Abstract
In this paper, we study a 2 × 3 temporomandibular joint (TMJ) underdetermined blind source separation (UBSS). This particular UBSS has been subject to an empirical experiment performed previously on two sparse TMJ sources and a non-sparse source modelled as super-Gaussian noise. In this study, we found that FastICA algorithm tends to separate the two highly super-Gaussian sources when applied to the mixtures. When these two mixtures were filtered, FastICA focused on the non-sparse source (i.e. noise). Previously, we did not examine why such filtering approach would lead to estimation of the nonsparse source. To this end, the objective is to provide an extensive set of simulations to demonstrate why this filtering approach fully solve this particular underdetermined blind identification. We have employed the shape parameter α of the generalized Gaussian distribution (GGD) as a measure of sparseness and Gaussianity. This parameter was also utilized to illustrate the convergence of our filtering approach and the sub-Gaussian effect of the filter on the mixtures. Moreover, we have also considered the case where the noise source is modelled as sub-Gaussian and Gaussian as an extension of our previous work. Simulation studies show that our filtering approach is robust and performs well in this particular TMJ UBSS application.