Data Driven Prognostics

Maintenance prognostics is concerned with the prediction of an abnormal operation (fault) within a system's component. The fault can be due to a breakdown of a part within the component, lack of syncronization internally or with other components, and soforth. Furthermore, being notified that a failure will occur when it is too late to prevent the failure is generally not of value. Thus, a prognosis system should be capable of predicting that a fault will occur, identifying the fault type andforecasting when the fault will occur.In this proposal, we seek to determine if neural network-based prognostics methods can be constructed for a component of the Airborne Laser System, which will utilize only information currently provided by the component. In particular, we seek wish todetermine if a fault will occur and the type of the fault, well in advance of the actual fault. During this investigation, we will attempt to produce meaningful rules from the neural networks, to assist in the validation of neural network-basedprognostics methods. Based upon the results of this study, we will recommend both prognostic methods and a framework for a prognostic system capable of monitoring the Airborne Laser System in real-time. The key benefit of this research is the utilizationof general learning methods that produce interpretable rules for the purpose of prognostics. These methods can form the heart of a prognostics system, which can be configured to handle various types of data, with little expert knowledge required.This capability should be beneficial to a wide range of customers of which two will be briefly mentioned. First, the ability to detect problems well in advance of serious consequences is invaluable. For instance, knowing that a weapon system is showingsigns of breakage in advance would permit repair or replacement. Or, for manufacturers, knowing a vital piece of equipment will cause a large stoppage in the near future, if a less time consuming maintenance doesn't take place, saves money.Second, the ability to validate the rules used by the prognostics system can be a large advantage. For instance, the users of the prognostics system can check to ensure the rules generated make sense. Second, the rules can indicate which data sources areof utility, when they are useful, and in which combinations they are educational. This latter information can lead to better choices of the types of data to acquire.