Research on the Construction and Resource Optimization of a UAV Command Information System Based on Large Language Models

Highlights

What are the main findings?

• A three-layer “cloud-edge-terminal” UAV command information system has been developed, deploying LLMs of different scales (cloud 175B, edge 70B, terminal 7B) to achieve optimal resource allocation and support elastic expansion from 10 to over 1000 UAVs. An improved Grey Wolf Optimization Algorithm (ILGWO) integrating Lévy flight, adaptive weighting, elite learning, and chaotic initialization strategies has been proposed to enhance system performance.
• The system achieves remarkable improvements compared to traditional methods: 34.2% reduction in task latency, 29.6% optimization in energy consumption, and 31.8% improvement in resource utilization. In urban rescue scenarios, response latency decreased by 44.7% and coordination efficiency increased by 39.5%, while LLM integration enhanced decision-making accuracy across multiple dimensions—task priority adjustment (76.3% to 94.7%), dynamic resource allocation (68.9% to 91.2%), and anomaly handling (71.5% to 93.8%).

What is the implication of the main finding?

• System Performance and Generalization Capabilities. Integrating large language models into unmanned aerial vehicle command information systems enables autonomous decision-making and intelligent planning without relying on preset rules, effectively addressing the shortcomings of existing information systems in handling dynamic scenarios and complex environments. The system demonstrates superior generalization ability with 82.5% performance retention in new scenarios, significantly outperforming the reinforcement learning method’s 56.7%, enabling rapid deployment in rescue, emergency response, inspection, and reconnaissance scenarios.
• Technical Specifications and Practical Implementation. The proposed system architecture and optimization algorithm provide a scalable and robust solution for large-scale UAV swarm operations. With fault tolerance capabilities maintaining an 88.4% task completion rate even under 20% node failure conditions and a real-time inference delay of only 156 ms for edge-deployed 7B models, the system meets practical requirements for emergency rescue, environmental monitoring, and intelligent surveillance applications. This work establishes a theoretical and practical foundation for integrating cutting-edge AI technologies into unmanned aerial vehicle systems, promoting the development of intelligent, adaptive, and resilient autonomous systems for critical mission scenarios.

Abstract

As UAVs are increasingly deployed in complex scenarios such as disaster monitoring, emergency rescue, and power-line inspection, traditional command and control systems face severe challenges in intelligent decision-making, resource allocation, and elastic scalability. To address these issues, we first propose a distributed UAV command and control system based on large language models of the LLaMA2 family. The system adopts a “cloud–edge–terminal” architecture, using 5G as the backbone network and the Internet of Things as a supplement, with edge computing serving as the computing platform. LLMs of various parameter scales are deployed on demand at different hierarchical levels to support both training and inference, enabling intelligent decision-making and optimal resource allocation. Second, we establish a multidimensional system model that integrates computation, communication, and energy consumption, providing a theoretical analysis of network dynamics, resource constraints, and task heterogeneity. Furthermore, we develop an improved Grey Wolf Optimizer (ILGWO) that incorporates adaptive weights, an elite learning strategy, and Lévy flights to solve the multi-objective optimization problem posed by the system. Experimental results show that the proposed system improves task latency, energy efficiency, and resource utilization by 34.2%, 29.6%, and 31.8%, respectively, compared with conventional methods. Real-world field tests demonstrate that, in urban rescue scenarios, the system reduces response latency by 44.7% and increases coordination efficiency by 39.5%. This work offers a reference for the optimized design and practical deployment of UAV command and control systems in complex environments.

1. Introduction

In recent years, rapid technological advancements have enabled unmanned aerial vehicles (UAVs) to be widely employed in emergency rescue, power-line inspection, disaster monitoring, and intelligence reconnaissance. Nevertheless, as application scenarios and mission requirements expand, existing UAV command and control (C2) systems confront significant challenges.

First, they exhibit insufficient intelligent decision-making capabilities. Most current systems rely on preset rules and heuristic algorithms, which struggle to cope efficiently with dynamic and complex environments. In multi-UAV cooperative missions, heavy human intervention is required, and capabilities such as intelligent situational awareness, dynamic resource adjustment, and task allocation are lacking.

Second, resource allocation efficiency is limited. Computing resources, communication channels, and energy budgets are configured independently without holistic optimization, resulting in a pronounced gap between resource utilization and system performance. This shortcoming is especially critical in emergency rescue operations, resource-constrained environments, and regions with deficient infrastructure, where efficient resource orchestration is indispensable.

Third, scalability is inadequate. Traditional centralized architectures scale poorly, making it difficult to support large-scale UAV swarms executing heterogeneous missions. As system size increases, overload risks emerge, leading to exponential growth in computational complexity, extended communication latency, and amplified fault impacts.

Consequently, scholars have made various efforts to address these challenges. Rule-engine-based approaches [1] offer rapid decision-making and straightforward implementation yet suffer from poor adaptability to unfamiliar scenarios. Reinforcement-learning-based methods [2,3,4], including DQN, PPO, and DPO, learn optimal policies via trial-and-error interactions with the environment; however, they entail high training costs, low sample efficiency, and limited generalization. Traditional machine learning techniques perform well in specific contexts but are constrained by limited adaptability, insufficient generalization, and dependence on manual feature engineering.

Recently, large language models (LLMs) have emerged as a novel AI paradigm, demonstrating remarkable intelligence and being regarded as a viable route toward artificial general intelligence [5,6,7,8]. Pre-trained on massive unsupervised corpora, LLMs acquire high-dimensional representations of general knowledge. Leveraging in-context learning (ICL) and few-shot prompting, they enable UAV systems to adapt to diverse task scenarios through succinct natural-language interactions and to accomplish situational awareness, task allocation, and command decision-making effectively in complex missions.

Motivated by these insights, this paper proposes a cloud–edge–terminal UAV command and control system empowered by LLMs. The principal contributions are as follows:

(1)

Architectural innovation: A three-tier distributed architecture integrated with large language models is designed, optimizing the entire pipeline from data acquisition to intelligent decision-making.

(2)

Modeling innovation: A multidimensional system-effectiveness optimization model is established that explicitly considers network dynamics, task heterogeneity, and resource constraints.

(3)

Algorithmic innovation: An improved Lévy-flight-enhanced Grey Wolf Optimizer (ILGWO) incorporating multiple enhancement strategies is devised to solve complex multi-objective optimization problems effectively.

(4)

Evaluation innovation: Comprehensive performance assessments and practical validations are conducted via simulations and real-world scenarios, demonstrating the system’s efficacy and utility.

2. Related Work

2.1. Unmanned Command and Control Systems

Command and control (C2) systems are defined differently across industry, telecommunications, and defense, yet they share a common essence: a system that supports personnel, equipment, or other systems in executing assigned tasks by issuing directives and maintaining control over assets [9]. Zhong et al. [10] constructed a two-stage optimization model for unmanned weapon systems that maximizes mission-resilience metrics, thereby improving the coordination between central and edge nodes after mission interruptions. Johansson et al. [11] advocate integrating UAV command and control into a single, cohesive system rather than treating them as separate functions, arguing that such unification yields synergistic advantages. Radovanović et al. [12] evaluate the feasibility of incorporating UAV mission execution into modern air-warfare C2 information systems and illustrate recent worldwide armed-conflict use cases. Wang et al. [13] propose a reinforcement-learning-based task-offloading method for post-disaster rescue scenarios.

Nevertheless, existing studies predominantly concentrate either on task offloading and communication-resource optimization on the UAV-task side or in vertical-domain applications. There remains a dearth of dedicated command-and-information systems tailored for UAVs. Moreover, most solutions rely on simple preset rules or heuristic algorithms, rendering them ill-suited for highly dynamic, complex environments where multiple heterogeneous tasks must be processed in parallel.

2.2. Large Language Model Compression Techniques

Large-scale models entail enormous parameter counts, data volumes, and computational demands. Although they exhibit near-general artificial intelligence capabilities, their heavy reliance on computational resources severely constrains their deployment scenarios. Model compression has emerged as an effective remedy. Techniques such as quantization and pruning drastically reduce model size while preserving predictive accuracy. Quantization maps high-precision floating-point weights to low-bit-width representations. LLM-QAT [14] introduces a data-free distillation scheme that leverages data synthesized by the pre-trained model itself, thereby minimizing dependence on external datasets during quantization-aware training and enhancing model extensibility.

Pruning identifies and eliminates redundant parameters to achieve lightweighting. LLM-Pruner [15] proposes a task-agnostic structured pruning method that constructs a dependence graph using gradient information. By leveraging first-order statistics and an approximated Hessian matrix, it evaluates parameter importance and interdependencies to selectively remove non-critical structures. Collectively, these techniques lay the groundwork for deploying large language models in resource-constrained environments.

2.3. Intelligent Decision-Making Approaches

Intelligent decision-making is one of the core capabilities of unmanned aerial vehicle command and control systems, and it is a concentrated reflection of the overall intelligence level of the system. The existing intelligent decision-making methods mainly include reinforcement learning, rule-based engines, game theory methods, and traditional machine learning methods.

Reinforcement-learning-related methods include the Deep Q-Network (DQN), Near-End Policy Optimization (PPO), and Actor Critic (SAC). This type of method learns the optimal strategy through trial-and-error interaction with the environment, demonstrating higher accuracy and faster speed in intelligent decision-making than human operators. However, there is a significant dependence on the number of training samples and reward function, and the model’s generalization ability is limited. The method based on a rule engine constructs a rule base based on domain knowledge, which has the advantages of fast decision-making speed and clear and accurate logic. However, it relies heavily on rules and has poor adaptability to application scenarios. The game theory method models the UAV decision-making problem as a differential game or multi-agent game and seeks optimization through a Nash equilibrium solution. It is naturally suitable for adversarial scenarios, but modeling dynamic and high-capacity scenarios is difficult, and optimization is challenging. Traditional machine learning methods mainly include support vector machines and random forests, which perform well in specific scenarios but have a high dependence on artificial features and scene characteristics.

The LLM method introduced in this article is a method that is convenient to use, has strong generalization ability, and performs well in intelligence. The LLM is mainly based on the transformer architecture, which learns general knowledge representations through pre-training on massive unsupervised data, empowering unmanned aerial vehicle systems to reduce the threshold of human–computer interaction, associate a wide range of relevant knowledge, and accurately solve intelligence situations. We compared the characteristics of several methods in

Table 1. Comparative analysis of different intelligent decision-making methods.

Characteristic Dimension	Reinforcement Learning	Rule Engine	Traditional Machine Learning	Large Language Models
Knowledge source	Environment interaction	Expert experience	Annotation data	Pre-training fine-tuning
Learning paradigm	Trial and error	Reward symbol reasoning	Supervised learning	Self-monitoring
Reasoning mechanism	Value and Strategy Function	Logic reasoning	Feature Mapping	Attention mechanism
Time complexity	$O\left(n^2\right)$	$O\left(n\right)$	$O\left(nlogn\right)$	$O\left(n^{2}d\right)$
Space complexity	$O$	$S$	$x$	$A$
Sample efficiency	$\mathrm{Low} ({10}^6$ levels)	No need for samples	$\mathrm{Medium} ({10}^3$ levels)	$\mathrm{High} ({10}^2$ levels)

to provide quick screening for researchers in this field.

In the above table, $n$ is the input dimension, $d$ is the feature dimension, $\left|S\right|$ and $\left|A\right|$ are the state and action space sizes, $\left|R\right|$ is the number of rules, and $L$ is the sequence length.

3. System Architecture Design

This chapter analyzes the system architecture proposed in this article. Section 3.1 provides a schematic diagram of the overall system framework and introduces the functions of each level. Section 3.2 summarizes and evaluates the key technologies covered by the system, helping readers quickly grasp the prerequisite knowledge in the relevant fields. Section 3.3 provides a schematic diagram of the system workflow, visualizing the system’s operational process. Section 3.4 and Section 3.5, respectively, discuss the scalability and reliability of the system.

3.1. Overall Architecture Overview

Figure 1 depicts the proposed LLM-based UAV command and control system, structured as a three-tier “cloud–edge–terminal” architecture. The design uses 5G as its backbone network and edge computing servers as its computational substrate [16]. By co-locating Multi-access Edge Computing (MEC) nodes with the 5G base-station CU (Central Unit), the architecture pushes both computing power and intelligent services to the mission edge, ensuring an unimpeded pipeline from data acquisition to intelligent decision-making.

The central layer of this architecture establishes a system center cloud by deploying high-performance general-purpose servers, base stations, and data centers for cloud deployment of UAVs and related business services, as well as the full LLM (LLaMA-175B). Cloud resources are abundant, suitable for handling large-scale and highly complex tasks. Through methods such as fiber optic cable network, 5G mobile communication, and satellite communication, the system maintains high-speed interconnection with other layers by routing and jumping through the aggregation layer, providing global situational awareness, intelligent decision-making, device control, and task distribution services for the system.

