METHOD OF PROCESSING VECTOR SIGNALS FOR PATTERN RECOGNITION BASED ON WAVELET ANALYSIS

摘要

FIELD: calculating; counting.;SUBSTANCE: invention relates to data processing methods. Method includes: A) primary processing of signal x(t), normalization of preprocessed signal to obtain array of signal samples to form training sample, divided into training, validation and test set; B) for each sample of sampling array, windows of current level of detail (WCLD) are determined, which correspond to specified value of width parameters s and position of their centers - t, with provision of overlapping of adjacent windows; C) each sample of sampling array is processed by wavelet transformation; D) selecting a reference function, a maximum number of its variables with subsequent construction of a family of models for displaying a wavelet coefficient function (a₁…a_n) by one target value; E) after which each family model according to claim D) is trained on a training set with selection of weight parameters of models w₁…w_n and subsequent selection of the best models based on the criterion calculated on the validation set; F) checking selected models at step D) on a test sample by calculating an algorithm convergence evaluation criterion; G) selection of significant sections (SS) corresponding to windows of said level of detail, H) transition to the next level of detail inside selected by item G) SS - WCLD corresponding to better models containing wavelet coefficients a_i, previously calculated inside these windows; I) for each windows of sections, determined at the first level of detail, applying a corresponding wavelet transformation with a smaller scale parameter to obtain a detailed representation of the set of SS, after which steps C)–I) are repeated until a convergence criterion is obtained, determined from the value of the target function on the test set, resulting in a set of paired combinations of s and t corresponding to the SS of the measured signal, from which the desired target parameter is determined.;EFFECT: technical result consists in improvement of target parameter determining accuracy.;17 cl, 3 dwg