Preliminary study of classification of defects by ultrasonic pulse-echo signals using an artificial neural network

The ultrasonic pulse-echo. technique, is widely employed in industry for the detection of discontinuities in materials. In spite of all the progress in automation, defect classification is still a difficult task and it demands an expert with good knowledge. This study proposes a methodology based on artificial neural networks (ANN) for the automatic classification of the defect type. After the signals were acquired, the efficiency of an artificial neural network was analysed in classifying four conditions of the welded joint: no defect (ND), lack of fusion (LF), lack of penetration (LP) and porosity (PO). The signals were used as input to the neural network in four conditions: radio frequency (RF), full-wave rectifier, preprocessed full-wave rectifier and preprocessed RE One low-pass filter was used in the preprocessing to smooth the noise existing in the signal.