Real-time novelty detection of an industrial gas turbine using performance deviation model and extreme function theory

Novelty detection is crucial to ensure the availability and reliability of an industrial gas turbine. With the application of modern health monitoring systems, there is an ample amount of data gathered from gas turbines, however they are usually from normal events with limited knowledge of any novelty. In current practice, the unknown event is detected by comparing with a model of normality through pointwise approaches, which is inefficient in terms of false alarms or missing alarms. This paper proposes an accurate novelty detection approach using performance deviation model and extreme function theory. The model is established from the multi-sensor real-time performance data. Outputs of the model, that is, the deviation curves, are considered as functions instead of individual data points to test the status of the system as ‘normal’ or ‘abnormal’ by the extreme value theory. The effectiveness of the proposed approach is demonstrated by the monitoring data from a single shaft gas turbine on site. Compared with other traditional methods, the proposed approach is superior in terms of high detection accuracy and high sensitivity with a good balance between the false alarm rate and missing alarm rate. This paper provides a reliable approach for the real-time health monitoring of the industrial gas turbines.