Exploratory Fault Detection with Multivariate Data: A Case Study on Engine Bearing

This paper presents a case study on using statistical method for detecting impending bearing failures using in-situ field data. We first explore the relationships between a few variables of interest using a matrix plot. By focusing on variables with consistent profile, we analyze the change in these multivariate data over time and propose a way to pinpoint impending failure. Due to the way data are generated and the inherent large variation, a Gaussian mixture model (GMM) is proposed and methods analogous to multivariate SPC are then applied to detect "out-of-control" signal. In particular, a phase I analysis using variances corresponding to the within and between sorties variations so that the correct control limits can be determined. From the actual failure and known conditions from field data, it was found that the proposed method is able to signal impending failure before it occurred.