Predictive repair scheduling of wind turbine drive-train components based on machine learning

Abstract

We devise a methodology to predict failures in wind turbine drive-train components and quantify its utility. The methodology consists of two main steps. The first step is the set up of a predictive model for shutdown events, which is able to raise an alarm in advance of the fault-induced shutdown. The model is trained on data for shutdown events retrieved from the alarm log of an offshore wind farm. Here, it is assumed that the timely prediction of low-severity events, typically caused by abnormal component operation, allows for an intervention that can prevent premature component failures. The prediction models are based on statistical classification using only supervisory control and data acquisition (SCADA) data. In the second step, the shutdown prediction model is combined with a cost model to provide an estimate of the benefits associated with implementing the predictive maintenance system. This is achieved by computing the maximum net utility attainable as a function of the model performance and efficiency of intervention carried out by the user. Results show that the system can be expected to be cost-effective under specific conditions. A discussion about potential improvements of the approach is provided, along with suggestions for further research in this area.