Full-field reconstruction from scanned measurements without references: the latent variable approach
Résumé
In inverse acoustics, the spatial resolution of the source field that can be reconstructed from microphone array measurements is fundamentally limited by the size of the array and the microphone density. One approach to mimic large arrays with high microphone density is to scan the object of interest by moving a small array around it. This typically requires that a number of fixed references are available in order to preserve the phase relationships between consecutive snapshots, or that measurements are made in the near-field in order to allow the successive reconstruction of patches of the source in front of the array by deliberately neglecting radiation of the source field not covered by the array. In the present work, a solution is introduced that can reconstruct the full source field of the radiating object from consecutive snapshots, but without the need for fixed references and without the edge effects of the patch approach. Under the stationary assumption, the problem is shown to boil down to the factorisation of a structured covariance matrix. This is undertaken within a Bayesian probabilistic approach, where the source field is encoded by latent variables which are iteratively reconstructed by using an Expected-Maximisation (EM) algorithm.
Origine : Fichiers éditeurs autorisés sur une archive ouverte
Loading...