Neural Networks and Dictionary Learning for Compressive, Phase-less, mm-Wave Beam Alignment

Fast, efficient, and scalable initial beam alignment algorithms with large antenna arrays are critical for mm-wave and sub-THz communications. With low-power mobile devices requiring phased array architectures and phase-less power measurements during the initial connection with a base station, traditional beam alignment methods require exhaustive sweeps of all possible beam directions and scale poorly with large arrays. Compressed sensing methods can better scale overhead with array size but are unreliable without perfect array calibration. We first discuss our neural-network-based beam-alignment algorithm, mmRAPID, that reduces measurement overhead even with poor factory array calibration. We then introduce a systematic method to optimize fixed compressive measurement beams for beam alignment with phase-less measurements and array mismatch. This “codebook learning” algorithm is based on the compressive sensing concept of dictionary learning and improves heuristic sounding beam designs, especially when paired with mmRAPID.