DSpace DSpace

University of Minnesota Digital Conservancy >
University of Minnesota - Twin Cities >
Dissertations and Theses >
Dissertations >

Please use this identifier to cite or link to this item: http://hdl.handle.net/11299/139897

Title: Sparsity-cognizant algorithms with applications to communications, signal processing, and the smart grid.
Authors: Zhu, Hao
Keywords: Compressive sampling
Lasso
Multiuser detection
Power system analysis
Issue Date: Aug-2012
Abstract: Sparsity plays an instrumental role in a plethora of scientific fields, including statistical inference for variable selection, parsimonious signal representations, and solving under-determined systems of linear equations - what has led to the ground-breaking result of compressive sampling (CS). This Thesis leverages exciting ideas of sparse signal reconstruction to develop sparsity-cognizant algorithms, and analyze their performance. The vision is to devise tools exploiting the ‘right’ form of sparsity for the ‘right’ application domain of multiuser communication systems, array signal processing systems, and the emerging challenges in the smart power grid. Two important power system monitoring tasks are addressed first by capitalizing on the hidden sparsity. To robustify power system state estimation, a sparse outlier model is leveraged to capture the possible corruption in every datum, while the problem nonconvexity due to nonlinear measurements is handled using the semidefinite relaxation technique. Different from existing iterative methods, the proposed algorithm approximates well the global optimum regardless of the initialization. In addition, for enhanced situational awareness, a novel sparse overcomplete representation is introduced to capture (possibly multiple) line outages, and develop real-time algorithms for solving the combinatorially complex identification problem. The proposed algorithms exhibit nearoptimal performance while incurring only linear complexity in the number of lines, which makes it possible to quickly bring contingencies to attention. This Thesis also accounts for two basic issues in CS, namely fully-perturbed models and the finite alphabet property. The sparse total least-squares (S-TLS) approach is proposed to furnish CS algorithms for fully-perturbed linear models, leading to statistically optimal and computationally efficient solvers. The S-TLS framework is well motivated for grid-based sensing applications and exhibits higher accuracy than existing sparse algorithms. On the other hand, exploiting the finite alphabet of unknown signals emerges naturally in communication systems, along with sparsity coming from the low activity of each user. Compared to approaches only accounting for either one of the two, joint exploitation of both leads to statistically optimal detectors with improved error performance.
Description: University of Minnesota Ph.D. dissertation. August 2012. Major: Electrical/Computer Engineering. Advisor: Georgios B. Giannakis. 1 computer file (PDF0; x, 138 pages, appendices A-C.
URI: http://purl.umn.edu/139897
Appears in Collections:Dissertations

Files in This Item:

File Description SizeFormat
Zhu_umn_0130E_13001.pdf1.45 MBPDFView/Open

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.