System for separating periodic amplitude peaks from non-periodic amplitude peaks in machine vibration data

Abstract

A statistical method is used to separate periodic from non-periodic vibration peaks in autocorrelation spectra. Generally, an autocorrelation spectrum is not normally distributed because the amplitudes of periodic peaks are significantly large and random relative to the generally Gaussian noise. In a normally distributed signal, the statistical parameter kurtosis has a value of 3. The method sequentially removes each largest amplitude peak from the peaks in the spectrum until the kurtosis is 3 or less. The removed peaks, which are all considered to be periodic, are placed into a set. The total energy of the peaks in the set is considered to be the total periodic energy of the spectrum. As the process of building the peak set proceeds, if its total energy becomes greater than or equal to a predefined energy threshold before its kurtosis reaches 3 or less, the process stops and the periodic peak set is defined.