Nonparametric model-based prognostics

Abstract

Equipment, process, and system prognostic techniques can be classified as belonging to one of three major classes of methods: 1) conventional reliability-based using failure times (Weibull), 2) population based with environmental considerations (e.g. proportional hazards modeling), and 3) individual based (e.g. general path model). A new individual-based prognostic algorithm, termed the path classification and estimation (PACE) model, has been developed and is based entirely on failure data. This model recasts the general path model (GPM), which is the foundation of the majority of the modern individual based prognosis algorithms, as a classification problem, where a current device's degradation path is classified according to a series of exemplar paths and the results of the classification are used to estimate the remaining useful life (RUL) of the device. The requirement of the existence of a failure threshold is removed, thereby enabling the PACE to be applied to ldquoreal worldrdquo systems, where a single failure threshold is not likely to occur. If the failure threshold is known, simple formatting may be applied to the degradation paths such that they can be easily used with the PACE. The newly proposed method was applied to data collected from the hydraulic steering system of a drill used for deep oil exploration with the objective of detecting, diagnosing, and prognosing faults. The PACE was used to predict the RUL for several failure modes using actual data. For this work, a three tiered architecture was implemented, where conventional reliability methods were used to estimate the population-based RUL, PACE population-based prognosers were trained to map the cause of a failure mode to the RUL, and PACE individual prognosers were trained to map the effects of a failure mode to the RUL. It was found that the population based prognoser produced RUL estimates with large errors (75 hours) and uncertainties (261 hours). The individual prognosers were found to significantly outperform the population based prognoser, with errors ranging from 1.2 to 11.4 hours with 95% confidence intervals ranging from 0.67 to 32.02 hours.