Data-driven neural networks for source localization and reconstruction using a planar array

Abstract

This study uses specialized deep neural networks comprising dense and convolutional neural networks to localize noise sources and reconstruct acoustic data on a reconstruction plane. The networks are trained on simulated acoustic data free from any form of noise in the signal. It is observed that neural networks can effectively localize monopole and dipole sources and reconstruct the acoustic data in reconstruction planes with higher accuracy than conventional methods. Performance of the networks is consistent over changes in some parameters like the source strength, noise in the input signal, and frequency range. Various tests are performed to assess the individual network performance. Results indicate that neural networks trained on a subset of the data are effective over the entire data set without significant bias or variance. Errors as low as 1% are observed, and the maximum error observed is below 5%. While reconstruction error decreased with an increase in the frequency of monopole sources, it increased with an increase in frequency for dipole sources.