Signal Quality Aware Deep RF Sensing

Radio Frequency (RF) sensing with deep learning is evolving rapidly, resulting in numerous task networks that enable various applications. However, the notable noise and interference in RF sensing lead to variations in signal quality. If not handled properly, the low-quality signals can undermine the efficacy of task networks. In this paper, we propose a signal quality aware deep RF sensing framework to quantify the signal quality and incorporate it for performance enhancement. Specifically, we design a unified plug-in module for classic task networks which projects signal quality-related representations to quality scores and dynamically modulates the training loss of our framework to avoid the impact of low-quality signals. A novel training strategy is also developed to jointly optimize quality scores and task networks. The experiment results show our framework effectively mitigates low-quality signal interference during training. For two typical RF sensing tasks, compared to ignoring signal quality, task network performance improved by 4.2% to 5.9%.