Sensing to Learn: Deep Learning Based Wireless Sensing via Connected Digital and Physical Experiments

Abstract

With the advancement of wireless technologies and sensing methodologies, wireless sensing empowers wireless hardware with the additional ability to learn the target location, activity, gesture, and vital signs. By analyzing the target’s influence on surrounding wireless signals, deep learning-based wireless sensing has attracted great attention due to its excellent performance in extracting discriminative sensing patterns. Nevertheless, it is labor-intensive to generate massive amounts of data for deep sensing study. Most existing efforts are focused on better exploiting the acquired sensing information, which, however, can hardly address the knowledge limitation fundamentally. In view of this, we propose an innovative learning framework by strategically integrating digital and physical experiments, alleviating data collection's intensive efforts. Specifically, one adaption module is introduced to connect the digital simulator and field tests for producing high-quality digital sensing data. This framework is expected to enable automatic, reliable, and efficient sensing data generation for future wireless sensing studies.