A novel weight reset strategy for the LMS algorithm subject to abrupt channel variations

Abstract

We discuss a novel strategy to reset the weight vector of the Least Mean Square (LMS) adaptive algorithm when abrupt channel variations occur. Such variations may be caused by line impedance sudden changes, typically due to switching events, and may significantly impact the performance of the LMS algorithms used for adaptive echo cancellation and equalization. Specifically, they may trigger a convergence transient period, whose duration depends on the distance between the current weight vector and the final one. In previous work, a weight reset strategy based on sign inversion was demonstrated in the context of echo cancellation for a power line modem. In this paper, we extend this work by investigating weight reset strategies based on a group of reinitialization vectors, some of which will be likely close to the final LMS optimal weights. The results show that it may indeed be possible to promptly detect abrupt channel variations and properly reset the LMS weights in order to shorten the induced transient period.