Classificação de descontinuidades em juntas soldadas utilizando máquinas de vetores-suporte treinadas a partir de sinais de ultrassom simulados numericamente.

Abstract

Ultrasonic testing is a fundamental tool for quality and safety controls in the industry. Welded joints, for example, are elements that need a careful inspection since they are sensitive to the formation and the spread of discontinuities. However, the analysis of the signals resulting from these inspections requires highly qualified and experienced operators to properly characterize and classify the discontinuities. This task can be accomplished by classification models, which, combined with numerical simulations, can be used as quick, inexpensive and effective alternatives to solve these problems. In this work, it was evaluated the performance of support vector machines in the classification of ultrasonic signals referring to three classes of discontinuity in welded joints: incomplete penetration, porosity and cracks. Classifications were performed involving experimental signals captured by the time of flight diffraction technique and signals obtained by simulations of this technique. Besides, the use of a hierarchical classification approach compared to the conventional One-Against-All classification approach was evaluated. The results obtained are promising and indicate that good success rates can be obtained in the classification of ultrasonic signals performed by support vector machines trained by simulated signals.