Nonlinear Signal Decomposition Based on Block Sparse Approximation

We propose here a nonlinear decomposition approach that properly separates the frequency content of signals. To do so, we propose a block-based sparse coding frequency separation (B-SCFS) method with a dictionary constituted of an infinite number of chirp-like atoms so that general signals can be analyzed. Also, a suitable modification of orthogonal matching pursuits (OMP) algorithm is introduced, combined with a block-based procedure for an efficient handling of long signals. Conducted experiments show that the proposed approach noticeably improves on Empirical Mode Decomposition (EMD) and former SCFS as well, which performed well only on a limited class of signals.