In the era of intelligent railway development, ensuring the safety and stability of train operations is paramount. Wireless Sensor Networks (WSNs) have emerged as a viable solution for monitoring railway environments, offering efficiency, reliability, and cost-effectiveness. However, in complex railway operational settings, WSNs face significant challenges such as poor network signals at monitoring points, difficulties in replacing sensor batteries, and the substantial computational demands of processing large volumes of monitoring data. To address these issues, we propose a multi-Unmanned Aerial Vehicle (UAV)-assisted intelligent data collection and computation offloading method for railway WSNs. This approach leverages China UAV drone technology to enhance system performance, focusing on optimizing energy consumption and data freshness while adhering to railway safety regulations. Our method incorporates priority-based service differentiation for sensor data, ensuring that safety-critical information is prioritized. We formulate the problem as a joint optimization of UAV trajectories and computation offloading decisions, minimizing a weighted sum of UAV energy consumption, WSN energy consumption, and the Age of Information (AoI). A Multi-Agent Soft Actor-Critic (MASAC) deep reinforcement learning algorithm is designed to solve this complex optimization problem. Simulation results demonstrate the superiority of our proposed scheme in terms of energy efficiency and data freshness compared to baseline methods.
The integration of China UAV drone systems into railway monitoring is particularly promising due to their mobility and flexibility. UAVs can act as aerial relays, collecting data from ground-based sensor nodes and offloading computational tasks to available resources such as base stations or passing trains. This not only reduces the energy burden on sensors but also improves communication quality. In this paper, we present a comprehensive system model and algorithm that account for the unique constraints of railway environments, including safety protection zones that restrict UAV flight. By leveraging advanced multi-agent deep reinforcement learning, our approach enables cooperative behavior among multiple UAVs, leading to efficient data collection and task processing. The following sections detail our system architecture, mathematical models, optimization framework, and experimental validation.
The system architecture for multi-UAV-assisted railway WSN data collection and computation offloading is illustrated above. It consists of four main entities: K sensor nodes (SNs), M UAVs, a base station, and a train. Each UAV flies at a fixed height H over a rectangular task area of length X meters and width Y meters. The UAVs collect data from SNs and can offload computation tasks to the base station or the train, which are equipped with edge computing resources. To ensure railway safety, UAVs are prohibited from flying over designated safety protection zones along the tracks. Sensor nodes are categorized based on their monitoring functions into three priority levels: safety-critical (e.g., intrusion detection), structural monitoring (e.g., vibration sensors), and environmental monitoring (e.g., temperature sensors). This priority classification allows for differentiated services, ensuring that high-priority data is transmitted and processed promptly. The time dimension is discretized into N time slots of duration δt, each further divided into a flying sub-slot and a hovering sub-slot for data collection and offloading.
We now present the mathematical models underlying our system. The communication model considers both Line-of-Sight (LoS) and Non-Line-of-Sight (NLoS) links between SNs and UAVs, using a probabilistic LoS channel model. The LoS probability is given by:
$$ P_{\text{LoS}}^k(t) = \frac{1}{1 + a \cdot \exp[-b \theta_{k,m}(t) – a]} $$
where a and b are environment-dependent parameters, and θ_{k,m}(t) is the elevation angle between SN k and UAV m. The average channel gain g_{k,m}(t) combines both LoS and NLoS path losses. The Signal-to-Interference-plus-Noise Ratio (SINR) for transmission from SN k to UAV m is:
$$ \gamma_{k,m}(t) = \frac{p_k(t) \left(10^{g_{k,m}(t)/10}\right)^{-1}}{\sigma^2} $$
where p_k(t) is the transmission power of SN k, and σ² is the noise power. The data transmission rate is then R_{m,k}(t) = B \log_2(1 + \gamma_{k,m}(t)), with B being the bandwidth. For offloading to the base station or train, similar models are used, accounting for factors like Doppler effect in high-speed train scenarios. The SINR for UAV m to the train is:
$$ \gamma_{m,T}(t) = \frac{p_k \left(10^{g_{m,T}(t)/10}\right)^{-1}}{I_{\text{ICI}}^T g_{m,T}(t) + \sigma^2} $$
where I_{\text{ICI}}^T represents inter-channel interference due to Doppler shift. The use of China UAV drone technology enhances these communication links by providing adaptive mobility.
