Channel-aware distributed detection in wireless networks with correlated observations

Thesis (Ph. D.)--University of Rochester. Department of Electrical and Computer Engineering, 2017.We study data fusion in a distributed detection system, consisting of several sensors and a fusion center (FC), that is tasked with solving an underlying binary hypothesis testing problem (e.g., detection of a signal source or a target in a field being monitored). Each sensor makes a binary local decision based on its local observation, where these local decisions are digital modulated and transmitted over wireless channels to neighboring sensors and/or the FC. A global binary decision is made at the FC by fusing the data received from sensors. But due to additive Gaussian noise and multipath fading, effect of wireless communication channel, the binary local decisions are corrupted, causing the global decision to be less reliable. Our goal is to maximize the reliability of the global decision. We ask the question: what is the optimal distributed detection system design in the presence of multipath fading and additive Gaussian noise in wireless communication channel? To address this question in this thesis, we identify and address three subproblems as the following: P1) We propose a new class of integrated distributed detection, which harvests cooperative gain (enabled by at most 1-bit information exchange among one-hop neighboring nodes) and improves the performance of the integrated distributed detection in the presence of fading, via pushing the communication bounds. In particular, we propose three schemes: (i) cooperative fusion architecture with Alamouti’s space-time coding (STC) scheme at sensors, (ii) cooperative fusion architecture with signal fusion at sensors, and (iii) parallel fusion architecture with local threshold changing at sensors. For these schemes, we derive the likelihood ratio test (LRT) and majority fusion rules at the FC, and provide upper bounds on the average error probabilities for homogeneous sensors, subject to uncorrelated Gaussian sensing noise, in terms of signal-to-noise ratio (SNR) of communication and sensing channels. Our numerical results show that, when the FC employs the LRT rule, unless for low communication SNR and moderate/high sensing SNR, performance improvement is feasible with the new cooperative and parallel fusion architectures, while scheme (iii) outperforms others. When the FC utilizes the majority rule, such improvement is possible, unless for high sensing SNR. In particular, for very high sensing SNR scheme (i) outperforms, whereas for moderate/low sensing SNR scheme (ii) outperforms others. P2) We consider a binary hypothesis testing problem in an inhomogeneous wireless sensor network, where a fusion center (FC) makes a global decision on the underlying hypothesis. We assume sensors’ observations are correlated Gaussian and sensors are oblivious of such correlation when making decisions. Sensors send their modulated decisions over nonideal fading channels, subject to individual and/or total transmit power constraints. Considering parallelaccess channel (PAC) and multiple-access channel (MAC) models, we derive modified deflection coefficient (MDC) of the test statistic at the FC for coherent and non-coherent reception. We propose a transmit power allocation scheme, which maximizes MDC of the test statistic, under three different sets of transmit power constraints. When analytical solutions to our constrained optimization problems are elusive, we discuss how these problems can be converted to convex ones. We study how correlation among sensors’ observations, reliability of local decisions, communication channel model and channel qualities, transmit power constraints, and reception at the FC, affect the reliability of the global decision and power allocation of inhomogeneous sensors. P3) We consider a wireless sensor network that is tasked with detecting a known signal in correlated Gaussian noise. We consider two system-level constrained optimization problems for three transmission approaches, which are deterministic, randomized (i), and randomized (ii) transmission approaches. The observation space at each sensor is divided to three regions R-1, R0, R1. In deterministic approach the symbol transmitted to the FC is determined only by the region that the observation at each sensor belongs to. In randomized approaches, the transmitted symbol is determined based on the region and the realizations of two independent Bernouli random vari131 ables with two different parameters. For deterministic transmission approach, we have shown that, similar to previous work for detecting a known signal in uncorrelated Gaussian noise, the optimal three regions R-1, R0, R1 are three continuous and disjoint intervals. We have shown that if the correlation coefficient is high enough such that the sensors observations are in two consecutive intervals, then using either of the two randomized transmission approaches leads into an enhanced detection performance, compared with using deterministic transmission approach. Moreover, if transmit power is low enough such that the smaller censoring threshold is negative, using randomized transmission approach (ii) provides a higher detection performance, compared with using deterministic transmission approac