The edge layer is located between the center layer and the edge layer. As a middleware for instruction parsing, data distribution, and task offloading, it integrates the MEC system with the 5G mobile communication system and deploys them in the same base station through the 5G base-station control unit (CU) and the edge computing node to build the edge cloud for processing delay-sensitive and medium-sized computing tasks.

The terminal layer includes unmanned aerial vehicle platforms, ground control stations, vehicle-mounted base-station nodes, and sensor devices. UAVs carry different types of payloads, such as electro-optical pods, radar systems, and communication relays, to carry out the corresponding tasks. Terminal devices have limited computing power and energy storage, so they mainly undertake tasks such as data collection, data transmission, and instruction execution.

3.2. Key Technology Architecture

3.2.1. Distributed Large Language Model Deployment Architecture

We adopt a layered deployment strategy for the system architecture designed in this article. We deploy a 175B full-scale model in the central layer to handle complex inference and global optimization tasks; deploy a 70B large language model at the edge layer to handle general complex tasks and latency sensitive tasks, saving the latency overhead of task feedback; deploy the 7B lightweight large language model together with the base-station CU at the terminal layer, and provide simple data-parsing tasks at the task site. The LLM scale selected for the three-layer network not only adapts to the existing edge computing servers but also adapts to the upper limit of hardware computing power of intelligent terminal devices.

3.2.2. Fifth-Generation (5G) Network Slicing and QoS Assurance

Leveraging 5G network-slicing technology [17], the system can logically share a single physical infrastructure among heterogeneous tasks. Depending on whether the UAV is engaged in data collection, reconnaissance, swarm control, or data dissemination, the appropriate slice—tailored in terms of latency, bandwidth, and compute—is selected on demand. A QoS framework further performs dynamic bandwidth allocation, task-queue optimization, and adaptive modulation and coding. By continuously matching resource allocation to task priority and type, the system guarantees end-to-end service quality.

3.2.3. Edge-Intelligence Computing Framework

Edge computing is realized through a containerized architecture that delivers elastic compute resources and scheduling. Built on Kubernetes [18], the platform automatically scales pod resources in response to real-time system load, ensuring efficient utilization of constrained hardware. A task scheduler allocates workloads according to task requirements, resource availability, and network conditions. By consulting a pre-built task-feature library and a demand-prediction model, the framework forecasts the performance of different scheduling policies. Reinforcement-style algorithms then learn the optimal policy, enhancing the system’s adaptability to evolving task environments.

3.2.4. LLM Empowers Intelligent Decision Analysis

In the previous section on intelligent decision-making methods, we briefly introduced a comparison between large language models and other intelligent decision-making methods. This section will analyze the applicability of LLMs in empowering intelligent decision-making for unmanned aerial vehicle command and control systems from a requirements perspective. There are three main requirements for UAVs to perform diverse tasks.

One is the real-time requirement. Unmanned aerial vehicle control has extreme requirements for system decision delay, and in complex scenarios, obstacle avoidance and emergency response usually require millisecond-level response delays. The rule engine method not only has an efficient reasoning process but also has fast decision-making ability. But it is difficult to effectively cope with complex task scenarios. Compressing the LLM while maintaining model performance can effectively reduce deployment thresholds and exhibit acceptable inference latency.

The second is the demand for environmental adaptability. Emergency rescue scenarios involving unmanned aerial vehicles are usually dynamic, complex, and uncertain. Reinforcement learning methods perform well in limited space training, but they have difficulty dealing with unfamiliar scenarios and interference factors outside the data distribution. And large language models rely on in-context learning and few-shot capabilities, which allow them to effectively cope with out-of-distribution scenarios.

The third requirement is multitask collaboration. With the application of UAVs in industries such as manufacturing agriculture, and power, their tasks are becoming more diverse. Tasks such as disaster reconnaissance, agricultural sowing, and power inspection require training models based on corresponding scenarios. An LLM, relying on its strong generalization ability, can quickly adapt to different scenarios and switch task scenarios through prompt engineering and model training.

In addition, during the selection phase of the system technology roadmap, we conducted extensive simulation experiments to test similar methods and compare the performance of different intelligent decision-making methods in unmanned aerial vehicle systems, as shown in

Table 2. Performance comparison of intelligent decision-making methods.

Method	Decision Accuracy	Response Time	Generalization Ability	Training Cost	Reasoning Memory
DQN	89.3% ± 2.1%	120 ms	68.5%	2000 GPU-h	4 GB
PPO	91.2% ± 1.8%	135 ms	71.2%	1500 GPU-h	3.5 GB
Rule engine	68.9% ± 3.5%	23 ms	45.3%	0	0.5 GB
Random forest	82.4% ± 2.3%	85 ms	76.8%	50 CPU-h	2 GB
This article’s LLM-7B	91.2% ± 1.5%	156 ms	85.6%	200 GPU-h	27 GB
This article’s LLM-3B	87.4% ± 1.8%	89 ms	82.3%	100 GPU-h	12 GB

.

The training cost of the LLMs in the table mainly includes the fine-tuning cost for unmanned aerial vehicle task scenarios and does not include pre-training costs. As this article uses an open-source LLM (LLaMA2) as the base model, most of the work expenses are in the post-training stage. In summary, this article has decided to use an LLM as the intelligent decision-making core, 5G MEC as the computing communication platform, and distributed deployment of large-scale models of different scales to support unmanned aerial vehicles in carrying out diverse tasks.

3.3. System Workflow

The system is designed for emergency response scenarios such as earthquakes, mudslides, and floods. Field intelligence is collected by UAVs, unmanned ground vehicles, and other sensors at the terminal tier. At this tier, the collected data are first preprocessed or feature-extracted by either a sensor-aggregation node or the lightweight LLM co-located with the base station. Based on the task’s type, requirements, and complexity, an execution policy is then selected: the task is offloaded to the lowest tier that still meets its minimal resource and latency constraints. After the central or edge tier has completed processing, the results are distributed back to the terminal endpoints via the 5G downlink. The overall workflow is illustrated in Figure 2.

3.4. System Scalability Design

In order to enhance the flexibility of the system in different scenarios, this article fully considers the scalability requirements during the system design phase, ensuring that the system can adapt to different application scenarios while meeting constantly changing application requirements. Therefore, we comprehensively consider four dimensions: horizontal expansion, vertical expansion, functional module expansion, and cross-domain collaborative expansion.

The first is the ability to expand horizontally. The central layer adopts a distributed cluster configuration, while the cloud computing center adopts containerized deployment and microservice architecture, supporting agile scaling of computing and storage nodes. The edge computing server of the edge layer adopts co-station deployment with the 5G base station and supports node plug and play, and the edge nodes are interconnected through standardized interfaces to build a flexible and elastic edge cloud. The terminal layer supports device access for different network standards, with devices equipped with 5G baseband and RF chips for on-demand access. Heterogeneous terminals access the 5G backbone network through protocol converters or IoT aggregation nodes.

The second is vertical scalability. In terms of computing resources, nodes at all levels support hardware upgrades and expansion. The central layer can increase GPU clusters to enhance computing power. The edge layer can expand computing resources by increasing computing power or upgrading processors and memory. Terminal devices can upgrade payloads or expand sensor computing and corresponding functions. In terms of storage resources, a distributed storage architecture is adopted, which can support linear expansion of storage resources for nodes at all levels as needed.

The third is the expansion of functional modules, mainly considering the expansion of intelligent engine models and other business functions. Large language models can be upgraded online and deployed locally, and vertical-domain models can be deployed according to task requirements, such as image segmentation, image recognition, and remote sensing image interpretation. The system deployment model management system can efficiently manage version iteration and rollback backtracking. The extension of other business functions supports plug-in deployment, enabling on-demand loading and agile deployment without the need for downtime upgrades.

The fourth is cross-domain collaborative expansion. On the one hand, the system supports distributed deployment, achieving regional linkage, efficient collaboration, and resource sharing. On the other hand, the system supports industry expansion, and the system architecture design fully considers the common needs of different scenarios while considering universality. Therefore, it can efficiently adapt to industrial applications such as power inspection, disaster monitoring, and logistics distribution.

3.5. System Reliability Design

UAVs require high reliability in critical scenarios such as emergency rescue, line inspection, and communication support. To this end, we enhance system reliability through three aspects of design: network security, redundancy design, and fault resolution.

The first is network security protection. On the one hand, security measures are designed for the access layer, network layer, and application layer. At the access layer, identity authentication mechanisms are set up through hardware certificates and dynamic keys, and access control whitelists and behavior patterns are dynamically controlled. In addition, end-to-end AES-256 encryption is used to ensure data transmission security. At the network layer, logical isolation is achieved through VLANs and software-defined networks, while a real-time network traffic monitoring mechanism is set up to monitor abnormal communication modes. An API gateway is built at the application layer for agreeing to API access control and traffic restrictions, and data anonymization and role-based granular permission control mechanisms are set up at the edge layer to ensure network layer security. On the other hand, intelligent intrusion detection systems are deployed at the edge and center layers, with built-in intrusion detection algorithms that perform abnormal behavior detection, threat intelligence fusion, and automatic response mechanisms.

The second is redundant design. On the one hand, design redundancy mechanisms are established for communication links. Multiple independent communication links are established, with 5G as the backbone network for the main link and satellite or shortwave communication as the backup link. On the other hand, redundancy design is implemented for computing resources, deploying N + 2 service instances at the central layer and deploying the same instance as a backup in data centers in different geographical regions to achieve automatic transfer in case of failure. The edge layer adopts an adjacent node backup scheme, and tasks are automatically offloaded to the optimal node based on optimization algorithms according to resource usage. In addition, data storage also adopts redundant settings, maintaining three copies of critical data backup, backing up data in data centers (DCs) in different regions, and keeping all three modified synchronously.

The third is the fault resolution mechanism. A multidimensional fault detection and resolution mechanism is established. One is the fault detection system, which regularly monitors the health status of hardware devices, monitors real-time indicators such as CPU, memory, and network, and uses large language models for predictive maintenance based on historical data. It conducts regular usability checks on software, analyzes abnormal logs, and monitors performance baseline deviations. Regular connectivity testing is conducted to check for network failures, monitor real-time latency changes, and measure packet loss rates. The second is automatic fault transfer. A multi-level fault transfer strategy is designed, with service-level transfer using automatic restart, instance switching, and downgrade service strategies. The node level adopts task migration, data synchronization, and resource reallocation strategies; system-level transfer adopts strategies such as backup system activation, data recovery, and business continuity support. The third is the service recovery strategy. Measures such as fault isolation, service reconstruction, and state recovery are adopted to support rapid recovery from system failures, ensuring data consistency through transaction rollback, data verification, and incremental synchronization.

4. System Modeling

This section establishes corresponding mathematical models for the designed system architecture and workflow, providing support for the subsequent introduction of LLMs and algorithm empowerment. Section 4.1 describes the network model; Section 4.2 and Section 4.3 establish corresponding models from the perspectives of communication and computation; Section 4.4 models the inference and accuracy of the LLM; Section 4.5 establishes the system energy consumption model, including communication, computation, and flight energy consumption; Section 4.6 and Section 4.7 model the system performance indicators and robustness; and Section 4.8 describes the optimization problem, creating conditions for subsequent algorithm introduction.

4.1. Network Model

Consider a three-layer architecture system consisting of $N$ UAV terminals, $M$ edge nodes, and 1 cloud center. Let $\mathcal{N}=\{1,2,\dots ,N\}$ be the set of UAVs, $\mathcal{M}=\{1,2,\dots ,M\}$ the set of edge nodes, and e the set of tasks to be processed.

Each task $k\in \mathcal{K}$ has the following attributes: $D_k$ represents input data size (bits), $C_k$ represents computational complexity (CPU cycles), $T_k^{max}$ represents maximum tolerable latency (seconds), ${\omega }_k$ represents task priority weight, and ${ϵ}_k$ represents result accuracy requirement.

The basic attributes of UAV $i$ include the following: $f_i^{local}$ represents local CPU frequency (Hz), $P_i^{max}$ represents maximum transmit power (watts), $E_i^{battery}$ represents battery capacity (joules), $v_i$ represents flight speed (m/s), and $\left(x_i, y_i, z_i\right)$ represents 3D coordinate position.

Since the large language models deployed at different levels of cloud, edge, and end are responsible for solving different types of tasks, the system uses a dynamic scheduling mechanism to determine the specific level of offloading based on the estimated amount of computation, response requests, and queuing sequence, combined with the advanced rules of edge computing and cloud computing centers, as well as the requirements of the task scenario for delay, network conditions, resource availability, and so on.

Table 3. Example of offloading UAV tasks in emergency rescue scenarios.

