Robust multivariate statistical ensembles for bearing fault detection and identification

Abstract

This paper presents a novel methodology for the identification and detection of faults on based upon high frequency bearing data collected from a 6205-2RS JEM SKF bearing. A robust derivative of the Mahalanobis distance is employed to accurately and precisely encapsulate varying fault behaviours, which can then be exploited for the purposes of fault detection and identification. Domain knowledge in the form of failure mode and effect analysis (FMEA) can be incorporated into the model, to determine potential failure modes. Seeded fault data was employed to derive the shape and location estimates to enable the use of a multivariate distance function. To reduce the computational complexity whilst simultaneously increasing sensitivity to the faults, the high frequency (48KHz) accelerometer data was pre-processed into a 4-tuple consisting of the Skewness, Kurtosis, Root mean square (RMS) and Shannon Entropy. This 4-tuple is shown to encapsulate and discriminate all fault modes identified through the FMEA, whilst reducing the data to 1Hz, allowing for the both exact, and meta-heuristic algorithms to be employed for robust analysis. Sensitivity to minimal fault development is demonstrated, with the technique accurately identifying 0.007" diameter inner race, outer race and roller element faults which had been seeded to the bearing through electro-discharge machining. To demonstrate the practicalities of the approach, the trained system is employed for analysis of an independent dataset, collected under different conditions. The technique is shown to accurately detect and identify the relevant fault mode pre-emptively, before catastrophic failure occurred, with 28.6% of bearing life remaining.