Arc welding defect detection by means of Principal Component Analysis and Artificial Neural Networks

The introduction of arc and laser welding in the aerospace, automotive and nuclear sectors, among others, has led to a great effort in research concerning quality assurance of these processes. Hence, an on-line, real-time welding monitor system able to detect instabilities affecting the welding quality would be of great interest, as it would allow to reduce the use of off-line inspection techniques, some of them by means of destructive-testing evaluation, improving process productivity. Among several different approaches, plasma optical spectroscopy has proved to be a feasible solution for the on-line detection of weld defects. However, the direct interpretation of the results offered by this technique can be difficult. Therefore, Artificial Neural Networks (ANN), due to their ability to handle non-linearity, is a highly suitable solution to identify and detect disturbances along the seam. In this paper plasma spectra captured during welding tests are compressed by means of Principal Component Analysis (PCA) and, then, processed in a back propagation ANN. Experimental tests performed on stainless steel plates show the feasibility of the proposed solution to be implemented as an on-line arc-welding quality monitor system.