Task Domain	Specific Task Type	Specific Task Type	Task Offloading Hierarchy
Natural Language Understanding	Instruction translation	Fly to a certain coordinates	Terminal Layer
	Semantic understanding	Multistep task planning in fire scenarios	Edge layer
	Complex reasoning	Emergency response plan generation for this scenario	Central layer
Image Recognition	Basic object detection	Real-time obstacle recognition	Terminal Layer
	Scene understanding analysis	Disaster situation assessment	Edge layer
	Visual reasoning	Global situation analysis	Central layer

lists common task types and example offloading levels.

4.2. Communication Model

4.2.1. Air-to-Ground Channel Model

Communication between the UAV and the terrestrial base station is modeled using an Air-to-Ground (A2G) channel [19]. Both Line-of-Sight (LoS) and Non-Line-of-Sight (NLoS) propagation paths are considered, and the channel gain is expressed as

$$h_{i,j}=\left\{\begin{array}{ll}{\eta }_{\mathrm{LoS}}·d_{i,j}^{-{\alpha }_{\mathrm{LoS}}}·{\xi }_{i,j},\hfill & \mathrm{if} \mathrm{LoS}\hfill \\ {\eta }_{\mathrm{NLoS}}·d_{i,j}^{-{\alpha }_{\mathrm{NLoS}}}·{\xi }_{i,j},\hfill & \mathrm{if} \mathrm{NLoS}\hfill \end{array}\right.$$

(1)

where ${\eta }_{\mathrm{LoS}}$ and ${\eta }_{\mathrm{NLoS}}$ denote the path-loss constants for the LoS and NLoS links, respectively; ${\alpha }_{\mathrm{LoS}}$ and ${\alpha }_{\mathrm{NLoS}}$ are the corresponding path-loss exponents; and ξ represents the shadow-fading component, which follows a log-normal distribution.

The LoS probability is a function of the UAV’s altitude and its horizontal distance from the base station and is expressed as

$$P_{\mathrm{LoS}}(h_i,d_{i,j}^{hor})={\displaystyle \frac{1}{1+a·\exp (-b({\theta }_{i,j}-a))}}$$

(2)

where ${\theta }_{i,j}=\arctan (h_i/d_{i,j}^{hor})$ is the elevation angle, and $a$ and $b$ are environment-dependent parameters.

4.2.2. Data Transmission Rate

Based on Shannon’s theorem, the uplink transmission rate from UAV $i$ to edge node $j$ is given by

$$R_{i,j}^{up}=B_{i,j}{\log }_2\left(1+{\displaystyle \frac{P_{i,j}·h_{i,j}}{{\sigma }^2+{\displaystyle \sum _{k\in {\mathcal{I}}_{i,j}}{P}_{k,j}}·h_{k,j}}}\right)$$

(3)

where $B_{i,j}$ denotes the bandwidth allocated to link $\left(i,j\right)$, $P_{i,j}$ is the transmit power of UAV $i$ on this link, ${\sigma }^2$ represents the noise power, and ${\mathcal{I}}_{i,j}$ is the set of UAVs that create interference toward link $\left(i,j\right)$.

Accounting for the OFDMA technology in 5G networks, the bandwidth allocation must satisfy

$${\displaystyle \sum _{i=1}^N{B}_{i,j}}\le B_j^{total}, \forall j\in \mathcal{M}$$

(4)

$$B_{i,j}^{min}\le B_{i,j}\le B_{i,j}^{max},\hspace{0.33em}\hspace{0.33em}\hspace{0.33em} \forall i\in \mathcal{N},j\in \mathcal{M}$$

(5)

4.2.3. Transmission Delay Model

The delay for task $k$ transmitted from UAV $i$ to edge node $j$ comprises propagation delay, transmission delay, and protocol overhead:

$${\tau }_{k,i,j}^{trans}={\displaystyle \frac{d_{i,j}}{c}}+{\displaystyle \frac{D_k}{R_{i,j}^{up}}}+{\tau }_{\mathrm{protocol}}$$

(6)

where $c$ is the speed of light, and ${\tau }_{\mathrm{protocol}}$ is the protocol-processing overhead.

Owing to the UAV’s mobility, the channel state varies; thus, the average rate during transmission must account for the Doppler effect:

$$R_{i,j}^{avg}={\displaystyle {\int }_0^{T_{trans}}{R}_{i,j}}(t)dt/T_{trans}$$

(7)

where $R_{ij}(t)$ is the instantaneous transmission rate at time $t$.

4.3. Computation Model

4.3.1. Local Computation Model

The latency for UAV $i$ to process task $k$ locally is

$${\tau }_{k,i}^{local}={\displaystyle \frac{C_k}{f_i^{local}}}+{\delta }_k^{overhead}$$

(8)

where ${\delta }_k^{overhead}$ represents system overhead, including operating-system scheduling and memory-access latency.

The energy consumption model for local computation accounts for both dynamic and static power:

$$E_{k,i}^{local}=\kappa ·{(f_i^{local})}^2·C_k+P_i^{static}·{\tau }_{k,i}^{local}$$

(9)

where $\kappa$ is the effective switched capacitance, and $P_i^{static}$ is the static power consumption.

4.3.2. Edge Computing Model

Edge node $j$’s computing resources are dynamically allocated through virtualization. Let the total computing capacity of edge node $j$ be $F_j^{total}$, and let the amount of computing resources allocated to task $k$ be $f_{k,j}^{edge}$.

Edge computing latency comprises the queuing delay plus the computation delay:

$${\tau }_{k,j}^{edge}={\tau }_{k,j}^{queue}+{\displaystyle \frac{C_k}{f_{k,j}^{edge}}}$$

(10)

The queuing delay is modeled as an M/M/1 queue that takes task priorities into account:

$${\tau }_{k,j}^{queue}={\displaystyle \frac{{\lambda }_j}{{\mu }_j({\mu }_j-{\lambda }_j)}}·\left(1+{\displaystyle \sum _{l\in {\mathcal{H}}_k}\frac{{\rho }_l}{1-{\rho }_l}}\right)$$

(11)

where ${\lambda }_j$ is the task arrival rate, ${\mu }_j$ is the service rate, ${\mathcal{H}}_k$ is the set of tasks with higher priority than task $k$, and ${\rho }_l$ is the traffic intensity of task $l$.

The allocation of computing resources at the edge node must satisfy

$${\displaystyle \sum _{k\in {\mathcal{K}}_j}{f}_{k,j}^{edge}}\le F_j^{total}$$

(12)

$$f_{k,j}^{min}\le f_{k,j}^{edge}\le f_{k,j}^{max}$$

(13)

where ${\mathcal{K}}_j$ is the set of tasks assigned to edge node $j$.

4.3.3. Cloud Computing Model

The cloud has ample computing resources; the dominant factor is network transmission delay.

$${\tau }_k^{cloud}={\tau }_{k,i,cloud}^{trans}+{\displaystyle \frac{C_k}{f_k^{cloud}}}+{\tau }_{k,cloud,i}^{back}$$

(14)

Let ${\tau }_{k,cloud,i}^{back}$ denote the round-trip delay for returning results.

Because the backhaul link from the cloud to the edge typically offers higher bandwidth, the return delay is comparatively small:

$${\tau }_{k,cloud,i}^{back}={\displaystyle \frac{D_k^{result}}{R_{cloud,i}^{down}}}$$

(15)

where $D_k^{result}$ is the size of the result data, which is usually much smaller than the input data.

4.4. Large Language Model Inference Model

4.4.1. Hierarchical Model Deployment

Based on model complexity and computational demand, the large language model is divided into three tiers:

Cloud full model: ${\mathrm{\Theta }}^{cloud}=175\times {10}^9$ parameters, supporting complex inference tasks.

Edge compressed model: ${\mathrm{\Theta }}^{edge}=7\times {10}^9$ parameters, obtained via knowledge distillation.

Terminal lightweight model: ${\mathrm{\Theta }}^{local}=0.7\times {10}^9$ parameters, dedicated to emergency decisions.

4.4.2. Inference Latency Modeling

The inference delay of an LLM is mainly composed of three factors: linear computational complexity, quadratic computational complexity, and hardware limitations [20]. The mainstream LLMs, including LLaMA used in this article, are usually based on the transformer architecture [21], and their total parameter count can be represented by the following equation:

$$\mathrm{\Theta } \approx L_{layers} \times d_{model} \times \left(4d_{model} + d_{ff}\right)$$

(16)

where $d_{model}$ represents the model dimension, $d_{ff}$ represents the feedforward layer dimension, and $L_{layers}$ represents the number of layers. In the self-attention mechanism, the projection of query, health, and value is $3 \times L_{seq} \times {d_{model}}^2$, the attention weight is represented as ${L_{seq}}^2 \times d_{model}$, and the output projection is $L_{seq} \times {d_{model}}^2$. The feedforward network has two projection layers with different dimensions: the first layer is represented as $L_{seq} \times d_{model} \times d_{ff}$, and the second layer is represented as $L_{seq} \times d_{ff} \times d_{model}$. Therefore, the total computational complexity is represented as

$${\mathrm{FLOP}}_{\mathrm{total}}=L_{layers}\times \left(6\times L_{seq}\times {d_{model}}^2+{L_{seq}}^2\times d_{model}+2\times L_{seq}\times d_{model}\times d_{ff}\right)$$

(17)

The complexity of feedforward networks and self-attention mechanisms can be expressed as

$$O(L_{seq}·d_{model}·d_{ff})$$

(18)

$$O(L_{seq}^2·d_{model})$$

(19)

The former has a linear relationship with sequence length and parameter quantity, while the latter has a quadratic relationship with sequence length. Hardware bottlenecks are mainly limited by computational frequency and memory bandwidth and are related to model parameter quantity and hardware performance [22]. Therefore, inference delay can be modeled as

$${\tau }_{LLM}=\alpha ·L_{seq}·\mathrm{\Theta }+\beta ·L_{seq}^2+\gamma ·{\displaystyle \frac{\mathrm{\Theta }}{f_{compute}}}$$

(20)

where $L_{seq}$ is the length of the input sequence; $\alpha$, $\beta$, and $\gamma$ are model-related constants; and $f_{compute}$ is the frequency of the computing unit. For the transformer architecture, the computational complexity of the self-attention mechanism is $O(L_{seq}^2·d_{model})$, where $d_{model}$ is the model dimension.

4.4.3. Model Accuracy Modeling

The commonly used methods for compressing large language models include quantization, pruning, and knowledge distillation. Model compression is based on information theory, and the relationship between model capacity and parameter quantity follows the logarithmic law:

$$Capacity \propto log\left(\mathrm{\Theta }\right)$$

(21)

Therefore, when the model is compressed from the original parameter quantity ${\mathrm{\Theta }}_{original}$ to ${\mathrm{\Theta }}_{compressed}$, the loss can be modeled as

$$Los{s}_{info}=log\left({\mathrm{\Theta }}_{original}/{\mathrm{\Theta }}_{compressed}\right)$$

(22)

The performance can be maintained at $Retentio{n}_{rate} = 1 - {\lambda }_{distill} \times Los{s}_{info}$; therefore the final accuracy of the LLM after compression can be expressed as

$$Ac{c}_{compressed}=Ac{c}_{original}·(1-{\lambda }_{distill}·\log ({\displaystyle \frac{{\mathrm{\Theta }}_{original}}{{\mathrm{\Theta }}_{compressed}}}))$$

(23)

where ${\lambda }_{distill}$ is the knowledge distillation loss coefficient. In addition, we have compiled experimental compression data from mainstream LLMs [23,24,25,26], which are summarized in

Table 4. Experimental data of mainstream LLM compression.

Original Model	Model Compression	Compression Ratio	Accuracy Retention Rate	${\mathit{\lambda }}_{\mathit{d}\mathit{i}\mathit{s}\mathit{t}\mathit{i}\mathit{l}\mathit{l}}$
GPT-3 175B	GPT-3 13B	13.5	0.87	0.051
BERT-Large	BERT-Base	3.0	0.95	0.042
T5-11B	T5-3B	3.7	0.91	0.068
LLaMA-65B	LLaMA-7B	9.3	0.82	0.088

.

According to statistical data, an average of ${\lambda }_{distill}=0.062\pm 0.018$ can be obtained, while the goodness of fit $R^2 = 0.94$.

4.5. Energy Consumption Model

4.5.1. Communication Energy

The UAV’s communication energy comprises the power consumed by the RF power amplifier and baseband processing:

$$E_{comm}={\displaystyle \frac{P_{tx}}{{\eta }_{PA}}}·T_{trans}+P_{circuit}·T_{trans}$$

(24)

where $P_{tx}$ is the transmit power, ${\eta }_{PA}$ is the power-amplifier efficiency, and $P_{circuit}$ is the circuit power.

