The classification of weld defects from ultrasonic images: a neural network approach

Neural networks are shown to be effective in being able to distinguish crack-like weld defects from more benign volumetric defects by directly analysing the images collected from ultrasonic scanning. The performance is similar to that of existing methods based on extracted feature parameters. In each case around 94% of the defects in a database derived from 84 artificially-produced defects of known type are placed correctly into one of four classes; rough and smooth cracks, slag and porosity. However, the methods based on classification directly from the ultrasonic image are faster, and the speed is sufficient to allow on-line classification during data collection. A prototype based on the Harwell ZIPSCAN ultrasonic scanning system is described.