A new unsupervised health index estimation method for bearings early fault detection based on Gaussian mixture model

Abstract

Bearings are indispensable components of machinery, playing a critical role in effective health monitoring. This monitoring is vital in detecting equipment incipient failure and reducing maintenance costs. The bearing degradation is a complex nonlinear process that defies straightforward description using physical models or predefined degradation patterns. Bearings exhibit inherent fault frequencies, and the early faults of bearings can alter their monitoring data distribution. Gaussian Mixture Modeling (GMM) can effectively visualize these changes in data distribution. This research focuses on the development of a new unsupervised method for constructing the bearing Health Index (HI) using GMM to estimate vibration signal distributions. Firstly, we introduce a new unsupervised HI construction method, named GMM-HI, designed to provide insight into the bearing degradation process. Secondly, the Wasserstein distance is adopted as the bearing HI, measuring the distance between initial and current health data. Thirdly, isotonic regression is utilized to address spurious fluctuations in bearing monitoring signals. Through extensive experimentation on three bearing datasets, our results demonstrate that the newly introduced GMM-HI allows for accurate detection of bearing early failures. It effectively categorizes the health stages of bearings and provides a unified approach for establishing bearing failure thresholds in estimating their remaining useful life. In comparison to other well-known HI methods, GMM-HI effectively and accurately characterizes the health index of bearings.