Welding defect pattern recognition in TOFD signals - Part 2. Non-linear classifiers

摘要

The time-of-flight diffraction (TOFD) technique has been widely used for automatic weld inspection. Despite the high speed of inspection, high reliability in sizing and the low rate of false indications, the classification of defects obtained by the TOFD technique is still frequently questioned since it depends mainly on the knowledge and experience of the operator. This dependence on the operator has led to attempts to automatically classify the defects detected during inspection and neural networks are powerful tools that can be used for this objective. The use of non-linear classifiers to improve the performance reached by linear classifiers, presented in previous works, is the main objective of this work. A-scan signals were obtained during the inspection of test samples containing well-controlled weld defect previously characterised by X-rays. After the inspection, the signals were divided into three classes; according to the type of defect found (lack of fusion, lack of penetration and porosity). Another class of signal was acquired from regions with no defects. Non-hierarchical and hierarchical non-linear classifiers, implemented by an artificial neural network, were used in the classification of these signals. The backpropagation learning rule was used to train the neural network. The performance of these two different classifiers was evaluated and compared. The non-linear classifiers had good results in the recognition of welding defect patterns of TOFD signals. The rate of success reached 100% and 98% for training and test respectively, against 99% and 96% for the linear classifiers, which were obtained in the last works.