Size- and lifetime-aware disjoint dominating set formation and multi-armed-bandit-based scheduling for energy-efficient IoT-WSNs in sustainable smart city applications Authors: Ahmed Salim, Ahmed M Khedr, Ahmed A El-Sawy, Bader Alwasel, Walid Osamy Published: 2026 ABSTRACT AbstractThe versatility and broad applicability of IoT-based heterogeneous wireless sensor networks (HWSNs) make them key enablers of sustainability objectives in sustainable smart cities (SSCs). However, their heterogeneous architecture must be carefully managed to ensure reliable and energy-efficient operation. To prolong network lifetime and avoid premature node depletion, effective energy management mechanisms are essential. One promising approach is to exploit the concept of dominating sets (DSs), whereby sensor nodes are partitioned into disjoint DSs, and only one set is activated at a time. In this work, we propose EAAS-S4C-MAB, an enhanced framework for IoT-based HWSNs in SSCs that combines skyline-based DS formation, four-case, size- and lifetime-aware scheduling, and energy-aware data collection within the active DS. In the DS formation phase, the proposed Energy-Attentive Algorithm with Skyline (EAAS) generates multiple valid disjoint DS candidates per iteration, evaluates them jointly in terms of lifetime and size, and applies a BNL_Skyline procedure to retain only non-dominated candidates. In the scheduling phase, the proposed S4C-MAB algorithm classifies DSs into four cases according to size and lifetime and uses a multi-armed bandit to learn the relative priority of these cases online, replacing fixed case weights with adaptive case-level preferences. These learned priorities are combined with a dynamically recomputed adjusted lifetime, along with case-dependent activation caps and cooldown intervals, to guide DS activation. During communication, the active DS nodes form a chain, and the node nearest to the sink serves as the chain leader for final data delivery. Simulation results show that EAAS-S4C-MAB achieves more balanced DS utilization, reduces the premature depletion of fragile sets, and significantly extends network lifetime compared to other baseline methods, while preserving network coverage and connectivity.Keywords: Dominating set, Heterogeneous wireless sensor networks, Network lifetime, Sustainable urbanization, Multi-armed-bandit, Energy consumption, Sustainable smart city, Modern citiesSubject terms: Energy science and technology, Engineering, Mathematics and computing Introduction With rapid global urbanization, the concept of the Sustainable Smart City (SSC) has emerged as an important approach for addressing the challenges associated with modern urban development. As the urban population continues to grow, SSCs are expected to provide efficient, reliable, and sustainable services across multiple domains, including smart governance, smart water and waste management, smart transportation, and smart connectivity. One of the key technological enablers of these services is the Internet of Things (IoT), particularly through Wireless Sensor Networks (WSNs), which support data acquisition, monitoring, and communication for a wide range of smart applications. In IoT-based WSNs, sensor nodes are typically deployed to monitor physical environments and forward collected data to a sink node through multi-hop communication. Unlike regular sensor nodes, sink nodes are generally equipped with abundant energy, stronger computing capabilities, wider communication ranges, and larger memory capacity1–4. Despite their broad adoption in agricultural monitoring, healthcare, smart homes, and industrial automation5, WSNs still face several practical limitations, including restricted battery capacity, limited communication range, constrained storage, and limited processing capability6,7. These limitations become even more critical in Heterogeneous WSNs (HWSNs), where nodes differ in energy level, sensing capability, and communication resources. As a result, efficient energy management remains one of the most important design challenges for ensuring network stability, survivability, and long operational lifetime in SSC environments5,8. Many strategies have been proposed to conserve energy and prolong the lifetime of WSNs. Since sensing and communication are the major sources of energy consumption, effective energy-aware scheduling is essential for extending network service time. Existing approaches include clustering-based routing, machine-learning-assisted cluster-head selection, and sleep–awake scheduling9,10. For example, Juwaied and Jackowska-Strumillo9 proposed ML-enhanced variants of the DEEC protocol, namely DEEC-KNN and DEEC-KM, by integrating K-Nearest Neighbours and K-Means into the cluster-formation process to improve energy efficiency, packet delivery, and network stability. Although such clustering-based methods improve energy utilization, they do not explicitly coordinate the activation of disjoint node sets for sleep–awake scheduling. One common approach is to alternate sensor nodes between sleep and active states11. Because WSNs are often densely deployed, temporarily switching a subset of nodes to sleep mode can significantly reduce energy consumption without compromising monitoring functionality. Under this strategy, only the nodes in the active set perform sensing and communication, while the remaining nodes stay in a low-power sleep state. This work builds on the concept of Dominating Sets (DSs) in graph theory to regulate the sleep–awake schedule of heterogeneous sensor nodes. A DS is a subset of nodes such that every node in the network either belongs to the subset or has at least one neighbor in it. To further reduce energy consumption, the network can be partitioned into disjoint DSs, only one set is active at a time12. This sleep–awake scheduling strategy prolongs network lifetime by distributing sensing and data-gathering tasks across multiple DSs. The main contributions of this paper can be summarized as follows: EAAS: we propose an energy- and size-aware dominating-set construction method that generates multiple feasible DS candidates at each iteration, evaluates them jointly in terms of lifetime and size, and applies a BNL_Skyline procedure to retain only non-dominated candidates. The final DS is then selected using either a composite-score rule or a knee-point heuristic, enabling a more balanced trade-off between robustness and compactness. EAAS: we propose an energy- and size-aware dominating-set construction method that generates multiple feasible DS candidates at each iteration, evaluates them jointly in terms of lifetime and size, and applies a BNL_Skyline procedure to retain only non-dominated candidates. The final DS is then selected using either a composite-score rule or a knee-point heuristic, enabling a more balanced trade-off between robustness and compactness. S4C-MAB: we develop a size- and lifetime-aware DS scheduling algorithm that classifies DSs into four categories (small–high, large–high, small–low, and large–low), applies case-dependent activation caps and cooldown intervals, dynamically updates adjusted lifetimes, and learns case priorities online through a multi-armed bandit mechanism. This allows the scheduler to adaptively balance DS usage rather than relying on fixed scheduling rules. S4C-MAB: we develop a size- and lifetime-aware DS scheduling algorithm that classifies DSs into four categories (small–high, large–high, small–low, and large–low), applies case-dependent activation caps and cooldown intervals, dynamically updates adjusted lifetimes, and learns case priorities online through a multi-armed bandit mechanism. This allows the scheduler to adaptively balance DS usage rather than relying on fixed scheduling rules. Comprehensive two-level validation: we conduct a two-level experimental evaluation. At the DS-formation level, we compare EAAS with existing DS construction methods and with a non-skyline ablation variant to isolate the effect of skyline-based selection. At the full-framework level, we compare EAAS-S4C-MAB with representative WSN routing and scheduling protocols under different densities and sink locations, showing consistent gains in stability and energy efficiency. Comprehensive two-level validation: we conduct a two-level experimental evaluation. At the DS-formation level, we compare EAAS with existing DS construction methods and with a non-skyline ablation variant to isolate the effect of skyline-based selection. At the full-framework level, we compare EAAS-S4C-MAB with representative WSN routing and scheduling protocols under different densities and sink locations, showing consistent gains in stability and energy efficiency. By combining Skyline-based multi-objective DS formation with adaptive case-aware scheduling, EAAS-S4C-MAB achieves more balanced DS utilization, reduces premature depletion of fragile sets, and significantly extends HWSN lifetime. The remainder of this paper is organized as follows. “Related work” reviews related work. “Proposed EAAS-S4C-MAB technique” details the proposed EAAS-S4C-MAB technique, including the EAAS formation stage and S4C-MAB scheduling stage. “Results and discussion” presents simulation results and compares performance with existing approaches. Finally, “Conclusion” concludes the paper. Related work The rapid increase of smart sensing devices within Sustainable Smart Cities (SSCs) has generated a pressing need for algorithmic approaches that are both resource-efficient and adaptive to the limited resources and heterogeneous characteristics of Wireless Sensor Networks (WSNs)13–16. The successful realization of technological advancements in SSCs relies on robust infrastructure as well as the integration of e-governance policies for efficient resource management17–19. Integrating WSNs with analogous technologies offers a wide array of applications in SSCs and has the potential to yield comprehensive strategies for sustainable communities20–22. Data insights from WSNs can be used to make SSCs more sustainable by enabling city planners and decision-makers to make informed choices that promote environmentally friendly urban development. Extending the lifetime of WSNs not only reduces environmental impact but also improves data reliability and frees up resources for other sustainable initiatives23–25. Consequently, there is a growing need to design paradigms that simultaneously conserve resources and address multiple facets of SSCs. While much of the literature focuses on energy-saving methods for homogeneous WSNs–where devices have identical energy reserves, connectivity, and processing capabilities–heterogeneous networks present added challenges due to diverse resource capacities26–28. Efficient design and resource management in heterogeneous WSNs (HWSNs) can reduce environmental impact, lower costs, and improve quality of life in smart cities29–31. To address energy constraints in WSNs, various solutions have been proposed. Since sensing and transmission are the most energy-intensive operations, many studies adopt a sleep–awake scheme as an effective means of conserving node energy19–21. For example, Threshold LEACH (T-LEACH)32 extends network lifetime by reducing energy consumption in data transmission and aggregation. Unlike LEACH, which selects cluster heads in every round, T-LEACH retains the current cluster heads until their residual energy drops below a dynamic threshold. This reduces cluster-head changes and the associated advertisement overhead. It also considers node heterogeneity and residual energy when selecting the optimal CH. EEM-TDMA33 is an energy-efficient TDMA scheduling method for LEACH-based wireless sensor networks. In this approach, each cluster head sends its cluster size to the base station, the base station determines the maximum cluster capacity, and then broadcasts that value back so all cluster heads can build a common TDMA schedule without inter-cluster-head communication. This reduces control overhead, saves energy, and improves packet delivery and network lifetime. T-LEACHSAS34is an improved LEACH-based protocol for wireless sensor networks that saves energy in two ways: it keeps a cluster head in place until its energy falls below a threshold, and it uses sleep–awake scheduling so sensor nodes stay asleep except when they need to sense or transmit in their TDMA slot. This reduces control overhead, idle listening, and unnecessary transmissions, so the network lasts longer than with basic LEACH and T-LEACH32. CDS-LEACH35 is a modified LEACH protocol for wireless sensor networks that selects cluster heads not only by energy level, but also by whether they belong to a Connected Dominating Set (CDS). This improves connectivity inside clusters, helps data reach cluster heads more reliably, and reduces overall energy use. The paper reports that CDS-LEACH performs better than standard LEACH, especially in denser networks, with later node death, lower cluster-head energy expenditure, and more data delivered. Enhanced LEACH36is a modified LEACH protocol for wireless sensor networks that improves cluster-head selection by considering nodes’ remaining energy and allowing all nodes to rejoin cluster-head election instead of excluding recent cluster heads. This leads to better energy use, later first-node death, more surviving nodes, and more packets delivered than standard LEACH, so the network lasts longer. Recent studies have also explored machine learning (ML)-assisted clustering and routing protocols to improve energy efficiency and network stability in WSNs. Juwaied and Jackowska-Strumillo9 proposed two ML-enhanced variants of the Distributed Energy-Efficient Clustering (DEEC) protocol, namely DEEC-KNN and DEEC-KM. These approaches integrate K-Nearest Neighbours (K-NN) and K-Means clustering into the cluster-head selection process to improve adaptive cluster formation, packet delivery, and network lifetime. Their results showed improved network stability and up to 110% improvement in packet delivery compared with the original DEEC protocol. Similarly, Juwaied10 introduced EMO-PEGASIS, a dual-phase machine-learning-based PEGASIS enhancement that combines K-means clustering with K-NN-based routing decisions to reduce energy consumption and end-to-end delay. Although these ML-assisted protocols demonstrate the potential of adaptive learning methods for WSN optimisation, they remain clustering- or routing-oriented frameworks and do not explicitly address the coordinated formation of multiple disjoint dominating sets, DS size–lifetime balancing, or activation scheduling. Although LEACH-based, PEGASIS-based, and ML-assisted routing protocols improve energy efficiency, delay, and load balancing, they do not explicitly address the coordinated formation and activation of multiple disjoint dominating sets. This limitation has motivated more structured DS-based approaches. One promising strategy leverages the concept of Dominating Sets (DSs), where nodes are partitioned into disjoint DSs and only one set is active at a time27,28. One promising strategy leverages the concept of Dominating Sets (DSs), where nodes are partitioned into disjoint DSs and only one set is active at a time27,28. Determining the minimum DS is, however, an NP-Hard problem23. Early works such as27,28 proposed centralized approximation techniques with approximation factors O(łog n) and , respectively. The DS technique employing a dual-greedy approach for sleep–awake scheduling was first presented in27 and later extended in28 through the concept of maximum disjoint DSs (Domatic Partition, DP) in Unit Disk Graphs (UDGs) for homogeneous networks. To overcome scalability issues in centralized schemes,30,37 explored distributed algorithms for DS construction. In29, DSs were generated using a distributed Maximal Independent Set (MIS) principle. The work in38 proposed a three-stage approach with Pseudo DSs to form smaller Connected DSs (CDSs) while improving connectivity and coverage. In20, CDS connectivity was enhanced using a non-trivial potential function, and a greedy algorithm was applied to generate three-connected components from Tutte’s decomposition of a two-connected graph. Additional DS-based enhancements have been proposed. For example,24 presented a clustering approach (EBDSC) relying on DSs, and25 introduced a decentralized selection method with a Wait Before Start (WBS) policy to allow dominant nodes to self-announce based on neighborhood size. In31, a cell layout was used to address the DP problem by constructing cliques per cluster and implementing a distributed nucleus algorithm (DNA) that executes in finite iterations under the congested message-passing model. Although these algorithms improved DS construction efficiency, most focused on homogeneous settings or static energy assumptions. Some works, such as22, generated CDSs in heterogeneous environments using local search strategies; however, their complexity remains high, and scalability is limited. A significant step toward addressing energy heterogeneity in SSC-based WSNs was taken in the original EADDSA framework39, which introduced the Energy Attentive Algorithm (EAA) to form energy-aware disjoint DSs and a scheduling strategy to extend network lifetime. While EADDSA improved lifetime compared to prior methods, it optimized DS formation using energy only, without explicitly considering DS size