Time-Frequency Trade-Offs in Synchrosqueezing Transform-Based Waveform Recognition

The spectrum is a shared, congested, and densely used resource. The emerging sixth-generation wireless (6 G) systems intend to alleviate this problem by integrating radio frequency sensing and wireless communications functionalities, enabling spectrum, hardware, and antenna resource sharing. Agile spectrum use requires situational awareness (SA) about the state of the radio environment. Automatic waveform recognition (AWR) is an important part of building and maintaining SA needed in cognitive radios, radars, and signal intelligence. Commonly, a convolutional neural network (CNN) has been paired with time-frequency (TF) imaging to distinguish among different low probability of intercept/detection (LPI/LPD) waveforms. In this paper, we thoroughly investigate the effect of the TF resolution in the classification accuracy of twelve LPI/LPD radar waveforms using different variants of the Fourier-based synchrosqueezing transform (FSST) to generate time-frequency images (TFIs). The pros and cons of downsampling in time and frequency domains are analyzed at the Heisenberg-Gabor limit. Practical classification results are presented at different sensing channel conditions.