Neural Networks: An Overview - Neural Networks: Methodology and Applications - page 55

1.2

1.0

0.8

0.6

0.4

0.2

0

1 5 9 13 17 21 25 29 33 37 41

Pairs of classes

Pixels Features

Fig. 1.35. Distances between classes for two representations: the feature-map based
representation makes the classes more widely separated in input space, thereby
making the classification task easier

1.4.3 An Application in Nondestructive Testing: Defect Detection
by Eddy Currents

The example that was presented in the previous section used classification for
picture recognition. Of course, patterns that can be recognized automatically
vary widely in nature. The application that we describe in the present section
pertains to nondestructive testing, where the patterns to be classified are
signals. The objective is the automatic detection of defects in the rails of
the Paris subway. It was developed by the National Research Institute on the
Safety of Transportation Systems for RATP, the company that operates the
Paris underground system [Oukhellou 1997].
Defect detection in metal parts by eddy currents is a standard nondestruc-
tive testing technique. An electromagnetic coil creates an alternating magnetic
field, which generates eddy currents in the metal part to be tested. Those cur-
rents are detected by a second coil, and the presence of defects in the metal
alters the amplitude and the phase of the resulting signal. Thus, that signal

Table 1.1.

Correct
classification Rejection Misclassification

rate (%) rate (%) rate (%)

Pixel representation 70.9 28.1 1

Feature-map-based
representation 90.3 8.7 1