or workload balance, and employed a scheduling policy that could overuse certain sets. Motivated by these gaps, we propose EAAS-S4C-MAB, an enhanced DS-based framework for IoT-enabled HWSNs in SSC applications. EAAS-S4C-MAB integrates EAAS, a multi-objective DS formation algorithm that balances lifetime and size using the BNL_Skyline procedure, with S4C-MAB, a size- and lifetime-aware scheduling strategy employing a four-case policy with activation caps and cooldowns. By jointly optimizing DS formation and scheduling, EAAS-S4C-MAB achieves more balanced energy consumption, mitigates premature depletion of fragile sets, and extends network lifetime beyond that of the original EADDSA. Table 1 summarizes the key contributions and limitations of existing DS-based energy conservation approaches in WSNs, positioning our proposed method within the context of prior work. Table 1Summary of related work on energy conservation and scheduling approaches in WSNs.ReferencesApproach typeKey features / methodsLimitations9 ML-assisted DEEC clusteringIntegrates K-NN and K-Means with DEEC for adaptive cluster-head selectionNot DS-based; does not optimize disjoint DS construction, DS size-lifetime trade-off.10 ML-assisted PEGASIS routingK-means clustering for spatial partitioning; K-NN classification for adaptive routingNot DS-based; does not optimize disjoint DS formation, DS size-lifetime trade-off.20Connectivity-enhanced CDSGreedy algorithm from Tutte’s decomposition for 3-connected CDSsFocus on connectivity over energy efficiency22Heterogeneous CDS formationLocal search with variable initial energiesHigh complexity; limited scalability24Clustering via DSEBDSC clustering approachNo explicit DS size/lifetime optimization25Decentralized DS selectionWait Before Start (WBS) method; time-weighted announcementsNo energy–size trade-off consideration27Centralized DS constructionDual-greedy approach to form disjoint DSs; sleep–awake scheduling for homogeneous WSNsIgnores heterogeneity; centralized control limits scalability; fault tolerance is not explicitly addressed28Centralized DS partitioningDomatic Partition on Unit Disk Graphs; maximizes number of disjoint DSsHomogeneous assumption; high computation in large networks; fault tolerance is not explicitly addressed29Distributed MIS-based DSMaximal Independent Set principle to indirectly form DSsDoes not optimize for energy or lifetime30,37Distributed DS constructionLocalized algorithms for DS generationLimited optimization for energy balance or size; fault tolerance is not explicitly addressed31Distributed nucleus algorithmClique per cluster; executes in finite iterationsOverhead in dense networks; ignores heterogeneity38Three-stage pseudo-DS CDSSmaller Connected DSs with improved coverage and connectivityIncreased complexity; not energy-aware39EADDSA (Energy-aware DS)Energy Attentive Algorithm (EAA) + scheduling for IoT-HWSNs in SSCsConsiders energy only; fault tolerance is not explicitly addressed.;no size/lifetime trade-off; scheduling may overuse certain DSsProposed workEAAS-S4C-MABEAAS for lifetime–size balanced DS formation using BNL_Skyline; S4C-MAB for four-case scheduling based MAB; improves balanceFramework does not explicitly address fault tolerance mechanism Summary of related work on energy conservation and scheduling approaches in WSNs. Proposed EAAS-S4C-MAB technique The proposed EAAS-S4C-MAB (Energy-Aware Disjoint Dominating Set Algorithm with Skyline and Four-Case Multi-Armed Bandit Scheduling) consists of three tightly coupled components that operate sequentially to maximize the lifetime of heterogeneous wireless sensor networks (HWSNs) while maintaining coverage and connectivity. EAAS: energy- and size-balanced disjoint ds construction. In this phase, the network is modeled as a graph, and the algorithm iteratively constructs disjoint dominating sets (DSs) that satisfy connectivity, disjointness, and positive-energy constraints. EAAS generates multiple valid DS candidates per iteration and evaluates them based on two objectives: (i) lifetime L(D), defined by the minimum residual energy in the set divided by the per-round consumption rate, and (ii) size |D|, i.e., the number of active nodes. The BNL_Skyline procedure is then applied to identify the set of non-dominated DSs in the lifetime–size space. The final DS for the iteration is selected from the Skyline set using either a composite-score rule or a knee-point heuristic, thereby balancing robustness and energy efficiency. EAAS: energy- and size-balanced disjoint ds construction. In this phase, the network is modeled as a graph, and the algorithm iteratively constructs disjoint dominating sets (DSs) that satisfy connectivity, disjointness, and positive-energy constraints. EAAS generates multiple valid DS candidates per iteration and evaluates them based on two objectives: (i) lifetime L(D), defined by the minimum residual energy in the set divided by the per-round consumption rate, and (ii) size |D|, i.e., the number of active nodes. The BNL_Skyline procedure is then applied to identify the set of non-dominated DSs in the lifetime–size space. The final DS for the iteration is selected from the Skyline set using either a composite-score rule or a knee-point heuristic, thereby balancing robustness and energy efficiency. S4C-MAB: size- and lifetime-aware DS scheduling with four-case policy. In this scheduling phase, the disjoint DSs produced by EAAS are classified into four cases according to their size and residual lifetime: small–high, large–high, small–low, and large–low. Each case is associated with a specific activation cap and cooldown interval in order to prevent premature depletion of fragile sets and to maintain steady use of robust sets. In each scheduling round, the scheduler recomputes an adjusted lifetime for every DS, incorporating a size penalty and, when available, a load penalty. A lightweight multi-armed bandit mechanism then learns a time-varying priority weight for each case based on the observed scheduling reward, and the DS with the maximum case-weighted adjusted lifetime is selected for activation. In this way, scheduling decisions adapt online as node energies and case utilities evolve. S4C-MAB: size- and lifetime-aware DS scheduling with four-case policy. In this scheduling phase, the disjoint DSs produced by EAAS are classified into four cases according to their size and residual lifetime: small–high, large–high, small–low, and large–low. Each case is associated with a specific activation cap and cooldown interval in order to prevent premature depletion of fragile sets and to maintain steady use of robust sets. In each scheduling round, the scheduler recomputes an adjusted lifetime for every DS, incorporating a size penalty and, when available, a load penalty. A lightweight multi-armed bandit mechanism then learns a time-varying priority weight for each case based on the observed scheduling reward, and the DS with the maximum case-weighted adjusted lifetime is selected for activation. In this way, scheduling decisions adapt online as node energies and case utilities evolve. Data collection and intra-DS communication. Once a DS is activated, it serves as the temporary sensing and forwarding backbone of the network. A communication chain is formed over the nodes of the active DS. Data are then aggregated and forwarded hop-by-hop along the chain toward a chain leader. This communication model ensures that the selected DS not only preserves domination coverage, but also acts as an energy-aware data gathering and forwarding structure for the entire network. Data collection and intra-DS communication. Once a DS is activated, it serves as the temporary sensing and forwarding backbone of the network. A communication chain is formed over the nodes of the active DS. Data are then aggregated and forwarded hop-by-hop along the chain toward a chain leader. This communication model ensures that the selected DS not only preserves domination coverage, but also acts as an energy-aware data gathering and forwarding structure for the entire network. By integrating Skyline-based DS formation with four-case, size- and lifetime-aware bandit scheduling and an active-DS communication backbone, EAAS-S4C-MAB explicitly addresses DS robustness, workload balance, and data forwarding efficiency. As a result, the framework improves network stability and extends the operational lifetime of the HWSN. Table 2 summarizes the commonly used symbols, along with their meanings and descriptions. Figure 1 depicts the overall flow of the EAAS-S4C-MAB technique, illustrating the interaction between the EAAS construction stage, the S4C-MAB scheduling stage, and the communication stage within the activated DS. Table 2Table of symbols and notations.SymbolMeaningDescriptionG[C]Induced subgraphSubgraph generated from the node subset Ckth dominating setThe disjoint dominating set generated by the frameworkOptimal DS collectionOptimal collection of dominating sets maximizing network lifetimeDS lifetimeEstimated operational lifetime of dominating set Residual energyRemaining energy of node u at time tCurrent node energyCurrent energy level of node u during schedulingMinimum residual energyMinimum residual energy among all nodes in δEnergy depletion rateConstant per-round energy consumption rateEstimated consumption rateEstimated energy consumption of node u when is activatedSize penalty functionPenalty function associated with the DS sizeβSize penalty coefficientParameter controlling the impact of DS size in optimizationηSize penalty parameterPenalty factor used to discourage inefficient DS sizesγLoad penalty factorWeight factor associated with communication load balancingNormalized loadNormalized communication load of dominating set Adjusted lifetimeLifetime adjusted according to size and load penaltiesCommunication radiusMaximum communication range between sensor nodesSkyline setSet of non-dominated DS candidates after Skyline filteringComposite scoreCombined lifetime-size evaluation metricKnee-point scoreBalanced trade-off score between lifetime and DS sizeξSelector typeDS selection mechanism (composite or Knee selector)Size thresholdThreshold separating small and large DSsLifetime thresholdThreshold separating high and low lifetime DSsCase weightLearned priority weight for case c at time tEmpirical mean rewardAverage observed reward of case cSelection countNumber of times case c has been selectedExploration parameterExploration coefficient in the UCB-based MAB algorithmImmediate rewardReward obtained after activating a DSPre-activation utilityAggregate network utility before DS activationPost-activation utilityAggregate network utility after DS activationEligible DS setSet of DSs eligible for scheduling at round tValid residual DSsRemaining valid DSs after removing dead nodesActivation capsMaximum consecutive activations allowed per caseCooldown intervalsWaiting intervals before reactivating a DS caseLifetime estimateEstimated lifetime of dominating set Skyline tuplesCandidate tuples evaluated in skyline processingDominance relationIndicates that tuple dominates tuple Table of symbols and notations. Overall workflow of EAAS-S4C-MAB. We begin with a description of the system model and the formulation of the problem. Subsequently, we offer a detailed description of the proposed EAAS-S4C-MAB framework. System model and problem formulation In our proposed framework, we represent the WSN as a graph, denoted as , where V and represent the set of nodes and the communication links between them. If we take nodes i and j ∈ V, an edge between them exists when i is within the communication range of j. A fundamental assumption in our model is the symmetry of communication links between nodes within the WSN, meaning that all nodes in the network share an identical communication range. Despite this symmetry, the battery-powered nodes exhibit heterogeneity in their initial energy levels. Additionally, we consider a constant energy usage rate across all nodes in the network. A subset D ⊆ V is called a dominating set (DS) if every node outside D is adjacent to at least one node in D. Formally, D is a DS of G if, for every node , there exists a node u ∈ D such that . In the context of WSNs, the nodes selected in the DS can act as active sensing, aggregation, or forwarding nodes on behalf of their neighboring nodes, while non-selected nodes can remain in a low-power sleep state. Therefore, rotating the activation of different DSs can reduce redundant communication and extend the network lifetime. A connected dominating set (CDS) is a stricter form of a DS. A subset C ⊆ V is a CDS if it is a dominating set of G and the subgraph induced by C, denoted by G[C], is connected. Thus, while a DS guarantees that every non-selected node is covered by at least one selected neighbor, a CDS additionally guarantees connectivity among the selected nodes 40. In this work, the selected sets are required to be DSs rather than CDSs because communication among active nodes is handled separately through the chain-based data-gathering phase. A DS with a smaller size tends to exhibit superior energy efficiency, as fewer active nodes per round allow more nodes to enter a low-power sleep state. However, smaller DSs can increase the workload per node, potentially reducing the DS lifetime if low-energy nodes are selected. Conversely, larger DSs distribute the load over more nodes but keep more devices active, consuming more total energy per round. Therefore, a trade-off exists between DS size and DS lifetime. To address this balance, we allow DS selection to consider both the minimum residual energy of the set and its size, promoting solutions that are energy-aware while avoiding excessively large or small DSs. A strategy involves substituting an existing DS with another disjoint DS and transitioning the nodes in the previous set to sleep mode. This entails employing as many disjoint DSs as possible, utilizing each for a specific duration, and subsequently replacing it with another set. The primary goal is to improve the HWSN lifespan by carefully selecting disjoint DSs based on both energy and size considerations. The lifespan of a DS is dependent on the node with the lowest energy, factoring in its energy consumption rate, while its size affects per-node workload. The cumulative network lifetime is calculated as the sum of the lifetimes of all disjoint DSs. The following assumptions are adopted: The network is heterogeneous in terms of initial energy; each node u∈ V is assigned an initial energy , which may differ from that of other nodes. The network is heterogeneous in terms of initial energy; each node u∈ V is assigned an initial energy , which may differ from that of other nodes. The sink location is fixed, and the sensor nodes are static. Each node has power-control capability and can adjust its transmission power according to the communication distance. The sink location is fixed, and the sensor nodes are static. Each node has power-control capability and can adjust its transmission power according to the communication distance. The communication radius is used to define the neighborhood graph for the domination/coverage constraint. The communication radius is used to define the neighborhood graph for the domination/coverage constraint. After a DS is selected, data gathering is performed through a chain-based communication phase inspired by PEGASIS-like data gathering 10,41. Therefore, the DS is required to satisfy the domination constraint for coverage, but it is not required to be connected as an induced subgraph. After a DS is selected, data gathering is performed through a chain-based communication phase inspired by PEGASIS-like data gathering 10,41. Therefore, the DS is required to satisfy the domination constraint for coverage, but it is not required to be connected as an induced subgraph. Chain links are evaluated using Euclidean inter-node distances and the adopted radio-energy model. Thus, the chain-based communication cost is handled separately from the graph-theoretic domination constraint. Chain links are evaluated using Euclidean inter-node distances and the adopted radio-energy model. Thus, the chain-based communication cost is handled separately from the graph-theoretic domination constraint. Constraints The following considerations are taken into account: Energy constraint: the total expended energy of all nodes in a DS should not exceed the complete available energy of the WSN, denoted as : 1 where represents the energy of node i in DS , and j is the disjoint DS count. The energy of each node in a DS should be greater than zero, considering a positive constant δ: for all , with δ > 0. Energy constraint: the total expended energy of all nodes in a DS should not exceed the complete available energy of the WSN, denoted as : 1 where represents the energy of node i in DS , and j is the disjoint DS count. The energy of each node in a DS should be greater than zero, considering a positive constant δ: for all , with δ > 0. Disjointness constraint: each pair of DSs should have no common nodes: for all . Disjointness constraint: each pair of DSs should have no common nodes: for all . Connectivity constraint: every node outside of a DS should have at least one neighbor in some DS: for each , there exists a DS D ⊂ V where v has at least one neighbor in D. Connectivity constraint: every