Ultrasonic grain signals classification using autoregressive models

Autoregressive (AR) analysis is used to characterize the ultrasonic microstructure (i.e. grain) scattering of materials. Grain scattering results in an upward shift in the expected frequency of broadband echoes, while attenuation results in a downward shift. Both the upward and downward shifts are correlated to grain size distribution. In order to evaluate the spectral shift in grain signals, the authors have adopted low-order AR models and extracted features such as AR coefficients, resonating frequency and maximum energy frequency. A Euclidean distance classifier based on these features is implemented to classify grain scattering characteristics. Computer-simulated and experimental data give a probability of correct classification about 75% for the second-order AR model and 88% for the third-order AR model when the expected frequency shift is less than 4%