Doctor of Philosophy (PhD)


Electrical and Computer Engineering

Document Type



Advances in microelectronics, communication and signal processing have enabled the development of inexpensive sensors that can be networked to collect vital information from their environment to be used in decision-making and inference. The sensors transmit their data to a central processor which integrates the information from the sensors using a so-called fusion algorithm. Many applications of sensor networks (SNs) involve hypothesis testing or the detection of a phenomenon. Many approaches to data fusion for hypothesis testing assume that, given each hypothesis, the sensors' measurements are conditionally independent. However, since the sensors are densely deployed in practice, their field of views overlap and consequently their measurements are dependent. Moreover, a sensor's measurement samples may be correlated over time. Another assumption often used in data fusion algorithms is that the underlying statistical model of sensors' observations is completely known. However, in practice these statistics may not be available prior to deployment and may change over the lifetime of the network due to hardware changes, aging, and environmental conditions. In this dissertation, we consider the problem of data fusion in heterogeneous SNs (SNs in which the sensors are not identical) collecting dependent data. We develop the expectation maximization algorithm for hypothesis testing and model estimation. Copula distributions are used to model the correlation in the data. Moreover, it is assumed that the distribution of the sensors' measurements is not completely known. we consider both parametric and non-parametric model estimation. The proposed approach is developed for both batch and online processing. In batch processing, fusion can only be performed after a block of data samples is received from each sensor, while in online processing, fusion is performed upon arrival of each data sample. Online processing is of great interest since for many applications, the long delay required for the accumulation of data in batch processing is not acceptable. To evaluate the proposed algorithms, both simulation data and real-world datasets are used. Detection performances of the proposed algorithms are compared with well-known supervised and unsupervised learning methods as well as with similar EM-based methods, which either partially or entirely ignore the dependence in the data.



Committee Chair

Naraghi-Pour, Mort