node outside of a DS should have at least one neighbor in some DS: for each , there exists a DS D ⊂ V where v has at least one neighbor in D. Chain-based communication constraint: for each activated DS , a chain is constructed over the active nodes in . Data are forwarded along the chain toward a selected leader node, and the leader transmits the aggregated packet to the sink. The resulting transmission, reception, aggregation, and leader-to-sink costs are included in the energy update. Chain-based communication constraint: for each activated DS , a chain is constructed over the active nodes in . Data are forwarded along the chain toward a selected leader node, and the leader transmits the aggregated packet to the sink. The resulting transmission, reception, aggregation, and leader-to-sink costs are included in the energy update. Objective function Let denote the set of selected node-disjoint dominating sets. The objective is to select such that it maximizes the cumulative lifetime contribution of the selected DSs while penalizing inefficient set sizes: subject to the positive-energy, domination/coverage, and disjointness constraints. Here, denotes the estimated lifetime of , and is a size-penalty function. In this work, we use . Therefore, the objective becomes The DS lifetime is defined as , where is the residual energy of node u, and is the estimated energy consumed by node u when is activated. During candidate generation, a simplified constant-depletion approximation δ is used: Under this approximation, the same objective reduces to The depletion parameter δ is set to 0.05 and is used by EAAS to estimate the nominal lifetime of a candidate dominating set as . In the simulation, the same value is used as the initial depletion estimate for computing initial DS lifetimes and thresholds. Actual energy consumption during each round is then computed by the radio-energy model according to the selected communication scheme. Overview of multi-armed bandits A multi-armed bandit (MAB) is a sequential decision-making framework in which a learner repeatedly selects one action, called an arm, and observes a reward only for the chosen action. The goal is to maximize cumulative reward over time while balancing exploration of less-sampled arms and exploitation of arms that currently appear most rewarding. The stochastic MAB formulation was introduced by Robbins, with later regret-optimal allocation results established by Lai and Robbins, and practical finite-time upper-confidence-bound (UCB) policies developed by Auer et al.42,43. In a standard UCB-based MAB with arms , the priority index of arm c at round t is computed as: where is the empirical mean reward of arm c, is the number of times it has been selected, and β >0 controls the exploration level. This rule favors arms with high observed reward while still assigning higher priority to under-explored arms43. In the EAAS-S4C-MAB framework, the four DS cases are modeled as the four bandit arms, i.e., Thus, instead of learning a separate policy for each dominating set, the scheduler learns only a case-level priority weight for each DS category. This keeps the scheduler lightweight while still allowing it to adapt online according to observed scheduling reward. Overview of skyline operator The Skyline operator44 identifies points (tuples) in a dataset that are optimal with respect to multiple, often conflicting, criteria. A classic example is selecting a hotel based on price and distance to the beach. Typically, hotels closer to the beach are more expensive, and cheaper hotels are farther away. The Skyline operator returns only those hotels that are not worse in both criteria than any other hotel. Definition Let represent an n-dimensional tuple, where each dimension corresponds to a selection criterion. Assuming minimization objectives (e.g., lower is better), a tuple tp dominates another tuple tq (denoted tp ≻ tq) if: The Skyline is the set of all tuples not dominated by any other tuple in the dataset. Computation approaches Two broad strategies exist for computing Skylines: Database-extended approach: integrating Skyline queries into database systems. While expressive, this can lead to complex queries and poor performance in high-dimensional data. Database-extended approach: integrating Skyline queries into database systems. While expressive, this can lead to complex queries and poor performance in high-dimensional data. Algorithmic approach: employing dedicated algorithms such as Block-Nested Loops (BNL) or Divide-and-Conquer. Algorithmic approach: employing dedicated algorithms such as Block-Nested Loops (BNL) or Divide-and-Conquer. We adopt the Block-Nested Loops (BNL) algorithm due to its simplicity and efficiency, particularly when the Skyline set is small45. BNL algorithm overview BNL maintains a Skyline list in main memory. Each incoming tuple from the input list is compared against those in the Skyline list: If an existing Skyline tuple dominates the new tuple, the new tuple is discarded. If an existing Skyline tuple dominates the new tuple, the new tuple is discarded. If the new tuple dominates one or more Skyline tuples, those dominated tuples are removed, and the new tuple is added. If the new tuple dominates one or more Skyline tuples, those dominated tuples are removed, and the new tuple is added. If the new tuple is incomparable to all Skyline tuples, it is added to the Skyline list. If the new tuple is incomparable to all Skyline tuples, it is added to the Skyline list. At the end of processing, the Skyline list contains only non-dominated tuples. EAAS: energy and size balanced disjoint DS construction Candidate DSs are generated using the EAAS-Skyline procedure. The algorithm starts from the remaining node set V and initializes a working set D=V. The nodes are sorted in ascending order of residual energy, and the algorithm attempts to remove low-energy nodes one by one. Whenever the reduced set remains a valid dominating set, it is stored in a temporary candidate pool , together with its estimated lifetime L(D), size |D|, and node set D. A BNL skyline filter is then applied to using two objectives: maximizing L(D) and minimizing |D|. The nondominated candidates are temporarily stored in . One candidate is then selected from using either the knee selector or the composite lifetime–size score. The selected DS is appended to the final disjoint DS set, and its nodes are removed from V to enforce disjointness. This process is repeated until no additional valid DS can be formed. After skyline filtering, one DS must be selected from the nondominated candidate set. Two selection rules are supported. The composite selector assigns each candidate D the scalar score where L(D) is the estimated lifetime of D, |D| is its size, and η is a size-penalty parameter. This rule explicitly penalizes very small candidate sets that may overload their active nodes. The knee selector, in contrast, normalizes the lifetime and inverse-size objectives and selects the candidate with the largest product of the normalized objectives: where is the normalized lifetime objective and is the normalized inverse-size objective. This rule selects a balanced trade-off solution without requiring a manually tuned size-penalty coefficient. In this work, the knee selector is used to choose the final DS from the skyline set. The detailed operations of EAAS are as follows. EAAS operates in iterations. In iteration r (r \ge 1), the goal is to determine a DS such that for r ≠ k, r, k \ge 1, and . Iteration process In the first iteration, we initialize , where is a temporary set. Nodes in are sorted in ascending order of their residual energies. We iteratively test the elimination of the lowest-energy node v: If still satisfies DS constraints, v is removed from and added to (the candidate set for the next iteration). The neighbor lists are updated accordingly. If still satisfies DS constraints, v is removed from and added to (the candidate set for the next iteration). The neighbor lists are updated accordingly. If removing v violates DS coverage, v is retained in . If removing v violates DS coverage, v is retained in . Instead of immediately committing to the resulting , each valid intermediate DS encountered is stored in a candidate pool . Skyline selection Each DS is evaluated using two metrics: , where is the minimum residual energy among nodes in D. , where is the minimum residual energy among nodes in D. |D|, the cardinality (size) of the DS. |D|, the cardinality (size) of the DS. A DS A dominates DS B if L(A) \ge L(B) and |A| łe |B|, with at least one strict inequality. The Skyline set consists of all non-dominated DSs in . After processing all candidates, the final is chosen from the Skyline–e.g., using a knee-point heuristic or a composite score with a small size penalty factor η. Subsequent iterations For r > 1, we initialize and repeat the process: Sort by ascending energy. Sort by ascending energy. Iteratively attempt to remove low-energy nodes without breaking DS coverage, adding valid DSs to . Iteratively attempt to remove low-energy nodes without breaking DS coverage, adding valid DSs to . Maintain the Skyline set by discarding dominated DSs as new candidates are added. Maintain the Skyline set by discarding dominated DSs as new candidates are added. Select from the Skyline as above. Select from the Skyline as above. Nodes in are removed from the graph, and is formed from eliminated nodes for the next iteration. Termination conditions The process stops when: The remaining nodes cannot form a valid DS, or The remaining nodes cannot form a valid DS, or All nodes have been assigned to disjoint DSs. All nodes have been assigned to disjoint DSs. The EAAS algorithm ensures that in each iteration, the selected DS achieves a balanced trade-off between lifetime and active node count, improving overall WSN longevity and efficiency. The complete pseudocode is presented in Algorithm 1. Energy-attentive algorithm with skyline filtering (EAAS). BNL_Skyline on DS tuples. EAAS description The EAAS algorithm (Algorithm 1) begins by initializing the output collection of disjoint dominating sets as empty, setting the remaining node set to the full node set V, and initializing the iteration counter r to 1 (lines 1–3). The algorithm then proceeds while there are still remaining nodes that may form additional disjoint dominating sets (line 4). At the start of each iteration, the working dominating-set candidate D is initialized as the current remaining node set , and a temporary candidate pool is created (lines 5–6). The algorithm first checks whether D is a dominating set of the original graph G (line 7). If it is, the candidate D, together with its size |D| and lifetime value L(D), is inserted into (line 8). Otherwise, no further disjoint dominating set can be constructed from the remaining nodes, and the algorithm terminates (lines 9–10). Next, the nodes in are sorted in ascending order of residual energy e(u) (line 12). This ordering gives priority to the removal of low-energy nodes, thereby attempting to preserve higher-energy nodes in the selected dominating sets. The algorithm then examines each node s in the sorted list (lines 13–19). For each node, it forms a temporary set (line 14). If T still satisfies the dominating-set condition with respect to G (line 15), the working set is updated as D ← T (line 16), and this reduced dominating set is added to the candidate pool (line 17). In this way, accumulates feasible dominating sets obtained by progressively removing low-energy nodes while preserving coverage. After the candidate-generation phase, the lifetime of each candidate is computed as where δ is the per-round depletion rate (line 20). This lifetime represents the number of rounds for which the weakest-energy node in D can continue operating. The candidate pool is then passed to the BNL_Skyline procedure, which retains only the non-dominated candidates according to two objectives: maximizing lifetime L(D) and minimizing cardinality |D| (line 21). A candidate is dominated if another candidate has at least as large a lifetime and no larger size, with strict improvement in at least one objective. From the resulting Skyline set , the algorithm selects one dominating set according to the selector type ξ (lines 23–30). If the composite selector is used, the selected candidate maximizes where η controls the penalty imposed on larger dominating sets (lines 23–26). If the knee-based selector is used, the algorithm normalizes the lifetime objective and the inverse-size objective 1/|D|, then selects the candidate that maximizes the product of the normalized objectives (lines 27–30). The selected dominating set is then added to the output collection (line 31). To enforce disjointness among the selected DSs, all nodes in are removed from the remaining node set by updating (line 32). The iteration counter is then incremented (line 33), and the process repeats until either becomes empty or the remaining nodes can no longer form a valid dominating set. The final output is therefore a collection of mutually disjoint DSs. Each selected set is chosen from the Skyline of feasible candidates in its iteration, balancing two competing objectives: maximizing network lifetime and minimizing DS size. BNL_Skyline description The BNL_Skyline subroutine (Algorithm 2) applies the Block-Nested Loops (BNL) method to identify non-dominated DS candidates based on two attributes: lifetime L(D) and size |D|. It begins by initializing the Skyline list as empty (line 1). For each tuple (L, m, D) in the input list X (line 2), two temporary variables are created: a Boolean flag dom set to false, and a set toRemove initialized as empty. The algorithm then iterates over each tuple (L', m', D') currently in (lines 3–9) and performs pairwise dominance checks. If (L, m) dominates (L', m') – meaning that L \ge L' and m łe m', with at least one strict inequality – then (L', m', D') is added to the toRemove list (lines 5–6). Conversely, if (L', m') dominates (L, m) under the same rules, the flag dom is set to true and the inner loop is exited early (lines 7–8). After all comparisons for the current tuple are complete, if dom remains false, all tuples in toRemove are removed from , and the current tuple (L, m, D) is added to (lines 10–11). This ensures that the Skyline list always contains only non-dominated tuples. The procedure repeats for each input tuple, and once all have been processed, the subroutine returns the final Skyline set (line 13), which contains the DS candidates that offer the best trade-offs between lifetime and size. DS scheduling algorithm for EAAS After EAAS identifies DSs, the scheduling phase aims to maximize the HWSN lifetime by allocating DS activation rounds according to both residual lifetime and set size. Although a fixed scheduling strategy has low computational complexity, it ignores structural differences among DSs. As a result, some sets, particularly small sets with high per-node load, may be depleted prematurely. In EAAS, each DS is characterized by its estimated lifetime L(D) and size |D|. These two quantities allow the scheduler to classify each DS into one of four cases: Case A: small size and high lifetime, referred to as small–strong. Case A: small size and high lifetime, referred to as small–strong. Case B: large size and high lifetime, referred to as large–strong. Case B: large size and high lifetime, referred to as large–strong. Case C: small size and low lifetime, referred to as small–weak. Case C: small size and low lifetime, referred to as small–weak. Case D: large size and low lifetime, referred to as large–weak. Case D: large size and low lifetime, referred to as large–weak. These cases guide how frequently and for how long each DS should be activated. Small–strong sets can save energy by allowing many nodes to sleep, but they should be activated in short bursts with cooldown periods to avoid rapid depletion. Large–strong sets can operate for longer periods because the communication load is distributed over more nodes. Small–weak and large–weak sets are used more conservatively and mainly serve as backup coverage options. To formalize this case-dependent scheduling policy, we introduce S4C-MAB: Size- and Lifetime-Aware DS Scheduling with Four-Case Policy and Multi-Armed Bandit Learning. In S4C-MAB, the four DS cases are treated as bandit arms. Rather than assigning fixed priorities to the cases, the scheduler learns time-varying case weights: This preserves the original four-case scheduling structure while allowing the scheduler to adapt its preferences according to observed scheduling outcomes. At the beginning of each scheduling round t, the scheduler recomputes the residual form of each DS after removing dead nodes. For each valid residual DS , the scheduler computes its residual size and its lifetime estimate , where is the residual energy of node u at round t, and δ is the estimated per-node drain rate. The scheduler then computes the adjusted lifetime of each DS as where is the size penalty factor, is the load penalty factor, and is the normalized per-node load of . This adjustment penalizes DSs that are too small or heavily loaded, even when their raw lifetime estimate is high. Each residual DS is then checked for validity. A DS is considered valid only if, after pruning dead nodes, it still covers the alive