Underdetermined Blind Source Separation Based On Reliable Time-frequency Points And Its Application In Dynamic Case

Blind source separation (BSS) is the process of extracting or recovering source signals from mixed signals without knowing both the source signals and the transmission environment. Owing to its important theoretical value, the blind source separation has wide applications in various fields such as image processing, speech signal processing, biomedical signal processing, radar signal detecting, wireless communications, etc. In the intelligent robot’s auditory system, the number of observed signals is less than the number of sources because of the restriction on the number of sensors, resulting in the underdetermined blind source separation problem (UBSS). In addition, when the sources are moving, the transmission channel is not only unknown, but also varying, which makes the separation problem much more difficult. It forms complicated underdetermined blind source separation in dynamic case. This paper made research in the above field and the main contributions include:1. With the invariable transmission channel, i.e. in the case of stationary mixing mode, a separation algorithm based on the mixing parameter estimation from the selected reliable time-frequency (TF) points is proposed, which is successfully used to underdetermined BSS. When the source signals are sparse in time-frequency domain, the relative attenuation and the relative delay deduced by the ratio of two observed signals will show clustering features. Instead of directly using the method of clustering or histogram to estimate the peak in traditional DUET algorithm, we introduced the selection criterion by TF points’ power into Fuzzy C-Means (FCM) clustering to estimate the mixing parameters, which strengthened the clustering effect and greatly improved the separation performance.2. In the dynamic environment, i.e. with the time-varying mixing matrix, taking the advantages of both the batching separation algorithm and the adaptive online algorithm, the block-online adaptive algorithm is proposed to solve the UBSS in non-stationary case. The idea of the standard block-ICA is extended to the UBSS. In the continuous time-varying case, the above improved offline algorithm is employed within one frame. To track the continuous variable mixing parameters, the estimated ones of previous frame can be used to initialize FCM in the present frame, instead of random initialization in case of the conventional FCM.