Wideband Machine-Learning-Based Amplitude-Only Direction Finding With Spiral Antennas

A method for performing amplitude-only direction finding (DF) with a multiarm spiral antenna is demonstrated over multioctave (1.5–6 GHz) bandwidth, with 360° coverage in azimuth and 100° total coverage in elevation. A combined frequency model and compact neural network (NN) architecture are deployed to perform DF in both azimuth and elevation. Root mean square error (RMSE) of about 5° at 40 dB signal-to-noise ratio (SNR) is demonstrated in measurement for a single-snapshot, machine-learning-based DF system, across two octaves of bandwidth. The wideband system is deployed with no RF beamforming or phase compensation hardware. A multichannel receiver is designed, manufactured, and integrated with an additively manufactured, cavity-backed spiral. An average RMSE of less than 2.5° is achieved in experiment with cascaded calibration and maintaining at least 30 dB SNR, while obtaining nearly 90% reduction in (digital) system footprint.