Feature extraction and learning decision rules from ultrasonic signals-applicability in non-destructive testing

The author presents a supervised multiple-concept learning method for generating decision rules from a set of ultrasonic data for defect characterization purposes in nondestructive testing. The first step towards flaw discrimination is to extract relevant information from the collected defect signatures. The large-dimension signal space is mapped into a smaller feature space. The learning set consists of preclassified examples described by a set of continuous attributes measuring the selected features. A decision-tree based algorithm is used to build classification rules able to classify any object from its values of attributes