Statistical Classification of Spectral Data for Laser Weld Quality Monitoring

Abstract

Signals from several sensors were employed for real-time laser weld quality monitoring. Sheet-metal butt-joint laser welds of three quality classes (full penetration, partial penetration, gapped) were produced in experimental trials. Optical, air-born acoustic and plasma charge signals acquired during welding were subsequently Fourier-transformed and the spectra were analyzed individually to determine relationships to laser weld quality. The frequency bands most highly correlated to weld quality were identified by stepwise linear discriminant analysis (LDA) of the spectra. Testing of the quality discriminators formulated by LDA of the spectral data showed that the acoustic signal was most reliably correlated with weld quality. Fusing the data from all three sensors prior to LDA analysis produced a discriminator that had about the same reliability as one based on acoustic data alone.