The power-amplifier efficiency is a function of transmit power:

$${\eta }_{PA}={\eta }_{max}·\left(1-\exp \left(-{\displaystyle \frac{P_{tx}}{P_{sat}}}\right)\right)$$

(25)

where ${\eta }_{max}$ is the maximum efficiency, and $P_{sat}$ is the saturation power.

4.5.2. Computation Energy

Computation energy depends on processor operating frequency and computational complexity.

$$E_{comp}=\kappa ·f^2·C+P_{static}·T_{comp}$$

(26)

For GPU computation, memory-access energy must also be considered:

$$E_{memory}=N_{access}·E_{per\_access}$$

(27)

where $N_{access}$ is the number of memory accesses, and $E_{per\_access}$ is the energy per access.

4.5.3. Flight Energy

The UAV’s flight energy consumption is related to its speed, weight, and aerodynamic characteristics:

$$E_{flight}={\displaystyle \frac{1}{2}}\rho S{C}_D{v}^3·T_{flight}+{\displaystyle \frac{m{g}^2}{2\rho S}}·{\displaystyle \frac{1}{v}}·T_{flight}$$

(28)

where $\rho$ is air density, $S$ is the reference area, $C_D$ is the drag coefficient, $m$ is the mass, and $g$ is the gravitational acceleration.

4.6. System Performance Metrics

4.6.1. Latency Metric

The overall system latency is the weighted sum of all task latencies:

$${\mathrm{\Psi }}_{delay}={\displaystyle \sum _{k=1}^K{\omega }_k}·{\displaystyle \frac{{\tau }_k^{total}}{T_k^{max}}}$$

(29)

where ${\tau }_k^{total}$ denotes the total latency of task $k$, including transmission, queuing, computation, and return delays.

4.6.2. Energy Metric

The total system energy consumption comprises the communication, computation, and flight energy of all UAVs:

$${\mathrm{\Psi }}_{energy}={\displaystyle \sum _{i=1}^N\frac{E_i^{total}}{E_i^{battery}}}$$

(30)

where $E_i^{total}=E_i^{comm}+E_i^{comp}+E_i^{flight}$ is the corresponding energy weighting factor.

4.6.3. Reliability Metric

System reliability accounts for communication interruption, computation failure, and model accuracy degradation:

$${\mathrm{\Psi }}_{reliability}={\displaystyle \sum _{k=1}^K{\omega }_k}·(1-P_k^{success})$$

(31)

where the probability of successful task completion is

$$P_k^{success}=P_k^{comm}·P_k^{comp}·P_k^{model}$$

(32)

4.7. System Robustness Modeling

System robustness is an important indicator of the ability of unmanned aerial vehicles to perform tasks under complex conditions. This section defines and models system robustness.

4.7.1. Definition of Robustness Indicators

The robustness of the system is quantified through the following four dimensions:

$$\mathrm{Availability} \mathrm{indicator}: A_{sys}={\displaystyle \frac{MTBF}{MTBF+MTTR}}$$

where $MTBF$ is the average time between failures, and $MTTR$ is the average time to repair.

$$\mathrm{Performance} \mathrm{degradation} \mathrm{degree}: D_{perf}=1-{\displaystyle \frac{P_{fault}}{P_{normal}}}$$

where $P_{fault}$ represents the system performance under fault conditions, and $P_{normal}$ represents the system performance under normal conditions.

$$\mathrm{Fault} \mathrm{tolerance}: F_{tolerance} ={\displaystyle \frac{N_{success}}{N_{total}}}$$

where $N_{success}$ is the number of tasks successfully completed under fault conditions, and $N_{total}$ is the total number of tasks.

$$\mathrm{Recovery} \mathrm{ability}: R_{recovery}=e^{-\lambda t_{recovery}}$$

where $\lambda$ is the recovery rate parameter, and $t_{recovery}$ is the system recovery time.

4.7.2. Fault Model

(1)

Communication link failure model

Considering the randomness of link failures, the available probability of link a is modeled as

$$P_{link}(i,j)=e^{-{\lambda }_{link}·d_{i,j}}·(1-I_{interference})\mathrm{v}$$

(33)

where ${\lambda }_{link}$ is the link failure rate, $d_{i,j}$ is the distance between nodes, and $I_{interference}$ is the interference indicator function.

(2)

Calculate node fault model

The failure probability of edge node $j$ is related to its load:

$$P_{fail}(j)=1-e^{-\alpha ·U_j^{\beta }}$$

(34)

where $U_j={\displaystyle \frac{{\displaystyle \sum _{k\in {\mathcal{K}}_j}{f}_{k,j}^{edge}}}{F_j^{total}}}$ is the resource utilization rate of node $j$, and $\alpha$ and $\beta$ are fault model parameters.

(3)

Environmental interference model

The impact of environmental interference on system performance is modeled as follows:

$$I_{env}(t)=I_{base}+{\displaystyle \sum _{m=1}^M{A}_m}\sin (2\pi f_{m}t+{\varphi }_m)$$

(35)

where $I_{base}$ is the basic interference level, and $A_m$, $f_m$, and ${\varphi }_m$ are the amplitude, frequency, and phase of the $m$ interference source, respectively.

4.7.3. Cascade Fault Model

Considering the interdependence between nodes in the system, a single fault may trigger cascading effects:

$$P_{cascade}(t)=P_{initial}·{\displaystyle \prod _{k=1}^K(1+{\beta }_k·P_{dependency}^k(t))}$$

(36)

where $P_{initial}$ is the initial failure probability, ${\beta }_k$ is the cascading factor, and $P_{dependency}^k(t)$ is the $k$-th level dependent failure probability.

4.7.4. System Elasticity Model

System resilience reflects the ability of a system to recover from a faulty state to a normal state:

$$\mathrm{Resilience}={\displaystyle {\int }_0^{T}Q}(t)dt/T$$

(37)

where $Q(t)$ is the system performance quality function at time $t$:

$$Q(t)=\left\{\begin{array}{ll}1,\hfill & normal\hfill \\ Q_{degraded}(t),\hfill & degraded\hfill \\ 0,\hfill & fault\hfill \end{array}\right.$$

(38)

Performance quality in degraded state:

$$Q_{degraded}(t)=Q_{min}+(1-Q_{min})·e^{-\gamma (t-t_{fault})}$$

(39)

where $Q_{min}$ is the minimum performance level, $\gamma$ is the recovery rate, and $t_{fault}$ is the time of failure occurrence.

4.8. Optimization Problem Description

Integrating latency, energy, and reliability, the system optimization objective is

$$\min \mathrm{\Phi }={\lambda }_1{\mathrm{\Psi }}_{delay}+{\lambda }_2{\mathrm{\Psi }}_{energy}+{\lambda }_3{\mathrm{\Psi }}_{reliability}$$

(40)

Decision variables:

$x_{k,d}\in \{0,1\}$: deployment decision for task $k$ ($d\in \{local,edge,cloud\}$); $f_{k,j}^{edge}$: computing resources assigned by edge node $j$ to task $k$; $B_{i,j}$: bandwidth allocation on link $\left(i,j\right)$; $P_{i,j}$: power allocation on link $\left(i,j\right)$.

Constraints are as follows:

Deployment constraint: ${\displaystyle \sum _d{x}_{k,d}}=1,\hspace{0.33em}\hspace{0.33em}\hspace{0.33em} \forall k$; resource constraint: ${\displaystyle \sum _{k\in {\mathcal{K}}_j}{f}_{k,j}^{edge}}\le F_j^{total},\hspace{0.33em}\hspace{0.33em}\hspace{0.33em} \forall j{\displaystyle \sum _i{B}_{i,j}}\le B_j^{total},\hspace{0.33em}\hspace{0.33em}\hspace{0.33em} \forall j{\displaystyle \sum _j{P}_{i,j}}\le P_i^{max},\hspace{0.33em}\hspace{0.33em}\hspace{0.33em} \forall i$; QoS constraint: ${\tau }_k^{total}\le T_k^{max},\hspace{0.33em}\hspace{0.33em}\hspace{0.33em} \forall k{E}_i^{total}\le E_i^{battery},\hspace{0.33em}\hspace{0.33em}\hspace{0.33em} \forall i$.

5. Improved Grey Wolf Optimizer

5.1. Basic Grey Wolf Optimizer

The Grey Wolf Optimization Algorithm [27] simulates the hierarchical concept and hunting behavior of wolf packs. The algorithm divides the wolf pack into four levels: $\alpha$ (optimal solution), $\beta$ (suboptimal solution), $\delta$ (third optimal solution), and $\omega$ (remaining solutions). The hunting process includes three stages: search, encirclement, and attack.

5.1.1. Encircling Prey

Wolves encircle the prey by adjusting their positions; the position update equation is

$$\overrightarrow{D}=|\overrightarrow{C}·{\overrightarrow{X}}_p(t)-\overrightarrow{X}(t)|$$

(41)

$$\overrightarrow{X}(t+1)={\overrightarrow{X}}_p(t)-\overrightarrow{A}·\overrightarrow{D}$$

(42)

where ${\overrightarrow{X}}_p$ is the prey’s position, $\overrightarrow{X}$ is the wolf’s current position, and $\overrightarrow{A}$ and $\overrightarrow{C}$ are coefficient vectors.

5.1.2. Hunting Mechanism

Since the prey’s exact location is unknown, the algorithm assumes that the $\alpha$, $\beta$, and $\delta$ wolves possess better information about its position; the remaining wolves update their own positions based on these three leaders:

$${\overrightarrow{D}}_{\alpha }=|{\overrightarrow{C}}_1·{\overrightarrow{X}}_{\alpha }-\overrightarrow{X}|{\overrightarrow{D}}_{\beta }=|{\overrightarrow{C}}_2·{\overrightarrow{X}}_{\beta }-\overrightarrow{X}|{\overrightarrow{D}}_{\delta }=|{\overrightarrow{C}}_3·{\overrightarrow{X}}_{\delta }-\overrightarrow{X}|$$

(43)

$${\overrightarrow{X}}_1={\overrightarrow{X}}_{\alpha }-{\overrightarrow{A}}_1·{\overrightarrow{D}}_{\alpha }{\overrightarrow{X}}_2={\overrightarrow{X}}_{\beta }-{\overrightarrow{A}}_2·{\overrightarrow{D}}_{\beta }{\overrightarrow{X}}_3={\overrightarrow{X}}_{\delta }-{\overrightarrow{A}}_3·{\overrightarrow{D}}_{\delta }$$

(44)

$$\overrightarrow{X}(t+1)={\displaystyle \frac{{\overrightarrow{X}}_1+{\overrightarrow{X}}_2+{\overrightarrow{X}}_3}{3}}$$

(45)

5.1.3. Attacking and Searching for Prey

Parameter $\overrightarrow{a}$ decreases linearly from 2 to 0 to balance exploration and exploitation:

$$\overrightarrow{A}=2\overrightarrow{a}·{\overrightarrow{r}}_1-\overrightarrow{a}\overrightarrow{C}=2·{\overrightarrow{r}}_2$$

(46)

where ${\overrightarrow{r}}_1$ and ${\overrightarrow{r}}_2$ are random vectors in the range [0, 1].

5.2. Algorithmic Enhancement Strategy

5.2.1. Chaotic-Map Initialization

Traditional random initialization often yields uneven population distributions and degrades algorithmic performance. This paper adopts the Tent chaotic map [28] to generate the initial population:

$$x_{n+1}=\left\{\begin{array}{ll}{\displaystyle \frac{x_n}{\mu }},\hfill & 0\le x_n<\mu \hfill \\ {\displaystyle \frac{1-x_n}{1-\mu }},\hfill & \mu \le x_n\le 1\hfill \end{array}\right.$$

(47)

where $\mu =0.7$ is the control parameter. The chaotic sequence is then mapped to the search space as

$$X_i^{(0)}=X_{min}+(X_{max}-X_{min})·x_i$$

(48)

5.2.2. Adaptive Weight Strategy

To balance global exploration and local exploitation, an adaptive weight adjustment mechanism is designed:

$$w(t)=w_{max}-(w_{max}-w_{min})·{\left({\displaystyle \frac{t}{T_{max}}}\right)}^2$$

(49)

where $w_{max} = 0.9, w_{min} = 0.1, T_{max}$ is the maximum iteration count.

The weight parameters are dynamically tuned according to the algorithm’s convergence state:

$$w_{adaptive}=w(t)·\left(1+{\displaystyle \frac{{\sigma }_{fitness}}{\left|{\mu }_{fitness}\right|}}\right)$$

(50)

where ${\sigma }_{fitness}$ and ${\mu }_{fitness}$ are the standard deviation and mean of the current population fitness, respectively.

