Material classification through neural networks

Abstract

Ultrasonic back wall echoes received from copper and aluminium plates of varying thicknesses are classified through neural network analysis for in situ material identification. To reduce the effect of thickness variation on the time domain signals, and the dimensionality, the Karhunen-Loeve transform was explored. Enormous data compression was achieved; however, the dimensionality of the reduced space was not constant and increased with the incorporation of the new ultrasonic signals from samples of different thicknesses. The power spectra in the frequency domain, on the other hand, was concentrated in the initial few discrete frequency components independent of thickness. A multi-layered feed-forward artificial neural network was trained by the frequency domain signals of the two classes. It was found that the performance of the learned network was quite reliable on the test samples even in cases where the thickness of the test sample is different from the learned samples.