Robust Spectral-Based Techniques for Classification of Wldeband Transient Signals

We recently investigated various spectral-based classification schemes designed to separate wideband transient signals and compared their performances to those obtained using a back-propagation neural network implementation [2]. The spectral-based measures considered include the Bhattacharyya distance, the divergence, the normalized cross-correlation coefficient, and the modified normalized cross-correlation coefficient. Results showed that accurate classification may be obtained using spectral-based measures and that the performances compare, or are sometimes better, to those obtained using neural networks when the training data used to train the neural network is small. In this paper we investigate the robustness of the spectral measures and the neural network approximation classification schemes to white additive noise degradation in the testing sets. Results show that the spectral-based techniques are more robust when the testing sets are degraded with noise.