A considerable amount of attention has been focused in recent years towards the development of probability of detection (POD) models for a variety of nondestructive evaluation (NDE) methods. Interest in these models is motivated by a desire to quantify the variability introduced during the process of testing. As an example, sources of variability involved in eddy current methods of NDE include those caused by variations in liftoff, material properties, probe canting angle, scan format, surface roughness and measurement noise. Numerical models have been extensively used to model physical processes underlying NDE phenomena. Such models have been used, for example, to predict the transducer response for a given specimen geometry, defect configuration and test conditions. These models, however, are deterministic in nature and do not take into account variabilities associated with the inspection carried out in the field. This has limited the utility of deterministic models to practitioners in general since a considerable difference can exist between the nominal value of the transducer response predicted by the model and the actual measurement.

This thesis presents a comprehensive POD model for eddy current NDE. Eddy current methods of nondestructive testing are used widely in industry to inspect a variety of nonferromagnetic and ferromagnetic materials. The development of a comprehensive POD model is therefore of significant importance. The model incorporates several sources of variability characterized by a multivariate Gaussian distribution and employs finite element analysis to predict the signal distribution. The method of mixtures is then used for estimating optimal threshold values. The research demonstrates the use of a finite element model within a probabilistic framework to predict the spread in the measured signal for eddy current nondestructive methods.

Using the signal distributions for various flaw sizes the POD curves for varying defect parameters have been computed. In contrast to experimental POD models, the cost of generating such curves is very low and complex defect shapes can be handled very easily. The results are also operator independent.