Unsupervised robust recursive least-squares algorithm for impulsive noise filtering

Abstract

A robust recursive least-squares (RLS) adaptive filter against impulsive noise is proposed for the situation of an unknown desired signal. By minimizing a saturable nonlinear constrained unsupervised cost function instead of the conventional least-squares function, a possible impulse-corrupted signal is prevented from entering the filter’s weight updating scheme. Moreover, a multi-step adaptive filter is devised to reconstruct the observed “impulse-free” noisy sequence, and whenever impulsive noise is detected, the impulse contaminated samples are replaced by predictive values. Based on simulation and experimental results, the proposed unsupervised robust recursive least-square adaptive filter performs as well as conventional RLS filters in “impulse-free” circumstances, and is effective in restricting large disturbances such as impulsive noise when the RLS and the more recent unsupervised adaptive filter fails.