A transfer learning approach for remaining useful life prediction subject to hard failure considering within and between population variations

Abstract

Accurate prediction of remaining useful life (RUL) of a unit plays a critical role in condition-based maintenance, especially for hard failure cases. In industrial practice, due to differences in units’ types and working environments, there may exist multiple populations, and even within the same population, there are also variations among units. However, existing methods either assume that different units share the same population characteristics and ignore the between-population variations, or solely focus on between-population knowledge transfer while neglecting the within-population variations. To address this issue, this article proposes a transfer learning approach by integrating a Cox Proportional Hazards (PH) model with a Bayesian hierarchical model, which considers both within and between population variations. Specifically, a shared prior distribution is deployed to the parameters of the Cox model in each population, which builds the foundation for transfer learning across different populations. To model within-population variations, a linear mixed-effects model is utilized to represent heterogeneous degradation data of each unit. The effectiveness of the proposed method is demonstrated and compared with various benchmarks through a simulation study and a case study of turbine engines.