5.2.3. Lévy Flight Search Strategy

Lévy flight [29] exhibits long-range correlation and can enhance the global search capability of the algorithm. The probability density function of the Lévy distribution is

$$L(s)~s^{-1-\beta },\hspace{0.33em}\hspace{0.33em}\hspace{0.33em} 0<\beta \le 2$$

(51)

Lévy flight step lengths are generated using the Mantegna algorithm:

$$step={\displaystyle \frac{u}{|v{|}^{1/\beta }}}$$

(52)

where $u~N(0,{\sigma }_u^2)$, $v~N(0,{\sigma }_v^2)$, and

$${\sigma }_u={\left[{\displaystyle \frac{\mathrm{\Gamma }(1+\beta )\sin (\pi \beta /2)}{\mathrm{\Gamma }((1+\beta )/2)\beta 2^{(\beta -1)/2}}}\right]}^{1/\beta }$$

(53)

${\sigma }_v=1$ during position updating. Lévy flight is incorporated as follows:

$${\overrightarrow{X}}_{new}={\overrightarrow{X}}_{old}+\alpha ·Levy(\beta )\otimes ({\overrightarrow{X}}_{best}-{\overrightarrow{X}}_{old})$$

(54)

where $\alpha$ is a step-size control parameter, and $\otimes$ denotes element-wise multiplication.

5.2.4. Elite Learning Strategy

To accelerate convergence and prevent the loss of high-quality solutions, an elite learning mechanism is designed. The population is sorted by fitness, and the top 20% of individuals are selected as the elite group. The position update formula for elite individuals is

$${\overrightarrow{X}}_{elite}(t+1)={\overrightarrow{X}}_{elite}(t)+\varphi ·({\overrightarrow{X}}_{\alpha }-{\overrightarrow{X}}_{elite}(t))+\psi ·rand(-1,1)$$

(55)

where $\varphi$ is the learning factor, and $\psi$ is the perturbation intensity.

5.2.5. Dynamic Boundary Handling

When a search solution exceeds the boundary, a dynamic boundary-handling strategy is adopted:

$$X_i^{new}=\left\{\begin{array}{ll}{X}_{min}+rand·(X_{max}-X_{min}),\hfill & X_i<X_{min}\hfill \\ X_{max}-rand·(X_{max}-X_{min}),\hfill & X_i>X_{max}\hfill \\ X_i,\hfill & otherwise\hfill \end{array}\right.$$

(56)

5.3. The Framework of the Improved Algorithm

The overall framework of the Improved Grey Wolf Optimizer (ILGWO) (Algorithm 1) is presented below.

Algorithm 1. ILGWO

ILGWO pseudocode

Input: Population size $N$, maximum number of iterations $T_{max}$, problem dimension $D$. Output: Optimal solution $X_{best}$. 1. Initialization phase  Generate initial population using Tent chaotic mapping.  Calculate the fitness value of each individual.  Identify $X_{\alpha }$, $X_{\beta }$, $X_{\delta }$. 2. while $t < T_{max}$ do 3. for each wolf $X_i$ do  Calculate adaptive weight $w_{adaptive}$.  Update coefficients $\overrightarrow{A}$ and $\overrightarrow{C}$.  Update the current wolf position based on the $\alpha , \beta , \delta$ wolf positions.  Apply Lévy flight search strategy.  Perform dynamic boundary processing. 4. end for 5. Elite Learning Update  Sort populations by fitness.  Perform learning updates on elite individuals. 6. Update the leader  Recalculate fitness value.  Update $X_{\alpha }$, $X_{\beta }$, $X_{\delta }$. 7. $t = t + 1$ 8. end while 9. return $X_{\alpha }$

5.4. Algorithm Complexity Analysis

The time complexity of the ILGWO algorithm mainly consists of the following parts:

$$\mathrm{initialization}: O(N·D)$$

Fitness calculation: $O(N·f_{eval})$, where $f_{eval}$ is the complexity of evaluating the fitness function.

$$\mathrm{Location} \mathrm{update}: O(N·D)$$

$$\mathrm{Sorting} \mathrm{operation}: O(N\log N)$$

$$\mathrm{The} \mathrm{overall} \mathrm{time} \mathrm{complexity} \mathrm{is} O(T_{max}·N·(D+f_{eval}+\log N))$$

Space complexity is mainly used to store population individuals and temporary variables, known as $O(N·D)$.

Compared with the basic GWO algorithm, the increase in time complexity of ILGWO mainly comes from Lévy flight calculations and elite learning operations, but the complexity of these operations is still at a linear level and will not significantly affect the efficiency of the algorithm.

5.5. Algorithm Scalability Analysis

5.5.1. Parallel Computing Expansion

If the edge nodes are cluster servers, the ILGWO algorithm naturally supports parallel evaluation of individual populations. The fitness calculation of each individual is independent of other individuals and can be executed in parallel on multi-core CPUs or GPUs, achieving near-linear acceleration effects. In addition, the algorithm can be extended to a distributed version, distributing the population across multiple nodes. Through scheduled information exchange and migration operations, distributed collaborative optimization is achieved, which is suitable for large-scale optimization problems.

5.5.2. Dynamic Scale Adjustment

Algorithms can dynamically adjust population size based on task complexity and available resources, efficiently adapting to various task scenarios. In addition, for large-scale problems, algorithms can adopt a hierarchical optimization strategy, decomposing complex problems into subproblems, optimizing them separately, and then coordinating globally.

5.5.3. Multi-Objective Expansion

The algorithm can be extended to a multi-objective optimization version while optimizing multiple objectives, such as latency, energy consumption, and reliability, generating a Pareto-optimal solution set for decision selection. In addition, supporting more complex constraint conditions, including dynamic constraints, probabilistic constraints, etc., enhances the adaptability of the algorithm in practical applications.

6. Experimental Design and Result Analysis

6.1. Experimental Environment and Parameter Settings

6.1.1. Hardware Platform

The experiment was conducted on the following platform: server: 2 × Intel Xeon Gold 5218R CPU (3.2 GHz), 1 T DDR4 memory, 4 × NVIDIA A100 GPU. Edge node: Intel i9-11900K CPU (3.5 GHz), 64 GB DDR4 memory, NVIDIA RTX 3090 GPU. UAV simulation: based on AirSim platform, equipped with ARM Cortex-A78 core and 8 GB memory.

6.1.2. Network Environment Configuration

Fifth-generation (5G) network: Center frequency of 3.5 GHz, bandwidth of 100 MHz, transmission power of 43 dBm. WiFi backhaul: IEEE 802.11ax standard 6 GHz frequency band, bandwidth of 160 MHz. Satellite link: Ka band, uplink 30–31 GHz, downlink 20–21 GHz, bandwidth of 500 MHz [30].

6.1.3. Algorithm Parameter Settings

The ILGWO algorithm parameter settings are as follows:

Population size: $N = 50$; maximum iteration number: $T_{\max }= 500$; Lévy flight parameter: $\beta =1.5$; adaptive weights: $w_{\max } = 0.9$, $w_{\min } = 0.1$; elite ratio: 20%.

Comparative algorithms include the following:

Basic Grey Wolf Optimization Algorithm (GWO), Particle Swarm Optimization Algorithm (PSO), Genetic Algorithm (GA), Differential Evolution Algorithm (DE), and Artificial Bee Colony Algorithm (ABC).

6.1.4. Task Load Design

Five typical tasks were defined, as shown in

Table 5. Typical UAV testing tasks.

Task Type	Image Recognition	Video Analysis	Sensor Data Fusion	Path Planning	Natural Language Processing
Data size	1–5 MB	5–20 MB	50–200 KB	500 KB-2 MB	10–100 KB
Computational complexity	0.8–2.5 G cycle	2–8 G cycle	100–500 M cycle	500 M–3 G cycle	1–10 G cycle
Delay requirements	100–500 ms	0.5–2 s	10–50 ms	200 ms–1 s	50–200 ms
Accuracy requirements	Over 95%	Over 90%	Over 98%	Over 99%	Over 85%

.

6.2. Benchmark Function Testing

6.2.1. Test Function Set

Table 6. Characteristics of benchmark functions.

Function	Name	Type	Dimension	Search Interval	Global Optimal
F1	Sphere	Single peak	30	[−100, 100]	0
F2	Rosenbrock	Multimodal	30	[−30, 30]	0
F3	Rastrigin	Multimodal	30	[−5.12, 5.12]	0
F4	Griewank	Multimodal	30	[−600, 600]	0
F5	Ackley	Multimodal	30	[−32, 32]	0
F6	Weierstrass	Multimodal	30	[−0.5, 0.5]	0
F7	Schwefel	Multimodal	30	[−500, 500]	0
F8	Katsuura	Multimodal	30	[−100, 100]	0
F9	Lunacek	Multimodal	30	[−5, 5]	0
F10	Shifted Sphere	Single peak	30	[−100, 100]	100

shows some benchmark functions for CEC2017, from which we selected ten representative functions for testing.

6.2.2. Performance Evaluation Indicators

Algorithm performance was evaluated using the following metrics:

Mean fitness value (Mean): The average result of 30 independent runs.

Standard deviation (Std): Reflects the stability of the algorithm.

Best value: The best result among 30 runs.

Convergent Algebra (Conv): The number of iterations required to achieve threshold accuracy.

Success rate (SR): The proportion of runs that achieve the preset accuracy.

6.2.3. Benchmark Test Results

We show the results of running the relevant algorithms on the F1, F3, and F5 test functions for 30 rounds in

Table 7. Comparative experimental results of ILGWO and several classic algorithms.

Function	Algorithm	Mean	Std	Best	Conv	SR (%)
F1	ILGWO	2.31 × 10−15	8.45 × 10−16	1.23 × 10−18	78	100
	GWO	3.67 × 10−12	1.23 × 10−11	4.56 × 10−15	142	95
	PSO	1.45 × 10−8	3.21 × 10−8	2.31 × 10−10	298	78
	GA	2.34 × 10−5	1.12 × 10−4	8.91 × 10−7	456	45
F3	ILGWO	0.0024	0.0156	0	156	95
	GWO	2.45	3.78	0.0189	234	72
	PSO	15.67	12.34	3.45	387	23
	GA	45.23	23.45	12.78	489	8
F5	ILGWO	4.12 × 10−9	2.34 × 10−8	8.88 × 10−14	89	100
	GWO	1.23 × 10−6	4.56 × 10−6	3.45 × 10−8	167	89
	PSO	0.0034	0.0123	1.23 × 10−5	298	56
	GA	0.234	0.567	0.0345	445	12

. Obviously, the ILGWO algorithm has certain advantages.

6.2.4. Convergence Performance Analysis

Figure 3 selected five representative functions from the function test set, namely, F1 single peak, F3 multipeak, F5 multipeak (highly multimodal), F7 composite multipeak, and F10 with strong deception, and tested them separately, generating convergence curves. In terms of convergence speed, ILGWO converges quickly in the first 100 iterations, with a convergence speed about 35% faster than GWO and about 60% faster than PSO. In terms of convergence accuracy, ILGWO can converge to higher precision, reaching machine accuracy level on the F1 function. In terms of stability, the convergence curve of ILGWO is smoother and less volatile, indicating that the algorithm has good stability. In terms of global search capability, ILGWO is better able to jump out of local optima on the multimodal function F3, demonstrating the effectiveness of Lévy flight and elite learning strategies.

From the benchmark test results, it can be seen that the ILGWO algorithm is significantly better than the compared algorithms in terms of convergence accuracy, stability, and success rate. However, comparing the greatly improved ILGWO with the original algorithms, such as GWO, PSO, GA, etc., lacks a certain degree of fairness. In order to comprehensively evaluate the performance of the ILGWO algorithm, we conducted comparative experiments with improved metaheuristic algorithms proposed in recent years.

Table 8. Comparative experimental results of ILGWO and several advanced improved algorithms.

