A Coverage-Preserving Dual-Graph Neural Sleep Scheduler for Wireless Sensor Networks

Abstract

Sleep scheduling is a direct means of extending the lifetime of dense wireless sensor networks, but independently classifying each node as active or sleeping does not guarantee that the resulting subset preserves both sensing coverage and communication connectivity. The Dual-Graph Neural Sleep Scheduler (DGNSS) is proposed as a coverage-preserving method for dynamic sleep scheduling. The method represents a deployment by two complementary graphs: a communication graph encoding radio reachability and a coverage-overlap graph encoding sensing redundancy. Separate Graph Sample and Aggregate (GraphSAGE) encoders process the two structures, and a gated fusion module produces node-level keep scores from structural, energy, coverage-witness, articulation, and temporal duty-cycle features. Instead of applying a fixed probability threshold, the neural scores guide a deterministic feasibility-by-construction decoder. The decoder combines multi-start pruning, constructive scheduling, exact grid-coverage and sink-connectivity tests, two-for-one exchanges, temporal energy rotation, and a safe-improvement guardrail. Experiments use reproducible energy-stress scenarios with 50 and 100 nodes and compare the proposed method with Always Active, random and threshold scheduling, an energy-aware greedy scheduler, coverage-redundancy and connected-dominating-set heuristics, a Low-Energy Adaptive Clustering Hierarchy (LEACH)-like policy, a greedy oracle, the former independent-threshold decoder, and method ablations. In the 50-node scenario, DGNSS saves 50.87% energy, maintains the coverage requirement for all 140 rounds, and delays first-node death by 4.33 rounds relative to the strongest energy-aware greedy baseline. In the 100-node scenario, it saves 24.47% energy and extends coverage lifetime by one round relative to the same baseline. Additional ten-topology robustness runs show a statistically supported energy-saving gain over Energy-Aware Greedy at 50 nodes and a smaller, non-significant energy-saving trend at 100 nodes. A feature-only multilayer perceptron ranker is also competitive under the same decoder, indicating that the deterministic connected-cover decoder and coverage-aware features account for a substantial part of the observed gains. These results show that learned ranking and exact constrained decoding are complementary: the graph model supplies transferable structural preferences, whereas the decoder prevents infeasible schedules and improves the energy--lifetime trade-off.