On the Use of Machine Learning for Direction Finding with Circular Monopole Antenna Arrays

Abstract

The training of a deep neural network for amplitude-only (AO) direction finding (DF) is presented. Four circular arrays with 3-, 4-, 6-, and 8-element uniformly distributed monopoles are considered. The impact of a center-placed metallic post on the performance of the arrays in conjunction with the utilized DF algorithm is investigated. The 1 GHz resonant arrays are first designed in Altair FEKO and voltages produced at the terminals of the antennas from incident plane waves are determined. These signals, at a signal-to-noise ratio of 20 dB, are then used to train the neural network for AO-DF. The best performing configuration, the 8-element array, has 1.8° root mean squared error and 1.33° average error across the 360° field of view. While the 3- and 4-element arrays without a post performed poorly, all other considered topologies yielded an average error less than 4°. These results allow system designers to make educated decisions in the context of a cost vs accuracy tradeoff.