Neural network approach to maximum likelihood estimation for eddy-current back-scattering NDE data inversion

The aim of this paper is to present a neural network approach to crack location based on eddy-current back-scattering measured data inversion. A deep defect inside a conductive object may be revealed by sliding an electromagnetic probe over the object's accessible surface: this operation gives a set of differential impedance measures, whose configuration carries information on both the shape and the location of the crack. By inverting the measured impedance data it is thus possible to reconstruct the geometry of the defect. Commonly employed data inversion techniques, such as the one based on maximum likelihood theory, require the availability of a forward model which describes the way the data are generated by the system under test. When a physical model is not available or is too much difficult to be handled, a suitable black-box model could be used instead. In this paper we propose the use of a multilayer perceptron for this purpose, which proved to be effective because of its well-known function approximation and system identification capabilities.