On confidence interval-based anomaly detection approach for temperature predictions of wind turbine drivetrains to assist in lifetime extension assessment

With the number of wind turbines being installed increasing, due to the commitment of a large number of countries investing more in renewable energy, an informative method to determine when a drivetrain is coming to the end of it’s life can be extremely useful. This paper investigates the uncertainty of an output of a methodology used for life extension evaluation of a generator bearing in the drivetrain. A method has been developed to determine when the non-drive end generator bearing is coming to the end of it’s life, based upon temperature data extracted from seven years of 10-minute averaged SCADA data. Data from Kelmarsh wind farm was used, which consists of six onshore 2.05 MW Senvion MM92 wind turbines. A number of parameters from the SCADA data are used as the inputs for the model, in order to predict the component temperature and then in turn determine a threshold value, in which if the component’s temperature passes, indicates that it is reaching the end of it’s life. Due to the consequences that can occur if a component fails, such as loss of power, cost of repair etc. it is extremely important for the model to be as accurate as possible by taking into account any error or uncertainty. Other than the uncertainty of the measurements recorded in the SCADA data, which may be due to noise and/or sensor failure, the other major source of uncertainty comes from the predictive machine learning model that has been developed. Therefore, the model uncertainty is evaluated by a sensitivity analysis, where the input parameters are changed to see how much the output changes. The contribution of this work has investigated the error propagated in the component’s remaining life, that have originated from the uncertainty of the machine learning model, as well as the model input parameters/data. The results show that the error arising from the machine learning model and the input data, should fall within a certain range in order to obtain the level of accuracy of the methodology.