Composite modeling of nonstationary signals (Q1093611)

A model for discrete-time nonstationary signals is discussed. The proposed model is of the following form: \(x(n)=m_ x(n)+\tilde x(n)\), where n belongs to a finite set of moments, \(m_ x(n)\) is the (time- varying) mean value of x(n) and \(\tilde x(n)\) a zero mean nonstationary signal. Time-varying estimators for the mean value and the variance functions are obtained and used to produce a zero-mean, constant-variance signal which can be modeled by autoregressive systems with time-varying coefficients. An identification method is given to obtain the filter coefficients. Some examples shows the advantages of the proposed method with respect to time-invariant models for many signals of practical interest, such as speech and radar return signals.