Handling sparse data problems in the context of monitoring multiple parameters in complex systems

The notion of utilising models of normality for health monitoring applications is now well understood^[1]. Robust models can be constructed from large volumes of normal data and then used as a reference to test for abnormality against new observed data. Such techniques have already proved to be very effective for monitoring individual parameters of complex systems^[2,3,4]. The introduction of extreme value theory (EVT)^[5,6], and its variant Bayesian EVT^[7], has also improved the mechanism for establishing novelty thresholds in situations where only small quantities of running data exist^[8]. These thresholds are then allowed to adapt as additional data is observed and a more complete representation of the tails in the underlying distribution is understood. The focus of this paper is with the practical considerations of applying these ideas to the monitoring of multiple parameters of gas turbines. More specifically, our interest lies with the problem of establishing a corresponding multi-dimension novelty threshold for such complex systems during their early stages of running and hence when relatively little service data is available.