Multi-Scale Rank-Permutation Change Localization

Prediction of equipment remaining useful life (RUL) is of considerable economic benefit to industry, by permitting the avoidance of unscheduled downtime and costly secondary damage. Detection of change is an important first step in building a prognostic system: when a detectable fault occurs, it will cause changes in one or more sensed parameters of the system. Once a change has been detected, localizing the time of change (presumably the time of fault onset), can contribute to the estimate of RUL in two ways. First, it may make the RUL estimate more accurate. Second, and independent of estimate accuracy, it may make the prognostic estimate more precise by reducing the variance of the estimate. This paper describes an approach for localizing the time of change in time series data. The performance of the system is assessed using synthetic data that closely matches the characteristics of real-world data. However, the synthetic data is deterministically labeled, so algorithm performance can accurately be assessed. The approach presented requires low computational power at runtime, an important feature for on-wing application that combines the rank transformation of data, randomization tests inspired by the work of Fisher and Pitman, and fusion of multi-scale estimates to result in a fast and accurate localization of change.