Crossterm-Free Time-Frequency Analyses Exploiting Deep Neural Networks

Nonstationary signals are naturally found and exploited in various applications, such as radar, sonar, radio astronomy, seismology, and electroencephalogram (EEG) [1]-[6]. Time-frequency (TF) analysis of nonstationary signals is a key enabling technology for radar-based human-machine interface through gesture recognition using hands and arms [7], [8]. Many nonstationary signals can be described as frequency modulated (FM) signals and characterized by their time-varying instantaneous frequencies (IFs). Compared to single-domain signal representations with respect to either time or frequency, joint TF-domain representations are better suited for the analyses and classification of such signals as they provide a time-varying spectrum [1], [2].