Forecasting turbine problems by means of the state space framework

Abstract

A forecasting system is set up to improve the diagnosis in a Condition Monitoring Programme of a critical turbine placed at an industrial plant. The system is based on a statistical model in a State Space framework, such that the local mean level of the vibration state of the equipment is estimated directly from the data, based on a continuous-time set up. This model is combined with a cost model in Conditioned Monitoring, by which the time of preventive replacement is produced when the minimum of the expected cost per unit of time is reached into the future. Such measure is a combination of the costs of failure, the costs of a preventive replacement and the probabilities of reaching the alarm levels fixed by some criteria. The system is estimated by Maximum Likelihood and thoroughly tested on the equipment. The main tests relate to statistical properties of the model residuals and a comprehensive comparison with an alternative system, namely a linear trend regression model in continuous time. The system produced a reasonable forecasting performance and sensible time of preventive replacement prediction and outperformed the alternative forecasting system.