Open Collections

Wu, Pengxia

University of British Columbia

Sparse channel estimation has great potential to address the challenges caused by the large pilot training overhead and the high dimensionality of the estimating channels when acquiring downlink channel state information (CSI) for massive multiple-input multiple-output (MIMO) systems. In practice, the conventional sparse channel estimation schemes employing compressive sensing techniques are unable to achieve satisfactory performance due to the sub-optimality when designing a random measurement matrix and the inherent inefficiency of iterative reconstruction algorithms. This thesis aims to improve sparse channel estimation by incorporating deep learning techniques into the classical compressive sensing framework to optimize the design of the measurement matrix and sparse channel reconstruction algorithms. First, novel sparse reconstruction algorithms are proposed. The sparse reconstruction problem is reformulated as a least-squares regression penalized by a trimmed l1-norm regularizer, which removes penalties on several large-magnitude elements and thus can lead to more accurate solutions compared to the widely-used Lasso regularizer. The difference of convex functions programming framework and gradient projection descent are applied to develop sparse reconstructions algorithms. The proposed algorithms have improved accuracy and require a shorter runtime compared to existing reconstruction algorithms. After developing the novel sparse reconstruction algorithms, we improve the measurement matrix design by using a data-driven method. The measurement matrix is closely associated with the pilot design and is crucial to channel estimation performance. Several model-based autoencoders are proposed to acquire data-driven measurement matrices. When compared with the existing random matrices, the acquired data-driven measurement matrices can achieve more accurate reconstructions and consume fewer pilots, thereby allowing higher achievable rates in massive MIMO systems. After optimizing the measurement matrix and the reconstruction individually, we propose a data-driven scheme to optimize jointly the pilot design and sparse reconstruction. The trimmed ridge regression, i.e., a least-squares regression penalized by a trimmed l2 -norm regularizer, is proposed for sparse reconstructions. The gradient projection descent algorithms of trimmed ridge regression are derived and then unfolded into deep networks to construct model-based autoencoders for joint data-driven pilot acquisition and channel reconstructions. Compared with other deep learning methods, the proposed autoencoders can achieve faster channel reconstruction with higher accuracy using fewer pilots.

Text

2022-01-05

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/80512

Electrical Engineering

University of British Columbia

UBCO

http://creativecommons.org/licenses/by-nc-nd/4.0/