Automatic Defect Characterization using Artificial Neural Networks and Deconvolution Techniques

Abstract

An automatic defect testing system is dealt with in this article. It includes two parts, defect feature extraction, and defect classification and sizing. Defect feature extraction is carried out by adaptive filter deconvolution. The time delay between two consecutive taps of the adaptive filter is one half cycle of the ultrasonic echo to be processed. Wideband ultrasonic defect echoes of center frequency 1.2 MHz generally have 2-4 cycles and 100-200 data points for a 50 MHz sampling rate. After deconvolution, data are reduced to 4-8 points, and the frequency bandwidth is greatly extended. As a result, the defect features stand out. The deconvolved defect echoes are presented to an artificial neural network (ANN) for automatic defect classification and sizing. Two application examples are given in this article, exact classification and reasonable sizing accuracy have been achieved.