We study optimal distributed detection strategies for wireless sensor networks under the assumption of spatially and temporally i.i.d. observations at the sensor nodes. Each node computes a local statistic and communicates it to a decision center over a noisy channel. The performance of centralized detection (noise-free channel) serves as a benchmark. We address the following fundamental question: under what network resource constraints can distributed detection achieve the same error exponent as centralized detection? Two types of constraints are considered: 1) transmission power constraints at the nodes, and 2) the communication channel between the nodes and the decision center. Two types of channels are studied: 1) a parallel access channel (PAC) consisting of dedicated AWGN channels between the nodes and the decision center, and 2) an AWGN multiple access channel (MAC). We show that for intelligent sensors (with knowledge of observation statistics) analog communication of local likelihood ratios (soft decisions) over the MAC is asymptotically optimal (for large number of nodes) when each node can communicate with a constant power. Motivated by this result, we propose an optimal distributed detection strategy for dumb sensors (oblivious of observation statistics) based on the method of types. In this strategy, each node appropriately quantizes its temporal observation data and communicates its type or histogram to the decision center. It is shown that type-based distributed detection over the MAC is also asymptotically optimal with an additional advantage: observation statistics are needed only at the decision center. Even under the more stringet total power constraint, it is shown that both soft decision- and type-fusion result in exponentially decaying error probability. 1

The problem of communicating sensor readings over a multiaccess channel for detecting a target using Type-Based Multiple Access (TBMA) is considered. TBMA is analyzed in a general framework by considering non-i.i.d. data and non-identical channel gains. An asymptotically optimal detector is proposed and its error-exponents for detection probabilities are characterized using tools from large deviations theory. In case of i.i.d. channel gains, it is shown that the performance of TBMA presents two distinct behaviors depending on whether the channel gains have zero mean. Numerical simulations are used to demonstrate that the error exponents provide reasonably accurate estimates of the performance of TBMA.

We study the problem of communicating sensor readings over a Gaussian multiaccess (MAC) channel. We focus on the scenario that each sensor observes a single random variable, and transmits it using certain signaling in a shared channel. The objective is the design of channel waveforms (i.e., signal constellation) to facilitate the estimation of field parameters from the channel output. We propose a new approach named Histogram-Delivering Multiple Access (HDMA). In case of symmetric channel gains, it is shown that the HDMA is asymptotically optimal in the limit of large number of sensors. In particular, we show that the HDMA together with a variant of the maximum-likelihood estimator achieves the Cramer-Rao lower bound asymptotically. We then compare the performance of HDMA with other approaches that allocate orthogonal channels to sensors such as TDMA.