Amer Al-Habsi

Optimal Reduced-Order Moment Estimation for Non-Stationary Signals

Thursday, August 27, 1998
3:30 PM
306 Egan

Abstract

The statistical moments of a stationary and ergodic signal can be estimated by uniform (or exponentially-weighted) time averaging, which is the same as filtering with a narrow-band lowpass filter. In contrast, moment estimation for nonstationary signals requires, in general, a wideband averaging filter whose response has to be adapted to the dynamics of the time-variant moment we wish to estimate. A recently proposed approach provides a simple characterization (in terms of certain average spectra) of the optimized averaging filter for any particular nonstationary moment estimation problem.

In this thesis we introduce several techniques for constructing low-order IIR approximants of the optimized averaging filter (which usually has infinite order). We show that the such low-order IIR averaging filters offer a very attractive cost-performance tradeoff: they can achieve essentially the same level of moment estimation error as the optimized averaging filter but at a significantly lower implementation cost. In contrast, past attempts to realize the optimized averaging filter employed high-order FIR approximants. We present several examples to demonstrate the cost savings achieved with our IIR approximants.

Thesis Committee:
Prof. H. Lev-Ari (advisor)
Prof. D.H. Brooks
Prof. R. Raghavan