A Comparative Study of Infomax, Extended Infomax and Multi-User Kurtosis Algorithms for Blind Source Separation

Abstract

In this article for the separation of Super Gaussian and Sub-Gaussian signals, we have considered the Multi-User Kurtosis(MUK), Infomax (Information Maximization) and Extended Infomax algorithms. For Extended Infomax we have taken two different non-linear functions and new coefficients and for Infomax we have taken a single non-linear function. We have derived MUK algorithm with stochastic gradient update iteratively using MUK cost function abided by a Gram-Schmidt orthogonalization to project on to the criterion constraint. Amongst the various standards available for measuring blind source separation, Cross-correlation coefficient and Kurtosis are considered to analyze the performance of the algorithms. An important finding of this study, as is evident from the performance table, is that the Kurtosis and Correlation coefficient values are the most favorable for the Extended Infomax algorithm, when compared with the others.

Nonlinear Blind Source Separation and Fault Feature Extraction Method for Mining Machine Diagnosis

“…When the nonlinearity of the mixed observation signals increases, ICA is not effective and, in some cases, even fails to identify the fault vibration source [14]. Consequently, the linear methods are not appropriate for complex nonlinear separation problems, such as in steam turbines, generators and mining machines [15]. Therefore, nonlinear blind separation has become a hot issue in fault vibration source separation.…”

“…Li et al [20] introduced a separator for marine gearbox fault source extraction based on the Chaos algorithm. However, the main limitation of nonlinear blind source separation (BSS) methods is that their solution is not unique [15]. In view of the fact that the neural networks can effectively identify the uncertainty of a nonlinear system, it is possible to combine the neural network and ICA to separate the nonlinear mixtures.…”