Recurrent Neural Network Based Beam Prediction for Millimeter-Wave 5G Systems

• Main Authors: Khunteta, Shubham, Chavva, Ashok Kumar Reddy
• Format: Conference Proceeding
• Language: English
• Subjects: 5G mobile communication; angle of arrival; beam gain; Beamforming; device orientation; LSTM; machine learning; Millimeter wave measurements; Millimeter wave technology; millimeterwave; Performance evaluation; Recurrent neural networks; RNN; sensors; Spectral efficiency; Throughput
• ISSN: 1558-2612
• DOI: 10.1109/WCNC49053.2021.9417509

5G millimeter-wave (mmWave) system provides ultra low latency and higher peak data rate with a major drawback of higher path loss at mmWave spectrum. Multiple beams are formed at base station (BS) and user equipment (UE) to compensate excessive path loss. To help find the best beam pair for data transmission, beam measurements are performed continuously, typically in round robin fashion. Time taken for the measurement of full beam pair set can be large which results in delay in finding the best beam pair, which in turn results in poor data rate and link quality. In this paper, we analyse the key factors affecting signal strength of beam pairs such as device orientation and angle of arrival (AoA) at system level. Further, we propose a method to predict top-K candidates for the best beam pair using recurrent neural networks (RNN) with sensor data and beam measurements as inputs. We evaluate the performance of the proposed method with a performance metric showing the number of times the best beam pair is in top-K predicted candidates. Further, we show gain in the throughput by scheduling the predicted candidates for measurement compared with conventional scheduling. We show that for an UE changing its orientation even at the rate of 90 degree per second, best UE beam is in Top-5 predicted UE beams 99% of the times and gain in the throughput is more than 50% compared to conventional methods.