Neural networks for array processing: from DOA estimation to blind separation of sources

Gilles Burel (1) , N. Rondel (1)

1 Thomson CSF-LER - Laboratoires Electroniques de Rennes de THOMSON-CSF

In many signal processing applications, signals are received on an array of sensors, and the problem consists in estimating the Directions Of Arrival (DOA) of the signals, and/or in estimating the sources. Basically, the techniques proposed for its solution use either information about the geometry of the array, or information about the statistics of the sources. Efficient neural-based approaches for both kinds of situations are proposed in this paper. When geometrical knowledge is available, the weights and structure of the neural networks are constrained according to the geometry of the array. When statistical information is available, neural networks which optimize a statistical criterion (namely the measure of dependence) are developed. Furthermore, neural networks provide the opportunity to fuse both approaches in a unified framework, and to take profit simultaneously of both kind of information.

Mots clés: Array Processing; Blind Separationof Sources; Higher Order Moments; Neural Networks; Backpropagation

Domaines: Traitement du signal et de l'image [eess.SP]; Réseau de neurones [cs.NE]; Réseaux et télécommunications [cs.NI]

Gilles Burel, N. Rondel. Neural networks for array processing: from DOA estimation to blind separation of sources. IEEE Systems Man and Cybernetics Conference - SMC, Oct 1993, Le Touquet, France. pp.601-606, ⟨10.1109/ICSMC.1993.384940⟩. ⟨hal-03223336⟩