Neural Networks to Select Ultrasonic Data in Non Destructive Testing

Abstract

In recent years, research concerning the automatic interpretation of data from non destructive testing (NDT) is being focused with an aim of assessing embedded flaws, quickly and accurately in a cost effective fashion. This is because data yielded by NDT techniques or procedures are usually in the form of signals or images which often do not present direct information of the structure’s condition. Signal processing has provided powerful techniques to extract the desired information on material characterization and defect detection from ultrasonic signals. The imagery available can add additional and significant dimension in NDT information. The task of this work is to minimize the volume of data to process replacing ultrasonic images type TOFD by sparse matrix, as there is no reason to store and operate on a huge number of zeros, especially when large structures are inspected. A combination of two types of neural networks, a perceptron and a Self Organizing Map (SOM) of Kohonen is used to distinguish between a noise signal from a defect signal in one hand, and to select the sparse matrix elements which correspond to the locations of the defects in the other hand. This new approach to data storage will provide an advantage for the implementations on embedded systems as it allows the normalization of the sparse matrix by fixing its dimension.