Statistical Learning Methods for Neutron Transmission Tomography of Fuel Cells

Published

July 8, 2008

Author(s)

Kevin Coakley, Dominic F. Vecchia, Daniel S. Hussey

Abstract

In a fuel cell, water is formed as a by-product of the reaction between hydrogen and oxygen. As a neutron beam passes through a fuel cell, it can undergo s-wave scattering. For the dry and wet states of a fuel cell, there are spatially varying neutron attenuation images. Ideally, the difference of these attenuation images is proportional to the water density in the fuel cell. We estimate a nonnegative residual attenuation image from joint analysis of the wet and dry state projection data using a penalized likelihood (PL) method with a Huber penalty function that has two adjustable regularization parameters. We also reconstruct using a Filtered Back Projection (FBP) method where we apodize the ramp filter by a window corresponding to a Gaussian kernel with an adjustable standard deviation. We determine the adjustable regularization parameters by a statistical learning method called two-fold cross-validation. We simulate projection data based on Hussey phantom images and neglect beam hardening effects, backgrounds, and model the number of detected photoelectrons as a compound Poisson process. For simulated data, the PL approach yields sharper reconstructions than does the FBP approach.

Proceedings Title

Proceedings of 8th World Conference on Neutron Radiography.

Conference Dates

October 16-19, 2006

Conference Location

Gaithersburg, MD, USA

Pub Type

Conferences

Keywords

cross-validation, filtered back projection, fuel cells, neutron transmission tomography, penalized maximum likelihood, statistical learning.

Citation

Coakley, K. , Vecchia, D. and Hussey, D. (2008), Statistical Learning Methods for Neutron Transmission Tomography of Fuel Cells, Proceedings of 8th World Conference on Neutron Radiography., Gaithersburg, MD, USA (Accessed March 13, 2025)