Estimating the frequencies of vibration signals using a machine learning algorithm with explained predictions

Abstract

Signals of short duration and containing a small number of cycles require special procedures if the precise estimation of their frequencies is intended. In this paper, we present an algorithm that allows accurate estimation of frequencies and simultaneously explains the decision regarding the prediction made. We first show why predictions regarding the frequency of signals mentioned above can contain significant errors and the prediction dependency on the analysis time. We then prove that the errors are systematic, and it is possible to train a neural network to quantify the errors and later correct the predictions. The algorithm also indicates the level of error by analyzing the signal-to-noise ratio. The algorithm was tested for numerous similar cases and proved to be reliable. At the end of the paper, we present how to use the algorithm using a signal generated with a known frequency.