Prognosis of wind turbine gearbox failures by utilising robust multivariate statistical techniques

In this paper we present a new methodology for the prognosis of a wind turbine gearbox. The statistically robust Mahalanobis distance was used to determine multivariate outliers within low frequency SCADA data without the need for manual labelling. Domain knowledge (meta-knowledge) was used to determine the multivariate vectors which encapsulate the condition of the wind turbine gearbox, providing a means to model anomalous gearbox behaviour whilst quantifying the severity of a monitored fault. A prognostic horizon of over 146 days was achieved using a new 3 degrees of freedom model, with a strong trend observed within the presented prognostic. This allowed for the quantification of fault severity, an estimation of the rate of fault development and also a means to quantify the quality and effectiveness of maintenance. In order to reduce noise inherent within SCADA data, an expert system was developed to transform the prognostic capability into actionable intelligence. This reduced the potential cognitive load placed upon the maintenance operator, whilst providing the knowledge required to optimise available maintenance resources. Due to the statistically robust nature of the approach, no gearbox fault data was required for training, enabling prognostic capability without the capital expense incurred through destructive testing. Furthermore, no additional capital expenditure is required due to data being collected from the pre-existing SCADA system available on all of the latest generation of wind turbines.