Beams Selection for MmWave Multi-Connection Based on Sub-6GHz Predicting and Parallel Transfer Learning

Abstract

To meet the increasing wireless data demands, leveraging millimeter wave(mmWave) frequency band has become imperative for 5G systems due to the rich spectrum resources and greater bandwidth. In mmWave communication systems, multi-connection is an indispensable key technology, where the coordinated service of multiple links will enable users to get more wireless resources and ensure mobile robustness. However, mmWave multi-connections face challenges in beams selection process: (i) The time of multi-link serial search is long relative to single link, and the search overhead is large and the hardware complexity is high; (ii) In the case of multi-connection parallel transmission, the mutual interference between beams results in low multiplexing gain; (iii) The conventional codebook produces non-standard (non-pencil-shaped) beam shapes, which makes it difficult to reduce inter-beam interference only by relying on different codebooks. In response to the above problems, this paper uses sub-6GHz channel and deep neural network (DNN) to enhance beam search for mmWave multi-connection. Specifically, the spatial correlation between the low frequency band and the mmWave frequency band is exploited to map the sub-6GHz channel information to the mmWave beam index. To speed beams search process, a parallel deep neural network with transfer learning is proposed to predict the best beams for multi-links of a user. Simulation results show that the sub-6G Hz channel information can be used to effectively predict the optimal mmWave beams for multi-connected user, and the parallel transfer learning structure can facilitate in reducing interference and training overhead. As a result, near-optimal system sum-rate can be achieved.