Enhanced Angle of Arrival (AoA) Prediction through Neural Networks and Primary and Adjacent Antennas Representation Transformation (PAAR)

In this study, we present an enhanced Angle of Arrival (AoA) prediction method using neural networks and the Primary and Adjacent Antennas Representation (PAAR) transformation. PAAR leverages rotational symmetry in segmented antenna data, to transform the data into a rotation-invariant form. Thus, PAAR improves predictive accuracy and stability, especially with limited training data. We conducted experiments with two Digital Video Broadcasting - Terrestrial (DVB-T) datasets and evaluated four different neural network models, each varying in parameter size. Results demonstrate that the PAAR method clearly outperforms the traditional unprocessed approach, which we refer to as the plain approach, in data-limited scenarios, reducing the mean absolute angular error (MAAE) by up to $\mathbf{4 0 \%}$. However, with extensive training data, the plain approach can surpass PAAR due to error propagation. The study demonstrates that the PAAR method effectively enhances AoA prediction, especially with sparse training data.