Robust system identification for non-persistently exciting input signals

The performance of adaptive identification algorithms is constrained by the spectral characteristics of the input signals. Too few frequency components may result in parameter wander, and possible instability. Direct sequence spread-spectrum techniques may be used to increase the spectral richness of the training signal. Computer simulations of gradient descent algorithms for FIR (finite impulse response) and IIR (infinite impulse response) systems indicate that parameter wander is eliminated and the rate of convergence is dramatically increased. The convergence of least squares algorithms is unaffected, but it is conjectured that their numerical properties are improved.