Crossterm-free time-frequency representation exploiting deep convolutional neural network

Abstract

Bilinear time-frequency (TF) analyses provide high-resolution time-varying frequency characterization of nonstationary signals. However, because of their bilinear natures, such TF representations (TFRs) suffer from crossterms. TF kernels, which amount to low-pass weighting or masking in the ambiguity function domain, are commonly used to reduce crossterms. However, existing fixed and adaptive kernels do not guarantee effective crossterm suppression and autoterm preservation, particularly for signals with overlapping autoterms and crossterms in the ambiguity function. In this paper, we develop a new method that offers high-resolution TFRs of nonstationary signals with desired autoterm preservation and crossterm mitigation capabilities, especially for signals with slowly time-varying instantaneous frequencies. The proposed method exploits a convolutional autoencoder network which is trained to construct crossterm-free TFRs. For the signals being considered, the proposed technique with properly trained networks offers the capability to outperform state-of-the-art TF analysis algorithms based on adaptive kernels and compressive sensing techniques.