Frequency discrimination using neural networks with applications in ultrasonics microstructure characterization

Neural networks based on the propagation algorithm are used to discriminate time and frequency signatures inherent in grain signals. The samples of grain signals are applied directly or preprocessed for feature selection before being applied to the neural network. The methods of feature selection are signal power spectrum, autocorrelation and autoregressive coefficients. These methods are applied to both simulated and experimental data. Overall recognition performance as high as 100% for simulated data and 87% for experimental data is obtained, although this high performance has not occurred for some feature selection methods