Prognosis of faults in gas turbine engines

A problem of interest to aircraft engine maintainers is the automatic detection, classification, and prediction (or prognosis) of potential critical component failures in gas turbine engines. Automatic monitoring offers the promise of substantially reducing the cost of repair and replacement of defective parts, and may even result in saving lives. Current processing for prognostic health monitoring (PHM) uses relatively simple metrics or features and rules to measure and characterize changes in sensor data. An alternative solution is to use neural nets coupled with appropriate feature extractors. Neural nets have been shown to give superior statistical performance when compared to more traditional techniques, and some net architectures are capable of "novelty detection" (identifying an event as dissimilar to any previously observed). On the negative side, neural nets used to model and predict engine status necessarily will be complicated and will contain many "black box" components that do not offer insight into the "why" the net came up with the results that it did. There is currently a wide variety of data mining (or rule extraction) tools commercially available. All of these tools come up with readily-understood rules by performing an exhaustive search on training data. However, in most cases the simple rule derived does not have the statistical performance realized with the neural models. Rule extraction tools are also incapable of novelty detection. We have developed techniques that couple neural nets with automated rule extractors to form systems that have: Good statistical performance Easy system explanation and validation Potential new data insights and new rule discovery Novelty detection Real-time performance We apply these techniques to data sets data collected from operating engines. Prognostic examples using the integrated system will be shown and compared with current PHM system performance. Rules for pet-forming the prognostics will be developed and the rule performance compared as well.