network. Therefore, scheduling decisions are made over valid residual DSs rather than over the original DS pool. The scheduler scores each eligible DS by combining its adjusted lifetime with the learned weight of its case: where is the strict eligible set at round t. Thus, captures the DS-specific quality, while captures the learned preference for the corresponding case. For each case , the scheduler maintains a selection count and an empirical mean reward . The learned case weight is computed using an upper-confidence-bound rule: where controls the exploration–exploitation trade-off. The empirical rewards may be warm-started using the initial heuristic priorities: Bandit-based scheduling justification and reward meaning The DS scheduling problem is naturally a sequential decision-making problem. At each round, the scheduler selects one DS for activation, but the long-term effect of this activation is not fully known before the action is taken. Activating a DS reduces the residual energy of its active nodes, may deplete bottleneck nodes, can reduce the lifetime of other candidate DSs, and may invalidate some residual DSs after pruning. Therefore, the quality of a scheduling decision can only be evaluated after observing its effect on the remaining candidate pool. A fixed policy does not use this feedback and may repeatedly activate fragile or energy-expensive DSs. Conversely, a purely greedy policy may repeatedly select the DS with the highest immediate EAAS score, which can accelerate the depletion of critical nodes and reduce future scheduling flexibility. The MAB formulation provides a lightweight online mechanism that balances these two behaviors. It exploits DS cases that have historically preserved future network utility while still exploring less frequently selected cases whose usefulness may change as the network evolves. In this work, the bandit is applied at the DS-case level rather than at the individual-DS level. This reduces the learning dimension and avoids unstable estimates when the number or validity of candidate DSs changes over time due to energy depletion. The bandit component does not replace the energy-aware EAAS score; instead, it adaptively adjusts the scheduler’s preference according to the observed ability of each DS case to preserve future network operating capability. Before activation, the aggregate utility of all valid residual DSs is computed as where denotes the set of valid residual DSs at round t. After the selected DS is activated, node energies are updated, dead nodes are pruned, residual DSs are recomputed, and DS validity is checked again. The post-activation aggregate utility is then computed as The immediate reward is defined as where \varepsilon >0 is a small constant used to avoid division by zero. A reward close to one indicates that the selected DS preserved most of the remaining valid DS utility, whereas a smaller reward indicates that the selected action significantly reduced future scheduling options. Hence, the reward measures the preservation of future DS-based network service capability rather than only the energy consumed in the current round. Let denote the case of the selected DS. After computing , the scheduler updates the bandit statistics of that case as follows: The updated empirical reward is then used in the next round to recompute the UCB-based case weight. Scheduling procedure Algorithm 3 summarizes the proposed S4C-MAB scheduling procedure. The algorithm applies case-dependent activation caps and cooldown intervals while learning case weights online from observed rewards. This enables robust DSs to be used effectively without overuse, preserves fragile DSs for critical coverage, and extends the overall network lifetime through balanced DS participation. S4C-MAB: size- and lifetime-aware DS scheduling with four-case policy and MAB. The proposed policy operates as follows: Small–strong sets (A) are preferred because they provide high lifetime with fewer active nodes, but their activation is limited by short caps and cooldowns. Small–strong sets (A) are preferred because they provide high lifetime with fewer active nodes, but their activation is limited by short caps and cooldowns. Large–strong sets (B) provide high endurance and distribute the communication load across more nodes, allowing longer permissible activation periods. Large–strong sets (B) provide high endurance and distribute the communication load across more nodes, allowing longer permissible activation periods. Small–weak sets (C) are used sparingly because they are fragile and may deplete quickly. Small–weak sets (C) are used sparingly because they are fragile and may deplete quickly. Large–weak sets (D) serve mainly as fallback options when stronger sets are unavailable. Large–weak sets (D) serve mainly as fallback options when stronger sets are unavailable. Unlike a strict scheduler, S4C-MAB does not terminate merely because the strict eligible set is empty. Instead, it applies a two-stage rescue rule. First, it relaxes temporary cooldown and consecutive-use restrictions and reconsiders all valid residual DSs with positive adjusted lifetime. If such candidates exist, it still selects according to the learned weighted score. If no relaxed candidate exists but valid residual DSs remain, the scheduler activates the valid residual DS with the largest residual lifetime. Therefore, the algorithm terminates only when no valid residual DS remains. This rescue mechanism prevents premature stopping caused by temporary activation constraints. It also aligns the execution logic with the main objective of lifetime-aware DS scheduling: maintaining feasible coverage for as long as possible while favoring DSs that offer strong lifetime, suitable size, and good long-term observed reward. Table 3 gives an illustrative example of how the four-case policy ranks valid DSs under the learned weighted score during the normal S4C-MAB selection stage. Table 3Illustrative scheduling example for the four DS cases under the S4C-MAB learned weighted score. Parameters: η =2, γ =0, thresholds and . Illustrative learned case weights at round t: , , , and .DSCasePriority415A (small-strong)1.007.5021016B (large-strong)0.9512.16148C (small-weak)0.702.804107D (large-weak)0.603.363 Illustrative scheduling example for the four DS cases under the S4C-MAB learned weighted score. Parameters: η =2, γ =0, thresholds and . Illustrative learned case weights at round t: , , , and . Using and , the sets are classified as follows: is small–strong (A), is large–strong (B), is small–weak (C), and is large–weak (D). Since γ =0 in this example, the load penalty is inactive and the adjusted lifetime reduces to . After applying the size-aware adjustment with η =2, obtains the highest weighted score, , and is therefore selected first whenever it belongs to the strict eligible set . Intuitively, combines high lifetime with larger size, which distributes the communication load across more nodes and makes it an attractive early-round choice. With activation caps and cooldowns, for example , , , and , , , , the scheduler may continue favoring whenever it remains valid and strictly eligible. If the strict eligible set becomes empty because of cooldown or consecutive-use constraints, the scheduler does not terminate. Instead, it first relaxes these temporary filters and reconsiders all valid residual DSs with positive adjusted lifetime. If such candidates exist, it still selects using the learned weighted score, thereby preserving the bandit-guided preference for stronger cases. If no positive-adjusted-lifetime relaxed candidate exists but valid residual DSs still remain, the scheduler activates the valid residual DS with the largest residual lifetime . In this sense, stronger sets such as and dominate normal operation, whereas weaker sets such as and may still appear as rescue choices when they are among the few remaining valid residual DSs. Over time, the repeated recomputation of from residual energies, the continual validity check over residual DSs, and the online updates of , , , and allow the schedule to adapt to changing network conditions without sacrificing domination coverage prematurely. As a result, the scheduler better aligns execution behavior with the lifetime objective: continue scheduling as long as any valid residual DS remains, while preferentially selecting DSs that preserve future network capacity most effectively. Clarification example To clarify the roles of EAAS and the S4C-MAB scheduler, consider the 20-node network scenario in Fig. 2, with communication radius , depletion parameter δ =0.05, and size-penalty parameter η =2. EAAS is first used as a DS-generation method; the generated disjoint dominating sets are then passed to S4C-MAB, which schedules one DS per round. Thus, EAAS constructs the DS pool, while S4C-MAB performs online DS activation. Example HWSN network with node energy values and EAAS-generated dominating sets. Nodes belonging to the same dominating set are highlighted using the same color. Step 1: EAAS candidate generation. EAAS starts with all nodes available, . At iteration r, it initializes and an empty candidate pool . If D is a valid dominating set, EAAS stores the candidate (D, |D|, L(D)) in , where . Thus, each stored candidate contains the node set, its size, and its estimated lifetime. EAAS then sorts nodes in by increasing residual energy and attempts to remove weak nodes first. For each sorted node s, it forms . If T remains a valid dominating set, the removal is accepted, D← T, and the new candidate (T, |T|, L(T)) is added to . If T is not a dominating set, node s is kept. Therefore, contains only valid candidate DSs produced during energy-aware pruning. For the first EAAS iteration, the sorted node order is: 6, 3, 15, 9, 18, 10, 2, 11, 4, 20, 5, 1, 19, 14, 7, 8, 12, 16, 17, 13. The accepted removals are 6, 3, 15, 9, 18, 10, 2, 11, 4, 20, 1, 19, 14, 7, 8, 13, while the rejected removals are 5, 12, 16, 17. Hence, the final remaining candidate is , which is still a valid dominating set and is retained in the candidate pool. Step 2: Skyline filtering and DS selection. After generating , EAAS applies skyline filtering. A candidate dominates if and , with at least one strict inequality. Therefore, only candidates that are not dominated with respect to maximizing lifetime and minimizing DS size are retained in . If the composite selector is used, EAAS selects . The selected DS is appended to the global DS pool, , and its nodes are removed from the remaining set, . This removal enforces disjointness among the generated DSs. For this example, EAAS generates , where , , , and . For example, for , the weakest node is node 5 with energy 0.518173, so . With η =2 and , the composite EAAS score is 10.363(1-2/4)=5.182. Similarly, contains node 15 with energy 0.042848, giving , which makes the weakest generated DS. Step 3: Scheduling-state construction for all four cases. The actual EAAS-generated DS pool in Table 4 contains DSs with sizes m=[4,4,5,4]. Under the median threshold , these DSs only instantiate Cases B and D because no DS satisfies . To show how the S4C-MAB case mechanism works for all four cases, we now use a controlled scheduling-state example derived from the same DS-pool concept. This controlled example is used only to illustrate the case logic; it does not change the EAAS-generation procedure. Table 4EAAS-generated disjoint dominating sets for the example network.DSNodesWeakest nodeEAAS score4Node 510.3635.1824Node 26.2273.1135Node 105.7533.4524Node 150.8570.428 EAAS-generated disjoint dominating sets for the example network. Let , , and δ =0.05. Consider four available DS states: These lifetimes correspond to weakest-node energies , , , and . Therefore, this scheduling state explicitly covers all four S4C-MAB cases. Step 4: Adjusted lifetime score. For each DS, S4C-MAB computes the adjusted lifetime score . In this controlled scheduling state, let η =2 and γ =0, so . Therefore, , , , and . Step 5: Case classification. Using and , the four DS states map to all four cases: satisfies and , so it is Case A; satisfies and , so it is Case B; satisfies and , so it is Case C; and satisfies and , so it is Case D. Step 6: MAB weighting and round-1 selection. Assume the warm-start empirical rewards are , , , and , with exploration parameter β =0.25. At round t, the UCB-style case weight is , where is the number of previous selections from case k. At t=1, all , so the exploration term is . Hence, , , , and . The final scheduling score is . Therefore, , , , and . Since all DSs are initially valid and off cooldown, , and the scheduler selects because it has the largest weighted score. Step 7: Reward computation. Before activation, the scheduler computes the pre-activation schedulable quality as . After selecting , the nodes in consume energy according to the communication model. In the full method, the selected DS communicates using the PEGASIS-chain model. The residual energies are updated, dead nodes are pruned if necessary, and the adjusted lifetime scores are recomputed as , , , and . The post-activation schedulable quality is , and the immediate reward is . This reward measures how much future DS schedulability remains after activating the selected DS. A higher value means the selected activation preserved more future scheduling capacity. Step 8: Bandit update. The reward updates only the selected case. Since belongs to Case A, the scheduler increments and updates the empirical mean reward using . The other counts and means remain unchanged. At the next round, the scheduler recomputes , and similarly for the other cases. Therefore, the MAB component rewards the selected case, not only the selected DS index. Step 9: Cooldown and repetition control. The scheduler also applies case-dependent activation caps and cooldowns. For example, with and , selecting from Case A updates the Case-A consecutive-use counter and cooldown. In the original hard-cooldown policy, a DS under cooldown is temporarily removed from the eligible set. In the adaptive soft-cooldown policy, the DS is not removed; instead, its score is penalized as , where and are penalties in (0, 1]. Thus, recent use is discouraged without forcing the scheduler to activate a much weaker DS. This example separates the two roles of the framework. EAAS-Skyline first generates the actual disjoint DS pool by storing valid candidate DSs in , filtering them using skyline dominance, selecting the best skyline candidate using the composite lifetime-size score, and removing its nodes to enforce disjointness. Then, S4C-MAB schedules among the available DSs online. The controlled scheduling-state example shows how all four cases are handled by the bandit: Case A represents small–strong DSs, Case B large–strong DSs, Case C small–weak DSs, and Case D large–weak DSs. The selected DS in round 1 is the one with the largest product of adjusted lifetime score and learned case weight. Compared with Round-Robin, which activates DSs cyclically regardless of energy state, S4C-MAB biases activation toward stronger DSs while still retaining the ability to learn case usefulness online. Intra-DS communication phase After a DS is selected for activation, data collection is performed based on chain-based forwarding within the active DS. This phase is designed to ensure that sensed data from all alive nodes are gathered efficiently and delivered to the sink. The nodes in the active DS form a communication chain for data forwarding. Let the ordered chain be denoted by where the last node in the chain, , is selected as the chain leader. In the proposed model, the chain leader is chosen as the active DS node nearest to the sink: where s denotes the sink position. Thus, the chain is constructed so that all intra-DS data forwarding progresses toward the DS node with the smallest distance to the sink. Once the chain is formed, data are aggregated and forwarded hop by hop along the chain toward the leader. Specifically, each node forwards one aggregated packet to its successor for . The leader then transmits the final aggregated packet to the sink. Based on this model, the communication cost of one activation round consists of two components: reception and aggregation of collected packets at the active DS nodes, together with hop-by-hop chain forwarding within the active DS; reception and aggregation of collected packets at the active DS nodes, together with hop-by-hop chain forwarding within the active DS; final transmission from the chain leader to the sink. final transmission from the chain leader to the sink. This communication phase complements the proposed S4C-MAB scheduler by ensuring that, once a DS is activated, it not only provides domination coverage but also serves as the temporary data-collection and forwarding backbone of the network. Since the chain leader is selected as the DS node nearest to the sink, the final long-range transmission cost is reduced, while the nearest-DS association rule limits the uplink burden on non-dominating nodes. Theoretical and computational analysis Lemma 1 The EAAS algorithm guarantees that it can always generate a valid dominating set in each iteration where candidates exist. Proof Let be the communication graph and let D⊆ V. A set D is a dominating set if In EAAS, each iteration starts from a working set D that is first checked against this domination