Date of Award

December 2019

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical Engineering and Computer Science

Advisor(s)

Pramod K. Varshney

Keywords

Copula theory, Distributed inference, Multimodal data fusion, Statistical dependence, Wireless sensor networks

Subject Categories

Engineering

Abstract

Fusing heterogeneous data from multiple modalities for inference problems has been an attractive and important topic in recent years. There are several challenges in multi-modal fusion, such as data heterogeneity and data correlation. In this dissertation, we investigate inference problems with heterogeneous modalities by taking into account nonlinear cross-modal dependence. We apply copula based methodology to characterize this dependence.

In distributed detection, the goal often is to minimize the probability of detection error at the fusion center (FC) based on a fixed number of observations collected by the sensors. We design optimal detection algorithms at the FC using a regular vine copula based fusion rule. Regular vine copula is an extremely flexible and powerful graphical model used to characterize complex dependence among multiple modalities. The proposed approaches are theoretically justified and are computationally efficient for sensor networks with a large number of sensors.

With heterogeneous streaming data, the fusion methods applied for processing data streams should be fast enough to keep up with the high arrival rates of incoming data, and meanwhile provide solutions for inference problems (detection, classification, or estimation) with high accuracy. We propose a novel parallel platform, C-Storm (Copula-based Storm), by marrying copula-based dependence modeling for highly accurate inference and a highly-regarded parallel computing platform Storm for fast stream data processing. The efficacy of C-Storm is demonstrated.

In this thesis, we consider not only decision level fusion but also fusion with heterogeneous high-level features. We investigate a supervised classification problem by fusing dependent high-level features extracted from multiple deep neural network (DNN) classifiers. We employ regular vine copula to fuse these high-level features. The efficacy of the combination of model-based method and deep learning is demonstrated.

Besides fixed-sample-size (FSS) based inference problems, we study a distributed sequential detection problem with random-sample-size. The aim of the distributed sequential detection problem in a non-Bayesian framework is to minimize the average detection time while satisfying the pre-specified constraints on probabilities of false alarm and miss detection. We design local memory-less truncated sequential tests and propose a copula based sequential test at the FC. We show that by suitably designing the local thresholds and the truncation window, the local probabilities of false alarm and miss detection of the proposed local decision rules satisfy the pre-specified error probabilities. Also, we show the asymptotic optimality and time efficiency of the proposed distributed sequential scheme.

In large scale sensors networks, we consider a collaborative distributed estimation problem with statistically dependent sensor observations, where there is no FC. To achieve greater sensor transmission and estimation efficiencies, we propose a two-step cluster-based collaborative distributed estimation scheme. In the first step, sensors form dependence driven clusters such that sensors in the same cluster are dependent while sensors from different clusters are independent, and perform copula-based maximum a posteriori probability (MAP) estimation via intra-cluster collaboration. In the second step, the estimates generated in the first step are shared via inter-cluster collaboration to reach an average consensus. The efficacy of the proposed scheme is justified.

Access

Open Access

Included in

Engineering Commons

Share

COinS