Lubricating Oil Remaining Useful Life Prediction Using Multi-Output Gaussian Process Regression

Abstract

Lubricant condition monitoring (LCM) is a preferred condition monitoring (CM) technology for fault diagnosis and prognosis owing to its ability to derive a wide range of information from the system (machine/equipment) state and lubricant state. Given the importance of LCM for maintenance decision support, an accurate and reliable remaining useful life (RUL) prediction framework is necessary. The LCM health information in the form of degradation trends is therefore evaluated using numerous statistical, model-based, and artificial intelligence approaches by various researchers. A multitude of factors widely affects the degradation trends viz. operating conditions, environmental variations, oil replenishments, oil loss, chemical breakdown, etc. These factors increase the complexity of the time-series degradation trends making RUL prediction intractable using several of the standard statistical approaches. Therefore, limited research is available on lubricating oil RUL prediction with these influential factors accounted for. Focusing on the complexity of the degradation trend with oil replenishment effects (ORE), we propose the use of the Gaussian process regression (GPR) model for RUL prediction in this study. The model has an advantage over other data-driven approaches as it is a non-parametric Bayesian method. To exploit prior information and historical data collected, the approach is extended to multi-output GPR (MO-GPR) which effectively defines the correlations between historical degradation trends for similar lubrication systems with the current degradation pattern of a system being monitored in real-time. Three different oil replenishment strategies are considered under MO-GPR to demonstrate the applicability and flexibility of this method.

Keywords

• Gaussian process regression
• lubrication condition monitoring
• prognostics
• remaining useful life