Function	Algorithm	Mean	Std	Best	Conv	SR(%)
F1	ILGWO	2.31 × 10−15	8.45 × 10−16	1.23 × 10−18	78	100
	IGWO-2023	4.56 × 10−14	1.23 × 10−13	2.34 × 10−16	98	95
	APSO-2024	1.23 × 10−12	3.45 × 10−12	5.67 × 10−15	134	90
	HGA-2023	2.34 × 10−11	6.78 × 10−11	1.23 × 10−13	156	85
	MOGWO-2024	3.45 × 10−13	8.91 × 10−13	4.56 × 10−16	112	92
	QPSO-2023	5.67 × 10−12	1.45 × 10−11	8.91 × 10−15	145	88
F2	ILGWO	0.0156	0.0234	0.0012	167	93
	IGWO-2023	0.0923	0.1456	0.0145	234	85
	APSO-2024	0.456	0.789	0.0892	298	72
	HGA-2023	0.234	0.567	0.0234	267	78
	MOGWO-2024	0.189	0.345	0.0189	245	82
	QPSO-2023	0.345	0.678	0.0456	289	75
F3	ILGWO	0.0024	0.0156	0	156	95
	IGWO-2023	0.0189	0.0234	0.0012	189	87
	APSO-2024	0.234	0.456	0.0234	234	78
	HGA-2023	0.156	0.289	0.0178	278	73
	MOGWO-2024	0.123	0.234	0.0045	198	82
	QPSO-2023	0.189	0.367	0.0289	256	80
F4	ILGWO	0.0012	0.0089	0	134	97
	IGWO-2023	0.0234	0.0456	0.0023	167	90
	APSO-2024	0.145	0.289	0.0156	245	75
	HGA-2023	0.089	0.167	0.0089	223	82
	MOGWO-2024	0.056	0.123	0.0034	189	85
	QPSO-2023	0.078	0.145	0.0067	198	83
F5	ILGWO	4.12 × 10−9	2.34 × 10−8	8.88 × 10−14	89	100
	IGWO-2023	2.34 × 10−6	8.91 × 10−6	1.23 × 10−8	145	88
	APSO-2024	0.0045	0.0123	2.34 × 10−6	267	70
	HGA-2023	0.0023	0.0067	4.56 × 10−7	234	75
	MOGWO-2024	1.23 × 10−5	4.56 × 10−5	3.45 × 10−8	178	85
	QPSO-2023	1.23 × 10−7	4.56 × 10−7	2.34 × 10−9	123	93
F6	ILGWO	1.234	3.456	0.0234	145	92
	IGWO-2023	12.34	23.45	2.345	189	83
	APSO-2024	45.67	78.91	8.912	298	68
	HGA-2023	34.56	56.78	5.678	267	72
	MOGWO-2024	23.45	34.56	3.456	234	78
	QPSO-2023	28.91	45.67	4.567	256	75

shows the comparative experimental results between ILGWO and several advanced improved algorithms in recent years. Including Improved Grey Wolf Optimization Algorithm (IGWO) [31], which combines elite reverse learning and Lévy flight strategy; Adaptive Particle Swarm Optimization (APSO) algorithm [32], which adopts dynamic parameter adjustment and diversity maintenance mechanism; Hybrid Genetic Algorithm (HGA) [33], which combines simulated annealing and a local search strategy; Multi-Objective Grey Wolf Algorithm (MOGWO) [34], which was specifically improved for multi-objective optimization problems; and Quantum Behavioral Particle Swarm Optimization (QPSO) [35], which is an improved PSO based on Quantum Mechanics Principles.

Figure 4 shows the convergence results of the algorithm performance comparison experiment.

Statistical significance test:

The Wilcoxon signed-rank test (α = 0.05) was used to statistically compare ILGWO with other improved algorithms, and the results showed the following:

Comparison between ILGWO and IGWS-2023: p = 0.0023 < 0.05, significant difference.

Comparison between ILGWO and APSO-2024: p = 0.0012 < 0.05, significant difference.

Comparison between ILGWO and other algorithms: p-values are all less than 0.05, indicating significant differences.

6.3. System Performance Simulation Experiment

6.3.1. Experimental Scene Setting

A simulation environment was constructed consisting of 20 UAVs, five edge nodes, and one cloud center. UAVs perform tasks within a 100 km × 100 km area, with edge nodes covering a radius of 20 km. The specific parameter configuration is shown in

Table 9. System parameter configuration.

Parameter Category	Parameter Name	Numerical Value
UAV	CPU frequency	1.5–2.0 GHz
	Battery capacity	500–800 Wh
	maximum power	20–30 W
	Flight speed	15–25 m/s
Edge node	CPU frequency	20–40 GHz
	Memory capacity	32–64 GB
	Coverage radius	20 km
Network	5G bandwidth	100 MHz
	Noise	−174 dBm/Hz
	Path-Loss Exponent	2.0–4.0

.

6.3.2. Task Load Change Experiment

Changes in system performance under different task numbers were tested, increasing the number of tasks from 50 to 500.

Table 10. System performance under different task scales.

Number of Tasks	Parameter Name	ILGWO	GWO	PSO	GA
100	Average latency (ms)	145.2	168.7	198.3	234.5
	Energy consumption ratio	0.732	0.821	0.889	0.923
	Success rate (%)	98.5	95.2	91.7	87.3
300	Average latency (ms)	287.4	345.6	412.8	489.2
	Energy consumption ratio	0.756	0.847	0.912	0.945
	Success rate (%)	96.8	92.1	86.4	79.8
500	Average latency (ms)	398.7	487.3	589.6	698.4
	Energy consumption ratio	0.779	0.863	0.934	0.967
	Success rate (%)	94.2	88.7	81.2	72.6

shows the system performance under different task scales.

The results indicate that as the number of tasks increases, the performance of all algorithms decreases, but the ILGWO algorithm consistently maintains optimal performance. In large-scale task scenarios, the advantages of ILGWO are especially pronounced. However, as the demand for tasks continues to rise, the system may face overload risks in certain extremely special scenarios. Therefore, we have conducted experiments on extreme scenarios and summarized their patterns as follows:

(1)

Performance degradation law

As the number of tasks increases, the performance of all algorithms shows a non-linear degradation trend. When the number of tasks exceeds 3000, the success rate of traditional algorithms drops to below 70%, while ILGWO can still maintain a success rate of over 80%.

(2)

System bottleneck identification

Under extreme loads, system bottlenecks manifest as a surge in transmission delay caused by communication bandwidth saturation, intensified competition for edge node computing resources, and exponential growth in task scheduling complexity.

(3)

Scalability strategy

In extreme scenarios, the system can increase its scalability by dynamically increasing the number of edge nodes, performing hierarchical task scheduling, and dynamically adjusting task priorities.

6.3.3. Network Condition Impact Experiment

The impact of different network conditions on system performance was tested, including changes in signal strength and interference levels.

Figure 5 shows the impact of network conditions on system performance. From the experimental results, it can be seen that when the signal strength decreases from −80 dBm to −110 dBm, the performance of all algorithms decreases, but the decrease in ILGWO is the smallest, showing better robustness. Secondly, in high-interference environments, ILGWO can maintain good performance by dynamically adjusting power allocation and offloading strategies. In addition, ILGWO can adaptively adjust optimization strategies according to changes in network conditions, reflecting the intelligence and adaptability of the algorithm.

6.3.4. Effect of Integrating Large Language Models

In order to evaluate the improvement effect of the integrated large language models on the intelligent decision-making capability of unmanned aerial vehicle command and control systems, we defined the decision quality indicators shown in

Table 11. Decision quality evaluation indicators.

Indicator Name	Definition	Calculation Formula	Evaluation Criteria	Test Scenario
Accuracy of task priority adjustment	In a dynamic environment, the system adjusts the proportion of task priorities correctly	Accuracy = number of correct adjustments/total adjustments × 100%	A gold-standard dataset based on expert annotation, containing optimal decisions from 500 typical scenarios	Priority ranking when multiple distress signals are received simultaneously in emergency rescue
Accuracy of dynamic resource allocation	The accuracy of the system in making optimal allocation decisions under resource constraints	Accuracy = number of optimal allocation schemes/total allocation schemes × 100%	Compared with the theoretical optimal solution obtained by the linear programming solver, a deviation of less than 5% is considered correct	Dynamic allocation of computing resources among 5 edge nodes, considering load balancing and minimizing latency
Accuracy of handling abnormal situations	The ability of the system to correctly identify and handle abnormal situations	Accuracy = number of correctly handled exceptions/total number of exceptions × 100%	Based on the historical fault case library, including 300 abnormal scenarios such as communication interruption, equipment failure, and environmental mutation	Autonomous decision-making and backup plan activation in case of communication interruption of unmanned aerial vehicles
Accuracy of collaborative path planning	Optimal path planning for multi-UAV collaborative operations	Accuracy = number of quasi-optimal paths/total number of paths × 100%	Compared with the optimal path length solved by the A* algorithm, a difference of less than 10% is considered accurate	Collaborative search-and-rescue path planning of 20 UAVs in urban environments

.

Based on this, we tested the improvement effect of the integrated LLM on the system’s intelligent decision-making capability, as shown in

Table 12. Comparison of decision quality before and after integration of LLM.

Decision Scenario	Traditional Rules	This Article’s Proposal	Increase Margin	Number of Test Samples	Standard Deviation	95% Confidence Interval
Task priority adjustment	76.3% ± 3.2%	94.7% ± 1.8%	24.1%	500 scenes	σ1 = 3.2%, σ2 = 1.8%	[70.0%, 82.6%] vs. [91.2%, 98.2%]
Dynamic allocation of resources	68.9% ± 4.1%	91.2% ± 2.3%	32.4%	200 allocation cases	σ1 = 4.1%, σ2 = 2.3%	[60.8%, 77.0%] vs. [86.7%, 95.7%]
Anomaly handling	71.5% ± 3.7%	93.8% ± 2.1%	31.2%	300 abnormal cases	σ1 = 3.7%, σ2 = 2.1%	[64.2%, 78.8%] vs. [89.7%, 97.9%]
Collaborative path planning	74.2% ± 2.9%	89.6% ± 2.4%	20.8%	100 path planning tasks	σ1 = 2.9%, σ2 = 2.4%	[68.5%, 79.9%] vs. [84.9%, 94.3%]
Average decision time (ms)	234 ± 28	156 ± 15	33.3%	1000 decision tests	σ1 = 28 ms, σ2 = 15 ms	[179, 289] vs. [127, 185]

.

From the experimental results, it can be seen that in emergency rescue scenarios, rule-based systems for task priority adjustment cannot take into account various changes in conditions, such as changes in the disaster situation, number of casualties, geographical location, and resources, resulting in secondary solutions. After the LLM is integrated into the system, the large language model can analyze dynamically changing external conditions based on prior knowledge and generate better priority sequences. In terms of dynamic resource allocation, most existing methods are based on load balancing, while the introduction of the LLM can combine multiple constraints, such as task requirements, remaining resources, and energy consumption limitations, to achieve efficient and accurate resource allocation. In terms of handling abnormal situations, the LLM relies on pre-trained knowledge representations to efficiently handle abnormal situations outside the distribution and has good generalization ability.

After integrating the large language model, the intelligent decision-making ability of the system significantly improved, and all decision quality indicators improved by more than 20%. Moreover, the method proposed in this article not only improves the average performance in relevant experimental scenarios but also reduces the standard deviation, indicating that stability also has certain advantages. And the confidence intervals do not overlap, indicating that this approach has significant differences from traditional approaches.

6.3.5. Comparative Experiment with Reinforcement Learning Methods

1.

Experimental setup

This section conducts comparative experiments with mainstream methods to provide specific differences between our method and reinforcement learning methods. We have selected DQN, PPO, SAC, QMIX, and LLM-7B (edge side). The experiment uses AirSim + Gym as the simulation platform, with a map of 10 km × 10 km and 10 UAVs. The obstacle density is set to 50 per square kilometer, and four subtasks are set: target search, collaborative reconnaissance, dynamic obstacle avoidance, and resource allocation. The climate conditions are set to clear, strong wind, and rain and fog, and the communication conditions are set to normal, partially interrupted, and strongly interfered with.

The hyperparameters for reinforcement learning are set as follows: the epoch number is 50,000, the number of steps per round is 1000, the learning rate is 3 × 10⁻⁴, the batch size is 256, the discount factor is 0.99, and the experience pool size is 1e6. Evaluation metrics include task success rate, convergence speed, average reward, sample efficiency, generalization performance, and computational cost.

2.

Task scenario design

Based on the characteristics of emergency rescue scenarios, we have designed four tasks, namely, dynamic target search, multi-aircraft collaborative reconnaissance, complex environment obstacle avoidance, and emergency resource scheduling, in order to approximate the actual scenario requirements as closely as possible.

3.

Analysis of experimental results

(1)

Task performance comparison

Figure 6 shows the performance comparison of different methods in four task scenarios. From the results, it can be intuitively concluded that the LLM method has certain advantages in specific scenarios or performs at average levels. The experimental results are shown in

Table 13. Comparison of experimental results in four scenarios.

Method	Scenario 1: Target Search	Scenario 2: Collaborative Reconnaissance	Scenario 3: Obstacle Avoidance Navigation	Scenario 4: Resource Scheduling
Key Metrics	Success Rate/Average time (s)	Coverage Rate/Repetition Rate	Success Rate/Collision Frequency	Optimization Degree/Latency (ms)
DQN	83.2%/245	91.3%/18.5%	79.8%/2.3	0.76/320
PPO	85.6%/232	93.2%/15.2%	82.4%/1.8	0.79/298
SAC	86.3%/228	94.1%/14.3%	84.1%/1.6	0.81/285
QMIX	84.5%/238	95.8%/11.2%	81.2%/2.0	0.78/305
LLM-7B	91.2%/198	96.7%/8.5%	89.3%/1.2	0.87/216

.

