The high-dynamic, chaotic nature of disaster scenes, such as sudden and intense urban fires, poses extreme challenges for traditional communication infrastructures. Often crippled by physical damage, channel congestion, and complex terrain, these conventional systems fail, severely hampering the efficiency and coordination of firefighting and rescue operations. Unmanned Aerial Vehicle (UAV) swarms, or ‘fire drone’ collectives, emerge as a transformative solution, offering unparalleled advantages in mobility, rapid deployment, and expansive coverage to establish resilient communication networks over disaster zones. However, the practical efficacy of these ‘fire drone’ swarms is fundamentally constrained by the core challenge of dynamic communication scheduling. This encompasses the volatile state of wireless links due to environmental interference and node mobility, the complex interplay and potential conflict in scheduling multiple nodes for diverse tasks, and the heterogeneous nature of missions ranging from reconnaissance to real-time video streaming. This article focuses on the critical problem of cooperative communication scheduling for UAV swarms within fire rescue scenarios. We design a mechanism that is task-load-aware and integrates multi-hop link reconstruction, establish a corresponding system architecture, and conduct rigorous performance validation to enhance the information assurance capability essential for effective disaster response.
To meet the stringent requirements of high dynamism, high interference, and low latency inherent to firefighting communications, it is crucial to construct a UAV swarm communication scheduling architecture endowed with both task-driving and link-sensing capabilities. As illustrated conceptually, our proposed architecture adopts a three-layer design philosophy of “distributed scheduling decision, multi-layer communication perception, and task-adaptive relay reconstruction,” comprising the Task-Driven Layer, the Link Scheduling Layer, and the Execution & Communication Layer. In the Task-Driven Layer, a command node decomposes overarching rescue missions—such as search, structure assessment, and hotspot monitoring—into granular subtasks like target identification, image/video backhaul, and path reconnaissance. These subtasks are then explicitly mapped onto specific communication resource requests, defining the data flow demands. The core Link Scheduling Layer introduces a Spatio-Temporal Joint Perception Model. This module continuously collects real-time data streams including node geolocation, communication parameters (e.g., transmit power), and channel state information (CSI). It fuses these metrics to quantify link availability and stability, accounting for both temporal fluctuations and spatial relational changes. A key output is a dynamic Link Stability Coefficient, where a lower value indicates a higher-priority, more robust link for scheduling purposes.
The effective operation of the overall architecture hinges on the implementation of several core key technologies. The first is the Time-Varying Link State Perception and Evaluation Model. Operating under the functional constraints of the system, this module must perform continuous measurement of link quality between ‘fire drone’ nodes, which is perpetually perturbed by environmental factors and the swarm’s own dynamic maneuvers. It provides quantifiable decision basis for scheduling. The system constructs a comprehensive Link Stability Coefficient, denoted as \( S_{ij}(t) \), over a sliding time window \( \Delta T \) using a weighted evaluation model that characterizes time-variant link usability. Its mathematical formulation is defined as:
$$ S_{ij}(t) = \alpha \times \frac{1}{SNR_{ij}(t)} + \beta \times PER_{ij}(t) + \gamma \times \frac{|d_{ij}(t) – d_{ij}(t-\Delta t)|}{\Delta t} $$
Here, \( S_{ij}(t) \) represents the stability metric for the link between node \( i \) and node \( j \) at time \( t \). \( SNR_{ij}(t) \) is the Signal-to-Noise Ratio, \( PER_{ij}(t) \) is the Packet Error Rate, and the term \( \frac{|d_{ij}(t) – d_{ij}(t-\Delta t)|}{\Delta t} \) captures the relative distance change rate between nodes, reflecting their spatial motion trend. The coefficients \( \alpha \), \( \beta \), and \( \gamma \) are weighting factors. This model synthesizes SNR, PER, and relative mobility into a single dynamic stability coefficient, effectively capturing the communication adaptability of the ‘fire drone’ swarm in harsh disaster environments. Channel gain is the dominant factor influencing SNR and PER. The system avoids scheduling links with critically low gain into relay paths to prevent the erosion of path diversity and transmission stability.
