New Approaches for Reliability Analysis of Industrial Systems Subject to Multivariate Degradation

Description

Degradation refers to a cumulative change of performance characteristics of a product/system over time, such as wear of tires and capacity reduction of batteries. The current big data era facilitates the simultaneous monitoring of multiple degradation signals that are closely related to the health status of the system. Investigation of degradation behaviour offers a promising way for reliability assessment and management of modern reliable systems. System reliability is also heavily affected by dynamic usage intensity as well as operating environment and can be significantly improved with preventive maintenance. Existing research on multivariate degradation mainly focuses on modelling dependences of multiple degradation characteristics while effects of lifetime-affecting factors and maintenance are overlooked. How to properly incorporate dynamic heterogeneities and preventive maintenance into the modelling of multivariate degradation is an important yet challenging problem. Corresponding effective reliability analysis and maintenance plans are also worth exploring to prevent catastrophic failures and ensure system safety. Motivated by interests of our industry collaborators, this project proposes to bridge the gap by developing a novel reliability analysis and management framework. A stochastic process model is first developed to describe the multivariate degradation and factor in the influence of observable dynamic covariates like relative humidity and machine operating modes, which can be used to predict system lifetime subject to complex time-varying environment. Considering the commonly scheduled maintenance in practice (e.g., lubrication reduces the vibration amplitude/degradation level of rolling bearings), we will formulate new statistical models to quantify the effect of maintenance actions. The corresponding parameter estimation and statistical inference for the proposed models will be investigated in a data-driven manner, to combine and improve standard statistical techniques. Based on the proposed multivariate degradation models incorporating dynamic covariates and preventive maintenance, we will develop an online/real-time Bayesian inference framework for remaining useful life prediction. The ultimate objective is to reduce downtime losses from unexpected breakdowns and recommend better maintenance actions and schedules. State-ofthe- art machine learning techniques will be studied and enhanced to address several challenges from the multivariate and dynamic nature of our problem. The results of remaining useful life prediction will be compared with various existing methods in terms of prediction accuracy, and the resulting maintenance cost will also be compared with those from other classic maintenance policies. The proposed methods in this project will advance the state-of-the-art of the degradation modelling and system health assessment and management from both theoretical and practical perspectives.