On Array Processing in Non-Stationary Random Fields

Abstract

The paper discusses semi-stationary vector processes as non-stationary models for the outputs of a passive-sonar array. The processes are characterized by time-evolutionary spectral-density matrices. Careful smoothing of short-period spectrograms is a reasonable method for estimating the spectral-density matrices. Other models recently introduced and basing on similar ideas are semi-homogeneous random fields characterized e.g. by frequency-angle spectral densities which depend on time and position. Tracking of short-period time-space spectrograms yields useful estimates of the densities. Typically, one cannot obtain simultaneously a high resolution in both the time-space domain and the frequency-angle domain. We report on investigations if principle component analysis of the estimated evolutionary spectral-density matrices can be used to obtain a better resolution of the estimates for the angle distribution.