Feature extraction techniques using adaptive wavelets for signal classification

Neural Networks have shown great success in solving classification problems. One of the most important issues in classification is related to the particular set of features to be extracted using a particular extraction method. Recently, the theory of wavelets has emerged as an alternative time-frequency analysis tool to the Fourier transform. Wavelets have been applied to a variety of problems, most notably data compression and noise reduction. The application of the multi-resolution analysis developed by Mallat to signal classification is explored in this paper. Feature extraction techniques based on the traditional dyadic wavelet decomposition and on the adaptive wavelet representation, have been developed and presented. These methodologies are then tested on a neural network classifier, while results on low frequency Fourier coefficients are provided as baseline comparisons for this study.