Interpretation of NDE data can be unreliable and difficult due to the complex interaction
between the instrument, object under inspection and noise and uncertainties about the system or
data. A common method of reducing the complexity and volume of data is to use thresholds.
However, many of these methods are based on making subjective assessments from the data or
assumptions about the system which can be source of error. Reducing data whilst retaining
important information is difficult and normally compromises have to be made. This thesis has
developed methods that are based on sound mathematical and scientific principles and require
the minimum use of assumptions and subjective choices.
Optimisation has been shown to reduce data acquired from a multilayer composite panel and
hence show the ply layers. The problem can be ill-posed. It is possible to obtain a solution close
to optimum and obtain confidence on the result. Important factors are: the size of the search
space, representation of the data and any assumptions and choices made. Further work is
required in the use of model based optimisation to measure layer thicknesses from a metal
laminate panel. A number of important factors that must be addressed have been identified.
Two novel approaches to removing features from Transient Eddy-Current (TEC) data have been
shown to improve the visibility of defects. The best approach to take depends on the available
knowledge of the system.
Principal Value Decomposition (PVD) has been shown to remove layer interface reflections
from ultrasonic data. However, PVD is not suited to all problems such as the TEC data
described. PVD is best suited in the later stages of data reduction.
This thesis has demonstrated new methods and a roadmap for solving multivariate problems,
these methods may be applied to a wide range of data and problems.