Distributed chemical plume process detection: MILCOM 2005 #1644

We have developed a new approach for detecting and tracking chemical or biological plumes in distributed sensor networks, with the objective of solving the inverse location problem. The canonical plume tracking problem suffers from the challenge of a large state space, and we seek reduced dimensionality using information theoretic and stochastic methods. Working with an airborne plume we model a plume using multiple hypothesis tracking (MHT) techniques as opposed to transport based methods rooted in solutions to differential equations. The simple plume model attributes include: diffusion constant, wind direction, and wind magnitude. The location of the plume prior to current observations is calculated statistically with the use of an estimator-based joint probability. The main contribution of this work is the predictor model - a required step of the MHT algorithm. A customized predictor for plumes (as opposed to Kalman filtering) allows the MHT-like algorithm to treat the plume tracking problem as the extreme instance of the multi-target tacking (MTT) problem. The central question: how can a MHT-like method be implemented for plumes in a sensor network of simple sensors capable of rudimentary binary detection, wind speed, and wind direction. The predictor must handle the problem of data association for plume observations. The context for this work is the development of multiple competing models which is correlated to incoming observations in real time. The models run in a generic multi-purpose framework called PQS (process query system). Simulations were performed demonstrating the viability of the MHT approach with the use of a customized predictor for plume target tracking