condition. A node s∈ D is removed only if the reduced set still satisfies the domination condition. Therefore, by induction over the accepted removal operations, every candidate inserted into the candidate pool is a valid DS. Since the BNL skyline step only filters candidates according to lifetime and size and does not modify their node membership, every skyline candidate is also a valid DS. The final selected set is therefore a valid DS. After is selected, its nodes are removed from the remaining available node set: Thus, no node selected in can appear in any later selected set. Consequently, the generated sets are pairwise node-disjoint: Therefore, EAAS generates a collection of node-disjoint valid DSs, provided that a valid DS exists in the current iteration. The proof is by contradiction. A set S is a DS if and only if there exists no node s in the graph G that is not connected to a node in S (or covered by a node in S). Suppose S is a DS and assume that at least one node s in G is not connected to S. This implies that s or some of its neighbors are not connected to any node in S, which contradicts the assumption that S is a DS in a connected graph G. In EAAS, each iteration starts with covering all remaining nodes in G (for r=1, this is V). The algorithm eliminates low-energy nodes only when doing so does not violate DS coverage. Every valid intermediate set encountered during this elimination is stored in the candidate pool . The BNL_Skyline procedure then filters to retain only non-dominated candidates based on lifetime and size. Since at least one valid DS (e.g., the unmodified ) is always in unless no DS is possible, the Skyline set will contain a valid . This ensures that when candidates exist, the algorithm produces a valid DS in that iteration. Computational complexity analysis This section analyzes the computational complexity of the proposed framework, which includes the EAAS candidate-generation phase and the S4C-MAB scheduling phase, as follows. Let n=|V| be the number of sensor nodes, be the number of communication links, R be the number of EAAS iterations, be the number of remaining nodes at EAAS iteration r, be the number of candidates generated in that iteration, m be the number of generated disjoint DSs, and H be the number of online scheduling rounds. EAAS candidate generation At iteration r, EAAS sorts the current node set by residual energy, which costs . It then performs up to node-removal attempts. For each attempt, the algorithm checks whether the resulting set still satisfies the domination constraint. Using neighbor-list scanning, one domination check costs . Hence, all removal attempts in iteration r cost . For each valid candidate, EAAS stores its node set, size, and estimated lifetime. Computing lifetime values for candidates costs . The BNL skyline filter then compares candidates using two objectives, namely lifetime maximization and size minimization. Since the number of objectives is constant, the worst-case skyline-filtering cost is . Therefore, the total cost of EAAS iteration r is Summing over all R iterations gives Using the upper bounds and , this can be written as Since EAAS can add at most one new candidate after each successful node-removal attempt, . S4C-MAB scheduling During each scheduling round, the scheduler prunes dead nodes from the residual DSs, recomputes minimum residual energy and lifetime values, checks whether each residual DS still satisfies the domination constraint, computes the EAAS utility, classifies each DS into a case, computes UCB weights, and selects the eligible DS with the highest score. Because the generated DSs are node-disjoint, recomputing lifetime-related values over all DSs costs at most . However, checking the domination validity of all m residual DSs by scanning neighbor lists costs . Computing EAAS utilities, case labels, UCB weights, and selecting the best eligible DS require . After selecting a DS , the chain-based data-gathering phase constructs a chain over the active nodes. If , the greedy chain construction costs . We denote the maximum chain-construction cost over all rounds by . Therefore, the per-round cost of S4C-MAB is Over H scheduling rounds, the online scheduling cost is: Overall complexity Combining the offline and online phases, the total computational complexity of the proposed framework is: For sparse WSNs with bounded node degree, . Under this common assumption, the complexity simplifies to: Thus, both the EAAS formation phase and the S4C-MAB scheduling phase run in polynomial time with respect to the network size and the number of generated candidate DSs. Results and discussion This section is divided into two subsections. First, we evaluate the proposed EAAS method. Next, we evaluate the complete EAAS-S4C-MAB framework, which integrates EAAS with S4C-MAB scheduling. MATLAB is used to evaluate the performance of the proposed EAAS-S4C-MAB framework. The MAX-DPA technique46 is employed for the random deployment of WSN nodes. Figure 3 depicts sample graphs generated for n=100 nodes under different density levels. Examples of randomly generated graphs with n=100 nodes under different densities (expected degrees). EAAS evaluation This subsection presents the simulation results and comparative analysis of the proposed EAAS method against several baseline algorithms, namely the Minimum Dominating Set algorithm (MDP)28, the Max-Energy-First Minimum Dominating Set algorithm (MFMDP), Local Search (LSearch), Variable Depth (VDepth), and Fixed Depth (FDepth)12, as well as the original Energy-Aware Disjoint Dominating Set Algorithm (EADDSA)39. The MFMDP algorithm is a modified version of MDP in which the highest-energy node in the network is selected as the initial dominating node. Multiple deployment scenarios are considered, with network sizes of 50, 100, 150, and 200 nodes to represent different network scales. For each network size, both sparse and dense topologies are generated. The initial energy of each node is assigned a random value r ∈ (0,1], and, following12, the energy depletion rate δ is fixed at 0.05 per round. All results are obtained by averaging the outcomes over 50 independently generated WSN instances using the MAX-DPA algorithm46. The performance metric adopted in this part of the evaluation is the average network lifetime. As in EADDSA, VDepth, FDepth, LSearch, MDP, and MFMDP, the network lifetime is defined as where denotes the minimum residual energy among all sensor nodes in set , and δ represents the per-round energy consumption rate. The evaluation is conducted from two perspectives: Different network sizes: the network size varies from 50 to 200 sensor nodes, in increments of 50 nodes; that is, 50, 100, 150, and 200 nodes. Different network sizes: the network size varies from 50 to 200 sensor nodes, in increments of 50 nodes; that is, 50, 100, 150, and 200 nodes. Different network densities: the performance is evaluated under different network density values, where the density is computed as with denoting the communication radius, n the number of nodes, and and the dimensions of the deployment area. Different network densities: the performance is evaluated under different network density values, where the density is computed as with denoting the communication radius, n the number of nodes, and and the dimensions of the deployment area. Evaluation results In this test, we have 4 scenarios, each scenario undergoes 50 executions, and the obtained results are averaged for analysis. The performance of the proposed EAAS algorithm is compared with EADDSA, VDepth, FDepth, LSearch, MDP, and MFMDP across four network sizes: 200, 150, 100, and 50 nodes. Figures 4, 5, 6, 7 illustrate the network lifetime variation with changing network densities for each scenario. Comparison: varying network densities and resultant network lifetime (scenario 1: 50 nodes). Comparison: varying network densities and resultant network lifetime (scenario 2: 100 nodes). Comparison: varying network densities and network lifetime for the algorithms (scenario: 150 nodes). Comparison: varying network densities and resultant network lifetime (scenario: 200 nodes). Scenario 1: 50 nodes In the 50-node scenario (Fig. 4), MDP achieves network lifetime values ranging from 16.24 to 28.94 across all network densities. MFMDP records lower values between 9.24 and 13.34, maintaining relatively stable performance but consistently underperforming compared to MDP. LSearch attains values from 27.21 to 56.70, showing a wider range and generally outperforming both MDP and MFMDP. VDepth achieves values between 33.10 and 66.24, demonstrating higher performance than the preceding algorithms. FDepth records values from 32.68 to 57.50, performing similarly to VDepth and consistently better than MDP and MFMDP. EADDSA achieves values between 30.1137 and 77.8870, surpassing all baseline algorithms. The proposed EAAS attains the highest network lifetime values overall, ranging from 31.2091 to 79.8466, outperforming all other evaluated methods. Scenario 2: 100 nodes In the 100-node scenario (Fig. 5), MDP achieves network lifetime values ranging from 18.70 to 28.81, while MFMDP records lower values between 12.42 and 18.10, indicating a shorter lifetime and confirming that MFMDP performs worse than MDP. LSearch outperforms MFMDP but falls behind MDP. VDepth demonstrates better performance than LSearch, and FDepth performs similarly to VDepth but slightly worse. EADDSA attains values ranging from 35.4694 to 141.5906, clearly surpassing all the aforementioned baseline algorithms. EAAS achieves the highest values overall, ranging from 43.8428 to 145.039, and thus demonstrates the best performance among all evaluated algorithms in terms of network lifetime. Scenario 3: 150 nodes In the 150-node scenario (Fig. 6), the proposed EAAS algorithm consistently achieves the highest network lifetime values across all network densities, ranging from 68.4295 to 165.9834. It outperforms all other algorithms, demonstrating superior performance in terms of network longevity. EADDSA follows with values ranging from 57.5948 to 159.7112. The VDepth algorithm performs consistently well (36.06 to 45.81), surpassing all other baseline methods but remaining slightly behind EADDSA. FDepth achieves comparable performance to VDepth (35.17 to 44.14), albeit marginally lower. LSearch performs moderately well but is outperformed by both VDepth and FDepth. MDP records values between 20.59 and 29.73, performing adequately but falling behind the aforementioned methods. MFMDP exhibits the lowest network lifetime values (13.17 to 19.30), representing the weakest performance among all evaluated algorithms. Scenario 4: 200 nodes In the 200-node scenario (Fig. 7), the proposed EAAS algorithm consistently achieves the highest network lifetime values across all network densities, ranging from 96.4651 to 195.6841. It outperforms all other algorithms, demonstrating superior performance in terms of network longevity. EADDSA follows with values ranging from 81.7451 to 185.5408. The VDepth algorithm performs consistently well (38.23 to 45.07), surpassing most competing methods but still falling short of EADDSA. FDepth achieves a similar level of performance to VDepth (37.14 to 43.40), albeit slightly lower. LSearch shows moderate performance (35.10 to 43.17), outperforming MDP and MFMDP but remaining behind VDepth and FDepth. MDP performs adequately (22.17 to 29.73) yet lags behind the aforementioned methods. MFMDP records the lowest network lifetime values (13.91 to 19.30), making it the weakest performer among all evaluated algorithms. Results summary Figure 8 presents the average network lifetime for all algorithms across different network sizes (50, 100, 150, and 200 nodes) averaged over all densities. The results indicate that EAAS consistently achieves the highest average lifetime across all scenarios, followed closely by EADDSA. Both methods show a clear upward trend as the network size increases, with EAAS maintaining a consistent lead due to its balanced DS formation and size/lifetime-aware scheduling. Comparison: average network lifetime for the algorithms across all densities. Among the baseline methods, VDepth and FDepth maintain stable performance across network sizes, with only minor variations. LSearch remains relatively constant but underperforms compared to VDepth and FDepth. MDP and MFMDP achieve the lowest average lifetimes; MFMDP in particular shows only marginal gains as network size increases. Table 5 reports the relative improvement in network lifetime achieved by the proposed EAAS compared to the baseline algorithms (MDP, MFMDP, LSearch, VDepth, FDepth, and EADDSA) for various node counts and network densities. The values indicate the factor by which EAAS outperforms each reference algorithm, with positive values denoting improvement and negative values indicating cases where the baseline slightly outperformed EAAS. Table 5Times of improvement of EAAS over other algorithms.Network densityMDPMFMDPLSearchVDepthFDepthEADDSANumber of nodes =50150.852.260.11-0.09-0.080.036201.663.670.590.280.320.083252.134.500.870.490.560.080302.524.940.550.210.32-0.043351.694.840.370.180.350.025Number of nodes =100200.901.860.140.040.070.236301.973.540.950.790.850.115402.764.521.511.331.400.144503.605.712.222.052.140.065603.916.822.592.272.410.024Number of nodes =150301.803.370.680.600.640.188402.834.931.331.241.280.156503.705.851.991.831.870.082604.486.742.512.292.440.099704.377.282.702.492.620.039Number of nodes =200402.694.881.331.141.200.180503.486.151.811.721.750.183604.827.472.542.422.440.047705.317.913.042.742.850.061805.248.613.303.123.280.055 Times of improvement of EAAS over other algorithms. For the 50-node scenario, EAAS shows consistent improvement over MDP and MFMDP across all densities, with gains ranging from 0.85× to 2.52× over MDP and from 2.26× to 4.94× over MFMDP. Improvements over LSearch, VDepth, and FDepth are smaller (generally below 1×) and in a few cases (densities 15 and 30) slightly negative, indicating that the proposed method performs comparably to these heuristics in sparse low-density networks. When the network size increases to 100 nodes, the improvements become more pronounced. EAAS outperforms MDP and MFMDP by factors up to 3.91× and 6.82× respectively, and achieves consistent but smaller margins over LSearch, VDepth, and FDepth, peaking at around 2.59× over LSearch. The performance gain over the original EADDSA is modest but generally positive, confirming that the Size- and Lifetime skyline filter provides a measurable lifetime extension on top of energy-aware DS construction. In the 150-node scenario, improvement factors over MDP and MFMDP remain strong, reaching up to 4.48× and 7.28× respectively. Gains over LSearch, VDepth, and FDepth are in the range of 0.6× to 2.7×, with the largest differences observed in higher-density settings. Relative to EADDSA, improvements remain positive but small, reflecting the fact that both methods share the same DS formation phase. For the largest network (200 nodes), EAAS again achieves the highest improvement over MFMDP (8.61×) and MDP (5.31×), while also delivering notable gains over LSearch, VDepth, and FDepth (up to 3.3×). Improvements over EADDSA are again moderate but consistent, highlighting that Size- and Lifetime skyline filter yields benefits even when DS formation quality is already high. Overall, the results confirm that EAAS provides substantial performance advantages over MDP and MFMDP in all configurations, moderate but consistent gains over LSearch, VDepth, and FDepth, and measurable improvements over EADDSA, especially in larger and denser networks. Ablation study: effect of skyline-based DS selection To evaluate the contribution of skyline-based filtering in the DS formation stage, we compare the proposed EAAS with a non-skyline variant, denoted as EAA-NS. Both methods use the same DS candidate generation procedure, but differ in the selection strategy. In EAA-NS, the DS is selected directly from the full candidate pool using a scalar objective, whereas EAAS first applies skyline filtering to retain only non-dominated DS candidates before final selection. Table 6 shows that skyline filtering consistently improves the average network lifetime across all tested network sizes. For 50 nodes, the average lifetime increases from 54.81 to 56.33, corresponding to an improvement of approximately 2.78%. For 100 nodes, the lifetime increases from 87.52 to 95.21, corresponding to an improvement of about 8.79%. The gain becomes more pronounced for larger networks: for 150 nodes, the average lifetime increases from 110.16 to 120.79, corresponding to an improvement of approximately 9.64%, while for 200 nodes it increases from 133.95 to 146.06, corresponding to an improvement of about 9.03%. Table 6Average lifetime comparison with and without skyline filtering.NodesWithout skylineWith skyline50 nodes54.8106656.33284100 nodes87.5194695.21142150 nodes110.16360120.78508200 nodes133.95178146.05580 Average lifetime comparison with and without skyline filtering. These results indicate that skyline filtering becomes increasingly beneficial as the network grows. In small networks, the number of feasible DS candidates is relatively limited, so the opportunity for skyline filtering to eliminate dominated solutions is smaller. As the number of nodes increases, however, the