a multi-objective evolutionary algorithm. After selection, crossover, mutation, and Pareto non-dominated sorting, the results were ranked according to task completion. Total energy consumption of the cluster Network resilience Using fitness metrics, we screen for optimal scheduling plans based on the Pareto front. Utilizing the fuzzy inference logic built into the human-machine collaborative switching unit, the task complexity is reduced. Environmental uncertainty and operator trust As input variables, the corresponding main control level is adapted through fuzzy rule reasoning and centroid method defuzzification, and the optimal scheduling plan is constrained and modified by combining preset safety thresholds and manual intervention permissions. The revised scheduling plan is converted into standardized instructions and issued to each unmanned node. The movement instructions are smoothed to generate continuous path points, and the communication switching instructions are executed by the network management module. The unmanned nodes transmit back the running status update of the global real-time situation matrix, continuously iterate in the OODA closed loop, and output the final unmanned cluster collaborative scheduling command after convergence.

7. The control method for an AI-driven software-defined unmanned cluster control system according to claim 3, characterized in that, The AI ​​trajectory prediction submodule predicts the risk of network splitting caused by node movement in the following way: Collect the historical motion trajectory of each unmanned node, input it into the time series prediction model, and output the predicted coordinates of each node at future time. A global real-time situational awareness matrix is ​​constructed based on predicted coordinates. This matrix is ​​then input into a graph neural network threat assessment model to calculate cluster connectivity risk. If If node movement is determined to cause network splitting risk, the topology adjustment process is initiated in advance.

8. The control method for an AI-driven software-defined unmanned cluster control system according to claim 3, characterized in that, The execution steps of the region coverage optimization algorithm are as follows: Discretize the target region into Uniform grid, defining a single unmanned node The detection radius is Set the grid coordinate identifier as If grid With any single unmanned node plane coordinates The Euclidean distance between them is less than or equal to the detection radius. If so, the corresponding raster is recorded as a valid covering raster; Calculate the overall regional coverage rate based on all valid coverage grids. Construct an adaptive particle swarm optimization model, where a single particle... A corresponding set of unmanned node spatial location configurations; Population iterative updates are performed using particle velocity and particle position iteration formulas, with adaptive inertia weights introduced during the iteration process. Dynamically adjust the search step size; Based on the particle's current fitness Average fitness of the population and minimum fitness of the population Update inertia weights For iterations that exceed the region boundary Perform boundary truncation correction on the particle coordinates. The maximum boundary dimension parameter of a single side of the target monitoring area; Continue iterative calculations until full coverage is achieved. Once the preset threshold is reached or the number of iterations reaches the upper limit, the optimal deployment coordinates of each unmanned node are output, and a regional coverage movement control command is issued.

9. The control method for an AI-driven software-defined unmanned cluster control system according to claim 3, characterized in that, The operation process of the topology reconstruction operation is as follows: S301. Based on the optimized output of the area coverage of each unmanned node, the node's heartbeat information is used to divide multiple independent communication partitions using the connected component algorithm. S302. For each independent communication partition, determine the corresponding representative node; aggregate the representative nodes of all partitions to form a global representative node set. ; S303, Based on the global representative node set Each representative node within the region is used to construct a convex quadrilateral by combining any four representative nodes, and the planar coordinates of the Steiner relay nodes used to connect different partitions are solved. S304. Traverse all idle unattended nodes in the cluster except for the representative node, and use the scheduling cost function to traverse and select the idle node with the smallest cost value as the cross-partition communication relay carrier. The virtual artificial potential field method is used to control the autonomous motion of the relay node. The node moves to the target relay coordinates by the resultant force formed by the superposition of the target gravity vector and the obstacle repulsion vector. S305. After the relay node arrives at the target coordinates and all isolated communication partitions restore full connectivity, the cluster topology connectivity reconstruction is completed, and the set of representative nodes for the entire domain is iteratively updated. This enables continuous maintenance of dynamic topology.

10. The control method for an AI-driven software-defined unmanned swarm control system according to claim 6, wherein the matching process for the three levels of human-led, machine-led, and hybrid human-machine autonomous control is as follows: Task complexity Environmental uncertainty and operator trust Quantize to intervals respectively ; Pre-determine the task complexity Environmental uncertainty and operator trust Set corresponding low-fuzzy sets, high-fuzzy sets, and membership functions respectively, and obtain the corresponding membership values ​​based on the membership functions: Task complexity membership value. Membership value of environmental uncertainty Operator trust membership value ; If the task complexity and membership value Adapting to its own low-fuzzy set and environmental uncertainty membership value Adapt to its own low fuzzy set and operator trust membership value If the set of high fuzziness adapts to itself, it switches to the machine-dominated control level; If the task complexity and membership value Adapt to its own highly fuzzy set or the membership value of environmental uncertainty Adapt to its own highly fuzzy set, or operator trust membership value If the low-fuzzy set that adapts to itself is used, then switch to the hybrid master control level; If the overall topology risk assessment results are output, then switch to the personnel primary control level.

Images