AoA Estimation With Practical Antenna Arrays Using Neural Networks

We study the problem of estimating angle-of-arrival (AoA) using practical antenna arrays. In practical antenna arrays, the antenna elements have nonuniform radiation patterns. However, most existing AoA estimation algorithms are developed based on the assumption that antenna elements are isotropic sources with uniform radiation patterns. Hence, the performance of existing algorithms can be significantly degraded when they are deployed on practical antenna arrays. To overcome the difficulty from nonuniform radiation patterns, we propose a fundamentally different approach of AoA estimation, where we train a neural network to learn the mapping from received signals to AoAs. Since we can generate a large amount of training data from propagation models of electromagnetic waves, our neural network can achieve high estimation accuracy (e.g., 10% estimation errors), which greatly improves the performance of state-of-the-art algorithms (e.g., up to 30% estimation errors). Moreover, by training with data generated from different radiation patterns, our neural network is agnostic of the specific radiation patterns of antenna elements. Since most computation is done in the training phrase, our trained neural network can perform AoA estimation in real time and can be used for tracking AoAs of mobile targets.