A clustering approach using a time-frequency entropy measure of wavelet transform coefficients

A non-parametric clustering approach that uses a time-frequency (TF) entropy measure taken from the signal wavelet transform coefficients is introduced (TFEWT). The TFEWT feature vector represents a concatenation of two vectors obtained from the projection of the signal wavelet entropy in TF space onto both the time and frequency axes. A signal-to-noise ratio criterion is evaluated to obtain the best clustering result by changing the signal time-frequency decomposition both by the basis set and the wavelet type. In comparison with FFT and different well known TF features like wavelet or wavelet packet coefficients, TFEWT renders a compact feature vector that optimizes the clustering criterion for distinct transient clusters, even when they have very similar TF energy distributions