Speakers
Description
The development of robust Radio Frequency Machine Learning (RFML) systems has been constrained by the suitability of realistic training data and the lack of models trained against real-world channel effects. While recent efforts focus on precise mathematical models for RF propagation, the most accurate approach is to operate directly in the real-world electromagnetic environment. This work presents an open-source RF Environment Generator (RFEG) that enables procedurally-generated over-the-air (OTA) datasets for online machine learning training for a variety of downstream tasks. The system employs distributed USRPs controlled via GNU Radio and general purpose processor (GPP) nodes, orchestrated by a central node generating JSON-formatted signal parameter files to emulate a user-specified environment.
The RFEG implements matched filtering and time-synchronization protocols enabling precise alignment between received signals and their reference copies, which is essential for downstream RFML tasks including symbol-level information recovery, channel estimation, equalization, interference mitigation and demodulation. We demonstrate the system through OTA validation using 3 distributed USRP emitters. By providing fully open-source, reconfigurable infrastructure, the RFEG enables research groups to perform cognitive sensing research embedded in their electromagnetic environment, rather than being constrained to generic datasets that may not reflect their operational scenarios.
| Talk Length | N/A |
|---|---|
| Link to Open Source Code | Coming soon |
| Acknowledge | Acknowledge In-Person |