The present work evaluates the application of artificial neural networks for pattern recognition of ultrasonic signals using pulse-echo and TOFD (Time of Flight Diffraction) techniques in weld beads. In this study pattern classifiers are implemented by artificial neural network of backpropagation type using MATLAB®. The ultrasonic signals acquired from pulse-echo and TOFD were introduced, separately, in the artificial neural network with and without preprocessing. The preprocessing was only used to smoothen the signal improving the classification. Four conditions of weld bead were evaluated: lack of fusion (LF), lack of penetration (LP), porosity (PO) and non-defect (ND). The defects were intentionally inserted in a weld bead of AISI 1020 steel plates of 20 mm thickness and were confirmed using radiographic tests. The results obtained show that it is possible to classify ultrasonic signals of weld joints by the pulse-echo and TOFD techniques using artificial neural networks. The results showed a performance superior a 72% of success for test. Although the preprocessing of the signal improved the classification performance of the signals acquired by the TOFD technique considerably, the same didn't happen with the signals acquired by the pulse-echo technique.
Nondestructive tests; ultrasonic technique; artificial neural network; pattern recognition