From the table, it can be seen that the lightweight large language model has significant advantages in target search and obstacle avoidance navigation scenarios, but it leads with a slight advantage in collaborative reconnaissance and resource scheduling. This is because we roughly selected the lightweight 7B LLM with edge-side quantization compression, and the computing and planning capabilities of the large language model in complex scenarios have not been fully realized.

(2)

Comparison of Learning Efficiency

As shown in Figure 7, the LLM method converges after only 2000 rounds in the same experimental setup. In zero-shot scenarios, the LLM can achieve a success rate of 82.4% without post-training. The specific results are shown in

Table 14. Convergence comparison experimental results.

Method	Convergence Rounds	Final Success Rate	Sample Complexity	Training Time (h)
DQN	35,000	86.3% ± 2.1%	3.5 × 107	168
PPO	28,000	88.7% ± 1.8%	2.8 × 107	142
SAC	25,000	89.5% ± 1.6%	2.5 × 107	135
QMIX	30,000	87.2% ± 2.0%	3.0 × 107	185
LLM-7B (Fine-tuning)	2000	91.2% ± 1.5%	2.0 × 105	24
LLM-7B (zero-shot)	0	82.4% ± 2.3%	0	0

.

(3)

Comparison of generalization ability

To accurately reflect the generalization advantage of the LLM method, we adjusted the terrain, weather, task, and scale to test the performance retention level of different methods in new scenarios. The experimental results are shown in

Table 15. Comparison of generalization experimental results.

Method	New Terrain	New Climate	New Task	Scale Adjustment	Average Retention Rate
DQN	62.3%	58.7%	45.2%	51.3%	54.4%
PPO	65.8%	61.2%	48.6%	54.7%	57.6%
SAC	67.2%	63.5%	51.3%	56.2%	59.6%
QMIX	64.5%	60.8%	47.9%	53.4%	56.7%
LLM-7B (fine-tuning)	85.6%	83.2%	79.8%	81.5%	82.5%
LLM-7B (zero-shot)	62.3%	58.7%	45.2%	51.3%	54.4%

.

From the experimental results, it can be seen that the large language model method has significant advantages in unfamiliar scenarios.

(4)

Real-time comparison

Considering that the average delay is more sensitive to extreme values, resulting in significant discrepancies between the actual experience and experimental values of the UAV system, we introduce P95 delay and P99 delay to better capture daily delay experiences and delays in extreme scenarios. Figure 8 is a visualization of comparative experiments using different methods. The left subgraph in Figure 8 reflects the central trend and dispersion of different methods through a block diagram, and points out outliers that exceed the IQR range by 1.5 times. The subgraph on the right of Figure 8 displays the density distribution of delay values using different methods through a violin plot, and marks the real-time threshold and ideal delay (green line), as well as the P95 and P99 percentages. Overall, it reflects that the large language model method has a concentrated distribution and advantages in real-time performance. The specific experimental results are shown in

Table 16. Real-time comparison experiment results.

Method	Average Latency	P95 Latency	P99 Latency	Standard Deviation
DQN	120 ± 15	145	168	15.2
PPO	135 ± 18	162	185	18.3
SAC	142 ± 20	175	198	20.1
QMIX	158 ± 22	190	215	22.4
LLM-7B	156 ± 12	172	186	12.3
LLM-3B	89 ± 8	98	105	8.1

.

4.

Key findings and discussion

From the experimental results, the LLM method has certain advantages in sample efficiency, generalization ability, and real-time performance. On the one hand, because the large language model learned a lot of prior knowledge during the pre-training stage, it can quickly adapt to the data distribution of new scenes in an unfamiliar environment. On the other hand, in future work, we will consider mixing the LLM with reinforcement learning methods, using the LLM to provide initial strategies to enhance early optimization of reinforcement learning and then improving the decision accuracy of the LLM in specific scenarios through online reinforcement learning. Theoretically, better results can be achieved.

6.4. Ablation Experiment

6.4.1. Analysis of Contribution of Algorithm Components

By gradually removing the key components of the ILGWO algorithm, the contribution of each component to algorithm performance can be analyzed. The average fitness in the table refers to the comprehensive average fitness value of system-level multi-objective optimization, which can be expressed as $f= {\lambda }_1 \times {\mathrm{\Psi }}_{delay}+{\lambda }_2 \times {\mathrm{\Psi }}_{energy} + {\lambda }_3 \times {\mathrm{\Psi }}_{reliability}$, where ${\mathrm{\Psi }}_{delay}$ represents the standardized delay index, ${\mathrm{\Psi }}_{energy}$ represents the standardized energy consumption index, ${\mathrm{\Psi }}_{reliability}$ represents the standardized reliability index, and ${\lambda }_1, {\lambda }_2, {\lambda }_3$ represent weight coefficients, the sum of which is 1. We have presented the specific results of the ablation experiment in

Table 17. Experimental results of algorithm component ablation.

Configuration	Average Fitness	Convergent Algebra	Success Rate	Degradation
Complete ILGWO	0.1247	89	96.8%	-
Remove Lévy flight	0.1432	134	89.3%	14.8%
Remove adaptive weights	0.1389	118	91.7%	11.4%
Remove elite learning	0.1356	107	93.2%	8.7%
Remove chaos initialization	0.1321	98	94.5%	5.9%

.

The results indicate that the Lévy flight strategy leads to the most significant improvement in algorithm performance, followed by adaptive weighting and elite learning mechanisms.

6.4.2. Parameter Sensitivity Analysis

Figure 9 illustrates the impact of key parameters on algorithm performance. Population size: When the population size is between 30 and 60, the algorithm performs better; being too small can affect diversity, while being too large can increase computational costs. Lévy flight parameter β: The optimal range is 1.3–1.7, and the algorithm performs best when β = 1.5. Elite ratio: The optimal range is 15–25%, and a high ratio can lead to insufficient diversity.

6.4.3. Impact of Large Language Model Configuration

The impact of different model configurations on system performance was tested. We have presented the test results of different LLMs in

Table 18. Performance comparison of different LLM configurations.

Model Configuration	Inference Delay (ms)	Decision Accuracy	Memory Usage (GB)	Power Consumption (W)
GPT-4 Full Version	1250	96.8%	350	450
7B compressed version	156	91.2%	27	85
3B Ultra Lightweight Edition	89	87.4%	12	35
Rule engine	23	68.9%	2	8

.

The results indicate that the 7B compressed version model achieves the best balance between performance and resource consumption.

6.5. System Robustness Experiment

6.5.1. Experimental Design

To comprehensively verify the robustness of the system, four types of fault scenario experiments were designed:

(1)

Node Failure Experiment

Fault type: Randomly select edge nodes for fault injection.

Failure rates: 5%, 10%, 15%, 20%, 25%.

Fault duration: 30 s, 60 s, 120 s, 300 s.

Evaluation indicators: Task completion rate, average latency, resource utilization rate.

(2)

Network Interruption Experiment

Interrupt types: Link interruption, bandwidth limitation, high latency.

Interruption intensity: Mild (10–30%), moderate (30–60%), severe (60–90%).

Scope of impact: Local interruption, regional interruption, and network-wide interruption.

Evaluation indicators: Communication success rate, data integrity, retransmission frequency.

(3)

Load impact test

Impact mode: Sudden load, sustained high load, periodic load.

Load intensity: 1.5×, 2×, 3×, 5× normal load.

Impact duration: 10 s, 30 s, 60 s, 180 s.

Evaluation indicators: Response time, throughput, resource overflow rate.

(4)

Environmental interference experiment

Interference sources: Electromagnetic interference, meteorological interference, obstacle obstruction.

Interference intensity: −5 dB, −10 dB, −15 dB, −20 dB, signal-to-noise ratio decrease.

Interference modes: Continuous interference, intermittent interference, random interference.

Evaluation indicators: Signal quality, bit error rate, connection stability.

6.5.2. Experimental Environment Configuration

Simulation parameter settings: Simulation time: 1800 s; number of UAVs: 30; edge nodes: 8; cloud center: 1; task load: 800 concurrent tasks; fault detection cycle: 5 s; recovery strategy: active redundancy + passive reconfiguration. We present the comparative results of task completion effectiveness under different failure rates in Table 19. In addition, we set up other differentiated experimental scenarios, and Table 20 shows the system performance under different network interruption intensities; Table 21 shows the changes in performance under different loads; Table 22 shows the communication quality under different interference strengths.

6.5.3. Experimental Results

(1)

Node fault tolerance test

Table 19. System performance under different failure rates.

Failure Rate	Algorithm	Task Completion Rate (%)	Average Latency (ms)	Resource Utilization Rate (%)	Recovery Time (s)
5%	ILGWO	97.3	178.4	89.2	12.3
	GWO	94.8	203.7	83.6	18.7
	PSO	91.2	245.3	78.1	25.4
	GA	87.6	287.9	72.8	32.1
10%	ILGWO	94.7	198.6	85.4	15.8
	GWO	89.3	238.2	78.9	24.3
	PSO	84.6	291.7	71.2	34.7
	GA	78.9	345.8	65.3	43.2
15%	ILGWO	91.8	223.5	81.7	19.4
	GWO	83.2	278.9	73.2	31.6
	PSO	76.4	342.1	64.8	45.9
	GA	68.7	412.3	56.9	58.7
20%	ILGWO	88.4	251.2	77.3	24.1
	GWO	76.8	324.6	66.8	39.8
	PSO	67.9	398.7	57.1	58.3
	GA	57.2	489.4	47.6	76.5

Table 19. System performance under different failure rates.

Failure Rate	Algorithm	Task Completion Rate (%)	Average Latency (ms)	Resource Utilization Rate (%)	Recovery Time (s)
5%	ILGWO	97.3	178.4	89.2	12.3
	GWO	94.8	203.7	83.6	18.7
	PSO	91.2	245.3	78.1	25.4
	GA	87.6	287.9	72.8	32.1
10%	ILGWO	94.7	198.6	85.4	15.8
	GWO	89.3	238.2	78.9	24.3
	PSO	84.6	291.7	71.2	34.7
	GA	78.9	345.8	65.3	43.2
15%	ILGWO	91.8	223.5	81.7	19.4
	GWO	83.2	278.9	73.2	31.6
	PSO	76.4	342.1	64.8	45.9
	GA	68.7	412.3	56.9	58.7
20%	ILGWO	88.4	251.2	77.3	24.1
	GWO	76.8	324.6	66.8	39.8
	PSO	67.9	398.7	57.1	58.3
	GA	57.2	489.4	47.6	76.5

(2)

Analysis of the Impact of Network Interruptions

Table 20. System performance under different network interruption intensities.

Interruption Strength	Communication Success Rate (%)	Data Integrity (%)	Average Number of Retransmissions	End-to-End Delay (ms)
Mild (10–30%)	94.6	98.7	1.23	189.4
Moderate (30–60%)	87.3	95.2	2.67	267.8
Severe (60–90%)	76.8	89.6	4.58	398.2
Network-wide interruption	45.2	67.3	8.94	756.3

Table 20. System performance under different network interruption intensities.

Interruption Strength	Communication Success Rate (%)	Data Integrity (%)	Average Number of Retransmissions	End-to-End Delay (ms)
Mild (10–30%)	94.6	98.7	1.23	189.4
Moderate (30–60%)	87.3	95.2	2.67	267.8
Severe (60–90%)	76.8	89.6	4.58	398.2
Network-wide interruption	45.2	67.3	8.94	756.3

(3)

Load impact adaptability test

Table 21. System response under sudden load.

Load Multiplier	Response Time (ms)	Throughput (ops/s)	Resource Overflow Rate (%)	Task Loss Rate (%)
1.5×	198.7	1847.3	2.1	0.8
2×	267.4	1623.8	5.7	2.3
3×	389.6	1289.4	12.4	6.8
5×	567.8	892.1	23.6	15.7

Table 21. System response under sudden load.

Load Multiplier	Response Time (ms)	Throughput (ops/s)	Resource Overflow Rate (%)	Task Loss Rate (%)
1.5×	198.7	1847.3	2.1	0.8
2×	267.4	1623.8	5.7	2.3
3×	389.6	1289.4	12.4	6.8
5×	567.8	892.1	23.6	15.7

(4)

Environmental interference resistance test

Table 22. Communication quality under different interference intensities.

Signal-to-Noise Ratio	Decrease Signal Quality (RSSI)	Error Rate	Connection Stability (%)	Switching Frequency (Times/min)
−5 dB	−67.3 dBm	1.2 × 10−4	96.8	0.7
−10 dB	−72.1 dBm	3.8 × 10−4	92.4	1.9
−15 dB	−78.6 dBm	9.6 × 10−4	85.7	4.2
−20 dB	−85.2 dBm	2.3 × 10−3	76.3	8.5

Table 22. Communication quality under different interference intensities.

