Estimation of the number of signals from features of the covariance matrix: A supervised approach

The purpose of this paper is to provide a fast and simplified detection test for use in the presence of a small number of sources (from 0-2), which is able to accommodate correlated paths and nonwhite noise; conventional eigenvalue-based criteria are unable to do so. For a uniform linear array, using common sense arguments, a small set of significant features of the covariance matrix are used as inputs to a neural net. The nonlinear transfer function of the neural net is adjusted by supervised training to provide the discriminant functions for order selection in its outputs. Results from the net are then compared with conventional criteria and demonstrate superior performance, in particular, for correlated sources and small sample sizes. Training may be introduced for known nonwhite noise, which serves to maintain high performance for reasonable correlation lengths.