Identification of Models and Signals Robust to Occasional Outliers

Abstract

In this paper estimation algorithms derived in the sense of the least sum of absolute errors are considered for the purpose of identification of models and signals. In particular, off-line and approximate on-line estimation schemes discussed in the work are aimed at both assessing the coefficients of discrete-time stationary models and tracking the evolution of time-variant characteristics of monitored signals. What is interesting, the procedures resulting from minimization of absolute-error criteria appear to be insensitive to sporadic outliers in the processed data. With this fundamental property the deliberated absolute-error method provides correct results of identification, while the classical least-squares estimation produces outcomes, which are definitely unreliable in such circumstances. The quality of estimation and the robustness of the discussed identification procedures to occasional measurement faults are demonstrated in a few practical numerical tests.