Data Centric Approach to Modulation Classification

Deep learning (DL) for modulation classification has shown significant performance improvements. The focus has been model centric, where newer architectures are demonstrated on benchmark data RADIOML.2016.10A (RML16). In contrast, we use a data centric DL approach where focus is on data quality. RML16 has shortcomings such as errors and ad-hoc choice of parameters. We build upon RML16 and provide realistic methodology of generating dataset. The errors in RML16 and appropriate corrections are discussed. A benchmark dataset RML22 is introduced that incorporates these corrections. The Python source code used to generate RML22 is shared to enable further improvements.

Automated summary

Deep learning has achieved significant performance improvements in modulation classification. Previous research mainly focused on model construction and benchmark dataset RML16, while this study adopted a data-centric deep learning approach. By addressing the errors and ad-hoc parameter choices in RML16, a more realistic dataset RML22 was introduced, and the Python source code used to generate RML22 was shared for further improvements.