A New Clustering Algorithm Based on Normalized Signal for Sparse Component Analysis

To the underdetermined sparse component analysis (SCA) model with noise, a new robust clustering algorithm based on normalized signal for mixture matrix estimation is addressed in this paper. This approach consists of two parts: signal clustering and matrix recovery. In the first step, according to the feature of normal signals clustering intensively on the unit observed signal hyper-sphere, we propose a criterion to detect and cluster dense observed signal sets, which is the conclusion of deduction from a fit mathematical statistics model. To the second stage for estimating the mixture matrix, Principal Component Analysis is introduced to process dense signal sets. Experiment simulations illustrate that new clustering algorithm´s performance on determination of the source numbers and precision of mixing matrix recovery.