Speaker
Description
TorchSig is software for radio frequency (RF) machine learning (ML) that generates signals and performs model training for signal detection and modulation recognition. The software is also packaged with the gr-spectrumdetect GNU Radio block which runs an ML model for detection and modulation recognition. The talk and paper discuss the latest advancements in the software: the ability to train against custom, user-provided datasets and a new set of RF-based transforms.
TorchSig has a library of 57 signal types, including frequency shift keying (FSK), quadrature amplitude modulation (QAM), orthogonal frequency division multiplexing (OFDM) and others. Custom signals allow a user the ability to supplement these built-in signals for ML training. The signals can be added in natively within the software or through reading in binary IQ files.
In order to make the signals and environment more realistic, TorchSig implements RF-based transforms which are real-world effects from hardware impairments and channel impairments such as IQ imbalance and quantization. TorchSig also implements ML-based transforms such as time-reversal and IQ swap. The transforms are applied to the signals which are then used to train an ML model against more realistic scenarios. A new, expanded set of transforms is presented.
Custom signals and datasets allow users the ability to integrate specific signals of interest into the automated training pipeline. This capability expands the scope of TorchSig and enables use of the software with any type of signal a user can create or imagine.
TorchSig is free, open-source software available at github.com/torchdsp/torchsig. For more information please see torchsig.com.
| Talk Length | 30 Minutes |
|---|---|
| Link to Open Source Code | github.com/torchdsp/torchsig |