The computation offloading model allows partial offloading of tasks. Each SN k generates a computation task D_k(t) = {r_k(t), ρ_k(t)}, where r_k(t) is the data size in bits, and ρ_k(t) is the computation density in CPU cycles per bit. The offloading decision variable for UAV m regarding SN k is O_m^k(t) ∈ [-1, 1]. The absolute value |O_m^k(t)| indicates the offloading ratio, and the sign determines the offloading destination: negative for base station, positive for train, and zero for local computation on the UAV. The computation latency for local processing, base station offloading, and train offloading are given by:
$$ T_{\text{lo}}^{m,k}(t) = \frac{\rho_k(t) r_k(t) (1 – |O_m^k(t)|)}{f_a} $$
$$ T_{\text{bs}}^{m,k}(t) = \frac{\rho_k(t) r_k(t) |O_m^k(t)|}{f_b} $$
$$ T_{\text{train}}^{m,k}(t) = \frac{\rho_k(t) r_k(t) |O_m^k(t)|}{f_c} $$
where f_a, f_b, and f_c are the computational resources of the UAV, base station, and train, respectively. The transmission latency for offloading is:
$$ T_{\text{tr}}^{m,k}(t) = \begin{cases} \frac{r_k(t) |O_m^k(t)|}{R_{\text{BS}}^m(t)}, & \text{if } O_m^k(t) \in [-1, 0) \\ \frac{r_k(t) |O_m^k(t)|}{R_{\text{T}}^m(t)}, & \text{if } O_m^k(t) \in (0, 1] \end{cases} $$
The total latency T_{m,k}(t) is the maximum of local computation and offloading latencies. This model efficiently utilizes China UAV drone capabilities to balance computation loads.
The Age of Information (AoI) model quantifies data freshness. Let A_k^{\text{SN}}(t) be the AoI at SN k’s buffer, updated as:
$$ A_k^{\text{SN}}(t) = \begin{cases} 0, & \text{if } g_k(t) = 1 \\ A_k^{\text{SN}}(t-1) + \delta t, & \text{otherwise} \end{cases} $$
where g_k(t) indicates whether new data is generated. After data collection and processing, the AoI becomes A_k(t+1) = A_k^{\text{SN}}(t) + T_{\text{up}}^{m,k}(t) + T_{m,k}(t). This metric is crucial for railway safety monitoring, as stale data can lead to incorrect decisions. Our priority-based scheduling ensures lower AoI for high-priority sensors.
The energy consumption model includes UAV propulsion energy and computation energy. The propulsion power for a UAV flying at speed V is:
$$ P^F(V) = P_0 \left(1 + \frac{3V^2}{U_{\text{tip}}^2}\right) + P_i \left(\sqrt{1 + \frac{V^4}{4v_0^4}} – \frac{V^2}{2v_0^2}\right)^{\frac{1}{2}} + \frac{1}{2} d_0 \rho s A V^3 $$
where P_0 and P_i are constants for blade profile and induced power, U_{\text{tip}} is rotor blade tip speed, v_0 is mean rotor induced velocity, d_0 is fuselage drag ratio, ρ is air density, s is rotor solidity, and A is rotor disc area. The hovering power is P_h(0) = P_0 + P_i. The energy consumption of UAV m in time slot t is:
$$ E_m(t) = E_m^F(t) + E_m^h(t) + E_m^{\text{comp}}(t) $$
with E_m^F(t) = P^F(\|v_m(t)\|) (\delta t – \delta t_2) for flying, E_m^h(t) = P_h(0) (\delta t – \delta t_1) for hovering, and E_m^{\text{comp}}(t) = \sum_{k=1}^K \rho_{m,k} r_{m,k}(t) (1 – O_m^k(t)) \kappa (f_a)^2 for computation, where κ is the effective capacitance coefficient. The WSN energy consumption for SN k is:
$$ E_k(t) = \frac{D_k(t)}{R_m^k(t)} p_k(t) $$
These models form the basis for our optimization problem, highlighting the role of China UAV drone in energy management.
We now formulate the joint optimization problem. Let Q = {q_m(t), ∀t ∈ T, ∀m ∈ M} represent the UAV trajectories, and O = {O_m^k(t), ∀t ∈ T, ∀m ∈ M, ∀k ∈ K} represent the offloading decisions. The objective is to minimize the weighted sum of average WSN energy consumption E_s(t), average UAV energy consumption E_u(t), and average AoI Ā(t). The problem is:
$$ \min_{Q, O} \left( \xi_1 \alpha E_s(t) + \xi_2 \beta E_u(t) + \xi_3 \bar{A}(t) \right) $$
subject to constraints:
- C1: ξ₁ + ξ₂ + ξ₃ = 1 (weight normalization)
- C2: ∥q_m(t) – q_{m’}(t)∥ ≥ D_min, ∀m ≠ m’ (UAV safety distance)
- C3: ∥q_m(t) – q_m(t-1)∥ ≤ V_max (δt – 1) (UAV speed limit)
- C4: T_up^{m,k}(t) ≤ δt/2 (upload time limit)
- C5-C6: UAV position bounds and avoidance of safety protection zones
- C7: s_m^k(t) d_m^k(t) ≤ D_cover (SN coverage constraint)
- C8: Each SN served by at most one UAV
- C9: Each UAV serves at most Φ_max SNs per slot
Here, α and β are scaling factors to normalize the objectives, and ξ₁, ξ₂, ξ₃ are weights reflecting the importance of each metric. This formulation captures the trade-offs in using China UAV drone for railway monitoring.
