Adaptive neural networks for robust estimation of signal parameters

Tadeusz Łobos, Paweł Kostyła, Zbigniew Wacławek, Andrzej Cichocki

Research output: Contribution to journalArticlepeer-review

4 Citations (Scopus)


In many applications, very fast methods are required for estimating of parameters of harmonic signals distorted by noise. Most of the known digital algorithms are not fully parallel, so that the speed of processing is quite limited. In this paper new parallel algorithms are proposed, which can be implemented by analogue adaptive circuits employing some neural networks principles. Algorithms based on the least-squares (LS) and the total least-squares (TLS) criteria are developed and compared. The problems are formulated as optimization problems and solved by using the steepest descent continuous-time optimization algorithm. The corresponding architectures of analogue neuron-like adaptive processors are also shown. The developed networks are more robust against noise in the measured signal than other known neural network algorithms. The network based on the TLS criterion optimizes the estimation under the assumption that the signal model can also be perturbated (frequency or sampling interval fluctuation and so forth). The TLS estimates are better and more reliable than the corresponding LS estimates, when applying a higher sampling frequency and a wider sampling window. The TLS algorithm is a generalization of the well known LMS rule and could be in some applications superior to the family of LMS algorithms. Extensive computer simulations confirm the validity and performance of the proposed algorithms.

Original languageEnglish
Pages (from-to)903-912
Number of pages10
JournalCOMPEL - The International Journal for Computation and Mathematics in Electrical and Electronic Engineering
Issue number3
Publication statusPublished - 2000
Externally publishedYes


  • Estimation
  • Neural networks
  • Signal processing


Dive into the research topics of 'Adaptive neural networks for robust estimation of signal parameters'. Together they form a unique fingerprint.

Cite this