Blind Identification of Noisy Non-stationary Sources Using a Binary Mask

In most functional blind source separation (BSS) applications, the observations contain additive noise that limits the performance of most existing BSS algorithms, especially in the case where the noise is not modeled by a random process (e.g., electromagnetic power supply noise). In this paper, we describe an algorithm where a new cost function is fed with a frequency profile of the noise. In this way, the coefficients of the separating matrix are identified without the bias introduced by the noise. The proposed cost function is based on a frequency domain binary mask and the coherence function. The binary mask selects the frequencies where the signal-to-noise ratio (SNR) is relatively high, while the coherence is minimized to obtain the inverse system. Moreover, any frequency noise profile, whether given a priori or estimated, can be applied to the binary mask to achieve the identification of the inverse matrix. Computer simulations show that the proposed algorithm exhibits better performance under different SNR scenarios compared to methods developed previously.