Bayesian Estimation for CBRN Sensors with Non-Gaussian Likelihood

Abstract

Many sensors in chemical, biological, radiological, and nuclear (CBRN) applications only provide very coarse, integer outputs. For example, chemical detectors based on ion mobility sensing typically have a total of eight outputs in the form of bar readings. Non-Gaussian likelihood functions are involved in the modeling and data fusion of those sensors. Under the assumption that the prior distribution is a Gaussian density or can be approximated by a Gaussian density, two methods are presented for approximating the posterior mean and variance. The Gaussian sum method first approximates the non-Gaussian likelihood function by a mixture of Gaussian components and then uses the Kalman filter formulae to compute the posterior mean and variance. The Gaussian-Hermite method computes the posterior mean and variance through three integrals defined over infinite intervals and approximated by Gaussian-Hermite quadrature.