This paper demonstrates the use of deep learning and time series data generated from user equipment (UE) beam measurements and positions collected by the base station (BS) to enable handoffs between beams that belong to the same or different BSs. We propose the use of long short-term memory (LSTM) recurrent neural networks with three different approaches and vary the number of lookbacks of the beam measurements to study the performance of the prediction used for the proactive beam handoff. Simulations show that while UE positions can improve the prediction performance, it is only up to a certain point. At a sufficiently large number of lookbacks, the UE positions become irrelevant to the prediction accuracy since the LSTMs are able to learn the optimal beam based on implicitly defined positions from the time-defined trajectories.
翻译:本文展示了使用基地站收集的用户设备(UE)光束测量和位置产生的深学习和时间序列数据,以便能够在属于同一或不同BS的波束之间分配。我们提议使用三种不同方法的长期短期内存(LSTM)经常性神经网络,并改变光束测量的回顾次数,以研究用于主动波束交接的预测的性能。模拟显示,虽然UE位置可以改进预测性能,但只能达到某一点。在足够多的回望中,UE位置与预测的准确性无关,因为LSTMS能够从时间定义的轨迹中根据隐含的定型位置学习最佳波束。