Prognostics, a new look at statistical life prediction for condition-based maintenance

Abstracf- The experience-based. statistical approach to prognostics is the simplest to implement when repair and failure histories are available. Furthermore this may he the preferred approach for classes of systems that are noncritical, too complex to model, or highly susceptible to variation due to the operating regimes or environment, such as with industrial motors for example. Today industry and military databases are becoming available that describe the histories of inspections. repairs. and failures for many types of systems. In this paper. a systematic approach is described for deriving and maintaining accurate Weibull distributions used to determine when maintenance actions should be performed. The approach starts by data mining the maintenance database through discovery of subpopulations of related systems so as to derive mnre accurate probability distributions. The power of the approach arises from the flexibility to simultaneously analyze and inspect the data through an interactive graphical interface. Related subpopulations of systems in the database are discovered through a graphical tree representation of the data according to arbitrarily definable sets of factors affecting the failure rates in the population under study. Once the subpopulations are identified then an approach to "realtime" data mining of sensor data is applied to account for variations in the operating environment. For illustration, a case study describes this approach as deployed for condition-based maintenance of critical industrial motors in steel and aluminum manufacturing plants. TABLE OF CONTENTS