Efficient block-wise independent component analysis with optimal learning rate

Abstract

Independent component analysis (ICA) is a potential solution for blind source separation (BSS) problems and has proven to be effective in many practical applications. One major drawback for ICA algorithms is its heavy computation load which is especially deteriorated in long-term signals. To solve this problem, the block-wise approach is proposed to improve the computational efficiency in this study. As ambiguities of variances and order of the separated components exist for ICA algorithms, signal reconstruction is an issue for block-wise approach. In this study, correlation coefficients (CC) of the short-term Fourier transform (STFT) among those acquired from the neighboring blocks are adopted for order permutation and phase correction of the separated signals. The weights of the separated signals are then normalized by the maximum amplitude of the corresponding one in previous block to avoid the variance ambiguity. Two robust ICA algorithms based on iterative matrix inversion approach, one with fixed nonlinear functions whereas another one with flexibly selected nonlinear functions, are implemented in both batch and block-wise processing for performance evaluation. Linearly mixed biopotential signals (ECG and EMG) and 60-Hz sinusoid are incorporated for the experiments. The results reveal that the proposed block-wise approach can attain the desired requirements in a more efficient way without loss of the separation performance. The proposed approach may be a potential tool for which the computational efficiency is an important issue.