Date of Award

December 2015

Degree Type


Degree Name

Doctor of Philosophy (PhD)


Electrical Engineering and Computer Science


Pramod K. Varshney


Copula Theory, Dependent observations, Detection, Estimation, Sensor Fusion

Subject Categories



Inferring events of interest by fusing data from multiple heterogeneous sources has been an interesting and important topic in recent years. Several issues related to inference using heterogeneous data with complex and nonlinear dependence are investigated in this dissertation. We apply copula theory to characterize the dependence among heterogeneous data.

In centralized detection, where sensor observations are available at the fusion center (FC), we study copula-based fusion. We design detection algorithms based on sample-wise copula selection and mixture of copulas model in different scenarios of the true dependence. The proposed approaches are theoretically justified and perform well when applied to fuse acoustic and seismic sensor data for personnel detection. Besides traditional sensors, the access to the massive amount of social media data provides a unique opportunity for extracting information about unfolding events. We further study how sensor networks and social media complement each other in facilitating the data-to-decision making process. We propose a copula-based joint characterization of multiple dependent time series from sensors and social media. As a proof-of-concept, this model is applied to the fusion of Google Trends (GT) data and stock/flu data for prediction, where the stock/flu data serves as a surrogate for sensor data.

In energy constrained networks, local observations are compressed before they are transmitted to the FC. In these cases, conditional dependence and heterogeneity complicate the system design particularly. We consider the classification of discrete random signals in Wireless Sensor Networks (WSNs), where, for communication efficiency, only local decisions are transmitted. We derive the necessary conditions for the optimal decision rules at the sensors and the FC by introducing a "hidden" random variable. An iterative algorithm is designed to search for the optimal decision rules. Its convergence and asymptotical optimality are also proved. The performance of the proposed scheme is illustrated for the distributed Automatic Modulation Classification (AMC) problem. Censoring is another communication efficient strategy, in which sensors transmit only "informative" observations to the FC, and censor those deemed "uninformative". We design the detectors that take into account the spatial dependence among observations. Fusion rules for censored data are proposed with continuous and discrete local messages, respectively. Their computationally efficient counterparts based on the key idea of injecting controlled noise at the FC before fusion are also investigated.

In this thesis, with heterogeneous and dependent sensor observations, we consider not only inference in parallel frameworks but also the problem of collaborative inference where collaboration exists among local sensors. Each sensor forms coalition with other sensors and shares information within the coalition, to maximize its inference performance. The collaboration strategy is investigated under a communication constraint. To characterize the influence of inter-sensor dependence on inference performance and thus collaboration strategy, we quantify the gain and loss in forming a coalition by introducing the copula-based definitions of diversity gain and redundancy loss for both estimation and detection problems. A coalition formation game is proposed for the distributed inference problem, through which the information contained in the inter-sensor dependence is fully explored and utilized for improved inference performance.


Open Access

Included in

Engineering Commons