This paper describes work on an extension of a previous neural network
classifier ((1)) that is able to perform an automatic assessment of t
he severity of cracks in the walls of a steel tube through the detecti
on of the leaked magnetic field. In the present report, the authors an
alyse the performance of this neural network for realistic conditions
in the measuring process. This is modelled with a population of signal
s in which both the shape of the cracks and other physical parameters
that are inherent to the detection process are allowed to change rando
mly. The neural network classifier is able both to assess the potentia
l danger of the crack and determine its location in the internal or ex
ternal wall of the tube. The neural network behaves as a robust classi
fier-showing in all cases a significant improvement over the tradition
al method of estimating the potential danger of the crack through the
amplitude of the detected signal.