Feature Engineering for DOA Estimation using a Convolutional Neural Network, for Sparse Arrays
- Creators
- Kulkarni, Pranav
- Vaidyanathan, P. P.
Abstract
In the past few years, there has been an emerging use of deep neural networks for improving the direction of arrival (DOA) estimation performance. This paper demonstrates how such methods can be applied for sparse arrays such as nested arrays, by adapting a recent method based on convolutional neural network (CNN). Many possible alternative inputs (proxy spectra) to the network are suggested here, and experiments show that even simple modifications of the input lead to improved DOA estimation performance without changing the network structure. Additionally, the experiments also show that, with the modified input proxy spectra it is possible to identify more sources than the number of physical sensors, as one would expect with nested arrays. This opens up further avenues of using coarray principles in conjunction with machine learning methods.
Additional Information
© 2021 IEEE. This work was supported in parts by the ONR grant N00014-21-1-2521, and the California Institute of TechnologyAdditional details
- Eprint ID
- 113936
- Resolver ID
- CaltechAUTHORS:20220317-376187000
- N00014-21-1-2521
- Office of Naval Research (ONR)
- Caltech
- Created
-
2022-03-18Created from EPrint's datestamp field
- Updated
-
2022-03-18Created from EPrint's last_modified field