Elaboration of Feature Systems for Procedures of Recognition of Noise Signals

Abstract

There is described the methods of informative features systems selection for various statistical recognition procedures. The features informativeness is determined according to Kullback S. statistics, two binomial distributions comparison criterion, dispersion comparison method. The classification is carried out according to maximum likelihood, probabilistic ratio generalization sequential analysis, etc. criteria. The methods described are applied for the recognition of noise-like (seismic, acoustic)signals according to spectral and cepstral features. The effectiveness of the features is proved by the recognition results. The probability of correct recognition of the noise-like signals according to cepstral features is by 10-15% greater than according to spectral ones. Besides, there has been determined that the most effective is the classifier of the maximum likelihood. In addition, there has been investigated the influence of the observed signal interval duration and features number upon the recognition results.