Neural networks for ultrasonic NDE signal classification using time-frequency analysis

Ultrasonic nondestructive evaluation (NDE) of material defects typically involves signals which are nonstationary in nature. Whether we are doing deconvolution or signal classification, time-frequency analysis, instead of the frequency or time domain analysis alone, is much needed. In this paper, we examine features derived from the Wigner distribution and its derivatives, and features derived from subband coding wavelet decomposition. Both the traditional nearest neighbor decision rule and the neural network classifiers, the backpropagation trained network and the Nestor's RCE network, are considered to classify the ultrasonic pulse echoes into one of three hidden geometrical defect classes. Neural network classifiers using features properly derived from the time-classification results. Although the data set employed is small, the conclusion is fairly consistent with experiments in other large data sets.