Signal-to-Noise Ratio	Decrease Signal Quality (RSSI)	Error Rate	Connection Stability (%)	Switching Frequency (Times/min)
−5 dB	−67.3 dBm	1.2 × 10−4	96.8	0.7
−10 dB	−72.1 dBm	3.8 × 10−4	92.4	1.9
−15 dB	−78.6 dBm	9.6 × 10−4	85.7	4.2
−20 dB	−85.2 dBm	2.3 × 10−3	76.3	8.5

6.5.4. Cascade Fault Analysis

To verify the system’s ability to resist cascading faults, a cascading fault experiment was designed. Figure 10 shows the dynamic topology of cascading fault propagation. In addition,

Table 23. Analysis of cascade fault propagation.

Initial Number of Faulty Nodes	Final Number of Faulty Nodes	Level of Contact	Recovery Time (s)	Performance Retention Rate (%)
1	1.2	0.20	18.4	94.7
2	2.7	0.35	34.6	86.3
3	4.8	0.60	52.1	75.8
4	7.3	0.83	78.9	62.4

presents the results of cascading fault propagation analysis.

Figure 11 shows the performance evolution and recovery performance comparison during cascading faults, visually demonstrating the performance advantages of the proposed algorithm ILGWO in this paper.

6.5.5. Analysis of Experimental Results

(1)

Stability advantage analysis

The experimental results show that the ILGWO algorithm has significant advantages in system robustness: (1) Fault tolerance. At a node failure rate of 20%, ILGWO can still maintain a task completion rate of 88.4%, which is 11.6 percentage points higher than GWO. (2) Quick recovery. The average fault recovery time is reduced by 35–40% compared to traditional algorithms, reflecting the algorithm’s adaptive reconfiguration capability. (3) Slow performance degradation. Under moderate interference, the performance degradation amplitude is controlled within 15%, demonstrating good anti-interference ability. (4) Cascade control. The propagation of cascading faults is effectively controlled, and the number of cascading connections is kept at a low level.

(2)

Key technological contributions

(1) Multi-layer redundancy mechanism. The cloud–edge–terminal architecture provides multiple layers of backup, and a single point of failure will not cause system crashes. (2) Dynamic load balancing. The Lévy flight search strategy enables the algorithm to quickly find alternative resource allocation solutions. (3) Intelligent fault detection. Anomaly detection based on large language models can identify potential faults in advance. (4) Adaptive reconfiguration. The elite learning mechanism ensures rapid resource reallocation after a malfunction.

(3)

Practical application significance

The robustness experiment verified the reliability of the system in actual complex environments, mainly reflected in the following aspects. (1) Applicable to critical scenarios of the task. In critical tasks such as emergency rescue, the system can continue to operate in the event of partial equipment failure. (2) Strong adaptability to the environment. It can still maintain basic functions under conditions such as electromagnetic interference and severe weather. (3) Good scalability. As the system scale expands, the robustness indicators can still be maintained within an acceptable range. (4) Low maintenance cost. The self-healing ability reduces the need for manual intervention and lowers operational costs.

6.6. Discussion of Experimental Results

6.6.1. Analysis of Algorithmic Performance Advantages

The superior performance of the ILGWO algorithm is mainly attributed to the following aspects. The first is initialization optimization: chaotic-map initialization provides a better population distribution, offering the algorithm a favorable starting point. The second is enhancement of the search strategy: the Lévy flight mechanism strengthens the algorithm’s global search capability, enabling it to escape local optima. The third is adaptive mechanisms: dynamic weight adjustment allows the algorithm to achieve a better balance between exploration and exploitation. The fourth is elite preservation: the elite learning strategy ensures the retention and evolution of high-quality solutions.

6.6.2. Analysis of System Architecture Advantages

The “cloud–edge–terminal” architecture aligns with the existing network information system, and its advantages are reflected in the following. The first is hierarchical processing: tasks of different complexities are handled at appropriate layers, improving overall efficiency. The second is intelligent decision-making: the integration of large language models significantly enhances the system’s intelligence. The third is resource optimization: dynamic resource allocation strategies improve resource utilization efficiency. The fourth is strong robustness: multi-layer backup mechanisms ensure system reliability.

6.6.3. Practical Application Value

Experimental verification shows that the designed system possesses practical application value, as it introduces the intelligent capabilities of large language models into UAV command and control information systems. First, performance is significantly improved: key metrics exhibit gains of 20–60%. Second, adaptability is strong: the system can adapt to different application scenarios and environmental conditions. Third, scalability is excellent: it supports coordinated operations of large-scale UAV swarms. Fourth, deployment is convenient: based on standardized interfaces, it can be easily integrated with existing systems.

6.6.4. Scalability Analysis

The system designed in this article has three advantages in scalability: system layer, algorithm layer, and data layer, as follows:

(1)

Scalability advantage at the system level. This is mainly reflected in the following aspects: Elastic architecture: The three-layer architecture, consisting of the cloud, edge, and end, naturally supports elastic expansion, and each layer can be independently expanded without affecting the operation of the other layers. Standardized interfaces: Using standardized communication protocols and interface specifications ensures seamless integration of heterogeneous devices and systems. Service-oriented design and microservice architecture: Functional modules can be independently developed, deployed, and expanded, improving system flexibility.

(2)

Scalability advantage at the algorithmic level. This is mainly reflected in the following aspects: Parallel processing capability: The parallel nature of the ILGWO algorithm enables it to fully utilize multi-core and distributed computing resources. Parameter adaptation: The algorithm parameters can be dynamically adjusted according to the size of the problem and resource conditions, ensuring optimization effectiveness at different scales. Decomposition strategy: The decomposition and solution of large-scale problems are supported, reducing the complexity of individual optimization tasks.

(3)

Scalability advantage at the data flow level. This is mainly reflected in the following aspects: Hierarchical processing: Data are preprocessed and filtered at different levels to reduce the processing pressure on core nodes. Cache mechanism: The multi-level caching strategy improves data access efficiency and supports large-scale concurrent access. Flow control: Intelligent traffic scheduling and load balancing ensure the stable operation of the system under high loads.

6.6.5. System Model Validation and Deep Analysis

(1)

Communication model validation

To verify the accuracy of the communication model established in Section 4.2, we conducted actual testing and simulation comparison experiments.

Validation of air–ground channel model. The testing environment includes three typical environments: urban, suburban, and rural. The altitude of the UAV is 50 m–500 m, with an interval of 50 m. The horizontal distance is 1 km–10 km, with an interval of 1 km. The test indicators are path loss, channel gain, and LoS probability.

Table 24. Communication model validation results.

Environmental Type	Altitude (m)	Distance (km)	Measured Path Loss (dB)	Model Prediction (dB)	Relative Error (%)
City	100	2	92.3	90.8	1.6
City	200	5	108.7	107.2	1.4
Suburb	150	3	88.9	87.5	1.6
Countryside	100	4	85.2	84.1	1.3

shows the specific results of the communication model validation.

The verification results show that the average relative error of the established communication model is less than 2%, which meets the requirements of engineering applications.

(2)

Computational model validation

Edge computing delay model validation. By deploying computing tasks of different complexities on the actual MEC platform, the computing model established in Section 4.3.2 is verified.

Table 25. Calculation delay model validation.

Task Type	Calculation Complexity (G Cycles)	Actual Delay (ms)	Model Prediction (ms)	Relative Error (%)
City	1.5	145.6	142.3	2.3
City	5.2	487.9	475.2	2.6
Suburb	1.8	167.4	163.8	2.2
Countryside	0.8	89.7	87.1	2.9

shows the validation results of the delay model.

(3)

Energy consumption model validation

Flight energy consumption model validation. The flight energy consumption model established in Section 4.5.3 is verified using actual UAV flight data.

Table 26. Flight energy consumption model validation.

Flight Speed (m/s)	Flight Time (min)	Actual Energy Consumption (Wh)	Model Prediction (Wh)	Relative Error (%)
15	30	234.6	228.9	2.4
20	45	421.8	408.7	3.1
15	30	234.6	228.9	2.4
20	45	421.8	408.7	3.1

shows the validation results of the flight energy consumption model.

From the above system model experiments, it can be seen that the model can be further improved in the following directions in the future: the first improvement is to introduce machine learning methods to enhance the model’s adaptive ability; the second is to increase the dynamic adjustment of model parameters by environmental factors; the third is to establish a multi-level model validation system to improve the credibility of the model. These verification results indicate that the system model established in this paper has good accuracy and practicality, providing a reliable theoretical basis for the design and optimization of unmanned aerial vehicle command and control systems.

6.7. Actual Scenario Verification Experiment

In order to verify the performance of the system in real environments, we collaborated with Beihang University to deploy and test the system in simulated urban emergency rescue scenarios. The specific experimental environment deployment, test scenario design, and key performance indicator testing are provided in Appendix A.

In addition, we selected an emergency rescue drill in a practical scenario experiment to verify the performance of the system proposed in this paper in complex urban rescue. The scene of the experimental site includes factors such as building obstruction, multipath effects, and high-density interference, which have a strong degree of realistic simulation. Figure 12 shows a comparison of system performance in urban emergency rescue scenarios.

The experimental results show that among several methods, the method proposed in this paper has certain advantages in urban rescue scenarios. Capable of handling over 30 concurrent tasks simultaneously, with a response time of less than 500 ms, the system has strong parallel capability. Secondly, in an environment with building obstructions and electromagnetic interference, the system maintains communication quality through intelligent routing and power control. However, there are still some challenges in actual testing that highlight aspects that need to be optimized and improved in future work. The first goal is to increase the wind resistance and positioning accuracy of the UAV, the second is to expand the training data by including Chinese dialects and professional terminology, and the third is to optimize the system architecture to achieve larger-scale concurrency.

7. Summary

Aiming to solve the problem of the UAV command and control system’s inability to perform intelligent decision-making and resource allocation, this paper proposes a “cloud, edge, and end” command and control system based on a large language model. With 5G as the backbone network, the Internet of Things as a supplement, edge computing as the computing platform, and the large language model as the intelligent engine, this system enhances the system’s intelligent decision-making, environmental adaptation, and resource allocation abilities. The main contributions are as follows:

(1)

System architecture contribution. A three-layer distributed architecture empowered by large language models has been proposed, which achieves the optimal matching of intelligent decision-making capabilities and computing resources through the layered deployment of LLMs of different scales (cloud 175B, edge 70B, terminal 7B). The system supports elastic expansion from 10 to over 1000 UAVs, with good scalability.

(2)

Contribution to modeling theory. A multidimensional system efficiency optimization model covering communication, computing, and energy consumption has been established, which, for the first time, incorporates the inference delay and accuracy loss of the large language model into the system modeling, providing a theoretical basis for the quantitative analysis of unmanned aerial vehicle command and control systems.

(3)

Contribution of algorithmic technology. The ILGWO algorithm is proposed, which integrates Lévy flight, adaptive weighting, elite learning, and other strategies, and outperforms traditional algorithms in terms of convergence accuracy and speed. The benchmark function test shows that the algorithm’s convergence accuracy and convergence speed are improved by 35% and 60%, respectively.

(4)

Application verification contribution. The effectiveness of the system has been demonstrated through large-scale simulation experiments and practical scenario verification. Task latency, energy efficiency, and resource utilization are improved by 34.2%, 29.6%, and 31.8%, respectively. In actual emergency rescue exercises, the response time was shortened by 44.7%, and collaborative efficiency was improved by 39.5%.

In addition, the system proposed in this article has the following advantages:

(1)

Advantages of intelligent decision-making. By introducing the LLM, the accuracy of the system in task priority adjustment, dynamic resource allocation, exception handling, collaborative path planning, and other aspects exceeds 90%, which is 20–30% higher than that of traditional rule systems.

(2)

System robustness advantage. Despite a 20% node failure rate, it can still maintain a task completion rate of 88.4%, with the average fault recovery time reduced by 35–40% compared to traditional algorithms, demonstrating excellent fault tolerance.

(3)

Real-time advantage. The inference delay of the 7B model deployed on the edge side is only 156 ms, which can meet the real-time decision-making requirements of emergency scenarios.

(4)

Generalization ability advantage. In scenarios such as new terrain, new climate, new tasks, and scale adjustments, the performance retention rate reaches 82.5%, significantly better than the reinforcement learning method’s 56.7%.

However, the system still faces the following challenges in practical applications:

(1)

Calculation resource requirements. The deployment of large language models requires significant computing resources and storage space, which may pose challenges in environments with extremely limited resources.

(2)

Communication bandwidth dependence. Cloud–edge collaboration requires a stable, high-bandwidth communication link, and system performance may decrease in harsh communication conditions.

(3)

Standardization level. The compatibility between the system and existing equipment needs to be further verified, and industry-standardized interface specifications need to be established.

With the continuous maturation and improvement of related technologies, this system is expected to play an important role in more fields, providing strong support for social and economic development and the improvement of people’s lives.