Expert Fusion with Meta-Adaptation for Signal Technology Recognition
Ife Olalekan Ebo, Idowu Ajayi, Lina Mroueh, Youmni Ziade · 2025
Low Power Wide Area Networks (LPWANs) enable cost-effective, low power consumption, and long-range IoT applications but often operate under low Signal-to-Noise Ratio (SNR) conditions. Conventional deep learning models for signal recognition struggle to generalize in such dynamic environments. We propose a lightweight framework combining ensemble learning, a Mixture of Experts (MoE) with uncertainty-aware soft gating, and meta-adaptation using the Almost No Inner Loop (ANIL) method. Expert models trained at distinct lowSNR levels share frozen base weights, while only the classifier head is adapted during few-shot learning for rapid specialization. The uncertainty-aware gating mechanism produces sparse, temperature-controlled expert weights, enhancing robustness and decision reliability. Experiments demonstrate improved accuracy and generalization across both seen and unseen SNR levels, highlighting the framework's effectiveness for real-world wireless signal technology recognition in spectrum monitoring systems.