candidate pool becomes richer and more diverse, which increases the likelihood of dominated DSs appearing. In such cases, skyline filtering becomes more effective because it removes inferior candidates and preserves only high-quality trade-offs between DS lifetime and DS size. The observed improvements can therefore be explained by the role of skyline filtering in enhancing DS selection quality. Without skyline filtering, EAA-NS selects DSs directly using a scalar objective, which may favor candidates that perform well in one objective but remain suboptimal in the overall lifetime–size trade-off. This can lead to the early selection of inefficient DSs and to less balanced energy depletion across nodes. In contrast, EAAS considers only Pareto-efficient DS candidates after skyline pruning. By eliminating candidates that are inferior in both lifetime and size, skyline filtering prevents clearly suboptimal DSs from being selected and leads to more balanced DS formation. Another important observation is that the gain from skyline filtering is stable across different network sizes and becomes stronger in medium and large networks. This suggests that skyline filtering does not merely exploit a particular parameter setting, but systematically improves the quality of the selected DSs. As network connectivity and the number of feasible DS options increase, skyline filtering ensures that these additional choices are exploited more effectively. Overall, the ablation results demonstrate that skyline-based filtering is a critical component of the proposed framework. The performance gains arise not from changes in candidate generation or parameter tuning, but from improved decision-making during DS selection. By enabling multi-objective selection without requiring manual adjustment of objective weights, skyline filtering improves DS robustness and provides a stronger foundation for subsequent scheduling. These results confirm that the effectiveness of EAAS-S4C-MAB depends not only on adaptive scheduling, but also on high-quality DS formation enabled by skyline-based selection. EAAS–S4C-MAB comparative performance analysis and discussion This section analyzes the comparative performance of the proposed EAAS–S4C-MAB framework. The discussion begins with a sensitivity study of the S4C-MAB control parameters, and , followed by an ablation study and baseline comparison to isolate the contribution of each framework component. The proposed method is then compared with five representative state-of-the-art baselines: T-LEACH 32, EEM-TDMA 33, T-LEACHSAS 34, CDS-LEACH 35, and Enhanced LEACH 36. Finally, statistical significance analysis is reported to verify the reliability of the observed performance differences. Sensitivity study of S4C-MAB parameters A sensitivity study was conducted to tune the two main control parameters of the S4C-MAB scheduler: the UCB exploration coefficient and the DS size-penalty coefficient . The tuning objective was to prioritize first-node-death delay (FND), while maintaining a high accumulated dominating-set lifetime (AccLifetime). FND denotes the first-node-death round, i.e., the first round in which any sensor node exhausts its residual energy. AccLifetime denotes the accumulated residual DS lifetime/capacity over the simulation horizon and is used to assess how well the scheduler preserves residual DS capacity. The study used a high-connectivity scenario with n=150, , sink position (50,150), and a maximum simulation horizon of 6000 rounds. This scenario was selected because it provides a sufficiently large DS pool, allowing the MAB scheduler to exhibit meaningful exploration–exploitation behavior. A two-dimensional grid search was performed using and . Each parameter pair was evaluated over 30 independent scenarios using the same random seeds to support paired comparison. The reference configuration was and . Candidate settings were ranked by prioritizing higher FND, then higher AccLifetime, followed by lower average energy consumption and lower energy-consumption variability. Table 7 summarizes the full 2D parameter grid evaluated for and . The highest FND was obtained with and , achieving 4019.07± 850.30 rounds. Compared with the reference setting , which achieved 3981.80± 816.99 rounds, the selected configuration increased the mean FND by 37.27 rounds, corresponding to approximately 0.94%. However, AccLifetime decreased from to , indicating a small trade-off between delaying first-node death and maximizing accumulated lifetime. The relatively large FND standard deviations are expected because FND is an event-based metric that is sensitive to random topology generation, initial node-energy distribution, and the location of early-depleted nodes. Therefore, all configurations were evaluated using identical scenario seeds, and paired comparisons were used to reduce the effect of scenario-to-scenario variability. Table 72D sensitivity study of and . Each cell reports FND/AccLifetime as mean±SD over 30 runs.0.00.51.02.00.054012.20± 849.87 3.025± 0.816 3.021± 0.8144012.40± 855.61 3.026± 0.8084016.17± 860.34 3.017± 0.8100.103992.53± 852.82 3.052± 0.8123988.73± 846.29 3.054± 0.8153998.70± 844.45 3.045± 0.8254002.17± 849.37 3.042± 0.8220.253963.87± 825.81 3.089± 0.8453966.87± 823.36 3.087± 0.8443957.70± 824.82 3.081± 0.8463981.80± 816.99 3.058± 0.8470.503977.53± 833.06 3.089± 0.8483973.70± 840.03 3.088± 0.8443978.60± 841.21 3.083± 0.8443982.03± 819.76 3.071± 0.8511.003972.67± 825.04 3.091± 0.8513963.37± 829.49 3978.43± 857.29 3.073± 0.8283968.97± 836.59 3.081± 0.842 indicates the selected FND-prioritized configuration. The bold AccLifetime value indicates the highest AccLifetime. 2D sensitivity study of and . Each cell reports FND/AccLifetime as mean±SD over 30 runs. 4012.20± 849.87 3.025± 0.816 3.021± 0.814 4012.40± 855.61 3.026± 0.808 4016.17± 860.34 3.017± 0.810 3992.53± 852.82 3.052± 0.812 3988.73± 846.29 3.054± 0.815 3998.70± 844.45 3.045± 0.825 4002.17± 849.37 3.042± 0.822 3963.87± 825.81 3.089± 0.845 3966.87± 823.36 3.087± 0.844 3957.70± 824.82 3.081± 0.846 3981.80± 816.99 3.058± 0.847 3977.53± 833.06 3.089± 0.848 3973.70± 840.03 3.088± 0.844 3978.60± 841.21 3.083± 0.844 3982.03± 819.76 3.071± 0.851 3972.67± 825.04 3.091± 0.851 3963.37± 829.49 3978.43± 857.29 3.073± 0.828 3968.97± 836.59 3.081± 0.842 indicates the selected FND-prioritized configuration. The bold AccLifetime value indicates the highest AccLifetime. Figure 9 visualizes the 2D sensitivity results for and . Figure 9a shows that the highest FND is obtained at and , confirming the FND-prioritized configuration selected from the grid search. In contrast, Fig. 9b shows that AccLifetime increases for larger values, with the maximum observed at and . This visual trend confirms the trade-off between delaying first-node death and preserving accumulated residual DS lifetime: lower favors FND, whereas larger favors AccLifetime. The square marker denotes the selected FND-prioritized setting, while the circle marker denotes the maximum value in each panel. Heat-map visualization of the 2D sensitivity study for and : (a) FND and (b) AccLifetime. The square marks the selected FND-prioritized configuration (0.05,0.5), while the circle marks the maximum value in each panel. The paired comparison against the reference setting showed a positive but modest FND gain of 37.27 rounds, with a 95% confidence interval of [-17.95,92.48]. Since this interval includes zero, the improvement should not be interpreted as statistically decisive. Rather, the sensitivity study indicates that S4C-MAB is relatively stable over the tested parameter range, with mainly controlling the trade-off between FND and AccLifetime. Lower values favored FND by encouraging more conservative exploitation of reliable DSs, whereas larger values increased AccLifetime by encouraging broader exploration of the DS pool. The effect of was less pronounced, although provided the best FND-prioritized balance. Based on this study, and were selected as the tuned S4C-MAB parameters for the final comparison against the baselines. This setting should be interpreted as FND-prioritized rather than globally optimal for all lifetime metrics. Ablation study and baseline comparison An ablation study was conducted to evaluate the contribution of the main components in the proposed EAAS–S4C-MAB framework. The tuned configuration selected from the sensitivity study, and , was used as the full method. In this study, S4C control refers specifically to the cooldown/cap control mechanism, which regulates repeated DS activation through cooldown restrictions and activation caps. Six alternative variants were considered: S4C-MAB without EAAS scoring, EAAS–S4C without MAB/UCB, EAAS–MAB without cooldown/cap control, EAAS–S4C-MAB without the size penalty, a round-robin (RR) baseline, and a random-selection baseline. The RR and random baselines use the same feasible DS pool as the proposed method, but do not use EAAS scoring, MAB/UCB learning, cooldown/cap control, or the DS-size penalty. Two connectivity regimes were evaluated. The low-connectivity regime uses n=100 and , representing a sparse topology with limited DS diversity. The high-connectivity regime uses n=150 and , representing a denser topology with a richer DS pool. In both regimes, the sink was placed at (50,150), the maximum simulation horizon was 6000 rounds, and 30 independent scenarios were generated using identical seeds for all variants. The feasible DS pool was fixed across variants for each scenario seed to ensure a fair comparison. Therefore, DSCount is treated as a regime-level contextual metric rather than a method-specific performance metric. The low-connectivity regime produced 2.43± 0.86 feasible DSs on average, whereas the high-connectivity regime produced 8.47± 1.80. This difference is important because DSCount determines the number of available scheduling arms for the MAB policy. The main reported metrics are FND and AccLifetime. FND measures the first-node-death round and is used as the primary lifetime metric. AccLifetime measures the accumulated residual DS lifetime/capacity and is used as a secondary metric to interpret lifetime-distribution effects. Table 8 summarizes the ablation and baseline results under both connectivity regimes. Table 8Ablation and baseline results under low- and high-connectivity regimes. FND and AccLifetime are reported as mean±SD over 30 runs. AccLifetime is scaled by .RegimeVariantFNDAccLifetime LowFull EAAS-S4C-MAB897.77± 331.770.599± 0.307LowS4C-MAB w/o EAAS scoring794.00± 246.150.608± 0.313LowEAAS-S4C w/o MAB/UCB898.37± 332.020.598± 0.307LowEAAS-MAB w/o cooldown/cap control0.270± 0.108LowEAAS-S4C-MAB w/o size penalty897.60± 331.570.599± 0.307LowRR baseline791.73± 236.150.673± 0.392LowRandom baseline789.70± 232.080.673± 0.393HighFull EAAS-S4C-MAB4019.07± 850.303.021± 0.814HighS4C-MAB w/o EAAS scoring2404.10± 625.883.967± 1.098HighEAAS-S4C w/o MAB/UCB4002.57± 840.253.039± 0.832HighEAAS-MAB w/o cooldown/cap control1.278± 0.273HighEAAS-S4C-MAB w/o size penalty4012.20± 849.873.025± 0.816HighRR baseline1976.00± 678.089.317± 3.369HighRandom baseline1975.53± 700.639.302± 3.361 Ablation and baseline results under low- and high-connectivity regimes. FND and AccLifetime are reported as mean±SD over 30 runs. AccLifetime is scaled by . In the high-connectivity regime, the full EAAS–S4C-MAB method achieved an FND of 4019.07± 850.30, substantially outperforming RR and random selection, which achieved 1976.00± 678.08 and 1975.53± 700.63, respectively. This corresponds to an FND gain of approximately 2043 rounds over both baselines. A similar trend is observed in the low-connectivity regime, where the full method achieved 897.77± 331.77, compared with 791.73± 236.15 for RR and 789.70± 232.08 for random selection. These results show that the proposed guided scheduling strategy is more effective than unguided or fixed-rotation DS selection in delaying first-node death. The most significant component effect is observed for the EAAS scoring mechanism. Removing EAAS scoring reduced FND from 4019.07 to 2404.10 in the high-connectivity regime, corresponding to a decrease of approximately 40.18%. In the low-connectivity regime, removing EAAS scoring reduced FND from 897.77 to 794.00, corresponding to a decrease of approximately 11.56%. The paired comparison confirms this effect: the FND difference relative to the full method was -1614.97 rounds in the high-connectivity regime and -103.77 rounds in the low-connectivity regime. Therefore, EAAS scoring is the most influential component for improving early-node survival, particularly when the DS pool is sufficiently rich. The contribution of MAB/UCB was more limited under the tuned setting. In the ”EAAS–S4C w/o MAB/UCB” variant, the adaptive MAB/UCB learning term was disabled while EAAS scoring, cooldown/cap control, and the size penalty were retained. This variant produced nearly identical FND in the low-connectivity regime and only a small reduction in the high-connectivity regime. This behavior is partly explained by DSCount. In the low-connectivity regime, only 2.43± 0.86 feasible DSs were available on average, with a range of 1–4 DSs, leaving few scheduling arms for the MAB policy to exploit. In contrast, the high-connectivity regime produced 8.47± 1.80 feasible DSs on average, with a range of 5–11 DSs, providing greater opportunity for adaptive selection. However, because the tuned value is conservative, the observed UCB contribution remains moderate even in the high-connectivity case. Therefore, MAB/UCB is expected to become more beneficial as DSCount increases, but its effect is constrained when the feasible DS pool is small or when exploration is intentionally limited. The cooldown/cap control mechanism reveals an important trade-off. Disabling cooldown/cap control produced the highest FND in both regimes, increasing FND from 897.77 to 1102.00 in the low-connectivity regime and from 4019.07 to 4204.40 in the high-connectivity regime. However, this improvement came with a substantial reduction in AccLifetime. In the high-connectivity regime, AccLifetime decreased from to , while in the low-connectivity regime it decreased from to . This indicates that cooldown/cap control does not maximize FND alone; instead, it regularizes repeated DS activation and preserves accumulated residual DS lifetime/capacity. Thus, the full EAAS–S4C-MAB configuration should be interpreted as a balanced scheduler rather than an absolute FND-maximizing variant. The DS-size penalty had only a marginal effect under the tested settings. Removing the size penalty changed FND by less than seven rounds in the high-connectivity regime and less than one round in the low-connectivity regime. This confirms the sensitivity-study observation that has a weaker effect than EAAS scoring and cooldown/cap control. The size penalty is therefore retained in the full method as a regularization term, but it is not the dominant driver of performance. An additional observation is that Avg. E/R and the standard deviation of per-round energy consumption were identical across variants within each connectivity regime. In the low-connectivity regime, Avg. E/R was 0.007605± 0.000356, while in the high-connectivity regime it was 0.005106± 0.000213. This indicates that the ablation variants mainly affect the distribution of energy depletion across nodes and DSs rather than the average per-round communication cost. Therefore, the ablation discussion focuses on FND, AccLifetime, and DSCount rather than Avg. E/R. Overall, the ablation study supports three main conclusions. First, EAAS scoring is essential for delaying first-node death, especially in high-connectivity networks with richer DS diversity. Second, the proposed full EAAS–S4C-MAB method substantially outperforms RR and random selection in FND under both connectivity regimes. Third, cooldown/cap control introduces a deliberate trade-off: disabling it improves FND but substantially reduces accumulated residual DS lifetime. Therefore, the full EAAS–S4C-MAB framework provides a balanced scheduling strategy that improves early-node survival relative to conventional baselines while preserving residual DS lifetime more effectively than the no-cooldown/cap variant. The results also indicate that EAAS–S4C-MAB can be tuned according to the target lifetime objective. The FND-prioritized configuration uses and , which achieved the highest FND among the tested full-method configurations. In contrast, larger values favored AccLifetime, indicating that the exploration coefficient controls a trade-off between delaying first-node death and preserving accumulated residual DS lifetime. Therefore, the selected configuration should be interpreted as FND-prioritized rather than globally optimal for all lifetime metrics. Comparison with state-of-the-art methods The proposed framework, EAAS-S4C-MAB, was evaluated against five representative baselines: T-LEACH32, EEM-TDMA33, T-LEACHSAS34, CDS-LEACH35, and Enhanced LEACH36. These protocols represent different lifetime-extension strategies in WSNs, including threshold-based cluster-head retention, sink-assisted TDMA scheduling, sleep–awake duty cycling, connectivity-aware cluster-head selection, and residual-energy-aware cluster-head election. However, unlike the proposed method, they remain fundamentally cluster-based rather than DS-centric. Accordingly, only the network-level