IoT applications—smart cities, healthcare monitoring, industrial automation [8–10]—impose stringent requirements on WSNs: intensive computation, continuous sensing, and multi-task operations under 6G constraints (limited memory, bandwidth [11,12]). MEC mitigates cloud latency by enabling edge processing [13], yet battery-constrained sensor nodes face critical power management challenges.

Key limitations of traditional approaches include:

Static power allocation ignoring dynamic topologies and node heterogeneity.

We propose a hierarchical multi-agent power management framework addressing these gaps: a.

Trust-driven redistribution: Adaptive power allocation via MA→CoA→TA hierarchy with real-time trust metrics (Eq. (2)).

Table 1 summarizes notation. The paper is organized as follows: Section 2 positions our novelty compared to existing works, Section 3 describes the proposed model and system assumptions. Section 4 describes proposed algorithms and related analysis. Section 5 is to demonstrate the simulation results and finally the paper is concluded with discussion of results, supporting our proposed method in Section 6, discussing the results of our proposed method.

Traditional clustering (LEACH [14], HEED, EECS [15]) elects cluster heads probabilistically but suffers CH election overhead and fails in non-uniform deployments. Enhanced variants like firework-optimized LEACH [16,17] and gradient routing [18] improve load balancing but retain clustering rigidity.

Multi-agent systems offer distributed alternatives. NPC [19,20] uses threshold-based power reduction across agents but lacks hierarchy and trust. Three-layer clustering [21] combines centralized/distributed selection but requires base station coordination.

Trust-aware routing (TARM [22]) computes direct/indirect trust via edge nodes but uses static allocation. Recent MARL approaches [23,24] enable decentralized control through Q-learning, achieving power efficiency but requiring extensive training without trust-power coupling.

MEC frameworks address offloading [25–27] and ESN design [26], optimizing latency/cost. DDPG-based allocation [16] handles continuous actions but lacks WSN-specific trust. Federated learning [28,29] distributes training but assumes stable topologies, unsuitable for dynamic WSNs.

NN-ILEACH [30] uses neural networks for CH selection (11,361 rounds lifetime) but centralizes training and ignores trust.

Table 2 compares our approach across seven criteria. Unlike clustering methods (overhead), single-agent trust (static), or MARL (training-heavy), we provide:

1.

Hierarchical non-clustered agents: MA→CoA→TA avoids election overhead.

It is evident that there are many works that exist where different approaches and methods were taken into consideration to attain energy-efficient sensor networks, and also some of the literature tried to enhance energy efficiency in edge environments. But in most cases, it lacks a technique that can be considered a systematic approach for energy-efficient smooth communication over the edge in the IoT-based WSN paradigm for V-IoT in beyond 5G network.

A resource-efficient system relies on multiple factors, and towards energy-efficient communication through a resource-constrained IoT network, the first step is establishing a network architecture with proper deployment of nodes, considering system requirements and infrastructure. The network architecture comprises multiple nodes, and in the IoT and IIoT scenario, it is possible to have multiple tasks over a network simultaneously, and it is evident that those can be different applications. Therefore, there are two possibilities to consider: a) application-specific operation by selecting heterogeneous nodes (where source and sink nodes are predefined) and b) selecting multiple paths for operation, where similar types of applications can occur at the same time within a homogeneous network. Therefore, in our work, we have considered the second option. In the next part of this subsection, detail description of the hierarchical agent-based WSN architecture is presented in Fig. 1.

In the proposed scheme the power management depends mainly on trust-based node selection and power distribution. Fig. 2 summarizes the proposed agent-based power management cycle.

The dynamic power management module implements the power allocation model of Section 3.2.2 as per the Dynamic Power Distribution Algorithm 3. When a Task Agent (TA) requests power for a task, the Master Agent (MA) first evaluates the energy condition of the MA, CoA, and TA through likelihood ratios, then adjusts the requested power to obtain Padj. If the system can satisfy this adjusted demand, power is allocated along the path; otherwise, a redistribution procedure attempts to gather additional energy from high-power nodes Power Redistribution Algorithm 4. If redistribution fails, the TA is placed in power-saving mode.

