Go To Program
44

Self-Configuring System Identification via Evolutionary Frequency-Domain Adaptive Filters

Marcus Zeller 1 Walter Kellermann 1
1 Multimedia Communications and Signal Processing, University of Erlangen-Nuremberg, Erlangen, Germany

This paper proposes a novel strategy for adaptive frequency-domain filtering in arbitrary system identification tasks. Employing a convex combination of two competing filters with different size, the lengths of the components are dynamically adjusted, following a performance-based soft decision. Thus, a doubly-adaptive, evolutionary filter structure is realized that can find and track the optimum filter length so as to automatically match the adaptive model size to that of the unknown system. Regarding the proposed DFT-domain algorithm, a focus is put on the various implications on coefficient updates and memory organization for different choices of the length increment. Results for the effectiveness of the estimation and identification are shown for noise and speech signals and several systems in an AEC scenario.


View pdf