Robust subspace estimation in non-Gaussian noise

Subspace methods are common in array processing, but standard schemes typically perform poorly when the noise is non-Gaussian and/or impulsive. Zero-memory nonlinear (ZMNL) functions may be applied to limit the influence of impulsive noise, but ZMNL pre-processing generally destroys the low-rank signal subspace. We develop a robust covariance matrix estimate that suppresses impulsive noise while also performing a model-based interpolation to restore the signal subspace. The approach is based on modeling the noise with a finite Gaussian mixture distribution, and an expectation-maximization (EM) algorithm is used for parameter estimation. The method is robust to noise model mismatch and works well with infinite-variance noise. Simulation results are included that illustrate the improved performance in detecting the number of sources and estimating the angles of arrival