Advanced statistical methods for analysis of NDE data

Abstract

<p>Nondestructive Evaluation (NDE) is a quality-ensuring technique widely used in modern industry. For example, ultrasonic inspection is a routine NDE method to detect flaws/defects in rotating components of jet engines. However, there are random factors that can affect the performance and reliability of NDE systems. Probability of detection (POD) is an important metric for quantifying NDE capability and reliability. The most commonly used POD assessment method is known as the a versus a method. However, the standard a versus a method can not be directly applied to some new modern NDE applications. The objective of this research is to (1) extend the a versus a method to handle bivariate response allowing for data censoring and truncation. (2) extend the standard method to adjust for bias in POD estimates due to flaw sizing errors. (3) develop a more complete understanding of inspection variability by identifying and quantifying the variance components in NDE operations;In Chapter 1, the standard a versus a method is extended to handle bivariate responses allowing for data censoring and truncation. In addition, for one inspection data, extra modelling efforts were made to accommodate the flaw misses that could not be directly accounted for by the bivariate a versus a model;In Chapter 2 of this thesis, we develop two statistical models for adjusting for bias in POD estimates that is caused by flaw sizing errors. We present the results of simulation studies that show how the use of our models will reduce flaw-sizing bias and we demonstrate the use of the methods with simulated inspection data based on the collected real inspection data;There are strong needs to identify and quantify variability sources in NDE applications, as such information is needed to properly decide on strategies to reduce inspection variability and thus to improve inspection quality. In the Chapter 3 of this thesis, we develop the Bayesian hierarchial model to identify and quantify the variance components of inspection in the presence of data censoring. The Bayesian approach is demonstrated with simulated data and experimental data. The computations use MCMC simulation implemented in the in WinBUGS software.</p>