A Time-Frequency Algorithm for Noisy ICA

The performance of standard algorithms for Independent Component Analysis (ICA) quickly deteriorates when the signals are contaminated by additive noise. In this paper, we propose an ICA approach exploiting the difference in the time-frequency (t-f) signatures of noisy signals to be separated. The approach uses a high-resolution t-f distribution to obtain the t-f matrices of mixed signals, then localizes the signal energy by Hough transform and obtains the estimated signals based on the diagonalization of a combined set of auto-term matrices. Furthermore, its performance is evaluated using the Signal-Noise-Ratio (SNR) as it is commonly employed to assess the ICA algorithms. Both the results of mathematical analysis and numerical simulations indicate that we could enhance the ICA performance by improving the input SNR or increasing the number of sampling points. The approach could increase the ICA robustness by spreading the noise power and localizing the source energy in the t-f domain.