Identificação de defeitos reais de soldagem a partir de sinais produzidos por simulação computacional do ensaio ultrassônico.

Abstract

The detection, location, and sizing of internal discontinuities in materials can be performed non-invasively using ultrasonic testing, which is based on the propagation of high-frequency mechanical waves. Among the techniques based on this principle, time-of-flight diffraction (TOFD) stands out for its high sensitivity and ability to detect defects from the diffracted waves at the ends of discontinuities. However, the classification of defects obtained by the technique from ultrasonic signals still strongly depends on the operator’s interpretation and experience, and is subject to inaccuracies. In this context, classification algorithms have proven to be promising tools for the automatic interpretation of these signals, enabling the identification of characteristic patterns associated with different defect types. In this work, a finite element simulation environment was developed for generating ultrasonic signals using the TOFD technique. Parts with three defect types present in welded joints were modeled in the COMSOL software: lack of penetration, porosity, and lack of fusion. The resulting signals were pre-processed and used in pattern recognition algorithms, including K-Nearest Neighbors (KNN), quadratic classifier, and artificial neural networks. Initially considering the scenario where the algorithms are trained and tested only with simulated signals, the best performance was obtained by the KNN classifier, which achieved an average accuracy of 87.47% and a maximum accuracy of 100% when using the signal envelope as input. In the stage where the models were trained with simulated data and evaluated using experimental data, the standout was the MLP artificial neural network with two hidden layers, which achieved an average accuracy rate of 63.13% and a maximum of 80.57% in classifying the complete set of welding defects.