Building upon the real-time stability coefficient matrix provided by the perception model, the Adaptive Communication Scheduling Algorithm performs dynamic resource allocation under multi-dimensional constraints including task priority, node load, and link availability. The algorithm constructs a scheduling objective function \( F(t) \) to select the optimal set of links amidst fluctuating communication demands and changing node topology. The expression is:
$$ F(t) = \sum_{(i,j) \in E(t)} [\lambda_1 S_{ij}(t) – \lambda_2 L_{ij}(t) – \lambda_3 Q_j(t)] $$
Here, \( E(t) \) is the set of candidate links at the current scheduling epoch. \( S_{ij}(t) \) is the link stability coefficient, \( L_{ij}(t) \) represents the link delay, and \( Q_j(t) \) signifies the communication load on node \( j \). The parameters \( \lambda_1 \), \( \lambda_2 \), and \( \lambda_3 \) are scheduling weights satisfying the normalization constraint \( \lambda_1 + \lambda_2 + \lambda_3 = 1 \). At the start of each scheduling cycle, the scheduler builds a candidate link graph based on the system state matrix, computes the cost function value for each potential link, and selects the path set that minimizes \( F(t) \), thereby jointly ensuring link quality, load balance, and controlled latency. To counteract environmental perturbations, a gradient-descent-inspired strategy is employed to adaptively update the weighting parameters based on historical performance feedback, enhancing the model’s robustness and responsiveness in highly dynamic ‘fire drone’ operations.
In the volatile communication environment of a fireground, direct links between ‘fire drone’ nodes can frequently break due to obstacle occlusion, mobility-induced interference, or channel degradation. The Multi-Hop Communication Link Reconstruction and Relay Strategy is activated to ensure topological connectivity and mission continuity. Leveraging the stability coefficients \( S_{ij}(t) \) and the optimized outcomes from the scheduling function \( F(t) \), the system models the network as a weighted directed graph \( G(V, E) \), where vertices \( V \) represent swarm units and edges \( E \) represent available links. A Minimum-Cost Relay Path Search algorithm is employed, aiming to select the optimal path from the candidate set that minimizes the total relay communication cost. This cost for a path from source \( i \) to destination \( d \) is defined as:
$$ C(i, d, t) = \min_{P_{i \to d}} \sum_{(u,v) \in P} [\omega_1 \times \frac{1}{S_{uv}(t)} + \omega_2 \times L_{uv}(t) + \omega_3 \times Q_v(t)] $$
In this equation, \( P_{i \to d} \) denotes a candidate relay path from source \( i \) to target \( d \). The terms \( S_{uv}(t) \), \( L_{uv}(t) \), and \( Q_v(t) \) retain their earlier definitions for link stability, delay, and node load, respectively, with \( \omega_1 \), \( \omega_2 \), \( \omega_3 \) as the corresponding weighting coefficients for path computation. Upon detection of a link failure by the perception module, a path reconstruction is triggered. An enhanced Dijkstra’s algorithm performs a weighted traversal of the relay graph, dynamically updating the minimum-cost path from the faulty source to the intended destination. This mechanism ensures rapid path switching and network self-healing, significantly bolstering the disturbance resistance and scheduling stability of the ‘fire drone’ swarm during critical firefighting missions.
During high-intensity firefighting operations, a traditional centralized scheduling architecture struggles with the demands of real-time task offloading and dynamic adaptation of communication resources. Therefore, an Edge Task Dispatching and Collaboration Mechanism is integrated into the scheduling framework. Leveraging each node’s locally perceived link state and resource status (compute, memory), this mechanism enables rapid local decision-making and collaborative inference among multiple ‘fire drone’ nodes. An edge node evaluates its suitability for processing an incoming task based on a local offloading cost function:
$$ T_i(t) = \eta_1 \times \frac{R_i(t)}{C_i(t)} + \eta_2 \times D_i(t) + \eta_3 \times \frac{1}{B_i(t)} $$
Here, \( T_i(t) \) is the offloading cost for node \( i \) at time \( t \). \( R_i(t) \) is the size of the pending task, \( C_i(t) \) is the node’s available computational capacity, \( D_i(t) \) is the current experienced delay, and \( B_i(t) \) is the available link bandwidth. The coefficients \( \eta_1 \), \( \eta_2 \), and \( \eta_3 \) are weights for task dispatching. Within a scheduling cycle, nodes engage in a lightweight broadcast negotiation based on their locally computed cost. A distributed minimum-finding mechanism elects the optimal node for task offloading, effectively constructing a dynamic edge scheduling graph where edge weights reflect bandwidth and responsiveness. This facilitates the collaborative processing of heterogeneous tasks with minimal central intervention, yielding low communication overhead and high response speed. In the system architecture, this edge collaboration mechanism operates in concert with the multi-hop link reconstruction strategy, forming a closed-loop scheduling system that synergizes communication and computational resources, markedly improving the overall resource scheduling efficiency and task adaptability of the ‘fire drone’ swarm in complex rescue scenarios.
