Feature extraction techniques for ultrasonic signal classification

In this paper, we present two feature extraction techniques for the classification of ultrasonic NDE signals acquired from weld inspection regions of boiling water reactor piping of nuclear power plants. The classification system consists of a pre-processing block that extracts features from the incoming patterns, and of an artificial neural network that assigns the computed features to a particular class of defect present on the inspected pipe. The two techniques are respectively based on the discrete Gabor transform (DGT) and on the discrete wavelet transform (DWT); a third feature extraction technique, based on the clustering of the wavelet coefficients, is also presented. The results carried out by artificial neural networks trained and tested using the described feature extraction techniques, demonstrate the usefulness of the clustered DWT method with respect to the well known techniques of DGT and DWT.