Pattern recognition of weld defects in preprocessed TOFD signals using linear classifiers

Abstract

The TOFD (“Time of Flight Diffraction”) technique is being widely used for automatic weld inspection, especially in the petrochemical industry, where welding quality is essential to avoid productivity losses. Although it provides high speed inspection, high sizing reliability and low rate of false defect indications, the classification of defects using ultrasound signals generated by the TOFD technique is still frequently questioned, because it depends heavily on the knowledge and experience of the operator. However, the use of computational tools for signal preprocessing and pattern recognition, such as the artificial neural networks, improves the classification reliability of defects detected by this technique. In this present work, three kinds of defects: lack of fusion (LF), lack of penetration (LP) and porosity (PO) were inserted into the specimens during the welding process, generating pattern defects. The position, type and dimension of each inserted defect were recorded using conventional ultrasonic and radiographic techniques. The Fourier Transform and Wavelet Transform were used for preprocessing Ascan signals acquired during weld inspection by TOFD technique. This study was able to show the versatility of Wavelet Transform to preprocess these kinds of signals, since the correct scale in Continuous Wavelet Transform had been selected to supply a neural network. Hierarchical linear classifiers were implemented into the neural network in order to distinguish the main defects in welded joints detected by the TOFD technique. The results show the good success rate of welding defect recognition in preprocessed TOFD signals, mainly using Wavelet Transform. On the whole, the results obtained were very promising and could give relevant contributions to the development of an automatic system of detection and classification of welding defects inspected by the TOFD technique.