Improving DOA Estimation via an Optimal Deep Residual Neural Network Classifier on Uniform Linear Arrays

Abstract

The main objective of this work is to improve and evaluate the effectiveness of the neural network (NN) architecture in the domain of estimation of direction of arrival (DOA), with an emphasis on a multi-class classification task with grid resolutions of 0.25 and 0.1. Specifically, a comprehensive assessment is performed to determine the competence of a residual NN (ResNet) in predicting the angle of arrival (AOA) of intercepted signals. Such signals are received by a 16-element uniform linear array and are subjected to real-world noise conditions. To this end, the superiority of the ResNet architecture in DOA estimations is substantiated through a comparison analysis with two other highly recognized NNs, namely, the feed-forward NN and the convolutional NN. Numerical results indicate that the ResNet model exhibits notable precision in estimating the AOAs, across various classes within a broad spectrum, along with a rapid temporal response. Finally, it remains consistent and maintains its superior performance even for diverse incoming signals and significantly reduced SNRs.

Keywords

• synthesis
• antenna optimization
• machine learning
• direction-of-arrival (DOA) estimation
• residual neural networks