metrics FND and AvgEconsPerRound were used for fair cross-protocol comparison, since these metrics are directly comparable across DS-based and cluster-based methods. To ensure a fair and consistent evaluation, all baseline protocols and the proposed EAAS-S4C-MAB framework were implemented under the same simulation environment. Specifically, all methods used the same first-order radio energy model, the same node deployment area, the same randomly generated network instances, the same sink locations, the same initial energy distribution, and the same stopping conditions. Two connectivity regimes were considered over a fixed deployment region with n=100 nodes. These regimes correspond to communication radii of 18 m and 25 m. For each regime, 30 independent network instances were generated using different random seeds, and the same set of instances was reused for every compared protocol. The sink was placed at two representative locations, namely the corner position (0, 0) and the center position (50, 50). Initial node energies were heterogeneous and independently sampled from , giving an expected total initial network energy of 30 J across the 100 nodes. For all compared methods, the first-order radio model was adopted using identical transmission, reception, amplification, and data-aggregation energy parameters. No protocol was assigned protocol-specific radio parameters or a different initial energy budget. For protocols originally designed for homogeneous WSNs, the same heterogeneous initial energy values were assigned to their nodes, and their original cluster-head selection or scheduling mechanisms were applied under this common energy setting. All simulations were executed using the same stopping rule: a run was continued until the network could no longer sustain valid data collection according to the adopted lifetime criterion. The FND metric was recorded as the first round in which any node exhausted its energy, while AvgEconsPerRound was computed using the same per-round energy accounting procedure for all protocols. Therefore, the reported differences arise from the protocol mechanisms themselves rather than from differences in radio assumptions, deployment conditions, energy initialization, or termination criteria. For fair evaluation, the first-order radio model was adopted for all compared methods. For a packet of size k bits transmitted over a distance d, the transmission energy is given by: where is the electronics energy, and and are the free-space and multipath amplifier parameters, respectively. The threshold distance is given by If a node receives and aggregates q packets, the corresponding energy cost is where denotes the data aggregation energy per bit. Table 9 summarizes the main simulation parameters. Table 9Simulation settings.ParameterValueDeployment areaNumber of nodes100Sink position(0, 0), (50, 50)Initial node energyExpected total energyCommunication radius18 m, 25 m50 nJ/bit10 pJ/bit/m0.0013 pJ/bit/m m5 nJ/bit/signalPacket size4000 bits Simulation settings. To ensure reproducibility, all experiments were conducted using fixed random seeds. For each density regime, 30 independent Monte Carlo scenarios were generated using seeds 1,… ,30. For a given seed, the same scenario file was used for all compared methods. This setting enables paired comparisons and reduces variability caused by different topology realizations. The MATLAB random number generator was initialized using rng(seed,’twister’) before scenario generation and simulation execution. Each simulation was executed until either the maximum number of rounds was reached or no valid residual dominating set remained after node-energy depletion and pruning. The maximum number of rounds was set to . For every run, the stopping reason, the number of completed rounds, residual energy, remaining lifetime, node-death statistics, and energy-consumption metrics were recorded. For the EAAS-S4C-MAB framework, the per-node energy drain was estimated online using an exponentially weighted moving average (EWMA). For each node u, the drain estimate was updated whenever the node incurred positive observed energy consumption in a round: where is the measured energy consumed by node u during round t, and is the smoothing factor. Accordingly, the instantaneous lifetime of a dominating set was defined as which represents the expected number of future activations before the first node in depletes its energy. This formulation avoids imposing a fixed per-DS energy-consumption assumption and instead adapts naturally to evolving routing and communication costs. The strongest directly comparable evidence is provided by the FND results in Table 10 and Fig. 10. As shown in Fig. 10, EAAS-S4C-MAB achieves the highest FND in all four deployment scenarios, namely lower-density and higher-density networks under sink locations (0, 0) and (50, 50). The strongest baseline is consistently T-LEACHSAS34, yet the proposed framework still maintains a large margin. At lower density with sink (0, 0), EAAS-S4C-MAB achieves an FND of 959.20, compared with 349.47 for T-LEACHSAS, corresponding to an improvement of approximately 174.5%. At lower density with sink (50, 50), the proposed method reaches 1038.57 versus 667.00, corresponding to an improvement of about 55.7%. The advantage becomes much larger at higher density: at sink (0, 0), the proposed framework reaches 2376.37 against 366.53, corresponding to a gain of about 548.3%, and at sink (50, 50) it reaches 2472.65 against 629.40, corresponding to a gain of about 292.9%. The strong visual separation in Fig. 10 confirms that the proposed framework does not merely improve stability incrementally; rather, it shifts the network into a substantially longer stable operating regime. Table 10FND comparison across protocols and deployment scenarios. Higher values are better.ProtocolLower density,Lower density,Higher density,Higher density,sink (0, 0)sink (50, 50)sink (0, 0)sink (50, 50)T-LEACH116.30192.83120.73175.23T-LEACHSAS349.47667.00366.53629.40Enhanced LEACH146.72334.17179.81317.37CDS-LEACH81.00122.52100.80194.68EEM-TDMA71.3175.2871.7377.11EAAS-S4C-MAB959.201038.572376.372472.65 FND comparison across protocols and deployment scenarios. Higher values are better. Comparison of FND results under four deployment conditions. The energy results in Table 11 and Fig. 11 reinforce the same conclusion. Figure 11 shows that EAAS-S4C-MAB consistently yields the lowest AvgEconsPerRound among all compared protocols. Relative to the strongest baseline, T-LEACHSAS34, the proposed framework reduces average per-round energy consumption by approximately 21.5% at lower density with sink (0, 0), 12.0% at lower density with sink (50, 50), 51.7% at higher density with sink (0, 0), and 46.9% at higher density with sink (50, 50). This is particularly important because it shows that the lifetime advantage is not obtained by increasing the energy spent per round. On the contrary, the proposed framework achieves both longer stability and lower average dissipation, which indicates better long-term energy balancing and more efficient use of the available network resources. Table 11AvgEconsPerRound comparison across protocols and deployment scenarios. Lower values are better.ProtocolLower density,Lower density,Higher density,Higher density,sink (0, 0)sink (50, 50)sink (0, 0)sink (50, 50)T-LEACH0.018500.016380.020170.01805T-LEACHSAS0.009450.008440.009620.00875Enhanced LEACH0.010170.010150.011950.01005CDS-LEACH0.016410.017000.016430.01863EEM-TDMA0.122390.115390.123310.11642EAAS-S4C-MAB0.007420.007430.004650.00465 AvgEconsPerRound comparison across protocols and deployment scenarios. Lower values are better. Comparison of average energy consumption per round (AvgEconsPerRound) across all protocols under four deployment conditions. Lower values indicate better energy efficiency. The energy results in Table 11 and Fig. 11 reinforce the same conclusion. Table ? reports the average per-round energy consumption as mean ± standard deviation, while Fig. 11 visualizes the corresponding mean values. Across all deployment scenarios, EAAS-S4C-MAB consistently yields the lowest AvgEconsPerRound among all compared protocols. Relative to the strongest baseline, T-LEACHSAS 34, the proposed framework reduces average per-round energy consumption by approximately 21.5% at lower density with sink (0, 0), 12.0% at lower density with sink (50, 50), 51.7% at higher density with sink (0, 0), and 46.9% at higher density with sink (50, 50). The associated standard deviations are small, indicating that the observed energy advantage is consistent across runs rather than being caused by isolated trials. This is particularly important because it shows that the lifetime advantage is not obtained by increasing the energy spent per round. On the contrary, the proposed framework achieves both longer stability and lower average dissipation, which indicates better long-term energy balancing and more efficient use of the available network resources. The ranking observed in the figures is also technically meaningful. T-LEACHSAS34 is the strongest baseline because it combines threshold-based cluster-head retention with sleep–awake scheduling, thereby reducing both cluster-head reselection overhead and idle listening. Enhanced LEACH36 generally follows because of its residual-energy-aware cluster-head election mechanism. T-LEACH32 performs moderately well by reducing the cost of repeated cluster-head changes. CDS-LEACH35 improves connectivity, but its CDS role remains tied to cluster-head selection rather than explicit DS scheduling. EEM-TDMA33 focuses mainly on centralized TDMA coordination and therefore cannot exploit structural redundancy in the same way as the proposed framework. The superiority of EAAS-S4C-MAB is thus the direct consequence of operating at the DS-management level rather than only at the cluster-head or TDMA-scheduling level. An especially important systems-level result is the low sink-position sensitivity of the proposed framework. This effect is visualized in Fig. 13 for lower density and Fig. 14 for higher density, and numerically summarized in Table 12. At lower density, most baselines show substantial percentage gains in FND when the sink is moved from (0, 0) to (50, 50); for example, Enhanced LEACH improves by about 127.8%, T-LEACHSAS by about 90.9%, T-LEACH by about 65.8%, and CDS-LEACH by about 51.3%. By contrast, EAAS-S4C-MAB changes by only about 8.27%. A similar pattern appears at higher density: CDS-LEACH changes by about 93.1%, Enhanced LEACH by about 76.5%, T-LEACHSAS by about 71.7%, and T-LEACH by about 45.1%, whereas the proposed framework changes by only about 4.05%. The two sink-effect figures therefore provide strong visual evidence that the baselines are much more geometry-sensitive than the proposed method. Percentage change in FND caused by moving the sink from (0, 0) to (50, 50) under lower-density deployment. Lower sensitivity indicates stronger robustness to sink placement. Percentage change in FND caused by moving the sink from (0, 0) to (50, 50) under higher-density deployment. Lower sensitivity indicates stronger robustness to sink placement. Table 12Sensitivity to sink relocation from (0, 0) to (50, 50). Positive indicates improved stability after sink relocation; negative indicates reduced energy consumption after sink relocation.Protocol (%), (%), (%), (%),lower densityhigher densitylower densityhigher densityT-LEACH65.8145.14-11.45-10.49T-LEACHSAS90.8671.72-10.70-9.04Enhanced LEACH127.7676.51-0.15-15.88CDS-LEACH51.2693.133.6013.42EEM-TDMA5.577.49-5.72-5.59EAAS-S4C-MAB8.274.050.080.03 Sensitivity to sink relocation from (0, 0) to (50, 50). Positive indicates improved stability after sink relocation; negative indicates reduced energy consumption after sink relocation. The density effect is equally revealing. By comparing the lower-density and higher-density results in Fig. 10, it becomes clear that EAAS-S4C-MAB benefits from increased node density far more than the baseline protocols. Specifically, the proposed framework improves from 959.20 to 2376.37 in FND when density increases at sink (0, 0), corresponding to approximately 147.7% growth, and from 1038.57 to 2472.65 at sink (50, 50), corresponding to approximately 138.1% growth. At the same time, Fig. 11 shows that its AvgEconsPerRound drops from about 0.00742–0.00743 to about 0.00465, which is roughly a 37.4% reduction. This is a highly favorable scaling pattern: higher density produces both longer stability and lower average energy dissipation. Such behavior strongly suggests that the proposed method is able to transform additional nodes into useful structural redundancy and exploit them adaptively through skyline-based selection and MAB-driven scheduling. Overall, the figures and tables consistently support the same conclusion. Figure 10 shows that EAAS-S4C-MAB achieves by far the best stability period; Fig. 11 shows that it also yields the lowest average energy dissipation; Fig. 12 confirms that it dominates the joint performance space across all scenarios; and Fig. 13 together with Fig. 14 demonstrate that it is much less sensitive to sink relocation than the baselines. Therefore, the main contribution of EAAS-S4C-MAB is not simply that it reduces energy consumption relative to improved LEACH variants, but that it converts network redundancy into balanced, long-lived, and deployment-robust operation through DS-aware formation and scheduling. Normalized composite heatmap based on FND and AvgEconsPerRound, summarizing the relative joint performance of all protocols across deployment conditions. Results summary The observed performance differences across protocols can be directly explained by their underlying design principles. Among the baselines, T-LEACHSAS34 consistently achieves the best performance because it combines threshold-based cluster-head retention with sleep–awake scheduling. This dual mechanism reduces both cluster-head reselection overhead and idle listening, leading to lower energy dissipation and delayed node death. Enhanced-LEACH36 performs next best due to its residual-energy-aware cluster-head election, which improves energy distribution but still relies on periodic clustering and does not exploit structural redundancy. T-LEACH32 provides moderate gains by avoiding frequent cluster-head changes, but its static threshold mechanism cannot adapt to dynamic energy variations across the network. CDS-LEACH35 improves connectivity by incorporating connected dominating set (CDS) membership into cluster-head selection; however, this CDS usage remains auxiliary and does not support explicit DS scheduling or lifetime-aware activation. As a result, its performance is limited compared to DS-based approaches. EEM-TDMA33 exhibits the weakest performance because its primary focus is on reducing control overhead through centralized TDMA scheduling, without addressing energy balancing or structural redundancy. Consequently, energy consumption remains high, and node depletion occurs early. The sensitivity to sink relocation from (0, 0) to (50, 50) can be explained by the communication structure of each protocol. In T-LEACH, T-LEACHSAS, Enhanced LEACH, and CDS-LEACH, data delivery depends strongly on the cluster-head layer, so moving the sink toward the center reduces the average cluster-head-to-sink transmission distance and significantly improves FND. T-LEACHSAS benefits the most among these methods because the shorter sink distance complements its threshold-based cluster-head retention and sleep–awake energy saving. Enhanced LEACH also shows good improvement because, although it improves cluster-head election fairness, it still remains highly dependent on sink-facing cluster-head transmissions. CDS-LEACH benefits from sink relocation because its connectivity-aware cluster-head placement becomes more effective when the final transmission distance to the sink is reduced. By contrast, EEM-TDMA shows only limited sensitivity because its primary improvement lies in reducing scheduling overhead through sink-centralized TDMA construction rather than in changing the structural routing backbone. The proposed EAAS-S4C-MAB framework exhibits the lowest sensitivity because its lifetime performance is driven mainly by skyline-based DS formation, adaptive DS scheduling, and balanced use of alternative DS backbones, while the final sink transmission is handled by the active DS node nearest to the sink. Consequently, its performance depends far less on sink geometry alone than that of the other baselines. The superior performance of EAAS-S4C-MAB stems from its DS-centric operation, where multiple disjoint DS candidates are generated, filtered via skyline dominance, and scheduled adaptively using a multi-armed bandit policy. This enables balanced utilization of network resources, prevents overuse of weak structures, and continuously adapts to changing energy conditions. In contrast, all baseline protocols operate at the cluster-head level and therefore lack explicit mechanisms for managing multiple alternative backbones. Statistical significance analysis To validate the robustness of the proposed EAAS-S4C-MAB protocol, Welch’s two-sample t-tests were conducted against five baseline protocols: EEMTDMA, CDS-LEACH, T-LEACH, ENHANCED-LEACH, and T-LEACHSAS. The tests were performed across all network configurations, including