Instantaneous Frequency Estimation for Multicomponent Signals Based on Conditional Adversarial Generation of Time-Frequency Distribution

Instantaneous frequency (IF) estimation plays an important role in multicomponent signal processing. However, most existing methods achieve poor performance when IF curves of signal components are closely spaced or intersecting. To address the above problem, this paper proposes a multicomponent signal IF estimation method which tries to approach ideal time-frequency (TF) distribution by exploiting conditional adversarial generation. Firstly, this method adopts TF distribution (TFD) images (e.g., masked Wigner-Ville distribution in this paper) via TF analysis as the original dataset of conditional generative adversarial network (CGAN). Based on the produced dataset, the CGAN is trained to learn features of input images and then we use the trained network to generate new TFD images, which are closer to ideal TFDs. According to these optimized TFD images, an improved Viterbi algorithm is employed to extract IF curves of different components. This improved algorithm introduces a penalty function corresponding to line segment gradient in order to guarantee the estimation accuracy when signal components are intersecting in TF domain. Experimental results show the effectiveness and superiority of the proposed method in the case where these IF curves are closely spaced or intersecting.