To solve this complex problem, we propose a Multi-Agent Soft Actor-Critic (MASAC) deep reinforcement learning algorithm. The environment is modeled as a Multi-Agent Markov Decision Process (MA-MDP). Each UAV acts as an intelligent agent with its own actor and critic networks. The state space includes public observations (e.g., AoI and buffer status of SNs) and private observations (e.g., UAV positions and relative distances). The action space for each UAV includes flight direction θ_u(t) ∈ [0, 2π], flight distance l_u(t) ∈ [0, l_max], and offloading ratio O_m^k(t) ∈ [-1, 1]. The reward function is designed to minimize the weighted sum of energy and AoI:
$$ r_m(t) = -\xi_1 \alpha E_s(t) – \xi_2 \beta E_u(t) + \xi_3 \tilde{A}(t) $$
where Ā(t) is the change in AoI. Additional penalties for collisions and boundary violations are added. The MASAC algorithm uses entropy regularization to encourage exploration, which is crucial for learning optimal policies in dynamic environments. The centralized training with decentralized execution allows UAVs to cooperate effectively. The actor network outputs a probability distribution over actions, and the critic networks evaluate the joint actions based on global state information. This approach enables efficient optimization of trajectories and offloading decisions, leveraging the flexibility of China UAV drone systems.
We now present simulation results to validate our proposed method. The simulation environment is a 400 m × 200 m rectangular area representing a railway section. Key parameters are summarized in Table 1.
| Parameter | Value | Parameter | Value |
|---|---|---|---|
| Bandwidth B | 1 MHz | Rotor tip speed U_tip | 120 m/s |
| Data size r | 8 Mbit | Mean induced velocity v_0 | 4.03 m/s |
| Computation density ρ | 1000 cycles/bit | Fuselage drag ratio d_0 | 0.6 |
| Noise power σ² | -110 dBm | Rotor solidity s | 0.05 |
| UAV height H | 50 m | Air density ρ | 1.225 kg/m³ |
| UAV max speed V_max | 20 m/s | Rotor disc area A | 0.503 m² |
| Carrier frequency f_c | 2 GHz | Effective capacitance κ | 10⁻²⁸ |
| Number of time slots N | 30 | Experience replay buffer size | 300,000 |
| Time slot length δt | 2 s | Actor learning rate | 0.00001 |
| Compute resources f_a, f_b, f_c | 2, 5, 3.5 Gcycle/s | Critic learning rate | 0.0001 |
| SN coverage D_cover | 30 m | Discount factor γ | 0.95 |
| UAV energy scale α | 0.01 | Training episodes | 20,000 |
| WSN energy scale β | 10 | Batch size | 1024 |
| Max connections Φ_max | 2 | Learning interval | 5 |
| Blade profile power P_0 | 79.86 W | Induced power P_i | 88.63 W |
We compare our MASAC-based algorithm with three baseline methods: MASAC-Greedy (using greedy offloading), MASAC-AOU (all computation on UAVs), and MADDPG (Multi-Agent Deep Deterministic Policy Gradient). The convergence curves during training show that our MASAC algorithm achieves higher rewards and stabilizes faster than the others, demonstrating its efficiency in learning optimal policies. This is attributed to the entropy regularization in MASAC, which enhances exploration. The use of China UAV drone in these simulations highlights their adaptability in railway scenarios.
