Array and Multichannel Signal Processing Using Nonparametric Statistics

In array signal processing a group of sensors located at distinct spatial locations is deployed to measure a propagating wave eld. The multichannel output is then processed to provide information about parameters of interest. Application areas include smart antennas in communications, radar, sonar and biomedicine. When deriving array signal processing algorithms the noise is typically modeled as a white Gaussian random process. A shortcoming of the estimation procedures derived under Gaussian assumption is that they are extremely sensitive to deviations from the assumed model, i.e. they are not robust. In real-world applications the assumption of white Gaussian noise is not always valid. Consequently, there has been a growing interest in estimation methods which work reliably in both Gaussian and non-Gaussian noise. In this thesis, new statistical procedures for array and multichannel signal processing are developed. In the area of array signal processing, the work concentrates on high-resolution subspace-based Direction Of Arrival (DOA) estimation and estimation of the number of source signals. Robust methods for DOA estimation and estimation of the number of source signals are derived. Spatial-smoothing based extensions of the techniques to deal with coherent signals are also derived. The methods developed are based on multivariate nonparametric statistics, in particular sign and rank covariance matrices. It is shown that these statistics may be used to obtain convergent estimates of the signal and noise subspaces for a large family of symmetric noise distributions. Simulations reveal that the techniques developed exhibit near-optimal performance when the noise distribution is Gaussian and are highly reliable if the noise is non-Gaussian. Multivariate nonparametric statistics are also applied to frequency estimation and estimation of the eigenvectors of the covariance matrix. Theoretical justi cation for the techniques is shown and their robust performance is illustrated in simulations.