Real-time arc-welding defect detection and classification with principal component analysis and artificial neural networks

A novel system which allows arc-welding defect detection and classification is presented in this paper. The spectroscopic analysis of the plasma spectra produced during the welding process is a well-known technique to monitor the quality of the resulting weld seams. The analysis of specific emission lines and the subsequent estimation of the electronic temperature TeTe profile offers a direct correlation between this parameter and the corresponding weld seams. However, the automatic identification and classification of weld defects has proven to be difficult, and it is usually performed by means of statistical studies of the electronic temperature profile. In this paper, a new approach that allows automatic weld defect detection and classification based in the combined use of principal component analysis (PCA) and an artificial neural network (ANN) is proposed. The plasma spectra captured from the welding process is processed with PCA, which reduces the processing complexity, by performing a data compression in the spectral dimension. The designed ANN, after the selection of a proper data training set, allows automatic detection of weld defects. The proposed technique has been successfully checked. Arc-weld tests on stainless steel are reported, showing a good correlation between the ANN outputs and the classical interpretation of the electronic temperature profile.