An improved recursive least squares algorithm robust to input power variation

This paper proposes a new recursive least-squares adaptive algorithm that improves the steady-state performance of the recently proposed Variable Memory Length (VML) algorithm. Most RLS-type algorithms tend to increase the cri-or in the estimated weight vector during reduced power situations. Like VML, the new algorithm, called Robust VML (RVML), is robust in system identification applications in which the input power is significantly reduced during operation. The RVML algorithm, however, improves the robustness of the VML algorithm when there is significant input power variations during convergence. It should encounter application in systems such as automotive suspension fault detection and adaptive control, and system identification using speech signals. In both cases, considerable periods of power variation during operation are common.