Joint Electromagnetic-Network Perception in WA-IoT: A Cross-Domain Reciprocal Learning Approach

Abstract

Wide-Area Internet of Things (WA-IoT) situational awareness requires joint modeling of electromagnetic spectrum occupancy and network quality states under low-SNR interference. Existing single-domain or shallow-fusion methods often underuse the statistical dependence between spectrum observations and link-level QoS. This paper proposes RLNet, a reciprocal cross-domain learning framework that combines adaptive spatial graph construction, dual-stream information exchange, gated feature reshaping, diversity regularization, and adaptive fusion. In a simulated WA-IoT scenario, RLNet achieves 89.81% detection probability and $\mathbf{1. 4 6}$% false alarm probability at an extremely low SNR of -15 dB, and reduces delay, packet-loss, and effective-rate prediction errors compared with graph, Transformer, and fusion baselines. The results show that reciprocal electromagnetic-network representation learning improves joint perception in low-SNR WA-IoT monitoring.