Robust Classification Of Radar And Speech Signals Using A Fault-Tolerant Feed-Forward Neural Network
Project Award Date: 0000-00-00
Over the past few years, feed-forward artificial neural networks (FANN) with large number of interconnections between the nodes in the adjacent layers have been successfully used in system classification problems. In this investigation, the performance of the FANN is tested in the presence of both system faults using noisy radar and speech signals. In the first part of the investigation, realistic time domain responses from five different aircraft are simulated, and a FANN is trained using the spectral samples of these responses. The network is then tested with noisy spectral samples as the links between the nodes get increasingly disconnected. It is found that the network can provide 94% success rate at a 10 dB SNR, even when 10% of its interconnections failed. A similar investigation is performed with the speech signal where a speaker is allowed to speak five different words, and the LPC coefficients of the time series are used to train a FANN. It is found that the network can provide 98% success at a SNR of 20 dB, even when 10% of its interconnections get disconnected. Thus, it appears that a properly trained FANN might be able to perform as a robust classifier in the presence of failures of its components.