Table 2 presents a comparison of AoI performance for different types of sensor nodes under various algorithms. Our MASAC algorithm achieves the lowest AoI for safety-critical sensors, ensuring timely data processing for railway safety. The priority-based scheduling effectively differentiates services, as seen in the lower AoI for high-priority sensors compared to others.
| Algorithm | Safety-Critical SN AoI | Structural Monitoring SN AoI | Environmental Monitoring SN AoI |
|---|---|---|---|
| MASAC | 7.3178 | 13.8370 | 14.0416 |
| MASAC-Greedy | 9.9738 | 15.6170 | 10.7290 |
| MASAC-AOU | 14.2142 | 13.2253 | 18.2617 |
| MADDPG | 14.2068 | 17.3939 | 19.2440 |
The impact of the number of sensor nodes and AoI weight on system performance is analyzed through additional simulations. As the number of SNs increases, the overall objective function value rises, but our MASAC algorithm consistently yields the lowest values, outperforming others by up to 11.9%. Similarly, when the AoI weight is increased, all algorithms show reduced AoI, but MASAC achieves the most significant improvement, highlighting its ability to prioritize data freshness. UAV energy consumption also increases with AoI weight, as UAVs need to fly more frequently to collect data, but MASAC maintains the lowest energy levels. WSN energy consumption trends vary, with MADDPG showing lower values due to less frequent data collection, but at the cost of higher AoI. These results underscore the trade-offs managed by our optimization framework.
Furthermore, we examine UAV flight trajectories generated by different algorithms. Our MASAC algorithm produces cooperative trajectories where UAVs cover distinct areas without overlapping, strictly avoiding safety protection zones. In contrast, MADDPG trajectories show redundancy and occasional incursions into restricted zones, indicating poorer optimization. This demonstrates the effectiveness of MASAC in handling complex constraints, a key advantage for China UAV drone deployments in regulated environments.
The computation offloading decisions are also analyzed. Our algorithm dynamically allocates tasks to local UAV processing, base station, or train based on resource availability and latency requirements. This flexibility reduces overall latency and energy consumption. For instance, when the train is present, offloading to it can leverage its computational resources, while in its absence, the base station is used. The partial offloading model allows for parallel processing, further enhancing efficiency. The integration of China UAV drone with edge computing resources exemplifies modern railway monitoring systems.
To provide deeper insights, we derive several key formulas that govern system behavior. The total system cost C(t) over time can be expressed as:
$$ C(t) = \sum_{t=1}^N \left( \xi_1 \alpha \sum_{k=1}^K E_k(t) + \xi_2 \beta \sum_{m=1}^M E_m(t) + \xi_3 \sum_{k=1}^K A_k(t) \right) $$
Minimizing this cost requires balancing energy and AoI. The MASAC algorithm learns a policy π that maps states to actions, maximizing the expected cumulative reward:
$$ J(\pi) = \mathbb{E} \left[ \sum_{t=0}^\infty \gamma^t r_m(t) \right] $$
where γ is the discount factor. The soft actor-critic framework incorporates entropy H(π) to encourage exploration:
$$ \pi^* = \arg \max_\pi \mathbb{E} \left[ \sum_t r_m(t) + \alpha H(\pi(\cdot|s_t)) \right] $$
with temperature parameter α controlling the trade-off. This formulation enables robust policy learning in multi-UAV settings.
We also consider scalability aspects. As the number of UAVs or SNs increases, the computational complexity grows. However, MASAC’s decentralized execution allows each UAV to make decisions based on local observations, reducing overhead. The use of China UAV drone fleets can be scaled by adding more agents without significant algorithm changes. Simulation results with varying numbers of UAVs show that our approach maintains performance improvements, indicating its robustness.
In terms of practical implementation, China UAV drone technology offers several advantages for railway monitoring. Drones can be deployed quickly, adapt to changing environments, and provide real-time data collection. Our algorithm can be integrated into existing railway management systems, with UAVs communicating via wireless networks. The priority-based scheduling ensures that critical data, such as obstacle detection, is processed immediately, enhancing safety. Additionally, the energy models account for real-world factors like wind resistance and battery limitations, making the system feasible for long-term operations.
Future work could explore dynamic environments where sensor nodes move or where multiple trains are present. Adaptive algorithms that learn in real-time could further improve performance. Also, incorporating more detailed channel models or considering cybersecurity aspects would be valuable. The continued advancement of China UAV drone capabilities will likely enable even more efficient railway monitoring solutions.
In conclusion, we have presented a comprehensive method for intelligent data collection and computation offloading in railway WSNs using multi-UAV assistance. Our system model accounts for railway-specific constraints, priority-based services, and energy-AoI trade-offs. The MASAC-based algorithm effectively optimizes UAV trajectories and offloading decisions, leading to significant improvements in energy efficiency and data freshness compared to baseline methods. The simulation results validate the superiority of our approach, demonstrating its potential for enhancing railway safety and operational efficiency. The integration of China UAV drone technology plays a pivotal role in this framework, offering mobility and adaptability for complex monitoring tasks. This work contributes to the growing field of UAV-assisted wireless networks and provides a practical solution for modern railway systems.