The Power Redistribution algorithm presents a systematic approach to reallocating power resources within a wireless sensor network when additional power (Pneed) is required. This algorithm implements a two-phase strategy to meet power demands while maintaining network stability and operational efficiency.

In the first phase, the algorithm searches for donor nodes whose power exceeds 1.5Pth, while excluding Master Agents (MAs) so that the hierarchy remains intact. These candidates are sorted in descending order of their power Pi, and for each node Nj the algorithm first checks that its associated MA still has at least Pth before any transfer is made. The amount of power moved is then chosen as: (4)Ptransfer=min(Pj−Pth,Pneed)which prevents the donor from dropping below the threshold and reduces the remaining demand as much as possible. If this first phase does not fully satisfy the requirement (Pneed>0), a second phase considers ordinary Task Agents (TAs). In this step, each TA has its power reduced by 50%, and the recovered energy is used to further decrease Pneed via Pneed←Pneed−0.5⋅Tk.power. The procedure is considered successful when the remaining demand satisfies Pneed≤0. In this way, the algorithm meets urgent power requests while still protecting critical roles and sustaining the network over the long term.

The trust update mechanism implements the behavior described in Section 3.2.1. Active nodes gain trust when they use power efficiently and maintain good-quality links, whereas inactive nodes gradually lose trust over time. After each update, the new trust value is clipped to lie within [0, 1]. The complete procedure is summarized in Algorithm 5.

The Fault Detection and Recovery Algorithm 6 provides a structured way to monitor node health and react to failures. Each node N sends heartbeat messages at regular intervals Δt, and the algorithm combines a timeout check with a trust-based check. Two thresholds are defined: a minimum trust value τmin=0.4 and a maximum allowable gap Tmax between heartbeats. If a node does not send a heartbeat within Tmax, it is marked as faulty and its tasks are reassigned using the power reallocation procedure in Algorithm 4. If a node continues to send heartbeats but its trust τi falls below τmin, its role is downgraded to TA; if it previously served as a CoA, a new CoA is selected according to Algorithm 2. Nodes that maintain regular heartbeats and τi≥τmin keep their current roles and power allocations. This hierarchical strategy allows the network to detect problems early and recover gracefully while keeping overall performance stable.

Heartbeat Monitoring: Node i is faulty if: t−tlast>Tmax or τi<τmin(=0.4)

Sensitivity to Intermittent Link Failures (Tmax, τmin)

To assess robustness under intermittent links, we conducted a sensitivity study by varying the heartbeat timeout Tmax∈{2,5,10} and the trust threshold τmin∈{0.3,0.4,0.6} under a heartbeat failure probability of 0.01. Table 3 reports the resulting Lifetime@0.9N and final average trust. Across all tested settings, Lifetime@0.9N remains 500 rounds, while final trust varies only mildly (0.7636–0.7815), indicating that the proposed trust/fault handling is stable and does not require strict parameter tuning.

The Energy-Efficient Data Routing Alogorithm 7 computes routes using both distance and trust information. Given a packet P, a source node S, a destination D, and a hop limit hmax=5, it builds a route R that respects energy and reliability constraints. The route is initialized with the source, and while the destination has not been reached and h<hmax, the algorithm looks for candidate next hops {Ci} within the transmission radius r of the current node R[−1], keeping only nodes with τCi>0.5. For each candidate, the cost is computed as (5)Cost(Ci)=w1⋅d(Ci,D)+w2⋅(1−τCi)where d(Ci,D) is the Euclidean distance to D and w1,w2 control the trade-off between path length and reliability. The node with the smallest cost is added to R. If no suitable candidate exists at some step, routing fails. This approach produces paths that are both short and trustworthy, reducing energy use while keeping the network reliable.

Routing Weight Sensitivity

Table 4 analyzes w1∈{0.2,0.5,0.8} impact (w2=1−w1, 50 routes, N=1000). Trust emphasis (w1=0.2, w2=0.8) yields optimal balanced performance: minimum hops (3.42), perfect path trust (1.0), with delivery (42%).

