Automatic alarm setup using extreme value theory

Abstract

In the present paper we outline and demonstrate the application of a more robust solution to monitoring the health state of rotating machinery. In order to avoid triggering false alarms, the individual operational conditions are identified using unsupervised learning system and the threshold levels are set for each condition indicator. The data are also screened for anomalous outliers by using data quality assessment techniques.Warning and alarm thresholds are determined by applying Generalised Extreme Value (GEV) theory to an established baseline for each operational condition. GEV theory does not make any a priori assumption about the distribution when the baseline statistics are established unlike other statistical techniques for deriving limit thresholds.The applicability and the benefits of the proposed approach are demonstrated using three data sets containing both process parameters and vibration measurements from industrial power generation plants. The results show that determining alert and alarm thresholds alongside the relevant operational conditions provides a more robust indication of the machine components health state compared to using only uniform threshold levels.