To validate the practicality and robustness of the proposed communication scheduling mechanism under high-dynamic fire rescue conditions, a comprehensive experimental platform was established. The platform is built upon the Robot Operating System (ROS) and the Gazebo simulation environment, with algorithmic cores implemented and results visualized using MATLAB. The hardware backbone consists of an Intel Xeon W-2265 processor with 64 GB RAM and an NVIDIA RTX A6000 GPU with 48 GB VRAM, running on Ubuntu 20.04 LTS to support containerized deployment and virtual link mapping. The simulated scenario is a 400 m × 400 m urban disaster zone with an obstacle density of 0.17, containing 10 fire source points. The ‘fire drone’ swarm size varies between 20 and 50 nodes. Simulated tasks include image backhaul, path reconnaissance, and fire source detection. Each scheduling cycle is set to 1.2 seconds. The pipeline involves Gazebo generating node states, the ROS-based perception stack estimating link quality and building the state matrix, which then drives the task offloading and relay modules. Scheduling decisions are transmitted via ROS messages to actuate UAV behaviors, with outcomes fed back into the next cycle, realizing a complete “perception-scheduling-execution-feedback” closed loop.
The performance of the scheduling mechanism was rigorously assessed within the simulated urban fireground. Comparative tests were conducted against traditional strategies, focusing on key performance indicators: average communication delay, task completion rate, link switching frequency, and data transmission success rate. Average communication delay is the mean time from task command generation to successful data delivery completion. Link switching frequency measures the number of triggered relay path reconstructions per scheduling cycle, indicating network topological stability. Tests were conducted with swarm sizes of 20 and 50 ‘fire drone’ nodes, under a dynamic task density of 30 commands per minute, over a simulation duration of 1200 seconds. The compared scheduling mechanisms include a traditional Round-Robin strategy, a Fixed-Priority strategy, and our proposed Adaptive Collaborative Scheduling mechanism. As summarized in Table 1, our mechanism demonstrates significant advantages across multiple metrics, particularly excelling in controlling communication delay and ensuring data integrity under conditions of high node density and link volatility.
| Metric Type | Round-Robin Mechanism | Fixed-Priority Mechanism | Adaptive Collaborative Mechanism | |||
|---|---|---|---|---|---|---|
| 20 Nodes | 50 Nodes | 20 Nodes | 50 Nodes | 20 Nodes | 50 Nodes | |
| Average Communication Delay (ms) | 78.4 | 145.2 | 63.7 | 113.5 | 45.3 | 68.7 |
| Task Completion Rate (%) | 86.1 | 79.5 | 91.4 | 84.6 | 96.2 | 91.3 |
| Link Switching Frequency (count) | 8.7 | 16.5 | 6.1 | 12.3 | 3.2 | 7.4 |
| Data Success Rate (%) | 91.5 | 85.3 | 94.2 | 89.8 | 98.6 | 96.9 |
For a more comprehensive evaluation, the Adaptive Collaborative Scheduling mechanism was benchmarked against the Fixed-Priority mechanism and a Link-Aware Heuristic mechanism. Tests were conducted under a unified environment of 50 nodes, an instruction density of 30 per minute, and a 1200-second simulation. Performance was compared across four dimensions: communication delay, path stability (link reconstruction frequency), task response efficiency, and packet loss rate. The results indicate that our mechanism maintains the lowest communication delay, stably under 70 ms. Its link reconstruction frequency is 8.1 per minute, which is 31% lower than the heuristic mechanism, demonstrating significantly enhanced topology preservation capability. The task response efficiency reaches 92.6%, outperforming the fixed-priority mechanism’s 86.4%. The packet loss rate is 1.8%, nearly 50% lower than the link-aware heuristic mechanism. In terms of resource cost, the spectral occupancy per task is reduced by approximately 23% compared to the heuristic mechanism, and the average CPU utilization remains below 41%, saving over 19% compared to the fixed-priority approach. These results confirm that the proposed scheduling mechanism exhibits superior link preservation capability and scheduling precision under conditions of high-frequency tasks and rapid node maneuvering, while also being more efficient in resource consumption, showcasing promising deployment potential and practical value for advanced ‘fire drone’ swarm operations.
In conclusion, addressing the critical communication needs of high-dynamic disaster scenarios, this research proposes and validates a UAV swarm communication scheduling mechanism equipped with link perception, adaptive scheduling, multi-hop reconstruction, and edge collaboration capabilities. Simulation experiments confirm its performance advantages in delay control, task completion, and network stability over conventional strategies. Future research should focus on integrating spatio-temporal prediction models to enhance the foresight of link variations, extending the framework to multi-swarm cross-domain scheduling systems, and exploring practical deployment strategies within real-world firefighting rescue contexts. Investigating the mechanism’s generalization capability across heterogeneous ‘fire drone’ platforms will be crucial for further enhancing the reliability and operational adaptability of next-generation emergency communication support systems.