EdgeDR-MR: A Dynamic Resource-Aware Lightweight Modulation Recognition Method for Integrated Sensing and Control in Consumer IoT

The automatic modulation recognition (AMR) based on deep learning models faces significant challenges in resource-limited consumer IoT scenarios due to their high computational complexity and the dynamic nature of device resources. While existing works primarily focus on static lightweight designs, they lack the ability for runtime adaptation. To address the conflict between model structures and fluctuating resource conditions, we propose a dynamic resource-aware lightweight AMR method named EdgeDR-MR. It features a dual-mode model architecture, comprising a high-accuracy mode and a high-efficiency mode that share common feature encoder, alongside dynamic resource scheduler. This scheduler monitors real-time device resources (CPU, latency, power, memory) to intelligently switch between modes, enabling a fine-grained trade-off. Experiments on the public RML22 dataset demonstrate that our method significantly reduces average inference latency and energy consumption while maintaining high recognition accuracy, outperforming static benchmarks. The results prove its effectiveness and practicality for enhancing the autonomy of consumer IoT devices in real-world dynamic scenarios.