The design of a neural network that performs a complex mapping for phase-sensitive detection and characterization of eddy-current impedance plane data

There are many signal processing applications for which Multilayer Feed-Forward Networks (MFFN's) performing non-linear, non-parametric classification would be valuable, but where the data is complex and phase or temporal information is an essential discriminant. This is particularly true when utilizing the output from a Fourier transform or time sensitive impedance plane information, such as that created from phase quadrature detection of an eddy current signal.;This dissertation details the design of a multi-level network structure and learning algorithm that together form a complex mapping that solves the particular problem of phase sensitive eddy current defect detection and characterization in nickel and titanium-based alloys. This network and algorithm extend a single neuron, linear model (Widrow's complex LMS) to a multi-level, multi-neuron network with non-linear, complex-valued transfer or activation functions, and a chain-rule-based learning algorithm (Rumelhart's, et al., and Werbos' Backpropagation). In addition, predictive measures associated with classification accuracy are investigated, a network's internal representation is studied, and methods are explored to improve network performance when training set size is limited.;To assess network performance, classic tests, such as the exclusive-or and encoder problems, were extended to the complex plane, providing results similar to those for a real- valued MFFN, with two real inputs and nodes, for each complex input and node. More importantly, for the base sensitive problem of eddy current defect detection and characterization in IN100 and titanium specimens, the CMFFN (Complex Mapping Feed-Forward Network) demonstrated a clear performance advantage over both real-valued MFFNS and human test subjects, with an overall classification accuracy improvement of 45% (to a 99% accuracy level) and 48%, respectively.;This network structure and learning algorithm are general and should provide similar results in other signal processing applications where phase or time considerations are critical for class discrimination.