two node-density settings, two sink-location scenarios, and two evaluation metrics: first node death (FND) and average energy consumption per round (AvgEconsPerRound). Each comparison was based on 30 independent simulation runs for EAAS-S4C-MAB and 30 independent simulation runs for the corresponding baseline protocol. In total, 40 pairwise statistical comparisons were conducted: A Bonferroni correction was applied to control the family-wise error rate across the 40 comparisons. Therefore, the corrected significance threshold was For each comparison, the mean, standard deviation, and 95% confidence interval were computed. Welch’s two-sample t-test was used because it does not assume equal variances between EAAS-S4C-MAB and the baseline protocols. The resulting Welch t-statistic, degrees of freedom, and two-sided p-value were then reported for each comparison. The statistical results show that EAAS-S4C-MAB outperforms all baseline protocols in every tested scenario. All 40 Welch t-tests remained statistically significant after Bonferroni correction, providing strong evidence that the observed improvements are not due to random variation. A summary of statistical improvements achieved by EAAS-S4C-MAB is presented in Table 13. For the FND metric, higher values indicate better network lifetime. EAAS-S4C-MAB achieved a mean FND improvement of 1109.15% across all baseline comparisons, with improvements ranging from 55.71% to 3212.83%. The largest improvement was observed against EEMTDMA under the higher-density scenario with the sink located at (0, 0), where EAAS-S4C-MAB improved FND by 3212.83%. Even the smallest FND gain, recorded against T-LEACHSAS under the lower-density scenario with the sink located at (50, 50), remained statistically significant after Bonferroni correction. Table 13Summary of statistical improvements achieved by EAAS-S4C-MAB.MetricComparisonsEAAS betterSignificantImprovement rangeFND2020/2020/2055.71%-3212.83%AvgEconsPerRound2020/2020/2012.00%-96.23% reduction Summary of statistical improvements achieved by EAAS-S4C-MAB. For AvgEconsPerRound, lower values indicate better energy efficiency. EAAS-S4C-MAB reduced energy consumption by an average of 60.21% across all comparisons. The reductions ranged from 12.00% to 96.23%. The strongest energy saving was achieved against EEMTDMA under the higher-density scenario with the sink located at (0, 0), where EAAS-S4C-MAB reduced average energy consumption per round by 96.23%. Table 14 reports the mean, standard deviation, 95% confidence interval, Welch t-statistic, degrees of freedom, and two-sided p-value for all 40 pairwise comparisons. Table 14Detailed statistical comparison between EAAS-S4C-MAB and baseline protocols across all density and sink-location configurations. Values are reported as mean ± standard deviation with 95% confidence intervals. The p-values correspond to two-sided Welch t-tests. AvgEcons denotes AvgEconsPerRound.MetricDensitySinkBaselineEAAS mean ± SDEAAS 95% CIBaseline mean ± SDBaseline 95% CI tdf p-valueFNDLower(0, 0)CDS-LEACH959.20 ± 357.42[825.74, 1092.66]81.00 ± 18.88[73.95,88.05]13.4429.16EEMTDMA959.20 ± 357.42[825.74, 1092.66]71.31 ± 7.87[68.37,74.25]13.6029.03ENHANCED-LEACH959.20 ± 357.42[825.74,1092.66]146.72 ± 33.35[134.27,159.17]12.4029.50T-LEACH959.20 ± 357.42[825.74, 1092.66]116.30 ± 41.01[100.99,131.61]12.8329.76T-LEACHSAS959.20 ± 357.42[825.74, 1092.66]349.47 ± 108.63[308.90,390.03]8.9434.31(50, 50)CDS-LEACH1038.57 ± 382.28[895.82, 1181.31]122.52 ± 27.31[112.32,132.72]13.0929.30EEMTDMA1038.57 ± 382.28[895.82, 1181.31]75.28 ± 8.67[72.05,78.52]13.8029.03ENHANCED-LEACH1038.57 ± 382.28[895.82,1181.31]334.17 ± 16.62[327.96,340.37]10.0829.11T-LEACH1038.57 ± 382.28[895.82,1181.31]192.83 ± 48.03[174.90,210.77]12.0229.92T-LEACHSAS1038.57 ± 382.28[895.82,1181.31]667.00 ± 151.92[610.27,723.73]4.9537.94Higher(0, 0)CDS-LEACH2376.37 ± 572.94[2162.43,2590.31]100.80 ± 27.41[90.57,111.03]21.7329.13EEMTDMA2376.37 ± 572.94[2162.43,2590.31]71.73 ± 7.14[69.07,74.40]22.0329.01ENHANCED-LEACH2376.37 ± 572.94[2162.43,2590.31]179.81 ± 54.18[159.58,200.04]20.9129.52T-LEACH2376.37 ± 572.94[2162.43,2590.31]120.73 ± 35.77[107.38,134.09]21.5229.23T-LEACHSAS2376.37 ± 572.94[2162.43,2590.31]366.53 ± 112.06[324.69,408.38]18.8631.22(50, 50)CDS-LEACH2472.65 ± 489.41[2289.90,2655.40]194.68 ± 51.00[175.64,213.72]25.3629.63EEMTDMA2472.65 ± 489.41[2289.90,2655.40]77.11 ± 7.93[74.14,80.07]26.8129.02ENHANCED-LEACH2472.65 ± 489.41[2289.90,2655.40]317.37 ± 19.54[310.08,324.67]24.1029.09T-LEACH2472.65 ± 489.41[2289.90,2655.40]175.23 ± 45.27[158.33,192.14]25.6029.50T-LEACHSAS2472.65 ± 489.41[2289.90,2655.40]629.40 ± 140.56[576.91,681.89]19.8333.75AvgEconsLower(0, 0)CDS-LEACH0.007419 ±0.000386[0.007275,0.007563]0.016409 ±0.002307[0.015548,0.017271]-21.0530.62EEMTDMA0.007419 ±0.000386[0.007275,0.007563]0.122389 ±0.005248[0.120429,0.124348]-119.6629.31ENHANCED-LEACH0.007419 ±0.000386[0.007275,0.007563]0.010165 ±0.001019[0.009784,0.010545]-13.8037.14T-LEACH0.007419 ±0.000386[0.007275,0.007563]0.018501 ±0.001774[0.017838,0.019163]-33.4331.74T-LEACHSAS0.007419 ±0.000386[0.007275,0.007563]0.009448 ±0.000129[0.009400,0.009496]-27.3235.39(50, 50)CDS-LEACH0.007425 ±0.000423[0.007267,0.007583]0.017000 ±0.002202[0.016177,0.017822]-23.3831.14EEMTDMA0.007425 ±0.000423[0.007267,0.007583]0.115391 ±0.005124[0.113478,0.117305]-115.0229.40ENHANCED-LEACH0.007425 ±0.000423[0.007267,0.007583]0.010149 ±0.000922[0.009805,0.010494]-14.7140.69T-LEACH0.007425 ±0.000423[0.007267, 0.007583]0.016383 ±0.001323[0.015889,0.016877]-35.3334.88T-LEACHSAS0.007425 ±0.000423[0.007267,0.007583]0.008437 ±0.000261[0.008340,0.008535]-11.1548.26Higher(0, 0)CDS-LEACH0.004646 ±0.000242[0.004555,0.004736]0.016426 ±0.001726[0.015781,0.017071]-37.0130.14EEMTDMA0.004646 ±0.000242[0.004555,0.004736]0.123314 ±0.003281[0.122089,0.124539]-197.5729.32ENHANCED-LEACH0.004646 ±0.000242[0.004555,0.004736]0.011952 ±0.003656[0.010587,0.013317]-10.9229.25T-LEACH0.004646 ±0.000242[0.004555,0.004736]0.020168 ±0.001641[0.019555,0.020781]-51.2630.26T-LEACHSAS0.004646 ±0.000242[0.004555,0.004736]0.009617 ±0.000136[0.009566,0.009667]-98.1345.66(50, 50)CDS-LEACH0.004647 ±0.000200[0.004572,0.004721]0.018631 ±0.000549[0.018426,0.018836]-131.0836.53EEMTDMA0.004647 ±0.000200[0.004572,0.004721]0.116420 ±0.003417[0.115144,0.117696]-178.8629.20ENHANCED-LEACH0.004647 ±0.000200[0.004572,0.004721]0.010054 ±0.000383[0.009911,0.010197]-68.5543.66T-LEACH0.004647 ±0.000200[0.004572,0.004721]0.018053 ±0.001548[0.017475,0.018632]-47.0429.96T-LEACHSAS0.004647 ±0.000200[0.004572,0.004721]0.008747 ±0.000189[0.008677,0.008818]-81.7757.82 Detailed statistical comparison between EAAS-S4C-MAB and baseline protocols across all density and sink-location configurations. Values are reported as mean ± standard deviation with 95% confidence intervals. The p-values correspond to two-sided Welch t-tests. AvgEcons denotes AvgEconsPerRound. [825.74, 1092.66] [73.95, 88.05] [825.74, 1092.66] [68.37, 74.25] ENHANCED- LEACH [825.74, 1092.66] [134.27, 159.17] [825.74, 1092.66] [100.99, 131.61] [825.74, 1092.66] [308.90, 390.03] [895.82, 1181.31] [112.32, 132.72] [895.82, 1181.31] [72.05, 78.52] ENHANCED- LEACH [895.82, 1181.31] [327.96, 340.37] [895.82, 1181.31] [174.90, 210.77] [895.82, 1181.31] [610.27, 723.73] [2162.43, 2590.31] [90.57, 111.03] [2162.43, 2590.31] [69.07, 74.40] ENHANCED- LEACH [2162.43, 2590.31] [159.58, 200.04] [2162.43, 2590.31] [107.38, 134.09] [2162.43, 2590.31] [324.69, 408.38] [2289.90, 2655.40] [175.64, 213.72] [2289.90, 2655.40] [74.14, 80.07] ENHANCED- LEACH [2289.90, 2655.40] [310.08, 324.67] [2289.90, 2655.40] [158.33, 192.14] [2289.90, 2655.40] [576.91, 681.89] 0.007419 ± 0.000386 [0.007275, 0.007563] 0.016409 ± 0.002307 [0.015548, 0.017271] 0.007419 ± 0.000386 [0.007275, 0.007563] 0.122389 ± 0.005248 [0.120429, 0.124348] ENHANCED- LEACH 0.007419 ± 0.000386 [0.007275, 0.007563] 0.010165 ± 0.001019 [0.009784, 0.010545] 0.007419 ± 0.000386 [0.007275, 0.007563] 0.018501 ± 0.001774 [0.017838, 0.019163] 0.007419 ± 0.000386 [0.007275, 0.007563] 0.009448 ± 0.000129 [0.009400, 0.009496] 0.007425 ± 0.000423 [0.007267, 0.007583] 0.017000 ± 0.002202 [0.016177, 0.017822] 0.007425 ± 0.000423 [0.007267, 0.007583] 0.115391 ± 0.005124 [0.113478, 0.117305] ENHANCED- LEACH 0.007425 ± 0.000423 [0.007267, 0.007583] 0.010149 ± 0.000922 [0.009805, 0.010494] 0.007425 ± 0.000423 [0.007267, 0.007583] 0.016383 ± 0.001323 [0.015889, 0.016877] 0.007425 ± 0.000423 [0.007267, 0.007583] 0.008437 ± 0.000261 [0.008340, 0.008535] 0.004646 ± 0.000242 [0.004555, 0.004736] 0.016426 ± 0.001726 [0.015781, 0.017071] 0.004646 ± 0.000242 [0.004555, 0.004736] 0.123314 ± 0.003281 [0.122089, 0.124539] ENHANCED- LEACH 0.004646 ± 0.000242 [0.004555, 0.004736] 0.011952 ± 0.003656 [0.010587, 0.013317] 0.004646 ± 0.000242 [0.004555, 0.004736] 0.020168 ± 0.001641 [0.019555, 0.020781] 0.004646 ± 0.000242 [0.004555, 0.004736] 0.009617 ± 0.000136 [0.009566, 0.009667] 0.004647 ± 0.000200 [0.004572, 0.004721] 0.018631 ± 0.000549 [0.018426, 0.018836] 0.004647 ± 0.000200 [0.004572, 0.004721] 0.116420 ± 0.003417 [0.115144, 0.117696] ENHANCED- LEACH 0.004647 ± 0.000200 [0.004572, 0.004721] 0.010054 ± 0.000383 [0.009911, 0.010197] 0.004647 ± 0.000200 [0.004572, 0.004721] 0.018053 ± 0.001548 [0.017475, 0.018632] 0.004647 ± 0.000200 [0.004572, 0.004721] 0.008747 ± 0.000189 [0.008677, 0.008818] Conclusion This paper presented EAAS-S4C-MAB, an enhanced DS-based framework for extending the lifetime of heterogeneous wireless sensor networks (HWSNs). In the formation stage, the proposed Energy-Attentive Algorithm with Skyline, denoted as EAAS, jointly considers energy and DS size and retains non-dominated DS candidates through skyline selection. In the scheduling stage, S4C-MAB classifies DSs into four cases and applies size- and lifetime-aware activation policies to guide DS usage more effectively. By jointly enhancing both DS formation and DS scheduling, the framework reduces premature DS depletion and achieves more balanced energy utilization across the network. In this work, we first evaluated the proposed EAAS DS formation method and then evaluated the complete EAAS-S4C-MAB framework, which integrates EAAS with S4C-MAB scheduling. For EAAS, comprehensive simulations and comparisons with well-known DS-based algorithms, including MDP, MFMDP, LSearch, VDepth, and FDepth, showed that the proposed method consistently achieved the highest network lifetime across all tested network sizes and densities. Specifically, the proposed method achieved lifetime ranges of 96.47–195.68 for 200 nodes, 68.43–165.98 for 150 nodes, 43.84–145.04 for 100 nodes, and 31.21–79.85 for 50 nodes. These results represent clear improvements over the original EADDSA, whose corresponding lifetime ranges are 81.75–185.54, 57.59–159.71, 35.47–141.59, and 30.11–77.89, respectively. In addition, the comparative results showed that EAAS-S4C-MAB consistently outperformed T-LEACH32, EEM-TDMA33, T-LEACHSAS34, CDS-LEACH35, and Enhanced LEACH36 in the two directly comparable network-level metrics, namely FND and AvgEconsPerRound. Across all evaluated deployment scenarios, the proposed framework achieved the highest stability period and the lowest average per-round energy consumption, with especially large gains in higher-density settings. The results further showed that its performance was markedly less sensitive to sink relocation than that of the compared baselines, indicating stronger robustness to deployment geometry. Overall, by integrating energy- and size-aware DS construction with adaptive four-case scheduling and energy-aware communication, EAAS-S4C-MAB delivers substantial lifetime extension without introducing prohibitive computational overhead. The results confirm that the proposed framework effectively transforms network redundancy into longer lifetime, lower energy dissipation, and more balanced network operation. Limitation and future work Although the proposed EAAS-S4C-MAB framework improves energy balancing and network lifetime, several limitations remain. First, the current framework focuses mainly on energy-aware DS formation and scheduling, while fault tolerance, link reliability, and critical-node protection are not explicitly optimized. Second, the candidate DS generation, skyline filtering, and online scheduling procedures may become computationally expensive in large-scale and dense WSN deployments. From a practical deployment perspective, the framework can be integrated into IoT monitoring systems as an energy-aware backbone scheduling layer. The selected DS nodes serve as active sensing and aggregation nodes, while non-selected nodes remain in low-power mode. This reduces redundant communication and extends the operational lifetime of battery-powered IoT deployments. Since the framework is modular, the sink, edge controller, or SDN controller can execute the candidate-generation, skyline-selection, and scheduling modules, while sensor nodes only follow the resulting activation decisions. In future work, the framework can be extended by developing a hybrid optimization algorithm that jointly considers energy efficiency, load balancing, DS size, and fault tolerance during DS formation and scheduling. Critical-node awareness can also be incorporated by assigning each node a criticality score based on coverage dependency, DS invalidation impact, residual energy, communication burden, and topological importance. This score can then be used as an additional skyline objective, a penalty in the EAAS utility function, a cooldown adjustment, or a reward penalty in the scheduler. Such an extension would discourage repeated activation of structurally important nodes and help preserve network robustness. Another promising direction is to extend the current UCB-based scheduler toward reinforcement learning. Unlike the present MAB design, which offers a low-complexity solution suitable for resource-constrained WSNs, an RL-based scheduler could incorporate richer state information, including residual-energy distribution, node centrality, link quality, predicted traffic load, and historical activation patterns. This may improve long-term scheduling decisions, although it would introduce additional computational and training overhead. Scalability and dynamic-network adaptation remain important future directions. In large dense networks, the cost of candidate DS generation, skyline filtering, and online scheduling may increase substantially. This limitation can be addressed through edge/cloud-assisted computation, where the sink or an edge server performs computationally intensive tasks such as skyline generation, candidate pruning, and scheduling decisions, while sensor nodes only follow lightweight activation commands. Cluster-based decomposition can also partition the network into smaller regions, allowing EAAS–S4C-MAB to operate locally within each cluster. In addition, parallel skyline filtering, hierarchical PEGASIS chains, and multi-leader aggregation can further reduce preprocessing and communication overhead. The current framework assumes a static topology; node mobility, sink mobility, and stochastic link failures are not explicitly modeled. In highly dynamic networks, the candidate DS pool may need to be regenerated or updated online. The framework also assumes that residual energy and energy-consumption estimates are available with sufficient accuracy. In practice, measurement noise, delayed reporting, or hardware variation may affect scheduling decisions. Therefore, while the proposed framework is suitable for moderate-sized WSNs, large-scale and highly dynamic deployments require distributed, hierarchical, adaptive, or edge-assisted extensions. Acknowledgements The authors would like to thank the Research Chair of Prince Dr. Faisal bin Mishaal bin Saud bin Abdulaziz for Artificial Intelligence (CPFAI) at Qassim University for supporting this research work (QU-CPFAI-25EF205) Author contributions Conceptualization, Ahmed Salim , Ahmed M. Khedr , Ahmed A El-Sawy, Bader Alwasel , Walid Osamy ; Methodology: Bader Alwasel , Ahmed Salim , Ahmed A El-Sawy, Ahmed M. Khedr , Walid Osamy ; Software: Bader Alwasel , Ahmed Salim; Validation: Bader Alwasel , Ahmed Salim; Formal analysis: Ahmed Salim , Ahmed A El-Sawy, Ahmed M. Khedr , Walid Osamy ; Investigation: Ahmed A El-Sawy, Ahmed M. Khedr , Walid Osamy ; Resources: Ahmed Salim , Ahmed A El-Sawy ; Data curation: Bader Alwasel , Ahmed Salim ; Writing—original draft: Walid Osamy , Ahmed A El-Sawy, Ahmed M. Khedr ; Writing—review & editing, Ahmed A El-Sawy, Ahmed M. Khedr , Walid Osamy ; Visualization: Bader Alwasel , Ahmed Salim ; Supervision: Ahmed Salim ,Ahmed M. Khedr ; Project administration, Bader Alwasel , Ahmed Salim ; Funding acquisition, Bader Alwasel ; All authors have read and agreed to the published version of the manuscript. Funding The authors would like to thank the Research Chair of Prince Dr. Faisal bin Mishaal bin Saud bin Abdulaziz for Artificial Intelligence (CPFAI) at Qassim University for supporting this research work (QU-CPFAI-25EF205). Data availability All data generated or analysed during this study are included in this published article [and its supplementary information files] Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement All data generated or analysed during this study are included in this published article [and its supplementary information files]