A general anomaly detection approach applied to rolling element bearings via reduced-dimensionality transition matrix analysis

Rolling element bearings are vital components in most rotating machines. Bearings often operate in harsh environments where manufacturing imperfections, misalignments, and fatigue can result in reduced component lifespan. These failures are often preceded by changes in the normal vibration of the system. Modeling and detecting these vibrational anomalies is common practice in predicting machine failure. This paper develops and implements a novel approach to detecting bearing vibration anomalies in the time-frequency domain. The performance of the new approach is quantified using both simulated and experimental bearing vibration data. In these ground-truth experiments, the proposed time-frequency method successfully detects anomalies (>98% true positive) using short time spans (<0.1 s) with low false alarm rates (<1% false positive). Using experimental data, this time-frequency approach is shown to outperform one-dimensional time series analysis techniques.