Feature extraction and classification of nonstationary signals based on the multiresolution signal decomposition

A new classification method based on the multiresolution decomposition analysis of nonstationary signals is proposed. The signal is decomposed into (M+1) octave bands in order to reduce the degree of nonstationarity. A separate feature vector is extracted from each band and a “time-frequency feature matrix” is formed using these vectors. This feature matrix is then inputted to the first stage of a two-stage classifier including (M+1) individual classifiers where a generalized version of the k-nearest neighbor classification rule and soft decision method is used. Final decision is given at the second stage by combining the decisions of the (M+1) classifiers using a decision combination function. The intended application of the method is the classification of respiratory sound signals into healthy and pathological classes