文摘
Nowadays,the advancement in scientific technologies has provided the information and technological platforms for many engineering problems that involve massive sensing and high dimensional inferences,especially those associated with a large network. In particular,concrete applications in wireless networks and power systems are considered in this dissertation,both of which entail wide area infrastructure and scalable data processing. Specifically,my doctoral research casts them as generalized theoretical problems and proposes practical solutions by leveraging recent discoveries and theories made in compressive sensing sparse recovery) and distributed optimizations. Compressive sensing has been proven extremely useful in many applications for capturing and recovering sparse signals. Its main goal is to minimize sampling overhead and to reduce computational complexities at a processing terminal. Although fundamental limits and practical algorithms are well understood for such design at a single terminal,the task of efficiently managing and processing network-wide data in a parsimonious fashion still presents several open questions. Hence,this study further deals with the sensing and management of large scale distributed data sources in a network,considering both the compressive acquisition of local data at individual terminals,as well as the fusion and processing of such data across the network on a higher level. The ultimate goal is to equip the network with low complexity data sensing front-end at a sub-system level,an efficient network-wide fusion and inference scheme or in-network data processing scheme for the purpose of scalable and robust computations of data analytics. Thus,this dissertation is divided into three parts. The first part of this dissertation deals with data sensing problems at network nodes,in particular on the design of wireless receivers. For wireless receivers,one of the most demanding tasks for a terminal is to receive communication signals,which today needs to be sensed in a broad spectrum and processed at an exceedingly high rate. Thus,this dissertation studies one of the most challenging sensing applications related to synchronization and communications with distributed sensing devices,and presents low-rate solutions for multiuser signal detection/synchronization as well as GPS signal acquisition using only sub-Nyquist samples of the analog signals. To reflect the practical gains and losses of using this popular technique,the advantages of the proposed solution are also analyzed in terms of complexity vis-a-vis other common alternatives. An interesting conclusion is that one can use compressive sensing to scale down storage requirement and processing complexity with greater flexibility than conventional architectures,even though in general they both have overall complexities that scale linearly with the search space. The second part of this dissertation considers another critical problem in network systems consisting of numerous terminals,which is the fusion of distributed information gathered from these terminals for various purposes. This dissertation studies the specific tasks of network anomaly detection and network topology identification respectively. For network anomaly detection,the network does not have any prior knowledge of the anomaly and typical distributions and wants to identify the anomalous patterns from a large population without performing combinatorial composite hypothesis tests. Thus,a practical scheme called ``Sparsity-Aware Least Squares Anomaly SALSA) detector that leverages upon the sparsity of anomalies is then proposed to identify the data pattern of each agent network sensor) and detect the presence and locations of anomalies by polling the agents strategically. For network topology identification,a specific scenario of power systems is considered,where the power network topology is to be determined from the power consumption data gathered at each network node. Rigorous analysis is performed to shed light on the identifiability issues given the limited data available for such inferences,and a sparsity-regularized approach with graphical constraints is used to learn the network topology effectively. The last part of the dissertation is focused on solving large scale data processing problems. Different from the first two parts,we move from the sub-system level to the network/application level,where we consider the problem of processing distributed data efficiently without aggregating them at one place for fusion. A motivating example associated with massive data sensing is the task of power grid monitoring,which calls for sensor information gathering and fusion over large scale networks in a robust,scalable and distributed manner. First,a gossip-based Gauss-Newton algorithm with convergence guarantees is proposed to solve general non-linear least squares NLLS) problems in a network with an information architecture that suits the need of decentralized sensor fusion in power systems. It is then further extended in an adaptive setting to also robustify its performance against bad data injections. Furthermore,it is shown that with the appropriate deployment of Phasor Measurement Units PMUs),one can ensure the algorithm convergence even in the face of sensing errors and random failures with only local knowledge of the physical and cyber networks. Last but not least,the performance of the proposed scheme is compared with existing decentralized approaches,followed by discussions of future works on other sensing and processing methods for other smart grid applications.