Fig. 3 visualizes the trade-off. This approach produces short, trustworthy paths reducing energy while maintaining reliability.

To avoid overusing particular nodes and to extend network lifetime, the framework periodically rotates roles Algorithm 8 and enables power-saving behavior for low-power TAs. In role rotation, well-powered, highly trusted TAs are promoted to MA positions, while exhausted nodes are relieved from leadership. The Dynamic Role Rotation algorithm formalizes this idea. At every interval trotate, it evaluates each TA Ti using (6)Fitness(Ti)=ατTi+βPTiPth,where τTi is the node’s trust and PTi its power level. The TA with the highest fitness, Tbest, is selected, the current MA S is demoted to TA, and Tbest becomes the new MA. A new CoA for this MA is then chosen using Algorithm 2. This periodic rotation spreads the leadership burden, avoids single points of failure, and keeps the network balanced. Also to validate stability, we tested trotate∈{30,60,120} and report sensitivity in Table 5. The results show <3.5% energy oscillation variation and zero routing churn across this range, confirming robustness. We use trotate=60 as the baseline setting.

The Power-Saving Mode Activation Algorithm 9 provides an adaptive strategy for controlling TA activity based on energy levels. It uses two thresholds: a low-power threshold Plow=0.3Pth and the normal operating threshold Pth. When a TA’s power PT falls below Plow, the node reduces its sensing rate (for example, by half), disables non-essential communications, and may request a power boost through Algorithm 3. When PT≥Pth, the node returns to normal operation. For intermediate levels Plow≤PT<Pth, the node remains in its current reduced state. This three-level behavior saves energy without abruptly disconnecting nodes that still have usable battery.

The Network-Wide Optimization Algorithm 10 coordinates resource management and role tuning across all N nodes, {Ni}i=1N, at regular intervals topt. At each optimization step, it first identifies TAs with power above 1.2Pth and redistributes their excess energy using Algorithm 4. It then refreshes all trust values using Algorithm 5 and, based on these updated metrics, decides whether to perform role rotations via Algorithm 8. Finally, it recomputes routing paths using Algorithm 7 so that routes remain consistent with the new power and trust landscape. This periodic cycle keeps resource usage efficient and helps maintain good performance as the network and its traffic conditions evolve.

The network lifetime of the network is shown with the number of alive nodes with respect to simulation rounds in Fig. 4. It can be seen the proposed method, along with EE-LEACH and FEDLEARN maintain all 1000 nodes alive for 500 rounds, demonstrating effective energy management and robustness. In contrast, PEGASIS breaks down much earlier: by around 130 rounds only 8 nodes are still alive, which shows how vulnerable a pure chain-based routing scheme is when power is limited. The Federated Learning baseline keeps all nodes alive for longer, but it does so at the cost of higher overall energy consumption.

The stability of the proposed scheme comes from how its main mechanisms work together. The hybrid CoA selection (Algorithm 2) places coordinators in energy-balanced locations so that no small region is overused, while the dynamic power redistribution (Algorithm 4) recovers and reallocates excess energy when needed. On top of this, the power-saving mode (Algorithm 9) prevents sudden failures of nodes by automatically throttling low-energy nodes, eventually help the network to maintain service for a longer time.

Th next Fig. 5 shows how the average trust value evolves over time. The proposed method keeps trust high and stable at around 0.78, which indicates reliable communication and a robust update process in Algorithm 5 that blends past behavior, power efficiency, and link quality. PEGASIS, on the other hand, cannot sustain trust because many nodes die early, so the average trust eventually drops to zero. EE-LEACH reaches the highest trust level (about 0.98), reflecting the benefits of clustering, but likely at the cost of higher energy use. The Federated Learning approach achieves a middle ground, maintaining trust at roughly 0.90 while still consuming more energy than the proposed scheme.

Therefore, by demoting low-trust nodes and reassigning their roles, the trust mechanism limits the influence of unreliable nodes and strengthens overall network reliability, which is crucial for consistent edge-computing performance.

