Reducing Uncertainty in PHM by Accounting for Human Factors - A Case Study in the Biopharmaceutical Industry

The ultimate goal of prognostics within Through-life Engineering Services (TES) is to accurately predict the remaining useful life (RUL) of components. Prognostic frameworks inherently presume that there is predictability in the failure rate of the system, i.e. a system experiencing exclusively stochastic failure events cannot, by definition, be predictable. Prediction model uncertainties must be bound in some logical way. Therefore, to achieve an accurate prognostic model, uncertainty must first be reduced through the identification and elimination of the root causes of random failure events. This research investigates human error in maintenance activities as a major cause of random failure events, using a case study from the biopharmaceutical industry. Elastomer failures remain the number one contamination risk in this industry and data shows unexplained variability in the lifetime of real components when compared to accelerated lifetime testing in the lab environment. Technician error during installation and maintenance activities of elastomers is one possible cause for this and this research explores how these errors can be eliminated, reduced, or accounted for within the reliability modeling process. The initial approach followed was to improve technician training in order to reduce errors and thereby reduce the variability of random failure events. Subsequent data has shown an improvement in key metrics with failures now more closely matching data from lab testing. However, there is scope for further improvements and future research will explore the role of performance influencing factors in the maintenance task to identify additional causes of variation. These factors may then be incorporated as a process variable in a prognostics and health management (PHM) model developed for the system. The paper will present these data fusion approaches accounting for human factors as a roadmap to improving PHM model reliability.