Performance analysis of deep neural networks for direction of arrival estimation of multiple sources

Recently, popular machine learning algorithms have successfully been applied to the direction of arrival (DOA) estimation. An implementation of determination of DOA estimation is presented based on deep neural networks (DNNs) to reduce the computational complexity of traditional superresolution DOA estimation methods. The classical DOA estimation algorithms have limitations due to unforeseen effects, such as array perturbations. Instead of computing an inverse mapping based on the incomplete forward mapping that relates the signal directions to the array outputs, the DOA problem is approached as a mapping, which can be approximated using a suitable DNN trained with input output pairs. The neural network architecture is based on a multilayer perception and a group of parallel DNNs to perform detection and DOA estimation, respectively. Simulation results are performed to investigate the effect of network parameters on estimation accuracy so that they can be roughly determined in the case of one signal scenario. Based on a set of simulations and experimental measurements, the performance of the optimum network is also assessed and compared to that of the classical DOA estimation methods for multiple signals. It has been shown that the proposed method can not only achieve reasonably high DOA estimation accuracy, but also dramatically reduce the computational complexity and the memory space.

Automated summary

Recently, machine learning algorithms have been successfully applied to direction of arrival (DOA) estimation. A deep neural network (DNN) based implementation is proposed to reduce the computational complexity of traditional superresolution DOA estimation methods. Instead of using inverse mapping, the DOA problem is approached as a mapping that can be approximated using a suitable DNN trained with input-output pairs. The proposed method achieves high accuracy in DOA estimation while reducing computational complexity and memory space.