Pattern classification of decomposed wavelet information using ART2 networks for echoes analysis

Abstract

The Ultrasonic Pulse-Echo technique has been successfully used in a non-destructive testing of materials. To perform Ultrasonic Non-destructive Evaluation (NDE), an ultrasonic pulsed wave is transmitted into the materials using a transmitting/receiving transducer or arrays of transducers,that produces an image of ultrasonic reflectivity. The information inherent in ultrasonic signals or image are the echoes coming from flaws, grains, and boundaries of the tested material. The main goal of this evaluation is to determine the existence of defect, its size and its position; for that matter, an innovative methodology is proposed based on pattern recognition and wavelet analysis for flaws detection and localization. The pattern recognition technique used in this work is the neural network named ART2 (Adaptive Resonance Theory) trained by the information given by the time-scale information of the signals via the wavelet transform. A thorough analysis between the neural network training and the type wavelets used for the training has been developed, showing that the Symlet 6 wavelet is the optimum for our problem.