A neural network approach to broadband beamforming

Beamforming techniques are commonly applied to signals captured by sensor arrays to enhance signals receivedfrom desired directions while reducing background noise and localized interference. Where the directions of thedesired and interfering sources are known, this knowledge, combined with assumptions on the background noisecharacteristics, is used to derive the beamformer coefficients for each sensor. Usually, this is done by optimizing thesecond-order statistics of the beamformer response, e.g., minimizing the energy of the output signal while preserv-ing the signals from desired directions. The beamformer coefficients areindependentlyderived for each discretefrequency, as an approximation to the true broadband response. Hereby, the complex inter-frequency interactions,e.g., due to windowing and spectral aliasing, are not modelled, leading to sub-optimal filter characteristics. Fur-thermore, for standard designs, the mainlobe also narrows with frequency, leading to a non-uniform beamwidth.These shortcomings can be overcome by data-driven approaches. As a first attempt towards such approaches, wefocus on the problem ofuniformbeamwidth and propose a beamforming neural-network architecture. We comparethe spatial characteristics of such an architecture to standard (second-order) beamformers.

ISBN:9783939296157

Jaar van publicatie:2019

Toegankelijkheid:Open