The decision fusion in the wireless network with possible transmission errors (Q712690): Difference between revisions

The paper deals with data fusion problems in wireless sensor networks. In such networks data are gathered from distributed agents and are processed collectively at a fusion center. The main goal of the paper is to formulate and partly investigate the fault-tolerant fusion decision model, assuming erroneous channels. The authors propose a Bayesian approach to decision fusion and develop methods allowing them to achieve fault-tolerant decisions. They concentrate on the decentralized fusion problem in the presence of channel transmission errors, considering channels prone to fading effects. They generalize \textit{B. Chen} and \textit{P. K. Willett}'s [IEEE Trans. Inf. Theory 51, No. 2, 693--699 (2005; Zbl 1270.94008)]. The authors consider a parallel sensor network which makes decisions, denoted by H0 and H1, respectively, about the presence or absence of a phenomenon of interest. The sensors compress their observations to 0 in the case of H0, and 1 in the case of H1, and send them through a channel to the fusion center, where the final decision, based on \(l\) local sensors' transmissions, is taken according to a given rule. The authors assume that the appropriate conditional probability density functions are known. They consider and solve a two-sensor bi-decision problem under AND as well as OR fusion rules, using Bayesian cost as the criterion function and looking for conditions minimizing this cost. Next, the considerations are generalized to a multiple-bit \(m\)-ary decision problem. The authors show necessary conditions for optimal sensor compression rules that allow them to use these rules for designing iterative algorithms for specific (e.g., AND/OR) sensor rules. Both independent and correlated transmission bits are considered, relative to whether they are transmitted over the same or different transmission channels. Next, the authors provide some numerical results using receiver operating characteristics, and show the performance of their algorithms. Finally, they compare their results to earlier ones described in the literature.