Zuzana Dvořáková (Czech Technical University and Czech Academy of Science): Informational divergences and their application in material defectoscopy

Información (pendiente de confirmación)

In material defectoscopy, the choice of classification methods and signal attributes is crucial for acoustic signal classification. We have tested several classical methods of signal classification (Fuzzy, Support Vector Machines, Model Based Clustering) under newly designed low-dimensional attribute vector, but our focus is on the newly developed clustering based on φ-divergences called Divergence Decision Tree (DDT) and its supervised version Supervised Decision Divergence Tree (SDDT). The φ-divergences, which originate from the field of information theory, are also successfully used as specific signal attributes to discriminate signal spectra in our classification algorithms. These newly proposed approaches are subjected to thorough quality tests and applied in several non-destructive testing (NDT) experiments (pressure vessel, bluntness of drill bits, etc.) carried out under laboratory conditions at the Czech Academy of Sciences. We develop the concept of model and data robustness and apply them to the selected well-known distances and φ-divergences (Kolmogorov, Cramér, Hellinger, blended divergences). This key property of robustness has been tested in computer simulation under minimum distance estimation setup. The effectiveness of the classification based on φ-divergences is shown and verified in the so-called time reversal (TR) techniques used in non-destructive testing of materials.