Blind Separation of Frequency Overlapped Sources Based on Constrained Non-Negative Matrix Factorization

Abstract

The separation of unobserved sources from the observed signals is a fundamental signal processing problem. Most of the proposed techniques for solving this problem rely on independence or at least uncorrelation assumption of source signals. However in some complex systems, the vibration sources are always correlative, and this does not satisfy the assumption condition. Here, a new method based on constrained non-negative matrix factorization (CNMF) is introduced for the case that the sources are correlated only through the overlapping frequencies. In contrast with other reported methods, the proposed method separates source signals in frequency domain without a parametric model of their dependent structure, and is mainly based on the good property of non-negative matrix factorization (NMF) that the sources do not need to be statistically independent. Some numerical simulations are provided to illustrate the feasibility and effectiveness of the proposed method.