Training data augmentation for deep learning radio frequency systems

Applications of machine learning are subject to three major components that contribute the final performance metrics. Within category neural networks, and deep specifically, first two architecture for model being trained training approach used. This work focuses on third component, data used during training. The primary questions arise “what is in data” within matters?” looking into radio frequency (RFML) field automatic modulation classification (AMC) as an example a tool situational awareness, use synthetic, captured, augmented examined compared provide insights about quantity quality available necessary achieve desired levels. Three discussed this work: (1) how useful synthetically system expected be when deployed without considering environment synthesis, (2) can augmentation leveraged RFML domain, and, lastly, (3) what impact knowledge degradations signal caused by transmission channel contributes system. In general, types each make contributions application, but captured germane intended case will always more significant information enable greatest performance. Despite benefit data, difficulties costs from live collection often needed peak impractical. paper helps quantify balance between real synthetic offering concrete examples where parametrically varied size source.