Statistics of signature images for arc welding fault detection

Abstract

The present work covers the statistical analysis of signature images for weld quality monitoring and fault detection. Welding signature images are two-dimensional histograms of voltage and current data. Signatures from production welding are compared with reference good welding using a statistical analysis of both the principal components, which are fitted with shifted normal distributions, and the remainders not explained by the principal components, which are fitted with a multidimensional normal distribution. A signature basis set is employed, allowing efficient real time computations. The statistical data fits are illustrated with data from overlap welds, and examples of fault detection with out of position welds are given.