Nonlinear statistics for bearing diagnosis

Abstract

This document presents the preliminary results of an ongoing study related to the use of nonlinear statistics for bearing diagnosis. In this study, we propose a methodology based on the K-nearest neighbor algorithm to test the ability of a group of nonlinear statistic to differentiate between vibration signals obtained from rotatory machines with bearings in good and in bad condition. Results showed that statistics such as Lempel-Ziv complexity, Sample Entropy, and others derived from the recurrence plot, unlike the correlation dimension, are good at detecting a failure in a bearing. Additionally, we found that the Sample Entropy is exceptionally good at this task.