On the Selection of the Proper Number of Classes in TFD Segmentation for Extraction of Useful Information Content from Noisy Signals

In this paper, the influence of the selection of the number of classes as defining parameter of 1D the segmentation into the procedure of evaluation of noise classes and useful information classes applied to time-frequency distributions (TFD) of noisy signals is investigated. The method for identification and separation of the useful information content, from the noisy background nonstationary signals, is based on an local Renyi entropy evaluation of the individual classes. The initial segmentation of the TFD information content by the K-means clustering algorithm, is, however, a decisive step of the denoising procedure, since it irreversibly affects the input data of the entropy estimation, and hence the overall algorithm's performance. The paper compares results, in terms of error rate, obtained for different fixed values of the parameter K, and for an adaptive method for determination of the optimal parameter K. Results tend to show that the increasing of the parameter K acts beneficially on the algorithm's performance, while no similar behavior has been observed using the adaptive approach.