It can be seen in Fig. 6 that the energy consumption the proposed method preserves battery power more effectively with balanced energy consumption compare to PEGASIS. In PEGASIS long chain transmissions drain nodes quickly, whereas EE-LEACH also performs well due to its clustering structure. But compare to all of these methods the Federated Learning approach introduces extra coordination overhead and therefore uses more energy overall.

In the proposed method combination of efficient power allocation, timely activation of power-saving mode, and dynamic redistribution of excess energy, extends node lifetime and cuts unnecessary energy use, which is an essential property for IoT edge devices that operate with very limited power budgets.

Fig. 7 shows the system throughput over time. PEGASIS reaches the highest peak throughput (about 18.94 Mbits) because it exploits long chain transmissions, but this behavior is short-lived and leads to an early network collapse, so it is not sustainable. The proposed scheme achieves a moderate throughput of roughly 1.96 Mbits per round, which ensures efficient data delivery rate and better energy efficiency for long-term operation at the edge. EE-LEACH and the Federated Learning baseline deliver higher throughput (around 4.9 Mbits), reflecting their different communication patterns, but they do so with additional costs in either scalability or resource usage.

This trade-off demonstrates a critical multi-objective optimization, where acceptable throughput is maintained while keeping the network alive for much longer, rather than maximizing data rate at the expense of network lifetime.

In Fig. 8, comparison of the average number of hops per data path is shown. The proposed scheme keeps the hop count at a moderate level (about 102), which is roughly 59% fewer hops than Federated Learning (249) and about 90% fewer than PEGASIS (996), showing that the CoA-based routing finds much shorter routes. EE-LEACH achieves the smallest hop count (around 20) because clustered nodes often communicate almost directly.

Fewer hops translate into lower delay and less cumulative energy use, indicating that the proposed routing successfully balances trust, distance, and power.

Fig. 9 shows how often nodes enter power-saving mode. In the proposed method, about 20%–30% of nodes are typically in power-saving state, which helps conserve energy and extend the network lifetime while still keeping the topology connected. PEGASIS, by contrast, rarely benefits from power-saving because nodes die quickly, leaving little opportunity for controlled savings.

This behavior illustrates how the scheme reacts dynamically to node energy levels and traffic demand.

Fig. 10 plots the remaining global battery power. The proposed method preserves more residual energy throughout the simulation than both Federated Learning and PEGASIS, while EE-LEACH also performs well in terms of battery conservation. Effective power redistribution and fault-tolerant operation prevent unnecessary drain, which is critical for prolonging the lifetime of IoT edge nodes.

To evaluate resilience against malicious behavior, we implemented a composite attack combining trust inflation (malicious nodes initialize τ=1.0 despite bad behavior) and selective packet dropping (malicious nodes drop with probability 0.5). Table 9 shows performance under 20% malicious nodes (mal_ratio=0.2).

The scheme maintains stable trust (0.778) and full lifetime despite attack, confirming that the weighted trust update (Eq. (2)) and periodic decay effectively counter inflation attacks. Packet dropping reduces throughput proportionally to malicious density, but does not cascade to premature node death or trust collapse.

The baseline performance comparison metrics after 500 rounds are summarized in Table 10.

From the earlier discussed simulation results, analysis, and performance table, it can be found that the proposed scheme is robust and mainly contributes with three key advantages, which are essential for scalable, energy-efficient edge WSNs:

It provides an optimal tradeoff between network longevity and throughput rates, which offers more sustainable operation compared to PEGASIS, which quickly drains the network with short bursts of high throughput.

Therefore, this non-cluster, agent-based approach delivers better results compared to cluster-based EE-LEACH in terms of greater flexibility for heterogeneous, multi-path deployments and also cost-effectiveness with respect to the overhead cost and longer routes compared to the Federated Learning baseline. These strong claims with comprehensive results validate the applicability and advantages of the agent-based power management framework as a multi-objective solution for real-world wireless sensor networks operating in energy-constrained edge computing environments.