Degree Type

Dissertation

Date of Award

2016

Degree Name

Doctor of Philosophy

Department

Aerospace Engineering

Major

Aerospace Engineering

First Advisor

Stephen D. Holland

Abstract

This thesis tackles the challenge of using nondestructive evaluation data as a sensor measurement input for a state estimation scheme in order to estimate the current state a part that is changing over time. This is made particularly challenging because of the multidisciplinary nature of the problem. The estimation solution incorporates work from statistics, computer vision, and the physics of nondestructive evaluation.

This thesis discusses the basis for spatio-temperal Kalman filtering and uses a simple version of spatio-temperal filtering to simulate material state tracking on composite spec- imens. The simulation illustrates that by using the algorithm presented here, even with very naive inputs, it is possible to track a dynamic material state and provide estimates that better reflect the true state of the part as compared with the most recent sensor measurement alone.

Additionally, this thesis demonstrates the algorithm on laboratory test data. Com- posite panels were manufactured and then intentionally impacted to induce subsurface delaminations. The composite panels were then loaded multiple times in four-point bending to induce incremental damage growth. After each damage event (initiation and loading) flash thermography and computed tomography data was collected. The flash thermography data was used as a sensor measurement in the spatio-termporal Kalman filter and the computed tomography data was used as a ‘truth' value for comparison. For four out of five data sets, at every time step, the spatio-temporal Kalman filter estimate matched the computed tomography ‘truth' better than the most recent single sensor measurement. For the fifth data set, the estimate better matched the truth at most time steps.

Finally, we present a method that uses a probabilistic approach to identifying the location and orientation, or pose, of a specimen or part within an image. This process found the most likely transformation from the object coordinate frame to the camera coordinate frame but also ranked less probable transformations by likelihood. The discus- sion is continued with an exploration of different, non-Gaussian uncertainty distributions that result from the process of registering two dimensional images to three dimensional part models. A method for approximating non-Gaussian distributions using Gaussian mixtures is presented and discussed.

The work presented in this thesis successfully demonstrates that Bayesian estimation with nondestructive evaluation data will provide superior and more meaningful state estimates while discussing the issues that must be considered in doing this estimation properly.

Copyright Owner

Elizabeth Dimmitt Gregory

Language

en

File Format

application/pdf

File Size

124 pages

Share

COinS