Agglomerative clustering of defects in ultrasonic non-destructive testing using hierarchical mixtures of independent component analyzers
Addisson Salazar, Jorge Igual, Luis Vergara · 2014
This paper presents a novel procedure to classify materials with different defects, such as holes or cracks, from mixtures of independent component analyzers. The data correspond to the ultrasonic echo recorded after an impact by several sensors on the surface of the material. These signals are modelled by independent component analysis mixture models (ICAMM) for every kind of defect. After the ICAMM model is estimated for every defect, these are merged according to a distance measure that is obtained from the Kullback-Leibler divergence. The hierarchy obtained from the impact-echo data and the learning process allow different kinds of defective materials to be grouped consistently.