Probabilistic Neural Network-based Sensor Configuration Management in a Wireless Ad-hoc Network

This paper describes a novel application of a probabilistic neural network for overcoming the computational complexity involved in performing sensor configuration management in a collaborative sensor network. We consider the problem of reliably tracking a target moving through a field of stationary sensors by fusing the measurements returned from the distributed array of sensors while conserving power by minimizing the number of sensors participating in the decision-making at each step, which is a challenging problem of significant current interest. The twin, and often conflicting, requirements of high tracking accuracy (achievable by recruiting more sensors in order to develop fused decisions) and minimization of network latency (performing decisions using measurements from only a small subset of sensors) place a major emphasis on developing optimal strategies for sensor configuration management in such application scenarios. Recently suggested approaches to this problem typically employ Bayesian Networks and Influence Diagrams, which are computationally intensive and are often prohibitive for real time applications, particularly when the number of sensors involved is large. To overcome the computational complexity, we propose the use of a probabilistic neural network (PNN). The task for the PNN is to produce a distance measure (a radius, for instance) about a target location estimate within which to query sensors for observations by using the previous state estimate of the target as input. By integrating the PNN with a particle filter implementation of a tracking algorithm, we develop a collaborative distributed tracking scheme. Performance evaluation results are presented to demonstrate the benefits from sensor fusion (improvement of tracking accuracy) and reduction of latency (saving in the number of sensors deployed for accomplishing the task) in chosen tracking scenarios.