A Self-Evolving Wireless Sensor Network Architecture Integrating Temporal Graph Neural Networks and Multi-Agent Deep Reinforcement Learning

Wireless sensor networks (WSNs) operating in dynamic, interference-prone, and energy-limited environments must cope with continuously evolving topologies, uneven energy depletion, and unpredictable node and link failures. Despite significant progress, most existing approaches remain fundamentally reactive : they adapt only after disruptions occur or treat prediction and control as loosely coupled components, preventing proactive resilience and long-term stability. We present SE-WSN , a self-evolving network architecture that unifies temporal graph intelligence with coordinated multi-agent control. SE-WSN models the WSN as a time-evolving graph and employs a Temporal Graph Neural Network (Temporal-GNN) to forecast short-horizon structural risks—ranging from node-failure likelihood to link-quality degradation and cluster-role instability. These predictions are distilled into a compact structural risk context that conditions a MADDPG-based controller, enabling distributed nodes to make risk-aware decisions for routing, transmission-power regulation, duty-cycling, and cluster-role adaptation. A lightweight self-healing engine translates these decisions into continuous, online topology reconfiguration, forming a closed-loop, prediction-driven adaptation cycle unprecedented in prior WSN designs. Comprehensive simulations across heterogeneous densities, traffic patterns, and failure scenarios demonstrate that SE-WSN delivers consistent and statistically robust gains in network lifetime, reliability, and energy balance over both classical protocols and state-of-the-art learning-based baselines—while maintaining low computational and communication overhead. Ablation studies further reveal the complementary roles of temporal graph forecasting and risk-aware reward shaping in stabilizing learning dynamics and preventing cascading failures. These results indicate that tightly coupling temporal graph representation learning with coordinated multi-agent control establishes a powerful foundation for next-generation autonomic and self-healing WSNs.