Bluetooth Low Energy (BLE) RF Dataset for Machine Learning in WBANs

The lack of availability of real-world RF datasets has often impeded physical layer research relating to IoT and Health IoT sectors. Towards this end, this paper presents an unprecedented open-source Bluetooth Low Energy (BLE) dataset, encompassing 100 MSps radio signal recordings from 13 ESP32 IoT nodes in both on-body and over-the-air settings within an anechoic chamber, utilizing the dual receiver setup on USRP x310. Wideband (complete 2.4 GHz ISM band) acquisition and high sampling frequency ensure exceptional resolution and room for use case-specific sampling scalability in the dataset. We delineate the dataset's acquisition methodology, spotlight its inherent advantages, and offer a valuable Python tool to process and scrutinize the raw IQ samples. To facilitate the use of datasets in machine learning applications for the physical layer, we have incorporated this tool within a dataset hosted on a public GitHub repository. Here, we explore preliminary feature selection and conduct tests and training of machine learning models, including Random Forest and K-Nearest Neighbors (KNN) that achieve 81% accuracy in node classification for static on-body frames. This demonstrates the dataset's versatility in various scenarios, such as authentication and localization.