Kernel-based learning of decision fusion in wireless sensor networks

The problem of decision fusion in wireless sensor networks for distributed detection applications has mainly been considered in scenarios where sensor observations are conditionally independent and both local sensor statistics as well as wireless channel conditions are available for fusion rule design. In this paper, kernel-based learning algorithms for the design of decision fusion rules are presented when no such prior knowledge is available. The fusion center receives a collection of labeled decision vectors from the sensor nodes and employs a discrete version of the method of kernel smoothing which exploits the ordinal nature of local sensor decisions. The aim is to arrive at fusion rules which are Bayes risk consistent, i.e., asymptotically optimal as the number of training samples tends to infinity. The kernel-based learning approach is applied to the problem of distributed detection of a deterministic signal in correlated Gaussian noise. Numerical results obtained by simulation show that the kernel-based fusion rules show good performance also for finite sample sizes.