RF Intelligence for Health: Classification of SmartBAN Signals in overcrowded ISM band

TL;DR

This study tackles the challenge of classifying ultra-low-power SmartBAN RF signals within the crowded 2.4 GHz ISM spectrum to enable interference-aware coexistence in wearable health monitoring. It introduces an open-source framework that blends synthetic spectrogram generation with real SDR over-the-air data and trains CNN-based semantic segmentation models (encoder–decoder with ResNet backbones and attention-enabled U‑Net decoders) to label time-frequency components by signal class. The method achieves over 90% accuracy on synthetic data across architectures and demonstrates consistent performance on real-world captures, validating robustness to dense interference and power asymmetry. The work advances reliable SmartBAN signal recognition in dense spectral environments, offering a foundation for adaptive spectrum access and more dependable wireless health monitoring in cluttered ISM bands.

Abstract

Accurate classification of Radio-Frequency (RF) signals is essential for reliable wearable health-monitoring systems, providing awareness of the interference conditions in which medical protocols operate. In the overcrowded 2.4 GHz ISM band, however, identifying low-power transmissions from medical sensors is challenging due to strong co-channel interference and substantial power asymmetry with coexisting technologies. This work introduces the first open source framework for automatic recognition of SmartBAN signals in Body Area Networks (BANs). The framework combines a synthetic dataset of simulated signals with real RF acquisitions obtained through Software-Defined Radios (SDRs), enabling both controlled and realistic evaluation. Deep convolutional neural networks based on ResNet encoders and U-Net decoders with attention mechanisms are trained and assessed across diverse propagation conditions. The proposed approach achieves over 90% accuracy on synthetic datasets and demonstrates consistent performance on real over-the-air spectrograms. By enabling reliable SmartBAN signal recognition in dense spectral environments, this framework supports interferenceaware coexistence strategies and improves the dependability of wearable healthcare systems.

Figures (11)

• Figure 1: An example of a crowded RF signals environment with different co-existing technologies at ISM band.
• Figure 3: Comparative analysis of performance metrics for different network architectures.
• Figure 4: Comparison of SmartBAN classification performance at different transmitter–receiver distances in presence of various other interference signals for various neural network architectures. Accuracy, Dice score, and IoU are shown as functions of distance.
• Figure 5: ResNet50 network classification performance on synthetic captures with two ISM band conditions.
• Figure 6: Experimental setup and classification result on a real-world capture at ISM band.