.png){kind=icon}
.png){kind=icon}
Return
RML22: Realistic Dataset Generation for Wireless Modulation Classification
.png){kind=icon}
PRE
AI
DOI：10.1109/TWC.2023.3254490.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Application of Deep learning (DL) to modulation classification has shown significant performance improvements. The focus has been model centric, where newer architectures are attempted on benchmark dataset RADIOML.2016.10A (RML16). RML16 is a high impact effort that laid the foundation for generating a synthetic dataset for applying DL models to wireless problems. This encouraged development of newer architectures to RML16. We use a data centric DL approach where focus moves from model architectures to data quality. RML16 has shortcomings such as errors and ad-hoc choices of parameters. We build upon RML16 and provide realistic and correct methodology of generating dataset. A new benchmark dataset RML22 is generated. Going forward, we envision researchers to improve model quality on RML22. We attempt to improve data quality by studying the impact of information sources. Further, the choices of artifacts and signal model parameterization are analyzed carefully. The Python source code used to generate RML22 is shared to enable researchers to further improve dataset quality.
Keywords:
Deep learning
modulation classification
benchmark dataset
GNU radio
spectrum sensing
Journal
IF:
10.7
Papers:
1.3W
Citations:
5.3W
