Examination Committee

Prof Danny H K TSANG, ECE/HKUST (Chairperson)
Prof Khaled BEN LETAIEF, ECE/HKUST (Thesis Supervisor)
Prof Shaojie SHEN, ECE/HKUST

Abstract

As a promising technique to meet the dramatic growing demand for both high throughput and uniform coverage in the fifth generation (5G) wireless networks, massive multiple-input multiple-output (MIMO) systems have attracted significant attention in recent years. Accurate channel state information (CSI) is critical for the signal processing tasks in such systems. However, in massive MIMO systems, conventional uplink training methods to obtain CSI will incur prohibitively high training overhead, which is proportional to the number of MUs. Thus, innovative training techniques are necessary to improve the overall performance of multiuser massive MIMO systems.
 
In this thesis, we first propose a non-orthogonal pilot design based on the mean square error (MSE) minimization approach. In this method, the pilot length can be smaller than the number of MUs, i.e., the pilots are correlated, which significantly reduces the training overhead. We establish the relationship between the pilot correlation coefficients and the MSE, and then minimize the MSE by optimizing the pilot correlation coefficients. Numerical results show that the proposed non-orthogonal pilot design enjoys significant throughput improvement compared to the conventional methods. Moreover, when the MU density increases, the proposed training method obtains higher throughput gain.
 
Then, we propose a selective uplink training method for massive MIMO systems, where in each channel block only part of the MUs will send uplink pilots for channel training, and the channel states of the remaining MUs are predicted from the estimates in previous blocks, taking advantage of the channels’ temporal correlation. We propose an efficient algorithm to dynamically select the MUs to be trained within each block and determine the optimal uplink training length. Simulation results show that the proposed training method provides significant throughput gains compared to the existing methods, with much lower estimation complexity. It is observed that the throughput gain becomes higher as the MU density increases.