| ABSTRACT: | ESTIMATED ACCURACY OF CLASSIFICATION OF DEFECTS DETECTED IN
WELDED JOINTS BY RADIOGRAPHIC TESTS
R. R. de Silva, M. Siqueira, D. Mery, J. Rebello, Luiz Cal™ba
Federal University of Rio De Janeiro (ufrj), University City - Ilha do Fundão, Rio de Janeiro, Brazil
Radiographic inspection, despite the new ultrasound advanced techniques, hasn’t lost importance among the
non-destructive testing. In contrast, its improvement can be easy noted by computed radiography (CR),
which has arisen to replace the traditional film inspection. In additional to CR, the development of image
processing and pattern recognition techniques led the large number of researches in the way to develop an
automatic or semiautomatic system of radiographic analysis. In this work, the authors presents some results
about a study of classification of welding defects: undercut, lack of penetration, porosity, slag inclusion,
crack and lack of fusion. Geometric features were extracted through pattern radiographs of the International
Institute of Welding to be use as components of the input vectors. To implement the classifiers, artificial
neural networks were employed. Linear and nonlinear classifiers were developed and evaluated to estimate
their accuracy, which were done by using random re-sampling like bootstrap. Moreover, these features were
evaluated by relevancy criterion in the discrimination of the welding defect classes. Nonlinear discriminate
analysis was also carried out to build two- dimensional graphs using the two principal components of
nonlinear discrimination. These principal components were also used as input vectors in the way to reduce
the dimensionality of the input data. The accuracies estimated of the classifiers were about 80-85% using
test samples for situation involving 6 or 5 classes.
|