Equipment prognostics and health management as part of operations and maintenance is a critical technology for accurately predicting impending failures and providing a mechanism for replacing parts safely before failure. Key components of these prognostic systems are the algorithms for anomaly detection. Many of techniques for anomaly detection require knowledge of the system model. Attempting to use model based detection methods when dealing with complex systems is often infeasible because the approximations necessary to develop computationally tractable models of complex systems based on fundamentals of physics are difficult to make without introducing significant model inaccuracies in the time and length scale of interest.

An ideal general purpose prognostic system would use data driven prognostics approaches that do not require a priori knowledge of the system. The prognostic system with use would learn the characteristics of the monitored system so that anomalies could be detected more quickly as it learned and remaining life estimates could be given with smaller associated uncertainty.

In this work we are developing data driven methods for detection of the precursors to slowly approaching failures in order that the remaining life of critical components can be accurately predicted at an early stage. W e are using the tools of computational dynamics and pattern discovery. Computational mechanics provides a means to find patterns in symbol sequences that reconstruct minimal equations of motion from the recursive structure of measurement sequences. The goal is to let the process describe itself, without appealing to a priori assumptions about the process's structure.

This work is being done in collaboration with Asok Ray, Distinguished Professor of Mechanical Engineering at Pennsylvania State University.